content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Indentation has been one of the major techniques towards identifying mechanical behavior of a vast list of materials~\citep{oyen2010handbook}. As a technique, it may be labeled as a direct probe, since there is an application of an explicit mechanical force towards identifying a deformation. Moreover, compared with tension/compression tests, nanoindentation is ultra-local and less invasive for the material shape and properties. Over the last few decades, indentation has developed into an important means to interrogate size dependent plasticity. At the microscale, the indentation size effect has been primarily attributed to the plastic strain gradients that are induced by non-uniform deformation modes~\citep{nix1998indentation}, similar to strain gradients in bending and torsion. These strain gradients have been found to be the essential reason behind the conclusion that ``smaller is harder''~\citep{aifantis1999strain, pharr2010indentation}. Several strain gradient plasticity theories~\citep[e.g.,][]{Hutchinson-SGP, SGP-Gao, Huang2004} have been formulated that can capture the size effects in indentation and many other inhomogeneous
deformation processes~\citep{wei1997steady,huang2000study,particleSGP,3D-rough}.
However, compression experiments on micro and nanopillars uncovered that size effects are present even in the absence of strain gradients~\citep{Uchic03,Uchic04, Dimiduk2005,Dimiduk2007, Greer2005}. Such size effects are responsible for strengthening at the nanoscale, and are shown to have various origins, such as dislocation source length truncation~\citep{Parthasarathy2007}, dislocation starvation~\citep{Greer2005, Greer2006} and single-arm dislocation sources~\citep{oh2009situ,cui2014theoretical}.
Even though size effects on strength have been the major focus of nanoscale mechanical deformation, it has also become clear that the mechanical response is intermittent and that these too display size effects~\citep{Dimiduk2005,NG2008, Papanikolaou2011, Papanikolaou2012,Papanikolaou2017}. Bursts of plastic deformation~\citep{papanikolaou2017avalanches,sethna2016deformation,maass2017micro} are observed either as stress drops (when imposing displacement control) or strain jumps (when imposing load control). In load controlled indentation, intermittency is also known as ``pop-in"~\citep{Lorenz2003, Bradby2004}, and in recent years, the first pop-in event in relatively pristine crystals~\citep{Durst2006,Shim2008} has been identified to mark the transition from elasticity to plasticity. Following the trend of size effects at the nanoscale, the equivalent stress for the first pop-in is also size dependent: \cite{Morris2011} proposed a statistical model for pop-in during nanoindentation that explained the size-dependence of pop-in stresses in accordance with experimental results. Moreover, \cite{Xia2016} extended pop-in investigations into multiple types of pop-in modes: a primary pop-in with a large displacement excursion and a number of subsequent pop-ins with comparable and small displacement excursions. The distribution of pop-ins in single FCC crystals has also been investigated to
some degree~\citep{wang2012size} for Cu single crystals, finding that the distribution can be fitted by a Gaussian with mean and variance that depend on grain orientation.
It is common in indentation testing that the sample is kept on a stage without any pre-existing applied stresses. However, pre-application of stress on the sample has been used to detect the effect of residual stresses on sample properties~\citep{Suresh1998, Swadener2001, Lee2004, Jang2003}. Early on, \cite{Tsui1996} and \cite{Bolshakov1996} have revealed the effect of elastic tensile/compressive pre-stress on the measured hardness in microindentation by both experiments and FEM simulations. The increase of tensile pre-stress results in a decrease of the measured hardness up to 6\%, while the effect of compressive pre-stress is barely seen. Moreover, finite element simulations (Bolshakov et al., 1996) suggested that the hardness dependence on pre-stress might be due to the experimental analysis procedures underestimating the indentation contact area. In this paper, we ask the question whether these effects of pre-stress persist for nanoindentation and the pop-in noise?
Commonly, in nanoindentation experiments, due to nanoscale surface roughness, oxide layers, and other defect-like surface features, the regime of depths below 50nm is completely neglected. However, modern nanoindenters in a relatively noiseless environment, only contain 1-2nm depth resolution. Moreover, it is natural to believe that high throughput nanoindentation experiments on a material surface could generate data that may average out extrinsic surface features. Finally, surface polishing techniques have improved to the degree that modern samples may have atomic scale roughness and no other defect-like surface features~\citep{Barnoush2008}. Anyway, it is relatively common that the hardness value as provided by continuous stiffness estimates appears to diverge for depths below 20nm~\citep{li2002review}. Is it possible that this divergence corresponds to a nanomechanical mechanism that emerges below 50nm? In association, it is also quite common that the first pop-ins, when sharp tips are used (corner-cube or Berkovich), appear at a similar scale (5-10nm). Is the first pop-in burst connected to the divergence of hardness below 20nm? Recently, Bei et al. (2016) described pop-in strength in different regimes: When the stressed volume is free of any pre-existing defects (at very small indentation depth), the pop-in corresponds to the homogeneous nucleation of dislocation at the theoretical strength. At the bulk limit (large indentation depth), the pop-ins are difficult to observe and the deviation from the elastic load–-displacement curves is governed by the bulk yield stress of the material. In the intermediate stage, the pop-in strength shows significant scattering.
There appear to be significant differences for nanoindentation between load and displacement control. Load controlled indentation is normally preferred for the relative simplicity in the experimental setup that does not require any electronic feedback loop~\citep{Vanlandingham2003, Minor2006}. In contrast, nanoindentation discrete dislocation dynamics simulations are most easily performed for displacement control, due to the advantage of not requiring a precise control of the contact area~\citep{Deshpande2005,indentation-Widjaja}.
Moreover, it is suspected that the hardness and pop-in statistics display differences between load and displacement control since \cite{Cui2016, Cui2017} showed significant differences through 3D discrete dislocation dynamics simulations of nanopillar compression. In that case study, the two different loading modes give different strengths and event (strain burst/stress drops) statistics. Ultimately, it is natural to inquire for analogous differences between various loading modes in nanoindentation.
In this paper, we perform 2D plane-strain discrete dislocation dynamics (2D DDD) simulations of nanoindentation~\citep{Deshpande2005, widjaja2006effect} to explore and understand the nature of ISEs at the nanometer scale, in relation to the dislocation microstructure. The main application of these simulations is thin crystalline films. We use 2D DDD to investigate nanoindentation features of a single crystalline sample by both load and displacement controlled protocols. Moreover, in order to identify the sensitivity of ISEs to various dislocation microstructures, constant tensile pre-stress is assigned prior to indentation. This work is focused on an exhaustive statistical exploration of initial conditions in order to unveil the statistical character of the indentation hardness and pop-in effects. Overall, while 2D-DDD methodology appears restrictive, this is a systematic dislocation plasticity study of nanoindentation that focuses on the effect of pre-stress and different loading modes that should guide precise 3D simulations and also experiments.
The remainder of this paper is organized as follows: Section 2 describes the methodology of our 2D-DDD model for indentation. Section 3 explains the load controlled protocol and presents a comparison with the usual displacement controlled indentation~\citep{indentation-Widjaja}. Section 4 presents simulation results of indentation hardness at different indentation depths, both for load controlled and displacement controlled protocols. In addition, we focus on the effect of pre-stress on hardness. The effect of pre-stress on the pop-in statistics is discussed in Section 5. In Section 6, we discuss and summarize our results.
\section{Description of the problem}
\label{description}
The model problem is shown in Fig.~\ref{fig:schematic}. A 2D crystal sample with width $1000\mu \rm{m}$ and height $50\mu \rm{m}$ is indented by a rigid circular indenter of radius $R$ at the middle of the sample. Tensile stress with magnitude $\sigma_{xx}$ is assigned at two ends of the sample prior to indentation. For large enough $\sigma_{xx}$, the system will generate dislocations, and the indentation simulation is started once the total number of dislocations in the sample is relatively stable.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{schematic-new}
\caption{Schematic of indentation problem. The zoom in shows three slip systems (with different colors), red dots stand for dislocation sources while blue dots present dislocation obstacles.}
\label{fig:schematic}
\end{figure}
The finite element mesh is highly refined in the region close to the indenter with the element segment length being 0.005$R$ to accurately capture the evolution of the contact area. Plasticity is limited in a window with size $w_{p}=20\mu \rm{m}$ and $h_{p}=10\mu\rm{m}$ for computational efficiency; once a dislocation reaches the border, the simulation is stopped. Within the plastic window, three slip systems are considered, with slip directions at $\pm 30^\circ$ and $90^\circ$ respectively, relative to the horizontal direction. Slip planes are spaced at 10$b$ relatively to each other where $b=0.25\rm{nm}$ is the magnitude of the Burgers vector. Dislocation sources (indicated by red dots in Fig.~\ref{fig:schematic}) and obstacles (indicated by blue dots) are randomly distributed on the slip planes.
Plastic deformation of the crystalline sample is described using the discrete dislocation framework for small strains developed by \citep{vandergiessen1995},
where the determination of the state in the material employs superposition.
As each dislocation is treated as a singularity in a linear elastic background solid with Young's modulus $E$ and Poisson's ratio $\nu$,
whose analytical solution is known at any position, this field needs to be corrected by a smooth image field $(\hat{\ })$ to ensure that actual boundary conditions are satisfied.
Hence, the displacements $u_i$, strains $\varepsilon_{ij}$, and stresses $\sigma_{ij}$ are written as
\begin{equation}
\label{eq:superposition}
u_i = \tilde{u}_{i}+\hat{u}_{i}, \; \varepsilon_{ij} = \tilde{\varepsilon}_{ij}+\hat{\varepsilon}_{ij}, \; \sigma_{ij} = \tilde{\sigma}_{ij}+\hat{\sigma}_{ij},
\end{equation}
where the ($\tilde{\ }$) field is the sum of the fields of all $N$ dislocations in their current positions, i.e.
\begin{equation}
\label{eq:dislocation-field}
\tilde{u}_i=\sum_{J=1}^{N}\tilde{u}_i^{(J)}, \; \tilde{\varepsilon}_{ij}=\sum_{J=1}^{N}\tilde{\varepsilon}_{ij}^{(J)}, \; \tilde{\sigma}_{ij}=\sum_{J=1}^{N}\tilde{\sigma}_{ij}^{(J)}.
\end{equation}
The image fields, indicated by the superposed $\hat{\ }$, are smooth and are obtained by solving a linear elastic boundary value problem using finite elements with the boundary conditions changing as the dislocation structure evolves under the application of mechanical load.
At the beginning of the calculation (before pre-stressing the sample), the crystal is stress and dislocation free. This corresponds to a well-annealed sample, yet with pinned dislocation segments left that can act either as dislocation sources or as obstacles. Dislocations are generated from sources when the resolved shear stress $\tau$ at the source location is sufficiently high ($\tau_{\rm nuc}$) for a sufficiently long time ($t_{\rm nuc}$).
We only consider bulk sources. The bulk sources are randomly distributed over slip planes at a density $\rho_{\rm{nuc}}$, while their strength is selected randomly from a Gaussian distribution with mean value $\bar{\tau}_{\rm nuc} = 50$ MPa and 10\% standard deviation. The sources are designed to mimic the Frank-Read mechanism in two dimensions~~\citep{vandergiessen1995}, such that they generate a dipole of dislocations at distance $L_{\rm nuc}$, when activated. The initial distance between the two dislocations in the dipole is
\begin{equation}
\label{eq:L_nuc}
L_{\rm nuc}= \frac{E}{4\pi(1-\nu^2)}\frac{b}{\tau_{\rm nuc}},
\end{equation}
at which the shear stress of one dislocation acting on the other is balanced by the local shear stress. Once nucleated, a dislocation can either exit the sample through the traction-free sides, annihilate with a dislocation of opposite sign when their mutual distance is less than $6b$ or become pinned at an obstacle. Point obstacles are included to account for the effect of blocked slip caused by precipitates and forest dislocations on out-of-plane slip systems that are not explicitly described. The strength of the obstacles $\tau_{\rm obs}$ is taken to be $150$ MPa with 20\% standard deviation. They are randomly distributed over the slip planes with a
density that is eight times the source density~\citep{Papanikolaou2017}.
Dislocations are considered to move by glide only, driven by the component of the Peach-Koehler force in the slip direction. For the $I$th dislocation, this force is given by
\begin{equation}
\label{eq:P-K}
f^{(I)} = \boldsymbol{n}^{(I)}\cdot\left(\boldsymbol{\hat{\sigma}}+\sum_{J\neq I}{\boldsymbol{\tilde{\sigma}}}^{(J)}\right)\cdot\boldsymbol{b}^{(I)},
\end{equation}
where $\boldsymbol{n}^{(I)}$ is the slip plane normal and $\boldsymbol{b}^{(I)}$ is the Burgers vector of dislocation $I$.
This force will cause the dislocation $I$ to glide, following over-damped dynamics, with an instantanoeus velocity
\begin{equation}
\label{eq:B-v}
v^{(I)} = \frac{f^{(I)}}{B},
\end{equation}
where $B$ is the drag coefficient. In this paper, its value is taken as $B=10^{-4}$Pa s, which is representative for aluminum. A dislocation stays pinned until its Peach-Koehler force exceeds the obstacle-dependent value $\tau_{\rm obs}b$. Each sample contains a random distribution of bulk dislocation sources as well as dislocation obstacles; for each parameter study, we therefore studied 20 realizations.
The simulation is carried out in an incremental manner, using a time step that is a factor 20 smaller than the nucleation time $t_{\rm nuc}=10\:$ns. At the beginning of every time increment, nucleation, annihilation, pinning at and release from obstacle sites are evaluated. After updating the dislocation structure, the new stress field in the sample is determined, using the finite element method to solve for the image fields~\citep{vandergiessen1995}.
\section{Nanoindentation with load control protocol}
In displacement controlled indentation, plasticity is characterized by the drop of the indentation force, as seen in previous DD simulations~\citep{balint2006discrete, widjaja2007contact}. Contact between rigid indenter and crystal is sticky. The contact area is
computed as the total length of all finite elements that are in contact, like in~\cite{widjaja2007contact}.
In indentation experiments that are typically carried out using load control, plasticity is characterized by displacement bursts (pop-in). A real load controlled indentation simulation needs a contact algorithm to be implemented between indenter and sample. In order to avoid the numerical complexity associated with this, in this paper, we propose a hybrid loading protocol to mimic the load control protocol used in experiments. The hybrid protocol is essentially displacement controlled based, but with a feedback loop to the indentation rate in order to maintain the indentation force until the indentation force increases. As in load controlled indentation, the position of indenter is adjusted and the indentation force never drops. The assumption in our hybrid protocol is that the indenter can move as fast as needed in order to maintain the load level. In order to capture indentation force drops, the displacement rate is chosen as 0.004 m$\rm{s}^{-1}$ which is much smaller than the one (0.1 m$\rm{s}^{-1}$) used in previous indentation studies~\citep{balint2006discrete, widjaja2007contact}. This is done to avoid the suppression of force drops by very high displacement rates, see Appendix for more details.
A comparison between load controlled indentation and displacement controlled indentation is shown in Fig.~\ref{fig:load-control}.
\begin{figure}[ht!]
\centering
\subfigure[]{
\includegraphics[width=0.45\textwidth]{Fig3a}
}
\subfigure[]{
\includegraphics[width=0.45\textwidth]{Fig3b.png}
}
\caption{Comparison between load controlled protocol and displacement controlled protocol with dislocation source density $\rho_{\rm nuc}=75/\mu\rm{m}^2$. Two different protocols have the same initial dislocation structure (dislocation sources and obstacles), (a) indentation force versus indentation depth, (b) evolution of the total number of dislocations.}
\label{fig:load-control}
\end{figure}
It can be seen in Fig.~\ref{fig:load-control}(a), using load controlled protocol gradually gives rise to a larger indentation force. The reason is that our load control protocol uses a higher displacement rate when the indentation force tends to drop and that a higher displacement rate yields a higher indentation force. The inset is the zoom in of the curve between indentation depth 15 nm and 20 nm. It can be clearly seen that plasticity is presented as displacement burst under load control while it is force drop under displacement control. The difference on the total number of dislocation can be seen in Fig.~\ref{fig:load-control}(b). At small indentation depth, the two protocols have more or less the same number of dislocations, but due to the changing displacement rate (a higher rate), application of the load control protocol gradually leads to fewer dislocations.
Details on the effect of the load control protocol are discussed in the Appendix. The different loading protocols are also discussed through nano pillar compression, a problem where the contact area stays constant.
\section{Dependence of indentation hardness on pre-stress}
\label{hardness-dependence}
\cite{Tsui1996} observed that a tensile pre-stress below the yield strength has a very small influence on the measured hardness in micron-indentation. In this study, we therefore extend tensile pre-stress up to the yield strength to see the influence of pre-stress on measured hardness in nano-indentation.
The yield strength of the sample with given dislocation parameters ($\rho_{\rm nuc}$, $\rho_{\rm obs}$, $\tau_{\rm nuc}$, $\tau_{\rm obs}$) can be computed by either a force or displacement controlled tensile (or compression) simulation (see Appendix). A simple estimate can be obtained by first noting that for the chosen slip system orientations (cf. Fig. 1), the ratio between the critical resolved shear stress $\tau_{\rm Y}$ and the initial tensile yield stress is
$\tau_{\rm Y}/ \sigma_{\rm Y} = \frac{1}{2} \rm{sin60}^{\circ}$.
The critical resolved shear stress is controlled by the weak dislocation sources in the Gaussian distribution with mean value $\bar\tau_{\rm nuc} = 50$ MPa and 10\% standard deviation. Assuming $\tau_{\rm Y}$ to be somewhere between one and a half standard deviation below the mean source strength, we estimate a yield strength of about 100 MPa. This indeed is the value found for the simulations with $\rho_{\rm nuc}=60\:\mu\mbox{m}^{-2}$ in the Appendix. For the lowest source density considered in this study, $\rho_{\rm nuc}=15\:\mu\mbox{m}^{-2}$, the yield strength is likely to be somewhat larger, whereas for $\rho_{\rm nuc}=75\:\mu\mbox{m}^{-2}$ and $150\:\mu\mbox{m}^{-2}$ it may be somewhat lower. Some strain hardening caused by the high density of strong obstacles ensures a stable response up to $\sigma_{\rm Y}= 100$ MPa for the entire range of parameters used. As an additional check we have verified that even for the softest samples, the stress inside the plastic window prior to indentation is equal to the applied pre-stress.
We start indentation using the load control protocol with dislocation source density $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$ (with $\sigma_{\rm{Y}}$=100 MPa). For the same sample (i.e., same dislocation source/obstacle distribution), the indentation force (for indenter radius $R=1\mu\rm{m}$) versus indentation depth ($P-h$ curve) with different pre-stresses $\sigma_{xx}$ is shown in Fig.~\ref{fig:P-h}. We notice that increasing the pre-tensile stress results in a decrease in the indentation force, as one should expect from the enhanced resolved shear stress argument.
\begin{figure}[h!]
\centering
\includegraphics[width=0.45\textwidth]{phs.png}
\caption{Examples of load versus depth curves obtained under load controlled indentation of a sample with different pre-tensile stresses. $\sigma_{Y}$ is taken as 100 MPa.}
\label{fig:P-h}
\end{figure}
\subsection{Effect of pre-stress on hardness}
Indentation hardness is defined as the indentation force $P$ divided by the real contact area $a$ which we calculate following~\cite{widjaja2007contact}. Fig.~\ref{fig:hardness}(a) shows examples of hardness versus indentation depth at various levels of tensile pre-stress. It can be seen that the hardness increases with indentation depth at small indentation depth just as observed without pre-stress~\citep{swadener2002correlation, widjaja2006effect}.
It is important here to note the experimentally confirmed difference between
a wedge (Berkovich in 3D) and a circular (spherical in 3D) indenter:
the ISE is controlled by indentation depth in Berkovich indentation whereas it is controlled by the indenter radius in spherical indentation. Moreover, the hardness actually increases with
increasing depth for a spherical indenter. Detailed discussions can be found in~\citep{swadener2002correlation, widjaja2006effect}.
With different magnitudes of pre-stress, Fig.~\ref{fig:hardness}(b) shows the hardness measured at different indentation depths. For the same indentation depth, we see that pre-stress reduces the indentation hardness; a small effect at low pre-stress but increasingly more at higher pre-stress levels. In fact, there appears to be a hardness transition at pre-stresses of around 50 MPa (the mean dislocation source strength). Moreover, the effect of pre-stress changes with indentation depth, as can be clearly seen in Fig.~\ref{fig:hardness}(c) where the hardness is normalized by the hardness $H_{0}$ with no pre-stress. We see that the hardness decreases with increasing pre-stress at small indentation depth, however, the trend becomes less obvious for larger indentation depth.
\begin{figure}[ht!]
\centering
{
\includegraphics[width=1\textwidth]{density-R1}
}
\caption{(a) examples of Hardness versus indentation depth for three different tensile pre-stresses obtained by load controlled protocol with dislocation source density $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$. (b) Indentation hardness at different indentation depths for different pre-stresses, (c) hardness normalized by hardness with 0 pre-stress. Error bars represent standard deviation of 20 realizations.}
\label{fig:hardness}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=1\textwidth]{stress-field-density75}
\caption{Shear stress filed of indentation under different tensile pre-stresses. (a), (b), (c) are for zero tensile pre-stress, (d), (e), (f) are the results with tensile pre-stress 100 MPa. Contour plots are for different indentation depth as indicated between figures, (a) (d): 2 nm, (b) (e): 10 nm, (c) (f): 20 nm. The thick dashed line indicates the indentation center.}
\label{fig:dislocation-structure}
\end{figure}
The influence of the pre-stress can also be revealed through the stress field under the indenter, shown in Fig.~\ref{fig:dislocation-structure}. Figures~\ref{fig:dislocation-structure}(a), (b), (c) give the stress field at different indentation depths when the tensile pre-stress is zero. We see that with increasing depth, dislocations get nucleated and a high-stress area extends inside the sample. By contrast, when the tensile pre-stress is 100 MPa, shown in Fig.~\ref{fig:dislocation-structure}(d), (e), (f), a high-stress area has been induced by tension (because of nucleated dislocations). Increasing the indentation depth changes the stress field, but not as much as in the case of zero pre-stress.
\subsection{Effect of dislocation source density}
\label{densityeffect}
It has been reported in many previous 2D DDD studies of different problems~
\citep{widjaja2006effect, Papanikolaou2017} that the density of dislocation sources can restrict plasticity if the system does not have a surplus of dislocation sources. The computations at a source density $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$ reported above, reveal an effect of pre-stress on hardness. In this section, we report results for different dislocation source densities: $\rho_{\rm{nuc}}=15/\mu\rm{m}^2$ and $\rho_{\rm{nuc}}=150/\mu\rm{m}^2$. It is seen in the inset of Fig.~\ref{fig:density}(a) that the lower source density gives rise to the larger hardness at the same indentation depth, just like in the previous 2D DDD indentation study by \cite{widjaja2006effect}. Moreover, it is worthy to note that for the lower source density (Fig.~\ref{fig:density}(a)), the effect of pre-stress on hardness is still present at an indentation depth of 20 nm while the effect disappears for the highest source density (Fig.~\ref{fig:density}(b)).
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{density15-density150}
\caption{Normalized hardness (by hardness with 0 pre-stress) versus tensile pre-stress obtained by load controlled protocol with different dislocation source densities. (a) $\rho_{\rm{nuc}}=15/\mu\rm{m}^2$, (b) $\rho_{\rm{nuc}}=150/\mu\rm{m}^2$. The inset in (a) is the hardness with zero pre-stress at different depth for two different dislocation source density. Figure (a) and (b) share the same legend shown in (b). }
\label{fig:density}
\end{figure}
\subsection{Effect of indenter radius}
\label{radius}
The results reported so far are for an indenter radius $R=1\mu\rm{m}$; in Fig.~\ref{fig:radius} we present the results for a smaller indenter $R=0.5\mu\rm{m}$ and for a larger indenter $R=5\mu\rm{m}$. It can be seen in the inset of Fig.~\ref{fig:radius}(a) that at the same indentation depth, the smaller indenter gives rise to the larger hardness. This is consistent with the typical spherical (circular) indentation size effect~\citep{swadener2002correlation,widjaja2006effect}. Furthermore, we see that for a larger indenter, the hardness transition (decreasing hardness with increasing tensile pre-stress) disappears at a smaller indentation depth (compare Fig.~\ref{fig:radius}(a) and Fig.~\ref{fig:radius}(b)).
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{density75-different-R}
\caption{Normalized hardness (by hardness with 0 pre-stress) versus tensile pre-stress obtained by load controlled protocol with dislocation source density $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$, but different indenter radii. (a) $R=0.5\mu\rm{m}$, (b) $R=5\mu\rm{m}$.
The inset in (a) is the hardness with zero pre-stress at different depth for two different radii. Figure (a) and (b) share the same legend shown in (b).}
\label{fig:radius}
\end{figure}
\subsection{Comparison with displacement control}
\begin{figure}[ht!]
\centering
\subfigure[]{
\includegraphics[width=0.45\textwidth]{Fig-displacement.png}
}
\subfigure[]{
\includegraphics[width=0.45\textwidth]{displacement-prestress.png}
}
\caption{(a) Indentation force versus depth under displacement control with dislocation source density $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$, (b) effect of pre-stress on normalized hardness. $\sigma_{\rm Y}$ is taken as 100MPa.}
\label{fig:displacement-curves}
\end{figure}
Fig.~\ref{fig:displacement-curves}(a) shows the indentation force versus depth for displacement controlled indentation under different pre-stresses.
It is worthy to note that under displacement controlled indentation, plasticity is characterized by a drop of indentation force, while under load control, it is characterized by a jump (or burst) of the indentation depth.
Even though the difference in loading mode leads to the difference of indentation force depth curves, the effect of pre-stress on hardness is similar, see Fig.~\ref{fig:displacement-curves}(b).
\section{Pop-in statistics}
While the emphasis so far has been on the effect of pre-stress on the indentation hardness, in this section, we try to unveil the effect of pre-stress through the statistics of those signals.
In annealed crystals, the burst of indentation depth appearing during load control is normally called pop-in, and the first pop-in relates to the nucleation of dislocations~\citep{Morris2011}. Using 3D DDD, \cite{fivel-indentation} related pop-in to the fast multiplication of dislocations below the indenter. When the dislocation network formed by fast dislocation multiplication is broken by indentation, pop-in takes place. In our model, pre-stress can induce dislocations prior to indentation, in which case the burst is caused by those pre-stress induced dislocations, not by dislocation nucleation.
Fig.~\ref{fig:dislocation-density}(a) shows the average dislocation density $\rho_{\rm dis}$ over the entire plastic window (200$\mu\rm{m}^2$) at different indentation depths. It is seen that larger pre-stress introduces some dislocations, but the dislocation density becomes almost the same at larger indentation depth irrespective of the pre-stress level. In order to see the connection between pop-ins and the nucleation of dislocations, we plot the rate of change of the dislocation density during indentation in Fig.~\ref{fig:dislocation-density}(b). It can be clearly seen that at low pre-stress levels, the growth rate of dislocation density is higher, which means that dislocation nucleation is dominant. At high pre-stresses, plasticity is mainly dominated by pre-stress induced dislocations.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{density-evo}
\caption{(a) Evolution of dislocation density. (b) change of total number of dislocations. $\rho_{\rm dis}^{0}$ is the initial dislocation density. For large pre-stress, $\rho_{\rm dis}^{0}$ is determined by the pre-stress, for small pre-stress (no pre-stress induced dislocations), $\rho_{\rm dis}^{0}=0.01/\mu\rm{m}^2$. Legend indicates the magnitude of the pre-stress where $\sigma_{\rm Y}=100$MPa.}
\label{fig:dislocation-density}
\end{figure}
The most straightforward way to unveil pop-in statistics is to analyze the displacement burst or the force drop in the indentation force-depth curve. The definition of a pop-in event is indicated in Fig.~\ref{fig:pop-in-definition} on the response curves shown in the inset of Fig.~\ref{fig:load-control}(a).
\begin{figure}[ht!]
\centering
\includegraphics[width=0.45\textwidth]{events-definition}
\caption{Definition of event: displacement burst in load controlled indentation, force drop and equivalent displacement burst in displacement controlled indentation.}
\label{fig:pop-in-definition}
\end{figure}
Fig.~\ref{fig:pop-in-statistics} shows statistics of pop-in of samples under load control. Here, $\eta$ is the the magnitude of a single indentation depth burst, while $P(\eta)$ is the probability density.
\begin{figure}[b!]
\centering
\includegraphics[width=0.9\textwidth]{statistics-load-control}
\caption{Probability event distribution $P(\eta)$ as function of event size $\eta$ under different pre-stresses for different material properties and indenter size (a) $\rho_{\rm{nuc}}=150/\mu\rm{m}^2$, $R=1\mu \rm{m}$; (b) $\rho_{\rm{nuc}}=15/\mu\rm{m}^2$,
$R=1\mu \rm{m}$; (c) $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$, $R=0.5 \mu \rm{m}$; (d) $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$, $R=5\mu \rm{m}$. $\sigma_{Y}$ is taken as 100 MPa.}
\label{fig:pop-in-statistics}
\end{figure}
We see that for small pre-stress (up to 0.4$\sigma_{\rm Y}$, i.e., 40 MPa), pop-in has power law statistics spanning two orders of magnitude, except for very small $\eta$,. However, at a pre-stress $\ge 50$ MPa (i.e., the average dislocation source strength), dislocations induced by pre-stress yield higher probability for very small $\eta$ and at the same time, they give rise to higher probability for bigger $\eta$. This means that pre-stress induced dislocations not only introduce noise in statistics, but also contribute to big plasticity events.
Under the displacement controlled protocol, the signal of plasticity is a drop of the indentation force as defined in Fig.~\ref{fig:pop-in-definition}. Statistics of force drops $S$ are shown in Fig.~\ref{fig:force-drop}(a). Here we barely observe an effect of pre-stress except for very small force drops where pre-stress induced dislocations leads to more noise. We also define an equivalent indentation depth jump as the distance between the point where the indentation force starts to drop to the point where the force increases back to its initial value (Fig.~\ref{fig:pop-in-definition}). Statistics of the equivalent depth bursts are shown in Fig.~\ref{fig:force-drop}(b). We see that larger pre-stress can result in larger indentation depth bursts, which is consistent with the statistics for the load controlled protocol shown in Fig.~\ref{fig:pop-in-statistics}. The power law statistics spans roughly four orders of magnitude with an exponent $-1.86$.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{statistics-displacement}
\caption{(a) Statistics of indentation force drops, (b) statistics of equivalent indentation depth burst, for a material with source density $\rho_{\rm{nuc}}=75/\mu\rm{m}^2$ and indenter radius $R=1\mu \rm{m}$. The black solid line is a power law fit. $\sigma_{Y}$ is taken as 100 MPa. }
\label{fig:force-drop}
\end{figure}
\section{Discussion and Conclusions}
The main focus of this paper is the effect of pre-stress on the indentation response of single crystals. One important feature observed in section~\ref{hardness-dependence} is that there is a hardness transition (decreasing hardness with increasing pre-stress) depending on pre-stress and indentation depth. This phenomenon is consistent with observations in our recent experimental study~\citep{scripta}.
Tensile pre-stress basically has two effects on indentation: (i) the stress field reduces the indentation force that is needed to nucleate dislocations, (ii) depending on the magnitude of the pre-stress, it may have introduced pre-existing dislocations prior to indentation. These dislocations got pinned in tension and become mobile in indentation because of the highly concentrated stress field induced by indentation. \cite{Tsui1996} have shown that if the pre-stress is within the elastic regime, there is almost no effect on the indentation hardness measured at a depth of hundreds of nanometers. Consistent with this, the effect of pre-stress found in this paper originates from the pre-stress induced dislocations.
The effect becomes less obvious or disappears at the depth where tension induced dislocations no longer dominate the indentation response, i.e., when dislocations nucleated by indentation dominates.
For the same dislocation source strength distribution, dislocation nucleation depends on the high stress volume (area, in 2D) and the probability of having dislocation sources in the volume (area). The former depends on indenter geometry and indentation depth, while the latter depends on the source density. It is natural to expect that a smaller indenter (smaller stress volume) and lower dislocation density (lower probability of having sources) shift the transition from pre-existing dislocation-dominated to dislocation nucleation-dominated behavior to larger indentation depths, as observed in sections~\ref{densityeffect} and~\ref{radius}. Fig.~\ref{fig:all-hardness}(a) shows all computed hardness data for an indenter radius $R=1\mu\rm{m}$ (different indentation depth $h$ and different dislocation source density $\rho_{\rm{nuc}}$) reported in the above sections versus the tensile pre-stress and the dimensionless quantity $\rho_{\rm{nuc}}h^2$. The tensile pre-stress determines the number of dislocations prior to indentation while $\rho_{\rm{nuc}}h^2$ influences dislocation nucleation during indentation. The more dislocations are nucleated during indentation, the weaker the effect of pre-stress induced dislocations, and the transition of hardness fades. Fig.~\ref{fig:all-hardness}(b) shows, for indentation depth 10nm, normalized hardness as the function of pre-stress and dimensionless quantity $\rho_{\rm{nuc}}R^2$. It shows that increasing $\rho_{\rm{nuc}}R^2$ results in a weaker dependence of hardness transition on the magnitude of the pre-stress.
\begin{figure}[h!]
\centering
\includegraphics[width=1\textwidth]{hardness-trend}
\caption{$H/H_{0}$ as the function of pre-stress for different values of the dimensionless quantity $\rho_{\rm{nuc}}h^2$ at constant indented radius $R=1\mu\rm{m}$ in (a) and versus $\rho_{\rm{nuc}}R^2$ at constant indentation depth $h=10$nm in (b).}
\label{fig:all-hardness}
\end{figure}
In summary, the salient conclusions of this study are:
\begin{itemize}
\item Larger tensile pre-stress results in a smaller indentation hardness when the indentation depth is small.
\item The indentation depth that influences the role of pre-stress depends on indenter geometry and material properties (source strength). The effect of pre-stress depends on the competition between pre-stress induced dislocations and indentation nucleated dislocations.
\item Nanoindentation addresses source limited plasticity at small indentation depth and interrogates the strain hardening regime at larger depth.
\item Statistics for both load controlled and displacement controlled show that pre-stress introduces more {\it noise} and leads to more big events in indentation.
\end{itemize}
\section*{Acknowledgement}
This research was supported by the U.S. Department of Energy, Office of Sciences, Basic Energy Sciences, DE-SC0014109. We also acknowledge the use of the Super Computing System (Spruce Knob) at WVU, which are funded in part by the National Science Foundation EPSCoR Research Infrastructure Improvement Cooperative Agreement 1003907, the state of West Virginia (WVEPSCoR via the Higher Education Policy Commission) and WVU. This work also benefited greatly from the facilities and staff of the High Performance Computing Center at the Zernike Institute for Advanced Materials at the University of Groningen, the Netherlands.
\section*{References}
|
\section{Introduction}
\label{sec:intro}
The so-called power flow (or load flow) problem consists in finding the steady
state of a power network where sources and loads are specified in terms of
constant power. It is one of the cornerstones of power systems analysis, upon
which many other analyses and tools are built. Mathematically, the core problem
consists in a system of nonlinear algebraic equations for the bus voltages,
given simply by the power balance equations at every bus. There are however
additional controls, mostly dealing with voltage regulation, that need to be
taken into account. The most important and ubiquitous among these is voltage
regulation by synchronous generators, exerted either locally or at a remote
bus. Other important examples are under-load tap changers (ULTC) in
transformers, automatically switched shunt capacitor/reactor banks, HVDC line
controls, and modern FACTS devices. In all these cases the control device can be
represented as one or more additional \emph{equality constraint} equations
representing how the regulated magnitude is to be restricted; and
correspondingly, new variables (which can be thought of as the regulating
resources) are added to the system in order to keep a balance with between the
total numbers of equations and unknowns. The problem is that all controls have
resource limits: generators have reactive Mvar limits, ULTC have a limited
range, etc. When a regulating resource hits a limit, the behavior of the
control changes: the regulated magnitude is now allowed to vary from the
commanded setpoint, while the resource remains at its saturation value. This
switching behavior turns the power flow into a hard computational
problem, particularly for large and complex network models.
An early review by Stott~\cite{Stott74} on powerflow methods discusses the two
main approaches for enforcing limits, which are still used in most iterative
power flow engines today (for a more recent review,
see~\cite{GomezExpAlvarado09}). One straightforward method consists in
``type-switching'', where the equations change as needed: the controlling
parameter (e.g. the Q injection for a generator; or the tap ratio for an ULTC)
is initially an independent variable; but if, after solving the power flow, it
hits a limit, it becomes a fixed parameter and the corresponding constraint
equation is removed. This method requires solving different power flows as the
switching takes place. The other strategy is based on using ``feedback
adjustments''~\cite{AllanArruda82,ChangBrandwajn88}: the controlling resource is
considered always a parameter, but it is modified on-the-fly between iterations,
according to how the controlled voltage deviates from the setpoint. This has the
advantage that the equations remain fixed, so that the iterative scheme is
simpler.
However, both approaches face the problem of slow convergence due to
``rebounds'': when a given control saturates during an intermediate step of
these methods, it may later require to be reverted back to the unsaturated
state. This happens because of nonlinear interactions among controls,
particularly when several of them are close to their limits. Sometimes these
rebounds enter into cycles, preventing convergence~\cite{GomezExpAlvarado09}.
Many methods have been developed to avoid this sort of
interactions~\cite{ChangBrandwajn91,Vlachogiannis94}, but their justification
remains heuristic. To this day, power flow practitioners still have to rely on a
combination of personal experience and heuristics in order to deal with these
convergence problems~\cite{PowerWorld04,Siemens12}, which usually constitute a
time-consuming task. This state of affairs is somewhat surprising since
transmission planners always need to deal with peak cases, which are quite close
to feasibility boundaries and contain a large percentage of generators (more
than 30\% is not unusual) saturated at their Mvar limits.
There have been many studies characterizing the effects of limits in the
transition to voltage
collapse~\cite{DobsonLu1992,VuLiu92,HiskensChakrabarti96,YueVenkatasubramanian07,KataokaShinoda05,ZhaoZhang08},
but none of these have resulted in a practical power flow method. In order to
find further insights with actual application to the general power flow problem,
one needs to look into the related areas of Optimal Power Flow (OPF). Many
authors have approached the problem of calculating maximum loadability in the
presence of limits using OPF techniques~\cite{Irisarri97}, but it was
specifically the representation of control limits by means of
\emph{complementarity constraints}~\cite{Vournas00,RomanRosehart05,Rosehart05}
that triggered new approaches to the general power flow~\cite{Kataoka05,Zhao08},
most notably~\cite{Murray2015}. One key idea in these papers is the realization
that a power flow with control limits can be tackled as a constrained
optimization problem, even though its aim is not computing a loadability
limit. Of course this begs the question of what \emph{objective function} to
use, and in our opinion none of these works justify their choices
satisfactorily. Having said that, using at least \emph{some} minimization
criterion seems a more principled approach than just using traditional
adjustment heuristics or type-switching, since neither one of those techniques
allow for a characterization of the solution they arrive to. One should be aware
that there could be in general more than one valid power flow solution,
differing in how control devices end up either in their saturated or
un-saturated state.
This paper presents a general-purpose power flow method in which the problem of
enforcing control limits is solved rigorously, based on a minimization principle
that stems from the fundamental physics of electrical networks. The method is
developed for the case of Mvar limits in generators regulating voltage on both
local and remote buses. Since it is based on the framework of the Holomorphic
Embedding Loadflow Method (HELM)~\cite{Trias12,Trias15,TriasMarin16}, the method
also inherits the properties of being constructive and therefore produces
unequivocal results: it yields the correct operating solution when the case is
feasible, and it detects infeasibility otherwise.
The paper is structured as follows. Section~\ref{sec:theproblem} presents the
problem definition and lays down the nomenclature and conventions used in the
method. Section~\ref{sec:Lagrangian} develops a Lagrangian formulation of power
flow problem, starting from the nonlinear DC problem (not to be confused with
the linear ``DC'' approximation to AC problems) and extending it to lossless AC
networks. This provides the minimization criterion underlying the proposed
method, since it shows how to understand the complementarity constraints in the
framework of a constrained minimization procedure. Section~\ref{sec:HELMscheme}
then describes how the HELM scheme is implemented for this problem. In contrast
to barrier methods used in traditional nonlinear programming (such as Interior
Point Methods), the HELM technique is purely based on power series and their
analytic continuation. This section shows how to accommodate the constraint
equations and, more importantly, how to ensure that the embedding guarantees
that the HELM reference germ is consistent with reactive limits and voltage
setpoints. Having laid down the essentials of the method, Section~\ref{sec:PW}
then deals with a technical obstacle in the analytic continuation: the
unavoidable singularity found at $s=1$, which provokes a very slow convergence
of the Padé approximants. This non-trivial numerical problem has been overcome
thanks to a new technique based on certain changes of variables having the
property that they leave the embedded power flow equations formally invariant
under the change. This allows to construct the analytic continuation, still
based on Padé approximants, incrementally in stages, similarly to the classic
Weierstrass idea of analytic continuation along a path. Since this technique is
absolutely crucial to the success of practical implementations on a computer,
the method has been named after it. Finally, Section~\ref{sec:NumResults} shows
a sample of numerical results. Appendix Section~\ref{app:methsummary}
recapitulates all steps of the method in summarized form.
\section{The problem}
\label{sec:theproblem}
Let us establish the nomenclature by considering a general power system
comprised of constant-power injections $S_i=P_i+jQ_i$ and constant-current
injections $I_i$. The power flow equations are given by the current balance at
each bus,
\begin{equation}
\label{eq:PF}
\sum_{j=0}^n y_{ij} (V_i-V_j) + Y_i^\text{sh} V_i = I_i + \frac{S_i^*}{V_i^*}
\quad (i=1, \ldots, n),
\end{equation}
where $V_i$ are bus voltages, $y_{ij}$ are branch admittances, and
$Y_i^\text{sh}$ are shunt admittances to ground, which are assumed to group
contributions from line charging susceptances, shunts from transformer modeling,
shunt reactor/capacitor banks, and any loads modeled as a constant
admittance. As it is commonplace in power systems, the \emph{active} sign
convention will be used here, so that generators inject positive current $I_i$
and power $P_i$, while loads inject negative values for those same variables.
The network is composed of a total of $n+1$ buses where, without loss of
generality, a swing bus is assumed at index 0 (and a swing voltage $V_0$ should
be specified). It is also customary to define the transmission admittance matrix
as:
\begin{equation*}
Y_{ij} \equiv
\begin{cases}
\sum_{k=0}^n y_{ik} &\text{if } i=j \\
-y_{ij} & \mbox{if } i \ne j
\end{cases}
\quad ,
\end{equation*}
so that the powerflow equations can be expressed in a more compact form:
\begin{equation}
\label{eq:PF_allPQ}
\sum_{j=0}^n Y_{ij} V_j + Y_i^\text{sh} V_i = I_i + \frac{S_i^*}{V_i^*}
\quad (i=1, \ldots, n).
\end{equation}
These are the power flow equations in their most fundamental form, where all bus
injections on the r.h.s.\ are specified. Since all magnitudes are
complex-valued, system~\eqref{eq:PF_allPQ} consists of $2n$ equations in $2n$
unknowns, the complex voltages $V_i$. In order to contemplate voltage regulation
or any other type of control, one includes additional \emph{constraint}
equations and variables to~\eqref{eq:PF_allPQ}, always obeying the basic
algebraic rule that the number of independent equations and the number of
variables has to match---otherwise one ends up with an over/under-determined
system.
\subsection{Voltage regulation (PV buses) with no Q-limits}
\label{subsec:PVunlim}
In the following, the formulation will focus on the most ubiquitous type of
control: the automatic voltage regulation (AVR) performed by generators. This
kind of control may be either local or remote. When it is local to the bus, the
bus is commonly referred to as ``PV-type''. Whether local or remote, for the
purposes of steady-state power flow, the control materializes as an additional
equation constraining the voltage modulus of the controlled bus $i$,
$|V_i|=V_i^\text{sp}$ (the setpoint). Correspondingly, the reactive power
injection $Q_j$ of the controlling generator $j$ becomes a new variable in the
system. If there are several generators regulating the same voltage, some
sharing factors need to be used in order to obtain a single effective degree of
freedom, to avoid under-determinacy.
Therefore, when the controlling resources (in this case, generator Mvars) are
assumed unlimited, the system of power flow equations simply gets augmented with
new equations (equality constraints) and new variables. In the holomorphic
embedding method, unlimited controls are easily incorporated into the
formulation, after taking proper care of embedding the new constraints in a way
that preserves both holomorphicity and consistency with the reference state at
$s=0$. This can be done in a different number of ways, as shown
in~\cite{Subramanian2013,Trias15,Wallace2016}. For the treatment of standard PV
buses, our favored embedding scheme is as follows:
\begin{align}
\label{eq:PF_PQ}
i \in \text{PQ:} & \quad \sum_{j=0}^n Y_{ij} V_j(s) + sY_i^\text{sh} V_i(s) =
sI_i + \frac{sS_i^*}{V_i^*(s^*)} \\
\label{eq:PF_PV}
k \in \text{PV:} & \quad \sum_{j=0}^n Y_{kj} V_j(s) + sY_k^\text{sh} V_k(s) =
sI_k + \frac{sP_k-jQ_k(s)}{V_k^*(s^*)} \\
\label{eq:PVulimconstraint}
k \in \text{PV:} & \quad V_k(s)V_k^*(s^*) =
1 + s\left( W_k^\text{sp} - 1\right) .
\end{align}
Here $W_k^\text{sp}$ is defined as $\left(V_k^\text{sp}\right)^2$. The swing
voltage is assumed to be $V_0=1$; this can be achieved either by performing a
global normalization of variables, or more simply by embedding the swing as
$V_0(s)=1+s(V_0-1)$.
Considering the real and imaginary parts of voltages, this is a system of
$2\ensuremath{n_{\text{PQ}}}+3\ensuremath{n_{\text{PV}}}$ equations, in $2\ensuremath{n_{\text{PQ}}}+3\ensuremath{n_{\text{PV}}}$ unknowns. However, these equations can
be simplified and the dimensionality reduced to $2\ensuremath{n_{\text{PQ}}}+\ensuremath{n_{\text{PV}}}$ equations and
unknowns, as shown in \ref{app:methsummary}.
\subsection{Dealing with Q-limits: complementarity equations}
\label{subsec:PVlim}
The focus is now switched to the problem of resource limits, which is the main
subject of the paper. When a generator reaches either limit, $Q^\text{min}$ or
$Q^\text{max}$, its controlling injection saturates at that limit, and the
controlled voltage can no longer be sustained at the setpoint. Moreover the sign
of the control error is important, as it should be consistent with the sign of
the control sensitivity: if the generation saturates at $Q^\text{max}$, the
achieved voltage should be $V\le V^\text{sp}$, while if it saturates at
$Q^\text{min}$ one should have $V^\text{sp}\le V$. One should be careful with
the special case where the PV bus condition has reversed sensibility, which
corresponds to a power flow scenario on the ``wrong side'' of the generator's
Q-V curve. Although this is a mathematically valid power flow solution, it
cannot be considered operationally correct, as it makes voltage regulation
unstable. Further, it can be shown how, from the point of view of either HELM or
a continuation power flow, these anomalous cases graze a point of collapse
somewhere along the path. In the following, we will assume that the user has
corrected these anomalies beforehand, so that all generator voltage controls
have the correct sensitivity, $\partial V/\partial Q > 0$.
Given the complementary nature of the constraints, they may be expressed through
the equations:
\begin{align}
\label{eq:complementarity}
&\Bigl( V_iV_i^* - W^\text{sp} \Bigr) \Bigl( Q^\text{max}_i - Q_i \Bigr)
\Bigl( Q_i - Q^\text{min}_i \Bigr) = 0 \\
\label{eq:Qlims}
&\Bigl( Q^\text{max}_i - Q_i\Bigr) \ge 0; \qquad
\Bigl( Q_i - Q^\text{min}_i \Bigr) \ge 0
\end{align}
plus the additional consistency requirements regarding voltage control
sensitivities:
\begin{equation}
\label{eq:sensitiv_cond}
\begin{split}
Q_i=Q^\text{max}_i &\implies |V_i| \le V^\text{sp} \\
Q_i=Q^\text{min}_i &\implies |V_i| \ge V^\text{sp}
\end{split}
\end{equation}
The problem now is finding an equation or set or equations which, when
appropriately embedded, would yield a solution satisfying all of these
conditions~\eqref{eq:complementarity}--\eqref{eq:sensitiv_cond} at $s=1$. In
contrast with the unlimited case, where the equality
constraint~\eqref{eq:PVulimconstraint} can be embedded in quite a
straightforward way, the limited case contains a number of deep subtleties that
make it much more non-obvious. Attempting an ad-hoc embedding
of~\eqref{eq:complementarity} with no regard to the other conditions is likely
to end up in a faulty HELM scheme. In the following section, it will be
demonstrated how, thanks to a suitable Lagrangian formulation of the power flow
problem, this problem can be correctly framed as one of \emph{minimization}, and
thus obtain a HELM scheme that properly enforces all of the above equality and
inequality constraints.
\section{Lagrangian formulation of power flow}
\label{sec:Lagrangian}
This section shows how the power flow equations can be derived from a Lagrangian
formulation, much in the spirit of classical mechanics. This is then used to
show how the problem of limits can be tackled in terms of minimization, and from
there devise a suitable HELM scheme. Other authors have also approached the
same problem using OPF-like minimization techniques~\cite{Murray2015}, but here
the treatment specifically addresses the holomorphic embedding formulation. In
particular, it will be shown to have an intriguing resemblance to interior point
methods, but with the advantage of remaining a method based on holomorphic
functions and their analytic continuation.
\subsection{(Nonlinear) DC power flow}
In order to show the essential points of the Lagrangian formulation, the
attention is first turned to the pure DC problem, where equations are simpler.
The nonlinear power flow equations of an all-DC network may be written as:
\begin{equation}
\label{eq:DC_PF}
\sum_{j=0}^n g_{ij} (V_i-V_j) + G_i^\text{sh} V_i = I_i + \frac{P_i^*}{V_i^*}
\quad (i=1, \ldots, n),
\end{equation}
This is \emph{not} the commonly used ``DC power flow'' approximation to AC;
rather, it is the exact DC counterpart to the AC problem~\eqref{eq:PF}, where
voltages, conductances, and power and current injections are all real
magnitudes. Systems of this type are not common in utility grids, but they
do appear for instance in spacecraft, shipboard systems, and some modern
microgrids~\cite{TriasMarin16}. In any case, this system will be shown to be
quite useful as it clarifies up many issues that are otherwise obscured in the
AC formulation.
It is straightforward to verify that equations~\eqref{eq:DC_PF} can be derived
from the minimization of the following Lagrangian:
\begin{equation}
\label{eq:DCLagrangian}
\mathcal{L} = \frac{1}{2} \ensuremath{\sum_{\substack{ij\\j<i}}} g_{ij} \left( V_i - V_j \right)^2
+\frac{1}{2} \sum_i G_i^\text{sh} V_i^2 - I_i V_i - P_i \ln V_i \;,
\end{equation}
where the double summation of the first term avoids counting links twice, but
does include the swing (at index $j=0$); while the single summation of the rest
of the terms runs only over non-swing buses. For now the term Lagrangian is used
here to refer to what is essentially a potential energy function (in the spirit
of classical physics), but concepts of constrained optimization and
Lagrangian duality will be used shortly after.
The fact that the powerflow equations admit a description in terms of a
potential function is remarkable and should not be dismissed as a mere
curiosity. It follows that the powerflow problem can be seen as equivalent to a
problem of finding static equilibria in classical mechanics. It is thus
instructive to analyze the physical meaning of each term in this Lagrangian, as
it can provide fruitful insights. The first term is the sum of the power losses
($I^2R$) over all transmission branches of the network (divided by two). The
next term accounts for the power consumption due to shunt conductances. The next
term, $I_iV_i$, is the power supplied by constant-current injections at each
bus. The last term, $P_i \ln V_i$, is also power supplied into the network,
although the particular form of its expression will become clearer shortly.
Reading~\eqref{eq:DCLagrangian} as a problem in classical mechanics, one could
envision a hypothetical system made up of masses connected with springs and
sitting under some peculiar ``gravitational'' fields. Conductances $g_{ij}$ and
$G_i^\text{sh}$ would play the role of spring constants, while the rest of the
terms could be interpreted as some sort of on-site potential producing a force
field acting locally on each bus. It is also interesting to note that the
operational solution corresponds to a stable equilibrium of the system (a
minimum), while all other power flow solutions (``black
branches''~\cite{Trias12}) correspond to unstable equilibria (saddle points),
although this will not be proven here.
The role of constant power injections becomes clearer
when~\eqref{eq:DCLagrangian} is viewed as resulting from a slightly different,
\emph{constrained} minimization problem:
\begin{equation}
\label{eq:DCminimiz}
\begin{split}
\mathcal{E} \left( \{V_i\} \right) \equiv \frac{1}{2} &\ensuremath{\sum_{\substack{ij\\j<i}}} g_{ij}
\left(V_i-V_j\right)^2 + \frac{1}{2}\sum_i G_i^\text{sh} V_i^2 - I_i V_i \\
\underset{\{V_i\}}{\text{minimize:}} &\quad \mathcal{E}(\{V_i\}) \\
\text{subject to:} &\quad V_k - V_k^\text{sp} = 0
\qquad \left( k\in \{\text{V-regulated}\} \right)
\end{split}
\end{equation}
Here the voltage is regulated to a setpoint $V_k^\text{sp}$ on those buses for
which there was a non-zero constant-power injection specified in
problem~\eqref{eq:DC_PF}. Since negative voltages are not allowed, these
constraint equations can be written equivalently as
$\ln V_k^\text{sp}-\ln V_k=0$. This constrained problem is then solved by the
standard method of Lagrange multipliers, using the following Lagrangian:
\begin{equation}
\label{eq:DCLagrangian2}
\mathcal{L} \equiv \mathcal{E}
\; - \sum_k P_k \ln \left( \frac{V_k}{V_k^\text{sp}} \right) \;,
\end{equation}
which is the same as~\eqref{eq:DCLagrangian}, except for additional constant
terms $P_i\ln V_i^\text{sp}$ which do not depend on the voltages. The
stationarity conditions then take the same form as the original power flow
equations~\eqref{eq:DC_PF}. Therefore, constant power injections can be seen as
the Lagrange multipliers of minimization
problem~\eqref{eq:DCminimiz}. Additionally, this view has the advantage of
having a more direct physical interpretation in terms of a fundamental energy
balance, since now the primal minimization problem is on $\mathcal{E}$: power
losses, plus power consumption, minus power supply. Under this light, the
standard power flow problem~\eqref{eq:DC_PF}, in which constant power injections
(both load and generation) are specified, could be interpreted as minimizing the
transmission network losses. This is a key insight and a powerful guide when
approaching the AC case.
Adding now resource constraints to the regulation resources:
\begin{equation}
\label{eq:DCminLimited}
\begin{split}
\underset{\{V_i\}}{\text{minimize:}} &\quad \mathcal{E}\left( \{V_i\} \right) \\
\text{subject to:} &\quad V_k - V_k^\text{sp}= 0
\qquad \left( k\in \{\text{V-regulated}\} \right) \\
&\quad P_k^\text{min} \le P_k \le P_k^\text{max}
\end{split}
\end{equation}
Note how these inequality constraints are on the dual variables of
problem~\eqref{eq:DCminimiz}. Let us solve the maximization of the dual
problem by means of logarithmic barrier methods, by defining:
\begin{equation*}
\mathcal{B}_\mu \left( \{P_k\} \right) \equiv \mathcal{E}
- \! \sum_k P_k \ln \left( \frac{V_k}{V_k^\text{sp}} \right)
+ \mu \ln \left(P_k^\text{max}-P_k\right)
+ \mu \ln \left(P_k-P_k^\text{min}\right)
\end{equation*}
where $\mu>0$ is the barrier parameter that is made to converge to zero. The
stationarity conditions for maximizing $\mathcal{B}_\mu$ yield:
\begin{equation}
\label{eq:DCBarrier}
- \ln \left( \frac{V_k}{V_k^\text{sp}} \right)
- \frac{\mu}{P_k^\text{max}-P_k} + \frac{\mu}{P_k-P_k^\text{min}} = 0
\end{equation}
Rearranging,
\begin{equation*}
\left(\ln V_k-\ln V_k^\text{sp}\right)
\left(P_k^\text{max}-P_k\right) \left(P_k-P_k^\text{min}\right) =
\mu \left( P_k^\text{max} + P_k^\text{min} - 2P_k \right)
\end{equation*}
In the limit $\mu\rightarrow0$, this equation is essentially the same as the
complementarity constraint~\eqref{eq:complementarity} (when translated to the DC
problem), since $\ln V_k=\ln V_k^\text{sp}$ if and only if
$V_k=V_k^\text{sp}$. The rest of the conditions are also met: the solution will
satisfy $P_k^\text{min}\le P_k\le P_k^\text{max}$ because the barrier guarantees
that if the starting point is interior, the solution will remain so; and in case
of saturation, the correct control sensitivities are satisfied as well, as it is
readily verified by inspecting~\eqref{eq:DCBarrier}. Therefore this suggests a
HELM scheme where using $\mu=\mu_0(1-s)$ to embed~\eqref{eq:DCBarrier} can
accomplish all the desired goals. The next subsection will fully develop this
idea for the AC case.
\subsection{AC power flow}
This treatment finds an exact parallel in the AC case. However, the authors have
found this is only possible in the limit of an ideal, lossless transmission
network, where all branches (i.e.\ lines, transformers, shunts) have zero
resistance. This is only a limitation as far as it concerns the
\emph{characterization} of the solution provided by the method, not its
precision. For networks in which this ideality limit is strongly violated, one
can still take the proposed HELM scheme at face value and compute a feasible
power flow satisfying control limits---the only difference is that in this case
it is no longer strictly true that the resulting solution is minimizing
something. Still, resistances in transmission networks are typically low
($R/X\lesssim 0.1$), so that, simply invoking a continuity argument, one should
expect that using this method should lead to a solution that
\emph{approximately} minimizes the Lagrangian shown here, which will be shown to
be related to the net reactive power losses.
It will first be shown how the power flow equations of an ideal lossless AC
network can be derived from the minimization of a suitable Lagrangian function.
In a lossless network all line and transformer conductances vanish, so that all
admittances, including shunts, become pure imaginary. Using real and imaginary
components, equation~\eqref{eq:PF} becomes:
\begin{equation}
\label{eq:PF_cartesian}
\begin{split}
-b_{ij} \Im(V_i-V_j) - b_i^\text{sh} \Im V_i = \Re I_i +
\frac{P_i\Re V_i + Q_i\Im V_i}{|V_i|^2} \\
b_{ij} \Re(V_i-V_j) + b_i^\text{sh} \Re V_i = \Im I_i +
\frac{P_i\Im V_i - Q_i\Re V_i}{|V_i|^2}
\end{split}
\end{equation}
Written in this form, it is straightforward to verify that these equations can
be derived from the minimization of the following Lagrangian:
\begin{multline}
\label{eq:ACLagrangian}
\mathcal{L} = -\frac{1}{2} \ensuremath{\sum_{\substack{ij\\j<i}}} b_{ij} \left\{
\left(\Re V_i-\Re V_j\right)^2 + \left(\Im V_i-\Im V_j\right)^2 \right\} \\
-\frac{1}{2} \sum_i b_i^\text{sh} \left\{
\left(\Re V_i\right)^2 + \left(\Im V_i\right)^2 \right\}
-\frac{1}{2} \sum_i \left( \Re I_i \Im V_i - \Im I_i \Re V_i \right)\\
-\frac{1}{2} \sum_i Q_i \ln \left\{
\left(\Re V_i\right)^2 + \left(\Im V_i\right)^2 \right\}
- \sum_i P_i \tan^{-1} \left( \frac{\Im V_i}{\Re V_i} \right) \;,
\end{multline}
where the index $j$ in the double summation includes the swing bus. If the
complex voltages and complex current injections are now reinterpreted to be
vectors in 2-dimensional real space, \eqref{eq:ACLagrangian} can be written in
the more compact form:
\begin{multline}
\label{eq:ACLagrangianVec}
\mathcal{L} = -\frac{1}{2} \ensuremath{\sum_{\substack{ij\\j<i}}} b_{ij} \left\| V_i- V_j \right\|^2
-\frac{1}{2} \sum_i b_i^\text{sh} \left\| V_i \right\|^2 \\
-\frac{1}{2} \sum_i \left\| I_i \times V_i \right\|
-\frac{1}{2} \sum_i Q_i \ln \left\| V_i \right\|^2
- \sum_i P_i \theta_i \;,
\end{multline}
Let us analyze the meaning of each term in this expression. The first term is
the sum of the reactive power losses ($I^2X$) over all network branches (divided
by two). Since the susceptance $b_{ij}$ of transmission lines, except in
exceptional cases such as equivalents, is always negative, this first term is
always positive. The next term is similar in nature, as it accounts for the
reactive power injection/consumption due to shunt admittance terms. These
originate mainly from in line charging susceptances, but may also include
contributions from shunt banks and constant-impedance load models. The next
term, containing the cross product of ``vectors'' $I_i$ and $V_i$ at each bus,
shows the effect of local constant-current injections. Its interpretation in a
mechanical analogue may not be obvious, but note that it also has dimensions of
power. The next term may be interpreted as a sort of potential energy provided
by an external on-site ``field'' produced by a local ``charge'' $Q_i$. It is a
local logarithmic potential acting on each bus. The last term may also be
interpreted as produced by an on-site field, this time mediated by external
charges $P_i$ acting on each node and exerting a sort of transversal force that
tries to rotate the voltage vector. Therefore all terms in this Lagrangian
consist of magnitudes having dimensions of power, and can be interpreted as
being some sort of potential energy, if one forgets about the time dimension.
As in the DC case, one can see this as a pseudo-mechanical analogue, where the
power flow equations are re-interpreted as the equations for the static
equilibria of the system (one could think of particles moving in a 2D space,
interconnected with springs and having on-site potentials acting on them). This
analogy may be exploited in multiple ways by drawing on the techniques and
intuition from classical mechanics, but here the focus will be on the
constrained problem.
Analogously to the DC case, voltage regulation can be seen as the following
constrained minimization problem:
\begin{equation}
\label{eq:ACminimiz}
\begin{split}
\underset{\{V_i\}}{\text{minimize:}} &\quad \mathcal{E} \left( \{V_i\} \right) \\
\text{subject to:} &\quad |V_k|^2 - W_k^\text{sp}= 0
\qquad \left( k\in \{\text{V-regulated}\} \right)
\end{split}
\end{equation}
Here $\mathcal{E}$ is defined as $\mathcal{L}$ in~\eqref{eq:ACLagrangianVec},
but omitting the terms $\frac{1}{2} Q_k \ln \left\| V_k \right\|^2$ for
voltage-regulated buses $k$. Again, this problem is solved with the standard
technique of Lagrange multipliers:
\begin{equation}
\label{eq:ACLagrangian2}
\mathcal{L} \equiv \mathcal{E}
\; - \frac{1}{2} \sum_k Q_k \ln \frac{|V_k|^2}{W_k^\text{sp}} \;,
\end{equation}
which is the same as~\eqref{eq:ACLagrangianVec}, except for additional constant
terms $Q_k\ln W_k^\text{sp}$ which do not depend on the voltages. The
stationarity conditions then take the same form as the original power flow
equations~\eqref{eq:PF_cartesian}. Similarly to the DC case, reactive power
injections can be seen as the Lagrange multipliers of minimization
problem~\eqref{eq:ACminimiz}. The physical interpretation for the minimization
of $\mathcal{E}$ in terms of energy is in general not so evident as in DC, but
it can still be recovered if all active power injections $P_i$ are made zero. In
that limit, it can be shown that all angles are zero, all voltages and currents
are real, and the system actually becomes mathematically equivalent to the DC
case.
Let us now contemplate resource constraints, i.e., Mvar limits:
\begin{equation}
\label{eq:ACminLimited}
\begin{split}
\underset{\{V_i\}}{\text{minimize:}} &\quad \mathcal{E} \left( \{V_i\} \right) \\
\text{subject to:} &\quad |V_k|^2 - W_k^\text{sp}= 0
\qquad \left( k\in \{\text{V-regulated}\} \right) \\
&\quad Q_k^\text{min} \le Q_k \le Q_k^\text{max}
\end{split}
\end{equation}
These inequality constraints are on the dual variables of
problem~\eqref{eq:ACminimiz}. Consider then the maximization of the dual
problem by means of logarithmic barrier methods, by defining:
\begin{multline}
\label{eq:ACBarrier}
\mathcal{B}_\mu \left( \{Q_k\} \right) \equiv \mathcal{E}
- \frac{1}{2} \!\sum_k Q_k \ln \frac{|V_k|^2}{W_k^\text{sp}} \\
+ \mu \ln \left(Q_k^\text{max}-Q_k\right)
+ \mu \ln \left(Q_k-Q_k^\text{min}\right)
\end{multline}
where $\mu>0$ is the barrier parameter. The
stationarity conditions for maximizing $\mathcal{B}_\mu$ yield:
\begin{equation}
\label{eq:ACBarrierCond1}
- \frac{1}{2} \ln \frac{|V_k|^2}{W_k^\text{sp}}
- \frac{\mu}{Q_k^\text{max}-Q_k} + \frac{\mu}{Q_k-Q_k^\text{min}} = 0
\end{equation}
Rearranging, one obtains:
\begin{multline}
\label{eq:ACBarrierCond2}
\left(\ln |V_k|^2 - \ln W_k^\text{sp} \right)
\left(Q_k^\text{max}-Q_k\right) \left(Q_k-Q_k^\text{min}\right) = \\
2 \mu \left( Q_k^\text{max} + Q_k^\text{min} - 2Q_k \right)
\end{multline}
In the limit $\mu\rightarrow0$, this equation is essentially the same as the
complementarity constraint~\eqref{eq:complementarity}, since
$\ln |V_k|^2 = \ln W_k^\text{sp}$ if and only if $|V_k|^2 = W_k^\text{sp}$. The
rest of the conditions are also met: \eqref{eq:Qlims} is satisfied because the
barrier terms guarantee that the solution will remain interior if the starting
point is so; and the signs of the control errors, given
by~\eqref{eq:sensitiv_cond}, can be verified to be the correct ones by
inspecting~\eqref{eq:ACBarrierCond1}. When the solution is such that the
$Q_k^\text{max}$ constraint is active, one obtains $|V_k|^2 \le W_k^\text{sp}$;
and vice versa.
Therefore this suggests a HELM scheme where~\eqref{eq:ACBarrierCond1} is
embedded by using, for instance, $\mu=\mu_0(1-s)$. Together with some additional
precautions, to be described in the next section, this equation
replaces~\eqref{eq:PVulimconstraint} and provides a HELM scheme to solve the
problem of Q-limits. This establishes the relationship between the proposed
HELM scheme and the above minimization problem. The HELM solution is thus
characterized as minimizing $\mathcal{E}$ in~\eqref{eq:ACminLimited}, which, in
the limit of the ideal lossless network, is essentially the net amount of
reactive losses.
Finally, this result also suggests an apparent resemblance between so-called
Interior Point Methods (IPM) and the analytic continuation technique used in
HELM. However, just as in the case of homotopy methods, IPM and similar barrier
methods are based on continuity and differentiability, not holomorphicity (which
is a much stronger requirement).
\section{HELM scheme. Consistency requirements}
\label{sec:HELMscheme}
Following the above, a HELM scheme is proposed to solve the problem of voltage
regulation while observing Q-limits, using the canonically embedded
system~\eqref{eq:PF_PQ}--\eqref{eq:PF_PV} and
embedding~\eqref{eq:ACBarrierCond1} as follows:
\begin{equation}
\label{eq:compl_embed}
V_k(s) V_k^*(s^*) - W_k^\text{sp} = \frac{\mu_k(1-s)}{Q_k(s)-Q_k^\text{min}}
- \frac{\mu_k(1-s)}{Q_k^\text{max}-Q_k(s)}
\end{equation}
where $\mu_k$ are suitable constants to be discussed below (following IPM
methods, they are chosen to be the same for both terms). As discussed in the
previous section, in the limit $s\rightarrow0$ these equations reproduce the
complementarity
constraints~\eqref{eq:complementarity}--\eqref{eq:sensitiv_cond}. At this point,
it is important to remark that there are possibly many other methods to obtain
solutions satisfying the complementarity equations, but this is a method
\emph{informed by the Lagrangian}. In case there are several feasible solutions,
the method presented here will select the one minimizing the
Lagrangian~\eqref{eq:ACLagrangianVec}, which has shown to be rooted in the
physics of the transmission network.
On the other hand, the following treatment will consider the possibility that
the voltage regulation may be \emph{remote}, i.e.\ that the controlled bus and
its controlling injection are not on the same bus. There is nothing in the
constraint equations preventing this, as it would still produce the same,
well-defined algebraic problem. The only difference in such case is that, in a
strict sense, it would no longer be possible to interpret the resulting solution
as the minimizer of the constrained problem shown in the previous
section. However, if one restricts this to the most common case, which is
regulation at a distance of one bus, the interpretation is expected to still
hold approximately, in the sense that other feasible solutions would yield a
higher value of the Lagrangian.
Therefore from now on the notation will label the set of $\ensuremath{n_{\text{PV}}}$ reactive bus
injections with indexes $a$, and the set of their corresponding $\ensuremath{n_{\text{PV}}}$ regulated
bus voltages with $b$. Note that the treatment given here does not contemplate
concurrent control (i.e. shared responsibility regulation) of the same bus
originated from different buses. This problem has no unique solution in general,
as it depends on the details of the var-sharing policy, priority rules among
generators, etc.
HELM now requires that the embedding is designed in a way that a unique and
meaningful reference powerflow state is obtained at $s=0$: the zero-injection
state, in which all voltages are equal to the swing. It is then postulated that
the \emph{operational} solution, in contrast to the many other power flow
solutions, is the one that is analytically continued to $s=1$ from this
reference state (when it exists). This is based on physical arguments, discussed
at length in~\cite{Trias12,Trias15}. An important question then is
whether~\eqref{eq:compl_embed} can be consistent with the HELM reference
solution, given that limits $Q^{min},Q^{max}$ come from engineering constraints
and could have any value. This consistency can be examined by considering the
power series of the variables involved, and checking ~\eqref{eq:compl_embed} at
order zero. Using the notation $V_b(s)=\sum_nV_b[n]s^n, Q_a(s)=\sum_nQ_a[n]s^n$
for power series coefficients, one obtains at zero order:
\begin{equation*}
V_b[0] V_b^*[0] - W_b^\text{sp} = \frac{\mu_a}{Q_a[0]-Q_a^\text{min}}
- \frac{\mu_a}{Q_a^\text{max}-Q_a[0]}
\end{equation*}
and since the reference solution has $V_b[0]=1,\, Q_a[0]=0$ for all buses:
\begin{equation}
\label{eq:zero_order}
1 - W_b^\text{sp} =
-\mu_a \frac{Q_a^\text{max}+Q_a^\text{min}}{Q_a^\text{max}Q_a^\text{min}}
\end{equation}
There are two different consistency requirements to observe here. The first one
is that the reference state should not violate limits (or hit exactly a limit),
which implies $Q_a^\text{min}<0, Q_a^\text{max}>0$. In the parlance of interior
point methods, this is is the analogue of ensuring that the starting point is
``interior''. The second requirement is that~\eqref{eq:zero_order} should be
satisfied with a \emph{positive} value of the barrier constant $\mu_a$. Since
$Q_a^\text{min}Q_a^\text{max}<0$, this implies
$\sgn(1-W_b^\text{sp})=\sgn(Q_a^\text{max}+Q_a^\text{min})$. Neither one of
these two requirements are guaranteed to be satisfied, in general. Therefore one
is forced to embed both the limits and the setpoints, in a way that the above
conditions are satisfied at order zero. Note how this was already done in the
unlimited PV case, as $W_b^\text{sp}(s)=1+s(W_b^\text{sp}-1)$
in~\eqref{eq:PVulimconstraint}. In this case, the embeddings should be designed
to require that:
\begin{equation}
\begin{split}
& Q_a^\text{min}[0]<0 ,\quad Q_a^\text{max}[0]>0 \\
& \sgn\left(1-W_b^\text{sp}[0]\right) =
\sgn\left(Q_a^\text{max}[0]+Q_a^\text{min}[0]\right)
\end{split}
\end{equation}
For one-sided limits, the last expression needs to be replaced by
$W_b^\text{sp}[0]>1$ when only $Q_a^\text{max}$ is present, or by
$W_b^\text{sp}[0]<1$ when only $Q_a^\text{min}$ is present. All these
conditions can be accomplished by means of many possible embeddings. However, in
keeping with the general principle of introducing the minimum amount of changes
needed, our preferred method is to embed limits and setpoints linearly, choosing
suitable guard constants:
\begin{equation}
\label{eq:Qlims_embed}
\begin{split}
Q_a^\text{min}(s) &= Q_a^\text{min} + \delta Q_a^\text{min} (1-s) \\
Q_a^\text{max}(s) &= Q_a^\text{max} + \delta Q_a^\text{max} (1-s) \\
W_b^\text{sp}(s) &= W_b^\text{sp} + \delta W_b^\text{sp} (1-s)
\end{split}
\end{equation}
If both limits are present, one reasonable choice for these constants consists
in fixing $\mu_a=1$ and using the same embedding
for $W_b^\text{sp}$ as in the unlimited case, plus the following symmetrizing
embedding for the limits:
\begin{equation}
\label{eq:JLMembedding}
\begin{split}
Q_a^\text{min}(s) &= - \frac{Q_a^\text{max}-Q_a^\text{min}}{2}
+ s \frac{Q_a^\text{max}+Q_a^\text{min}}{2} \\
Q_a^\text{max}(s) &= \frac{Q_a^\text{max}-Q_a^\text{min}}{2}
+ s \frac{Q_a^\text{max}+Q_a^\text{min}}{2}
\end{split}
\end{equation}
which leads to the following embedding constants in~\eqref{eq:Qlims_embed}:
\begin{equation}
\label{eq:JLMembeddingconstants}
\begin{split}
\delta Q_a^\text{min} &= \delta Q_a^\text{max} = - \frac{Q_a^\text{max}+Q_a^\text{min}}{2} \\
\delta W_b^\text{sp} &= 1 - W_b^\text{sp}
\end{split}
\end{equation}
For buses with one-sided limits, one needs to choose values for the guard
constants in~\eqref{eq:Qlims_embed} so that the zero order coefficients satisfy
the sign requirements and have reasonable values. For instance, in case one has
only a lower limit $Q_a^\text{min}$:
\begin{equation}
\label{eq:JLMembeddingconstants1limmin}
\begin{split}
\delta Q_a^\text{min} &= Q_a^\text{min}[0] - Q_a^\text{min} \\
\delta W_b^\text{sp} &= W_b^\text{sp}[0] - W_b^\text{sp} \\
\mu_a &= Q_a^\text{min}[0] \left( W_b^\text{sp}[0] - 1 \right)
\end{split}
\end{equation}
so that one reasonable choice could be $W_b^\text{sp}[0]=0.9$ pu and
$Q_a^\text{min}[0]=-10$ pu. For buses where there is only an upper limit
$Q_a^\text{max}$, one obtains:
\begin{equation}
\label{eq:JLMembeddingconstants1limmax}
\begin{split}
\delta Q_a^\text{max} &= Q_a^\text{max}[0] - Q_a^\text{max} \\
\delta W_b^\text{sp} &= W_b^\text{sp}[0] - W_b^\text{sp} \\
\mu_a &= Q_a^\text{max}[0] \left( W_b^\text{sp}[0] - 1 \right)
\end{split}
\end{equation}
and the corresponding reasonable choice could be $W_b^\text{sp}[0]=1.1$ pu and
$Q_a^\text{min}[0]=+10$ pu. However, to keep the following treatment as general
as possible, no particular embedding constants will be assumed; it will only be
assumed that the embeddings are linear, as given by~\eqref{eq:Qlims_embed}.
The HELM procedure can now be constructed by obtaining the linear system that
allows obtaining all unknown coefficients at order $N$ from the knowledge of
coefficients at orders $N-1$ and lower. This system has been referred to as the
$N$-th order representation of the embeded equations~\cite{Trias12}. Solving
these linear systems in sequence is what constructs the germ of the reference
solution, which is later to be analytically continued into the final solution at
$s=1$ (if it exists). For the base power flow equations~\eqref{eq:PF_PQ} and
\eqref{eq:PF_PV}, \ref{app:methsummary} provides the $N$-th order representation
formulas (in the more general Pad\'e-Weierstrass framework). Here the focus is
put on the constraint equations, in order to show how their contribution results
in an enlarged but well-defined linear system.
The embedding selected in~\eqref{eq:Qlims_embed} directly provides the power
series corresponding to setpoints and limits. They have just two coefficients:
\begin{equation*}
\begin{split}
Q_a^\text{min}(s) &= Q_a^\text{min}[0] + Q_a^\text{min}[1] \, s \\
Q_a^\text{max}(s) &= Q_a^\text{max}[0] + Q_a^\text{max}[1] \, s \\
W_b^\text{sp}(s) &= W_b^\text{sp}[0] + W_b^\text{sp}[1] \, s
\end{split}
\end{equation*}
If you chose the symmetrizing embedding~\eqref{eq:JLMembedding}, the values of
$\mu_a$ can be chosen all $\mu_a=1$; otherwise they will be given by the
constraint equation~\eqref{eq:compl_embed} at zero order:
\begin{equation*}
\mu_a = \left( W_b^\text{sp}[0] - 1 \right)
\frac{Q_a^\text{max}[0]\,Q_a^\text{min}[0]}{Q_a^\text{max}[0]+Q_a^\text{min}[0]}
\end{equation*}
Now the following auxiliary series will be defined:
\begin{equation*}
\mathcal{B}_a^{(+/-)}(s) \equiv \frac{\mu_a(1-s)}{Q_a(s)-Q_a^\text{(min/max)}(s)} =
\sum_{m=0}^\infty \mathcal{B}_a^{(+/-)}[m] \, s^m
\end{equation*}
Since $Q_a[0]=0$, its zero-order values are:
\begin{equation*}
\mathcal{B}_a^{(+/-)}[0] = -\frac{\mu_a}{Q_a^\text{(min/max)}[0]}
\end{equation*}
For $N\ge 1$, equating the $N$-th power coefficients on each side of the constraint
equation~\eqref{eq:compl_embed} yields:
\begin{equation}
\label{eq:Nth_constraint}
\sum_{m=0}^N V_b[m]\,V_b^*[N-m] - W_b^\text{sp}[1]\,\delta_{1,N} =
\mathcal{B}_a^-[N] + \mathcal{B}_a^+[N]
\end{equation}
where $\delta_{1,N}$ is the Kronecker delta. Using the known values $V_b[0]=1$,
the convolution sum on the right hand side decomposes as:
\begin{equation*}
2\Re V_b[N] + \sum_{m=1}^{N-1} V_b[m]\,V_b^*[N-m]
\end{equation*}
Note that, even though this expression involves complex magnitudes, the sum is
real, given the symmetry of the convolution.
The power series coefficients of $\mathcal{B}_a^{(+/-)}$ can be obtained in
terms of those of $Q_a(s)$ from their definition. For instance, by equating the
$N$-th order coefficients of each side of equation
$\mathcal{B}_a^+(s)\left(Q_a(s)-Q_a^\text{max}(s)\right) = \mu_a(1-s)$, for
$N\ge 1$:
\begin{multline*}
\sum_{m=0}^N \mathcal{B}_a^+[m] \, Q_a[N-m]
- Q_a^\text{max}[0] \, \mathcal{B}_a^+[N] \\
- Q_a^\text{max}[1] \, \mathcal{B}_a^+[N-1] = -\mu_a\,\delta_{1,N}
\end{multline*}
and analogously for $\mathcal{B}_a^-$. Since $Q_a[0]=0$, one obtains:
\begin{multline*}
Q_a^\text{max}[0] \, \mathcal{B}_a^+[N] = \mathcal{B}_a^+[0]\, Q_a[N]
+ \sum_{m=1}^{N-1} \mathcal{B}_a^+[m] \, Q_a[N-m] \\
- \mathcal{B}_a^+[N-1] \, Q_a^\text{max}[1] + \mu_a\,\delta_{1,N}
\end{multline*}
Moving all terms in~\eqref{eq:Nth_constraint} that include coefficients at order
$N$ to the left hand side, one obtains:
\begin{equation}
\label{eq:Nth_constraint2}
2\Re V_b[N] - \left( \frac{\mathcal{B}_a^-[0]}{Q_a^\text{min}[0]}
+ \frac{\mathcal{B}_a^+[0]}{Q_a^\text{max}[0]} \right) Q_a[N] =
\mathcal{T}_a[N-1]
\end{equation}
where the symbol $\mathcal{T}_k[N-1]$ is used as a convenient shorthand to group
all other terms, which are either known quantities or involve coefficients of
order $N-1$ and lower. In the case of one-sided limits, the expressions are very
similar and straightforward to obtain.
This set of equations~\eqref{eq:Nth_constraint2} complement the rest, to form a
linear system of dimension $2\ensuremath{n_{\text{PQ}}}+3\ensuremath{n_{\text{PV}}}$, involving variables $\Re V_j[N]$,
$\Im V_j[N]$, and $Q_a[N]$. Actually, variables $Q_a[N]$ can be easily
eliminated from~\eqref{eq:Nth_constraint2} in terms of $\Re V_b[N]$ (it is
verified that their pivot cannot be zero), and substituted into the rest of the
equations to obtain a linear system of dimension $2(\ensuremath{n_{\text{PQ}}}+\ensuremath{n_{\text{PV}}})$. The method then
proceeds as usual~\cite{Trias12}: one solves these linear systems in sequence,
starting from $N=1$ until enough power series terms are obtained to either
obtain convergence of detect oscillation in the Padé approximant
sequences.
\section{The Padé-Weierstrass method for analytic continuation}
\label{sec:PW}
The preceding scheme is able to produce a well-defined germ, but its analytic
continuation to $s=1$ by means of Padé approximants suffers from an intrinsic
problem: by its very construction, the embedded complementarity
constraint~\eqref{eq:compl_embed} creates a singularity of the algebraic curve
at $s=1$, since this is a branching point for many other solutions. Such
solutions correspond to the many possible choices in which one can saturate the
controls (i.e.\ different selections of buses switched to PQ-type). All of them
satisfy the power flow equations as well as the complementarity constraint, but
most of them violate one or more of the other conditions, i.e.\ \eqref{eq:Qlims}
and \eqref{eq:sensitiv_cond}. And if there existed any configuration satisfying
all the requirements (something that cannot be ruled out, due to the
nonlinearity of the problem), it would be ``worse'' than the solution produced
by the proposed HELM scheme, in the sense of the minimization criterion provided
by the Lagrangian formulation on which the method is grounded (the HELM solution
approximately minimizes reactive power losses).
As it is well-known from Stahl's theory~\cite{Stahl97}, the near-diagonal
sequences of Padé approximants cease to converge at singularities (and on the
minimal cut-set joining them). For the HELM scheme developed in the previous
section this means that, even if the power flow is feasible, the approximants
are no longer guaranteed to converge at $s=1$. Actually, unless the case is
outright infeasible (in which case the approximants will clearly oscillate),
what happens is that the convergence rate gets very slow as $s$ approaches 1.
In the unlimited HELM method, this would only happen when calculating a case
\emph{exactly} at a point of collapse. In both situations, due to the limits of
floating point arithmetic (truncation and round-off), there is a point at which
high order Padé approximants can no longer be calculated with enough precision
to improve the results. The net result is a ceiling to the achievable precision
in the solution as $s\rightarrow 1$. Numerical experiments show that, in the
particular case of the Q-limits HELM scheme, the magnitude of the precision loss
is usually too large to obtain acceptable results. On the other hand, in case
of powerflow infeasibility, a voltage collapse point will be met earlier at some
$s_\text{crit}<1$, as shown in~\cite{Trias15}; this will be more easily detected
as oscillations in the approximants when evaluated at $s=1$.
In any case, given the nature of the HELM reference solution at $s=0$, it is
always possible to find a real value $0<s_0<1$ such that the Padé approximants
do converge to the required precision. This means it is always possible to
obtain a \emph{partial solution} $V_i(s_0)$ whose values are exact, within
machine precision. What follows is a method that exploits this property and
comes up with a method to perform the analytic continuation of the reference
germ in a series of stages that approach $s\rightarrow 1$, greatly enhancing the
numerical precision. This is effectively an analytic continuation along a path
on the real axis, in the same spirit as Weierstrass, but using a novel
technique. A classical Weierstrass procedure would use the initial power
series to construct a second power series at a point $s_0$ inside its radius of
convergence, and then repeat this process. However,
Henrici~\cite{Henrici93Vol1} shows that this naive approach yields in practice
very poor numerical precision. Here, by contrast, the method uses partial
solutions and the power flow equations in order to re-expand the new germs along
the path. Since the procedure makes use of Padé approximants as a key element
for obtaining partial solutions to very high precision, the method is termed
\emph{Padé-Weierstrass} (to be abbreviated as P-W).
\subsection{Two-bus DC system}
The essential elements of the Padé-Weierstrass method can be shown in the
simplest power flow problem, the two-bus DC system. For further simplicity, the
swing voltage will be normalized to 1. The embedded equation reads:
\begin{equation}
\label{eq:2busDC}
G\Bigl( V(s) - 1 \Bigr) = s\frac{P}{V(s)}
\end{equation}
As always, the aim is to obtain the analytic continuation of the HELM germ to
$s=1$. Assume now that the convergence of Padé approximants is so slow at $s=1$
that the achievable precision is degraded (though in this simple system, this
only happens when the solution is \emph{very} close to point of voltage
collapse). One can always find a value $0<s_0<1$ such that the approximants
converge to any required tolerance (within machine precision limits). Let us use
these values $s_0$ and $V(s_0)$ to rewrite the equation in terms of the
following renormalized parameter and voltage:
\begin{equation}
\label{eq:renorm}
\begin{split}
s &\equiv s_0 + (1-s_0)s' \\
V(s) &\equiv V(s_0) V'(s')
\end{split}
\end{equation}
The parameter change $s\rightarrow s'$ is a conformal mapping that takes the
range $[s_0,1]$ in $s$-space into the range $[0,1]$ in $s'$-space. Using these,
\eqref{eq:2busDC} reads:
\begin{equation*}
V(s_0) \, G \Bigl( V(s_0)V'(s') - 1 \Bigr) =
\frac{s_0 P}{V(s)} + s' \frac{(1-s_0)P}{V'(s')}
\end{equation*}
Using the fact that $G\bigl(V(s_0)-1\bigr)=s_0\frac{P}{V(s_0)}$, the right hand
side of this equation yields:
\begin{multline*}
V(s_0)\,G\Bigl(V(s_0)-1\Bigr) V'(s') + V(s_0)\,G\Bigl(V'(s')-1\Bigr) = \\
s_0 P\,V'(s') + V(s_0)\,G\Bigl(V'(s')-1\Bigr)
\end{multline*}
and the final equation becomes:
\begin{equation}
\label{eq:PWstep}
G' \Bigl( V'(s') - 1 \Bigr) = \Gamma' \Bigl( \frac{1}{V'(s')} - V'(s') \Bigr)
+ s' \frac{P'}{V'(s')}
\end{equation}
where the new constants are defined as $G'\equiv V(s_0)G$,
$\Gamma'\equiv s_0 P$, and $P'\equiv(1-s_0)P$. This equation describes a new
power flow problem, very similar to~\eqref{eq:2busDC} except for the appearance
of two new terms which will be jointly referred to as a $\Gamma$-term. Before
analyzing in detail its meaning and role, let us remark that the term itself is
completely invariant under this renormalization procedure. That is, assume that
instead of~\eqref{eq:2busDC} one is given the following initial power flow problem:
\begin{equation}
\label{eq:2busDC_withGamma}
G\Bigl( V(s) - 1 \Bigr) = \Gamma \Bigl( \frac{1}{V(s)} - V(s) \Bigr)
+ s\frac{P}{V(s)}
\end{equation}
Then it can be verified that, under the change~\eqref{eq:renorm}, this system
also transforms exactly into system~\eqref{eq:PWstep}, with these new
parameters:
\begin{equation*}
\begin{split}
G' &\equiv V(s_0)G \\
\Gamma' &\equiv \Gamma + s_0 P \\
P' &\equiv (1-s_0)P
\end{split}
\end{equation*}
This result is quite remarkable. Conceptually, this is similar to the way
certain changes of coordinates leave physical laws invariant, i.e. the principle
of general covariance. For instance Newton's equations of classical mechanics
are invariant under Galilean transformations, and Einstein's equations of
Special Relativity are invariant under Lorentz transformations. What we have
here is that, when the embedded powerflow problem is written
as~\eqref{eq:2busDC_withGamma}, it is invariant under the ``change of reference
frame'' given by~\eqref{eq:renorm}.
This property can be exploited to good effect. The whole procedure described
here, which will be referred to as a single \emph{Padé-Weierstrass stage}, can
be repeated multiple times, for suitable values $s'_0, s''_0$, etc. At each
stage one always obtains a new powerflow problem that it is formally the same,
where only the parameters have changed. The great advantage is that, as it will
be shown, each successive stage produces a better-conditioned HELM problem from
the numerical point of view. Then, if at some stage $k$ the Padé approximants
are able to converge with the requested precision at $s^{(k)}=1$, one can simply
undo all the changes~\eqref{eq:renorm} and thus obtain the sought solution to
the initial problem.
Let us now turn to the $\Gamma$-term. On the first P-W stage it is zero, but it
will appear at all subsequent stages. The term can be thought of as a
combination of a constant-power injection and a constant-admittance shunt. Their
respective currents cancel out when $V'(s')=1$, which happens at $s'=0$. This
is consistent with the HELM reference state (no flows, all voltages equal to the
swing). However, the novelty is that this $\Gamma$-term is not embedded in $s$,
so it remains to be seen how the HELM procedure can work. It turns out that
given the specific form of this term, whereby their currents cancel out at
$s=0$, the standard HELM method can be carried through without any problem, as
shown now.
To kickstart the method on~\eqref{eq:2busDC_withGamma}, consider first the
equation obtained for the power series coefficients at zero order. It is
verified to be consistent with the HELM reference state, $V(0)=1$, since
$V^{-1}[0]=V[0]=1$. Recall that the power series coefficients for the function
$1/V(s)$, symbolically represented as $V^{-1}[N]$ (not to be confused with
$1/V[N]$), can be obtained from those of $V(s)$ by the relationship:
\begin{equation}
\label{eq:invSeries}
\sum_{m=0}^N V[m] \, V^{-1}[N-m] = \delta_{N,0}
\end{equation}
At orders $N\ge 1$, one obtains:
\begin{equation*}
G V[N] = \Gamma \Bigl( V^{-1}[N] - V[N] \Bigr) + P V^{-1}[N-1]
\end{equation*}
Using~\eqref{eq:invSeries}, one finally obtains:
\begin{equation}
\label{eq:2busDC_Norder}
\left( G + 2\Gamma \right) V[N] =
P V^{-1}[N-1] - \Gamma\sum_{m=1}^{N-1} V[m] \, V^{-1}[N-m]
\end{equation}
All terms on the right hand side are order $N-1$ or lower, so they can be
computed from coefficients obtained at previous steps. Therefore the HELM scheme
can be carried out without any problem.
There is one final aspect that deserves attention. The coefficient of $V[N]$
in~\eqref{eq:2busDC_Norder} could in principle become zero if left unchecked. In
a full $n$-bus system, the analogous problem would manifest itself as the matrix
of the HELM linear system becoming singular. Let us analyze for instance what
would happen when trying to solve the two-bus system~\eqref{eq:2busDC} for a
case right at the voltage collapse point, i.e.\
$P=P_\text{crit}=-G/4$. Performing one step of the P-W procedure, one chooses a
point $0<s_0<1$ for which the approximants converge to $V(s_0)$ within the
required precision. Since this problem can be solved in closed form, one finds
$V(s_0)=(1+\sqrt{1-s_0})/2$. Therefore, at the next P-W stage the pivot
coefficient in~\eqref{eq:2busDC_Norder} becomes:
\begin{equation*}
G' + 2\Gamma' = \frac{1}{2} G \left( 1-s_0 + \sqrt{1-s_0} \, \right)
\end{equation*}
From this expression one can observe that, when the target state at $s=1$ sits
near or at a singularity, there is a certain trade-off in the choice of
$s_0$. Even if the approximants allowed it, getting too close in one step
(i.e. $s_0\approx 1$) would produce a numerically unstable system in the next
P-W step; so it may be more advantageous to settle for a smaller value even if
this means taking a few more P-W steps, if the overall numerical stability
improves.
Next it will be shown how this method can be extended to the full AC power flow,
including voltage regulation and Q-limits.
\subsection{Full AC system: main equations}
The P-W method will now be developed for the full AC power flow problem, as
represented by the embedded system~\eqref{eq:PF_PQ}--\eqref{eq:PF_PV} and
including the observance of Q-limits by means of the complementarity
constraint~\eqref{eq:compl_embed} as shown in Section~\ref{sec:HELMscheme}. As
in the two-bus DC system, the end result will be a sequence of HELM subproblems
all formally identical to the initial one, where only the parameters get
re-defined. At each successive stage of the P-W process, the numerical stability
of each HELM subproblem improves, so that it is possible to reach much higher
levels of precision at $s=1$, even when a singularity is very close. When
dealing with limits, this is a must, since the constraint equation creates a
singularity exactly at $s=1$, as discussed at the beginning of this Section.
In order to simplify the exposition, constant-current injections (either load or
generation) will be omitted for now; their treatment is not given here but, it
can be shown that the method is easily adapted to contemplate them as well. Let
us first consider the equation for PQ buses, \eqref{eq:PF_PQ}, which here will
be written as: \begin{equation}
\label{eq:PW_PQ}
i \in \text{PQ:} \; \sum_{j=0}^n Y_{ij} V_j(s) + sY_i^\text{sh} V_i(s) =
\frac{sS_i^*}{V_i^*(s^*)} + \Gamma_i \left(\frac{1}{V_i^*(s^*)} - V_i(s)\right)
\end{equation}
Here $\Gamma_i$ is a parameter whose value is zero on the first stage of the
method, but in subsequent stages it will be non-null, in general. As in the DC
example, assume now that the system is solved with the standard HELM method and
a real value $0 < s0 <1$ is chosen such that the Padé approximants converge with
the required tolerance. The values $s_0$ and $V_i(s_0)$ will now be used to
rewrite the equation in terms of the renormalized parameter $s'$ and voltage
variables $V'(s')$ defined by~\eqref{eq:renorm}. Evaluating~\eqref{eq:PW_PQ} at
$s=s_0$ and multiplying both sides by $V_i^*(s_0^*)$, one has:
\begin{equation}
\label{eq:PW_PQ_at_s0}
\sum_{j=0}^n V^*_i(s_0) Y_{ij} V_j(s_0) + s_0 Y_i^\text{sh} |V_i(s_0)|^2 =
s_0 S_i^* +
\Gamma_i \left( 1 - |V_i(s_0)|^2 \right)
\end{equation}
where the fact that $s_0$ is chosen real has been used to simplify the
expression. A non-real value of $s_0$ could in principle be chosen instead, thus
leading to a path analytic continuation that would meander around the complex
plane in order to get from $s=0$ to $s=1$. Uniqueness of the result would still
be guaranteed, since by Stahl's theorem Padé approximants will keep avoiding the
minimal cut set, and therefore there will be no ``branch jumps'' no matter what
path is chosen. However, as shown in~\cite{Trias15}, physics-based arguments
dictate that analytic continuation of the white germ is actually \emph{required}
to exist along all points $0\le s_0\le 0$ on the real axis, in order to be
called the operational solution. Therefore the method requires that $s_0$ is
chosen real.
Writing~\eqref{eq:PW_PQ} in terms of~\eqref{eq:renorm}, also multiplying both
sides by $V_i^*(s_0^*)$:
\begin{multline}
\label{eq:PW_PQ_tmp}
\sum_{j=0}^n V^*_i(s_0) Y_{ij} V_j(s_0) \, V'_j(s') \\
+ s_0 |V_i(s_0)|^2 Y_i^\text{sh} \, V'_i(s')
+ s' (1-s_0) |V_i(s_0)|^2 Y_i^\text{sh} \, V'_i(s') = \\
\frac{s_0 S_i^*}{V_i'^*(s'^*)} + s'\frac{(1-s_0) S_i^*}{V_i'^* (s'^*)}
+ \Gamma_i \left(\frac{1}{V_i'^*(s'^*)} - |V_i(s_0)|^2 V_i(s')\right)
\end{multline}
In passing, note that if the swing voltage (index $j=0$) is embedded as
$V_0=1+s(V_\text{sw}-1)$, then it is straightforward to obtain its expression in
terms of the new parameter:
\begin{equation*}
V_0'(s') = \frac{V_0(s)}{V_0(s_0)} =
1 + s' \frac{(1-s_0)(V_\text{sw}-1)}{1+s_0(V_\text{sw}-1)}
\end{equation*}
It is useful to introduce now the shorthand notation:
\begin{equation*}
\hat{Y}_{ij} \equiv V^*_i(s_0) Y_{ij} V_j(s_0)
\end{equation*}
However, unlike the original matrix $Y_{ij}$, this new admittance matrix would not
satisfy the transmission condition (i.e., the sum of all columns does not yield
the zero column vector). In fact, the sum of its columns can be readily
calculated using~\eqref{eq:PW_PQ_at_s0}:
\begin{equation*}
\sum_{j=0}^n \hat{Y}_{ij} = - s_0 Y_i^\text{sh} |V_i(s_0)|^2 +
s_0 S_i^* + \Gamma_i \left( 1 - |V_i(s_0)|^2 \right)
\end{equation*}
This can be fixed by defining the new admittance matrix as follows:
\begin{equation}
\label{eq:Yrenorm}
Y'_{ij} \equiv
\begin{cases}
\hat{Y}_{ij} & \text{if } i \ne j \\
\hat{Y}_{ii} - \sum_{l=0}^n \hat{Y}_{il} & \text{if } i=j
\end{cases}
\end{equation}
With this definitions and after straightforward algebraic rearrangements and
simplifications of~\eqref{eq:PW_PQ_tmp}, one arrives at the following equation
for PQ buses:
\begin{equation}
\label{eq:PW_PQ_at_sprime}
\sum_{j=0}^n Y'_{ij} V'_j(s') + s' Y_i^{\prime\text{sh}} V'_i(s') =
\frac{s'S_i'^*}{V_i'^*(s'^*)}
+ \Gamma'_i \left(\frac{1}{V_i'^*(s'^*)} - V'_i(s')\right)
\end{equation}
where the new parameters are given by:
\begin{equation*}
\begin{split}
Y_i^{\prime\text{sh}} &\equiv (1-s_0) |V_i(s_0)|^2 Y_i^\text{sh} \\
\Gamma'_i &\equiv \Gamma_i + s_0 S_i^* \\
S_i' &\equiv (1-s_0) S_i
\end{split}
\end{equation*}
Equation~\eqref{eq:PW_PQ_at_sprime} is exactly the same as the original one,
\eqref{eq:PW_PQ}. In this sense, it can be said that the system is invariant
under the Padé-Weierstrass transformation given by~\eqref{eq:renorm}.
Analogously, the equations for PV buses can be shown to be invariant as
well. Following the notation introduced in Section~\ref{sec:HELMscheme}, which
allows for one-on-one remote voltage regulation, the equations corresponding to
\emph{regulating} buses are written as:
\begin{multline}
\label{eq:PW_PV}
a \in \text{PV:} \\ \sum_{j=0}^n Y_{aj} V_j(s) + sY_a^\text{sh} V_a(s) =
\frac{sP_a - jQ_a(s)}{V_a^*(s^*)} +
\Gamma_a \left(\frac{1}{V_a^*(s^*)} - V_a(s)\right)
\end{multline}
Note that these buses are not strictly speaking PV-type, as the formulation here
is more general. Buses labeled $a$ control the voltage of buses labeled $b$; but
a bus is said to be PV-type when $a=b$. However, for the sake of simplicity, the
set of buses labeled $a$ will be referred to as ``PV buses'' just to avoid
longer expressions such as ``the buses having voltage-controlling reactive
injections''.
In this case, one also needs to specify how the new variables $Q_a(s)$ should
transform. In contrast to voltages, it turns out that the transformation that
works is \emph{additive}:
\begin{equation}
\label{eq:renormQ}
Q_a(s) \equiv Q_a(s_0) + Q_a'(s')
\end{equation}
Following now a procedure completely analogous to the PQ case, it is
straightforward to verify that the equations for PV buses~\eqref{eq:PW_PV} are
also invariant under the transformation given by~\eqref{eq:renorm} and
\eqref{eq:renormQ}. In this case the new parameters are:
\begin{equation*}
\begin{split}
Y_a^{\prime\text{sh}} &\equiv (1-s_0) |V_a(s_0)|^2 Y_a^\text{sh} \\
\Gamma'_a &\equiv \Gamma_a + s_0 P_a - j Q_a(s_0)\\
P_a' &\equiv (1-s_0) P_a
\end{split}
\end{equation*}
\subsection{Full AC system: complementarity constraints}
Let us now analyze how the complementarity constraint equations
\eqref{eq:compl_embed} transform under the change given by~\eqref{eq:renorm},
\eqref{eq:renormQ}. Consider the most general case, where limits and
setpoints are embedded and voltage regulation may be remote, as described in
Section~\ref{sec:HELMscheme}:
\begin{equation}
\label{eq:compl_embed2}
V_b(s) V_b^*(s^*) - W_b^\text{sp}(s) =
- \frac{\mu_a(1-s)}{Q_a^\text{max}(s) - Q_a(s)}
+ \frac{\mu_a(1-s)}{Q_a(s) - Q_a^\text{min}(s)}
\end{equation}
It is readily verified that this equation is also invariant under the change, if
one defines the new transformed parameters as follows:
\begin{equation*}
\begin{split}
\mu_a' &\equiv \frac{(1-s_0)\,\mu_a}{\left|V_b(s_0)\right|^2} \\
W_b^{\prime\text{sp}}(s') &\equiv
\frac{W_b^\text{sp}}{\left|V_b(s_0)\right|^2}
+ \frac{(1-s_0)\,\delta W_b^\text{sp}}{\left|V_b(s_0)\right|^2} (1-s') \\
Q_a^{\prime\text{min}} (s') &\equiv
Q_a^\text{min} - Q_a(s_0) + \delta Q_a^\text{min} (1-s_0) (1-s') \\
Q_a^{\prime\text{max}}(s') &\equiv
Q_a^\text{max} - Q_a(s_0) + \delta Q_a^\text{max} (1-s_0) (1-s')
\end{split}
\end{equation*}
For the transformed problem, the consistency requirements at $s'=0$ are
automatically satisfied, as the corresponding equation is equivalent to the
original one when $s=s_0$, and $s_0$ is interior. Namely, one has
$Q_a^{\prime\text{min}}(s'=0)<0$, $Q_a^{\prime\text{max}}(s'=0)>0$, and:
\begin{equation*}
\sgn\left(1-W_b^{\prime\text{sp}}(s'=0) \right) =
\sgn\left( Q_a^{\prime\text{min}}(s'=0) + Q_a^{\prime\text{max}}(s'=0) \right)
\end{equation*}
This last condition may be satisfied even with a value of zero, i.e.\
$W_b^{\prime\text{sp}}(s'=0)=1$; this poses no problem since the transformed
coefficients $\mu_a'$ are already known. Therefore after the first P-W stage
there is no need to introduce any additional guards in the embedding of neither
the setpoint nor the limits.
This completes then the P-W procedure for the full AC system including Q-limits.
At each stage, one obtains (using standard HELM) a \emph{partial solution} at a
value $0<s_0<1$ such that the required precision is satisfied (which is always
possible, by Stalh). Then one uses this partial solution in $s$-space in order
to define a transformed problem in $s'$-space, using~\eqref{eq:renorm},
\eqref{eq:renormQ}. The transformed problem turns out to be formally the same
as the original one, so that the whole procedure can be repeated recursively,
until one is able to reach the required precision at $s^\text{(k)}=1$ at some
P-W stage $k$. The solution to the original problem is then obtained by undoing
all the transformations defined by using~\eqref{eq:renorm}, \eqref{eq:renormQ}.
In effect, this is a path-analytic continuation in the spirit of Weierstrass,
but the re-expansion of the power series at intermediate points is obtained
using the exact function values (within machine precision, using Padé
approximants) instead of the initial power series.
\subsection{Full AC system: HELM scheme}
It only remains to see how the HELM scheme is slightly modified by the presence
of the new terms $\Gamma$ introduced by the P-W procedure. This only affects the
main power flow equations; the complementarity constraint equations remain exactly
the same as in the first stage, so that the $N$-th order representation and the
resulting HELM scheme are exactly the same as those shown in
Section~\ref{sec:HELMscheme}.
From the equation for PQ buses, \eqref{eq:PW_PQ}, one obtains the following
relationship for the power series coefficients at order $N$ (the so-called
$N$-th order representation):
\begin{multline*}
\sum_{j=0}^n Y_{ij} V_j[N] + \Gamma_i \left( V_i[N] - V_i^{-1*}[N] \right) = \\
S_i^* V_i^{-1*}[N-1] - Y_i^\text{sh} V_i[N-1]
\end{multline*}
Recall that the coefficients of $1/V(s)$, $V_i^{-1}[N]$, can be obtained in
terms of those of $V(s)$ by~\eqref{eq:invSeries}. Initial values are
$V_i[0] = V_i^{-1}[0] = 1$. Therefore:
\begin{equation*}
\sum_{j=0}^n Y_{ij} V_j[N] + 2\Gamma_i \Re V_i[N] = \mathcal{R}_i[N-1]
\end{equation*}
where all terms on the right hand side depend only on coefficients of order
$N-1$ or less:
\begin{multline*}
\mathcal{R}_i[N-1] \equiv
S_i^* V_i^{-1*}[N-1] - Y_i^\text{sh} V_i[N-1] \\
- \Gamma_i \sum_{m=1}^{N-1} V_i^*[m] V_i^{-1*}[N-m]
\end{multline*}
Analogously, from the equation for PV buses, \eqref{eq:PW_PV}, one obtains:
\begin{multline*}
\sum_{j=0}^n Y_{aj} V_j[N] + \Gamma_a \left( V_a[N] - V_a^{-1*}[N] \right) = \\
P_a V_a^{-1*}[N-1] - Y_a^\text{sh} V_a[N-1] - j \sum_{m=0}^N Q_a[m] V_a^{-1*}[N-m]
\end{multline*}
Making use of initial values $V_a^{-1}[0]=1,\, Q_a[0]=0$, and moving all terms
of order $N$ to the left hand side, one obtains:
\begin{equation*}
\sum_{j=0}^n Y_{aj} V_j[N] + 2\Gamma_a \Re V_a[N] + jQ_a[N] = \mathcal{R}_a[N-1]
\end{equation*}
where all terms on the right hand side depend only on coefficients of order
$N-1$ or less:
\begin{multline*}
\mathcal{R}_a[N-1] \equiv
P_i V_a^{-1*}[N-1] - Y_a^\text{sh} V_a[N-1] \\
- \Gamma_i \sum_{m=1}^{N-1} V_a^*[m] V_a^{-1*}[N-m]
- j \sum_{m=1}^{N-1} Q_a[m] V_a^{-1*}[N-m]
\end{multline*}
and this concludes the description of the method. Recall that variables $Q_a[N]$
can be eliminated from the constraint equation~\eqref{eq:Nth_constraint2}, thus
leading to a linear system of dimension $2(\ensuremath{n_{\text{PQ}}}+\ensuremath{n_{\text{PV}}})$ in the variables
$\Re V[N]$, $\Im V[N]$.
\section{Numerical results}
\label{sec:NumResults}
\subsection{Improved accuracy and numerical stability near collapse}
Before numerical results are presented on power flow cases with Q-limits, let us
analize the effects of the P-W technique on the base HELM method, when no Mvar
limits are enforced. In order to do so, let us consider the kind of situations
under which the method may suffer from numerical precision problems. There are
two possible sources: very close proximity to the feasibility boundary (voltage
collapse), and ill-conditioning of the transmission admittance matrix. The last
case may occur when some anomalous branch admittances are much larger or much
smaller than the rest, ending up in a HELM linear system whose matrix has a bad
condition number. This will produce errors in the power series coefficients, and
eventually slower convergence. Just as it is done in iterative methods, this is
simply prevented by requiring the admittances to be within certain minimum and
maximum thresholds.
The other case has a more fundamental cause. A case that is feasible but very
near the critical point is a case for which the embedded HELM problem has a
singularity at $s_\text{crit}=1+\varepsilon$, where $0<\varepsilon<<1$. The
point $s_\text{crit}$ marks the transition, as the Padé approximants will start
to oscillate for $s>s_\text{crit}$. The problem is that, even though Stalh's
theorem still guarantees that the Padé approximant sequences converge at $s=1$
for this case, the convergence rate slows down so much that too many orders of
the power series are needed, and the approximants can no longer be computed with
precision due to the inherent limits of floating point arithmetic. In simpler
terms: as a case approaches the feasibility boundary, the method hits a ceiling
of maximum attainable numerical precision.
In order to quantify this effect, and show how the new P-W technique helps in
mitigating it, some numerical results are presented below. The important point
here is not so much the ability of P-W to obtain ``nose points'' with much
greater precision (after all, it can be shown that the attainable precision of
the base HELM method is sufficient for most practical purposes). Rather, the
aim here is to provide insight as to why P-W is needed when incorporating
limits, since in that case the complementarity constraints always produce a
singularity right at $s_\text{crit}=1$.
Figure~\ref{fig:max_attainable_precision} shows the concept of maximum
attainable precision in an actual example, \texttt{case9241pegase} available in
MATPOWER~\cite{MATPOWER}. All calculations are performed using standard IEEE
double precision throughout, i.e. about 15 digits of precision. The figure
plots the overall convergence rate by looking at the maximum update error in
voltages across all buses versus $N$, the number of power series terms. This
update error is calculated as the difference between two successive Padé
approximants down the staircase sequence of the Padé table, e.g. $[L/M]$ and
$[L+1/M]$ (in the notation of~\cite{BakerGraves96}), where one normally uses
$L=M=N/2$. The results show that in this case the standard HELM algorithm is not
able to obtain the voltages with more precision than about $\pm 10^{-7}$ pu, no
matter how many series terms are obtained. This happens because at around 25
terms the evaluation of Padé approximants starts suffering from the limitations
of floating point arithmetic (catastrofic cancellations and
round-off). Performing just one step of the P-W procedure, the new maximum
achievable precision in the transformed problem rises to $\pm 10^{-11}$ pu. A
second P-W step is sufficient to achieve about $\pm 10^{-15}$, thus exausting
the machine precision.
\begin{figure}[!t]
\centering
\includegraphics[width=0.98\columnwidth]{figs/Figs_errors_vs_order_as_PW-case9241pegase-noQlims}
\caption{Convergence rates for \texttt{case9241} (unmodified, no Mvar limits
enforced). Maximum attainable precision improves from about
$\pm 10^\text{-7}$ to $\pm 10^\text{-11}$ pu by performing just one P-W
step.}
\label{fig:max_attainable_precision}
\end{figure}
Figure~\ref{fig:case9collapse} demonstrates the effect of successive
applications of the P-W transformation in a more extreme case. The calculation
was performed using \texttt{case9} from MATPOWER, in which a uniform scaling was
applied to all loads and generators in order to bring the system as close as
possible to collapse. The location of the critical point was narrowed down to a
scaling factor of $\lambda_\text{crit}=2.48539267\pm 1.0e-8)$, which was
confirmed using both HELM and MATPOWER's CPF method. Since the total load at the
critical point is around 315 MW, this error margin in the determination of the
exact value of $\lambda_\text{crit}$ corresponds to about 3 Watts. Therefore,
since Mvar limits are not enforced here, the limiting singularity
$s_\text{crit}$ is very close to, but slightly above 1. At stage zero (base HELM
method), it can be observed that the convergence rate is quite slow and the
maximum achievable precision seems to be around $\pm 10^{-2}$ pu. After 8 P-W
steps, the convergence rate is so fast that just 9 terms are enough to achieve
maximum machine precision.
\begin{figure}[!t]
\centering
\includegraphics[width=0.98\columnwidth]{figs/Figs_errors_vs_order_vs_PW_case9}
\caption{Convergence rates for \texttt{case9} (no Mvar limits enforced), where
load and generation have been scaled up as close as possible to the point of
collapse. Maximum attainable precision improves from about
$\pm 10^\text{-2}$ to $\pm 10^\text{-15}$ pu by performing 8 P-W steps.}
\label{fig:case9collapse}
\end{figure}
This behavior can be qualitatively understood in terms of the geometry of the
P-W transformation in $s$-space. Let us recall the parameter transformation,
rewriting~\eqref{eq:renorm} in this form:
\begin{equation*}
s' = \frac{s - s_0}{1-s_0}
\end{equation*}
It is easy to see that this conformal mapping is a translation and a dilation of
$s$-space, which has the effect of advancing and ``zooming in'' towards the
point $s=1$. Using conformal maps in order to push unwanted singularities away
and thus improve the convergence properties of Padé approximants is a well-known
technique that has been applied to general cases~\cite{BrunoReitich94}. Here it
will be shown how this particular transformation works in the HELM problem.
Figure~\ref{fig:s-space_transform} shows graphcally the effect that the
successive P-W transformations have on the singularities of the problem: all
non-relevant singularities get pushed away, eventually leaving the only one that
matters: the first one to be encountered on the positive real
axis~\cite{Trias15}. When this critical singularity is $s_\text{crit}>1$, the
successive P-W transformations will keep pushing it away until it is so far that
the power series is sufficienly well-conditioned to allow its Padé approximants
to converge quickly (one could even obtain a series whose radius of convergence
included 1, although this is not strictly necessary). If on the other hand
$s_\text{crit}<1$, the P-W procedure would ``pull in'' the singularity towards 0
and quickly get stuck (the successive values $s_0$ would approach zero), thus
signaling infeasibility. As a by-product, this would also provide a precise
value for $s_\text{crit}$. Finally, if $s_\text{crit}=1$ then all other
singularities are pushed away but $s_\text{crit}$ will always remain at a
distance of 1. The Padé approximants of every HELM subproblem will still
\emph{not} converge exactly at 1, but all that is needed is that they converge
\emph{near} 1. Then, thanks to the zooming property, one can get extremely
close to 1 \emph{in the original} $s$-space. For instance, assume a case where
one always obtains sufficient convergence at $s_0=0.9$ in every P-W stage. At
the second stage, the starting point $s''=0$ represents $s=0.99$. After 10 P-W
stages, the partial solution will be just $10^{-10}$ away from 1 in the original
$s$-space. This qualitatively explains the good numerical results that are
obtained in the next section.
\begin{figure}[!t]
\centering
\includegraphics[width=0.95\columnwidth]{figs/Figs_splane_transform}
\caption{How the P-W transformation acts in $s$-space. The graphs show a set
of singularities in: the original $s$-space (top); the $s'$-transformed
space, for $s_0=0.44$ (center); the $s''$-transformed space, for $s'_0=0.75$
(bottom). A grid is shown to aid the eye in appreciating the dilating
effects of successive P-W transformations. After the second P-W step, the
Padé approximants are much better conditioned to converge for $s''=1$.}
\label{fig:s-space_transform}
\end{figure}
\subsection{Examples with Q-limits}
Figure~\ref{fig:case9241pegase_withQlims} shows the convergence rates for
\texttt{case9241pegase} when Mvar limits are enforced using the P-W
method. Contrast this with Figure~\ref{fig:max_attainable_precision}, where
limits are not enforced. The presence of the complementarity constraints
introduces an unavoidable singularity at $s=1$, whose effects can be clearly
seen in the convergence rates. However, the application of the P-W technique
achieves the solution with good precision.
\begin{figure}[!t]
\centering
\includegraphics[width=0.98\columnwidth]{figs/Figs_errors_vs_order_as_PW-case9241pegase}
\caption{Convergence rates for \texttt{case9241} (unmodified, with Mvar limits
enforced). A precision of about $\pm 10^\text{-12}$ pu is achieved ater 26
P-W steps.}
\label{fig:case9241pegase_withQlims}
\end{figure}
Table~\ref{tab:PWresults} shows the results of applying the HELM-PW method to
the top 25 largest cases available in MATPOWER. The numerical tolerances
required for all cases were $10^{-11}$ pu maximum update error in voltages (at
each P-W stage), and $10^{-8}$ pu maximum mismatch error in the original power
flow equations (nodal currents). The maximum number of power series terms used
at each P-W stage was $N_\text{max}=32$. Note that in all these examples
regulation is always local to the bus, and that $Q^\text{min},Q^\text{max}$
represent the aggregate limits of all generators connected to the bus, in case
there is more than one.
The first column shows the number of P-W stages needed in order to meet the
required tolerance; this provides a rough idea of the numerical difficulty
involved in arriving to the solution. The next three columns show the number of
PV buses that end up saturated at either $Q^\text{min}$ or $Q^\text{max}$,
comparing the results of HELM vs.\ the two limit-enforcing strategies available
in MATPOWER's Newton-Raphson method (strategy 1 consists in type-switching all
violated buses at once, on every NR solution; while strategy 2 consists in
type-switching one bus at a time). One can observe how there are many cases
where the solutions obtained with MATPOWER are different, and a few cases where
NR does not converge. Additionally, HELM's solutions seem to consistently yield
a lower total number of saturated buses.
The table also shows the percentage of saturated buses over the total number of
PV buses. Interestingly, there seems to be no clear correlation between this
number and the number of P-W steps required, which probably indicates that the
numerical difficulty of the case may be influenced by other factors, such as
proximity to voltage collapse. This issue has been left for further research.
\setlength\tabcolsep{5pt}
\begin{table}[t]
\centering
\begin{tabular}{lccccr}
\toprule
& P-W & \multicolumn{3}{c}{sat.\ $Q^\text{min}/Q^\text{max}$} \\
\cmidrule(lr){3-5}
Case & steps & HELM & MP-1 & MP-2 & \% PV\\
\midrule
128 & 6 & 5/1 & = & = & 11.3 \\
145 & 6 & 0/1 & = & = & 2.0 \\
illinois200 & 8 & 0/13 & = & = & 35.1 \\
300 & 15 & 0/11 & = & = & 16.2 \\
300\_PS & 15 & 0/13 & = & = & 19.1 \\
1354pegase & 14 & 0/25 & = & = & 9.7 \\
1888rte & 23 & 13/2 & 13/\textbf{3} & = & 5.5 \\
1951rte & 22 & 4/9 & \textbf{8/7} & = & 3.7 \\
2383wp & 16 & 149/222 & \textbf{NC} & \textbf{NC} & 75.8 \\
2736sp & 19 & 67/99 & \textbf{73/102} & = & 53.6 \\
2737sop & 23 & 63/87 & \textbf{NC} & = & 56.8 \\
2746wop & 23 & 206/155 & \textbf{NC} & \textbf{NC} & 76.8 \\
2746wp & 18 & 98/235 & \textbf{107/239} & \textbf{99}/235 & 67.2 \\
2848rte & 15 & 27/32 & \textbf{30/34} & = & 15.7 \\
2868rte & 13 & 12/31 & \textbf{20/33} & \textbf{13}/31 & 10.2 \\
2869pegase & 19 & 0/72 & = & = & 14.1 \\
3012wp & 19 & 112/194 & \textbf{113}/194 & = & 66.0 \\
3120sp & 15 & 124/143 & \textbf{NC} & \textbf{NC} & 67.6 \\
3375wp & 18 & 117/182 & \textbf{118/184} & = & 48.3 \\
6468rte & 53 & 7/40 & \textbf{10/38} & = & 16.2 \\
6470rte & 38 & 34/38 & \textbf{36/37} & = & 15.5 \\
6495rte & 36 & 16/56 & \textbf{22/53} & = & 14.6 \\
6515rte & 60 & 11/68 & \textbf{14/66} & \textbf{12}/68 & 16.0 \\
9241pegase & 20 & 6/190 & \textbf{7}/190 & = & 13.6 \\
13659pegase & 10 & 0/1 & = & = & 0.02 \\
\bottomrule
\end{tabular}
\caption{Numerical results of the P-W method applied to the top 25 largest cases
available in MATPOWER. The MP-1 and MP-2 columns report the MATPOWER results
when using NR with the two alternative limit-enforcing strategies
(\texttt{enforce\_q\_lims} 1 and 2, respectively). Results equal to
HELM's are marked with (=), differences are
highlighed in bold, and non-convergence shown as \textbf{NC}.}
\label{tab:PWresults}
\end{table}
\subsection{Some preliminary results on performance}
A performance evaluation of the original HELM method, which did not have Q-limit
enforcement, was presented in~\cite{TriasMarin16}, using a MATLAB-based
implementation and comparing results to several methods available in
MATPOWER. Since our current implementation of the P-W method has not yet
undergone a proper optimization, its performance will not be reviewed
here. Instead, some scaling arguments will be given in order to roughly
characterize the expected performance.
As seen in table~\ref{tab:PWresults}, it is observed that the method requires a
number of P-W steps that varies between 6 and 60, for this sample of cases and a
choice of $N_\text{max}=32$. Since each P-W step involves roughly one full HELM
computation, this provides an approximate idea of the kind of performance one
may expect. There is a certain trade-off between $N_\text{max}$ and the
resulting number of P-W steps needed to achieve the required tolerance. If one
uses a large value for $N_\text{max}$ (e.g., more than 60 power series terms),
the advancing values $s_0$ at each P-W step will be larger, and therefore the
total amount of P-W steps needed will be lower. However, the HELM problem being
solved at each of those steps will be more costly, because there will be more
series terms to solve. This would involve solving more linear systems, with
their corresponding right-hand sides (which involve convolutions); and also
computing Padé approximants at larger orders. The authors have not yet tried to
find rigorously the optimum point for this trade-off, but initial numerical
results indicate that $N_\text{max}\approx 30$ gives good results.
\section{Conclusions and further research}
A new HELM-based method has been presented that deals with control limits in a
rigorous way, based on a novel Lagrangian formulation of the power flow
equations. The method has been developed in full detail for the case of Mvar
limits of voltage-controlling generators in AC networks, which is the most
important problem of this type. Like HELM, the method is direct, constructive,
and deterministic.
The Lagrangian approach allows framing the problem in terms of constrained
minimization, leading to a HELM scheme for calculating that minimum. Two
additional key ingredients are also necessary: on the one hand, a proper
embedding for setpoints and Q-limit values, in order to ensure consistency of
the HELM reference state at $s=0$. On the other, the application of the P-W
technique, which enables the possibility of calculating the analytic
continuation of the reference state to $s=1$ with good numerical precision. P-W
exploits a remarkable property of the power flow equations, whereby certain
changes of variables and the embedding parameter leave the equations formally
invariant. This allows one to calculate the analytic continuation of the power
series in several steps, leveraging the power flow equations at each step to
re-expand new auxiliary series. The net result is an analytic continuation along
a path, with far greater numerical stability.
In contrast to traditional iterative methods, this one directly avoids the
convergence problems derived from mutual interactions among controls, which can
produce oscillatory behavior, particularly when several nearby controls are
close to their saturation limits. The method presented here is closer in spirit
to some OPF-like approaches to power flow, although the differences are
fundamental, since HELM is based on algebraic curves and complex analysis.
The authors have also applied this method to other types of regulating devices
with limits, in the context of DC microgrids~\cite{TriasMarin16}. Examples are
the sequential shunt unit (SSU) that regulates the output voltage of solar
panels aboard spacecraft, and the voltage regulation of DC-DC converters. In
those cases the regulation can also be expressed in terms of algebraic
constraints involving new regulating injections, and their limits can be
accommodated under the same P-W technique even though they are sometimes
dependent on voltage.
Future areas to explore include the extensions to other controls, such as
automatic tap changers (ULTC transformers), automatic phase-shifting
transformers, or inter-area flow constraints. Shared control of a given point
originated from different buses is also a challenging goal. On the other hand,
the resemblance between the method presented here and certain OPF methods
certainly calls for a more detailed study on the possible applications to
optimization problems.
|
\section{Introduction\label{Sec1}}
Clock frequency estimation forms the basis of physics for precision measurements~\cite{Hall2006,Hansch2006,Margolis2014}, spectroscopy~\cite{Takamoto2003}, and astronomy~\cite{Steinmetz2008} etc.
It is also the heart of the modern technologies, from global position system (GPS)~\cite{Grewal2013} and magnetometers~\cite{Kitching2011,Budker2007} to inertial sensors~\cite{Kitching2011} (e.g., gyroscopes and gravimeters).
A two-level system (TLS) coupled with an oscillating field is the prototype of a clock.
This may correspond to an ensemble of atoms or an ion driven by an electromagnetic field in the microwave, optical or ultraviolet regions.
The TLS's tick with the oscillations of the electromagnetic wave whose frequency is locked to that of the energy difference between two discrete levels.
The closer the frequency of the external field to the transition frequency of the TLS, the larger contrast of the Rabi oscillation occurs.
Hence, the clock frequency can be estimated according to the TLS's population dynamics.
This model was first proposed by Rabi~\cite{Rabi1937} and widely used in atomic clocks~\cite{Hinkley2013,Bloom2014,Nicholson2015}.
The performances of an atomic clock are affected by many factors~\cite{Porsev2006,Taichenachev2006,Gibble2009}.
In realistic experiments, decoherence is unavoidable and plays an important role, especially when the uncertainty is at the level of $10^{-17}$~\cite{Rosenband2008,Chou2010} or even smaller~\cite{Huntemann2016}.
For an atomic clock, one of the most essential mechanisms of decoherence is spontaneous emission.
The effects of spontaneous emission may have great influence on the TLS's population dynamics and eventually degrade the performance of the clock.
Therefore, it is of great interest to investigate how the spontaneous emission affects the population dynamics and the clock frequency estimation.
Theoretically, a TLS driven by an oscillating field under spontaneous emission is well described by a Lindblad equation~\cite{Weinberg2016}.
Although the driven TLS's are extensively investigated, the search for exactly solvable driven TLS's is still an active field~\cite{DasSarma2012}.
Two famous examples of solvable driven TLS's are the Landau-Zener~\cite{Landau1932,Zener1932} and Rabi~\cite{Rabi1937} problems.
Solvable models on driven TLS's provide the reliable explanations for experimental observations as well as numerical calculations in quantum optics~\cite{AllenBook1992,BoydBook1992,MeystreBook2001,HarocheBook2006}, quantum metrology~\cite{Ivanov2013,Malossi2013,Tian2015,Ivanov2015}, and quantum control~\cite{Greilich2009,Poem2011,Bason2012,Sauer2013,Daems2013}.
Among the existed analytical methods for solving TLS's~\cite{DasSarma2012,Wu2007,Wu20072,Xie2010,Hai2013,Barnes2013,Messina2014}, much attention is paid to the driven TLS's described by Schr\"{o}dinger equations which fail to include the spontaneous emission.
The solvable driven TLS obeying Lindblad equation~\cite{BreuerBook2002,NielsenBook2000} is still of rareness (see Ref.~\cite{HanggiRev1998} for a review on dissipative driven TLS's).
Furthermore, the analytical study of decoherence is always hard and nontrivial for practical applications~\cite{Thorwart2000,Mogilevtsev2008,Sergi2013,Sauer2013,Zlatanov2015}.
Here, we study a TLS driven by an oscillating field under spontaneous emission.
Within the Markovian approximation, the system can be described by a Lindblad master equation.
Then, we solve the Lindblad equation analytically and give its exact solutions.
We analyze different regimes for the population dynamics and investigate the clock frequency estimation in the presence of spontaneous emission.
Our analytical results illustrate that, the spontaneous emission does not induce any frequency shift but the estimated precision of the transition frequency would be reduced.
This article is organized as follows.
In Sec.~\ref{Sec2}, we give the Bloch equations for our system.
In Sec.~\ref{Sec2A}, by implementing a linear transformation, we obtain an ordinary differential equations with constant coefficients for the system.
In Sec.~\ref{Sec2B} and~\ref{Sec2C}, we give the analytical solutions for the Lindblad equation.
In Sec.~\ref{Sec3}, we discuss the population dynamics and the clock frequency estimation in the presence of spontaneous emission.
Finally, a brief summary of our results is given in Sec.~\ref{Sec4}.
\section{Lindblad equation and its analytical solutions\label{Sec2}}
\begin{figure}[t]
\includegraphics[width=1.0\columnwidth]{Fig1.eps}
\caption{\label{Fig1}(Color online). Schematic diagram of a TLS coupled with an oscillating electromagnetic field under spontaneous emission. Here, $\Omega$ and $\omega_\mathrm{D}$ are the Rabi frequency and oscillation frequency of the driving field, $\omega_0$ is the transition frequency of the TLS and $\gamma$ stands for the spontaneous emission rate.}
\end{figure}
We consider two stable states $\ket{g}$ and $\ket{e}$ in an atomic clock.
By introducing the pseudospin-$1/2$ operators
$\hat S^+=\ket{e}\bra{g}$,
$\hat S^-=\ket{g}\bra{e}$,
$\hat S^x=\frac{1}{2}(\hat S^++\hat S^-)$,
$\hat S^y=\frac{1}{2i}(\hat S^+-\hat S^-)$,
and $\hat S^z=\frac{1}{2}(\ket{e}\bra{e}-\ket{g}\bra{g})$,
within the Markovian approximations, the density matrix operator of the system, $\hat\rho(t)$, is governed by a Lindblad equation (throughout this article, we set $\hbar=1$),
\begin{equation}\label{Eq_lindblad_S}
\tfrac{\mathrm d}{\mathrm dt}\hat\rho(t)=-i\big[\hat H(t),\hat\rho(t)\big]
+\mathcal{L}\big[\hat\rho(t)\big].
\end{equation}
The time-dependent Hamiltonian reads,
\begin{eqnarray}
\hat H(t)&=&\hat H_0+\hat H_1(t), \\
\hat H_0&=&\omega_0\hat S^z, \\
\hat H_1(t)&=&\Omega\big[\hat S^x\cos(\omega_\mathrm{D}t)+\hat S^y\sin(\omega_\mathrm{D}t)\big],
\end{eqnarray}
where $\hat H_0$ is the Hamiltonian of the TSL with transition frequency $\omega_0$,
and $\hat H_1(t)$ corresponds to the coupling between the TSL and the external electromagnetic field with Rabi frequency $\Omega$ and driving frequency $\omega_\mathrm{D}$.
The schematic is shown in Fig.~\ref{Fig1}.
Here, $\mathcal{L}$ is a superoperator which describes the spontaneous emission~\cite{BreuerBook2002,NielsenBook2000},
\begin{equation}\label{Eq_SO_S}
\mathcal{L}\big[\hat\rho\big]=\gamma\big(\hat S^-\hat\rho\hat S^+
-\tfrac{1}{2}\hat S^+\hat S^-\hat\rho-\tfrac{1}{2}\hat\rho\hat S^+\hat S^-\big),
\end{equation}
with $\gamma$ being the spontaneous emission rate.
In this article, we derive the equations in the Dirac picture.
By applying the transformation $\hat\rho^\mathrm{D}(t)=e^{i\hat H_0t}\hat\rho(t)e^{-i\hat H_0t}$,
the Lindblad equation~\eqref{Eq_lindblad_S} becomes
\begin{equation}\label{Eq_lindblad_D}
\tfrac{\mathrm d}{\mathrm dt}\hat\rho^\mathrm{D}(t)
=-i\big[\hat H_1^\mathrm{D}(t),\hat\rho^\mathrm{D}(t)\big]
+\mathcal{L}^\mathrm{D}\!\big[\hat\rho^\mathrm{D}(t)\big].
\end{equation}
Here, by using the identity $e^{i\hat H_0 t}\hat S^\pm e^{-i\hat H_0 t}=e^{\pm i\omega_0 t}\hat S^\pm$,
\begin{equation}\label{Eq_Driving_D}
\hat H_1^\mathrm{D}(t)=\tfrac{1}{2}\Omega\big(\hat S^+e^{-i\Delta t}+\hat S^-e^{i\Delta t}\big)
\end{equation}
and
\begin{equation}\label{Eq_SO_D}
\mathcal{L}^\mathrm{D}\!\!\left[\hat\rho^\mathrm{D}\right]=\gamma\big(
\hat S^-\hat\rho^\mathrm{D}\hat S^+-\tfrac{1}{2}\hat S^+\hat S^-\hat\rho^\mathrm{D}
-\tfrac{1}{2}\hat\rho^\mathrm{D}\hat S^+\hat S^-\big),
\end{equation}
where the detuning is defined as $\Delta=\omega_\mathrm{D}-\omega_0$.
Eq.~\eqref{Eq_lindblad_D} in the basis of eigenstates of $\hat S^z$ becomes
\begin{equation}\label{Eq_Matrix_Form}
\!\left\{\!\!\begin{array}{l}
\frac{\mathrm d}{\mathrm dt}\rho^{\mathrm{D}00}
=\frac{1}{2i}\Omega\left(e^{-i\Delta t}\rho^{\mathrm{D}10}-e^{i\Delta t}\rho^{\mathrm{D}01}\right)-\gamma\rho^{\mathrm{D}00}, \\
\frac{\mathrm d}{\mathrm dt}\rho^{\mathrm{D}01}
=\frac{1}{2i}\Omega e^{-i\Delta t}\left(\rho^{\mathrm{D}11}-\rho^{\mathrm{D}00}\right)-\frac{1}{2}\gamma\rho^{\mathrm{D}01}, \\
\frac{\mathrm d}{\mathrm dt}\rho^{\mathrm{D}10}
=-\frac{1}{2i}\Omega e^{i\Delta t}\left(\rho^{\mathrm{D}11}-\rho^{\mathrm{D}00}\right)-\frac{1}{2}\gamma\rho^{\mathrm{D}10}, \\
\frac{\mathrm d}{\mathrm dt}\rho^{\mathrm{D}11}
=-\frac{1}{2i}\Omega\left(e^{-i\Delta t}\rho^{\mathrm{D}10}-e^{i\Delta t}\rho^{\mathrm{D}01}\right)+\gamma\rho^{\mathrm{D}00},
\end{array}\right.\!\!\!\!\!\!\!\!
\end{equation}
with
\begin{equation}
\rho^\mathrm{D}(t)=\left(\begin{array}{cc}
\rho^\mathrm{D00}(t) & \rho^\mathrm{D01}(t) \\
\rho^\mathrm{D10}(t) & \rho^\mathrm{D11}(t)
\end{array}\right).
\end{equation}
Next, introducing the Bloch vector,
$\mathbf{R}=(u,v,w)^\mathrm{T}$ (where the superscript ``T" denotes the operation of matrix transpose),
the density matrix can be written as
\begin{equation}
\hat\rho^\mathrm{D}=\tfrac{1}{2}\left(\hat 1+u\hat\sigma^x+v\hat\sigma^y+w\hat\sigma^z\right),
\end{equation}
where $\hat\sigma^x$, $\hat\sigma^y$ and $\hat\sigma^z$ are the Pauli matrices.
The three coefficients can be expressed in the form of
\begin{equation}\label{Eq_Pauli_uvw}
\left\{\!\!\begin{array}{ccccl}
u&\!\!=\!\!&\mathrm{tr}[\hat\rho^\mathrm{D}\hat\sigma^x]&\!\!=\!\!&
\cos(\omega_0t)\left<\hat\sigma^x\right>+\sin(\omega_0t)\left<\hat\sigma^y\right>,\\
v&\!\!=\!\!&\mathrm{tr}[\hat\rho^\mathrm{D}\hat\sigma^y]&\!\!=\!\!&
-\sin(\omega_0t)\left<\hat\sigma^x\right>+\cos(\omega_0t)\left<\hat\sigma^y\right>,\\
w&\!\!=\!\!&\mathrm{tr}[\hat\rho^\mathrm{D}\hat\sigma^z]&\!\!=\!\!&
\left<\hat\sigma^z\right>.
\end{array}\right.\!\!\!\!\!\!
\end{equation}
Here, $\left<\hat\sigma^\alpha\right>=\mathrm{tr}[\hat\rho(t)\hat\sigma^\alpha]$ is the expectation value of $\hat\sigma^\alpha$ for the evolved state $\hat\rho(t)$ ($\alpha=x,y,z$).
Then, we obtain the Bloch equations
\begin{equation}\label{Eq_Bloch_Original}
\left\{\!\!\begin{array}{l}
\frac{\mathrm{d}}{\mathrm{d}t}u
=-\frac{1}{2}\gamma u+\Omega\sin(\Delta t)w, \\
\frac{\mathrm{d}}{\mathrm{d}t}v
=-\frac{1}{2}\gamma v-\Omega\cos(\Delta t)w, \\
\frac{\mathrm{d}}{\mathrm{d}t}w
=-\Omega\sin(\Delta t)u+\Omega\cos(\Delta t)v-\gamma w-\gamma.
\end{array}\right.
\end{equation}
The Bloch equations~\eqref{Eq_Bloch_Original} can be written in a more compact form
\begin{equation}\label{Eq_Bloch_Original_Compact}
\tfrac{\mathrm{d}}{\mathrm{d}t}\mathbf{R}=\mathbf{M}\mathbf{R}+\mathbf{R}_0,
\end{equation}
where $\mathbf{R}_0=-\gamma(0,0,1)^\mathrm{T}$, and the coefficient matrix is defined as
\begin{equation}\label{Eq_Bloch_M}
\mathbf{M}=\left(\begin{array}{ccc}
-\frac{1}{2}\gamma & 0 & \Omega\sin(\Delta t) \\
0 & -\frac{1}{2}\gamma & -\Omega\cos(\Delta t) \\
-\Omega\sin(\Delta t) & \Omega\cos(\Delta t) & -\gamma
\end{array}\right).
\end{equation}
\subsection{Linear transformation\label{Sec2A}}
In general, when the coefficient matrix $\mathbf{M}$~\eqref{Eq_Bloch_M} is time-dependent, it is not easy to solve the Bloch equation~\eqref{Eq_Bloch_Original_Compact} directly.
Such differential equations with periodically oscillating coefficients can be dealt by Floquet theory~\cite{EasthamBook1973}.
However, the obtained solutions are usually given in a form of a series, which may be sometimes inconvenient.
Here, we present another simpler approach, which enables us to obtain the completely and analytical solutions in closed-form expressions.
First, we take a general reversible linear transformation on the Bloch vectors,
\begin{equation}\label{Eq_linear_trans}
\mathbf{R}=\mathbf{P}\mathbf{R}^Q,
\end{equation}
where $\mathbf{R}^Q=(u^Q,v^Q,w^Q)^\mathrm{T}$ is a new set of variables,
and $\mathbf{P}$ is an undetermined $3 \times 3$ matrix with time-dependent matrix elements.
Meanwhile, Eq.~\eqref{Eq_Bloch_Original_Compact} becomes
\begin{equation}\label{Eq_Bloch_RQ}
\tfrac{\mathrm{d}}{\mathrm{d}t}\mathbf{R}^Q=\mathbf{F}\mathbf{R}^Q+\mathbf{G},
\end{equation}
with
\begin{eqnarray}\label{FG}
&& \mathbf{F}=\mathbf{P}^{-1}\mathbf{M}\mathbf{P}
-\mathbf{P}^{-1}\tfrac{\mathrm{d}}{\mathrm{d}t}\mathbf{P}, \\
&& \mathbf{G}=\mathbf{P}^{-1}\mathbf{R}_0.
\end{eqnarray}
The differential equation~\eqref{Eq_Bloch_RQ} can be simplified into a linear one with constant coefficients
if $\mathbf{F}$ is time-independent,
or equivalently $\tfrac{\mathrm{d}}{\mathrm{d}t}\mathbf{F}=0$.
To let $\tfrac{\mathrm{d}}{\mathrm{d}t}\mathbf{F}=0$, it is not unique to choose the transformation matrix $\mathbf{P}$.
Here, as an example, we set $\mathbf{P}$ in the form of
\begin{equation}\label{Eq_linear_trans_P}
\mathbf{P}=e^{-\gamma t}
\left(\begin{array}{ccc}
\cos(\Delta t) & -\sin(\Delta t) & 0 \\
\sin(\Delta t) & \cos(\Delta t) & 0 \\
0 & 0 & 1
\end{array}\right),
\end{equation}
according to $\mathbf{M}$~\eqref{Eq_Bloch_M}.
Substituting $\mathbf{P}$~\eqref{Eq_linear_trans_P} into Eq.~\eqref{FG}, one can easily obtain
\begin{equation}\label{Eq_Matrix_F}
\mathbf{F}=\left(\begin{array}{ccc}
\frac{1}{2}\gamma & \Delta & 0 \\
-\Delta & \frac{1}{2}\gamma & -\Omega \\
0 & \Omega & 0
\end{array}\right)
\end{equation}
and $\mathbf{G}=-\gamma e^{\gamma t}(0,0,1)^\mathrm{T}$, respectively. Here, $\mathbf{F}$~\eqref{Eq_Matrix_F} satisfies the desired condition $\tfrac{\mathrm{d}}{\mathrm{d}t}\mathbf{F}=0$.
Until by now, we have simplified the Bloch equation~\eqref{Eq_Bloch_Original_Compact} by a straight linear transformation and the new differential equation can be solved much more easily.
\subsection{Solutions with zero detuning\label{Sec2B}}
At resonance $\omega_D=\omega_0$ (i.e., zero detuning $\Delta=0$), Eq.~\eqref{Eq_Bloch_RQ} can be reduced to
\begin{equation}\label{Eq_Bloch_Zero_detuning}
\left\{\!\!\begin{array}{l}
\frac{\mathrm{d}}{\mathrm{d}t}u^Q=\frac{1}{2}\gamma u^Q, \\
\frac{\mathrm{d}}{\mathrm{d}t}v^Q=\frac{1}{2}\gamma v^Q-\Omega w^Q, \\
\frac{\mathrm{d}}{\mathrm{d}t}w^Q=\Omega v^Q-\gamma e^{\gamma t}.
\end{array}\right.
\end{equation}
In this case, it is easy to find that $u^Q(t)=u_0 e^{\frac{1}{2}\gamma t}$, $v^Q(t)=\frac{1}{\Omega}[\frac{\mathrm d}{\mathrm dt}w^Q(t)+\gamma e^{\gamma t}]$,
and
\begin{equation}\label{Eq_Zero_detuning_f}
\tfrac{\mathrm d^2}{\mathrm dt^2}w^Q(t)
-\tfrac{1}{2}\gamma\tfrac{\mathrm d}{\mathrm dt}w^Q(t)
+\Omega^2 w^Q(t)
+\tfrac{1}{2}\gamma^2 e^{\gamma t}=0.
\end{equation}
A special solution to Eq.~\eqref{Eq_Zero_detuning_f} is
$w^Q_s(t)=-\tfrac{\gamma^2}{\gamma^2+2\Omega^2} e^{\gamma t}$.
The characteristic equation of Eq.~\eqref{Eq_Zero_detuning_f} is
\begin{equation}\label{Eq_Zero_detuning_chara}
\lambda^2-\tfrac{1}{2}\gamma\lambda+\Omega^2=0,
\end{equation}
We denote the discriminant of~\eqref{Eq_Zero_detuning_chara} as
$D_0=\frac{1}{4}\gamma^2-4\Omega^2$,
then, the general solution of Eq.~\eqref{Eq_Zero_detuning_f} is \\
(\romanNum{1}) for $D_0>0$ ($\gamma>4\Omega$),
\begin{equation}\label{Eq_Bloch_Zero_D0>0}
w^Q(t)=C_1e^{\lambda_1t}+C_2e^{\lambda_2t}+w^Q_s(t);
\end{equation}
(\romanNum{2}) for $D_0=0$ ($\gamma=4\Omega$),
\begin{equation}\label{Eq_Bloch_Zero_D0=0}
w^Q(t)=e^{\eta t}(C_1+C_2t)+w^Q_s(t);
\end{equation}
(\romanNum{3}) for $D_0<0$ ($0\le\gamma<4\Omega$),
\begin{equation}\label{Eq_Bloch_Zero_D0<0}
w^Q(t)=e^{\eta t}\big[C_1\cos(\omega t)+C_2\sin(\omega t)\big]+w^Q_s(t).
\end{equation}
Here, in Eq.~\eqref{Eq_Bloch_Zero_D0>0},
$\lambda_{1,2}=\frac{1}{4}\gamma\pm\frac{1}{2}\sqrt{D_0}$ ($\lambda_1>\lambda_2$)
are two distinct roots of Eq.~\eqref{Eq_Zero_detuning_chara};
in Eq.~\eqref{Eq_Bloch_Zero_D0=0}, $\eta=\frac{1}{4}\gamma$ is the double root;
and in Eq.~\eqref{Eq_Bloch_Zero_D0<0}, $\eta\pm i\omega=\frac{1}{4}\gamma\pm i\frac{1}{2}\sqrt{-D_0}$
are two conjugate complex roots
($\eta=\frac{1}{4}\gamma$ and $\omega=\frac{1}{2}\sqrt{-D_0}$).
The coefficients ($C_1$,~$C_2$) appear in Eqs.~\eqref{Eq_Bloch_Zero_D0>0}-\eqref{Eq_Bloch_Zero_D0<0}
are determined via the initial conditions (see Appendix~\ref{SecApp1}).
So far, we completely solve the Bloch equation~\eqref{Eq_Bloch_Original_Compact} in the zero detuning case ($\Delta=0$).
\subsection{Solutions with nonzero detuning\label{Sec2C}}
For the case of nonzero detuning ($\Delta\ne 0$), Eq.~\eqref{Eq_Bloch_RQ} becomes
\begin{equation}\label{Eq_Bloch_NoneZero_detuning}
\left\{\!\!\begin{array}{l}
\frac{\mathrm{d}}{\mathrm{d}t}u^Q=\frac{1}{2}\gamma u^Q+\Delta v^Q, \\
\frac{\mathrm{d}}{\mathrm{d}t}v^Q=-\Delta u^Q+\frac{1}{2}\gamma v^Q-\Omega w^Q, \\
\frac{\mathrm{d}}{\mathrm{d}t}w^Q=\Omega v^Q-\gamma e^{\gamma t}.
\end{array}\right.
\end{equation}
Here, we have $v^Q(t)=\frac{1}{\Omega}[\frac{\mathrm d}{\mathrm dt}w^Q(t)+\gamma e^{\gamma t}]$,
$u^Q(t)=\frac{1}{\Delta\Omega}[-\frac{\mathrm{d}^2}{\mathrm{d}t^2}w^Q(t)
+\frac{1}{2}\gamma\frac{\mathrm d}{\mathrm dt}w^Q(t)-\Omega^2w^Q(t)
-\frac{1}{2}\gamma^2e^{\gamma t}]$, and
\begin{eqnarray}\label{Eq_Nonzero_detuning_f}
\tfrac{\mathrm{d}^3}{\mathrm{d}t^3}w^Q(t)
&-&\gamma\tfrac{\mathrm{d}^2}{\mathrm{d}t^2}w^Q(t)
+\left(\Delta^2+\Omega^2+\tfrac{1}{4}\gamma^2\right)\tfrac{\mathrm{d}}{\mathrm{d}t}w^Q(t)
\nonumber \\
&-&\tfrac{1}{2}\Omega^2\gamma w^Q(t)
+\gamma\left(\Delta^2+\tfrac{1}{4}\gamma^2\right)e^{\gamma t}=0.
\end{eqnarray}
A special solution to Eq.~\eqref{Eq_Nonzero_detuning_f} is
$w^Q_s(t)=-\tfrac{4\Delta^2+\gamma^2}{4\Delta^2+\gamma^2+2\Omega^2}e^{\gamma t}$.
The characteristic equation of Eq.~\eqref{Eq_Nonzero_detuning_f} is
\begin{equation}\label{Eq_Nonzero_detuning_chara}
\lambda^3-\gamma\lambda^2+\left(\Delta^2+\Omega^2+\tfrac{1}{4}\gamma^2\right)\lambda-\tfrac{1}{2}\Omega^2\gamma=0.
\end{equation}
Let $\lambda=\lambda'+\frac{1}{3}\gamma$, the characteristic equation~\eqref{Eq_Nonzero_detuning_chara} becomes
\begin{equation}\label{Eq_Nonzero_detuning_chara_2}
\lambda'^3+p\lambda'+q=0,
\end{equation}
with $p=\Delta^2+\Omega^2-\frac{1}{12}\gamma^2$ and $q=\frac{1}{108}\gamma(36\Delta^2-18\Omega^2+\gamma^2)$.
We denote the discriminant of \eqref{Eq_Nonzero_detuning_chara_2} as $D=\frac{1}{4}q^2+\frac{1}{27}p^3
=\frac{1}{432}[\Delta^2\gamma^4+(8\Delta^4-20\Delta^2\Omega^2-\Omega^4)\gamma^2+16(\Delta^2+\Omega^2)^3]$, then the general solution of Eq.~\eqref{Eq_Nonzero_detuning_f} is \\
(\romanNum{1}) for $D=0$ and $p=0$,
\begin{equation}\label{Eq_Bloch_Nonzero_1}
w^Q(t)=e^{\lambda_1 t}(C_1+C_2t+C_3t^2)+w^Q_s(t);
\end{equation}
(\romanNum{2}) for $D=0$ and $p\ne0$,
\begin{equation}\label{Eq_Bloch_Nonzero_2}
w^Q(t)=C_1e^{\lambda_1 t}+e^{\lambda_2 t}(C_2+C_3t)+w^Q_s(t);
\end{equation}
(\romanNum{3}) for $D<0$,
\begin{equation}\label{Eq_Bloch_Nonzero_3}
w^Q(t)=C_1e^{\lambda_1 t}+C_2e^{\lambda_2 t}+C_3e^{\lambda_3 t}+w^Q_s(t);
\end{equation}
(\romanNum{4}) for $D>0$,
\begin{equation}\label{Eq_Bloch_Nonzero_4}
w^Q(t)=C_1e^{\lambda_1 t}+e^{\eta t}\big[C_2\cos(\omega t)+C_3\sin(\omega t)\big]+w^Q_s(t).
\end{equation}
Here, in Eq.~\eqref{Eq_Bloch_Nonzero_1}, $\lambda_1=\frac{1}{3}\gamma$ is the triple root of Eq.~\eqref{Eq_Nonzero_detuning_chara};
in Eq.~\eqref{Eq_Bloch_Nonzero_2}, $\lambda_1=\frac{1}{3}\gamma-(4q)^{1/3}$ is the single root and $\lambda_2=\frac{1}{3}\gamma+(\frac{1}{2}q)^{1/3}$ the double one;
in Eq.~\eqref{Eq_Bloch_Nonzero_3}, $\lambda_1=\frac{1}{3}\gamma+r\cos(\frac{1}{3}\phi)$, $\lambda_2=\frac{1}{3}\gamma+r\cos[\frac{1}{3}(\phi+2\pi)]$, $\lambda_3=\frac{1}{3}\gamma+r\cos[\frac{1}{3}(\phi-2\pi)]$ are the three distinct real roots [where, $r=2(-\frac{1}{3}p)^{1/2}$ and $\phi=\arccos(-4qr^{-3})$];
in Eq.~\eqref{Eq_Bloch_Nonzero_4}, $\lambda_1=\frac{1}{3}\gamma+R_1+R_2$ is the real root while $\eta\pm i\omega$ the two conjugate complex roots with $\eta=\frac{1}{3}\gamma-\frac{1}{2}(R_1+R_2)$ and $\omega=\frac{\sqrt{3}}{2}(R_1-R_2)$ [where $R_1=(-\frac{1}{2}q+\sqrt{D})^{1/3}$, $R_2=(-\frac{1}{2}q-\sqrt{D})^{1/3}$].
The coefficients ($C_1$,~$C_2$,~$C_3$) appear in Eqs.~\eqref{Eq_Bloch_Nonzero_1}-\eqref{Eq_Bloch_Nonzero_4}
are determined via the initial conditions (see Appendix~\ref{SecApp2}).
Till here, we completely solve the Bloch equation~\eqref{Eq_Bloch_Original_Compact} in the nonzero detuning case ($\Delta\neq0$).
\section{Clock frequency estimation\label{Sec3}}
In the previous section, we have mathematically solved the Bloch equation~\eqref{Eq_Bloch_Original_Compact} for both zero ($\Delta=0$) and nonzero ($\Delta\ne 0$) detuning cases and found out their exact analytical solutions.
In this section, we mainly discuss the physics behind the analytical solutions and show how the spontaneous emission affects the clock frequency estimation within our model.
\subsection{Rabi oscillating, Damped oscillating and overdamping regimes\label{Sec3A}}
In the absence of spontaneous emission ($\gamma=0$), the TLS performs the Rabi oscillation.
The population difference $w(t)$ evolves as
\begin{eqnarray}\label{Eq_gamma_0}
w(t)&=&\tfrac{1}{\omega_\mathrm{R}^2}\Big[
\Delta(-\Omega u_0+\Delta w_0)
+\omega_\mathrm{R}\Omega v_0\sin(\omega_\mathrm{R} t)\nonumber \\
&&+\Omega(\Delta u_0+\Omega w_0)\cos(\omega_\mathrm{R} t)\Big],
\end{eqnarray}
with the total Rabi frequency $\omega_\mathrm{R}=\sqrt{\Delta^2+\Omega^2}$.
For the initial state $\rho(0)=\ket{g}\bra{g}$, one has $\rho^{\mathrm{D}}(0)=\rho(0)$,
and $(u_0,v_0,w_0)^{\mathrm T}=(0,0,-1)^{\mathrm T}$.
The solution for $w(t)$~\eqref{Eq_gamma_0} is reduced to
\begin{equation}
w(t)=-\frac{\Delta^2}{\omega_\mathrm{R}^2}
-\frac{\Omega^2}{\omega_\mathrm{R}^2}\cos(\omega_\mathrm{R} t).
\end{equation}
The excited population in $\ket{e}$ is given as
\begin{equation}
P_{e}(t)=\frac{1+w(t)}{2}=\frac{\Omega^2}{\omega_\mathrm{R}^2}\sin^2\left(\frac{\omega_\mathrm{R} t}{2}\right),
\end{equation}
which is consistent with the well-known Rabi oscillation.
\begin{figure}[t]
\includegraphics[width=1.0\columnwidth]{Fig2.eps}
\caption{\label{Fig2}(Color online). The phase diagram of three different types of solutions. The black line lying at $\gamma/\Omega=0$ represents the Rabi oscillating solutions. The blue area is the damped oscillating regime. The red area as well as the yellow curves $b_1$, $b_2$, $b_3$, denotes the overdamping regime.}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=2.0\columnwidth]{Fig3.eps}
\caption{\label{Fig3}(Color online). (a) Evolution of excited population at resonance ($\Delta=0$) under different spontaneous emission rates $\gamma=0$ (black), $0.001$ (blue), $0.005$ (red), $0.01$ (green), and $0.02$ (magenta). (b) Evolution of excited population under spontaneous emission ($\gamma=0.005$) with different detunings $\Delta=0$ (black), $0.05$ (blue), $0.1$ (red), $0.15$ (green), and $0.2$ (magenta). (c) The spectrum of maximal excited population versus detuning $\Delta$ under different spontaneous emission rates $\gamma=0$ (black), $0.001$ (blue), $0.005$ (red), $0.01$ (green), and $0.02$ (magenta). Inset shows the relationship of the relative full width at half maximum (FWHM) of the spectrum $\Delta\tilde\omega_{\text{FWHM}}$ versus spontaneous emission rate $\gamma$. Here, we set $\omega_0=10$ and $\Omega=0.1$.}
\end{figure*}
In the presence of spontaneous emission ($\gamma\ne0$), there are two types of solutions corresponding to the damped oscillating and the overdamping regimes, respectively.
In damped oscillating regime, the excited population oscillates with damped amplitude. While in overdamping regime, the excited population decays exponentially.
These two regimes can be distinguished by the discriminant of the characteristic equation.
For $D_0<0$ with $\Delta=0$ or $D>0$ with $\Delta\ne0$, the system is in the damped oscillating regime.
On the contrary, for $D_0\ge0$ with $\Delta=0$ or $D\le0$ with $\Delta\ne0$, the system is in the overdamping regime.
After some algebra, the criteria for determining the damped oscillating regime can be summarized as
\begin{equation}\label{Eq_damp_osci_cond}
\Delta^2\gamma^4+(8\Delta^4-20\Delta^2\Omega^2-\Omega^4)\gamma^2+16(\Delta^2+\Omega^2)^3>0.
\end{equation}
Otherwise, when Eq.~\eqref{Eq_damp_osci_cond} violates, the system is overdamping.
In Fig.~\ref{Fig2}, we give the phase diagram of Rabi oscillating, damped oscillating and overdamping regimes.
The damped oscillating regime is colored in blue and bounded by four curves:\\
(i) the black line,
\begin{equation}
\gamma=0;
\end{equation}
(ii) the yellow curve $b_2$ ($0\le|\Delta/\Omega|\le\frac{1}{2\sqrt{2}}$),
\begin{equation}
\gamma=\sqrt{14\Omega^2-4\Delta^2+4\Omega\tfrac{\Omega^2-8\Delta^2}{\Omega+\sqrt{\Omega^2-8\Delta^2}}};
\end{equation}
(iii) and (iv) the yellow curves $b_1$ ($\Delta>0$) and $b_3$ ($\Delta<0$) (both with $0<|\Delta/\Omega|\le\frac{1}{2\sqrt{2}}$),
\begin{equation}
\gamma=\sqrt{14\Omega^2-4\Delta^2+4\Omega\tfrac{\Omega^2-8\Delta^2}{\Omega-\sqrt{\Omega^2-8\Delta^2}}}.
\end{equation}
The red area corresponds to the overdamping regime.
Here, the intersection point of $b_1$ and $b_2$ is $(\frac{\gamma}{\Omega},\frac{\Delta}{\Omega})=(\frac{9}{\sqrt{6}},\frac{1}{2\sqrt{2}})$ while the intersection point of $b_1$ and $b_3$ is $(\frac{\gamma}{\Omega},\frac{\Delta}{\Omega})=(\frac{9}{\sqrt{6}},-\frac{1}{2\sqrt{2}})$.
The phase diagram is useful for clock frequency estimation.
When the system is overdamping, the population will not oscillate and the clock frequency cannot be inferred.
Therefore, in order to extract the clock frequency via the population oscillation under spontaneous emission, one should make sure that the system is always in the damped oscillating regimes.
The spontaneous emission rate should satisfy the condition that
\begin{equation}\label{gamma}
0<\gamma<\sqrt{14\Omega^2-4\Delta^2+4\Omega\tfrac{\Omega^2-8\Delta^2}{\Omega+\sqrt{\Omega^2-8\Delta^2}}},
\end{equation}
i.e., $\gamma$ should be on the left side of curve $b_2$ in Fig.~\ref{Fig2}.
Meanwhile, the corresponding analytic solution can be given as (derived from Sec.~\ref{Sec2} B and C)
\begin{equation}\label{Eq_Bloch_damped_oscillating}
w^Q(t)=C_1e^{\lambda_1 t}+e^{\eta t}\big[C_2\cos(\omega t)+C_3\sin(\omega t)\big]+w^Q_s(t).
\end{equation}
Here, $w^Q_s(t)=-\tfrac{4\Delta^2+\gamma^2}{4\Delta^2+\gamma^2+2\Omega^2}e^{\gamma t}$, $\lambda_1=\frac{1}{3}\gamma+R_1+R_2$, $\eta=\frac{1}{3}\gamma-\frac{1}{2}(R_1+R_2)$ and $\omega=\frac{\sqrt{3}}{2}(R_1-R_2)$ with $R_1=(-\frac{1}{2}q+\sqrt{D})^{1/3}$, $R_2=(-\frac{1}{2}q-\sqrt{D})^{1/3}$, $p=\Delta^2+\Omega^2-\frac{1}{12}\gamma^2$, $q=\frac{1}{108}\gamma(36\Delta^2-18\Omega^2+\gamma^2)$, and $D=\frac{1}{4}q^2+\frac{1}{27}p^3$.
The coefficients $(C_1,C_2,C_3)$ are given by the solution of a linear equation (see Appendix~\ref{SecApp2}),
\begin{equation}
\left(\!\begin{array}{ccc}
\lambda_1^2 & \eta^2-\omega^2 & 2\eta\omega \\
\lambda_1 & \eta & \omega\\
1 & 1 & 0
\end{array}\!\right)\!\!\!
\left(\!\begin{array}{c}
C_1 \\ C_2 \\ C_3
\end{array}\!\right)
=\left(\!\begin{array}{c}
B_1 \\ B_2 \\ B_3
\end{array}\!\right),
\end{equation}
where $B_1=-\Delta\Omega u_0+\frac{1}{2}\Omega\gamma v_0-\Omega^2w_0-(w^Q_0+1)\gamma^2$, $B_2=\Omega v_0-(w^Q_0+1)\gamma$, and $B_3=w_0-w^Q_0$ with $w^Q_0=w^Q_s(0)$.
\subsection{Clock frequency estimation in damped oscillating regime\label{Sec3B}}
In this section, we discuss how to estimate the clock frequency of the TLS in the presence of spontaneous emission.
The TLS initials from $\ket{g}$ at time $t=0$, and the population would oscillates when the driving field is applied.
Substituting the initial condition $(u_0,v_0,w_0)^\mathrm{T}=(0,0,-1)^\mathrm{T}$ into Eq.~\eqref{Eq_Pauli_uvw}, we have
\begin{equation}\label{Eq_mean_Pauli}
\left\{\!\!\begin{array}{ccl}
\left<\hat\sigma^x\right>&\!\!=\!\!&\cos(\omega_0t)u(t)-\sin(\omega_0t)v(t),\\
\left<\hat\sigma^y\right>&\!\!=\!\!&\sin(\omega_0t)u(t)+\cos(\omega_0t)v(t),\\
\left<\hat\sigma^z\right>&\!\!=\!\!&w(t),
\end{array}\right.
\end{equation}
where $\left<\hat\sigma^z\right>$ represents the population difference between $\ket{e}$ and $\ket{g}$.
When $\gamma$ is modest which satisfies Eq.~\eqref{gamma}, the system is in damped oscillating regime and the evolved excited population $P_e(t)$ reads
\begin{equation}
P_e(t)=\frac{1+\left<\hat\sigma^z\right>}{2}
=\frac{1+w(t)}{2}
=\frac{1+w^Q(t)e^{-\gamma t}}{2}.
\end{equation}
with $w^Q(t)$ being given according to Eq.~\eqref{Eq_Bloch_damped_oscillating}.
Here, the evolution of the excited population $P_e(t)$ is determined by both detuning $\Delta$ and spontaneous emission rate $\gamma$.
If the system is at resonance, i.e., $\omega_D=\omega_0$, the spontaneous emission shrinks the amplitude (or the contrast) of the excited population, see Fig.~\ref{Fig3}~(a).
When $\gamma=0$, the excited population oscillates sinusoidally from 0 to 1 with Rabi frequency $\Omega$.
When $\gamma>0$, the excited population oscillates with damped amplitudes. The excited population would cease to oscillate when the evolution time is long enough.
As $\gamma$ increases, the maximal amplitude decreases rapidly.
The detuning affects both the amplitude (or the contrast) and the period of $P_e(t)$. The evolution of the excited population with different detuning under $\gamma=0.005$ is shown in Fig.~\ref{Fig3}~(b).
As the detuning $\Delta$ gets larger, the maximal amplitude drops and the period of the oscillation becomes shorter.
From the responses of the excited population with different driving frequency $\omega_D$, one can extract the information of clock frequency $\omega_0$.
We calculate the evolved excited population with different detuning $\Delta$ and pick up the maximal amplitude $P_{e,\text{max}}$.
In Fig.~\ref{Fig3}~(c), we show the spectrum of maximal excited population $P_{e,\text{max}}$ versus detuning $\Delta$ with different spontaneous emission rates $\gamma$.
The peaks with different $\gamma$ are all centered at $\Delta=0$, which indicates that no additional frequency shift is induced by the spontaneous emission and $\omega_0$ can be inferred by tuning the driving frequency $\omega_D$ with largest $P_{e,\text{max}}$.
However, the height of the peak decreases as the effects of spontaneous emission becomes stronger.
The peak also becomes less sharper when $\gamma$ increases.
It means that, the process of spontaneous emission would have a negative influence on the estimated precision of $\omega_0$.
To characterize the estimated precision of $\omega_0$ quantitatively, we use the full width at half maximum
(FWHM) of the spectrum denoted by $\Delta\omega_{\text{FWHM}}$.
In the inset of Fig.~\ref{Fig3}~(c), we show the dependence of relative FWHM $\Delta\tilde{\omega}_{\text{FWHM}}$ on the spontaneous emission $\gamma$.
Here, the relative FWHM is defined as the FWHM respect to the one for $\gamma=0$,
\begin{equation}
\Delta\tilde{\omega}_{\text{FWHM}}(\gamma)=\frac{\Delta\omega_{\text{FWHM}}(\gamma)}{\Delta\omega_{\text{FWHM}}(\gamma=0)}.
\end{equation}
It is shown that, $\Delta\tilde{\omega}_{\text{FWHM}}$ gets larger when $\gamma$ increases, which indicates that the estimated precision of the clock frequency becomes worse when the spontaneous emission gradually comes into play.
Although the spontaneous emission reduces the measurement precision, the clock frequency is not shifted and can still be estimated when the spontaneous emission rate is modest.
\section{Summary\label{Sec4}}
In summary, we have explored the dynamical evolution of a driven TLS under spontaneous emission and illustrate how to perform the clock frequency estimation based on this model.
We derive the Bloch equations for the Lindblad equations in the Dirac picture and give the completely and exactly analytical solutions with closed-form expressions.
In the absence of spontaneous emission, our results recover the perfect Rabi oscillation.
In the presence of spontaneous emission, the system may be in damped oscillating or overdamping regime dependent on both the spontaneous emission rate and the detuning.
We analytically give the boundaries of these regimes and show that how a driven TLS under spontaneous emission can be used for clock frequency estimation.
We find that, the spontaneous emission does not cause additional frequency shift but reduces the estimated precision.
Our results are of potential applications in quantum frequency estimation~\cite{Ivanov2013,Malossi2013,Tian2015,Ivanov2015} and quantum control~\cite{Greilich2009,Poem2011,Bason2012,Sauer2013,Daems2013} of driven TLS's under spontaneous emission.
\acknowledgments
This work is supported by the National Natural Science Foundation of China (Grant No. 11374375, 11574405, 11465008).
Jiahao Huang is partially supported by National Postdoctoral Program for Innovative Talents of China (Grant No. BX201600198).
|
\section{ Introduction}
The nature of dark matter (DM) remains one of the most enduring mysteries in physics.
One of the theoretically best motivated DM candidates is the axion.
This low mass pseudoscalar boson is required by the Peccei--Quinn (PQ) mechanism
which provides a possible resolution to the strong CP problem,
the absence of CP violation in quantum chromodynamics (QCD)~\cite{Peccei:2006as,
Kim:2008hd, Agashe:2014kda}.
The value of the axion mass $m_a$ for which axions provide all or a significant part of the relic DM density depends on the cosmological history.
In the scenario in which the PQ symmetry is broken after inflation~\cite{Sikivie:2009fv,Kawasaki:2013ae},
the observed DM density points to a favoured $m_a$ range of 50--200~$\mu$eV, with $m_a\simeq 100~\mu$eV as a preferred value.
While no experiment yet has been able to probe this parameter space~\cite{Hiramatsu:2012gg,Kawasaki:2014sqa},
a new experimental concept has recently been introduced to search for these `high-mass' axions:
the dielectric haloscope~\cite{TheMADMAXWorkingGroup:2016hpc}.
Axions interact with photons as described by the Lagrangian
${\mathcal L}_{a\gamma}=g_{a\gamma}{\bf E}\cdot {\bf B}a$,
with a dimensionful coupling $g_{a\gamma}\sim\alpha/(2\pi f_a)$,
the fine-structure constant $\alpha$,
the PQ scale $f_a$, and the electric and magnetic fields ${\bf E}$ and ${\bf B}$, respectively.
Accordingly, axions mix with photons in the presence of an external magnetic field, which we will denote as ${\bf B}_{\rm e}$.
Even for a strong ${\bf B}_{\rm e}$ field, the mixing is small and leaves two admixtures of axions and photons in the mass basis:
a mostly axion-like particle, which acquires a small electric and magnetic field,
and a mostly photon-like particle, which acquires a small axion component.
While the axion-induced electric and magnetic fields are extremely tiny,
the conversion of axions to photons can be enhanced by providing suitable geometries.
This has traditionally been attempted via microwave cavities,
which would allow the axion field to drive resonant modes of the cavity~\cite{Sikivie:1983ip}.
By using a cavity with a high quality factor ($Q$-factor) and a large volume $V$ inside a strong ${\bf B}_{\rm e}$ field,
one expects an enhanced conversion of axions to photons,
as exemplified by the axion DM search experiments ADMX and HAYSTAC \cite{Rybka:2014xca,Kenany:2016tta}.
However, since the dimensions of the cavity must be on the order of $\lambda/2=0.62~{\rm cm}\, (100~\mu{\rm eV}/m_a)$,
where $\lambda$ is the Compton wavelength of the axion, cavity haloscopes become increasingly ineffective towards high values of the axion mass. Recently there is increasing interest in building effective cavities operating at higher frequencies~\cite{Rybka:2014cya,Goryachev:2017wpw,McAllister:2017ern,McAllister:2017lkb}.
Rather than using a resonant cavity, a dielectric haloscope uses a series of dielectric disks with adjustable spacings
placed in a strong ${\bf B}_{\rm e}$ dipole field~\cite{TheMADMAXWorkingGroup:2016hpc}.
In such a setup, axions lead to the generation of electromagnetic (EM) power according to the following principle:
if one places an interface between two different dielectric media inside a magnetic field,
the axion field causes microwave radiation to be emitted from either side of the interface~\cite{Horns:2012jf}.
In essence, breaking translation invariance allows one to convert axion-like states to photon-like states.
One can enhance this effect to an observable level by using multiple interfaces,
using reflections and constructive interference to boost the signal~\cite{TheMADMAXWorkingGroup:2016hpc}. For an illustration, see figure~\ref{fig:haloscope}.
Such a dielectric haloscope could be used to search for DM axions in the well-motivated $m_a$ range of $40-400~\mu$eV for the first time.
The underlying theory of dielectric haloscopes has been explored in detailed studies,~\cite{Millar:2016cjp,Ioannisian:2017srr}, but in these works the velocity of the axion was assumed to be zero.
In the present study we now go beyond the zero-velocity limit and investigate the effects for a velocity of $v \sim 10^{-3}c$,
as expected for galactic cold DM axions on earth. These effects will come almost entirely from the velocity of the axion in the direction perpendicular to the disks, which provides a change in phase of the axion over the haloscope. %
We also discuss strategies to probe the galactic DM velocity distribution in the event of a discovery.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=10cm]{pics/haloscope.png}
\caption{Schematic of a dielectric haloscope. In the presence of an axion and a strong external magnetic field ${\bf B}_{\rm e}$, EM waves (blue lines) are emitted from each dielectric disk (gray bars), and then reflect and transmit through the device till they reach the receiver. The mirror (dark gray) ensures that the entire signal power is directed towards the receiver. For simplicity we have not depicted the reflections. A fraction of the axion de Broglie wavelength is also depicted: the effects of the changing phase of the axion throughout the device will be the primary concern of this paper. Figure adapted from \cite{TheMADMAXWorkingGroup:2016hpc} with permission.}
\label{fig:haloscope}
\end{center}
\end{figure}
This paper is organised as follows.
We first derive the mixing between axions and photons in matter
for arbitrary values of the axion velocity and general ${\bf B}_{\rm e}$ alignments in section~\ref{mixing}.
We then consider the radiation caused by a single dielectric interface inside a magnetic field in a 3D~setting in section~\ref{3D}.
In section \ref{multilayer} we generalise the transfer matrix formalism developed in the zero-velocity limit in reference~\cite{Millar:2016cjp}
to allow for a description of the more realistic case with a non-zero axion velocity.
Using this formalism, we study both analytically treatable special cases
and realistic examples of configurations of dielectric haloscopes in section~\ref{analytic}.
This includes for the first time a full and realistic 80~disk example of a dielectric haloscope,
as is being considered for the proposed MADMAX experiment~\cite{TheMADMAXWorkingGroup:2016hpc}.
From these examples we conclude that the velocity of the axion is only relevant when the haloscope is a large fraction ($\sim20$\%) of the axion's de Broglie wavelength. As such, it can be safely neglected in the MADMAX context. However, as discussed in section~\ref{vdist} if the axion were to be discovered, one could design a larger device capable of measuring the axion's velocity distribution with directional sensitivity.
We reach our conclusions in section~\ref{conclusion}.
\section{Axion-photon mixing in an external magnetic field}
\label{mixing}
Before exploring the effect of the axion's velocity in a complicated setup such as a dielectric haloscope, we must know how the axion mixes with the photon in an external magnetic field~${\bf B}_{\rm e}$. Here we consider the linearised axion-Maxwell equations in the picture of axion-photon mixing.
\subsection{Equations of motion}
The interaction between photons and axions is described by the Lagrangian density
\begin{equation}\label{lagrangian}
{\cal L} = -\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-J^\mu A_\mu+\frac{1}{2}\partial_\mu a \partial^\mu a
-\frac{1}{2}m_a^2a^2 -\frac{g_{a\gamma}}{4}F_{\mu\nu}\widetilde F^{\mu\nu}a,
\end{equation}
with $a$ the axion field with mass
$m_a$. The coupling strength between axions and photons is governed by $g_{a\gamma}$.
We use $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$ to denote the EM
field-strength tensor in terms of the vector potential
$A^\mu=(A_0,{\bf A})$ and $J^\mu=(\rho,{\bf J})$ is the electric 4-current
involving the charge density $\rho$ and the current density ${\bf J}$. The dual tensor is given by
\smash{$\widetilde F^{\mu\nu} = \frac{1}{2}\varepsilon^{\mu\nu\alpha\beta}F_{\alpha\beta}$}, where we use the Levi-Civita tensor with $\varepsilon^{0123}=\varepsilon_{123}=+1$. The electric and
magnetic fields are given by
\begin{equation}
{\bf E} = -{\bm \nabla} A_0-\dot {\bf A}
\qquad\hbox{and}\qquad
{\bf B} = {\bm \nabla}\times {\bf A}\,.
\end{equation}
Throughout this paper we will use natural units with $\hbar = c = 1$ and the Lorentz-Heaviside convention
$\alpha=e^2/4\pi$, giving
$1~{\rm V}/{\rm m}=6.5163\times 10^{-7}~{\rm eV}^2$ and
$1~{\rm T}=1~{\rm Tesla}=195.35~{\rm eV}^2$. The energy density of the EM field
is $\frac{1}{2}({\bf E}^2+{\bf B}^2)$.
While the mass and photon couplings are independent for a generic axion-like particle (ALP), we
are mostly interested in QCD axions. In this case, both parameters are largely decided by
the PQ scale, or axion decay constant, $f_a$ by
$m_af_a\sim m_\pi f_\pi$ and $g_{a\gamma}\sim\alpha/(2\pi f_a)$, where $m_\pi$ and
$f_\pi$ are the pion mass and decay constant, respectively. The most recent study
gives the numerical values \cite{diCortona:2015ldu}
\begin{subequations}
\begin{eqnarray}
m_a&=&5.70(6)(4)\, {\rm \mu eV}\,\left(\frac{10^{12}\rm\,GeV}{f_a}\right)\,,\label{eq:ma}\\
g_{a\gamma}&=&-\frac{\alpha}{2\pi f_a}\,C_{a\gamma}
=-2.04(3)\times10^{-16}~{\rm GeV}^{-1}\,\left(\frac{m_a}{1\,\mu{\rm eV}}\right)\,C_{a\gamma}\,,\label{eq:gag}\\
C_{a\gamma}&=&\frac{{\cal E}}{{\cal N}}-1.92(4)\label{eq:cag}\,,
\end{eqnarray}
\end{subequations}
where the numbers in brackets denote the uncertainty in the last digit,
with the first and second error given for $m_a$ referring to quark-mass uncertainties and higher order corrections, respectively. $C_{a\gamma}$ is a model dependent number of ${\cal O} (1)$, which is given by the EM and colour anomalies (${\cal E}$ and $\cal N$, respectively). In models where ordinary quarks and leptons do not carry PQ charges,
the axion-photon interaction arises
entirely from $a$-$\pi^0$-$\eta$ mixing and ${{\cal E}}/{{\cal N}}=0$,
the KSVZ model \cite{Kim:1979if, Shifman:1979if}
being the traditional example. In more general models,
${{\cal E}}/{{\cal N}}$ is a ratio of small integers, the DFSZ model \cite{Dine:1981rt, Zhitnitsky:1980tq}
with ${{\cal E}}/{{\cal N}}=8/3$ being
an often-cited example, although there exist many other cases \cite{Kim:2014rza}. Recent studies have explored the range of $C_{a\gamma}$, finding that for the QCD axion $0\lesssim C_{a\gamma}\lesssim 55$, though the extreme cases require judicious parameter choices~\cite{DiLuzio:2016sbl,DiLuzio:2017pfr}. ALPs can have any value of $C_{a\gamma}$ not ruled out by experiments or observations.
From \eqref{lagrangian} we can derive the axion-Maxwell equations in terms of electric and magnetic fields, \cite{Sikivie:1983ip, Wilczek:1987mv}
\begin{subequations}
\begin{eqnarray}
{\bm\nabla}\cdot {\bf E} &=& \rho - g_{a\gamma} {\bf B}\cdot{\bm\nabla} a\,,
\label{eq:Maxwell-a}\\
{\bm\nabla}\times {\bf B}- \dot{\bf E} &=& {\bf J}
+g_{a\gamma}\({\bf B}\, \dot a -{\bf E}\times{\bm\nabla} a\)\,,\label{eq:Maxwell-b}\\
{\bm\nabla}\cdot{\bf B}&=& 0\,,\\
{\bm\nabla}\times{\bf E}+\dot{\bf B}&=&0\,,\label{eq:Maxwell-d}\\
\ddot a-{\bm\nabla}^2 a +m_a^2 a &=&
g_{a\gamma} {\bf E}\cdot {\bf B}\,.\label{eq:Maxwell-c}
\end{eqnarray}
\end{subequations}
While the homogenous equations do not involve the axion field, they will be
used to set the EM boundary conditions for an
interface.
\subsection{Linearized axion-Maxwell equations}
\label{linearization}
Our situation of interest is one where a strong static external magnetic field
${\bf B}_{\rm e}$ has been set up by means of an external current ${\bf J}_{\rm e}$
and is used to mix axions (in our case DM axions) with EM waves. As shown explicitly, e.g., in Ref.~\cite{Millar:2016cjp},
such a setting can be described by linearised macroscopic axion-Maxwell equations, which have plane-wave solutions. This description in terms of linear particle mixing is possible due to the significant hierarchy of scales between the static ${\bf B}_{\rm e}$ field, the background axion field and the induced EM fields, i.e., ${\bf B}_{\rm e}$ is much larger than all other fields, and the background axion field is much larger than any induced EM fields.
For related linearisations, see also~\cite{Wilczek:1987mv,Das:2004qka,Das:2004ee,Ganguly:2008kh,VISINELLI:2013fia}.
Expanding the time-dependent fields in plane waves proportional to $\exp\[-i(\omega t - {\bf k}\cdot{\bf x})\]$ leads then
to the linearised macroscopic axion-electrodynamic equations in Fourier space~\cite{Millar:2016cjp}
\begin{subequations}
\begin{eqnarray}
\epsilon{\bf k}\cdot {\bf E} &=& -g_{a\gamma}{\bf k}\cdot{\bf B}_{\rm e}a\,,
\label{eq:Maxwell-aaa}\\
{\bf k}\times {\bf B}/\mu+\omega \epsilon{\bf E} &=&
-g_{a\gamma}\omega{\bf B}_{\rm e} a\,,\label{eq:Maxwell-bbb}\\
{\bf k}\cdot{\bf B}&=& 0\,,\\
{\bf k}\times{\bf E}-\omega{\bf B}&=&0\,,\\
\(\omega^2-{\bf k}^2-m_a^2\) a &=&
-g_{a\gamma} {\bf B}_{\rm e}\cdot{\bf E}\,,\label{eq:Maxwell-ccc}
\end{eqnarray}
\label{eq:Maxwell-macroscopic}
\end{subequations}
where ${\bf E}$, ${\bf B}$, $a$, $\epsilon$ and $\mu$ depend on $\omega$ and ${\bf k}$. In the macroscopic equations~\eqref{eq:Maxwell-macroscopic}, the media is assumed to be homogeneous and isotropic,
with a linear response that can be encoded
for the electric displacement field ${\bf D}=\epsilon{\bf E}$ in terms of the dielectric constant $\epsilon$
and for the magnetic field ${\bf H}={\bf B}/\mu$ in terms of the magnetic permeability $\mu$.
Notice that
in our natural units, $\epsilon$ and $\mu$ in vacuum are equal to unity,
i.e., they are what is usually called the {\em relative\/} dielectric
permittivity and {\em relative\/} magnetic permeability. In lossy media, $\epsilon$ and $\mu$ have imaginary parts.
We gain more insight by expressing ${\bf E}$ and ${\bf B}$ in terms of
their components along ${\bf k}$, which we call the $\ell$ (longitudinal)
component, and the transverse components. To describe the latter, we
use a coordinate system with one vector in the plane of ${\bf k}$ and
${\bf B}_{\rm e}$, which we call the \emph{b} direction, and one vector
transverse to both ${\bf k}$ and ${\bf B}_{\rm e}$, which we
call the \emph {t} direction, (see figure~\ref{axes}). In our practical applications,
typically ${\bf B}_{\rm e}$ will be very nearly perpendicular to
${\bf k}$ so that the ``\emph{b} direction'' then almost coincides with ${\bf B}_{\rm e}$.
We denote the angle between ${\bf B}_{\rm e}$ and ${\bf k}$ as $\varphi$ (as in figure~\ref{axes}) so that the external $B$-field in the
$b$ direction is $B_{\rm e}\sin\varphi$ and the one in the $\ell$ direction
is $B_{\rm e}\cos\varphi$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=7cm]{pics/axes-new.pdf}
\caption{Polarization basis for a field with wavevector ${\bf k}$,
defining the $\ell$ direction. Perpendicular to ${\bf k}$ and
${\bf B}_{\rm e}$ is the \emph{t} direction, whereas the \emph{b} direction is
perpendicular to ${\bf k}$ and spans the plane containing ${\bf B}_{\rm e}$.}
\label{axes}
\end{center}
\end{figure}
Equation~(\ref{eq:Maxwell-aaa}) applies to the $\ell$ component and simply
reads $k\epsilon E_{\rm \ell}=-g_{a\gamma}k B_{\rm e}a\cos{\varphi}$, whereas
the $\ell$ part of equation~(\ref{eq:Maxwell-bbb}) reads
$\omega\epsilon E_{\rm \ell}=-g_{a\gamma}\omega B_{\rm e}a\cos{\varphi}$. These
equations are equivalent except when $\omega=0$ or $k=0$. We only
consider solutions with $\omega\not=0$ because by assumption the only static field
is ${\bf B}_{\rm e}$.
On the other hand, we also consider homogeneous situations ($k=0$), so overall
we may drop equation~(\ref{eq:Maxwell-aaa}), which is trivial when $k=0$, and keep the $\ell$ part of
equation~(\ref{eq:Maxwell-bbb}).
It is convenient to return to the vector potential as a description of the
time-varying ${\bf E}$ and ${\bf B}$ fields. In Fourier space we have
${\bf B}=i\,{\bf k}\times{\bf A}$ and ${\bf E}=i(\omega{\bf A}-{\bf k}A_0)$.
In the longitudinal components of the equations,
only the combination $\omega A_{\rm \ell}-k A_0$ ever appears. Because in our case
$\omega\not=0$, it is consistent to set $A_0=0$ which amounts to using
the temporal gauge (also known as Weyl or Hamiltonian gauge). The
subtleties of this choice do not affect our case. Alternatively, because $\omega\not=0$,
we could define $A_{\rm \ell}$ simply as a notation for the
physical quantity $E_{\rm \ell}/i\omega$.
In terms of ${\bf A}$, the remaining linear equations of motion are
\begin{subequations}
\begin{eqnarray}
\(\epsilon\,\omega^2-{\bf k}^2/\mu\){\bf A}+({\bf k}\cdot{\bf A}){\bf k}/\mu
&=&ig_{a\gamma}\omega{\bf B}_{\rm e}\,a\,,\label{eq:Maxwell-b4}\\
\(\omega^2-{\bf k}^2-m_a^2\)\,a &=&
-ig_{a\gamma}\omega{\bf B}_{\rm e}\cdot{\bf A}\,.\label{eq:Maxwell-c4}
\end{eqnarray}
\end{subequations}
In terms of the different polarization components, our final equations
of motion in Fourier space are
\begin{equation}\label{mixingM}
\vvvv{\omega^2-k^2-m_a^2 &\Delta_{\rm \emph b}\omega & \Delta_{\rm \ell}\omega&0}
{\Delta_{\rm \emph b}\omega & \epsilon\omega^2-k^2/\mu & 0&0}
{\Delta_{\rm \ell}\omega & 0 & \epsilon\omega^2&0}{0 & 0 &0&\epsilon\omega^2-k^2/\mu}
\vvvv{a}{iA_{\rm \emph b}}{iA_{\rm \ell}}{iA_{\rm \emph t}}= 0\,,
\end{equation}
where the ``mixing energies'' are $\Delta_{\rm \emph b}=g_{a\gamma}
B_{\rm e}\sin{\varphi}$
and $\Delta_{\rm \ell}=g_{a\gamma} B_{\rm e}\cos{\varphi}$. We recall
that in general the response functions
$\epsilon$ and $\mu$ depend on $\omega$ and ${\bf k}$
and that in lossy media they have imaginary parts. The factor of $i$ in \eqref{mixingM} reflects a phase difference of $\pi/2$ between the components of $\bf A$ and $a$, which disappears when $\bf A$ is expressed in terms of $\bf E$.
\subsection{Eigenmodes of the interacting system}
The last line of the equations of motion (\ref{mixingM}) remains
unaffected by the axion-photon interaction and leads to the
dispersion relation $k^2=n^2\omega^2$, where we have introduced the usual
refractive index\footnote{We assume that the
imaginary part of $n$ is much smaller than its real part
and that the real part is positive. In this
case we can use relations of the type $x=\sqrt{x^2}$
with impunity because
we avoid the branch cut of the square-root function along the
negative real axis.}
\begin{equation}
n^2={\epsilon\mu}\,.
\end{equation}
Notice that in general $n$ can be a complicated function of $\omega$ and $k$,
although in our cases of practical interest, we ignore spatial dispersion
(no dependence on $k$) and temporal dispersion (the dependence on $\omega$), as we expect them to be weak. The same
dispersion relation applies, of course, to the \emph b polarization in the absence
of axion-photon mixing. In this case the $\ell$ dispersion relation is
$\epsilon\omega^2=0$ and, because we only consider $\omega\not=0$, would require $\epsilon({\omega,k})=0$ to support a propagating mode. A plasma can provide such a mode,
the longitudinal plasmon. A non-relativistic plasma has $\epsilon\simeq(1-\omega_p^2/\omega^2)$, where $\omega_p$ is the plasma frequency (i.e., the frequency of the longitudinal plasmon). While in principal axions can mix resonantly with longitudinal plasmons~\cite{Das:2004qka}, there are currently no proposed experimental efforts using such a mixing. The dielectric media we will consider later do not support such
dynamical $\ell$ modes. The axion dispersion relation in the absence of a magnetic field is $\omega^2-k^2=m_a^2$.
In the presence of axion-photon interactions, the eigenmodes of our system are
superpositions of the axion and photon field. Note that $A_{\rm t}$ seems to decouple in equation \eqref{mixingM} but when we consider a boundary between different dielectric media we will see that non-trivial alignments between the interface and the magnetic field can mix $A_{\rm t}$ with $A_{\rm b}$.
Nontrivial solutions of equation~(\ref{mixingM}) require the
determinant of the matrix to vanish. Because in general $\epsilon$
and $\mu$ depend on both $\omega$ and ${\bf k}$, the solution
depends on the material properties of the medium. Here we will neglect spatial dispersion, i.e., henceforth we assume that $\epsilon$ and $\mu$ depend
only on $\omega$. Moreover, we also assume isotropic media where
the response functions do not depend on the direction of ${\bf k}$.
We then find two solutions for the dispersion relation.
To lowest non-trivial order in the extremely small mixing energies they are
\begin{subequations}\label{eq:dispersionrelations}
\begin{eqnarray}
\hbox to 6em{Axion like: \hfil}~k^2&=& \omega^2-m_a^2- \frac{\Delta_b^2\omega^2\mu }{n^2\omega^2-\omega^2+m_a^2}
-\frac{\Delta_\ell^2\omega^2\mu}{n^2\omega^2}+{\cal O}(\Delta_{b,\ell}^4)\,\\
\hbox to 6em{Photon like:\hfil}~k^2&=&n^2\omega^2
+ \frac{\Delta_b^2\omega^2\mu }{n^2\omega^2-\omega^2+m_a^2}+{\cal O}(\Delta_{b,\ell}^4)\,.
\end{eqnarray}
\end{subequations}
A resonance can occur when $n<1$, i.e., when the ``squared effective photon mass''
$m_\gamma^2=\omega^2-k^2=(1-n^2)\omega^2$ is positive. However, as the mixing is typically exceedingly small, even tiny losses in the medium will limit the resonance enough so that higher order terms are not required. For example, for 100~$\mu$eV DM axions in a 10~T $B$-field $\Delta_{b,\ell}\sim 10^{-12}~\mu$eV, so ${\rm Im}( n)\lesssim 10^{-26}$ is required for $\Delta_{b,\ell}^4$ effects to become relevent in \eqref{eq:dispersionrelations}. Further, to get the resonance condition one would need $n\sim 10^{-3}$, so an incredibly special material would be needed to realise this. For higher order terms to be necessary, the numerator and denominator must be of similar size, which requires very large frequencies, low losses and large mixing energies. Such resonances can occur in the often-studied case of axion-photon oscillations \cite{Raffelt:1987im} at
high energies, where one typically assumes axions and photons
to be nearly degenerate, i.e., it is assumed that
\smash{$\Delta_b\omega\gg m_a^2$} and \smash{$\Delta_b\omega\gg |n^2-1|\omega^2$}.
Actually at the highest astrophysically relevant energies in the
10~TeV range, intergalactic magnetic fields provide this condition as the cosmic microwave background then produces
a non-negligible refractive index~\cite{Dobrynina:2014qba}.
Motivated by our case of an ordinary dielectric medium,
we have assumed a
photon refractive index of $n>1$, i.e., the photon dispersion relation is
space like, ignoring a possible small imaginary part.
In this case
the two branches of the
dispersion relation do not cross for any value of $\omega$; photons
and axions are never on resonance with each other, and so the
denominator in \eqref{eq:dispersionrelations} never becomes zero. Note that the velocities of the particles are given by the group velocity of the waves, $\partial\omega/\partial k$. For the axion this then gives us that ${\bf v}={\bf k}/\omega$
It is now straightforward to obtain the field configurations
corresponding to these propagating modes. To lowest order in the small
mixing energies we find
\begin{subequations}
\begin{eqnarray}
\hbox to 6em{Axion like: \hfil}~\vvv{a}{i A_{\emph b}}{iA_{\ell}} &=&
\vvv{1}{-\chit \mu \sin\varphi}{-\chil \mu \cos\varphi} +{\cal O}(\chi_{ b,\ell}^2),\\[1ex]
\hbox to 6em{Photon like:\hfil}~\vvv{a}{i A_{\emph b}}{iA_{\ell}}&=&
\vvv{\chit \sin\varphi}{1}{0} + {\cal O}(\chi_{ b,\ell}^2)\,,
\end{eqnarray}
\end{subequations}
where we have defined the small dimensionless mixing parameters
\begin{subequations}
\begin{eqnarray}
\chit &=& \frac{g_{a\gamma} B_{\rm e}\,\omega}{n^2\omega^2-\omega^2+m_a^2}\,, \\
\chil &=& \frac{g_{a\gamma} B_{\rm e}\,\omega}{n^2\omega^2}\,.
\end{eqnarray}
\end{subequations}
These vectors are normalized at linear order in $\chi_{b,\ell}$ (linear order
in $g_{a\gamma}$). On the other hand, the dispersion relations in
equation~(\ref{eq:dispersionrelations}) get modified only at order
$g_{a\gamma}^2$. Therefore, the dominant effect of the axion-photon interaction
is that the normal propagating photon and axion modes obtain a small admixture of the other flavor. Henceforth we will always work to linear order
in $g_{a\gamma}$ and thus ignore the modification of the dispersion relations.
Notice that these two solutions correspond to
different combinations of $\omega$ and $k$, i.e., those which obey
the respective dispersion relations. For a fixed choice
of $\omega$, these two solutions have different $k$ values.
Recall that in our
dielectric media, there is no longitudinal photon-like solution.
Regardless, through the mixing the axion-like mode carries a small electric field component
along ${\bf k}$.\footnote{Likewise, we could consider frequencies below the axion mass, $\omega< m_a$, where the axion has
no propagating mode, yet the photon-like mode exists and
still carries a small axion component.} This small longitudinal electric field is often neglected in the literature---while for most experimental concepts the $B$-field is transverse by design, there are exceptions such as the AMELIE helioscope proposal~\cite{Galan:2015msa}. As in this case the $B$-field is not aligned with the solar axions, it is possible that longitudinal axion induced $E$-fields could influence the experiment. We will show in section \ref{3D} that the axion induced longitudinal field can influence the generation of photon-like waves from an interface.
\subsection{Dark matter axions}
Our case of interest is very special in that the galactic DM axions
are highly non-relativistic ($v\lesssim 10^{-3}$). In view of the small axion mass and velocity, and concomitant large de Broglie wavelength,
\begin{equation}
\lambda_{\rm dB}
=\frac{2\pi}{m_a v}
=12.4~{\rm m}~\left(\frac{100~\mu{\rm eV}}{m_a}\right)\left(\frac{10^{-3}}{v}\right),
\label{eq:lambdadeBroglie}
\end{equation}
the axion field only exhibits a slow change of phase over the laboratory scale. As explained earlier, we ignore a frequency shift by
axion-photon mixing which is quadratic in the small coupling constant $g_{a\gamma}$.
The difference between the transverse and longitudinal mixing is negligible,
\begin{equation}
\chit\simeq\chil\equiv\chi=\frac{g_{a\gamma}B_{\rm e}}{n^2 m_a}
=3.98(5)\times 10^{-16}~\frac{C_{a\gamma}}{n^2}~\frac{B_{\rm e}}{10\,\rm T}\,.
\end{equation}
The axion-like wave is given by
\begin{equation}
a(t,{\bf x}) = a({\bf p}) e^{i({\bf p}\cdot {\bf x}-\omega t)} \quad ; \quad
i{\bf A}_a(t,{\bf x}) = - \chi \mu {\bf \hat B}_e\, a(t,{\bf x}) ,\label{dmwave}
\end{equation}
where ${\bf A}_a$ is the axion induced ${\bf A}$ field.
We see that the axion field develops a small electric field oscillating with the axion's frequency $\omega$,
\begin{equation}\label{eq:Ea}
{\bf E}_a(t,{\bf x})= -\dot {\bf A}_a(t,{\bf x})=-\chi\mu\omega\,a(t,{\bf x}){\bf \hat B}_{\rm e}=-\frac{g_{a\gamma}{\bf B}_{\rm e}}{\epsilon}\,a(t,{\bf x})\,.
\end{equation}
In addition, there is also a small induced $H$-field given by
\begin{equation}
\label{eq:Ha}
{\bf H}_a(t,{\bf x}) = \frac{1}{\mu}\nabla\times {\bf A}_a(t,{\bf x}) = - v \sin\varphi\, \chi \, \omega a(t,{\bf x}) {\bf \hat t}_a=-\frac{vg_{a\gamma}B_{\rm e}\sin\varphi}{\mu\epsilon}a(t,{\bf x}) {\bf \hat t}_a.
\end{equation}
Note that this is suppressed by the axion velocity; the axion induced magnetic field will only have a subdominant effect in dielectric haloscopes. However, there are proposals to use this field to detect DM axions by using an LC circuit~\cite{Sikivie:2013laa}.
The photon-like waves are very much like ordinary transverse EM waves (transverse up to $O(\chi^2)$ corrections) accompanied by a relativistic axion field proportional to the ${\bf \hat b}$ projection of the magnetic field, i.e., ${\bf \hat B}\cdot {\bf \hat b}=\sin\varphi$ (with $\varphi$ defined as in figure~\ref{axes})
\begin{equation}
{\bf A}(t,{\bf x}) = A_0{\bf \hat b} e^{i({\bf k}\cdot {\bf x}-\omega t)} \quad ; \quad
a_{\gamma}(t,{\bf x}) = i \chi {\bf \hat B}_e \cdot {\bf A}(t,{\bf x}) =
\frac{\chi}{\omega} {\bf \hat B}_e \cdot {\bf E}(t,{\bf x}). \label{photonlike}
\end{equation}
with $n^2\omega^2=|{\bf k}|^2$. As we will be converting axions to photons, all photon-like waves will be produced at linear order in $\chi$, so the axionic component is ${\cal O}(\chi^2)$ and thus negligible.
Apart from an overall phase, the local axion field is
\begin{equation}
a(t,{\bf x})= \int \frac{d^3p}{(2\pi)^3}a({\bf p})e^{-i[\omega t-{\bf p}\cdot {\bf x} +\alpha({\bf p})]},
\end{equation}
where $\alpha({\bf p})$ is a relative phase for each momenta. The axion field strength $a({\bf p})$ in general can give a non-trivial dependence on ${\bf p}$. The local axionic DM density in some large averaging volume $V$ is
\begin{equation}
\rho_a= \frac{1}{V}\int \frac {d^3p}{(2\pi)^3} \frac{\omega^2 |a({\bf p})|^2}{2}\simeq\frac{1}{V}\int \frac {d^3p}{(2\pi)^3} \frac{m_a^2 |a({\bf p})|^2}{2}=f_{\rm DM}\,\frac{300~{\rm MeV}}{{\rm cm}^3}\,, \label{eq:density}
\end{equation}
where $f_{\rm DM}$ is a fudge factor expressing the uncertainty in the local
dark-matter density\footnote{The local DM density has been estimated by various
authors using different data and assumptions \cite{Catena:2009mf, Strigari:2009zb, Weber:2009pt, Bovy:2012tw, Nesti:2013uwa, Bozorgnia:2013pua, Read:2014qva}. The value
$300~{\rm MeV}~{\rm cm}^{-3}$ is often used as a benchmark number, although
in the axion literature, $400~{\rm MeV}~{\rm cm}^{-3}$ is frequently used. The particle
data group \cite{Agashe:2014kda} gives a even larger value based on Ref.~\cite{Catena:2009mf}.} as well as the uncertainty
of the dark-matter fraction
consisting of axions relative to possible other forms of DM or
relative to the fraction gravitationally bound in axion mini clusters. Note that the factor of $1/2$ in \eqref{eq:density} comes from the cycle average of the axion field.
We find
\begin{equation}
E_{a}({\bf p})=-\frac{g_{a\gamma} B_{\rm e} a({\bf p})}{\epsilon}=1.3\times10^{-12}~{\rm V}/{\rm m}~\(\frac{B_{\rm e}}{10~{\rm T}}\)~
\frac{C_{a\gamma}f_{\rm DM}^{1/2}}{\epsilon}\frac{a({\bf p})}{a_0}\,
\end{equation}
where $a_0$ would be the field strength if axions were monochromatic with momentum ${\bf p}_0$, i.e., $a({\bf p})=a_0(2\pi)^3\delta^3({\bf p}-{\bf p}_0)$.
For the QCD axion, this result is independent of the axion mass.
Now that we have found the mixed axion and photon fields, including a non-zero velocity, we can consider a 3D axion field interacting with an interface. From this we will build up a transfer matrix formalism that includes velocity effects.
\section{Axion DM induced boundary EM radiation: 3D calculation with non-zero axion velocities}
\label{3D}
Now that we have discussed the general axion-photon mixing in matter, we can turn our attention to the effect of a change in media. Here we will consider a fully general 3D case, albeit with an interface of infinite extend, as depicted in figure~\ref{medium}.
We allow both the axion velocity and the $B$-field to have a non-trivial orientation with respect to the boundary. However, while this situation is technically more complex, as in \cite{Horns:2012jf} the solution still reduces to compensating discrepancies in the axion induced $E$ and $H$ fields with (mostly) regular EM waves.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=15.5cm]{pics/reflectio.png}
\caption{\small A non-relativistic axion DM wave of amplitude $a({\bf p})$ and momentum ${\bf p}$ crosses from medium~1 to medium~2 becoming the transmitted right moving wave of amplitude $R_a\simeq a({\bf p})$, with a negligible reflected left moving wave $L_a\simeq 0$. In general the change in media causes a discontinuity in the axions $E$ and $H$ fields that will be compensated by emission of EM waves $L_b,R_b$ (as well as the not pictured $L_t,R_t$). The momentum parallel to the boundary is conserved and the perpendicular momentum is fixed by the dispersion relations, leading the EM waves to be emitted almost perpendicular to the boundary. Different media are those with different axion-photon mixing energies, so any change in the $B$-field intensity, direction or dielectric constant induces radiation of EM waves from both sides of the interface (red and blue arrows). }
\label{medium}
\end{center}
\end{figure}
The EM and axion fields in medium 1 can be represented as a combination of the incoming axion-like wave, and reflected axion-like and photon-like waves
\begin{eqnarray}
\nonumber
\vvvv{a}{i A_{\emph b}}{iA_{\ell}}{iA_t}_1=&&a({\bf p})\vvvv{1}{-\chi_1\mu_1\sin\varphi}{-\chi_1\mu_1\cos\varphi}{0}e^{i({\bf p}\cdot {\bf x})}+
L_a\vvvv{1}{-\chi_1\mu_1\sin\varphi_{L_a}}{-\chi_1\mu_1\cos\varphi_{L_a}}{0}e^{i({\bf p}_L\cdot {\bf x})}\\
&& +
L_b\vvvv{\chi_1\sin\varphi_{L_b}}{1}{0}{0}e^{i({\bf k}_{L}\cdot {\bf x})}+L_t\vvvv{0}{0}{0}{1}e^{i({\bf k}_L\cdot {\bf x})}
\end{eqnarray}
while in medium 2 the transmitted axion-like and created photon-like waves are
\begin{eqnarray}
\vvvv{a}{i A_{\emph b}}{iA_{\ell}}{iA_t}_2=&&R_a\vvvv{1}{-\chi_2\mu_2\sin\varphi_{R_a}}{-\chi_2\mu_2\cos\varphi_{R_a}}{0}e^{i({\bf p}_R\cdot {\bf x})}+
R_b\vvvv{\chi_2\sin\varphi_{R_b}}{1}{0}{0}e^{i({\bf k}_R\cdot {\bf x})}\nonumber \\
&&+R_t\vvvv{0}{0}{0}{1}e^{i({\bf k}_R\cdot {\bf x})}.
\end{eqnarray}
We have chosen to write $L_b,~R_b ~(L_t,~R_t)$ to emphasize that we are referring to photon-like waves polarised in the $b~(t)$ direction. Every wave has in principle a different orientation with respect to the external magnetic fields and thus a different $\varphi$ angle (as defined in section~\ref{linearization}). While in this language it seems like $A_t$ plays no role, this is only because we are not in the natural basis of the system, which would be $(x,y,z)$ coordinates.
The boundary conditions on the EM waves are given by
\begin{subequations}
\begin{eqnarray}
{\bf E}_{||,1}&=&{\bf E}_{||,2},\\
{\bf H}_{||,1}&=&{\bf H}_{||,2},
\end{eqnarray}\label{eq:boundary}
\end{subequations}
so that without any additional symmetries there are four constraints, requiring four propagating waves to satisfy. The axion field is conserved across the boundary, to good approximation. This will mean that in general both polarisations (which we take to be $A_b, A_t$) will be required in each medium. Choosing $x=x_0$ as our boundary, we first impose
\begin{equation}
\label{BC1}
\vvv{a}{iA_{b,||}}{iA_{t,||}}_1(t,x_0,y,z) = \vvv{a}{iA_{b,||}}{iA_{t,||}}_2(t,x_0,y,z) ,
\end{equation}
where $A_{b,||},A_{t,||}$ are the projections of the $A_b,A_t$ components in the parallel plane at $x=x_0$. These equations can be only satisfied for all points $(y,z)$ in the boundary if all the phases vary with the same wavenumber along the boundary, i.e.,
\begin{equation}
{\bf p}_{||}={\bf p}_{||,L}={\bf p}_{||,R}={\bf k}_{||,L}={\bf k}_{||,R} .
\end{equation}
We can then divide out a common $e^{i {\bf p}_{||}\cdot {\bf x}}$ factor from the boundary condition \eqref{BC1}.
This implies that the angle of the reflected and transmitted photon-like waves with respect to the normal to the surface is suppressed by the small axion DM velocity,
\begin{equation}
\sin \kappa = \frac{|{\bf k}_{||}|}{|{\bf k}|} =\frac{|{\bf p}_{||}|}{n \omega} = \frac{|{\bf v}_{||}|}{n} ,
\end{equation}
i.e., photon-like waves are emitted \emph{perpendicularly} to the surface up to ${\cal O}(v/n) \sim 10^{-3}$ corrections. A kin of Snell's law relates the angles of the transmitted and reflected photon-like waves $n_1 \sin \kappa_1=n_2 \sin \kappa_2$.
For the axion-like waves, the transmission and reflection angles are the same because their dispersion relation is the same in both sides of the boundary, up to $\chi^2$ corrections.
Finally, note that the conservation of momentum and of energy for every wave defines the magnitude of the perpendicular momenta through the dispersion relations. This is particularly important for the photon-like waves, for which $|{\bf k}|^2 = n^2 \omega^2$. This imparts a large perpendicular momentum
\begin{equation}
|{\bf k}_{\perp}| = \sqrt{n^2\omega^2-|{\bf k}_{||}|^2} = \sqrt{n^2\omega^2-|{\bf p}_{||}|^2} = n\omega+\frac{|{\bf p}_{||}|^2}{2 n \omega}+...
=n\omega[1+{\cal O}(v^2)] .
\end{equation}
In order to continue we make the assumption that the $B$-field \emph{direction} is constant across the boundary. This is not necessary but serves for all the cases we want to cover and simplifies our derivations. We still allow for changes in the magnetic field strength across the boundary. We use ${\bf \hat B}_{\text e}=(\cos\beta,\sin\beta,0)$ with $\sin\beta>0$ in the ($x,y,z$) basis. That is, ${\bf \hat B}_{\text e}$ is in the $x,y$ plane with $z$ being the transverse direction. The $B$-field aligns with the boundary when $\beta=\pi/2$. Note that this assumption of a constant, arbitrary $B$-field direction requires that the $\mu$ of each media is the same. This is also desired for practical reasons, as moving magnetic materials inside strong $B$-fields is practically very difficult. Because of this we set $\mu_r=1$ from here on.
We also make the stronger assumption that ${\bf B}_{\rm e}$ is homogenous. Inhomogeneities in the $B$-field also would act as a source of momentum for the generated photons~\cite{Redondo:2010js}. We will leave such considerations for future work looking in more depth at this kind of three dimensional effects. In general we will have six boundary conditions, which will require six fields to satisfy them. It is because of this that we must include the transversely polarised photons $A_t$. We can use a change of basis to write the {\bf A} fields associated to the different waves in $(x,y,z)$ coordinates:
\begin{eqnarray}
\label{Afields}
i{\bf A}_{a} &=& -\chi_1 a({\bf p}) (\cos \beta,\sin\beta,0) \\ \nonumber
i{\bf A}_{L_a} &=& -\chi_1 L_a (\cos \beta,\sin\beta,0) \\ \nonumber
i{\bf A}_{R_a} &=& -\chi_2 R_a (\cos \beta,\sin\beta,0) \\ \nonumber
i{\bf A}_{L_b} &=& \frac{ L_b}{\sin\beta+v_y\cos\beta/n_1}\[(0,\sin\beta,0) +\frac{1}{n_1}(v_y\sin\beta,v_y\cos\beta,v_z\cos\beta)+...\] \\ \nonumber
i{\bf A}_{R_b} &=&\frac{R_b }{\sin\beta-v_y\cos\beta/n_2}\[(0,\sin\beta,0) -\frac{1}{n_2}(v_y\sin\beta,v_y\cos\beta,v_z\cos\beta)+...\] \\ \nonumber
i{\bf A}_{L_t} &=& \frac{ L_t}{\sin\beta+v_y\cos\beta/n_1}\[(0,0,-\sin\beta) -\frac{1}{n_1}(v_z\sin\beta,-v_z\cos\beta,v_y\cos\beta)+...\] \\ \nonumber
i{\bf A}_{R_t} &=&\frac{R_t }{\sin\beta-v_y\cos\beta/n_2}\[(0,0,\sin\beta) -\frac{1}{n_2}(v_z\sin\beta,-v_z\cos\beta,v_y\cos\beta)+...\] . \nonumber
\end{eqnarray}
Note that $A_{\ell}$ is a component of ${\bf A}_a$; mismatches in the longitudinal field must also be compensated. In writing the last four equations we have used that
\begin{subequations}
\begin{eqnarray}
{\bf \hat b}_\gamma &=&{\bf \hat t}_\gamma {\mathbf \times \hat \ell}_\gamma,\\
{\bf \hat t}_\gamma &=&-\frac{1}{\sin\varphi_\gamma} {\bf\hat B}_{\text e} \times {\bf \hat \ell}_\gamma,
\end{eqnarray}
\end{subequations}
with ${\bf \hat \ell}_\gamma=((-1)^{j},v_y/n_j,v_z/n_j)$ for each media $j$. Both photon polarisations have the same~${\bf k}$. Note $\bf\hat b$ is approximately the same for right and left moving waves, whereas from our definition the ${\bf \hat t}$ are almost equal and opposite. To simplify matters, we can express the $\varphi$ angles (that enter into the axion-photon mixing) as
\begin{subequations}
\begin{eqnarray}
\sin\varphi_{R_a} &=& \sin\varphi = ...\\
\cos\varphi_{R_a} &=& \cos\varphi = ...\\
\sin\varphi_{L_b} &=&\sin\varphi_{L_t}= \sin\beta+ \cos\beta \frac{v_y}{n_1} + O(v^2)\equiv \sin\varphi_L \\ \nonumber
\sin\varphi_{R_b} &=& \sin\varphi_{R_t}= \sin\beta- \cos\beta \frac{v_y}{n_2} + O(v^2)\equiv \sin\varphi_R \\
\end{eqnarray}
\end{subequations}
We just need to solve for the values of $L_a,R_a,L_b,R_b,L_t,R_t$ as a function of $a({\bf p})$.
Explicitly, the boundary conditions \eqref{eq:boundary} are
\begin{eqnarray}
a(x_0)_1&=&a(x_0)_2, \\ \nonumber
\partial_x a(x_0)_1 &=& \partial_xa(x_0)_2, \\ \nonumber
{\bf A}_{y}(x_0)_1 &=&{\bf A}_{y}(x_0)_2, \\ \nonumber
(\partial_x {\bf A}_{y}-\partial_y {\bf A}_{x})(x_0)_1 &=& (\partial_x {\bf A}_{y}-\partial_y {\bf A}_{x})(x_0)_2, \\ \nonumber
{\bf A}_{z}(x_0)_1 &=&{\bf A}_{z}(x_0)_2, \\ \nonumber
(\partial_x {\bf A}_{z}-\partial_z {\bf A}_{x})(x_0)_1 &=&(\partial_x {\bf A}_{z}-\partial_z {\bf A}_{x})(x_0)_2,
\end{eqnarray}
which translate into the following equations at first order in $v$
\begin{equation}
\small
\kern -2em\left( \begin{array}{ccccccc}
1 & 1 & -1 & \chi_1\sin\varphi_{L_b} & -\chi_2\sin\varphi_{R_b}&0&0\\
v_x & -v_x & -v_x & -n_1\chi_1\sin\varphi_{L_b} & -n_2\chi_2\sin\varphi_{R_b} &0&0\\
-\chi_1\sin\beta & -\chi_1 \sin\beta& \chi_2\sin\beta & 1
& -1 & \frac{v_z}{n_1}\cot\beta & -\frac{v_z}{n_2} \cot\beta \\
\chi_1g_{-} & \chi_1g_{+} & -\chi_2g_{-} & -n_1
& -n_2 & -v_z\cot\beta & -v_z\cot\beta\\
0 & 0 & 0 & \frac{v_z}{n_1}\cot\beta & \frac{v_z}{n_2} \cot\beta &-1&-1\\
v_z\chi_1\cos\beta & v_z\chi_1\cos\beta & -v_z\chi_2\cos\beta & -v_z\cot\beta & v_z\cot\beta &n_1 &-n_2\\
\end{array} \right)
\left( \begin{array}{ccccc}
a({\bf p})
\\ L_a
\\ R_a
\\ L_b
\\ R_b
\\
L_t
\\ R_t
\end{array} \right) = 0,
\end{equation}
where we have taken $x_0=0$ and used ${\bf v}={\bf p}/\omega$. We have defined $g_{\pm}=v_y\cos\beta\pm v_x\sin\beta$.
It is interesting to see that even neglecting the axion components there is an interaction term between the polarisations $b,t$ due to the non-trivial geometry. When these cross terms are non-zero this indicates that both $A_b$ and $A_t$ are required for a consistent solution to the axion-Maxwell equations. One can think of this as coming from the mismatch between the $b,t$ directions and the $y,z$ directions. The full solution is quite involved, but at first order in the tiny quantities $\chi_1, \, \chi_2$ and in the DM velocity $v$ it simplifies to:
\begin{eqnarray}
\label{3Dsolution}
L_a &=& 0, \\ \nonumber
R_a &=& a({\bf p}),\\ \nonumber
L_b &=& a({\bf p}) \(\chi_1-\chi_2\)\sin\beta\frac{n_2}{n_1+n_2}-a({\bf p})\(\chi_1-\chi_2\) \frac{v_x\sin\beta -v_y\cos\beta }{n_1+n_2},
\\ \nonumber
R_b &=& - a({\bf p}) \(\chi_1-\chi_2\)\sin\beta \frac{n_1}{n_1+n_2}-a({\bf p})\(\chi_1-\chi_2\)\frac{v_x\sin\beta -v_y\cos\beta}{n_1+n_2}
\\ \nonumber
L_t &=& L_b\frac{v_z}{n_1}\cot\beta -a({\bf p})\(\chi_1-\chi_2\)\frac{v_z\cos\beta}{n_1+n_2}
\\ \nonumber
R_t &=& R_b\frac{v_z}{n_2}\cot\beta+ a({\bf p})\(\chi_1-\chi_2\)\frac{v_z\cos\beta}{n_1+n_2} ,
\end{eqnarray}
which agrees with the results of \cite{Millar:2016cjp} when $v\to 0$ and $\sin\beta=1$.
While these solutions are not as elegant as in the $v=0$ case, the interpretations are just as straightforward. The first term in $L_b$ and $R_b$ is due to the mismatch in the axion-like $E$-field, whereas the second terms in $L_b$, $R_b$, as well as both terms in $L_t$ and $R_t$ come from a mismatch in the axion-like $H$-field. This can be seen more explicitly by rewriting the last four equations in terms of $E$ and $H$:
\begin{eqnarray}
\label{3DsolutionEfields2}
E_L^b &=& \hphantom{+} ({\bf E}_2^a-{\bf E}_1^a)_{y}\frac{n_2}{n_1+n_2}- ({\bf H}_2^a-{\bf H}_1^a)_z\frac{1}{(n_1+n_2)},
\\ \nonumber
E_R^b &=& - ({\bf E}_2^a-{\bf E}_1^a)_{y} \frac{n_1}{n_1+n_2}-({\bf H}_2^a-{\bf H}_1^a)_z\frac{1}{(n_1+n_2)}
\\ \nonumber
E_L^t &=& \hphantom{+} ({\bf H}_2^a-{\bf H}_1^a)_y\frac{n_2-n_1}{n_1(n_1+n_2)}
\\ \nonumber
E_R^t &=& \hphantom{+} ({\bf H}_2^a-{\bf H}_1^a)_y\frac{n_2-n_1}{n_2(n_1+n_2)} ,
\end{eqnarray}
where $E_{L,R}^{b,t}$ is the $E$-field traveling in the left and right direction polarised in the $b,t$ direction, respectively.
As ${\bf B}_{\rm e}$ is defined to lie in the $x,y$ plane there is no mismatch in ${\bf E}_{a,{\bf z}}$ so the axion field does not directly require $A_t\neq0$ as $B_{{\rm e},z}=0$. However, the cross terms between the two photon polarisations $b$ and $t$ do require $A_t$ to be present as well. The symmetry between the two is broken as the axions only interact with $A_b$ (and so the corrections to $A_b$ come at $v^2$ order).
Maximum emission of the $b$ polarisation waves happens when the $B$-field is exactly along the boundary direction $\bf \hat y$ ($\sin\beta =1$), as noted in \cite{Horns:2012jf}. The amplitude of the waves polarised along $A_t$, i.e., orthogonal to the external $B$-field, are suppressed by $v_z\sim O(10^{-3})$. While the longitudinal modes $A_{\ell}$ are not dynamical, the mismatch in the axion field induced $A_{\ell,||}^a$ does contribute to the production of photon-like waves. We now have the solution to the 3D boundary case.
\section{Multilayer analysis: dielectric haloscopes}
\label{multilayer}
Given the new dielectric haloscope concept~\cite{TheMADMAXWorkingGroup:2016hpc}, we are interested not only in the production of photons from a single interface, but rather in the cumulative effects of many dielectric layers (as depicted in figure~\ref{fig:haloscope}).
In the presence of several media with numerous interfaces we can arrange them to radiate coherently, enhancing the signal---a dielectric haloscope. With this one could detect axion DM~\cite{TheMADMAXWorkingGroup:2016hpc}. In this section we extend the transfer matrix formalism introduced in \cite{Millar:2016cjp} to include a non-zero axion velocity.
\subsection{Transfer matrix formalism}
We consider the idealised case of a plane parallel sequence of dielectric regions labelled $0,1... m$, where 0 and m are semi-infinite regions for the output/input, see figure~\ref{scheme}. The thickness of each layer is $d_1,d_2,...,d_{m-1}$ and we allow the indices of refraction to be different ($n_0,...,n_m$) so that the optical thicknesses are $\delta_1=\omega n_1d_1,...$. We consider $\mu=1$ throughout the system, i.e., we assume non-magnetic media.
\begin{figure}[t]
\begin{center}
\includegraphics[width=12cm]{pics/multi.png}
\caption{Several dielectric regions. The amplitudes $L_r$ and $R_r$ denote
the electric-field amplitudes of left and right moving EM waves in each homogeneous region $r$. The regions end at $x_{r+1}$, covering a distance $d_r$. The angle of photon emission has been exaggerated for visual effect.}
\label{scheme}
\end{center}
\end{figure}
We consider the external $B$-field aligned with the boundaries, as this corresponds to maximal photon emission, as noted in \cite{Horns:2012jf} and can be seen from taking $\sin\beta=1$ in \eqref{3DsolutionEfields2}.
Consider an axion DM wave travelling through such a set of dielectric regions. We have learned in the full 3D study of one interface that the axion wave gets transmitted completely through the surfaces, up to higher order effects, and that the momentum of each wave in the direction parallel to the surfaces is conserved in each crossing. As we need to solve for the propagation of the waves generated at each interface via axions, we will need to generalise the transfer matrix formalism used in~\cite{Millar:2016cjp}. The solution of the axion-Maxwell equations is a combination of two plane waves inside each dielectric region together with the axion wave. We have for each region $r$ at first order in the DM velocity ${\bf v}={\bf p}/\omega$ and in $\chi_r$
\begin{equation}
\vvvv{a}{iA_{x}}{iA_{y}}{i A_z} =
a({\bf p})\vvvv{1}{0}{-\chi_r }{0} e^{i v_x \omega x}+
R_r \vvvv{\chi_r}{-v_y}{1}{0} e^{i n_r \omega (x-x_r)}+
L_r \vvvv{\chi_r}{v_y}{1}{0} e^{-i n_r \omega (x-x_r)},
\end{equation}
where we have divided out a common factor of $e^{i(p_y y+p_z z -\omega t)}=e^{i\omega(v_y y+v_z z -t)}$, which does not affect the derived quantities. However, these terms will offset the angle of emitted EM waves.
The continuity of the axion and $\bf E_{||}$-field and its derivatives across the boundary $r,r+1$ imposes in principle four restrictions onto the amplitude of the right and left moving waves. At first order in $\chi\equiv {\rm Max}(\chi_r)$ only the continuity of the $E_{y}$-component and its derivative is relevant as $E_z$ vanishes, giving
\begin{eqnarray}
- a({\bf p})e^{iv_x \omega x_{r+1}} \chi_r +R_r e^{i \delta_r} + L_r e^{-i \delta_r} &=& - a({\bf p})e^{iv_x \omega x_{r+1}} \chi_{r+1}+R_{r+1} + L_{r+1}, \\
- a({\bf p})e^{iv_x \omega x_{r+1}} v_x \chi_{r}+n_rR_re^{i \delta_r} - n_r L_re^{-i \delta_r} &=& - a({\bf p})e^{iv_x \omega x_{r+1}} v_x \chi_{r+1}+n_{r+1} R_{r+1} -n_{r+1} L_{r+1} . \nonumber
\end{eqnarray}
Recalling that $\delta_r=\omega n_r(x_{r+1}-x_r)$, which can be complex if the medium is conducting or absorbing, the system solves to
\begin{eqnarray}
\vv{R_{r+1}}{L_{r+1}}&=&
\frac{1}{2n_{r+1}}\(\begin{array}{cc}
n_{r+1}+n_r & n_{r+1}-n_r \\
n_{r+1}-n_r & n_{r+1}+n_r \end{array}\)
\vv{e^{i \delta_r}R_r}{e^{-i \delta_r}L_r}\nonumber \\
&+& a({\bf p})e^{iv_x \omega x_{r+1}}\frac{\chi_{r+1}-\chi_r}{2} \vv{1+v_x/n_{r+1}}{1-v_x/n_{r+1}} ,
\end{eqnarray}
which we can write as
\begin{equation}
\vv{R_{r+1}}{L_{r+1}}=
{\sf G}_{r} {\sf P}_r \vv{R_r}{L_r} + a({\bf p})\chi {\sf S}_{r} \vv{1}{1},
\end{equation}
where
\begin{subequations}
\begin{eqnarray}
{\sf G}_r &=& \frac{1}{2n_{r+1}}\(\begin{array}{cc}
n_{r+1}+n_r & n_{r+1}-n_r \\
n_{r+1}-n_r & n_{r+1}+n_r \end{array}\) ,
\\
{\sf P}_r &=& \(\begin{array}{cc}
e^{i \delta_r} & 0 \\
0 & e^{-i \delta_r} \end{array}\) ,
\\
{\sf S}_r &=& e^{iv_x \omega x_{r+1}}\frac{\chi_{r+1}-\chi_r}{2\chi}
\(\begin{array}{cc}
1+v_x/n_{r+1} & 0 \\
0 & 1-v_x/n_{r+1} \end{array}\)\label{Sss} .
\end{eqnarray}
\end{subequations}
These equations have the same form as those derived in \cite{Millar:2016cjp}, but there are three key differences. The first and most important is that the phase of the axion changes between each interface, given by an phase in ${\sf S}_r$. The second is that the discontinuity in the axion induced $H$-field alters the matrix in ${\sf S}_r$ from the zero velocity limit, which simply has an identity matrix. This will generally give an overall linear shift of $10^{-3}$. Lastly, the velocity also influences the axion frequency $\omega=m_a(1+v^2/2)$. This primarily affects the propagation of the photon, slightly modifying ${\sf P}_r$. Unlike the other two changes, the transverse velocities of the axion also enter ${\sf P}_r$, rather than just $v_x$.
For multiple layers we iterate this equation to derive a relation between the amplitude of the $R,L$ waves in the external media, 0 and $m$:
\begin{eqnarray}
\vv{R}{L}_m
&=& {\sf G}_{m-1} {\sf P}_{m-1} \vv{R}{L}_{m-1} + a({\bf p})\chi {\sf S}_{m-1}\vv{1}{1} \nonumber \\
&=& {\sf G}_{m-1} {\sf P}_{m-1} \[{\sf G}_{m-2} {\sf P}_{m-2} \vv{R}{L}_{m-2} + a({\bf p})\chi {\sf S}_{m-2}\vv{1}{1}\]+ a({\bf p})\chi{\sf S}_{m-1}\vv{1}{1} \nonumber \\
&=& ... \nonumber \\
&=& {\sf T} \vv{R}{L}_{0} + a({\bf p})\chi {\sf M} \vv{1}{1}\label{transfereq}
\end{eqnarray}
where ${\sf T}={\sf T}_0^m$, with ${\sf T}^a_b$ the transfer matrix from surface $b$ to $a$
\begin{equation}
{\sf T}^a_b = {\sf G}_{a-1}{\sf P}_{a-1}{\sf G}_{a-2}{\sf P}_{a-2}{\sf G}_{a-3}{\sf P}_{a-3} \, ...\, {\sf G}_{b}{\sf P}_{b} .
\end{equation}
In this notation, ${\sf T}^a_a={\sf P}_0=\mathbb{1}$. Similarly, we have defined
\begin{equation}
{\sf M}\equiv\sum_{s=1}^m {\sf T}^{m}_s {\sf S}_{s-1}.\label{eq:mdef}
\end{equation}
As in \cite{Millar:2016cjp}, we have separated \eqref{transfereq} into two parts. The first part encodes the normal propagation of EM waves through dielectric media. The second gives the axion-DM induced source terms ${\sf S}_r$, one for each layer. Because each wave has to traverse different layers to reach the detector they will all come with different phases. These are the phases we shall arrange to be coherent to increase the axion DM signal. Interestingly, the main axion velocity effects only enter in ${\sf S}_r$, and then only by $v_x$. Essentially, the dielectric haloscope picks out the part of the axion's velocity aligned with the haloscope. Because of this, in situations where the velocity is non-negligible a dielectric haloscope should have a directional sensitivity to the axion.
\subsubsection*{Transmissivity and Reflectivity }
We can use ${\sf T}$ to define the usual transmission and reflection amplitude coefficients
\begin{subequations}
\label{eq:T-R-Coefficients}
\begin{eqnarray}
{\cal T}_L &=& \frac{R_m}{R_0}\Big{|}_{L_m=0} = \frac{{\rm Det} [{\sf T}]}{{\sf T}[2,2]}\,, \\[1ex]
{\cal T}_R &=& \frac{L_0}{L_m}\Big{|}_{R_0=0} = \frac{1}{{\sf T}[2,2]}\,, \\[1ex]
{\cal R}_L &=& \frac{L_0}{R_0}\Big{|}_{L_m=0} = -\frac{{\sf T}[2,1]}{{\sf T}[2,2]}\,, \\[1ex]
{\cal R}_R &=& \frac{R_m}{L_m}\Big{|}_{R_0=0} = \frac{{\sf T}[1,2]}{{\sf T}[2,2]}\,,
\end{eqnarray}
\end{subequations}
with Det$[{\sf T}]={n_0}/{n_m}$. These expressions correspond to decoupling the axion by taking $\chi\to 0$, though for $\chi> 0$ the RHS of the above equations \eqref{eq:T-R-Coefficients} then define the transmissivity and reflectivity. These can be useful quantities, as they will generally be correlated with the production of photons from axions. Note that non-zero axion velocities can make this correlation less manifest, as the transmissivity and reflectivity are unaffected by the axion. To regain the connection, one must also know the axion velocity.
\subsubsection*{Boost amplitude}
For a dielectric haloscope we only want the EM waves generated in the presence of axions. In this case we exclude any incoming EM waves, $R_0=L_m=0$, and solve for the outgoing waves to define the usual boost amplitude for dielectric haloscopes
\begin{subequations}
\begin{eqnarray}
{\cal B}_L= \frac{L_0}{\chi a({\bf p})} &=& -\frac{{\sf M}[2,1]+{\sf M}[2,2]}{{\sf T}[2,2]}, \\
{\cal B}_R = \frac{R_m}{\chi a({\bf p})}&=& {\sf M}[1,1]+{\sf M}[1,2] -\frac{{\sf M}[2,1]+{\sf M}[2,2]}{{\sf T}[2,2]}{\sf T}[1,2].
\end{eqnarray}
\end{subequations}
In general, we will not need to worry about the distinction between ${\cal B}_L$ and ${\cal B}_R$, as we only consider situations that are either symmetric, so ${\cal B}_L={\cal B}_R\equiv {\cal B}$, or maximally asymmetric, with the medium on one side being a mirror. In this case only one of ${\cal B}_L$ and ${\cal B}_R$ is non-zero.
To measure the gain in the amplitude of the emitted $E$-field, we define the boost factor $\beta\equiv |{\cal B}|$. Note that usually the detector will be outside the magnetic field, so $E_a$ is not directly measured. Thus the power per unit area emitted in EM waves is
\begin{equation}
\frac{P({\bf p})}{A}=\frac{E({\bf p})^2}{2}=\chi^2 \omega^2\beta^2({\bf p})|a({\bf p})|^2.\label{eq:power}
\end{equation}
Note that any measurement implicitly happens over some given frequency range (or alternatively, integrating over the corresponding momentum distribution).
\section{Velocity effects in dielectric haloscopes}
\label{analytic}
We can now turn our attention to the effects on the boost factor of dielectric haloscopes caused by the non-zero velocity of the axion. Here we will neglect the finite transverse size of realistic dielectric disks, as well as any potential tilting effects coming from nonparallel disks, both of which may lead to losses. To see the general behaviour of our systems, we revisit the analytic examples studied in~\cite{Millar:2016cjp}. We will start with some general considerations based on the structure of ${\sf M}$ in section~\ref{generalM}. Then we will study a single dielectric disk to show the small shift due to the axion induced $H$-field in section~\ref{singledisk}. We will then see that resonance conditions do not lead to additional velocity effects by studying a simple cavity setup in section~\ref{cavity}. We will then see from a transparent setup containing many disks that in section~\ref{transparent} the primary effect comes from the linear extent of the dielectric haloscope in the $x$ direction, i.e., the number of disks, their spacings and to a lesser extent the thickness of the disks. Velocity effects start having a ${\cal O}(10\%)$ effect when the length of haloscope in the $x$ direction ($L$) is $15-20$\% of the de Broglie wavelength of the axion. When the de Broglie wavelength is similar to the length of the haloscope the axion velocity can give an ${\cal O}(1)$ effect. To show that this remains true for non-trivial configurations of the type that would be used by an experiment, we explore the effect of a non-zero velocity on a realistic 80 disk setup in section \ref{80disk}.
\subsection{General behaviour}
\label{generalM}
Before we begin studying explicit examples, it would be instructive to gain a general understanding of at what order and from which terms velocity effects begin to affect dielectric haloscopes. To do this, we must look closer at ${\sf M}$ (as defined in \eqref{eq:mdef}). We note that it is possible to rewrite any element of a transfer matrix ${\sf T}_a^b[i,j]$ in the form
\begin{equation}
{\sf T}_a^b[i,j]=\sum_{p} W^{i,j}_{p}e^{i\delta_{p}},
\end{equation}
where $W_p^{i,j}$ is a constant depending only on the $n_r$, and $\delta_{p}$ is a combination of the phases with differing signs, i.e., each $\delta_p$ is of the form $\pm \delta_a\pm...\pm\delta_b$ with all $2^{b-a-1}$ possible combinations.
From this we can write
\begin{eqnarray}
{\sf M}[i,j]=\sum_{s=1}^m e^{iv_x \omega x_{s}}\frac{\chi_{s}-\chi_{s-1}}{2\chi}
\(1\pm\frac{v_x}{n_{s}}\)\sum_{p}W^{i,j}_{p}e^{i\delta_p},
\end{eqnarray}
where the $\pm$ signs depends on the exact matrix element. The axion velocity enters in three places: the change in the phase of the axion, the discontinuity in the axion induced $H$-field and the shift to the axion frequency.
To isolate the velocity effects, we can look at
\begin{eqnarray}
v_x\frac{\partial {\sf M}[i,j]}{\partial v_x}&\supset & v_x\sum_{s=1}^mi\omega x_s\frac{\chi_{s}-\chi_{s-1}}{2\chi}\sum_{p} W^{i,j}_{p}e^{i\delta_p} \big{|}_{v=0}\nonumber \\
&\pm &v_x\sum_{s=1}^m\frac{1}{n_s}\frac{\chi_{s}-\chi_{s-1}}{2\chi}\sum_p W^{i,j}_pe^{i\delta_p}
\big{|}_{v=0}\, \nonumber \\
&+&
v_x^2\sum_{s=1}^m\frac{\chi_{s}-\chi_{s-1}}{2\chi}\sum_p i\delta_p W^{i,j}_{p}e^{i\delta_p} \big{|}_{v=0}\, ,\label{eq:genv}
\end{eqnarray}
where we have only kept the leading contribution from each of the three effects mentioned above. This assumes that the phase appearing in the exponentials is small, which breaks down for physically large haloscopes. The three effects can be understood separately:
\begin{enumerate}
\item{{\em Axion phase:} The term in the first line of \eqref{eq:genv} is the most important, and comes from the change of phase of the axion itself. As it comes with factors of $x_s$, this term is roughly proportional to $m_a L v_x$, where $L$ is the length of the haloscope in the $x$ direction, leading it to dominate over the other terms as $L$ increases. When $L$ is very large, this expansion breaks down and the change in ${\sf M}$ is significant, potentially even ${\cal O}(1)$. Note that this is only the leading contribution coming from the phase of the axion: we neglect cross terms that may be of the same or higher order as the other two contributions, such as terms proportional to $m_a^2 L^2 v_x^2$ and $m_a Lv_x^2$.}
\item{{\em $H$-field:} The term in the second line of \eqref{eq:genv} is due to the discontinuity in the axion induced $H$-field. Note that there is no enhancement by the length scales of the haloscope. Because of this, terms coming purely from the axion induced $H$-field cannot have a significant effect on ${\sf M}$ as $v$ is very small.}
\item{{\em Frequency shift:} The contribution in the last line of \eqref{eq:genv} comes from the frequency shift of the axion appearing in the photon propagator. We can see from the combination of the phase $\delta_p$ and the factor of $v_x^2$ that this contribution is generically proportional $m_a Lv^2$. As this is a frequency shift, it can only large effect when the boost factor varies rapidly as a function of frequency, a case which is not usually considered in the dielectric haloscope context. For the cases we will study below, $\beta(m_a)\simeq\beta(\omega)$ to very good approximation.
Note that the transverse velocities also result in a similar frequency shift. For simplicity, we will assume throughout this section that the disk is perfectly aligned with the axion velocity, so that $v\equiv v_x$. As the shift to the frequency makes certain analytic cases more complex, but is numerically irrelevant, for the analytic cases we will define distances in terms of $\omega$, rather than the axion mass~$m_a$. Note that even though the change to the boost factor is negligible, a detector with an energy resolution better than ${\cal O }(10~{\rm kHz})$ would be able to resolve this frequency shift for $100~\mu$eV axions. }
\end{enumerate}
Thus we expect that the primary influence on the boost factor will come from the change of phase of the axion. This can be contrasted with the velocity effects considered in dish antenna setups~\cite{Jaeckel:2013sqa,Jaeckel:2015kea}. Dish antennas have only a single interface, essentially giving $L=0$ so that the dominant effects come from the axion velocities parallel to the interface.
To see exactly at what level the velocity comes into play, we must consider some more concrete examples.
\subsection{Single dielectric disk}
\label{singledisk}
As they form the building blocks of the dielectric haloscope, it is instructive to consider a single dielectric disk, as depicted in figure \ref{fig:disk}. This disk can be categorised by a thickness $d$ and an index of refraction $n\equiv n_1>1$, with vacuum on either side ($n_0=n_2=1$). The change of phase experienced by a perpendicular EM wave inside this disk is thus $\delta\equiv \delta_1=n\omega d$. Using transfer matrices, equation~\eqref{transfereq}, we can then solve for
\begin{subequations}
\begin{eqnarray}
{\cal T}_{\rm D}&=&\frac{i\,2n}{i\,2n\cos\delta+(n^2+1)\,\sin\delta}\, ,\\
{\mathcal B}_{\rm D}&=&\frac{(n^2-1)\[ \(e^{2i\delta}-1\)n^2+n(v-1)\(1+e^{2i\delta}-e^{\frac{i\delta(n+v)}{n}}\)+v+ ve^{2i\delta}\]}{n^2[(n-1)^2e^{2i\delta}-(n+1)^2]},
\end{eqnarray}
\end{subequations}
where we have used the subscript D to indicate that these quantities are for a single disk. These expressions agree with the corresponding ones in \cite{Millar:2016cjp} when $v=0$. Note that ${\cal B}_{\rm D}$ depends on $v$, whereas ${\cal T}_{\rm D}$ does not. While a strong correlation between ${\cal B}_{\rm D}$ and ${\cal T}_{\rm D}$ was found in \cite{Millar:2016cjp}, if velocity effects are significant this correlation is less manifest. To gain information about the boost factor from the transmissivity, one must also know the axion velocity.
\begin{figure}[t]
\begin{center}
\includegraphics[width=6cm]{pics/scheme-single-disk.pdf}
\caption{Schematic of a single dielectric disk with dielectric constant $\epsilon$ and disk thickness $d$.}
\label{fig:disk}
\end{center}
\end{figure}
For a simple example, we can consider a transparent disk, which is characterised by $\delta=\pi$.
Expanding to linear order in $v$ gives
\begin{eqnarray}
{\cal B}_{\rm D}&=&\frac{1-v}{2}\(1-\frac{1}{n^2}\)\(1+e^{\frac{i\pi v}{n}}\)\nonumber \\
&\sim & \(1-\frac{1}{n^2}\)\(1-v+\frac{i\pi v}{2n}\).
\end{eqnarray}
In the second line of this equation, the term proportional to $-v$ can be understood as coming from the direct modification to ${\sf S}_r$, i.e., the change to $H_a$. The second velocity dependent term, proportional to $\frac{i\pi v}{2n}$, comes from the change in phase the axion experiences over the disk, as seen from the factor of $\pi/2$. However, as this second term is the only contribution to $\rm {Im}({\cal B}_{\rm D})$, it will be relegated to second order in $\beta_{\rm D}=|{\cal B}_{\rm D}|$, leaving only the term coming from the axion induced $H$-field
\begin{equation}
\beta_{\rm D}\sim(1-v)\(1-\frac{1}{n^2}\).
\end{equation}
As one would expect, for a single dielectric disk with a reasonably small thickness $d$ the velocity of the axion has only a tiny effect on $\beta_{\rm D}$.
As they form the building blocks of the dielectric haloscope, it is instructive to consider a single dielectric disk, as depicted in figure \ref{fig:disk}. This disk can be categorised by a thickness $d$ and an index of refraction $n\equiv n_1>1$, with vacuum on either side ($n_0=n_2=1$). The change of phase experienced by a perpendicular EM wave inside this disk is thus $\delta\equiv \delta_1=n\omega d$. Using transfer matrices, equation~\eqref{transfereq}, we can then solve for
\begin{subequations}
\begin{eqnarray}
{\cal T}_{\rm D}&=&\frac{i\,2n}{i\,2n\cos\delta+(n^2+1)\,\sin\delta}\, ,\\
{\mathcal B}_{\rm D}&=&\frac{(n^2-1)\[ \(e^{2i\delta}-1\)n^2+n(v-1)\(1+e^{2i\delta}-e^{\frac{i\delta(n+v)}{n}}\)+v+ ve^{2i\delta}\]}{n^2[(n-1)^2e^{2i\delta}-(n+1)^2]},
\end{eqnarray}
\end{subequations}
where we have used the subscript D to indicate that these quantities are for a single disk. These expressions agree with the corresponding ones in \cite{Millar:2016cjp} when $v=0$. Note that ${\cal B}_{\rm D}$ depends on $v$, whereas ${\cal T}_{\rm D}$ does not. While a strong correlation between ${\cal B}_{\rm D}$ and ${\cal T}_{\rm D}$ was found in \cite{Millar:2016cjp}, if velocity effects are significant this correlation is less manifest. To gain information about the boost factor from the transmissivity, one must also know the axion velocity.
\begin{figure}[t]
\begin{center}
\includegraphics[width=6cm]{pics/scheme-single-disk.pdf}
\caption{Schematic of a single dielectric disk with dielectric constant $\epsilon$ and disk thickness $d$.}
\label{fig:disk}
\end{center}
\end{figure}
For a simple example, we can consider a transparent disk, which is characterised by $\delta=\pi$.
Expanding to linear order in $v$ gives
\begin{eqnarray}
{\cal B}_{\rm D}&=&\frac{1-v}{2}\(1-\frac{1}{n^2}\)\(1+e^{\frac{i\pi v}{n}}\)\nonumber \\
&\sim & \(1-\frac{1}{n^2}\)\(1-v+\frac{i\pi v}{2n}\).
\end{eqnarray}
In the second line of this equation, the term proportional to $-v$ can be understood as coming from the direct modification to ${\sf S}_r$, i.e., the change to $H_a$. The second velocity dependent term, proportional to $\frac{i\pi v}{2n}$, comes from the change in phase the axion experiences over the disk, as seen from the factor of $\pi/2$. However, as this second term is the only contribution to $\rm {Im}({\cal B}_{\rm D})$, it will be relegated to second order in $\beta_{\rm D}=|{\cal B}_{\rm D}|$, leaving only the term coming from the axion induced $H$-field
\begin{equation}
\beta_{\rm D}\sim(1-v)\(1-\frac{1}{n^2}\).
\end{equation}
As one would expect, for a single dielectric disk with a reasonably small thickness $d$ the velocity of the axion has only a tiny effect on $\beta_{\rm D}$.
\subsection{Cavity mode}
\label{cavity}
One might wonder whether velocity effects can be enhanced by multiple reflections forming a resonance. To check this, we can take the simple cavity setup from section 5.2 of \cite{Millar:2016cjp}, consisting of a single mirror with a dielectric disk of refractive index $n$ (shown in figure~\ref{fig:cavity}). The phase change in the vacuum gap $\delta_{\rm v}$ and the disk $\delta_\epsilon$ can be chosen so that a resonance occurs: $\delta_{\rm v}=(2m-1)\pi$, where $m\in{\mathds N}$, and $\delta_\epsilon=\pi/2$. Such a resonance can be handled analytically, giving
\begin{equation}
{\cal B}_{\rm C}=n\(e^{\frac{i\pi v}{2n}}+e^{\frac{i\pi v[ 1+2n(2m-1)]}{2n}}\)-\frac{ie^{\frac{i\pi v}{2n}}}{n}-v\(1-\frac{1}{n^2}\), \label{eq:bcav}
\end{equation}
where the C subscript stands for cavity.
There are two instructive limits of this equation: a strongly resonant setup and a physically large setup (i.e., large $m$).
Let us first check the strongly resonant limit of this equation. The resonance will give a sharp, high boost factor peak when the disk is as reflective as possible, i.e., when $n$ is large. Expanding \eqref{eq:bcav}, taking the lowest order terms in $mv$ and assuming that $n\gg 1$ we get
\begin{equation}
\beta_C=|{\cal B}_{\rm C}|\sim n\[2-\(\frac{v \pi [2m-1]}{2}\)^2\].
\end{equation}
We can see from this that enhancing the resonant conditions (by increasing $n$) does not change the proportional importance of velocity effects. Because of this we do not expect that resonant setups will be more prone to velocity effects than non-resonant setups, such as those discussed below. Note that here we have assumed that the resonance width is much larger than the axion line width, as is expected for dielectric haloscopes for practical reasons~\cite{Millar:2016cjp}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8cm]{pics/cavity.png}
\caption{Schematic of a simple cavity consisting of a mirror and a single dielectric disk. The distances $d_{\rm v}$ and $d_\epsilon$ denote the thicknesses of the vacuum gap and dielectric disk, respectively. Note that $R_0,L_0$ (not shown) are both zero inside the mirror.}
\label{fig:cavity}
\end{center}
\end{figure}
On the other hand, if the device is made large by increasing the spacing between the disk and the mirror, the velocity dependent terms can dominate $\beta_{\rm C}$. This can be seen immediately as the dependence on the size of the device (given by $m$) enters as the phase in a complex exponential in \eqref{eq:bcav}. Such a setup was proposed in~\cite{Irastorza:2012jq} to gain a directional sensitivity to the axion velocity. While in principle the axion velocity could enter at linear order, this does not occur due to a $\pi/2$ phase shift between the linear order $v$ terms and the $v$ independent terms in ${\cal B}_{\rm C}$.
If the phase change of the axion is an appreciable fraction of $2\pi$, i.e., the device is a large fraction of the de Broglie wavelength, velocity effects will be significant. In figure~\ref{fig:cbeta} we plot the boost factor against $m$, with the number of half wavelengths in the device given by $2m-1$. From figure~\ref{fig:cbeta} we see that these effects reach the $10\%$ level at a scale that roughly corresponds to 15\% of the de Broglie wavelength~\eqref{eq:lambdadeBroglie}. Here we have assumed that $n=5$, however this result does not depend strongly on the exact choice of $n$.
\begin{figure}[b]
\begin{center}
\includegraphics[width=9.5cm]{pics/Cavitybeta.pdf}
\caption{$\beta_{\rm C}$ as a function of $m$ for a resonant cavity consisting of a mirror and a single dielectric disk of optical thickness $\pi/2$ and refractive index $n=5$, with $2m-1$ half wavelengths between the mirror and reflective disk. Here we have used $v=10^{-3}$; if $v=0$ then $\beta_{\rm C}$ would be constant as we assume lossless materials. }
\label{fig:cbeta}
\end{center}
\end{figure}
Of course, this is not a particularly practical limit: one would never actually choose to place the disk hundreds of wavelengths away from the mirror. The losses caused by diffraction, and the narrowing of the boost factor bandwidth that accompanies higher order modes would make this a very inefficient experiment. However, one could ameliorate the losses due to diffraction by enclosing the experiment in a metallic cavity.
\subsection{Transparent mode}
\label{transparent}
Let us consider next a simple multidisk setup, that consists of a series of transparent disks, each with a refractive index $n$. For some frequency $\omega_0$, we can specify that each disk has an optical thickness $\delta_\epsilon=n\omega_0 d_\epsilon=\pi$, each equally spaced with a vacuum gap of $\delta_{\rm v}=\omega_0 d_{\rm v}=\pi$. A three disk example is shown in figure~\ref{fig:transparent}.
As we have presented the building blocks of the transparent mode in section \ref{singledisk}, we can see how one puts them together. We must be more careful than in the zero velocity limit explored in~\cite{Millar:2016cjp}: the boost factor for $N$ disks is no longer simply $N$ times the boost factor of a single disk. The result still is given by a sum of boost factors but we must also account for the change in the phase of the axion at each interface. Note that both the right and the left side of each interface must be summed over, giving two sums
\begin{eqnarray}
{\cal B}_{\rm T}&=&\(1-v\)\(1-\frac{1}{n^2}\)\[\sum_{j=1}^{N}e^{\frac{i\pi v j}{n}+(j-1)i\pi v}\nonumber -\sum_{j=1}^{N-1}e^{\frac{i\pi v j}{n}+i\pi v j}\] \\
&=&N\frac{1-v}{2}\(1-\frac{1}{n^2}\)\frac{\(1+e^{\frac{i\pi v}{n}}\)\(e^{\frac{i(1+n)N\pi v}{n}}-1\)}{e^{\frac{i(1+n)\pi v}{n}}-1}.
\end{eqnarray}
\begin{figure}[t]
\begin{center}
\includegraphics[width=13cm]{pics/transparent.png}
\caption{Schematic of a simple dielectric haloscope consisting of three dielectric disks, each with refractive index $n=\sqrt {\epsilon}$. The distances $d_{\rm v}$ and $d_\epsilon$ denote the thicknesses of the vacuum gaps and dielectric disks, respectively. }
\label{fig:transparent}
\end{center}
\end{figure}
We have used the subscript $\rm T$ to denote that this boost amplitude is for the transparent mode. Similar to the cavity case, as velocity effects come in from a phase enhanced now by a factor of $N$, for large numbers of disks, corresponding to a physically large device, they will govern ${\cal B}_{\rm T}$. As $N$ can be arbitrarily high in principle, to see the full behaviour we must in principal keep all orders of $Nv$. However, a series expansion can help us see at what $N$ velocity effects become important. The most relevant terms at large $N$ are
\begin{equation}
\beta_{\rm T}/N\sim 1-\frac{N^2\pi^2v^2}{12}\, .\label{eq:betaT2}
\end{equation}
From this it is clear that when $N$ is large the velocity dependent terms will dominate, entering at quadratic order. Again, there are no linear order terms in $\beta_{\rm T}$ due to a relative phase between the relevant linear order velocity term and the velocity independent terms.
To see when velocity effects become important we plot $\beta_{\rm T}$ in figure \ref{fig:tbeta}, using dielectric disks with $n=5$. While the exact point this occurs at depends weakly on $n$, for $v=10^{-3}$ velocity effects reach the $10\%$ level for $N\gtrsim 400$. This roughly corresponds to a linear distance $L=(N-1)\lambda/2+N\lambda/2n$ of $20\%$ of the axion de Broglie wavelength.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=9.5cm]{pics/Transparentbeta.pdf}
\caption{Boost factor $\beta_{\rm T}(N)$ for a setup consisting of $N$ transparent (optical thickness $\pi$) dielectric disks of refractive index $n=5$ equally spaced with an optical thickness $\pi$ between each disk. The blue curve corresponds to $v=10^{-3}$, the red to $v=0$.}
\label{fig:tbeta}
\end{center}
\end{figure}
Note that the transparent mode is an example of a broadband frequency response: in general one is also interested in the behaviour of frequencies adjacent to $\omega_0$.
By looking at the adjacent frequencies, as depicted in figure~\ref{fig:Tdisks}, we see that some frequencies are more prone to velocity effects than others. For 80 disks, as considered in \cite{TheMADMAXWorkingGroup:2016hpc,Millar:2016cjp}, we see that velocity effects are only on the percent level. Going up to 200 disks we start to see an effect, with 400 disks showing ${\cal O}(10\%)$ deviation from the $v=0$ case. At 800 disks there is an ${\cal O}(1)$ change to $\beta$. However, these changes are not uniform: for example for 800 disks there are frequencies that go from $\beta\sim 0$ to $\beta\sim 800$, and others that exhibit almost no change. The differing response to the axion velocity at different frequencies is due to the relative phase structure of the velocity independent and dependent terms. As there is still a high degree of symmetry to the system the velocity effects show a periodic variation of maximal constructive/destructive interference.
\begin{figure}[t]
\begin{center}
\includegraphics[width=7.6cm]{pics/T80.pdf}
\includegraphics[width=7.6cm]{pics/T200.pdf}
\includegraphics[width=7.6cm]{pics/T400.pdf}
\includegraphics[width=7.6cm]{pics/T800.pdf}
\caption{Boost factor $\beta(\omega/\omega_0)$ for 80, 200, 400 and 800 disks (top left, top right, bottom left and bottom right panels, respectively). Note that $\omega=\omega_0$ corresponds to the transparent mode, in which all disks are transparent. In each case, the dashed blue curve corresponds to $v=10^{-3}$, the solid red curve to $v=0$. }
\label{fig:Tdisks}
\end{center}
\end{figure}
While it appears that this effect is coming from adding disks, this is misleading. One can see the same effects by simply increasing the spacing between the disks, as in section~\ref{cavity}. As shown in figure~\ref{fig:ddisks} the same fractional deviation from the zero velocity limit occurs for 100~disks with $3\pi$ phase shifts as for 300 disks with $\pi$ phase shift (though the latter will achieve a higher boost factor as it uses more disks). This is because the exponentiated factors of $N$ come from the change of phase of the axion over the haloscope. Note that increasing the spacing of the disks also results in a narrowing of the boost factor peak, though also allows new structures to exist (i.e., at $\omega=\omega_0/3$). So again we see that the primary influence comes simply from the size of the de Broglie wavelength of the axion when compared to the physical size of the dielectric haloscope.
\begin{figure}[t]
\begin{center}
\includegraphics[width=9.5cm]{pics/tlambda2.pdf}
\includegraphics[width=9.5cm]{pics/tlambda32.pdf}
\caption{Boost factor $\beta(\omega)$ for 300 disks with $\delta_0\equiv\delta_{\rm v}\equiv\delta_\epsilon=\pi$ for $\omega=\omega_0$ (top) and 100 disks with $\delta_0=3\pi$ (bottom). Note that $\omega=\omega_0$ corresponds to the transparent mode, in which all disks are transparent. In each case, the blue curve corresponds to $v=10^{-3}$, the red to $v=0$. }
\label{fig:ddisks}
\end{center}
\end{figure}
\subsection{80 disk dielectric haloscope}
\label{80disk}
While the analytically tractable cases discussed above are illustrative, it is also important to check that realistic setups behave similarly. For an experiment it is desirable to have both
broad, rectangular (top-hat) responses and narrow resonant responses \cite{TheMADMAXWorkingGroup:2016hpc}.
To see this, we must consider how an experiment would be performed. The boost factor, while potentially broadband, does not cover the full frequency range we would like to search. Thus we must scan across frequencies. To do so, one must arrange the disks to enhance $\beta$ over a given frequency range, measure, then adjust the positions of the disks to measure an adjacent frequency range and repeat untill the full range is covered. For each measurement taking a time $\Delta t$, Dicke's radiometer equation gives the signal to noise ratio
\begin{equation}
\frac{S}{N}=\frac{P}{T_{\rm sys}}\sqrt{\frac{\Delta t}{\Delta\nu_a}},
\end{equation}
where the $T_{\rm sys}$ is the system noise temperature and $\Delta\nu_a\sim 10^{-6}\nu$ is the axion line width. Recall that the power $P$ is given by \eqref{eq:power}. As the Area Law for dielectric haloscopes implies that $P\Delta\nu_\beta$ is roughly constant (where $\Delta\nu_\beta$ is the frequency width over which $\beta$ is enhanced), naively one expects that a narrow resonance with high $P$ is desirable~\cite{Millar:2016cjp}. However, for an optimal scan rate, the measurement time should be similar to the time taken to adjust the disks. Broadband setups can compensate for the potentially lengthy time required to readjust the disk positions between each measurement.
Such responses would allow one to scan a large frequency range in a single measurement; one could then use narrow resonances to confirm or reject a potential discovery at a higher signal to noise ratio. As in general the axion frequency is not simply related to the disk thickness (such as by a half wavelength as in the transparent mode above), the disk configurations must be found by a numerical optimisation procedure, as described in~\cite{Millar:2016cjp}. In the above examples we saw that the scale at which velocity effects become important depends almost solely on the length of device relative to the de Broglie wavelength~\eqref{eq:lambdadeBroglie}. In particular, we found that resonance effects are unimportant. Thus we expect that our general considerations hold for both broadband and resonant disk configurations.
As it took several hundred disks or higher harmonics for the transparent mode to exhibit significant distortion due to the non-zero velocity, we expect that the extrapolation to $80$ disks in~\cite{TheMADMAXWorkingGroup:2016hpc,Millar:2016cjp} should not be marred by the inclusion of 1D velocity effects.
To check this, we explicitly generate a broadband configuration using 80 dielectric disks, 1~mm thick with $n=5$, with a mirror on one side. Such a large number of disks has not been calculated in a realistic, optimised setup before, and allows us to check the extrapolations made in \cite{TheMADMAXWorkingGroup:2016hpc}. For consistency with~\cite{TheMADMAXWorkingGroup:2016hpc}, we will also use 25 GHz as our benchmark frequency, and optimise for a bandwidth of 50 MHz. The boost factor $\beta$, reflectivity $|{\cal R}|$ and group delay $\frac{\partial}{\partial \nu}\text {Arg}(\mathcal R)$ can be seen in figure~\ref{fig:80disks1}. This 80 disk setup gives a power boost ($\beta^2$) within 5\% of the prediction made in \cite{TheMADMAXWorkingGroup:2016hpc} using the Area Law (that is, that the area under $\beta^2(\nu)$ should increase linearly with the number of disks). Increasing the number of disks increases the number of boundary conditions that must be satisfied, leading to configurations containing a more ``rippling" structure than the 20 disk configuration of the same width shown in~\cite{TheMADMAXWorkingGroup:2016hpc}.
Similar substructure can also be seen if one looks at the reflectivity of the setup. The group delay maps out resonances and reveals a similar rippling structure to $\beta$. In addition, assuming that the dielectrics are lossy, with ${\rm Im}(n)=10^{-3}$, $|{\cal R}|$ maps out resonant structure within the haloscope. To show this structure clearly, in figure~\ref{fig:80disks1} the losses were exaggerated an order of magnitude when compared to a realistic dielectric such as LaAlO$_3$ at low temperatures
\begin{figure}[t]
\begin{center}
\includegraphics[width=10cm]{pics/80disks.pdf}
\caption{Boost factor $\beta(\nu)$ (red curve), reflectivity $|{\cal R}(\nu)| $ (black curve) and group delay $\frac{\partial}{\partial \nu}\text {Arg}(\mathcal R)$ (gray curve) for 80 disks optimised for a bandwidth of 50 MHz, centred on 25 GHz. The reflectivity has been plotted assuming lossy dielectrics, with $\tan\delta\sim\,$few$\times 10^{-4}$. More specifically, we assume ${\rm Im}(n)=10^{-3}$. }
\label{fig:80disks1}
\end{center}
\end{figure}
With such a disk configuration, we can directly see the effects of including a non-zero axion velocity. This is shown by plotting $\beta_{v=10^{-3}}-\beta_{v=0}$ as a function of frequency in the bottom panel of figure \ref{fig:80disks2}. The difference is entirely negligible, with less than a percent level change in the region of interest (as seen in the top panel of figure \ref{fig:80disks2}).
This is actually somewhat less than one would expect from studying the transparent mode. As the spacing between each disk is close to $\lambda/2$, 80 disks in transparent mode (as shown in the top left panel of figure \ref{fig:Tdisks}) should be comparable with our 80 disk broadband configuration.
However, the transparent setup is actually more prone to velocity effects than the realistic configurations. This is probably because of the large levels of symmetry present in the transparent mode; destructive and constructive effects occur coherently. That being said, the size of the distortion is similar enough to conclude that the scale at which velocity effects becomes important depends only weakly on the configuration as long as the total size of the haloscope remains constant. Thus dielectric haloscopes of the size considered in~\cite{TheMADMAXWorkingGroup:2016hpc,Millar:2016cjp} are free to neglect the 1D effects of a finite axion velocity without negative consequence.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9.5cm]{pics/80d.pdf}
\includegraphics[width=9.5cm]{pics/80ddiff.pdf}
\caption{{\it Top:} Boost factor $\beta(\nu)$ for 80 disks optimised for a bandwidth of 50 MHz, centred on 25 GHz. Both $\beta_{v=0}$ and $\beta_{v=10^{-3}}$ are plotted as red and dotted blue curves, respectively. {\it Bottom:} Difference between $\beta$ for $v=10^{-3}$ and $v=0$ as a function of frequency. Inside the region of interest (the shown 50 MHz bandwidth), this difference corresponds to a less than percent level effect. }
\label{fig:80disks2}
\end{center}
\end{figure}
Note that errors in the disk placement made when readjusting the disks between measurements will likely result in larger changes to $\beta$ than the axion velocity. While the exact sensitivity to mispositioning errors depends on the distribution function of these errors, from numerical simulation we can see that one will require ${\cal O}(\mu{\rm m})$ placement position to keep the change in $\beta$ at the 10\% level. One can see an example assuming a top hat error function with a width of $5~\mu$m in figure~\ref{fig:error}. This gives a deviation in $\beta$ of order $10-20\%$, which should be an acceptable deviation for an experiment. Further, one can see that the reflectivity $|{\cal R}|$ remains correlated with $\beta$, allowing one to gain information about the deviation of $\beta$, allowing for errors to be corrected. For example, the large deviation of the gray curve can be seen in both the reflectivity and boost factor. While the distortion from mispositioning errors is a greater effect than that of non-zero axion velocities, the time-dependence of changes to the axion velocity may allow it to be extracted from a signal.
\begin{figure}[t]
\begin{center}
\includegraphics[width=10cm]{pics/errorbeta.pdf}
\includegraphics[width=10cm]{pics/errorref.pdf}
\caption{Boost factor $\beta(\nu)$ (top) and reflectivity $|{\cal R}(\nu)|$ (bottom) for 80 disks optimised for a bandwidth of 25 MHz, centred on 25 GHz. Five configurations are shown, with each disk perturbed from the configuration shown in figure~\ref{fig:80disks1} by a top hat error function with width $5~\mu$m. For the reflectivity we assume losses exaggerated by an order of magnitude over the realistic value for visual comparison, with ${\rm Im}(n)=10^{-3}$. Losses are neglected for the boost factor.}
\label{fig:error}
\end{center}
\end{figure}
\section{Probing the axion velocity with directional sensitivity}
\label{vdist}
Finally, we turn our attention to the benefits and drawbacks of velocity sensitivity. Whether or not the axion velocity plays a role depends on the type of experiment under consideration. Broadly speaking, there are two types of experimental setups: a search experiment, in which one is trying to verify the existence of DM axions, and an ``axion telescope", which would try to extract the most astrophysical information possible out a discovered DM signal.
\subsection{Search experiment} When searching for axions, the velocity effects must be small, i.e., $\beta(v)$ should be approximately independent of $v$. To see why we must require this, recall that the generated power at given axion momentum given by \eqref{eq:power} is
\begin{equation}
P({\bf p})\propto |a({\bf p})|^2\beta^2({\bf p}). \label{eq:power2}
\end{equation}
Note $P({\bf p})$ is our only observable: as we do not know the velocity distribution of the axion $a({\bf p})$ a priori, one cannot disentangle $a({\bf p})$ from $\beta({\bf p})$, unless $\beta({\bf p})$ is independent of $v$. Consider if an experiment had a strongly ${\bf p}$ dependent $\beta$. When no signal is seen in a given measurement, we could not tell whether there were no axions present, or if there were DM axions with an unexpected velocity distribution giving a different boost factor than expected. In other words, the experiment does not distinguish between $a({\bf p})
\sim 0$ and $\beta({\bf p})\sim 0$, so no limits can be easily set. Thus, a design requirement for this type of setup is that $\beta(\vec p)= \beta(m_a)$.
As we saw in section~\ref{analytic}, velocity effects become relevant when the haloscope is on the order of $20\%$ of the axion de Broglie wavelength. This gives us a limit on the potential size of the device, corresponding to several hundred disks, as can be seen from considering the transparent mode in section \ref{transparent}. Ultimately, due to the Area Law found in~\cite{Millar:2016cjp}, this will provide a (very high) limit on the possible boost factor/bandwidth achievable by a search experiment type dielectric haloscope.
Despite the lack of direct directional sensitivity, there is still a wealth of information that can be extracted after detection. When an axion signal is found, measuring $P(\omega)$ gives immediately the axion power spectrum $|a(\omega)|^2$ from \eqref{eq:power2}, as $\beta$ would be roughly constant and known by design. Note that $P$ only has a direct directional sensitivity if $\beta$ does. However, we could easily extract the mass and an angle averaged velocity distribution,
\begin{equation}
F(\omega) = \langle |a(\vec p)|^2\rangle =
\frac{v \omega^4}{2V\rho_{\rm a}} \int \frac{d\Omega}{(2\pi)^3} |a(\vec p)|^2.
\end{equation}
With current receiver technology the number of channels in a search experiment can easily be $\gtrsim 10^4$; once a signal is detected these can be easily refocused to within the axion linewidth, allowing a very precise measurement of $F$. This alone contains significant astrophysical information: for example, if the axion DM forms a Bose-Einstein condensate as proposed in \cite{Sikivie:2009qn}, we would expect a Dirac delta power spectrum rather than say a Maxwell-Boltzmann spectrum, as is often assumed. One could also look for transient phenomena, though detecting substructure such as axion miniclusters~\cite{Hogan:1988mp} and associated tidal streams~\cite{Tinyakov:2015cgg} could take many years of measurement~\cite{OHare:2017yze}. If found these would give strong hints about the origin and structural formation of axion DM. Recall that we only measure the axion generated $E$-field (like any other haloscope experiment), which depends on both the coupling $g_{a\gamma}$ and the DM density $\rho_a$ given by \eqref{eq:density}. Because of this a haloscope experiment cannot determine the absolute local axion DM density and thus the normalisation of $a({\bf p})$. Breaking this degeneracy requires an independent measurement of $g_{a\gamma}$, an extraordinary challenge in this mass range.
Interestingly, $F(\omega)$ depends on time, due to the relative motion of the experiment with respect to the DM wind. With sufficiently long measurements one could detect annual and diurnal modulation, though the latter would be difficult~\cite{Turner:1990qx,OHare:2017yze}. The time dependence of $F(\omega)$ enters via very small shifts in the frequency of the axion, coming from small changes to the relative velocity of the DM with respect to the lab location on Earth. Annual modulation gives a relatively large ($5\%$) change to the velocity, but of course happens over yearly timescales~\cite{Turner:1990qx,Ling:2004aj}. By contrast diurnal modulation is relatively rapid, but so slight (on the order of 0.1\%~\cite{Turner:1990qx,Ling:2004aj}) that a very large amount of accumulated statistics would be needed to detect it.
These measurements will help to confirm the veracity of the signal and give information about the local structure of DM. This would give the rough direction of the DM wind, as well as potentially detecting some substructure in the halo.
\subsection{Axion telescope} Once we know the axion mass and $F(\omega)$, we could become much more ambitious and move towards an ``axion telescope". In this case, we desire the extra information contained in the axion velocity distribution, the full $a({\bf p})$. To extract this, we need an experiment in which the output power is sensitive to ${\bf p}$, i.e., we require $\beta$ to have a strong ${\bf p}$ dependence.
To do this one would build an experiment with a size comparable to the axion de Broglie wavelength: as this is highly sensitive to $v_x$, we essentially measure the fraction of DM with a velocity in the $x$ direction (or a range of velocities, depending on the details of the setup). This is similar to the type of elongated cavity experiments proposed in \cite{Irastorza:2012jq}, which would also be sensitive to the change in phase of the axion over the device.
Using knowledge of the axion mass and power spectrum obtained from the search experiment, the dielectric disks could be placed to provide a very large signal, allowing for shorter measurements. For example, we could measure for a few hours rather than a day or a week as is currently considered~\cite{MADMAXinterestGroup:2017bgn}. These shorter measurements would be sensitive to velocity changes on shorter time scales, giving a more rapid accumulation of statistics.\footnote{Note that one does not actually have to preform shorter measurements, rather we simply want that the statistics accumulate at a faster rate. One could do a single long measurement and break the data down into different times during analysis.}
While the rotation of the Earth only changes the magnitude of the axion velocity by 0.1\%, the changing direction gives a sizeable modulation of $v_x$ depending on the alignment of the $x$-direction of the experiment and the DM wind. This directional modulation of the axion velocity can be used to probe the axion velocity structure in a 2D plane.\footnote{This 2D plane is given by the circle traced out by the $x$ direction of the experiment as the Earth rotates. Note that this circle changes over the course of the year, allowing for limited 3D coverage.} As the device would now be the same size as the de Broglie wavelength, there would be an ${\cal O}(1)$ variation in the signal power, sensitive to much smaller variations in $v_x$ than the ${\cal O}(10^{-9})$ shift in the axion frequency would allow. This would give a much more detailed map of the local DM than is possible with directionally insensitive devices, hence the name ``axion telescope".
We could build further evidence for the astrophysical origin of the signal by comparing very precisely the observed directionality of the signal with theory (i.e., the expected DM wind). Unlike a search experiment, any features observed would come with directional information; a search experiment can only infer the direction of DM wind using knowledge of the motion of the Earth. For example, if we see features such as axion miniclusters or tidal streams~\cite{OHare:2017yze} we would know the direction that these objects came from.
When combined with simulations of galaxy formation this directional information could be a test of structure formation. To aid in this, it might be useful to also have a velocity-insensitive device operating next to the axion telescope, giving a simultaneous measurement of the overall axion power spectrum and the $x$-moving component of DM at that time. Alternatively, one could use multiple experiments pointed in three directions to have a complete 3D local map at any given point in time.
Directional sensitivity can be partially achieved even with the search experiment. If the axion were discovered at the high mass end of the scanned range, as the corresponding Compton wavelength $\lambda$ is small there would be a lot of unused magnet volume. As the spacings between the dielectrics would be roughly $\lambda/2$ during a regular search~\cite{Millar:2016cjp}, as $\lambda$ decreases the total distance taken up by the haloscope decreases. We could then increase the spacing between the disks to a higher harmonic configuration where the distance is $\sim 3\lambda/2, 5\lambda/2$, etc. instead of $\sim \lambda/2$. In figure \ref{fig:80disks3} we show the $v_x$ dependence for a configuration optimised for a 25 GHz axion signal, using spacings of order $5\lambda/2$. For illustrative purposes,
we assumed a Dirac delta velocity distribution, and have neglected the effects caused by the ${\cal O}(10^{-9})$ change in frequency, i.e., we show only the boost factor dependence. Note that the exact distribution does not significantly change this result: if a Maxwell-Boltzmann distribution is used instead the effect of the directional change is similar. While the exact velocity variation during the Earth's rotation depends on the location and alignment of the experiment, figure~\ref{fig:80disks3} gives the rough scale of the boost factor modulation. Maximal diurnal modulations can be achieved by aligning the $x$-direction of the experiment with the DM wind at some time of the day: this point gives the maximum $v_x$, with the minimum occurring 12 hours later.
The power modulation caused by the axion velocity is typically 10\%, significantly more pronounced than for the configuration shown in figures~\ref{fig:80disks1} and \ref{fig:80disks2}, and potentially detectable.
\begin{figure}[t]
\begin{center}
\includegraphics[width=9.5cm]{pics/dirurn.pdf}
\caption{Boost factor $\beta(v_x)$ for 80 disks of 1~mm thickness optimised for a 25 GHz axion (with a bandwidth of 1 MHz). The disk spacings have been optimised around $5\lambda /2$. This gives the rough scale of diurnal modulations due to the changing axion velocity. Here we have assumed that $a({\bf p})$ is a delta function. }
\label{fig:80disks3}
\end{center}
\end{figure}
For most cases the effects of the DM velocity are slight for an experiment of a similar size to the planned MADMAX~\cite{MADMAXinterestGroup:2017bgn}, but in the event that axions are discovered these effects contain a wealth of information about the DM in our universe, and potentially the galactic structure. As dielectric haloscopes are sensitive not only to the overall velocity distribution, but in principal have directional sensitivity, they would be a very powerful tool to study the DM in our galaxy.
\section{Conclusions and outlook}
\label{conclusion}
The axion is one of the best motivated DM candidates, potentially resolving both the DM question as well as the strong CP problem. The parameter space in which axions can form the DM of our Universe can be heavily restricted by theoretical considerations; in particular, high mass axions in the range $50-200~\mu$eV are predicted in scenarios where the PQ symmetry is broken after inflation. Dielectric haloscopes seem to be a promising avenue for searching for these axions, allowing one to use a detection device with a much larger volume than a traditional resonant cavity.
The axion's non-zero velocity can in principal have non-trivial effects in devices where the linear dimensions are comparable to the axion de Broglie wavelength. To study this in detail, we derived the axion-photon mixing equations with a non-zero velocity, and then looked at the photon production due to changing media. From this, we saw that the main influence of the velocity of the axion comes from the change of phase of the axion itself over the haloscope: effects from the transverse velocities and axion induced $H$-field were always subdominant. With a generalised transfer matrix formalism, we could then study the change to the boost factor in detail. We found that for haloscopes smaller than $\sim15 - 20$\% of the axion de Broglie wavelength, the axions velocity could be safely ignored.
To see this in realistic setting, we studied a realistic 80 disk setup for the first time. While in normal situations velocity effects were entirely negligible, if one increased the spacing between each disk by odd multiples of $\lambda/2$ it is possible to see a noticeable shift in the boost factor. This could be used in the event of a potential discovery of the axion, allowing one to observe effects such as diurnal modulation. This would be both a useful check on the veracity of the signal, as well as allow one to preform astrophysical measurements, such as testing for diurnal modulation with directional sensitivity.
\section*{Acknowledgments}
We thank the MADMAX Working Group at the Max Planck Institute for Physics
for support and helpful discussions, in particular Georg Raffelt. We would also like to thank Edoardo Vitagliano for helpful discussions and Jan Schuette-Engel for pointing out typos and a sign error in section 3. We acknowledge partial support by the
Deutsche Forschungsgemeinschaft through Grant No.\ EXC 153 (Excellence
Cluster ``Universe'') and Grant No.\ SFB 1258
(Collaborative Research Center ``Neutrinos, Dark Matter, Messengers''), as well as the European Union through the Initial
Training Network ``Elusives'' Grant No.\ H2020-MSCA-ITN-2015/674896 and Grant No.\
H2020-MSCA-RISE-2015/690575 (Research and Innovation Staff Exchange project ``Invisibles Plus''). J.R.\ is supported by the Ramon y Cajal Fellowship 2012-10597 and FPA2015-65745-P (MINECO/FEDER). Part of this work was performed at the Bethe Forum ``Axions'' (7--18 March 2016), Bethe Center for Theoretical Physics, University of Bonn, Germany.
|
\section{Introduction}
Ordinary differential equations (ODEs) are the primary means to modelling dynamical systems in a wide range of natural and engineering sciences. When the complexity of the considered system is high, the number of equations required limits our capability of performing effective analyses. This has motivated a large body of research, across many disciplines, into abstraction techniques that provide smaller ODE systems preserving the original dynamics in some appropriate sense (e.g.,~\cite{1098900,Iwasa1987287,okino1998,ccdtt-sci-reports-2012}).
In particular, we refer to a number of techniques whose basic idea is to build a smaller ODE system with less variables and/or parameters, while preserving most of the original dynamics.
Depending on the specific technique used, the obtained reduced model may be used, e.g., to provide a more compact description that abstracts from low-level details, or to improve the performance and scalability of required analyses.
The development of abstraction techniques is widely investigated in many disciplines in the science and engineering domain that deal with dynamical systems, for instance:
\begin{itemize}
\item
Ecological systems typically regard massive populations of entities that interact with the environment, and adapt to it. For this reason, quantitative abstractions have been widely studied in ecology (e.g.,~\cite{Iwasa1987287}). In this domain, abstraction techniques mainly exploit the fact that entities can range from molecules to whole organisms in their ecosystem (hence they act at different time scales), and the fact that there is often a high heterogeneity among the entities, e.g., they can have different age and metabolism, or can be located in different points in space.
In particular, many techniques focus on the aggregation of large-scale models of ecosystems based on ODEs or difference equations (the time-discrete homologous of ODEs).
Aggregations are obtained as a coarsening of the state space in terms of \emph{macro-variables}, each describing a group of original variables (i.e., of entities in the original ecosystem). Aggregations have been developed to ignore, e.g., age differences in population models~\cite{Iwasa1987287} or to exploit the fact that entities act at different time scales (see~\cite{Auger2008} for a survey).
\item
Cellular automata are a classic example of models used in physics to describe dynamical systems consisting of simple agents with small state spaces, whose interactions result in complex phenomena. Abstraction methods identify sets of neighbouring cells for which a coarse-grained description can still describe the overall behaviour of the original model~\cite{PhysRevE.73.026203}.
\item
In computational systems biology, many abstraction techniques have been developed to cope with the combinatorial species and chemical reaction networks (CRN) that arise from the interactions of basic molecules which can undergo internal changes, or that can bind each other to form complex species.
In particular, we refer to techniques developed for systems of ODEs arising from protein interaction networks (e.g.~\cite{Borisov2005951,Feret21042009}). Here,
methods have been proposed to consider: a covering of the state space (where a variable may appear in more than one group)~\cite{Feret21042009}; quotienting, induced by a partition of the ODE variables (e.g.,~\cite{Feret2012137}); and aggregations exploiting time-scale separation (e.g.,~\cite{Murray2002}).
\item
Large-scale dynamical systems are common also in control engineering. Here, starting from the seminal paper~\cite{1098900}, many approaches have been proposed to reduce the original model in a way that preserves controllability, i.e., the capability of driving the system to a desired state by using appropriate control inputs~\cite{Pappas}.
\end{itemize}
More recently, in the computer science community there has been an increasing interest towards quantitative models of computation based on ODEs, for example to use formal languages to describe biochemical models~\cite{1999133,Danos200469,Blinov22112004,Calzone15072006,CardelliTCS08,DBLP:conf/cmsb/KwiatkowskiS08,biopepa,Pedersen:2010aa}
or as a deterministic approximation for languages with stochastic semantics~\cite{biopepa,DBLP:journals/tcs/HaydenB10,TribastoneTSE2012}.
Our own line of research~\cite{tacas17cttv,gecco17,DBLP:conf/lics/CardelliTTV16,DBLP:conf/popl/CardelliTTV16,DBLP:conf/sfm/VandinT16,DBLP:conf/tacas/CardelliTTV16,DBLP:conf/concur/CardelliTTV15,DBLP:conf/mfcs/IacobelliTV15} consists of a computer-science perspective to the abstraction problem, borrowing ideas from the concurrency theory community.
We recast the ODE reduction problem to that of finding an appropriate equivalence relation over ODE variables, akin to classical models of computation based on labelled transition systems.
We studied such \emph{differential equivalences} for two basic intermediate languages, trading expressivity for efficiency:
\begin{enumerate}
\item[i)] \emph{IDOL} (Intermediate Drift-Oriented Language)~\cite{DBLP:conf/popl/CardelliTTV16} covers a general class of non-linear ODEs with derivatives containing polynomials, rationals, minima/maxima, and absolute values. This is enough, e.g., to capture the existing ODE semantics of stochastic process algebras~\cite{DBLP:conf/mfcs/IacobelliTV15,Tschaikowski2013b,concur2012,CardelliTCS08}.
The largest equivalences of IDOL terms are computed using a symbolic partition-refinement algorithm that exploits an encoding into a satisfiability modulo theories (SMT) problem;
\item[ii)]
\emph{Reaction networks}~\cite{DBLP:conf/concur/CardelliTTV15,DBLP:conf/tacas/CardelliTTV16}, a slight generalization of chemical reaction networks, characterise ODEs with polynomial derivatives.
In this case, the partition refinement is based on Paige and Tarjan's seminal proposal~\cite{partitionref}, giving an efficient algorithm that runs in polynomial time.
\end{enumerate}
Our framework for ODE reduction has been implemented in the tool ERODE~\cite{tacas17cttv} (\url{http://sysma.imtlucca.it/tools/erode/}), allowing us to provide evidence of effective reductions in realistic models from the literature.
This paper briefly reviews our framework for ODE reduction.
A more detailed tutorial-like presentation unifying the two approaches can be found in~\cite{DBLP:conf/sfm/VandinT16}, while~\cite{DBLP:conf/lics/CardelliTTV16,gecco17} address the more general problem of computing all differential equivalences of a model.
\section{Framework Overview}
\paragraph{Differential equivalences.}
Differential equivalences induce a quotienting of the ODE variables. Two distinct notions of equivalence for ODEs have been provided in~\cite{DBLP:conf/popl/CardelliTTV16}. \emph{Forward differential equivalence} (FDE) allows us to abstract from single variables, and to consider only cumulative information on each equivalence class. In particular, it guarantees that the original ODE system can be described in terms of a smaller ODE system having only one macro-variable per equivalence class that describes the sum of the variables within the equivalence class. For instance, consider the following ODE system with variables $x_1$, $x_2$, and $x_3$:
\begin{align}\label{eq:simple}
\dot{x}_1 & = - x_1, & \dot{x}_2 & = k_1 \cdot x_1 - x_2, & \dot{x}_3 & = k_2 \cdot x_1 - x_3,
\end{align}
where $k_1$ and $k_2$ are constants and the ‘dot' operator denotes the derivative operator (with respect to time).
It can be shown that the partition of variables ${H}=\{\{x_1\},\{x_2,x_3\}\}$ is an FDE because
\begin{align}\label{eq:simple.fde}
\dot{x}_1 & = - x_1, & \dot{(x_2 + x_3)} & = \dot{x}_2 + \dot{x}_3 = (k_1 + k_2) \cdot x_1 - (x_2 + x_3) .
\end{align}
In fact, by applying the simple change of variable $x_{2,3} = x_2 + x_3$, this is equivalent to writing
\begin{align*}
\dot{x}_1 & = - x_1 & \dot{x}_{2,3} & = (k_1 + k_2) \cdot x_1 - x_{2,3} .
\end{align*}
This reduced ODE system contains one variable per equivalence class in ${H}$, each of which describes the sum of the solutions of the variables in the corresponding equivalence class. Thus, setting the \emph{initial condition} $x_{2,3}(0) = x_2(0) + x_3(0)$ yields that the solution satisfies $x_{2,3}(t) = x_2(t) + x_3(t)$ at all time points $t$.
As discussed in~\cite{DBLP:conf/concur/CardelliTTV15}, the notion of FDE has proved to be particularly useful in the context of computational systems biology, where modellers are interested in a few observations of interest, like the evolution of the concentration of few species only, or of the sums of certain species that represent different configurations of the same molecule.
\emph{Backward differential equivalence} (BDE) instead allows us to identify redundant dynamics in the original system. In fact, it equates variables that have the same solutions at all time points, if initialized equally. In the ODE system in Equation~\eqref{eq:simple}, if $k_1=k_2$, we have that ${H}=\{\{x_1\},\{x_2,x_3\}\}$ is also a BDE. Hence, we can reduce the original ODE system by removing either equation between $x_2$ and $x_3$, say $x_3$, and by rewriting every remaining occurrence of $x_3$ as $x_2$:
\begin{align*}
\dot{x}_1 & = - x_1 & \dot{x}_2 & = k_1 x_1 - x_2 .
\end{align*}
The notion of BDE has proved to be useful in the context of evolutionary biology, where one is interested in understanding whether a system has evolved into another one preserving the original functionality~\cite{cardelli14,ccdtt-sci-reports-2012,10.1371/journal.pcbi.1005100}.
In fact, it is well known that using larger networks instead of smaller ones of \emph{equal functionality} is beneficial for enhanced stability with respect to stochastic noise~\cite{ccdtt-sci-reports-2012}.
In particular, in~\cite{DBLP:conf/popl/CardelliTTV16,DBLP:conf/lics/CardelliTTV16} we have shown that BDE succeeded in tracing a certain functionality across two CRNs (by taking the CRN whose species and reactions are the disjoint union of those from the source and target one).
Differential equivalences guarantee that the relationship between the original model and the abstract one is exact. However, FDE leads to a loss of information, because information on the individual variables within an equivalence class may not be recovered in general.
Instead, BDE preserves all information, however it can be applied only when the initial conditions are coherent with the considered equivalence classes: i.e., when all backward equivalent variables are initialized equally.
Forward differential equivalence is closely related to the notion of exact ODE lumpability, thoroughly investigated in the chemistry domain (e.g.,~\cite{LiRabitz1997,okino1998,Li19891413}). However no automated procedure exists to reduce an ODE system using this approach (e.g.,~\cite{Turanyi14}). To cope with this, i.e., to guarantee that our framework can be instantiated in a family of fully automatic reduction techniques, restrictions are imposed to be able to develop minimisation algorithms.
\paragraph{Symbolic minimisation.}
In~\cite{DBLP:conf/popl/CardelliTTV16} each ODE variable is treated explicitly as a real function and a differential equivalence is encoded in a logical formula over ODE variables. Thus, checking whether a candidate partition is BDE/FDE can be done symbolically using an encoding into satisfiability modulo theories (SMT)~\cite{Biere:2009:HSV:1550723}. In fact, differential equivalences belong to the quantifier-free fragment of first-order logic. It is possible to restrict the admissible ODE systems to those for which an SMT solver for nonlinear real arithmetic — e.g.~Z3, \cite{DeMoura:2008:ZES:1792734.1792766} — is a decision procedure. This can be done by, roughly speaking, excluding trigonometric functions. Hence, e.g., we support rational functions, necessary to encode sophisticated biological kinetics like Hill's one~\cite{Biere:2009:HSV:1550723}, and minima/maxima, necessary to capture the ODEs arising from terms of stochastic process algebras~\cite{DBLP:journals/tcs/HaydenB10,TribastoneTSE2012,cj:fluidflow}, or from queuing networks~\cite{10.1109/TSE.2012.66}.
Hence, the technique is generic enough to be applied to a wide range of dynamical systems.
Let us consider the example in Equation~\eqref{eq:simple}, assuming $k_1 = k_2 = 1$, for which case we have shown that ${H}=\{\{x_1\},\{x_2,x_3\}\}$ is a BDE. The condition for ${H}$ to be a BDE has been shown to correspond to requiring that related variables with equal assignments always have equal derivatives~\cite{DBLP:conf/popl/CardelliTTV16}. This can be encoded in a quantifier-free first-order logical formula, which in the specific case of ${H}$ is:
\[
\phi_{{H}} \ := \ (x_2 = x_3) \implies (k_1\cdot x_1 -x_2 =k_2\cdot x_1 - x_3)
\]
The SMT check $sat(\neg\phi_{H})$ searches for an assignment of the variables $x_1$, $x_2$, and $x_3$ such that $\neg\phi$ holds. Hence, the partition is a BDE if and only if the procedure returns ``unsat'', meaning that $\phi_{H}$ can not be falsified. This is the case for our example. More interestingly, in case the solver returns an assignment which satisfies $\neg\phi_{H}$, a \emph{witness}, then we can use it within an iterative algorithm which computes the coarsest BDE partition that refines the initial one (i.e., that can be obtained by splitting the blocks of the initial one).
Assume for example that we start from the candidate BDE partition ${H}_2=\{\{x_1,x_2,x_3\}\}$. It can be shown that ${H}_2$ is not a BDE, and hence the SMT check $sat(\neg\phi_{{H}_2})$ is satisfiable. For instance, a witness for the satisfiability of $\neg\phi_{{H}_2}$ is $(x_1 = 1, x_2 = 1, x_3 = 1)$, which yields to different evaluations for the three derivatives:
\[
\dot{x}_1 = -x_1= −1
,\quad
\dot{x}_2 = k_1\cdot x_1-x_2 =0
, \quad
\dot{x}_3 = k_2\cdot x_1 - x_3 = 0
\]
This suggests the implementation of a partition-refinement algorithm that \emph{splits} the blocks of the current partition in sub-blocks of variables which cannot be distinguished using the returned witness. In other words, at the next iteration the candidate BDE partition would be $\{\{x_1\},\{x_2,x_3\}\}$.
A similar algorithm has been provided also for the FDE case.
Note that this algorithm meets an important property often required for reduction techniques: the reduced model preserves user-defined observables of the original system.
Indeed, each variable of interest for the modeller can be put in a singleton initial block. In the backward case, this allows us to provide an initial partition coherent with the initial conditions of the original model (that is, two variables are in the same initial block if their initial conditions are the same), a necessary condition for the reduced model to faithfully represent the original dynamics.
%
\paragraph{Syntactic minimisation.}
More efficient partition refinement algorithms can be provided for ODEs with derivatives that contain only multivariate polynomials of degree at most two~\cite{DBLP:conf/tacas/CardelliTTV16}. This is quite a general class of ODEs, covering linear systems as well as chemical reaction networks. This approach is based on
a syntactic representation of an ODE system in terms of a so-called \emph{reaction network} (RN), which slightly generalizes the notion of chemical reaction network by allowing reaction rates to be also negative. Therefore, one can readily encode any polynomial ODE system as an RN~\cite{DBLP:conf/tacas/CardelliTTV16}. In particular, an RN consists of a set of species/variables interacting by means of reactions parameterised by a real value.
Intuitively, we use reaction networks to generalize labelled transition systems (LTS) so to consider transitions among multisets of nodes rather than among individual nodes. Hence, there exists a loose connection between species and nodes of an LTS, and between reactions and transitions of an LTS.
We provide two bisimulation equivalences for such syntactic representation: the forward RN bisimulation (FB), and the backward RN bisimulation (BB), which are related to FDE and BDE, respectively~\cite{DBLP:conf/concur/CardelliTTV15}. Continuing the parallel between reaction networks and LTSs, our bisimulations are similar in spirit to quantitative bisimulations on LTSs, e.g., Larsen and Skou's probabilistic bisimulation~\cite{Larsen19911}. In particular, the largest RN bisimulations that refine a given input partition can be computed by generalizing Paige and Tarjan's famous algorithm~\cite{partitionref}. In~\cite{DBLP:conf/tacas/CardelliTTV16} we propose a partition refinement algorithm along the lines of the best-performing analogues for Markov chain lumping such as~\cite{Derisavi2003309} and~\cite{DBLP:conf/tacas/ValmariF10}, and for probabilistic transition systems~\cite{DBLP:journals/jcss/BaierEM00}. The algorithm computes the largest FB/BB refining a given input partition of variables in $O(mn\log n)$ time, where $m$ is the number of monomials in the ODE system and $n$ is the number of variables.
\paragraph{Tool support: ERODE.}
We implemented our minimisation techniques in ERODE~\cite{tacas17cttv}, a mature tool featuring a modern integrated development environment for the evaluation and reduction of ordinary differential equations.
ERODE is on the Eclipse framework, it is multi-platform and does not require any installation process. The tool is available together with a manual and sample models at \url{http://sysma.imtlucca.it/tools/erode}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth]{architecturewithlogo.eps}
\caption{ERODE{}'s Architecture.}\label{fig:architecture}
\end{figure}
Fig.~\ref{fig:architecture} depicts
ERODE{}'s architecture.
It is organized in the GUI layer, and the core layer.
Fig.~\ref{fig:screenshot} depicts the main components of the GUI layer, including
a fully-featured text editor based on the \textsl{XTEXT} framework which supports syntax highlighting, content assist, error detection, and fix suggestions (top-middle of Fig.~\ref{fig:screenshot}).
Additionally, this layer offers a number of views, including a \emph{project explorer} to navigate among different ERODE{} files (top-left of Fig.~\ref{fig:screenshot});
an \emph{outline} to navigate the parts of the currently open ERODE{} file (bottom-left of Fig.~\ref{fig:screenshot});
a \emph{plot view} to display ODE solutions (top-right of Fig.~\ref{fig:screenshot}); and a \emph{console view} to display diagnostic information (bottom-right of Fig.~\ref{fig:screenshot}).
The core layer implements the minimization algorithms and related data structures for FDE, BDE, FB and BB.
The algorithms for FB and BB reductions have been provided entirely in Java, while those for FDE and BDE reductions resort on the Z3 SMT solver, accessed via its Java APIs.
The core layer also features encoding capabilities from RN to ODE (or IDOL) representation, and vice versa, as well as export/import functionalities for third-party formats, like biochemical models for
BioNetGen~\cite{Blinov22112004} and Microsoft GEC~\cite{GEC}, or ODEs defined in MATLAB. Also, models can be exported in the SBML format, an XML-based language for describing biological systems.
Finally,
this layer provides support for the numerical solutions of ODEs by means of the Apache Commons Maths library~\cite{apache:maths}.
When the input is a CRN (i.e. an RN with only positive rates) it can also be interpreted as a CTMC, following an established approach~\cite{Gillespie77}. Using the \emph{FERN} library~\cite{Erhard2008}, ERODE{} features CTMC simulation.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\textwidth]{erodescreenshotnew.eps}
\caption{A screenshot of ERODE{}.
}\label{fig:screenshot}
\end{figure}
By using a standard laptop machine, in~\cite{tacas17cttv} we have shown that our syntactic minimisation algorithms can scale to models with up to 2.5 million variables and 25 million monomials, with runtimes of a few minutes only. Instead, the symbolic minimisation scales up to 70 thousand variables in the backward case, and to 1 thousand variables in the forward case.
As discussed in~\cite{DBLP:conf/popl/CardelliTTV16}, the difference in performance and scalability between the backward and forward symbolic minimisation techniques are due to the fact that combinatorially many SMT checks among each pair of related variables are required in order to establish if a partition is FDE, while the BDE case requires only one check.
\section{Conclusions}
This paper presented two equivalences for ordinary differential equations (ODEs), forward and backward differential equivalences.
In case the ODEs can be described in terms of the IDOL language, which essentially restricts to ODEs whose derivates do not contain exponential and trigonometric functions, then such ODE systems can be minimized up to these notions of equivalence using symbolic reduction techniques based on SMT solving.
If instead the derivatives of the considered ODE system are multivariate polynomials, then more efficient syntactic minimization techniques can be used, based on an encoding of the notions of differential equivalence in terms of two probabilistic bisimulations, forward and backward RN bisimulations.
In~\cite{tacas17cttv} we have shown that the syntactic minimisation algorithms can scale to systems with millions of variables and monomials, terminating in a few minutes also in some challenging models. There are, however, further challenges ahead which we wish to tackle.
Forward bisimulation is only a sufficient condition for FDE; while it provides significant reductions in practice~\cite{DBLP:conf/concur/CardelliTTV15,DBLP:conf/tacas/CardelliTTV16}, some other examples from the literature demonstrate that the algorithm may miss some FDE reductions (see~\cite{DBLP:conf/popl/CardelliTTV16}). In ongoing work we are relaxing the notion of forward bisimulation so to fully characterize FDE
(for multivariate polynomial derivatives of degree at most two). Instead, backward bisimulation characterises BDE, and hence the corresponding syntactic minimisation technique from~\cite{DBLP:conf/tacas/CardelliTTV16} should be used when dealing with polynomial ODEs of degree at most two. Also, we plan to extend the RN representation and the two notions of forward and backward RN bisimulations in order to be able to apply them ti to polynomial derivatives with arbitrary degree.
As regards our symbolic reduction techniques, we used SMT only in a black box fashion. It could be interesting to study ad-hoc heuristics for SMT solving that could lead to better performance or scalability. For the same reason, it would be interesting to consider parallelised versions of such symbolic reduction techniques.
Also, another interesting research line related to our symbolic reduction techniques is that of further exploiting the generality of SMT in order to handle models with uncertainties in rates (a well-known issue in mathematical biology): here the SMT framework can already be easily extended to compute partitions that are differential equivalences under all possible assignments of such uncertain parameters, left as free variables in the satisfiability problem (similarly to the SMT-based parametric minimisation approach of~\cite{DKP13} for probabilistic models written in PRISM~\cite{DBLP:conf/cav/KwiatkowskaNP11}).
Lastly, we remark that our techniques regard exact aggregations. In some cases, however, one might be interested in more permissive, approximate notions that do not discriminate ODE variables with nearby trajectories
(e.g.,~\cite{DBLP:conf/concur/PierroHW03,worrell:approximating,DBLP:journals/lmcs/GuptaJP06,tac15,dsn2013,DBLP:journals/tcs/DesharnaisGJP04,DBLP:conf/mfcs/LarsenMP12}).
Such approximate variants can be considered as weaker notions that, e.g., allow variability in the parameters, considering the exact versions as a degenerate case in which no such variability is needed.
We remark that, at least for the numerical benchmarks considered so far with realistic models, the exact reductions can already be quite effective. Of course, approximate ones might be able to provide even coarser descriptions. In this case, however, the main challenge is to be able to relate the variability in the parameters tolerated by the coarsening procedure with the error incurred when considering an approximate, smaller model, instead of the original one.
In an ongoing research we are developing approximate variants of our differential equivalences, aiming at maintaining computational tractability, and certified error bounds that do not grow fast with time.
As regards our tool ERODE{}, we also plan to enrich its family of offered analysis techniques. For example, we plan to extend ERODE{}, and in particular its stochastic simulator, with statistical model checking~\cite{legay2010statistical} capabilities. %
This can be obtained upon integration with the statistical analyser MultiVeStA~\cite{sebastio2013multivesta}, which has already proven useful in the analysis of a wide range of scenarios, including software product lines~\cite{DBLP:conf/splc/BeekLLV15,DBLP:conf/isola/BeekLLV16}, crowd-steering~\cite{DBLP:conf/hpcs/PianiniSV14},
public transportation systems~\cite{DBLP:conf/ifm/GilmoreTV14},
and swarm robotics~\cite{MisscelAndPirlo}.
\paragraph{Acknowledgement.}
The author is indebted to Luca Cardelli, Mirco Tribastone, and Max Tschaikowski for the collaboration on the development of the presented framework, and for many fruitful
discussions.
\bibliographystyle{eptcs}
|
\section{Introduction}
A rapidly advancing direction in the studies of multidimensional nonlinear
wave patterns deals with two-component systems supporting the spin-orbit
coupling (SOC), which are represented by derivative linear-mixing terms in
the underlying systems of nonlinear Schr\"{o}dinger/Gross-Pitaevskii
equations (GPEs) \cite{dal,gal,zha}. This topic has drawn a great deal of
interest after it has been demonstrated, both theoretically and
experimentally, that the SOC of nonlinear spinor fields, which has been
known for a long time as a fundamentally important interaction in the
quantum mechanics of electrons in semiconductor settings, can be emulated in
a variety of bosonic media, such as atomic Bose-Einstein condensates (BECs)
\cite{dal,gal,zha}, polariton microcavity systems \cite{she,sal}, and
photonics \cite{bli}.
Recent theoretical analysis of systems combining the cubic attractive
interactions of the fields and linear SOC has produced several unexpected
results. First, in the free-space two-dimensional (2D) geometry, the SOC
governed by the Rashba Hamiltonian for BEC\ gives rise to \emph{stable}
half-fundamental-half-vortex solitons of two types, \textit{viz}., \textit{%
semi-vortices}, with one fundamental (zero-vorticity) component, and one
with vorticity $S=\pm 1$, and \textit{mixed modes}, which combine
fundamental and vortex terms with $S=\pm 1$ in both field components \cite%
{sak}. Stability of nonlinear states in SOC systems is radically different
with respect to the ones without SOC \cite{skr,HPu}. Indeed, the GPE in the
free 2D space with the cubic self-attraction, or a system of two GPEs with
self- and cross attraction, gives rise to families of zero-vorticity, alias
fundamental (Townes' \cite{Townes}) and vorticity-carrying (\cite{Yank}-\cite%
{Minsk2}) bright solitons which are completely unstable either due to the
critical collapse \cite{coll1,coll2,coll3} or due to the ring-splitting
instabilities \cite{firth1,firth2}, respectively. On the other hand, the SOC
terms come with a coefficient which fixes an inverse length scale in the
system, thus breaking the scaling invariance which makes norms of all the
solitons belonging to the Townes' family exactly equal to the critical value
necessary for the onset of the collapse. As a result, norms of the
two-component solitons of both the semi-vortex and mixed-mode types, morphed
by the interplay of the cubic attraction and SOC, are pushed \emph{below}
the collapse threshold, which immediately makes them stable \cite%
{short-review}. Actually, the stabilized 2D solitons play the role of the
ground state, which was missing in the system with the cubic attraction in
the absence of SOC. Furthermore, although in the 3D free-space binary BEC
the SOC terms, unlike the 2D case, cannot create the missing ground state,
as the corresponding \textit{supercritical collapse} has zero threshold \cite%
{coll1,coll2,coll3}, the linear SOC terms readily create metastable solitons
of the same two types, semi-vortices and mixed modes \cite{HPu}.
In microcavity polariton systems, the largest contribution to SOC originates
from the energy splitting between the TE and TM modes. If written in the
spinor form, it is tantamount to the linear inter-component coupling
represented by the second-order spatial derivatives \cite{fla}, while the
above-mentioned Rashba SOC in BEC is accounted for by the first-order
derivatives, in the leading approximation \cite{zha,HPu}. Thus, depending on
physics of the SOC system under the consideration, vorticities in the two
spinor components differ by either $\Delta S=2$ or $1$, for the second- and
first-order-derivative coupling terms, respectively.
Adding trapping or lattice potentials to the nonlinear SOC systems is a very
active direction of the current research, see, e.g., Refs. \cite%
{kar,sch,kart}. In the polariton microcavities, which are the subject of the
present work, a variety of practical methods have been developed to secure
high quality of the engineered trapping landscapes \cite{sch}. As concerns
polaritons in a simple single potential well, there are only a few studies
that addressed nonlinear modes of these traps, largely disregarding SOC,
see, e.g., Refs. \cite{ask,tra,yul,bor1}. Nevertheless, a recent
experimental work on open polaritonic resonators supplies a system with
strong SOC effects \cite{duf}.
In this work, we address an effectively 2D polariton condensate modeled by a
system of two GPEs with the SOC terms represented by the above-mentioned
second spatial derivatives. Usual ingredients of microcavity models, \textit{%
viz}., the linear gain, viscosity (diffusion of the effective fields),
nonlinear cubic loss, and an isotropic harmonic-oscillator trapping
potential, are included. In the general case, the Zeeman splitting between
the two fields is present too. Our objective is to identify stable 2D states
in this system, which turn out to be mixed modes and vortex-antivortex bound
states, that tend to be stable, severally, under the action of weak and
strong SOC, with a small bistability area in the parameter space. These
results for the symmetric system, which does not include Zeeman splitting,
are obtained by means of analytical and numerical methods in Section II. In
addition to that, in Section III we demonstrate that sufficiently strong
Zeeman splitting creates stable semi-vortex modes, with vorticities $0$ and $%
2$ in its components. The latter states are often referred to as the
half-vortices in the polariton context \cite{fla,rub,nat}, and are also
known in models without SOC \cite{rub}, or with the spin-only coupling
(direct linear mixing of the Rabi-coupling type) \cite{bor1}; see also Ref.
\cite{bor2}, as concerns vortex lattices in a scalar model, which does not
include spin effects. The paper is completed by a summary in Section IV.
\section{Analytical and numerical results for the symmetric system}
\subsection{Basic equations}
The model that we consider in this work consists of coupled GPEs for the
two-component polariton spinor wave function $(\Psi _{+},\Psi _{-})$ \cite%
{fla,ask}. We assume that the polariton condensate is pumped by means of a
nonresonant optical scheme, which creates a density of coherent and
incoherent excitons \cite{wou,kee}. Adiabatic elimination of the latter
yields the following system of coupled equations \cite{kee}:
\begin{gather}
i\partial _{t}\Psi _{+}=-{\frac{1}{2}}(1-i\eta )(\partial _{x}^{2}+\partial
_{y}^{2})\Psi _{+}+(|\Psi _{+}|^{2}+\alpha |\Psi _{-}|^{2})\Psi _{+} \notag
\\
+\beta (\partial _{x}-i\partial _{y})^{2}\Psi _{-}+i(\varepsilon -\sigma
|\Psi _{+}|^{2})\Psi _{+}+\left[ \frac{1}{2}\left( x^{2}+y^{2}\right)
+\Omega \right] \Psi _{+}, \label{+} \\
i\partial _{t}\Psi _{-}=-{\frac{1}{2}}(1-i\eta )(\partial _{x}^{2}+\partial
_{y}^{2})\Psi _{-}+(|\Psi _{-}|^{2}+\alpha |\Psi _{+}|^{2})\Psi _{-} \notag
\\
+\beta (\partial _{x}+i\partial _{y})^{2}\Psi _{+}+i(\varepsilon -\sigma
|\Psi _{-}|^{2})\Psi _{-}+\left[ \frac{1}{2}\left( x^{2}+y^{2}\right)
-\Omega \right] \Psi _{-}. \label{-}
\end{gather}%
Here, gain $\varepsilon >0$ is determined by the rate of the pumping of the
polariton density, $\sigma >0$ is the gain-saturation coefficient that
prevents an unphysical blow-up and allows for the time-independent finite
density solutions to exist in the system, and $\eta $ is an effective
polariton diffusion coefficient (viscosity), linked to the carrier diffusion
\cite{kee,sie}. Further, the strength of the repulsive intra-component
interactions is normalized to be $1$, while $\alpha =-0.05$ accounts for the
inter-component attraction, and $\beta $ is the strength of the polariton
SOC \cite{fla}, while the strength of the harmonic-oscillator potential is
scaled to be $1$. The potential corresponds to the shift of the cavity
detuning, and originates from the trapping of the photonic component of
exciton-polariton condensate. The potential can be created, for example, if
either the above-mentioned open cavity with a parabolic top mirror is used,
or one of the many available patterning techniques is applied to the top
mirror in the traditional closed-resonator setting with Bragg reflectors
\cite{sch}. Lastly, $\Omega $ characterizes the Zeeman-splitting energy,
which is proportional to the applied magnetic field. Accessible physical
values of the parameters in the linear Hamiltonian part of Eqs. (\ref{+}), (%
\ref{-}), and, in particular, the SOC strength in the presence of the
harmonic-oscillator potential have been discussed in detail in Ref. \cite%
{duf}. That work demonstrates that SOC energy may reach $1$meV, which is
comparable to the characteristic harmonic-oscillator energy, thereby making
values of $\beta $ between $0$ up to $1$ physically relevant. Achieving
large polarization energy splitting has been important, since the beginning
of the studies of microcavity polaritons, in the context of the polariton
spin Hall effect \cite{kav,ley}, see also Ref. \cite{she} for a detailed
review of the previous work. Currently, the design of strong SOC is getting
crucially important for the work towards experimental observation and
applications of polariton topological insulators \cite{sal,bar,nal,kar1}.
Stationary solutions to Eqs. (\ref{+}) and (\ref{-}), with chemical
potential $\mu $, are sought in the form of
\begin{equation}
\Psi _{\pm }=e^{-i\mu t}\psi _{\pm }\left( r,\theta \right) , \label{mu}
\end{equation}%
where $\psi _{\pm }$ are complex functions of polar coordinates $\left(
r,\theta \right) $. The substitution of expressions (\ref{mu}) in Eqs. (\ref%
{+}) and (\ref{-}) leads to the following stationary equations:%
\begin{eqnarray}
&&\mu \psi _{+}=-{\frac{1}{2}}(1-i\eta )\left( \frac{\partial ^{2}}{\partial
r^{2}}+\frac{1}{r}\frac{\partial }{\partial r}-\frac{1}{r^{2}}\frac{\partial
^{2}}{\partial \theta ^{2}}\right) \psi _{+}+(|\psi _{+}|^{2}+\alpha |\psi
_{-}|^{2})\psi _{+} \notag \\
&&+\beta \left[ e^{-i\theta }\left( \frac{\partial }{\partial r}-\frac{i}{r}%
\frac{\partial }{\partial \theta }\right) \right] ^{2}\psi
_{-}+i(\varepsilon -\sigma |\psi _{+}|^{2})\psi _{+}+\left( \frac{1}{2}%
r^{2}+\Omega \right) \psi _{+}, \label{u+} \\
&&\mu \psi _{-}=-{\frac{1}{2}}(1-i\eta )\left( \frac{\partial ^{2}}{\partial
r^{2}}+\frac{1}{r}\frac{\partial }{\partial r}-\frac{1}{r^{2}}\frac{\partial
^{2}}{\partial \theta ^{2}}\right) \psi _{-}+(|\psi _{-}|^{2}+\alpha |\psi
_{+}|^{2})\psi _{-} \notag \\
&&+\beta \left[ e^{i\theta }\left( \frac{\partial }{\partial r}+\frac{i}{r}%
\frac{\partial }{\partial \theta }\right) \right] ^{2}\psi
_{+}+i(\varepsilon -\sigma |\psi _{-}|^{2})\psi _{-}+\left( \frac{1}{2}%
r^{2}+\Omega \right) \psi _{-}, \label{u-}
\end{eqnarray}%
which are used below as the basic system.
\subsection{Linear and quasi-linear modes: vortex-antivortex states}
Before tackling the full nonlinear system, it is relevant to address the
simplest version of stationary equations (\ref{u+}) and (\ref{u-}), which
neglects the nonlinearity, gain, and diffusion, as well as assumes the
symmetry between the components, i.e., $\Omega =0$ (no Zeeman splitting):
\begin{gather}
\mu \psi _{+}=-{\frac{1}{2}}\left( \frac{\partial ^{2}}{\partial r^{2}}+%
\frac{1}{r}\frac{\partial }{\partial r}-\frac{1}{r^{2}}\frac{\partial ^{2}}{%
\partial \theta ^{2}}\right) \psi _{+}+\beta \left[ e^{-i\theta }\left(
\frac{\partial }{\partial r}-\frac{i}{r}\frac{\partial }{\partial \theta }%
\right) \right] ^{2}\psi _{-}+\frac{1}{2}r^{2}\psi _{+}, \notag \\
\mu \psi _{-}=-{\frac{1}{2}}\left( \frac{\partial ^{2}}{\partial r^{2}}+%
\frac{1}{r}\frac{\partial }{\partial r}-\frac{1}{r^{2}}\frac{\partial ^{2}}{%
\partial \theta ^{2}}\right) \psi _{-}+\beta \left[ e^{i\theta }\left( \frac{%
\partial }{\partial r}+\frac{i}{r}\frac{\partial }{\partial \theta }\right) %
\right] ^{2}\psi _{+}+\frac{1}{2}r^{2}\psi _{-}. \label{lin}
\end{gather}
It is easy to see that Eq. (\ref{lin}) gives rise to two exact eigenmodes of
the linearized dissipation-free system, corresponding to independent signs $%
\pm $ in Eq. (\ref{vortex-antivortex}):%
\begin{equation}
\psi _{+}=A_{\max }r\exp \left( -i\theta -\frac{r^{2}}{2\sqrt{1\mp 2\beta }}%
\right) ,~\psi _{-}=\pm A_{\max }r\exp \left( i\theta -\frac{r^{2}}{2\sqrt{%
1\mp 2\beta }}\right) , \label{vortex-antivortex}
\end{equation}%
\begin{equation}
\mu _{\pm }^{(0)}=2\sqrt{1\mp 2\beta }, \label{mu0}
\end{equation}%
where $A_{\max }$ is an arbitrary constant, and superscript $(0)$ attached
to $\mu _{\pm }$ refers to the linear approximation. These exact eigenmodes,
which, as a matter of fact, are found following the pattern of standard
solutions for the 2D isotropic harmonic oscillator in quantum mechanics, may
be naturally called vortex-antivortex states. Note that the symmetric
vortex-antivortex state with the lower eigenvalue of the energy, $\mu
_{+}^{(0)}=2\sqrt{1-2\beta }$, exists if the SOC is not too strong, \textit{%
viz}., at $\beta \leq 0.5$, while its antisymmetric counterpart with higher
energy, $\mu _{-}^{(0)}=2\sqrt{1+2\beta }$, exists at all values of $\beta $%
. In fact, the disappearance of the eigenmode (\ref{vortex-antivortex})
corresponding to the top sign in $\mp $ at $\beta =0.5$, where the radial
size of the eigenmode shrinks to zero as $\left( 1-2\beta \right) ^{1/4}$,
may be considered as a quantum phase transition in the polariton condensate,
cf. Ref. \cite{phase-transition}.
Both types of the vortex-antivortex states, antisymmetric and symmetric
ones, were experimentally observed, in a quasi-linear regime, in Ref. \cite%
{duf}, where they were named, respectively, modes of types (i) and (iii). In
terms of the present notation, the respective SOC\ coupling, estimated in
Ref. \cite{duf}, was $\beta \approx 0.04$.
In a sense, the exact vortex-antivortex modes are similar to the
above-mentioned semi-vortices found in the nonlinear dissipation-free 2D\
SOC\ model of the atomic condensate considered in Ref. \cite{sak}, which
feature zero vorticity in one component and vorticity $\pm 1$ in the other.
Indeed, the difference of the vorticities in the components of the
semi-vortex is $\Delta S=1$ because (as mentioned above too) SOC in the
atomic condensates is represented by the linear differential operator of the
first order \cite{zha}, while in the present system the difference between
the components of the vortex-antivortex is $\Delta S=2$, as the SOC is
represented by the second-order operator in Eqs. (\ref{+}) and (\ref{-}).
If the nondissipative nonlinear terms are kept in Eqs. (\ref{u+}) and (\ref%
{u-}) as small perturbations, it is easy to find the first nonlinear
correction to the chemical potential, $\mu _{\pm }^{(1)}$, using the
commonly known method in quantum mechanics, that generates the first
correction as the spatial average of the perturbation potential \cite{skr,LL}%
. A simple calculation, which makes use of the unperturbed wave functions (%
\ref{vortex-antivortex}), yields
\begin{equation}
\mu _{\pm }^{(1)}=\frac{1+\alpha }{4}\sqrt{1\mp 2\beta }A_{\max }^{2},
\label{mu1}
\end{equation}%
where $A_{\max }^{2}$ should be small enough, to keep correction (\ref{mu1})
small in comparison with the result (\ref{mu0}), obtained in the linear
limit.
Furthermore, if the linear gain, diffusion, and nonlinear loss are also kept
as small perturbations in Eqs. (\ref{u+}) and (\ref{u-}), then amplitude $%
A_{\max }$ of the linear mode (\ref{vortex-antivortex}), which is treated in
Eqs. (\ref{vortex-antivortex}) and (\ref{mu1}) as a free parameter, is
uniquely selected by the obvious balance condition for the total norm (it is
defined below by Eq. (\ref{N})), written as%
\begin{equation}
\varepsilon \int_{0}^{\infty }\left\vert \psi _{\pm }(r)\right\vert
^{2}rdr=\sigma \int_{0}^{\infty }\left\vert \psi _{\pm }(r)\right\vert
^{4}rdr+\frac{\eta }{2}\int_{0}^{\infty }\left[ \left\vert \frac{d\psi _{\pm
}}{dr}\right\vert ^{2}+\frac{\left\vert \psi _{\pm }(r)\right\vert ^{2}}{%
r^{2}}\right] rdr, \label{balance}
\end{equation}%
for the mode $\left\vert \psi _{+}(r)\right\vert =|\psi _{-}(r)|$ taken as
per Eq. (\ref{vortex-antivortex}). The result of the simple calculation is%
\begin{equation}
\left( A_{\max }\right) _{\mathrm{balance}}^{2}=\frac{\varepsilon \sqrt{1\mp
2\beta }}{2\sigma \left( 1\mp 2\beta \right) +\eta }, \label{A^2}
\end{equation}%
which makes sense if the pump rate $\varepsilon $ is small enough. A more
sophisticated analytical approximation for the vortex-antivortex in the full
system is developed below, see Eqs. (\ref{ans})-(\ref{NN}).
\subsection{Mixed-mode states}
In addition to the two species of vortex-antivortex complexes, built as per
Eqs. (\ref{vortex-antivortex}), the linearized coupled GPEs (\ref{lin}) give
rise to another type of eigenstates, which, following Ref. \cite{sak}, may
be called mixed modes, as they contain terms with zero vorticity and
vorticities $\pm 2$ in each component. Like the symmetric vortex-antivortex
states (\ref{vortex-antivortex}), the mixed modes obey constraint $\psi
_{+}=\psi _{-}^{\ast }$, with $\ast $ standing for the complex conjugate.
Linear mixed modes cannot be represented by an exact solution (unlike Eq. (%
\ref{vortex-antivortex}) for the vortex-antivortex complexes), but they can
be easily constructed in an approximate form for $\beta \ll 1$, starting
from an obvious ground state of the harmonic-oscillator in each component at
$\beta =0$:
\begin{equation}
\psi _{\pm }\approx A_{\max }\exp \left( -\frac{r^{2}}{2}\right) \left[ 1\pm
i\beta r^{2}\sin (2\theta )\right] ,\;\mu \approx 1-\beta .
\label{mixed mode}
\end{equation}%
At larger $\beta $, the mixed mode eigenmodes were constructed numerically
by means of the imaginary-time method, applied to the time-dependent version
of linearized equations (\ref{lin}), starting with the ground-state input, $%
\Psi _{+}=\Psi _{-}=\exp (-r^{2}/2)$. The numerically found mixed-mode
shapes are displayed in Fig. \ref{f1}, and the comparison with the
analytical approximation (\ref{mixed mode}), for a relevant small value $%
\beta =0.1$, is presented in Fig. \ref{f1}(b).
Further, eigenvalues corresponding to the mixed modes were compared to the
analytically found eigenvalues for the vortex-antivortex states, given by
Eq. (\ref{mu0}), with the aim to identify the ground state, which must have
the smallest eigenvalue of energy, $\mu $, for given $\beta $. As a result,
it was found that, at $\beta <0.46$, the mixed modes realize the ground
state (for small $\beta $, this is evident from the comparison of Eqs. (\ref%
{mu0}) and (\ref{mixed mode})), but in a narrow interval of $0.46<\beta <0.5$%
, the symmetric vortex-antivortex complex, with the upper sign in Eq. (\ref%
{vortex-antivortex}), produces a slightly smaller value of $\mu $, which is
given by Eq. (\ref{mu0}), as shown in Fig. \ref{f1}(c); recall that
vortex-antivortex state with the upper sign does not exist at $\mu >0.5$.
The same figure \ref{f1}(c) shows that the mixed mode does not exists either
at $\beta >0.5$ (which is another essential manifestation of the
above-mentioned quantum phase transition), hence the only eigenmode which
survives at $\beta >0.5$ is the antisymmetric vortex-antivortex complex with
the lower sign in Eq. (\ref{vortex-antivortex}). It is worthy to mention
here that, as shown below, solely vortex-antivortex modes with \emph{%
opposite signs} of their components are found as stable states in the full
system of Eqs. (\ref{+}) and (\ref{-}), while vortex-antivortex bound states
with identical signs of its components cannot be generated by the full
system, even if they may realize the ground state of the linearized system
at $\beta $ close to $0.5$.
\begin{figure}[tbp]
\begin{center}
\includegraphics[height=10cm]{fig1-eps-converted-to.pdf}
\end{center}
\caption{(a) Profiles of $|\protect\psi _{+}|=|\protect\psi _{-}|$ for
mixed-mode eigenstates of the linearized system (\protect\ref{lin}),
obtained as numerical imaginary-time solutions for values of the SOC
strength $\protect\beta =0.5,0.48,0.4$, $0.35$ and $0$ (from bottom to top).
(b) Imaginary parts of the numerically found solution for $\protect\psi _{+}$
(solid curve), and $\protect\psi _{-}$ (dotted curve), shown along the
diagonal cross section, $y=x$ (i.e., $\protect\theta =\protect\pi /4$), for $%
\protect\beta =0.1$. The dashed line, which completely overlaps with the
solid one, is the analytical prediction, Im$\left( \protect\psi _{+}\right)
=2A_{\max }\protect\beta x^{2}\exp \left( -x^{2}\right) $, given by Eq. (%
\protect\ref{mixed mode}). (c) Energy eigenvalues of the mixed-mode and
symmetric vortex-antivortex states, shown by the solid and dashed lines,
respectively, versus the SOC strength, $\protect\beta $. The exact
vortex-antivortex' eigenvalue is given by Eq. (\protect\ref{mu0}) with the
upper sign. (d) Energy eigenvalues of the mixed-mode and vortex-antivortex
states for $0.45<\protect\beta <0.5$.}
\label{f1}
\end{figure}
\begin{figure}[tbp]
\begin{center}
\includegraphics[height=5cm]{fig2-eps-converted-to.pdf}
\end{center}
\caption{A set of profiles of components $A\equiv |\protect\psi _{+}|=|%
\protect\psi _{-}|$ of the ground state of the nonlinear dissipation-free
system (\protect\ref{nonlin}) at values of the SOC$\ $strength $\protect%
\beta =0.48,0.45,0.42$ (from bottom to top). A transition from the mixed
mode to the symmetric vortex-antivortex complex takes place as $\protect%
\beta $ increases from $0.45$ to $0.48$.}
\label{f2}
\end{figure}
\begin{figure}[tbp]
\begin{center}
\includegraphics[height=5cm]{fig3-eps-converted-to.pdf}
\end{center}
\caption{(a) Profiles of components $\left\vert \Psi _{+}\right\vert
=\left\vert \Psi _{-}\right\vert $ in a typical stable mixed mode, produced
by Eqs. (\protect\ref{+}) and (\protect\ref{-}) at $\protect\eta =0.1$, $%
\protect\varepsilon =0.4$, $\protect\beta =0.3$, and $\Omega =0$. (b) The
same for a typical semi-vortex, with vorticities $0$ and $2$ in components $%
\Psi _{+}$ and $\Psi _{-}$, respectively, which is produced by the numerical
solution of Eqs. (\protect\ref{+}) and (\protect\ref{-}) in the presence of
the Zeeman splitting with $\Omega =0.2$. Other parameters are $\protect\eta %
=0.4$, $\protect\varepsilon =0.6$, and $\protect\beta =0.6$.}
\label{f3}
\end{figure}
The next step is to consider Eqs. (\ref{+}) and (\ref{-}) in which nonlinear
terms are kept, but the dissipative ones are still ignored:
\begin{gather}
i\partial _{t}\Psi _{+}=-{\frac{1}{2}}(\partial _{x}^{2}+\partial
_{y}^{2})\Psi _{+}+\frac{1}{2}\left( x^{2}+y^{2}\right) \Psi _{+}+(|\Psi
_{+}|^{2}+\alpha |\Psi _{-}|^{2})\Psi _{+}+\beta (\partial _{x}-i\partial
_{y})^{2}\Psi _{-}, \notag \\
i\partial _{t}\Psi _{-}=-{\frac{1}{2}}(\partial _{x}^{2}+\partial
_{y}^{2})\Psi _{-}+\frac{1}{2}\left( x^{2}+y^{2}\right) \Psi _{-}+(|\Psi
_{-}|^{2}+\alpha |\Psi _{+}|^{2})\Psi _{-}+\beta (\partial _{x}+i\partial
_{y})^{2}\Psi _{+}. \label{nonlin}
\end{gather}%
Solving these equations by means of the imaginary-time method makes it
possible to produce stationary solutions with a fixed norm,%
\begin{equation}
N\equiv \int \int \left[ \left\vert \Psi _{+}\left( x,y\right) \right\vert
^{2}+\left\vert \Psi _{-}\left( x,y\right) \right\vert ^{2}\right] dxdy=1,
\label{N}
\end{equation}%
and eventually identify nonlinear ground states, which realize the minimum
of the Hamiltonian,%
\begin{gather}
H=\int \int \left\{ \frac{1}{2}\left( |\nabla \Psi _{+}|^{2}+|\nabla \Psi
_{-}|^{2}\right) +\frac{1}{2}\left( x^{2}+y^{2}\right) \left( \left\vert
\Psi _{+}\right\vert ^{2}+\left\vert \Psi _{-}\right\vert ^{2}\right)
+\alpha |\Psi _{+}|^{2}\left\vert \Psi _{-}\right\vert ^{2}\right. \notag
\\
\left. +\frac{1}{2}\left( |\Psi _{+}|^{4}+\left\vert \Psi _{-}\right\vert
^{4}\right) -\beta \left[ \left( (\partial _{x}+i\partial _{y})\Psi
_{+}\right) ^{\ast }(\partial _{x}-i\partial _{y})\Psi _{-}+\mathrm{c.c.}%
\right] \right\} dxdy \label{E}
\end{gather}%
(here c.c. stands for the complex-conjugate expression), for the given norm.
Figure \ref{f2}(a) shows the so found profiles of $|\Psi _{+}|$ and $|\Psi
_{-}|$ for the ground state at $\beta =0.48,0.45,0.42$. In agreement with
the above-mentioned transition of the ground state from the mixed mode to
the symmetric vortex-antivortex in the linearized system, as the SOC
strength changes from $\beta <0.46$ to $\beta >0.46$, Fig. \ref{f2}
demonstrates that, in the dissipation-free nonlinear system, the same
transition takes place as the SOC strength increases from $\beta =0.45$ to $%
\beta =0.48$.
\subsection{Stable modes in the full nonlinear dissipative system}
The systematic numerical analysis demonstrates that the full system of Eqs. (%
\ref{+}) and (\ref{-}) gives rise to stable stationary states \emph{only if}
the diffusion term is present, $\eta >0$ (at $\eta =0$, chaotic solutions
appear, which are not shown here in detail). For $\beta $ not two large, the
system with small values of $\eta >0$ generates stable patterns of the
mixed-mode type, which thus plays the role of the fundamental state in the
full system, see an example in Fig. \ref{f3}(a).
At larger values of the SOC strength, $\beta $, and $\eta >0$, the full
system readily produces the fundamental state in the form of stable
vortex-antivortex complexes with the opposite sign of the components, which,
in the linear limit, correspond to the lower sign (antisymmetric bound
state) in Eq. (\ref{vortex-antivortex}). A typical example is presented in
Fig. \ref{f4}, where panel (a) shows real parts of $\Psi _{+}$ and $\Psi _{-}
$ (with the opposite signs, as said here), and panel (b) shows $|\Psi _{\pm
}(x)|$ (the solid line) in the cross section of $y=0$. We stress that the
numerical solution of the full system, including the nonlinear and
dissipative terms, has not produced any stable vortex-antivortex complex
with identical signs of its two component, in spite of the fact that such
complexes play the role of the ground state in the dissipation-free system
close to $\beta =0.5$, as shown above.
The vortex-antivortex states produced by the full system can be quite
efficiently approximated in an analytical form. To this end, the following
ansatz is adopted for their two components:
\begin{equation}
\psi _{\pm }=\pm A_{\max }r\exp \left[ -(1/2)\rho r^{2}\mp i\theta \right] ,
\label{ans}
\end{equation}%
with free parameters $A_{\max }$ and $\rho $, cf. Eq. (\ref%
{vortex-antivortex}). First, ignoring the dissipative terms, it is possible
to predict values of $A_{\max }$ and $\rho $ by minimizing the system's
energy (\ref{E}) at a fixed value of norm $N$ (this time, we do not set $N=1$%
). The results is
\begin{equation}
\rho =\left[ 1+2\beta +(1+\alpha )N/8\right] ^{-1/2},~A_{\max }=\sqrt{N}\rho
. \label{rho}
\end{equation}%
Further, restoring the dissipative terms $\sim $ $\varepsilon ,\eta $, and $%
\sigma $, the norm, which appears as a free parameter in Eq. (\ref{rho}),
can be predicted by the balance equation (\ref{balance}). The final result is%
\begin{equation}
N=\frac{4}{\sigma ^{2}}\left\{ -\left( \sigma \eta -\epsilon ^{2}(1+\alpha
)/4\right) +\sqrt{\left[ \epsilon ^{2}(1+\alpha )/4-\sigma \eta \right]
^{2}-\sigma ^{2}\left[ \eta ^{2}-\epsilon ^{2}(1+2\beta )\right] }\right\} .
\label{NN}
\end{equation}%
The comparison of the vortex-antivortex modes predicted by Eqs. (\ref{ans})-(%
\ref{NN}) with their numerically found counterparts is presented in Figs. %
\ref{f4}(b) and (c), showing good accuracy of the analytical approximation.
Note that Fig. \ref{f4}(c) demonstrates that this family of the
vortex-antivortex states does not extend to $\beta <0.4$.
\begin{figure}[tbp]
\begin{center}
\includegraphics[height=4.7cm]{fig4-eps-converted-to.pdf}
\end{center}
\caption{(a) Real parts of components $\protect\psi _{+}$ and $\protect\psi %
_{-}$ (solid and dashed lines, respectively) of a stable vortex-antivortex\
produced by the full system at $\protect\varepsilon =0.6,\protect\eta =0.8,$
$\protect\sigma =0.3$, $\protect\beta =0.75$, and $\Omega =0$. (b) The
comparison of $|\protect\psi _{+}|=\left\vert \protect\psi _{-}\right\vert $%
, for the same numerically found vortex-antivortex (solid curve), and the
respective ansatz (\protect\ref{ans}) with $A_{\max }=0.9936$ and $\protect%
\rho =0.5942$ (dashed curve), as predicted by the analytical approximation.
(c) Maximum value $A_{\max }$ of the numerically found $|\protect\psi _{+}|$
component (rhombi) and the respective analytical prediction (the dashed
line) versus the SOC\ strength $\protect\beta $, for the same parameters as
in (a).}
\label{f4}
\end{figure}
\begin{figure}[tbp]
\begin{center}
\includegraphics[height=6cm]{f5a.jpg} \includegraphics[height=6cm]{f5b.jpg}
\end{center}
\caption{(a) Boundaries of areas of the existence of stable mixed mode and
vortex-antivortex states, in the plane of ($\protect\beta ,\protect\eta $).
Other parameters are $\protect\varepsilon =0.6,$ $\protect\sigma =0.3$, and $%
\Omega =0$. (b) The lower existence boundary for semi-vortices in the plane
of $\left( \protect\beta ,\Omega \right) $. Other parameters are $\protect%
\varepsilon =0.6$, $\protect\eta =0.4$, and $\protect\sigma =0.3$.}
\label{f5}
\end{figure}
Results of the systematic numerical analysis are summarized in Fig. \ref{f5}%
(a), which displays the existence areas (between the solid and dashed lines)
for stable mixed-mode and vortex-antivortex states, in the plane of the most
essential control parameters, \textit{viz}., the SOC strength, $\beta $, and
diffusion coefficient, $\eta $ . As mentioned above, the mixed modes and
antisymmetric vortex-antivortex complexes tend to be stable, as ground
states, at smaller and larger values of $\beta ,$ respectively, with a small
bistability region revealed by the figure. Above the dashed lines, the
solutions decay to zero, while below the solid lines, the evolution makes
them chaotic.
\section{Stable semi-vortex states supported \ by the Zeeman splitting}
The Zeeman splitting, represented by $\Omega \neq 0$ in Eqs. (\ref{+}) and (%
\ref{-}), can strongly impact properties of nonlinear modes in the present
system, as is known to happen in other polariton models, see, e.g., Ref.
\cite{gul}. We here aim to outline general trends imposed by the Zeeman
splitting, without producing full details. As suggested by the analysis of
the spin-orbit-coupled two-component atomic condensates \cite{Sherman}, the
enhancement of the Zeeman splitting tends to replace mixed modes by the
semi-vortices, in which, in the present case, one component has zero
vorticity, and the other carries vorticity $2$. The present system does not
give rise to semi-vortices at $\Omega =0$; however, they appear with the
increase of $\Omega $.
The case of very large $\Omega $ can be considered by means of an analytical
approximation. In this case, the chemical potential contains a large term, $%
-\Omega $, suggesting one to substitute
\begin{equation}
\Psi _{\pm }\left( x,y,t\right) \equiv e^{i\Omega t}\Phi _{\pm }\left(
x,y,t\right) , \label{Omega}
\end{equation}%
where the time dependence in $\Phi _{\pm }$ is assumed to be slow in
comparison to $e^{i\Omega t}$. Then, using an approximation similar to that
developed in Ref. \cite{Sherman}, we use Eq. (\ref{+}) to eliminate $\Phi
_{+}$ in favor of $\Phi _{\_}$:
\begin{equation}
\Phi _{+}\left( x,y,t\right) \approx -\frac{\beta }{2\Omega }(\partial
_{x}-i\partial _{y})^{2}\Phi _{-}\left( x,y,t\right) ~. \label{PsiPsi}
\end{equation}%
Finally, the substitution of this approximation in Eq. (\ref{-}) reduces it
to a single\textit{\ fourth-order} equation for the wave function $\Phi _{-}$%
:
\begin{eqnarray}
&&i\partial _{t}\Phi _{-}=-{\frac{1}{2}}\left( 1-i\eta \right) (\partial
_{x}^{2}+\partial _{y}^{2})\Phi _{-}+|\Phi _{-}|^{2}\Phi _{-}-\frac{\beta
^{2}}{2\Omega }(\partial _{x}^{2}+\partial _{y}^{2})^{2}\Phi _{-} \notag \\
&&+i(\varepsilon -\sigma |\Phi _{-}|^{2})\Phi _{-}+\frac{1}{2}%
(x^{2}+y^{2})\Phi _{-}. \label{Psi}
\end{eqnarray}%
In particular, in the case of $\eta =\varepsilon =\sigma =0$, we arrive at a
linear or nonlinear Schr\"{o}dinger equation with a kinetic \textit{%
super-energy} term, expressed by the fourth-order operator $\sim \beta
^{2}/\Omega $:%
\begin{equation}
i\partial _{t}\Phi _{-}=-{\frac{1}{2}}(\partial _{x}^{2}+\partial
_{y}^{2})\Phi _{-}-\frac{\beta ^{2}}{2\Omega }(\partial _{x}^{2}+\partial
_{y}^{2})^{2}\Phi _{-}+|\Phi _{-}|^{2}\Phi _{-}+\frac{1}{2}(x^{2}+y^{2})\Phi
_{-}. \label{super}
\end{equation}%
It is obvious that the present approximation gives rise to semi-vortices,
with a large zero-vorticity component produced by isotropic equation (\ref%
{Psi}), and a small vortex one generated from it by Eq. (\ref{PsiPsi})
An example of a stable semi-vortex is presented in Fig. \ref{f3}(b), which
shows profiles of absolute values of the zero-vorticity component (the solid
curve), $\left\vert \Psi _{+}\right\vert $, and its counterpart with
vorticity $2$ ($|\Psi _{-}|$, the dashed line). Results of the systematic
study of the semi-vortices are summarized in Fig. \ref{f5}(b), in the
parameter plane of $(\beta ,\Omega )$. Stable semi-vortices exist above the
boundary shown in the figure, provided that $\Omega $ exceeds a finite
threshold value, $\Omega _{\min }\approx 0.16$.
\section{Summary}
The aim of this work is to identify species of robust 2D localized modes
which play the role of fundamental states in two-component (spinor)
exciton-polariton condensates, subject to the action of SOC (spin-orbit
coupling), represented by the second-order linear differential operator
which mixes the two components. The system includes ordinary terms which are
common to semiconductor-microcavity models, such as the linear gain and
effective diffusion, nonlinear loss, self-repulsive nonlinearity in each
component, and the isotropic harmonic-oscillator trapping potential.
Starting from the analysis of the linearized dissipation-free system, and
then proceeding to its full form, we have found, by means of numerical and
analytical methods, that basic states supported by the system are
antisymmetric vortex-antivortex complexes and mixed modes. The latter ones
combine zero-vorticity and vorticity-carrying terms in each component. The
mixed modes and vortex-antivortex complexes tend to realize the fundamental
states (similar to the ground states of conservative systems) when SOC is,
respectively, weak or strong, with a small region of bistability. The
presence of the effective diffusion is a necessary condition for the
stability of modes of both types in the full system. We have also found that
the addition of the Zeeman splitting tends to replace these modes by stable
semi-vortices, with vorticities $0$ in the larger component and $2$ in the
smaller one.
\section*{Funding}
Royal Society (IE 160465); ITMO University through the Government of Russia
(074-U01); Binational (US-Israel) Science Foundation (2015616).
|
\section{Introduction}
\subsection{Background}
Let $M$ be an $n$-dimensional manifold and consider the $1$-jet space
$J^1(M)= T^*M \times {\mathbb R}$ of $M$. This space can be given the structure
of a contact manifold, with contact form $\alpha= dz - \sum_j y_j
dx_j$. Here $(x,y)$ are coordinates on $T^*M$ and $z$ is the
coordinate in the ${\mathbb R}$-direction. An $n$-dimensional submanifold
$\Lambda \subset J^1(M)$ is called \emph{Legendrian} if it is
everywhere tangent to the contact distribution $\xi = \operatorname{Ker} \alpha$,
and a \emph{Legendrian isotopy} is a smooth 1-parameter family of
Legendrian submanifolds. A major problem in contact geometry is to
determine whether two given Legendrian
submanifolds are \emph{Legendrian isotopic}, i.\ e. if there is a
Legendrian isotopy connecting them.
To that end, a number of \emph{Legendrian invariants} have
been introduced. These are objects associated to Legendrian
submanifolds, invariant under Legendrian isotopies.
One such invariant
is \emph{Legendrian contact homology}, which is the homology of a
differential graded algebra (DGA) associated to the Legendrian $\Lambda$. This
algebra is called the \emph{Chekanov-Eliashberg algebra of $\Lambda$},
and we denote it by $\mathcal{A}(\Lambda)$. It is a free, unital algebra
generated by the \emph{Reeb chords} of
$\Lambda$, which are flow segments of the \emph{Reeb vector field}
$\partial_z$, having their start and end points on $\Lambda$. We assume
that $\Lambda$ is chord generic, meaning that the \emph{Lagrangian
projection} $\Pi_{\mathbb C}: J^1(M) \to T^*M$ projects the Reeb chords of
$L$ to isolated double points of $\Pi_{\mathbb C}(\Lambda)$. The differential
of the DGA is defined by counting certain pseudo-holomorphic disks.
Legendrian contact homology fits into the machinery of
\emph{Symplectic field theory}, introduced by Eliashberg,
Givental and Hofer in \cite{sft}. In particular, let $L$
be an exact Lagrangian cobordism in the \emph{symplectization} $ ({\mathbb R} \times
J^1(M), d(e^t\alpha))$ of $J^1(M)$. Assume that $L$ is asymptotic to
cylinders ${\mathbb R} \times \Lambda_\pm$ at $\pm \infty$, where
$\Lambda_\pm\subset J^1(M)$
are Legendrians. According to \cite{ratsft}, if we choose the coefficient ring to be given by ${\mathbb Z}_2$, then $L$ induces a DGA-morphism $\Phi_L:\mathcal{A}(\Lambda_+)
\to \mathcal{A}(\Lambda_-)$ in a functorial way. Here $\Phi_L$ is defined via a modulo 2 count of
pseudo-holomorphic disks with boundary on $L$. This is used in \cite{tobhonda} to
derive results about isotopy classes of exact Lagrangians with
prescribed boundary. More precisely, these results were derived
from explicit descriptions of $\Phi_L$ in the case
$L$ is induced by the trace of an elementary Legendrian isotopy.
That Legendrian contact homology can be defined over ${\mathbb Z}$, provided $\Lambda$ is spin, is proven in
\cite{orientbok}. In that paper the differential of $\mathcal{A}(\Lambda)$ is
defined by a count of rigid pseudo-holomorphic disks in the Lagrangian
projection, with the disks having boundary
on $\Pi_{\mathbb C}(\Lambda)$. To get a signed count of these disks, it is shown that the moduli space of pseudo-holomorphic disks admits a coherent orientation. However, there is another way to define the
differential, which is more convenient if one wants to consider the
functorial properties in Symplectic field theory. That method is to
count rigid pseudo-holomorphic disks in the symplectization
of $J^1(M)$, with the disks having boundary on ${\mathbb R} \times \Lambda$.
In
\cite{Georgios} it is proven that
these two different counts give the same
DGA, given that we
work with ${\mathbb Z}_2$-coefficients. We will prove that this also holds with
${\mathbb Z}$-coefficients, provided that $\Lambda$ is spin.
More precisely, we will prove that the coherent orientation scheme
given in \cite{orientbok} can be lifted to give a coherent orientation for moduli spaces of pseudo-holomorphic disks in ${\mathbb R} \times
J^1(M)$ with boundary on ${\mathbb R} \times \Lambda$.
Then we prove that this lifted orientation scheme allows us to extend
the definition of $\Phi_L$ from \cite{ratsft} to ${\mathbb Z}$-coefficients, provided that $L$ is spin and that
$\Lambda_\pm$ are given the induced spin structure as boundary of
$L$. That this lift can be performed seems in particular important if
one wants to relate SFT theories with Floer theories, for example via
Seidel's isomorphisms which briefly says that if $\Lambda$ admits an
exact Lagrangian filling $L$, then there is an isomorphism between
$H_*(L)$ and the
linearized Legendrian contact cohomology of $\Lambda$ with respect to
the augmentation induced by $L$. Compare with \cite{Georgios} and \cite{georgios-roman}.
For other examples of applications of such a signed lift which allows $\Phi_L$ to be defined over the integers, see e.g.\ \cite{floerlag}, \cite{floerlag2}, \cite{toblose}, \cite{el}. Note that the existence of such a signed lift is indicated but not proved in these papers.
Yet another motivation for understanding DGA-morphisms with coefficients in ${\mathbb Z}$ comes from the connection between Legendrian contact homology and homological mirror symmetry, together with the machinery in \cite{rm}.
The coherent orientation scheme for the moduli spaces of pseudo-holomorphic
disks will be defined by using something called \emph{capping
operators}, which are $\bar{\partial}$-operators defined on the 1-punctured
unit disk in ${\mathbb C}$ with trivialized Lagrangian boundary
conditions. Using the DGA-morphism induced by the trivial cobordism
${\mathbb R} \times \Lambda$, we will derive an expression of how the DGA
changes if we change capping operators. In this way we can relate the
orientation scheme of pseudo-holomorphic disks in $T^*M$ given in
\cite{orientbok} with the one given in
\cite{orienttrees}.
The orientation scheme defined in \cite{orienttrees} is adapted to the situation
when the differential in Legendrian contact homology is defined by
counting rigid \emph{Morse flow trees} instead of pseudo-holomorphic
disks. We refer to \cite{trees} for the definition of these trees,
and for the proof that the trees can replace the pseudo-holomorphic disks
in the definition of the differential if we work with
${\mathbb Z}_2$-coefficients. In \cite{orienttrees}, this result is extended
to also hold for ${\mathbb Z}$-coefficients. The advantage of using Morse flow trees instead of
pseudo-holomorphic disks is that the former ones can be found using
finite-dimensional flow techniques, while the latter ones give rise
to non-linear PDE:s, which in general are hard to solve.
In \cite{tobhonda} it is shown that one can use Morse flow trees to
compute the DGA-morphism induced by an exact Lagrangian cobordism, in
the case when the coefficients are given by ${\mathbb Z}_2$. This is one of the
reasons why the DGA-morphisms induced by traces of
elementary Legendrian isotopies can be described explicitly when $n=1$. We sketch an argument that Morse flow trees can be used to compute DGA-morphisms also with integer coefficients, given our orientation scheme of moduli spaces.
\subsection{Organization of the paper}
In Section \ref{sec:mainresults} we give a definition of
the DGA associated to a Legendrian $\Lambda \subset J^1(M)$, and the
DGA-morphism induced by an exact Lagrangian cobordism. We also state
the main theorems.
In Section \ref{sec:holo} we recall the definition of punctured
pseudo-holomorphic disks, and give a more detailed definition of the
relevant moduli spaces. In Section \ref{sec:oc} we fix orientation
conventions, and prove that these conventions make it possible to define Legendrian contact homology with integer coefficients in the symplectization setting. In Section \ref{sec:cob-pf} we prove that this also gives the desired results for the DGA-morphisms induced by exact Lagrangian cobordisms. In Section \ref{sec:trees} we discuss how the
orientation scheme can be used to orient the moduli space of
Morse flow trees associated to exact Lagrangian cobordisms.
\subsection*{Acknowledgments}
This paper is built on parts of the author's PhD thesis at Uppsala University. The work was further developed when the author was a postdoc at the University of Nantes, and completed while the author was a postdoc at Stanford University.
The author would like to thank Tobias Ekholm and Paolo Ghiggini for useful discussions.
\section{Main results}
\label{sec:mainresults}
Here we formulate the main results. To be able to do this, we first
need to introduce some more notation.
\subsection{Legendrian contact homology}
As outlined in the Introduction, there are two different ways of defining the differential $\partial$ of $\mathcal{A}(\Lambda)$. One method is to compute punctured, rigid, pseudo-holomorphic disks in $T^*M$ with boundary on $\Pi_{\mathbb C}(\Lambda)$. I.e., the differential is defined on generators $a$ by
\begin{equation*}
\partial_l(a) = \sum_{\dim \mathcal{M}_{l,\Lambda}(a,{\bf b}) = 0} |\mathcal{M}_{l,\Lambda}(a, {\bf b})| {\bf b},
\end{equation*}
and extended by the Leibniz rule to the rest of the algebra. Here ${\bf b} = b_1\dotsm b_m$ is a word of Reeb chords, $\mathcal{M}_{l,\Lambda}(a,{\bf b})$ is the moduli space of rigid pseudo-holomorphic punctured disks with a positive puncture at $a$, negative punctures at $b_1,\dotsc,b_m$, and with boundary on $\Pi_{{\mathbb C}}(\Lambda)$, and $|\mathcal{M}_{l,\Lambda}(a, {\bf b})|$ denotes the algebraic count of disks in the moduli space. We refer to Section \ref{sec:holo} for more details.
We denote the DGA defined in this way by $(\mathcal{A}(\Lambda), \partial_l; R)$, where $R$ indicates the coefficient ring. In \cite{pr} it is proven that for a generic choice of compatible almost complex structure on $T^*M$, this differential satisfies $\partial_l^2 =0$, and the homology of this complex gives a well-defined Legendrian isotopy invariant if we choose the coefficient ring to be ${\mathbb Z}_2$. In \cite{orientbok} these results were extended to hold for ${\mathbb Z}$-coefficients in the case when $\Lambda$ is spin. In the special case $n=1$ and $M={\mathbb R}$, these results were first established in \cite{chekanov} for the case of ${\mathbb Z}_2$-coefficients, and in \cite{sabloffng} for ${\mathbb Z}$-coefficients.
The other method of computing the differential, which was discussed in \cite{sft}, and where the details were worked out in \cite{eliinv} for $n=1$, and further developed in \cite{ratsft} for higher dimensions, is to count rigid pseudo-holomorphic disks in the symplectization of $J^1(M)$. That is, in this case the differential is defined by
\begin{equation*}
\partial_s(a) = \sum_{\dim \hat\mathcal{M}_{s,\Lambda}(a,{\bf b}) = 1} |\hat\mathcal{M}_{s,\Lambda}(a, {\bf b})/{\mathbb R}| {\bf b}
\end{equation*}
on generators, and again extended by the Leibniz rule to the whole
algebra. Here $\hat\mathcal{M}_{s,\Lambda}(a,{\bf b})$ is the moduli space of
punctured pseudo-holomorphic disks with boundary on ${\mathbb R} \times \Lambda$,
having a positive puncture asymptotic to a strip over the Reeb chord
$a$ at $t=+ \infty$, and having negative punctures asymptotic to
strips over the Reeb chords $b_1,\dotsc,b_m$ at $t=-\infty$. Moreover,
we assume that the given almost complex structure is cylindrical, so that we get an induced ${\mathbb R}$-action on $\hat
\mathcal{M}_{s,\Lambda}$, given by translation in the $t$-direction. See Section
\ref{sec:holo}.
We let
\begin{equation*}
\mathcal{M}_{s,\Lambda}=\hat\mathcal{M}_{s,\Lambda}(a, {\bf b})/{\mathbb R}
\end{equation*}
be the space where we have divided out this ${\mathbb R}$-action. For a generic choice of cylindrical almost complex structure we have $\partial_s^2=0$, given that we are using ${\mathbb Z}_2$-coefficients,
and the homology of $\mathcal{A}(\Lambda)$ is invariant under Legendrian isotopies. See \cite{ratsft}.
In \cite{Georgios} it is shown that under certain, not too restrictive, choices of almost complex structures of $T^*M$ and ${\mathbb R} \times J^1(M)$ we have that
$(\mathcal{A}(\Lambda), \partial_s;{\mathbb Z}_2) \simeq (\mathcal{A}(\Lambda), \partial_l;
{\mathbb Z}_2)$,
where the isomorphism is induced by the projection
\begin{equation*}
\pi_P: {\mathbb R} \times (T^*M \times {\mathbb R}) \to T^*M.
\end{equation*}
In particular, it is proven that the induced map
\begin{align*}
\pi_P: \mathcal{M}_{s, \Lambda}(a, {\bf b}) \to \mathcal{M}_{l,\Lambda}(a, {\bf b}),
\quad u \mapsto \pi_P(u),
\end{align*}
is a diffeomorphism. In the present paper we extend this result to ${\mathbb Z}$-coefficients,
by proving that there is a choice of orientation conventions so that
the coherent orientation scheme given for $\mathcal{M}_{l,\Lambda}$ in
\cite{orientbok} can be lifted under $\pi_P$ to give a coherent
orientation scheme for $\mathcal{M}_{s,\Lambda}$. Compare [\cite{Georgios}, Remark 2.4].
\begin{thm}\label{thm:georgios}
Let $J_P$ and $\tilde J_p$ be almost complex structures on $T^*M$ and ${\mathbb R} \times J^1(M)$, respectively, satisfying the assumptions in [\cite{Georgios}, Theorem 2.1]\footnote{The conditions are that $(D\pi_p)\tilde J_p = J_p(D\pi_p)$, and that $J_p$ is regular (the 0-dimensional moduli spaces $\mathcal{M}_{l,\Lambda}$ are transversely cut out) and integrable in neighborhoods of the double points of $\Pi_{\mathbb C}(\Lambda)$.}. Further assume that $\Lambda \subset J^1(M)$ is a spin Legendrian submanifold. Then there are choices of coherent orientations of the moduli spaces $\mathcal{M}_{s,\Lambda}(a, {\bf b})$ and $\mathcal{M}_{l,\Lambda}(a, {\bf b})$ so that
\begin{align*}
\pi_P: \mathcal{M}_{s, \Lambda}(a, {\bf b}) \to \mathcal{M}_{l,\Lambda}(a, {\bf b}),
\quad u \mapsto \pi_P(u),
\end{align*}
is orientation preserving. Moreover, for $i=s,l$ we have that
\begin{equation}\label{eq:differential}
\partial_i(a) = \sum_{\dim \mathcal{M}_{i,\Lambda}(a,{\bf b}) = 0} |\mathcal{M}_{i,\Lambda}(a, {\bf b})| {\bf b}
\end{equation}
satisfies
$\partial_i^2 =0$. Here $ |\mathcal{M}_{i,\Lambda}(a, {\bf b})| $ denotes the
algebraic number of disks in the moduli space, where the signs of the
disks come from the coherent orientation scheme.
\end{thm}
\begin{rmk}
We also get that the stable tame isomorphism class of the DGA:s is
invariant under Legendrian isotopies. Compare [\cite{orientbok},
Section 4.3].
\end{rmk}
\begin{rmk}\label{rmk:dumsign}
We will use slightly different orientation conventions than in
\cite{orientbok}, to simplify the expression of the differential. In
that paper it is instead of \eqref{eq:differential} given by
\begin{equation*}
\partial_l(a) = \sum_{\dim \mathcal{M}_{l,\Lambda}(a,{\bf b}) = 0} (-1)^{(n-1)(|a|+1)}|\mathcal{M}_{l,\Lambda}(a, {\bf b})| {\bf b}.
\end{equation*}
Compare with the discussion in Subsection \ref{sec:clock}.
\end{rmk}
\subsection{Exact Lagrangian cobordisms}\label{sec:elc}
Here we describe how an exact Lagrangian
submanifold $L \subset {\mathbb R} \times J^1(M)$ with cylindrical Legendrian
ends induces a morphism between the DGA:s of the ends.
\begin{defi}\label{def:exactcob}
Let $\Lambda_+, \Lambda_- \subset J^1(M)$ be Legendrian submanifolds.
An \emph{exact Lagrangian cobordism from $\Lambda_+$ to $\Lambda_-$} is an exact Lagrangian submanifold $L$ of the symplectization of $J^1(M)$, satisfying
\begin{align*}
& \mathcal{E}_+(L) := L \cap ((T,\infty) \times J^1(M)) = (T, \infty) \times \Lambda_+, \\
& \mathcal{E}_-(L) := L \cap ((-\infty, -T) \times J^1(M)) = (-\infty, -T) \times \Lambda_-,
\end{align*}
for some $T>0$, and so that
\begin{enumerate}
\item each function $f$ that satisfies $d f = e^t \alpha|_L$, also satisfies that $f|_{\mathcal{E}_\pm(L)}$ is constant,
\item $L \setminus (\mathcal{E}_+(L) \cup \mathcal{E}_-(L))$ is compact with boundary $\Lambda_+ - \Lambda_-$.
\end{enumerate}
\end{defi}
An exact Lagrangian cobordism $L$ induces a DGA-morphism
\begin{equation*}
\Phi_L:(\mathcal{A}(\Lambda_+), \partial_+; {\mathbb Z}_2) \to (\mathcal{A}(\Lambda_-), \partial_-; {\mathbb Z}_2),\end{equation*}
where $\partial_{\pm}$ denotes the differential $\partial_s$ associated to $\mathcal{A}(\Lambda_{\pm})$.
Indeed, we can define $\Phi_L$ by
\begin{equation}\label{eq:phidisk}
\Phi_L(a) =
\sum_{\dim \mathcal{M}_L (a, {\bf b})=0} |\mathcal{M}_L(a, {\bf b})| {\bf b}, \qquad {{\bf b} = b_1\dotsm b_m, }
\end{equation}
if $a$ is a generator, and extend it to the rest of the algebra by
\begin{align}
\label{eq:ex1} &\Phi(a +b ) = \Phi(a) + \Phi(b) \\
\label{eq:ex2} &\Phi(ab) = \Phi(a)\Phi(b).
\end{align}
See \cite{ratsft} and [\cite{tobhonda}, Section 3.5].
Here $\mathcal{M}_L(a,{\bf b})$ denotes the moduli space of punctured pseudo-holomorphic disks with boundary on $L$, positive puncture mapped asymptotically to a strip over the Reeb chord $a$ at $t = + \infty$, negative punctures mapped asymptotically to strips over the Reeb chords $b_1,\dotsc,b_m$ at $t=-\infty$, and $|\mathcal{M}_L(a, {\bf b})|$ is the modulo $2$ count of elements. See Section \ref{sec:holo}. Note that in [\cite{tobhonda}, Section 3.5] the results are only stated for $n=1$, but tracing the proofs one sees that they can be extended word-by-word to arbitrary $n$.
We prove that we can replace the modulo 2 count by a signed count, so that $\Phi_L$ gives a DGA-morphism also with ${\mathbb Z}$-coefficients.
\begin{thm}\label{thm:kalman}
Let $L\subset {\mathbb R} \times J^1(M)$ be a spin, exact Lagrangian cobordism from $\Lambda_+$ to $\Lambda_-$. Then there are choices of coherent orientations of the moduli spaces $\mathcal{M}_{L}(a,{\bf b})$, $\mathcal{M}_{s,\Lambda_+}(a,{\bf b})$ and $\mathcal{M}_{s,\Lambda_-}(a,{\bf b})$ so that
\begin{equation*}
\Phi_L:(\mathcal{A}(\Lambda_+), \partial_+; {\mathbb Z}) \to (\mathcal{A}(\Lambda_-), \partial_-; {\mathbb Z}) \end{equation*}
defined by \eqref{eq:phidisk} -- \eqref{eq:ex2} is a DGA-morphism. Now $|\mathcal{M}_L(a, {\bf b})|$ represents the algebraic count of disks in the moduli space.
\end{thm}
Moreover, $\Phi_L$ satisfies SFT-functorality. That is, let $L_1$, $L_2 \subset {\mathbb R} \times J^1(M)$ be two exact Lagrangian cobordisms such that $L_1$ goes from $\Lambda_0$ to $\Lambda_1$ and $L_2$ goes from $\Lambda_1$ to $\Lambda_2$. Then we can form the \emph{concatenation} $L_1 \# L_2$, by gluing the negative end of $L_1$ to the positive end of $L_2$, as explained in [\cite{tobhonda}, Section 1.2]. This gives an exact Lagrangian cobordism from $\Lambda_0$ to $\Lambda_2$, which satisfies
$ \Phi_{L_1\#L_2} = \Phi_{L_2} \circ \Phi_{L_1}$
as a DGA-morphism from $(\mathcal{A}(\Lambda_0), {\mathbb Z}_2)$ to $(\mathcal{A}(\Lambda_2), {\mathbb Z}_2)$. See [\cite{tobhonda}, Lemma 3.13].
We prove that the functorial properties of $\Phi$ continue to hold with integer coefficients.
\begin{thm}\label{thm:functorial}
Assume that $L_1,L_2 \subset {\mathbb R} \times J^1(M)$ are two spin, exact Lagrangian cobordisms with fixed spin structures. Assume that $L_1$ goes from $\Lambda_0$ to $\Lambda_1$ and that $L_2$ goes from $\Lambda_1$ to $\Lambda_2$. Then there are choices of coherent orientations of the moduli spaces $\mathcal{M}_{s,\Lambda_i}(a, {\bf b})$, $i=0,1,2$, $\mathcal{M}_{L_i}(a, {\bf b})$, $i=1,2$, and $\mathcal{M}_{L_1\#L_2}(a, {\bf b})$ so that
\begin{equation}
\label{eq:funct1}
\Phi_{L_2} \circ \Phi_{L_1} = \Phi_{L_1\#L_2}
\end{equation}
as DGA-morphisms from $(\mathcal{A}(\Lambda_0), {\mathbb Z})$ to $(\mathcal{A}(\Lambda_2), {\mathbb Z})$.
Moreover, if $\Lambda\subset J^1(M)$ is a spin Legendrian then there are choices of coherent orientations of the moduli spaces $\mathcal{M}_{s,\Lambda}(a, {\bf b})$, $\mathcal{M}_{{\mathbb R} \times \Lambda}(a, {\bf b})$ so that
\begin{equation}
\label{eq:funct2}
\Phi_{{\mathbb R} \times \Lambda} = \operatorname{id}.
\end{equation}
\end{thm}
We will prove that the orientation scheme from Theorem
\ref{thm:georgios} can be used to derive these results.
As indicated in the
Introduction, this orientation scheme is defined using capping
operators. Briefly, this works as follows.
Let $u \in \mathcal{M}_{l,\Lambda}(a, {\bf b})$. We have a linearized $\bar{\partial}$-operator $\bar{\partial}_u$ associated to $u$, defined on the punctured unit disk in ${\mathbb C}$ and with a trivialized Lagrangian boundary condition induced by the spin structure of $\Lambda$. This boundary condition is ``closed up'' by
gluing capping disks to the punctures of $u$. That is, for each Reeb
chord $c$ of $\Lambda$ we define two different capping operators
$\bar{\partial}_{c,+}$ and $\bar{\partial}_{c,-}$. These are $\bar{\partial}$-operators defined on
the unit disk in ${\mathbb C}$ with one puncture. We glue $\bar{\partial}_{a,+}$ to $\bar{\partial}_u$ at the positive
puncture of $u$, and $\bar{\partial}_{b_{i},-}$ to $\bar{\partial}_u$ at the negative puncture
corresponding to the chord $b_i$, $i=1,\dotsc , m$. We require the
trivialized boundary conditions for the capping operators to be
defined in such a way so that these gluings induce a trivialized Lagrangian boundary
condition for the non-punctured unit disk in ${\mathbb C}$. Then we use the
fact that there is a canonical orientation of the determinant line
bundle for the $\bar{\partial}$-operator over the space of trivialized
Lagrangian boundary conditions for the unit disk in ${\mathbb C}$. This
canonical orientation is given via evaluation at the boundary, see
[\cite{fooo}, Section 8], and the canonical orientation for the $\bar{\partial}$-operator
associated to the capped boundary condition induces
an orientation of the determinant line $\det \bar{\partial}_u$, which in turn induces an orientation of $T_u \mathcal{M}_{l,\Lambda}(a, {\bf b})$. This is explained in more detail in
Section \ref{sec:holo} and Section \ref{sec:oc}.
Notice that the signs occurring in the differential of the DGA of $\Lambda$ depend on the
choice of capping operators. We will prove that for certain systems
of capping operators, the associated DGA:s are isomorphic.
\begin{thm}
\label{thm:differentcappings}
Let $\Lambda$ be a spin Legendrian submanifold of $J^1(M)$ with a fixed spin structure. Let $\mathcal{S}$ denote a system of capping operators for $\Lambda$
satisfying \ref{c1} -- \ref{c3} in Section \ref{sec:orientcap}. Let $\partial_{l,\mathcal{S}}$ denote the induced differential as defined in \eqref{eq:differential}, where the orientation of the moduli space is induced by the system $\mathcal{S}$. Then $(\mathcal{A}(\Lambda), \partial_{l,\mathcal{S}};{\mathbb Z})$ is a DGA whose homology is invariant under Legendrian isotopies.
Moreover, if $\mathcal{S'}$ is another system of capping operators for $\Lambda$ satisfying \ref{c1} -- \ref{c3}, then there is a DGA-isomorphism
\begin{equation}
\label{eq:isocap}
\Phi_{\mathcal{S},\mathcal{S}'}:(\mathcal{A}(\Lambda), \partial_{l,\mathcal{S}};{\mathbb Z}) \to (\mathcal{A}(\Lambda), \partial_{l,\mathcal{S}'};{\mathbb Z}).
\end{equation}
\end{thm}
We refer to Section \ref{sec:cob-pf} for an explicit description of the map (\ref{eq:isocap}).
\begin{rmk}
From the proofs of Theorem \ref{thm:georgios}, Theorem \ref{thm:kalman} and Theorem \ref{thm:functorial} it follows that any system of capping operators satisfying \ref{c1} -- \ref{c3} gives coherent orientation schemes so that the statements of the theorems hold.
\end{rmk}
\begin{rmk}
Note that the capping operators defined in [\cite{orientbok}, Section 3.3], [\cite{orientbok}, Section 4.5] and [\cite{orienttrees}, Section 3.4] all satisfy \ref{c1} -- \ref{c3}. Compare with Remark \ref{rmk:notv}.
\end{rmk}
\begin{rmk}\label{rmk:oras}
All orientation schemes above depend on choices of orientations of ${\mathbb R}^n$ and of ${\mathbb C}$, which we from now on assume to be fixed.
\end{rmk}
\section{Punctured pseudo-holomorphic disks}\label{sec:holo}
In this section we give a definition of punctured pseudo-holomorphic disks. We also define the moduli spaces that will be relevant for us.
\subsection{Pseudo-holomorphic disks}
An \emph{almost complex structure} $J$ on a symplectic manifold $(X, \omega)$ is an endomorphism $J: TX \to TX$ satisfying $J^2 =-\operatorname{id}$. We say that $J$ is \emph{compatible with $\omega$} if $\omega(\cdot, J \cdot)$ defines a Riemannian metric on $X$. If $(X, \omega) = ({\mathbb R} \times J^1(M), d(e^t \alpha))$, then $J$ is \emph{cylindrical} if it is compatible with $\omega$, is invariant under $t$-translation, and satisfies $J(\xi)= \xi$, $J(\partial_t) = R_\alpha$. Here $R_\alpha$ denotes the Reeb vector field of $\alpha$.
Let $D$ be the compact unit disk in ${\mathbb C}$ and let $D_{m+1}$ denote the disk with $m+1$ marked points $p_0,\dotsc,p_m \in \partial D$, cyclically ordered along the boundary in the counter-clockwise direction. Let $\dot D_{m+1}$ denote the corresponding punctured disk with the marked points removed.
We will assume that $p_0=1 \in {\mathbb C}$, and call it the \emph{positive puncture}. We say that $p_1,\dotsc,p_m$ are the \emph{negative punctures}.
A map $u:D_{m+1} \to X$ (or $u:\dot D_{m+1} \to X$ ) is \emph{J-holomorphic} if it satisfies
\begin{equation*}
\bar{\partial}_J(u) := du + J \circ du \circ i =0.
\end{equation*}
If we want to neglect the choice of $J$ we say that $u$ is \emph{pseudo-holomorphic}.
\subsection{Gradings}
Each Reeb chord $a$ of $\Lambda$ comes equipped with a grading $|a|$, given by
\begin{equation*}
|a| = CZ(a) - 1
\end{equation*}
where $CZ(a)$ is
the \emph{Conley-Zehnder index} of $a$. Since we will not perform any explicit calculations of the gradings in this paper we refer to [\cite{pr}, Section 2.2] for a proper definition.
\subsection{Moduli spaces}
In this section we give definitions of the relevant moduli spaces of pseudo-holomorphic disks.
\subsubsection{Moduli spaces in the Lagrangian projection}
Fix an almost complex structure $J$ on $T^*M$, compatible with $\omega$. We let $\mathcal{M}_{l,\Lambda}(a, \bf b)$, ${\bf b}=b_1\dotsm b_m$, denote the moduli space of pseudo-holomorphic maps $u:(D_{m+1}, \partial D_{m+1}) \to (T^*M, \Pi_{\mathbb C}(\Lambda))$ satisfying the following:
\begin{enumerate}
\item $u|_{\partial \dot D_{m+1}}$ has a continuous lift $\tilde u$ to $\Lambda$;
\item $u(p_0)= \Pi_{\mathbb C}(a)$, where $a$ is a Reeb chord of $\Lambda$, and the $z$-coordinate of $\tilde u$ makes a positive jump when passing through $p_0$ in the counterclockwise direction;
\item $u(p_i)= \Pi_{\mathbb C}(b_i)$, $i=1,\dotsc,m$, where $b_i$ is a Reeb chord of $\Lambda$, and the $z$-coordinate of $\tilde u$ makes a negative jump when passing through $p_i$ in the counterclockwise direction.
\end{enumerate}
Moreover, we consider two maps $u_1$, $u_2$ satisfying the above to be equal if they differ by a biholomorphism of $D_{m+1}$.
In \cite{legsub} it is proven that for generic $J$ the moduli spaces are transversely cut out manifolds
of dimension
\begin{equation}\label{eq:dim1}
\dim\mathcal{M}_{l,\Lambda}(a,{\bf b}) = |a| - \sum_{i=1}^m |b_i| -1.
\end{equation}
\subsubsection{Moduli spaces in the symplectization}
Fix a cylindrical almost complex structure $J$ on ${\mathbb R} \times J^1(M)$. We let $\hat\mathcal{M}_{s,\Lambda}(a, \bf b)$ denote the moduli space of pseudo-holomorphic maps $u:(\dot D_{m+1}, \partial \dot D_{m+1}) \to ({\mathbb R} \times J^1(M), {\mathbb R} \times \Lambda)$ satisfying the following:
\begin{enumerate}
\labitem{{(s1)}}{s1} in a neighborhood of the positive puncture $p_0$ the map $u$ is asymptotic to the Reeb chord strip $[0,\infty) \times c$;
\labitem{{(s2)}}{s2} in a neighborhood of the negative puncture $p_i$ the map $u$ is asymptotic to the Reeb chord strip $(-\infty,0] \times b_i$ $i=1,\dotsc,m$.
\end{enumerate}
Again, we consider two maps $u_1$, $u_2$ satisfying the above to be equal if they differ by a biholomorphism of $D_{m+1}$.
We let
$\mathcal{M}_{s,\Lambda}(a, {\bf b}) = \hat \mathcal{M}_{s,\Lambda}(a, \bf b)/{\mathbb R}$
where the ${\mathbb R}$-action is given by translation in the $t$-direction
For generic $J$ the moduli spaces are transversely cut out
manifolds of dimension
\begin{equation*
\dim\mathcal{M}_{s,\Lambda}(a,{\bf b}) = |a| - \sum_{i=1}^m |b_i|-1.
\end{equation*}
See [\cite{Georgios}, Section 4.2.4].
\subsubsection{Moduli spaces associated to an exact Lagrangian cobordism}
Fix a compatible almost complex structure $J$ on ${\mathbb R} \times J^1(M)$, and assume that it is cylindrical for $|t|>N$ for some $N$. We let $\mathcal{M}_{L}(a, \bf b)$ denote the moduli space of pseudo-holomorphic maps $u:(\dot D_{m+1}, \partial \dot D_{m+1}) \to ({\mathbb R} \times J^1(M), L)$ satisfying \ref{s1} and \ref{s2},
and again we consider two maps $u_1$, $u_2$ to be equal if they differ by a biholomorphism of $D_{m+1}$.
For generic $J$ the moduli spaces are transversely cut out manifolds of dimension
\begin{equation}\label{eq:dim2}
\dim\mathcal{M}_{L}(a,{\bf b}) = |a| - \sum_{i=1}^m |b_i|.
\end{equation}
See [\cite{tobhonda}, Lemma 3.7].
From now on we assume that the almost complex structures are chosen so that the relevant moduli spaces are transversely cut out manifolds of the expected dimension.
We call a disk $u \in \mathcal{M}_{i,\Lambda}(a,{\bf b})$, $i=l,s$, a \emph{pseudo-holomorphic disk of $\Lambda$ with positive puncture $a$ and negative punctures $b_1,\dotsc,b_m$}. If moreover $\dim \mathcal{M}_{i,\Lambda}(a,{\bf b}) =0$ we say that $u$ is \emph{rigid}. We use similar language for disks $u \in \mathcal{M}_{L}(a,{\bf b})$.
\subsection{The linearized \texorpdfstring{$\bar{\partial}$}{dbar}-operator}
The algebraic count of elements in \eqref{eq:differential} and \eqref{eq:phidisk} are defined by associating a sign to each rigid pseudo-holomorphic disk. This assignment of signs can be understood as an orientation of the moduli spaces, and since these spaces are assumed to be zero-dimensional they are always orientable. However, since we require that $\partial^2= 0$ and $\partial \circ \Phi_L = \Phi_L \circ \partial$, we need to choose the orientations in a coherent way. This is done by considering linearized $\bar{\partial}$-operators associated to the pseudo-holomorphic disks.
Let the Sobolev space $\mathcal{H}_k({D}_{m+1}, {\mathbb C}_n)$ be the closure of $\mathcal{C}_0^\infty (\dot{D}_{m+1}, {\mathbb C}^n )$ equipped with the
standard Sobolev $\|\cdot \|_{k,2}$-norm. That is, $\mathcal{H}_k({D}_{m+1}, {\mathbb C}_n)$ consists of all elements in $L^2({D}_{m+1}, {\mathbb C}_n)$ whose weak derivatives exist and belong to $L^2$, up to order $k$.
Choose local coordinates on $\dot D_{m+1}$ in a neighborhood of the puncture $p_j$, given by a half-infinite strip
$E_{p_j} = {(\tau, t) \in [0,\infty) \times [0, 1]},
(\tau, t) = \tau + it$. For each puncture $q_i \in D_{m+1}$ we define a \emph{weight vector}
\begin{equation*}
\nu_i = (\nu_i^1,\dotsc,\nu_i^n) \in (-\pi/2, \pi/2)^n
\end{equation*}
and let $\nu= (\nu_0,\dotsc,\nu_m)$. Let $w_{\nu}: D_{m+1} \to GL(n)$ be a smooth function satisfying
\begin{align*}
w_{\nu}(\tau,t) =
\operatorname{Diag}(e^{\nu_i^1 |\tau|},\dotsc,e^{\nu_i^n |\tau|})
\end{align*}
in $E_{p_i}(M)$, and assume that $w_{\nu}$ is close to the identity
matrix in compact regions of the disk.
Let the \emph{weighted Sobolev space $\mathcal{H}_{k,\nu}(D_{m+1},{\mathbb C}^n)$} be defined by
\begin{equation*}
\mathcal{H}_{k,\nu}(D_{m+1},{\mathbb C}^n) = \{f \in \mathcal{H}_k^{loc}(D_{m+1},{\mathbb C}^n); w_{\nu}f \in \mathcal{H}_k(D_{m+1},{\mathbb C}^n)\}
\end{equation*}
with norm
\begin{equation*}
\|f\|_{k,\nu} = \|w_{\nu}f\|_{k,2}.
\end{equation*}
If $u: (D_{m+1}, \partial D_{m+1}) \to (T^*M, \Pi_{\mathbb C}(\Lambda))$ is a pseudo-holomorphic disk, then $u^* T\Pi_{\mathbb C}(\Lambda)$ induces a Lagrangian boundary condition on $D_{m+1}$. Pick a complex trivialization of $u^*TT^*M$. Using that $\Lambda$ is spin, we get a well-defined trivialization of the Lagrangian boundary condition, following [\cite{orientbok}, Section 3.4.2] and [\cite{pr}, Section 4.4]. This gives a collection of maps
$A = (A_0,\dotsc,A_{m+1}): \partial D_{m+1} \to U(n),
$
where
\begin{equation*}
A_i : [p_i, p_{i+1}] \to U(n), \qquad i=0,\dotsc,m+1, \quad m+2 =0.
\end{equation*}
Let
\begin{equation*}
\mathcal{H}_{2,\nu}[A](D_{m+1}, u^*TT^*M)
\end{equation*}
denote the closed subspace of $\mathcal{H}_{2,\nu}(D_{m+1}, u^*TT^*M)$, consisting of elements $s$ that satisfy the linearized Lagrangian boundary condition $A$ along $\partial D_{m+1}$, and which satisfy
$\bar{\partial}_u s |_{\partial D_{m+1}} = 0$.
Similarly, let
\begin{equation*}
\mathcal{H}_{1,\nu}[0](D_{m+1}, T^{*0,1}D_{m+1} \otimes u^*TT^*M) \end{equation*}
be the closed subspace of $\mathcal{H}_{1,\nu}(D_{m+1}, T^{*0,1}D_{m+1} \otimes u^*TT^*M)$ consisting of elements $s$ satisfying $ s |_{\partial D_{m+1}} = 0$.
From this we get an associated linearized $\bar{\partial}$-operator
\begin{equation*}
\bar{\partial}_A = \bar{\partial}_{l,A}: \mathcal{H}_{2,\nu}[A](D_{m+1}, u^*TT^*M) \to \mathcal{H}_{1,\nu}[0](D_{m+1}, T^{*0,1}D_{m+1} \otimes u^*TT^*M).
\end{equation*}
The boundary condition $A$ lifts to a boundary condition $\operatorname{id} \oplus A$ under $\pi_P$, and gives rise to a similar operator
\begin{multline*}
\bar{\partial}_{s,A}:\mathcal{H}_{2,\nu}[\operatorname{id} \oplus A](D_{m+1},\tilde u^*T({\mathbb R} \times J^1(M))) \to\\
\mathcal{H}_{1,\nu}[0](D_{m+1},T^{*0,1}D_{m+1} \otimes \tilde u^*T({\mathbb R} \times J^1(M))),
\end{multline*}
where $\tilde u$ is the lift of $u$ under $\pi_P$. We extend the weight function to $GL(1+n)$ by redefining the weight vector at puncture $q_i$ to be given by
\begin{equation*}
\nu_i = (-\epsilon,\nu_i^1,\dotsc,\nu_i^n),
\end{equation*}
for some $\epsilon > 0$ small, $i= 0, \dotsc , m$.
\begin{thm}[\cite{pr}, Lemma 4.3 and Lemma 4.5;\cite{Georgios}, Lemma 8.2]\label{thm:fredholm}
There is a choice of weight vectors so that the operators $\bar{\partial}_{l,A}$, $\bar{\partial}_{s,A}$ are Fredholm, and so that for a generic choice of almost complex structures these operators are surjective after having stabilized their domains with the space of conformal variations from Section \ref{sec:confvar}.
\end{thm}
Similar constructions are done for the linearized $\bar{\partial}$-operator at a holomorphic disk $u \in \mathcal{M}_L(a, \bf b)$, and we get the same results about Fredholmness and surjectivity.
\begin{rmk}
Sometimes we write $\bar{\partial}_u$ instead of $\bar{\partial}_{l,A}$, $\bar{\partial}_{s,A}$, to simplify notation.
\end{rmk}
All this is related to orientations of moduli spaces in the following way.
If $u$ is a pseudo-holomorphic disk of $\Lambda$ (or of $L$), it follows from Theorem \ref{thm:fredholm} that $\bar{\partial}_u$ is Fredholm. That is, it has finite-dimensional kernel and cokernel. This means that we can consider its determinant line $\det \bar{\partial}_u$,
\begin{equation*}
\det \bar{\partial}_u = \bigwedge^{\max} \operatorname{Ker} \bar{\partial}_u \otimes \bigwedge^{\max} (\operatorname{Coker} \bar{\partial}_u)^*,
\end{equation*}
where $\bigwedge^{\max}V$ is the top exterior power of the vector space $V$.
In particular this means that we can give an orientation to $\det \bar{\partial}_u$. This orientation will in turn be related to the orientation of $T_u \mathcal{M}_{i,\Lambda}$ (or $T_u \mathcal{M}_L$), as we will explain in Section \ref{sec:confvar}.
\section{Orientation conventions}\label{sec:oc}
The signs in the algebraic count of elements in the DGA-morphisms, and also in the DGA-differentials, come from orientations of the moduli spaces of $J$-holomorphic disks. These orientations depend on several choices, which we fix in this section.
We mainly use the approach of \cite{orientbok} where the moduli spaces $\mathcal{M}_{l,\Lambda}$ are oriented, but we will make some slight modifications of these conventions to make them fit into the symplectization setting.
We close this section by proving that the chosen conventions imply the statements in Theorem \ref{thm:georgios}.
\subsection{Short exact sequences}
First of all, it is a standard fact that an exact sequence
\begin{equation}\label{eq:es}
0 \xrightarrow{} V_1 \xrightarrow{\alpha} W_1\xrightarrow{\beta} W_2
\xrightarrow{\gamma} V_2 \xrightarrow{} 0
\end{equation}
of finite-dimensional vector spaces induces an isomorphism
\begin{equation}\label{eq:phi}
\phi:\bigwedge^{\max}V_1 \otimes \bigwedge^{\max}V^*_2 \xrightarrow{\approx}
\bigwedge^{\max}W_1 \otimes \bigwedge^{\max}W^*_2.
\end{equation}
See e.g. [\cite{Floer-Hofer}, Appendix].
This isomorphism is not canonical, but depends on choices. For a deeper discussion on this, see \cite{z}.
We will use the following convention, described in terms of oriented bases:
First we identify $\bigwedge^{\max}V^*$ with $\bigwedge^{\max}V$ via
$v_1\wedge \dotsb \wedge v_k \mapsto v_1^* \wedge \dotsb \wedge v_k^*$,
where $(v_1,\dotsc,v_k)$ is any basis for $V$, and $v_i^*$ is the vector dual to $v_i$. Now pick a basis $(v_1,\dotsc,v_k)$ for $V_1$, and vectors $(u_1,\dotsc,u_l)\in W_2$ so that $(\gamma(u_1),\dotsc,\gamma(u_l))$ gives a basis for $V_2$. Then pick vectors $(w_1,..,w_m) \in W_1$ so that $(\alpha(v_1),\dotsc,\alpha(v_k), w_1,\dotsc,w_m)$ gives a basis for $W_1$. From the exactness of the sequence \eqref{eq:es} it then follows that $(\beta(w_1),\dotsc,\beta(w_m),u_1,\dotsc,u_l)$ gives a basis for $W_2$. We fix the isomorphism \eqref{eq:phi} to be given by
\begin{multline}\label{eq:o2}
v_1\wedge\dotsb \wedge v_k \otimes \gamma(u_1)\wedge\dotsb\wedge \gamma(u_l) \mapsto\\
v_1\wedge\dotsb \wedge v_k \wedge w_1 \wedge \dotsb \wedge w_m \otimes u_1\wedge\dotsb\wedge u_l \wedge \beta(w_1)\wedge \dotsb \wedge \beta(w_m),
\end{multline}
and extend by linearity. It is straightforward to check that this definition does not depend on the choice of oriented bases.
\begin{rmk}\label{rmk:kan}
This convention is slightly different than the one in [\cite{orientbok}, Section 3.2.1].
As a consequence of this choice, we get rid of the sign $(-1)^{\frac{1}{2}(n-1)(n-2)}$ in the statement of [\cite{orientbok}, Lemma 3.11].
\end{rmk}
\subsection{Exact gluing sequences, and order of gluing}\label{sec:gluing}
We will repeatedly make use of exact gluing sequences of pseudo-holomorphic disks. For a detailed description we refer to [\cite{orientbok}, Section 3.2]. Here we give an outline of the construction.
Let $D_{m_1+1}$ be a disk with punctures $(q_0,q_1,\dotsc,q_{m_1})$ and with an associated Lagrangian boundary condition $A: \partial D_{m_1+1} \to U(n)$. Similarly, let $D_{m_2+1}$ be a disk with punctures $(p_0,p_1,\dotsc,p_{m_2})$ and with an associated Lagrangian boundary condition $B: \partial D_{m_2+1} \to U(n)$. If $A$ and $B$ are asymptotically equal to the same constant map at the punctures $q_0$ and $p_k$, say, then we can glue $D_{m_1+1}$ to $D_{m_2+1}$ at $q_0 =p_k$, and get a trivialized Lagrangian boundary condition $A\#B$ on the glued disk $D_{m_1+m_2} = D_{m_1+1}\#D_{m_2+1}$.
This gluing induces an exact sequence for the kernels and cokernels of the associated operators $\bar{\partial}_A$, $\bar{\partial}_B$ and $\bar{\partial}_{A\#B}$, given by
\begin{equation}\label{eq:gluconv}
0 \to \operatorname{Ker}\dbar_{A\#B}
\xrightarrow{\alpha}
\begin{bmatrix}
\operatorname{Ker}\dbar_B \\
\operatorname{Ker}\dbar_A
\end{bmatrix}
\xrightarrow{\beta}
\begin{bmatrix}
\operatorname{Coker}\dbar_B \\
\operatorname{Coker}\dbar_A
\end{bmatrix}
\xrightarrow{\gamma}
\operatorname{Coker}\dbar_{A\#B} \to 0.
\end{equation}
Here we use the notation
\begin{equation*}
\begin{bmatrix}
V\\
W
\end{bmatrix} = V \oplus W.
\end{equation*}
\begin{rmk}\label{rmk:glumap}
The maps $\alpha$, $\beta$ and $\gamma$ are given as follows. First embed the kernels and cokernels of the $\bar{\partial}_A$ and $\bar{\partial}_B$-operators in the Sobolev spaces associated to $\bar{\partial}_{A \#B}$, by cutting off the elements with cut-off functions $\phi_A^\rho, \phi_B^\rho$.
Then $\alpha$ is $L^2$-projection
onto the space spanned by the cut-off kernel elements, $\beta$ is
$\bar{\partial}$ composed with $L^2$-projection and $\gamma$ is projection
to the quotient. See [\cite{orientbok}, Section 3.2.2] for a more detailed description.
\end{rmk}
Using the isomorphism \eqref{eq:phi} we see that orientations of $\det \bar{\partial}_A$ and $\det \bar{\partial}_B$ induce an orientation of $\det \bar{\partial}_{A\#B}$. Note that this induced orientation depends on the pairwise order of the vector spaces in the second and third column of the gluing sequence \eqref{eq:gluconv}, and that we have chosen the opposite order compared to [\cite{orientbok}, Section 3.2.2]. The reason for this change is that the order in \eqref{eq:gluconv} seems more feasible when working with an extra ${\mathbb R}$-direction, which shows up when we consider pseudo-holomorphic disks in the symplectization instead of in the Lagrangian projection. Compare with the discussion in Subsection \ref{sec:clock}.
\subsection{Orientations of the space of conformal variations}\label{sec:confvar}
Let $u:D_{m+1} \to X$, $X = T^*M$ or $X= {\mathbb R} \times J^1(M)$, be a rigid $J$-holomorphic disk of $\Lambda$ or of $L$. If $m>1$, then the linearized $\bar{\partial}$-operator at $u$, restricted to the Sobolev space of candidate maps, will have cokernel isomorphic to the tangent space of the space of conformal structures of $D_{m+1}$. We call this tangent space the \emph{space of conformal variations}, and the orientation (i.e.\ the sign) of $u$ will depend on which orientation we choose on this space. See [\cite{orientbok}, Lemma 3.17]. We fix this orientation as follows.
Let $\mathcal{C}_m$ denote the space of conformal structures on $D_{m+1}$. If we fix the positions of three of the punctures of $D_{m+1}$, then the position of the other punctures parameterize $\mathcal{C}_{m}$.
To describe the orientation of the tangent space $T_{\kappa} \mathcal{C}_m$ at a conformal structure $\kappa$, let $\partial_{p_j}$ denote the vector tangent to $\partial D_{m+1}$ at $p_j$, pointing in the counterclockwise direction. Then if we choose $m-2$ of the vectors $\partial_{p_0},\dotsc,\partial_{p_m}$ we get a basis for $T_{\kappa} \mathcal{C}_m$. We define the positive orientation of $T_{\kappa} \mathcal{C}_m$ to be given by
\begin{equation}\label{eq:confor}
(\partial_{p_m}, \dotsc, \partial_{p_3}).
\end{equation}
This somewhat unnatural choice of orientation is a consequence of the convention \eqref{eq:gluconv}. Compare with the discussion in Subsection \ref{sec:clock}.
\begin{rmk}\label{rmk:ob}
This gives the same orientation as the oriented basis
\begin{equation*}
(\partial_{p_m},\dotsc,\partial_{p_{k+1}},-\partial_{p_{k-1}},\dotsc,-\partial_{p_{j+1}}, \partial_{p_{j-1}},\dotsc,\partial_{p_1}).
\end{equation*}
\end{rmk}
\begin{rmk}
If $m \leq 1$ then we can add marked points to the boundary of $D_{m+1}$ to get the setting above. See [\cite{orientbok}, Section 4.2.3].
\end{rmk}
To see how the orientation of the space of conformal variations relates to the sign of a rigid pseudo-holomorphic disk, we consider the \emph{fully linearized} $\bar{\partial}$-operator $d\Gamma_u$ at a $J$-holomorphic disk $u$. Here
\begin{equation*}
d \Gamma_u: \mathcal{H}_{2,\nu}[A] \oplus T\mathcal{C}_{m} \to \mathcal{H}_{1,\nu}[0], \qquad d\Gamma_u(v,w) = \bar{\partial}_u(v) + \Psi(w),
\end{equation*}
where $\Psi: T\mathcal{C}_{m} \to \mathcal{H}_{1,\nu}[0]$ is a linear map which we will not specify in detail. The regularity assumptions on the almost complex structure $J$ implies that $d\Gamma_u$ is surjective and that the tangent space of the moduli space $\mathcal{M}$ to which $u$ belongs can be identified with the kernel of $d\Gamma_u$,
\begin{equation*}
T_u\mathcal{M} \simeq \operatorname{Ker} d\Gamma_u.
\end{equation*}
We see that an orientation of $\operatorname{Ker} d\Gamma_u$ induces an orientation of $T_u\mathcal{M}$, and in particular, if $u$ is rigid so that $\mathcal{M}$ is zero-dimensional this will just be a sign assigned to $\mathcal{M}$ at $u$.
By the proof of [\cite{orientbok}, Lemma 3.17] we have that
\begin{equation*}
\bigwedge^{\max}\operatorname{Ker} d\Gamma_u \simeq \det \bar{\partial}_u \otimes \bigwedge^{\max}T\mathcal{C}_{m}.
\end{equation*}
Thus an orientation of $\det \bar{\partial}_u$ and of $T\mathcal{C}_{m}$ induces an orientation of $\operatorname{Ker} d\Gamma_u$. In particular, if $u$ is rigid then $\operatorname{Ker} d\Gamma_u = 0 = \operatorname{Ker}\dbar_u$ (assuming $m>1$) and $\det \bar{\partial}_u \otimes \bigwedge^{\max}T\mathcal{C}_{m}$ is given by the sign of the isomorphism
\begin{equation}\label{eq:psibar}
\bar{\Psi}:T\mathcal{C}_{m} \to \operatorname{Coker}\dbar_u.
\end{equation}
Here $\bar \Psi$ is given by $\Psi$ composed with the projection to the cokernel of $\bar{\partial}_u$. In Section \ref{sec:orientdisk} we define an orientation of $\det \bar{\partial}_u$ .
Now recall the gluing of $\bar{\partial}_A$ and $\bar{\partial}_B$ described in Section \ref{sec:gluing}. The direct sum of the conformal structures of the disks $D_{m_1+1}$ and $D_{m_2+1}$ (which were joined at $q_0 = p_k$, with $q_0$ denoting the positive puncture of $D_{m_1+1}$) can be seen as an element of the boundary of the space $\mathcal{C}_{m}$, $m=m_1 + m_2-1$. In addition, the outward normal at this conformal structure can be given by $\partial_{q_1}=-\partial_{p_{k-1}}$, or alternatively $\partial_{p_{k+1}} = - \partial_{q_{m_1}}$. We orient the boundary by outward normal last.
\begin{lma}\label{lma:confvar}
We have
\begin{equation*}
T\mathcal{C}_{m_2} \oplus T \mathcal{C}_{m_1} \oplus {\mathbb R} = (-1)^{(m_1-1)k+1}T\mathcal{C}_m
\end{equation*}
as oriented vector spaces, where ${\mathbb R}$ is given the orientation from the outward normal.
\end{lma}
\begin{proof}
This is similar to the proof of [\cite{orientbok}, Lemma 4.7].
\end{proof}
\subsection{Canonical orientation of the closed disk, trivializations, and spin structures}\label{sec:canon}
The determinant line bundle of the $\bar{\partial}$-operator over the space of
trivialized Lagrangian boundary conditions on the non-punctured unit disk in
${\mathbb C}$ is orientable. Moreover, if we fix an orientation of ${\mathbb R}^n$ and
of ${\mathbb C}$, then
this induces an orientation, via evaluation at the boundary. See
[\cite{fooo}, Proposition 8.1.4]. We denote this induced orientation the \emph{canonical
orientation} (recall that we assume that we have fixed orientations
of ${\mathbb C}$ and ${\mathbb R}^n$ already, see Remark \ref{rmk:oras}) .
The following proposition follows from [\cite{orientbok}, Section 3.4.2], [\cite{pr}, Section 4.4] and [\cite{fooo}, Section 8.1].
\begin{prp}
If $\Lambda$ (or $L$) is spin, then a choice of spin structure induces a trivialized Lagrangian boundary condition of $u \in \mathcal{M}_{l,\Lambda}$ (or $u \in \mathcal{M}_L$), which is well-defined up to homotopy.
\end{prp}
In this paper we use the following conventions. If $\Lambda$ is a
spin Legendrian and $L$ is the Lagrangian cylinder ${\mathbb R} \times
\Lambda$, then we give $L$ the spin structure induced from the spin
structure of $\Lambda$ and the trivial spin structure on ${\mathbb R}$. If $L$
is a spin, exact Lagrangian cobordism with cylindrical ends
$\Lambda_\pm$, then we require that $\Lambda_\pm$ are given the
boundary spin structures induced by $L$. We refer to
[\cite{orientbok}, Section 4.4] for a discussion on how other choices of spin structure affect the orientations of the moduli space of pseudo-holomorphic disks.
From [\cite{orientbok}, Lemma 3.11] together with our orientation convention \eqref{eq:o2} we get the following useful result.
\begin{lma}\label{lma:kan}
Let $\bar{\partial}_A$, $\bar{\partial}_B$ be two problems defined on the non-punctured unit disk, where $A$ and $B$ are trivialized Lagrangian boundary conditions. Let $\bar{\partial}_{A\#B}$ denote the problem induced by gluing $\bar{\partial}_A$ to $\bar{\partial}_B$. If $\det \bar{\partial}_A$ and $\det \bar{\partial}_B$ are given their canonical orientation, and ${\mathbb R}^n$ is given its fixed orientation, then the gluing sequence
\begin{equation}\label{eq:gluconv2}
0 \to \operatorname{Ker}\dbar_{A\#B}
\xrightarrow{\alpha}
\begin{bmatrix}
\operatorname{Ker}\dbar_B \\
\operatorname{Ker}\dbar_A
\end{bmatrix}
\xrightarrow{\beta}
\begin{bmatrix}
\operatorname{Coker}\dbar_B \\
{\mathbb R}^n \\
\operatorname{Coker}\dbar_A
\end{bmatrix}
\xrightarrow{\gamma}
\operatorname{Coker}\dbar_{A\#B} \to 0
\end{equation}
induces the canonical orientation on $\det \bar{\partial}_{A \# B}$. Here the ${\mathbb R}^n$-summand comes from gluing non-punctured disks, compare [\cite{orientbok}, Lemma 3.1].
\end{lma}
\subsection{Capping operators}\label{sec:orientcap}
Let $u$ be a holomorphic disk of $\Lambda$ or of $L$. As pointed out above, to give a sign to $u$ is related to give an orientation to the determinant line of $\bar{\partial}_u$. All this must be done in a coherent way, so that
we get
$\partial^2 =0$ and $\Phi_L \circ \partial = \partial \circ \Phi_L$ in the very end.
The idea from \cite{orientbok} is to use the trivialized Lagrangian boundary condition of $u$, induced by the spin structure of $\Lambda$ or of $L$, together with the canonical orientation of $\det \bar{\partial}$ over the space of trivialized Lagrangian boundary conditions on the non-punctured disk. To make this work, we need to choose a way to close up the trivialized boundary conditions of $u$ at the punctures. In \cite{orientbok} this is done by using something called \emph{capping operators}, and this is the method that we will use. We give an outline of the constructions, and also explain the modifications needed to carry it over to the symplectization.
\subsubsection{Capping trivializations}
The capping operators are $\bar{\partial}$-operators defined on the
$1$-punctured unit disk in ${\mathbb C}$, and we have two operators,
$\bar{\partial}_{p,+}$ and $\bar{\partial}_{p,-}$, associated to each Reeb chord $p$ of
$\Lambda$. The reason for this is that we need one capping operator
for $p$ in the case when $p$ occurs as a positive puncture of a disk,
and another capping operator for $p$ when $p$ occurs as a negative puncture.
To each capping operator $\bar{\partial}_{p,\pm}$ we have an associated trivialized Lagrangian boundary condition $R_{p, \pm}$, which is chosen in a way so that we get a trivialized boundary condition on the non-punctured disk after having glued all the capping operators corresponding to the punctures of $u$ to $\bar{\partial}_u$. We call the boundary conditions $R_{p, \pm}$ the \emph{capping trivializations}.
There are different possibilities to define $R_{p,\pm}$. See for example [\cite{orientbok}, Section 3.3], [\cite{orientbok}, Section 4.5], and [\cite{orienttrees}, Section 3.4.C]. We will not fix a specific system of capping trivializations in the present paper, instead we consider any system that satisfies certain conditions, listed below. In particular, the systems defined in \cite{orientbok} and in \cite{orienttrees} satisfy these conditions. Before stating the conditions, we first have to discuss a stabilization of the tangent bundle of $\Lambda$ and of $L$, made by adding a trivial bundle.
\subsubsection{Auxiliary directions} In [\cite{orientbok}, Section 3.3.3], something called \emph{auxiliary directions} are introduced. These are artificial extra directions that are added to the capping trivializations and to the Lagrangian trivializations induced by the pseudo-holomorphic disks. The main reason for doing this is to get the invariance proof of Legendrian contact homology over ${\mathbb Z}$ to work out well. These extra directions also simplify the work of assuring that we get a trivialized boundary condition on the non-punctured disk when we glue the capping operators to $\bar{\partial}_u$.
In the case when we are considering Legendrian knots $\Lambda \subset
{\mathbb R}^3$ (i.e.\ when $n=1$) we add one auxiliary direction, to get the stabilized tangent space $\tilde T\Pi_{\mathbb C}(\Lambda) = T \Pi_{\mathbb C}(\Lambda) \oplus {\mathbb R}$.
In the more general setting when $n\geq 2$ we add
two auxiliary directions, to get the stabilized tangent space $\tilde T\Pi_{\mathbb C}(\Lambda) = T \Pi_{\mathbb C}(\Lambda) \oplus {\mathbb R}^2$.
In the case of an exact cobordism $L$ we do the similar thing, so that we get a stabilized tangent space $\tilde TL = T L \oplus {\mathbb R}^i$, where $i = 1$ if $n=1$ and $i=2$ if $n \geq 2$.
We will
in what follows use $d_A$ for the dimension of the auxiliary space
added. That is, if $n=1$ then $d_A=1$, and if $n>1$ then $d_A=2$.
\begin{rmk}
The reason of adding only one auxiliary direction for $n=1$ is to get compatibility with [\cite{orientbok}, Section 4.5]. We can as well consider the case of adding a $2$-dimensional auxiliary space for all possible $n$. This is easily seen by tracing the proofs in Subsection \ref{sec:pfger} and Section \ref{sec:cob-pf}.
\end{rmk}
If $u$ is a pseudo-holomorphic disk of $\Lambda$ or of $L$, then the
linearized $\bar{\partial}_u$-problem is extended to the auxiliary directions
so that it gives an isomorphism here. See [\cite{orientbok}, Section 3.3.3]. Thus we get a canonical isomorphism between the determinant line of the original $\bar{\partial}_u$-problem and the extended one. With abuse of notation, we let $\bar{\partial}_u$ denote the extended problem from now on.
\subsubsection{System of capping operators for disks in $T^*M$.}
The capping operators $\bar{\partial}_{p,\pm}$ are also extended to the
auxiliary directions, but will in general not give isomorphisms in
these directions. To describe the properties that we require the
capping operators to have, recall that we assume that we have fixed a
trivialization of the Lagrangian boundary conditions of $\bar{\partial}_u$ (now
also extended to the auxiliary directions, using the trivial spin
structure here. See [\cite{orientbok}, Section 3.4.2]).
If $p$ is a puncture of $u$, then notice that we have two
Lagrangian subspaces associated to $u$ at $p$, given by the two
stabilized tangent spaces of $\Pi_{\mathbb C}(\Lambda)$ at $p$. From the fixed
trivialization we then get oriented frames for these two spaces. Let
$p_+, p_-$ denote the endpoints of the Reeb chord of $\Lambda$
corresponding to $p$, where $p_+$ corresponds to the end with largest
$z$-coordinate. Let $X_\pm$ denote the oriented frame of the
stabilized tangent space of $\Pi_{\mathbb C}(\Lambda)$ at $p$ that lifts to $T_{p\pm}\Lambda$.
We define a \emph{system of capping operators for $\Lambda$} to be
a set $\mathcal{S}$ consisting of $\bar{\partial}$-operators defined on the
one-punctured unit disk in ${\mathbb C}$, such that for each Reeb chord $p$
of $\Lambda$ we have a pair of $\bar{\partial}$-operators $\bar{\partial}_{p,+}$,
$\bar{\partial}_{p,-} \in \mathcal{S}$ with associated trivialized boundary
conditions $R_{p,\pm}: \partial D_1 \to U(n+d_A)$. Moreover, as a
part of the data of $\mathcal{S}$ we choose an orientation of $\det \bar{\partial}_{p,-}$ for each Reeb chord $p$.
We say that the system is \emph{admissible} if the operators satisfy the following:
\begin{enumerate}
\labitem{{(c1)}}{c1} $R_{p,-}$ takes the oriented frame $X_+$ to the oriented frame $X_-$;
\labitem{{(c2)}}{c2} $R_{p,+}$ takes the oriented frame $X_-$ to the oriented frame $X_+$;
\labitem{{(c3)}}{c3}
\begin{align*}
&\dim \operatorname{Ker}\dbar_{p,+} \equiv 0, & \dim \operatorname{Coker}\dbar_{p,+} &\equiv |p| + n +d_A+1,& \\
&\dim \operatorname{Ker}\dbar_{p,-} \equiv 1, & \dim \operatorname{Coker}\dbar_{p,-} &\equiv |p|,&
\end{align*}
everything modulo 2.
\end{enumerate}
\begin{rmk}
The author has not been able to prove Theorem \ref{thm:georgios} -- \ref{thm:differentcappings} for capping
operators not satisfying \ref{c1} -- \ref{c3}, but believes it
should be possible.
\end{rmk}
\begin{rmk}\label{rmk:notv}
Note that \ref{c1} -- \ref{c3} is not vacuous, since it is satisfied by the capping operators from \cite{orientbok} and from \cite{orienttrees}. This follows from [\cite{orientbok}, Section 3.3.6] for $d_A=2$, [\cite{orientbok}, Section 4.5.2] together with [\cite{legsub}, Proposition 8.14] for $d_A=1$, and from [\cite{orienttrees}, Corollary 3.31].
\end{rmk}
\subsubsection{Capping trivialization in the symplectization-direction}
We extend the boundary conditions $R_{p,\pm}$ to the symplectization,
by defining them to be given by the identity in the
${\mathbb R}_t$-direction. We denote the induced capping operators by
$\bar{\partial}_{s,p,\pm}$, and we use the notation $\bar{\partial}_{l,p,\pm}$ for the
capping operators for disks in $T^*M$ defined above (that is,
$\bar{\partial}_{l,p,\pm}$ is the restriction of $\bar{\partial}_{s,p,\pm}$ to $T^*M$,
but still extended to the auxiliary space).
We need to consider weighted Sobolev spaces to get the capping operators
$\bar{\partial}_{s,p,\pm}$ to be Fredholm. To that end, we put a small positive
exponential weight at the puncture in the $\partial_t$-direction. We get the following.
\begin{prp}\label{cor:capiso}
For each Reeb chord $p$ of $\Lambda$ the projection $\pi_P$ extends to canonical isomorphisms
\begin{align*}
\pi_P : \operatorname{Ker}\dbar_{s,p,\pm} \to \operatorname{Ker}\dbar_{l,p,\pm}, \\
\pi_P : \operatorname{Coker}\dbar_{s,p,\pm} \to \operatorname{Coker}\dbar_{l,p,\pm}.
\end{align*}
\end{prp}
\begin{proof}
We need to prove that in the ${\mathbb R}_t$-direction the capping operators $\bar{\partial}_{s,p,\pm}$ are
isomorphisms. But this follows from [\cite{legsub}, Proposition 8.14 and Proposition 8.16] together with the fact that we are considering a $\bar{\partial}$-problem with a one-dimensional Lagrangian boundary condition given by ${\mathbb R}$. More directly, this can be seen by using Fourier expansion as in the proof of [\cite{jholo}, Theorem C.4.1].
\end{proof}
\subsubsection{Orientation of capping operators}
Next we define the orientation of the capping operators $\bar{\partial}_{s,p,\pm}$, $\bar{\partial}_{l,p,\pm}$, by slightly adjusting the constructions from [\cite{orientbok}, Section 3.3] to our situation.
Recall that for each Reeb chord $p$ we are assumed to fix an orientation of $\det
\bar{\partial}_{l,p,-}$ when we specify our system of capping operators. This will be the \emph{capping orientation of
$\bar{\partial}_{l,p,-}$}. Notice that this canonically induces an orientation
of $\det \bar{\partial}_{s,p,-}$ via the isomorphism in Proposition \ref{cor:capiso}.
To define the orientation of $\det
\bar{\partial}_{s,p,+}$, let $\bar{\partial}_{s,p}$ denote the $\bar{\partial}$-problem on the
non-punctured disk obtained by gluing the $\bar{\partial}_{s,p,+}$-problem to the
$\bar{\partial}_{s,p,-}$-problem, and consider the induced exact gluing sequence
\begin{equation}\label{eq:glucap}
0 \to
\operatorname{Ker} \bar{\partial}_{s,p} \to
\begin{bmatrix}
\operatorname{Ker} \bar{\partial}_{s,p,+} \\
{\mathbb R}_t\\
\operatorname{Ker} \bar{\partial}_{s,p,-}
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker} \bar{\partial}_{s,p,+}\\
\operatorname{Coker} \bar{\partial}_{s,p,-}
\end{bmatrix}
\to
\operatorname{Coker} \bar{\partial}_{s,p} \to 0.
\end{equation}
Here the ${\mathbb R}_t$-summand corresponds to a gluing kernel which is born when gluing positive weighted Sobolev spaces, compare [\cite{orienttrees}, Lemma 3.16].
The chosen capping orientation of $\bar{\partial}_{s,p,-}$ together with the canonical orientation of $\det \bar{\partial}_{s,p}$ and the natural orientation of ${\mathbb R}_t$, induces an orientation $\mathcal{O}(p_+)$ of $\det \bar{\partial}_{s,p,+}$, via the sequence \eqref{eq:glucap} and the isomorphism \eqref{eq:phi}.
\begin{defi}\label{def:capor}
We define the \emph{capping orientation of $\bar{\partial}_{s,p,+}$} to be
$(-1)^{|p|+n + d_A + 1} \mathcal{O}(p+)$, and we will refer to the $\bar{\partial}_{s,p}$ -problem as \emph{the glued capping disk at $p$}.
\end{defi}
We give the operator $\bar{\partial}_{l,p,+}$ the capping orientation induced by the capping orientation of $\bar{\partial}_{s,p,+}$ under the isomorphism $\pi_P$ from Proposition \ref{cor:capiso}. We let $\bar{\partial}_{l,p}$ denote the $\bar{\partial}$-problem on the non-punctured disk obtained from gluing $\bar{\partial}_{l,p,+}$ to $\bar{\partial}_{l,p,-}$, and we prove the following.
\begin{lma}\label{lma:lma1}
Assume that we have chosen an admissible system of capping operators
for $\Lambda$.
Then the capping orientations of $\bar{\partial}_{l,p,+}$ and $\bar{\partial}_{l,p,-}$ glue to the canonical orientation of $\bar{\partial}_{l,p}$, times $(-1)^{|p|+n+d_A+1}$, under the exact gluing sequence
\begin{equation}\label{eq:glucaplag}
0 \to \operatorname{Ker} \bar{\partial}_{l,p} \to
\begin{bmatrix}
\operatorname{Ker} \bar{\partial}_{l,p,+} \\
\operatorname{Ker} \bar{\partial}_{l,p,-}
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker} \bar{\partial}_{l,p,+}\\
\operatorname{Coker} \bar{\partial}_{l,p,-}
\end{bmatrix}
\to
\operatorname{Coker} \bar{\partial}_{l,p} \to 0.
\end{equation}
\end{lma}
\begin{rmk}
Since we use the gluing convention \eqref{eq:gluconv}, this gives the opposite convention of [\cite{orientbok}, Section 3.3.4]. Also notice that the sign $(-1)^{|p|+n+d_A+1}$ in the definition of the orientation of the capping operators is not used in that paper.
\end{rmk}
\begin{proof}[Proof of Lemma \ref{lma:lma1}]
By construction we have that $\operatorname{Ker}\dbar_{s,p} \simeq {\mathbb R}_t \oplus \operatorname{Ker}\dbar_{l,p}$, and
from Remark \ref{rmk:glumap} we see that the first nontrivial map in \eqref{eq:glucap} restricted to the ${\mathbb R}_t$-factors is given
by projection $v \mapsto v$. Indeed, the kernel that is born during the gluing is cut-off and embedded in the Sobolev space of the glued map, and the gluing map $\alpha$ in \eqref{eq:gluconv} was given by $L^2$-projection. Since by the assumptions \ref{c3} we have $\dim \operatorname{Ker}\dbar_{s,p,+} \equiv 0$ modulo 2, we can remove ${\mathbb R}_t$ from both
the first and second nontrivial column without affecting orientations on the remaining spaces. But after removing ${\mathbb R}_t$ we get the gluing sequence for the capping operators in the Lagrangian projection, and since the canonical orientation is given via evaluation the result follows.
\end{proof}
From now on, we use the notation $\bar{\partial}_{p,\pm}$ to denote the capping operators both in the symplectization-setting and in the setting of the Lagrangian projection.
\subsection{Capping orientation of disks}\label{sec:orientdisk}
Now we give the definition of the capping orientation of a punctured pseudo-holomorphic disk. Below $X$ denotes either $T^*M$ or ${\mathbb R}\times J^1(M)$, with almost complex structure $J$ as described in Section \ref{sec:holo}.
If $L$ is an exact Lagrangian cobordism with cylindrical Legendrian ends $\Lambda_\pm$, then assume that $\mathcal{S}_\pm$ gives a system of capping operators for $\Lambda_\pm$. This gives rise to an \emph{induced system of capping operators of $L$}, where the positive capping operators $\bar{\partial}_{p,+}$ are taken from the system $\mathcal{S}_+$, and the negative capping operators $\bar{\partial}_{p,-}$ are taken from $\mathcal{S}_-$. This system is \emph{admissible} if both $\mathcal{S}_\pm$ are admissible.
Let $u :D_m \to X$ be a pseudo-holomorphic disk of $\Lambda$ or of $L$, with positive puncture $a$ and negative punctures $b_1,\dotsc,b_m$. Assume that we have fixed a system of capping operators, and consider the exact gluing sequence
\begin{equation}\label{eq:caporient}
0 \to \operatorname{Ker} \bar{\partial}_{\hat u}
\to
\begin{bmatrix}
\operatorname{Ker} \bar{\partial}_u \\
\operatorname{Ker} \bar{\partial}_{a,+} \\
\operatorname{Ker} \bar{\partial}_{b_m,-} \\
\vdots\\
\operatorname{Ker} \bar{\partial}_{b_1,-}
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker} \bar{\partial}_u \\
\operatorname{Coker} \bar{\partial}_{a,+} \\
\operatorname{Coker} \bar{\partial}_{b_m,-} \\
\vdots\\
\operatorname{Coker} \bar{\partial}_{b_1,-}
\end{bmatrix}
\to
\operatorname{Coker} \bar{\partial}_{\hat u}
\to 0.
\end{equation}
Here $\bar{\partial}_{\hat u}$ denotes the $\bar{\partial}$-problem on the non-punctured
disk with trivialized boundary condition $\hat u$, which is obtained
by gluing the trivialized boundary condition induced by $u$ to the
positive capping trivialization of $a$ at the positive puncture of
$D_m$, and then to the negative capping trivializations of
$b_m,\dotsc,b_1$ at the corresponding negative punctures. We refer to
the $\bar{\partial}_{\hat u}$-problem as the \emph{fully capped problem
corresponding to $u$}, and to the sequence \eqref{eq:caporient} as the
\emph{capping sequence for $u$}.
\begin{rmk}
The reason that the capping operators are glued clockwise along $u$, instead of counterclockwise as in [\cite{orientbok}, Section 3.3.5], is because of our choice of convention \eqref{eq:gluconv}, which in turn depended on the fact that we have to take into account the extra ${\mathbb R}$-direction coming from the symplectization. This is explained further in Section \ref{sec:clock}.
\end{rmk}
Let $\operatorname{index} (u)$ denote the Fredholm index of $\bar{\partial}_u$, restricted to the space of candidate maps. That is, if $u$ is rigid and has $m$ negative punctures, then
\begin{equation*}
|\operatorname{index} (u)| \equiv m \pmod{2}.
\end{equation*}
\begin{defi}\label{def:capping}
We define the \emph{capping orientation} of $u \in \mathcal{M}_{l, \Lambda}(a, {\bf b}),\mathcal{M}_L(a, {\bf b})$ to be the orientation $\mathcal{O}(\bar{\partial}_u)$ on $\det \bar{\partial}_u$ induced by the gluing sequence \eqref{eq:caporient}, where $\det \bar{\partial}_{\hat u}$ is given the canonical orientation, and the capping operators are given their capping orientations. For $ u \in \mathcal{M}_{s, \Lambda}(a, {\bf b})$ we define the capping orientation to be given by $\det \bar{\partial}_{\pi_p(u)}\wedge \partial_{t} = (-1)^{\operatorname{index} (\pi_p(u))}\partial_{t} \wedge \det \bar{\partial}_{\pi_p(u)}$, where $\det \bar{\partial}_{\pi_p(u)}$ is given its capping orientation and $\partial_{t}$ gives the positive orientation in the symplectization direction.
\end{defi}
\begin{rmk}
We will use the notation $\partial_{\mathcal{S}}$ to indicate the dependence of the chosen capping system $\mathcal{S}$ in the definition for the DGA-differential.
\end{rmk}
In the case when $u$ is a rigid disk (and where we assume that we have
divided out the ${\mathbb R}_t$-action if $u$ is a disk of ${\mathbb R} \times
\Lambda$), the capping orientation of $u$ can be understood as an
orientation of the kernel or the cokernel of $\bar{\partial}_u$. Moreover, by
[\cite{orientbok}, Section 4.2.3], we may assume that we are in the
case when $\operatorname{Ker}\dbar_u$ is trivial, so that an orientation of $\det \bar{\partial}_u$
is nothing but an orientation of $\operatorname{Coker}\dbar_u$. Let $\kappa$ denote the
conformal structure of $u$, and recall that the space of
conformal variations at $\kappa$ was given a fixed
orientation in Section \ref{sec:confvar}. Thus, we can compare the capping
orientation of $u$ with
this orientation via the isomorphism \eqref{eq:psibar} and get a sign $\sigma(u) \in \{-1,1\}$. This sign is
the \emph{capping sign of $u$}, and is the one that we use in the the
algebraic count of the elements in the moduli spaces when defining the
DGA-morphisms and the DGA-differentials in (\ref{eq:differential}) and
(\ref{eq:phidisk}), respectively. See [\cite{orientbok},
Section 3.4.3].
\begin{rmk}\label{rmk:onedim}
In the case when $u$ belongs to a one-dimensional moduli space $\mathcal{M}$ (after having divided out the ${\mathbb R}_t$-action if $u$ is a disk of ${\mathbb R}\times \Lambda$) we get that the map given in \eqref{eq:psibar} is not an isomorphism, but is surjective with one-dimensional kernel $\operatorname{Ker} \bar \Psi$. Give this space the orientation so that
\begin{equation*}
T\mathcal{C}_{m} = \operatorname{Ker} \bar \Psi \oplus \operatorname{Coker}\dbar_u,
\end{equation*}
as oriented vector spaces. Here $\mathcal{C}_{m}$ is the space of conformal variations at $u$ and is given its fixed orientation, and $\operatorname{Coker}\dbar_u$ is given its capping orientation.
The \emph{capping orientation of $\mathcal{M}$ at $u$} is given by this orientation of $\operatorname{Ker} \bar \Psi$, after having identified $\operatorname{Ker} \bar \Psi$ with $\operatorname{Ker} d\Gamma_u$.
\end{rmk}
\subsection{Proof of Theorem \ref{thm:georgios}}\label{sec:pfger}
To prove Theorem \ref{thm:georgios} it only remains to establish the following.
\begin{lma}\label{lma:lma3}
Let $\mathcal{S}$ be an admissible system of capping operators and let $i=s$ or $i=l$. Then the map defined by
\begin{equation*}
\partial_{\mathcal{S},i} a = \sum_{\dim \mathcal{M}(a,{\bf b}) = 0} | \mathcal{M}_{i,\Lambda}(a, {\bf b})| {\bf b}
\end{equation*}
on generators and extended by the signed Leibniz rule to rest of the algebra, satisfies $\partial_{\mathcal{S},i}^2 =0$. Here $| \mathcal{M}_{i,\Lambda}(a, {\bf b})|$ is the algebraic count of disks in $ \mathcal{M}_{i,\Lambda}(a, {\bf b})$, where each disk is counted with its capping sign induced by $\mathcal{S}$.
\end{lma}
\begin{proof}
We follow the proof of Theorem 4.1 in \cite{orientbok}, and in
particular the notations therein. Briefly, the argument goes as
follows.
Let $a$ be a Reeb chord of $\Lambda$, and assume that $\partial^2(a)$ has a summand $N{\bf d}$, where $N \in {\mathbb Z}$ and $ {\bf d}= d_1 \cdots d_l$ is a word of Reeb chords. Then $\mathcal{M}(a, {\bf d})$ is one-dimensional and can be compactified by broken pairs of disks from zero-dimensional moduli spaces. Let $\mathcal{M} \subset \mathcal{M}(a, {\bf d})$ be a component with oriented boundary
\begin{equation}
\partial \mathcal{M} = \mathcal{M}_1 - \mathcal{M}_0,
\end{equation}
where $\mathcal{M}_0$ is given by 2 broken disks $(u_1, u_2)$ with
\begin{align*}
& u_2 \in \mathcal{M}(a; d_1 \cdots d_{k-1} c d_{k+m+1} \cdots d_l),\\
& u_1 \in \mathcal{M}(c; d_k \cdots d_{k+m}),
\end{align*}
and $\mathcal{M}_1$ is given by 2 broken disks $(u_1', u_2')$ with
\begin{align*}
& u_2' \in \mathcal{M}(a; d_1 \cdots d_{k'-1} c' d_{k'+m'+1} \cdots d_l),\\
& u_1' \in \mathcal{M}(c'; d_{k'} \cdots d_{k'+m'}).
\end{align*}
Let $\mu_i$ be the capping sign of $u_i$, and let $\mu_i'$ be the capping sign of $u_i'$, $i= 1,2$. Then the broken disk $(u_1, u_2)$ contributes with $(-1)^{\sum_{i=1}^{k-1}|d_i|} \mu_1 \mu_2 {\bf d}$ to $\partial^2(a)$, and the broken disk $(u_1', u_2')$ contributes with $(-1)^{\sum_{i=1}^{k'-1}|d_i|} \mu_1' \mu_2' {\bf d}$. We claim that
\begin{equation}\label{eq:claim}
(-1)^{\sum_{i=1}^{k-1}|d_i|} \mu_1 \mu_2 = - (-1)^{\sum_{i=1}^{k'-1}|d_i|} \mu_1' \mu_2'.
\end{equation}
To prove this claim we would like to use the orientation of $\mathcal{M}$. That is, by gluing, we have that $u_1\# u_2$ and $u_1'\# u_2'$ give elements in $\mathcal{M}$ with associated $\bar{\partial}$-operators $\bar{\partial}_{u_1\#u_2}$, $\bar{\partial}_{u_1'\#u_2'}$, equipped with capping orientations. Let $o_c(\bar{\partial}_{u_1 \# u_2})$ denote the capping orientation of $\det \bar{\partial}_{u_1 \# u_2}$ and let $o_i(\bar{\partial}_{u_1 \# u_2})$ denote the orientation induced on $\det \bar{\partial}_{u_1 \# u_2}$ by the gluing sequence
\begin{equation}\label{eq:ny01}
0 \to
\operatorname{Ker}\dbar_{u_1 \# u_2}
\to
\begin{bmatrix}
\operatorname{Ker}\dbar_{u_2} \\
\operatorname{Ker}\dbar_{u_1} \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{u_2} \\
\operatorname{Coker}\dbar_{u_1} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{u_1 \# u_2}
\to 0
\end{equation}
in the setting of the Lagrangian projection, and by the sequence
\begin{equation}\label{eq:nys01}
0 \to
\operatorname{Ker}\dbar_{u_1 \# u_2}
\to
\begin{bmatrix}
\operatorname{Ker}\dbar_{u_2} \\
\operatorname{Ker}\dbar_{u_1} \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{u_2} \\
{\mathbb R}_t \\
\operatorname{Coker}\dbar_{u_1} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{u_1 \# u_2}
\to 0
\end{equation}
in the symplectization setting. Here we assume that $\det \bar{\partial}_{u_i}$ is given its capping orientation for $i=1,2$.
The ${\mathbb R}_t$-summand in \eqref{eq:nys01} comes from a gluing cokernel that is born when we glue negative weighted problems, see [\cite{orientbok}, Lemma 3.1]. We define $o_c(\bar{\partial}_{u'_1 \# u'_2})$, $o_i(\bar{\partial}_{u'_1 \# u'_2})$ in a completely analogous way.
Let
\begin{equation*}
\delta = \dim \operatorname{Coker}\dbar_{u_1} + \dim \operatorname{Coker}\dbar_{u_2} +1 = \dim \operatorname{Coker}\dbar_{u_1'} + \dim \operatorname{Coker}\dbar_{u_2'} +1
\end{equation*}
if we are in the symplectization setting, and
\begin{equation*}
\delta = n + d_A+1
\end{equation*}
if we are in the setting of the Lagrangian projection. The claim \eqref{eq:claim} will follow if we can prove that
\begin{align}\label{eq:holm1}
& o_c(\bar{\partial}_{u_1 \# u_2}) = (-1)^{(\dim \operatorname{Coker}\dbar_{u_1}-1)k + \sum_{i=1}^{k-1}|d_i|+ \delta} o_i(\bar{\partial}_{u_1 \# u_2}) \\ \label{eq:holm2}
& o_c(\bar{\partial}_{u'_1 \# u'_2}) = (-1)^{(\dim \operatorname{Coker}\dbar_{u'_1}-1)k' + \sum_{i=1}^{k'-1}|d_i|+\delta} o_i(\bar{\partial}_{u'_1 \# u'_2}).
\end{align}
Before we derive these equations, we prove that they
imply \eqref{eq:claim}. Indeed, to relate the orientation of $\mathcal{M}$ with \eqref{eq:holm1}, we should consider the following commutative diagram
\begin{equation}\label{eq:twostar}
\xymatrix{
T\mathcal{C}_{u_2}\oplus T\mathcal{C}_{u_1} \oplus {\mathbb R} \ar[d] \ar[r] & T\mathcal{C}_{u_1\#u_2} \ar[d] \\
\operatorname{Coker}\dbar_{u_2} \oplus \operatorname{Coker}\dbar_{u_1} \ar[r] & \operatorname{Coker}\dbar_{u_1\#u_2}.
}
\end{equation}
Here $\mathcal{C}_v$ denotes the space of conformal structures of the domain of the disk $v$, $v=u_1, u_2, u_1\#u_2$. This commuting diagram is similar to the one given in the proof of [\cite{orientbok}, Lemma 4.11], and the maps in the diagram are given as follows.
\begin{enumerate}
\labitem{{(hu)}}{hu} The upper horizontal map is given by the map in Lemma \ref{lma:confvar}.
\labitem{{(hl)}}{hl} The lower horizontal map is the one induced by the gluing sequence \eqref{eq:ny01} in the setting of the Lagrangian projection and induced by the gluing sequence \eqref{eq:nys01} in the symplectization setting.
\labitem{{(vl)}}{vl} The left vertical map is given by $(x,y,t) \mapsto(\bar\Psi_1(x), \bar \Psi_2(y))$, where $\bar \Psi_i$ is the map \eqref{eq:psibar} associated to $u_i$, $i=1,2$.
\labitem{{(vr)}}{vr} The right vertical map is given by $z \mapsto \bar \Psi(z)$, where $\bar \Psi$ is the map \eqref{eq:psibar} associated to $u_1\#u_2$.
\end{enumerate}
Assume that $T\mathcal{C}_{u_1\#u_2}, T\mathcal{C}_{u_1}, T\mathcal{C}_{u_2}$ are given their fixed orientations and that ${\mathbb R}$ is endowed with the orientation corresponding to the outward normal of $T\mathcal{C}_{u_1\#u_2}$, as in Section \ref{sec:confvar}. Also assume that $\operatorname{Coker}\dbar_{u_1\#u_2}, \operatorname{Coker}\dbar_{u_1}, \operatorname{Coker}\dbar_{u_2}$ are given their capping orientations. Then Lemma \ref{lma:confvar} implies that \ref{hu} is an isomorphism between oriented spaces of sign $(-1)^{\nu_1}$,
\begin{equation*}
\nu_1 = (\dim T\mathcal{C}_{u_1}-1)\cdot k +1,
\end{equation*}
and \ref{vl} restricted to $T\mathcal{C}_{u_1} \oplus T\mathcal{C}_{u_2}$ is by definition an isomorphism between oriented spaces of sign $(-1)^{\nu_2}$,
\begin{equation*}
(-1)^{\nu_2} = \mu_1\mu_2.
\end{equation*}
From \eqref{eq:holm1} we get that the sign of \ref{hl} is given by $(-1)^{\nu_3}$,
\begin{equation*}
\nu_3 = (\dim \operatorname{Coker}\dbar_{u_1}-1)\cdot k + \sum_{i=1}^{k-1}|d_i|+ \delta.
\end{equation*}
Thus, by Remark \ref{rmk:onedim} together with the commutativity of \eqref{eq:twostar} we see that the orientation of $T_{u_1\#u_2}\mathcal{M}$ is given by
\begin{align*}
(-1)^{\nu_1 + \nu_2 + \nu_3}{\mathbb R} &= (-1)^{(\dim T\mathcal{C}_{u_1}-1)\cdot k +1 + (\dim \operatorname{Coker}\dbar_{u_1}-1)\cdot k + \sum_{i=1}^{k-1}|d_i|+ \delta}\mu_1\mu_2 {\mathbb R}\\
&= (-1)^{1+\sum_{i=1}^{k-1}|d_i|+ \delta}\mu_1\mu_2{\mathbb R},
\end{align*}
where ${\mathbb R}$ is given the orientation of the outward normal of $T\mathcal{C}_{u_1\#u_2}$ at $u_1\#u_2$.
Using the same argument at $u'_1\#u'_2$ we get in total that the orientation of $\mathcal{M}$ at the boundary component $\mathcal{M}_0$ is given by $(-1)^{1+\sum_{i=1}^{k-1}|d_i|+ \delta}\mu_1\mu_2$ times the outer normal and the orientation of $\mathcal{M}$ at the boundary component $\mathcal{M}_1$ is given by $(-1)^{1+\sum_{i=1}^{k'-1}|d_i|+ \delta}\mu'_1\mu'_2$ times the outer normal. But since the orientation of $\mathcal{M}_1$ is opposite to the orientation of $\mathcal{M}_0$ we get \eqref{eq:claim}.
To prove \eqref{eq:holm1} and \eqref{eq:holm2} , consider the general situation where
$A$ and $B$ are trivialized Lagrangian boundary conditions
associated to punctured disks $D_{m}$ and $D_{r+1}$,
respectively, and assume that we glue them together as described in
Section \ref{sec:gluing}. That is, we glue the positive puncture of
$\bar{\partial}_A$ to the $k$:th negative puncture of $\bar{\partial}_B$.
Note that $\operatorname{Ker}\dbar_{s,A} \simeq {\mathbb R}_t \oplus \operatorname{Ker}\dbar_{l,A}$, $\operatorname{Coker}\dbar_{s,A} \simeq \operatorname{Coker}\dbar_{l,A}$, and similarly for $\hat A$, $B$ and $\hat B$. We write $\bar{\partial}_{C}=\bar{\partial}_{l,C} $ to simplify notation, where $C$ is any trivialized Lagrangian boundary condition on the (possibly punctured) disk.
Assume that $\bar{\partial}_A$ corresponds to a disk $u_1 \in \mathcal{M}(b_k,f_1\cdots f_{m-1})$ and that $\bar{\partial}_B$ corresponds to a disk $u_2 \in \mathcal{M}(a,b_1\cdots b_r)$, and that both these disks are rigid. Also assume that we are in the symplectization setting (the computations in the setting of the
Lagrangian projection are similar), and that $m>2, r>1$.
The sequence \eqref{eq:nys01} now reads
\begin{equation}\label{eq:ny1}
0 \to
\operatorname{Ker}\dbar_{s,A \# B}
\to
\begin{bmatrix}
\operatorname{Ker}\dbar_{s,B} \\
\operatorname{Ker}\dbar_{s,A} \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{s,B} \\
{\mathbb R}_t \\
\operatorname{Coker}\dbar_{s,A} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{s, A \# B}
\to 0,
\end{equation}
which moreover can be simplified to
\begin{equation}\label{eq:ny001}
0 \to
{\mathbb R}_t^{A\#B}
\to
\begin{bmatrix}
{\mathbb R}_t^B \\
{\mathbb R}_t^A
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{B} \\
{\mathbb R}_t \\
\operatorname{Coker}\dbar_{A}
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{A \# B}
\to 0.
\end{equation}
Here we use the notation ${\mathbb R}_t^{A\#B}$ for the ${\mathbb R}_t$-factor in $\operatorname{Ker}\dbar_{A\#B}$, and similar for ${\mathbb R}_t^A$, ${\mathbb R}_t^B$. We have also used the assumption that $m>2, r>1$, so that $\operatorname{Ker}\dbar_A=\operatorname{Ker}\dbar_B=\operatorname{Ker}\dbar_{A\#B}=0$.
By Remark \ref{rmk:glumap} and the proof of [\cite{orientbok}, Lemma 3.11], we may assume that the first non-trivial map in \eqref{eq:ny001} is given by $t \mapsto (t,t)$ and that the second map is given by $(s,t) \mapsto (0,t-s,0)$. By our orientation conventions \eqref{eq:o2} it follows that if we move ${\mathbb R}_t$ over $\operatorname{Coker}\dbar_A$, with a cost of $(-1)^{\sigma_0}$,
\begin{equation}\label{eq:cokersign}
\sigma_0 = \dim \operatorname{Coker}\dbar_A,
\end{equation}
we can reduce \eqref{eq:ny001} to
\begin{equation}
0 \to
0
\to
0
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{B} \\
\operatorname{Coker}\dbar_{A}
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{A \# B}
\to 0,
\end{equation}
where the non-trivial map is the one that occurs in \eqref{eq:twostar} as \ref{hl}.
To compute the difference between the capping orientation of $\det \bar{\partial}_{s,A\#B}$ and the orientation induced by the capping orientations of $\det \bar{\partial}_{s,A}$ and $\det \bar{\partial}_{s,B}$ from the sequence \eqref{eq:ny001}, consider the following two gluing sequences
\begin{equation}\label{eq:4.9.1}
\begin{bmatrix}
{\mathbb R}_t \\
\operatorname{Ker}\dbar_{\hat A \# \hat B}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
{\mathbb R}_t^B \\
\hat{\bigoplus}_{j=1}^r \operatorname{Ker}\dbar_{b_{j},-} \\
\end{pmatrix}\\
\begin{pmatrix}
{\mathbb R}_t^A \\
\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Ker}\dbar_{f_{j},-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar_B \\
\operatorname{Coker}\dbar_{a,+} \\
\hat{\bigoplus}_{j=1}^r \operatorname{Coker}\dbar_{b_{j},-}
\end{pmatrix}\\
{\mathbb R}_t \\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_A \\
\operatorname{Coker}\dbar_{b_{k},+} \\
\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Coker}\dbar_{f_{j},-}
\end{pmatrix}
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{\hat A \# \hat B},
\end{equation}
\begin{equation}\label{eq:4.9.2}
\begin{bmatrix}
{\mathbb R}_t \\
\operatorname{Ker}\dbar_{\hat A \# \hat B}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
{\mathbb R}_t^B \\
{\mathbb R}_t^A\\
\hat{\bigoplus}_{j=k+1}^r \operatorname{Ker}\dbar_{b_{j},-} \\
\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Ker}\dbar_{f_{j},-} \\
\hat{\bigoplus}_{j=1}^{k-1} \operatorname{Ker}\dbar_{b_{j},-}
\end{pmatrix}\\
\begin{pmatrix}
{\mathbb R}_t \\
\operatorname{Ker}\dbar_{b_{k},-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar_B \\
{\mathbb R}_t\\
\operatorname{Coker}\dbar_A \\
\operatorname{Coker}\dbar_{a,+} \\
\hat{\bigoplus}_{j=k+1}^r \operatorname{Coker}\dbar_{b_{j},-}\\
\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Coker}\dbar_{f_{j},-} \\
\hat{\bigoplus}_{j=1}^{k-1} \operatorname{Coker}\dbar_{b_{j},-}\\
\end{pmatrix}\\
{\mathbb R}_t \\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{b_{k},+} \\
\operatorname{Coker}\dbar_{b_{k},-}
\end{pmatrix}
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{\hat A \# \hat B}.
\end{equation}
Here we use the notation $\hat{\bigoplus}_{j=1}^r V_j = V_r \oplus \dotsb \oplus V_1$. We have also omitted the trivial maps $0 \to {\mathbb R}_t \oplus \operatorname{Ker}\dbar_{\hat A \# \hat B}$ and $\operatorname{Coker}\dbar_{\hat A \# \hat B} \to 0$ to fit the sequences within page margins. In addition, we have dropped some spaces which by assumption have even dimension, since they will not affect the calculations below.
The first sequence corresponds to the gluing of $\bar{\partial}_{s,\hat A}$ to $\bar{\partial}_{s,\hat B}$, and the second one to the gluing of $\bar{\partial}_{s,\widehat{A\#B}}$ to the glued capping disk at $b_k$. Here we have used the associativity of orientations under gluing, see [\cite{orientbok}, Section 3.2.3], together with the capping sequences for $\bar{\partial}_{s,A}$, $\bar{\partial}_{s,B}$ and $\bar{\partial}_{s,A\#B}$, respectively. By Lemma \ref{lma:kan} we get that both sequences induce the canonical orientation on $\bar{\partial}_{s,\hat A \# \hat B}$, given that
$\bar{\partial}_{s,\hat A}$, $\bar{\partial}_{s,\hat B}$, $\bar{\partial}_{s,\widehat{A\#B}}$ and the glued capping disk at $b_k$ all are given their canonical orientation.
\begin{rmk}\label{rmk:forenkla}
Notice that, if we compensate by an overall sign \linebreak
$(-1)^{\operatorname{index} \bar{\partial}_A + \operatorname{index} \bar{\partial}_B + \operatorname{index} \bar{\partial}_{A\#B}} = 1$ at the end, we can instead of the capping orientations of $\bar{\partial}_{s,A}, \bar{\partial}_{s,B}, \bar{\partial}_{s,A\#B}$ consider the capping orientations of $\bar{\partial}_{A}, \bar{\partial}_{B}, \bar{\partial}_{A\#B}$ (assuming that all ${\mathbb R}_t$-summands are given their natural orientation). This follows from Definition \ref{def:capping}.
\end{rmk}
Thus, if we assume that $\bar{\partial}_A$ and $\bar{\partial}_B$ are given their capping orientations, we can compute the difference between the capping orientation of $\bar{\partial}_{A\#B}$ and the one induced by the sequence \eqref{eq:ny001} by rearranging the spaces in \eqref{eq:4.9.1} to be in the same order as the spaces in \eqref{eq:4.9.2}, in the same time as we keep track of the change in orientations that these rearrangements induce.
First, by Remark \ref{rmk:glumap} and the proof of [\cite{orientbok}, Lemma 3.11], we may assume that the bottom-most ${\mathbb R}_t$-summand in
the second column of \eqref{eq:4.9.2} is mapped by the identity to the
bottom-most ${\mathbb R}_t$-summand in the third column. Thus these two spaces
can be removed if we compensate by a sign
\begin{equation*}
(-1)^{\dim \operatorname{Ker}\dbar_{b_k,-} + \dim {\mathbb R}^{n+d_A} + \dim \operatorname{Coker}\dbar_{b_k,+} + \dim \operatorname{Coker}\dbar_{b_k,-}}=(-1)^{1 + n + d_A + |b_k| +n+1+d_A+|b_k|} = 1.
\end{equation*}
Here we have used \ref{c3} repeatedly.
Next, notice that by \ref{c3}, \eqref{eq:dim1} and by the assumption that $u_2$ is rigid, we have
\begin{equation}\label{eq:forenk}
\dim \operatorname{Coker}\dbar_{a,+} + \dim \left(\hat{\bigoplus}_{j=1}^{r}\operatorname{Coker}\dbar_{b_{j},-}\right) \equiv |a|+1 +n+d_A + \sum_{j=1}^{r}|b_j| \equiv n+d_A \pmod{2}.
\end{equation}
Thus, by moving ${\mathbb R}_t^A$ in the sequence \eqref{eq:4.9.1} to the
position right under ${\mathbb R}_t^B$, in the same time as we move the
${\mathbb R}_t$-factor in the third column to the position right under $\operatorname{Coker}\dbar_B$, we get a sign
$(-1)^{\sigma_1}$,
\begin{align*}
\sigma_1 &= 1\cdot \dim\left(\hat{\bigoplus}_{j=1}^{r}\operatorname{Ker}\dbar_{b_{j},-}\right) + 1 \cdot\left( \dim\left(\hat{\bigoplus}_{j=1}^{r}\operatorname{Coker}\dbar_{b_{j},-}\right) + \dim \operatorname{Coker}\dbar_{a,+} \right) \\
&\equiv r + n +d_A \pmod{2}.
\end{align*}
Again we have used \ref{c3}.
Then we move $\operatorname{Coker}\dbar_A$, which has dimension $m-1$ modulo 2, to the place just below $\operatorname{Coker}\dbar_B \oplus {\mathbb R}_t$. Using \eqref{eq:forenk} we see that this costs
\begin{equation*}
(-1)^{(m-1)(n+d_A +n+d_A )} =1.
\end{equation*}
Then we change places of $ \hat{\bigoplus}_{j=1}^{m-1} \operatorname{Ker}\dbar_{f_{j},-}$ and $\operatorname{Ker}\dbar_{b_{k},-}$, which costs $(-1)^{\sigma_2}$,
\begin{align*}
\sigma_2 &= \left(\dim \left(\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Ker}\dbar_{f_{j},-}\right) + \dim \operatorname{Ker}\dbar_{b_{k},-}\right) \cdot \dim \left(\hat{\bigoplus}_{j=1}^{k-1} \operatorname{Ker}\dbar_{b_{j},-}\right)\\
&\quad + \dim \left(\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Ker}\dbar_{f_{j},-}\right) \cdot \dim \operatorname{Ker}\dbar_{b_{k},-}\\
&\equiv m(k-1) + m-1 \equiv mk +1 \pmod{2},
\end{align*}
and then we do the similar thing for the cokernels. Using that
\begin{equation*}
\dim \left(\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Coker}\dbar_{f_{j},-}\right) \equiv \sum_{j=1}^{m-1}|f_j| \equiv |b_k| +1 \pmod{2},
\end{equation*}
again by \ref{c3}, \eqref{eq:dim1} and since $u_1$ is rigid, we see that this permutation costs $(-1)^{\sigma_3}$,
\begin{align*}
\sigma_3 &= \left(\dim \left(\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Coker}\dbar_{f_{j},-}\right) + \dim \operatorname{Coker}\dbar_{b_{k},-}\right) \cdot \\ &\quad \left(\dim \operatorname{Coker}\dbar_{b_{k},+} + \dim {\mathbb R}^{n+d_A} + \dim \left(\hat{\bigoplus}_{j=1}^{k-1} \operatorname{Coker}\dbar_{b_{j},-}\right)\right)\\
&\quad + \dim \left(\hat{\bigoplus}_{j=1}^{m-1} \operatorname{Coker}\dbar_{f_{j},-}\right) \cdot \dim \operatorname{Coker}\dbar_{b_{k},-}\\
&\equiv(|b_k| +1 + |b_k|)(|b_k| +1 + n +d_A + n + d_A + \sum_{i=1}^{k-1}|b_i|) + (|b_k| +1)|b_k|\\
&\equiv \sum_{i=1}^{k}|b_i| +1 \pmod{2}.
\end{align*}
Thus we get a total sign $(-1)^{\sigma}$,
\begin{equation*}
\sigma = \sigma_1 + \sigma_2 + \sigma_3 \equiv r +n +d_A+ mk + |b_k| + \sum_{i=1}^{k-1}|b_i| \pmod{2}.
\end{equation*}
Now, notice that the second gluing sequence does not induce the
canonical orientation of $\det \bar{\partial}_{\hat A \# \hat B}$, but
$(-1)^{|b_k|+ n+d_A+1}$ times the canonical orientation. This is because of Definition \ref{def:capor}, which implies that
\begin{equation*}
0 \to
\operatorname{Ker} \bar{\partial}_{s,b_{k}} \to
\begin{bmatrix}
\operatorname{Ker} \bar{\partial}_{s,b_{k},+} \\
{\mathbb R}_t\\
\operatorname{Ker} \bar{\partial}_{s,b_{k},-}
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker} \bar{\partial}_{s,b_{k},+}\\
\operatorname{Coker} \bar{\partial}_{s,b_{k},-}
\end{bmatrix}
\to
\operatorname{Coker} \bar{\partial}_{s,b_{k}} \to 0
\end{equation*}
induces the canonical orientation times $(-1)^{|b_k|+n+d_A+1}$ on $\det\bar{\partial}_{s,b_{k}}$, given that the capping operators are given their capping orientations. (In the Lagrangian setting, the similar conclusion follows from Lemma \ref{lma:lma1}.)
Hence the difference in orientation of $\det \bar{\partial}_{\hat A \# \hat B}$ given by \eqref{eq:4.9.1} and \eqref{eq:4.9.2}, assuming the capping operators are given their capping orientations, is given by $(-1)^{\tilde \sigma}$,
\begin{equation}\label{eq:sis}
\tilde \sigma = \sigma + |b_k|+n+d_A+1 \equiv mk + \sum_{i=1}^{k-1}|b_i| +r +1 \pmod{2}.
\end{equation}
Now recall form Definition \ref{def:capping} and Remark \ref{rmk:forenkla} that we would like to express the capping orientation of $\det \bar{\partial}_{s,A\#B}$ as a wedge $\partial_t \wedge \det \bar{\partial}_{A\#B}=\partial_t \wedge \bigwedge^{\max}\operatorname{Coker}\dbar_{A\#B}$. From \eqref{eq:cokersign} and \eqref{eq:sis} it follows that
\begin{equation*}
\partial_t \wedge \bigwedge^{\max}\operatorname{Coker}\dbar_{A\#B} = (-1)^{\tilde \sigma + \dim \operatorname{Coker}\dbar_A}\partial_t \wedge \bigwedge^{\max}\operatorname{Coker}\dbar_{B} \wedge \bigwedge^{\max}\operatorname{Coker}\dbar_{A}
\end{equation*}
given that $\operatorname{Coker}\dbar_A, \operatorname{Coker}\dbar_B$ and $\operatorname{Coker}\dbar_{A\#B}$ are given their capping orientations. Using Remark \ref{rmk:forenkla} again, noting that
\begin{equation*}
\tilde \sigma + \dim \operatorname{Coker}\dbar_A \equiv (\dim \operatorname{Coker}\dbar_{A}-1)\cdot k + \sum_{i=1}^{k-1}|b_i| + \dim \operatorname{Coker}\dbar_{A} + \dim \operatorname{Coker}\dbar_{B} +1
\end{equation*}
we deduce that \eqref{eq:holm1} holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:georgios}]
The first statement follows by [\cite{Georgios}, Theorem 2.1]
together with Definition \ref{def:capping}. The second statement follows from Lemma \ref{lma:lma3}.
\end{proof}
\section{DGA-morphisms induced by exact Lagrangian cobordisms}\label{sec:cob-pf}
Let $L$ be an exact Lagrangian cobordism with cylindrical Legendrian ends
$\Lambda_\pm$, and assume that we are given admissible systems $\mathcal{S}_\pm$ of
capping operators for $\Lambda_\pm$. Let $\partial_\pm =
\partial_{\mathcal{S}_\pm}$ to simplify notation.
Moreover, let $\mathcal{S}$ denote the induced system of $L$, and let $\Phi_L:
(\mathcal{A}(\Lambda_+), \partial_+, {\mathbb Z}) \to (\mathcal{A}(\Lambda_-), \partial_-, {\mathbb Z})$
be defined by \eqref{eq:phidisk} -- \eqref{eq:ex2}, where now $|\mathcal{M}_L(a,{\bf b})|$ denotes
the algebraic count induced by the capping orientation of $\mathcal{M}_L$
corresponding to the system $\mathcal{S}$. When we want to emphasize that
$\Phi_L$ is given with respect to the system $\mathcal{S}$ we write
$\Phi_{L,\mathcal{S}}$.
In this section we prove that this setup gives the statements in
Theorem \ref{thm:kalman}, Theorem \ref{thm:functorial} and Theorem \ref{thm:differentcappings}.
\subsection{Proof of Theorem \ref{thm:kalman} and Theorem
\ref{thm:differentcappings}}\label{sec:kalman-pf}
To prove Theorem \ref{thm:kalman}, we must show that
\begin{equation}\label{eq:samma}
\Phi_L \circ \partial_+ = \partial_- \circ \Phi_L.
\end{equation}
This uses similar techniques as the proof of Theorem
\ref{thm:georgios}. That is, we will use compactness results from
[\cite{compact}, Section 11.3] to pair up 2-level buildings that emanate from the
left hand side of \eqref{eq:samma} with
2-level buildings that emanate from the the right hand side. We perform
this pairing in a way so that each
pair can be interpreted as the boundary
of a 1-dimensional moduli space. Then we compute the difference
in orientations induced by gluings of such buildings, and argue that all
the signs cancel.
\begin{proof}[Proof of Theorem \ref{thm:kalman}.]
Assume that $L\subset {\mathbb R} \times J^1(M)$ is an exact Lagrangian
cobordism with cylindrical Legendrian ends $\Lambda_\pm$. Let $a$ be
a Reeb chord of $\Lambda_+$, and let $u_0 \in \mathcal{M}_{s,\Lambda_+}(a, b_1\cdots
b_m)$, $v_0 \in \mathcal{M}_L(a,c_1\cdots c_l)$ be rigid disks. Also pick
rigid disks $v_i \in \mathcal{M}_{L}(b_i, b^i_1\cdots b^i_{m_i})$ for
$i=1,\dotsc,m$, and $u_j \in \mathcal{M}_{s,\Lambda_-}(c_j,f_1\cdots f_k)$ for
some $j \in\{1,\dotsc,l\}$, such that we can interpret $u_0
\#_{i=1}^m v_i$ and $v_0 \# u_j$ as broken boundary components of a
1-dimensional component $\mathcal{M}\subset \mathcal{M}_L$. In particular, we assume that
\begin{equation*}
b^1_1\cdots b^1_{m_1} b_1^2\cdots b^2_{m_2}\cdots b^m_1 \cdots b_{m_m}^m = c_1\cdots
c_{j-1} f_1\cdots f_k c_{j+1} \cdots c_l.
\end{equation*}
Let $\bar{\partial}_T$ denote the dbar-problem we get when we glue the fully
capped $\bar{\partial}_{\hat u_0}$-problem to the fully capped $\bar{\partial}_{\hat
v_i}$-problems, $i=1,\dotsc,m$. This gives the same problem as if
we first glue $u_0$ to $v_1,\dotsc, v_m$, then glue the capping
operators to the punctures of this new problem, and then glue
this to the glued capping disks at $b_1,\dotsc,b_m$.
We also have to analyze the situation at the other boundary component of
$\mathcal{M}$. To that end, let $\bar{\partial}_{\tilde T}$ denote the dbar-problem that we get
if we glue the fully capped $\bar{\partial}_{\hat v_0}$-problem to the fully
capped $\bar{\partial}_{\hat u_j}$-problem. This is the same dbar-problem as we get
if we first glue $v_0$ to $u_j$, then glue the capping
operators to the punctures of this new problem, and then glue this fully capped problem to the glued capping
disk at $c_j$.
Now, following the arguments in the proof of Lemma \ref{lma:lma3}, we first calculate the difference in
orientation of $\det \bar{\partial}_T$, induced by the 2 different gluings
described above, and then we do the same thing for $\det \bar{\partial}_{\tilde T}$. If
these differences cancel when we take the orientation of the
space of conformal variations associated to $\mathcal{M}$ into account, then we
get that $\Phi_L \circ \partial_+ = \partial_- \circ \Phi_L$.
From the gluings for $\bar{\partial}_T$ we get the following two exact
sequences. To simplify notation, we assume that $u_i$, $v_j$, $i=0,\dotsc,l$, $j=0,\dotsc,m$,
all have at least 2 negative punctures.
The first exact sequence has the form
\begin{equation}\label{eq:glu1}
\operatorname{Ker}\dbar_T \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{u_0}\\
\operatorname{Ker}\dbar_{a,+}\\
\operatorname{Ker}\dbar_{b_m,-}\\
\vdots \\
\operatorname{Ker}\dbar_{b_1,-}
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{v_m}\\
\operatorname{Ker}\dbar_{b_m,+}\\
\operatorname{Ker} \operatorname{cap}(v_m,-)
\end{pmatrix}\\
\vdots\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{v_1}\\
\operatorname{Ker}\dbar_{b_1,+}\\
\operatorname{Ker} \operatorname{cap}(v_1,-)
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar_{u_0}\\
\operatorname{Coker}\dbar_{a,+}\\
\operatorname{Coker}\dbar_{b_m,-}\\
\vdots \\
\operatorname{Coker}\dbar_{b_1,-}
\end{pmatrix}\\
{\mathbb R}_{t,m}\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{v_m}\\
\operatorname{Coker}\dbar_{b_m,+}\\
\operatorname{Coker} \operatorname{cap}(v_m,-)
\end{pmatrix}\\
\vdots\\
{\mathbb R}_{t,1}\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{v_1}\\
\operatorname{Coker}\dbar_{b_1,+}\\
\operatorname{Coker} \operatorname{cap}(v_1,-)
\end{pmatrix}
\end{bmatrix}
\to
\operatorname{Coker}\dbar_T,
\end{equation}
where $\operatorname{Ker} \operatorname{cap}(v_i,-) = \hat \oplus_{j=1}^{m_i} \operatorname{Ker}\dbar_{b^i_j,-}$, and
similar for the cokernel, and where we use the notation ${\mathbb R}_{t,i}$ to
indicate the cokernel which is born when we glue the $\hat v_i$-disk to
the bigger problem.
Similarly, the second gluing sequence has the form
\begin{equation}
\label{eq:glu2}
\operatorname{Ker}\dbar_T \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{u_0}\\
\operatorname{Ker}\dbar_{v_m}\\
\vdots\\
\operatorname{Ker}\dbar_{v_1}\\
\operatorname{Ker}\dbar_{a,+}\\
\operatorname{Ker} \operatorname{cap}(v_{m},-)\\
\vdots\\
\operatorname{Ker} \operatorname{cap}(v_1,-)
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{b_m,+}\\
{\mathbb R}_t\\
\operatorname{Ker}\dbar_{b_m,-}
\end{pmatrix}\\
\vdots\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{b_1,+}\\
{\mathbb R}_t\\
\operatorname{Ker}\dbar_{b_1,-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar_{u_0}\\
{\mathbb R}_{t,m}\\
\operatorname{Coker}\dbar_{v_m}\\
\vdots\\
{\mathbb R}_{t,1}\\
\operatorname{Coker}\dbar_{v_1}\\
\operatorname{Coker}\dbar_{a,+}\\
\operatorname{Coker} \operatorname{cap}(v_{m},-)\\
\vdots\\
\operatorname{Coker} \operatorname{cap}(v_1,-)
\end{pmatrix}\\
{\mathbb R}_t\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{b_m,+}\\
\operatorname{Coker}\dbar_{b_m,-}
\end{pmatrix}\\
\vdots\\
{\mathbb R}_t\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{b_1,+}\\
\operatorname{Coker}\dbar_{b_1,-}
\end{pmatrix}
\end{bmatrix}
\to \operatorname{Coker}\dbar_T.
\end{equation}
Again we have omitted the maps $0 \to \operatorname{Ker}\dbar_T$ and $\operatorname{Coker}\dbar_T \to 0$ in \eqref{eq:glu1} and \eqref{eq:glu2} to fit the sequences within page margins.
Perform rearrangements similar to those in the proof of Lemma \ref{lma:lma3}. That is, first notice
that by construction of the gluing map, each ${\mathbb R}_t$-summand in
the second column in \eqref{eq:glu2} is mapped by the identity to
the corresponding ${\mathbb R}_t$-summand in the third column. That is, the
${\mathbb R}_t$-summand in the kernel of the glued capping disk at $b_i$ is mapped by
the identity to the ${\mathbb R}_t$-summand that corresponds to the cokernel
which is born when we glue the glued capping disk at $b_i$ to the larger
problem. Thus,
using that we require the capping operators to satisfy \ref{c3} we
can remove all these occurrences of ${\mathbb R}_t$ from the sequence
\eqref{eq:glu2} without changing any induced orientation. This is done
by moving all these space to the bottom of column 2 and 3,
respectively, and recalling our orientation convention \eqref{eq:o2}.
Next we rearrange vector spaces in the sequence \eqref{eq:glu1}, and we start with letting $\operatorname{Ker}\dbar_{b_i,-}$ switch position with $\operatorname{Ker} \operatorname{cap}(v_i,-)$ for
$i=m,\dotsc,1$. By \ref{c3} we have that $\dim \operatorname{Ker} \operatorname{cap}(v_i,-) = m_i$, so that we get that this rearrangement costs $(-1)^{\sigma_1}$,
\begin{align*}
\sigma_1 &= (m_m +1)\cdot(m-1) +m_m + \dotsc + (m_1+1)\cdot(m-1)+
m_1\\
&\equiv m \cdot \sum_{i=1}^m m_i + m(m+1) \equiv m \cdot \sum_{i=1}^m m_i \pmod{2}.
\end{align*}
Next we move ${\mathbb R}_{t,m}$ and $\operatorname{Coker}\dbar_{v_m}$ to the place just
below $\operatorname{Coker}\dbar_{u_0}$. Using a calculation similar to \eqref{eq:forenk} we see that this costs $(-1)^\nu$,
\begin{align}\nonumber
\nu &= \dim {\mathbb R}_{t,m}\cdot \dim {\mathbb R}^{n+d_A} + ( \dim {\mathbb R}_{t,m} + \dim \operatorname{Coker}\dbar_{v_m})\\ \nonumber
&\quad \cdot\left(\dim {\mathbb R}^{n+d_A} + \dim \left(\hat{\bigoplus}_{i=1}^{m} \operatorname{Coker}\dbar_{b_{i},-}\right) + \dim \operatorname{Coker}\dbar_{a,+}\right)\\ \label{eq:sehar}
&\equiv n + d_A + (1+m_m)\cdot(n+d_A+n+d_A) \equiv n +d_A \pmod{2}.
\end{align}
Perform the similar move with the other such spaces, so that we end up with
\begin{equation}
\label{eq:cokernelar}
\operatorname{Coker}\dbar_{u_0} \oplus {\mathbb R}_{t,m} \oplus \operatorname{Coker}\dbar_{v_m} \oplus \dotsc \oplus
{\mathbb R}_{t,1} \oplus \operatorname{Coker}\dbar_{v_1}
\end{equation}
at the top of the third column in the sequence \eqref{eq:glu1}. These
moves have a total cost $(-1)^{\sigma_2}$,
\begin{equation*}
\sigma_2 = m \cdot(n+d_A)+ \sum_{i=1}^m(m-i)(m_i+1) \equiv m \cdot(n+d_A) + m \cdot \sum_{i=1}^m m_i + m + \sum_{i=1}^mi\cdot(m_i+1),
\end{equation*}
modulo 2.
Indeed, to move ${\mathbb R}_{t,i}$ and $\operatorname{Coker}\dbar_{v_{i}}$ costs $(-1)^{n+d_A+(m-i)(m_i+1)}$. Here $n +d_A$ is calculated just as in \eqref{eq:sehar}, and $(m-i)(m_i+1)$ comes from moving ${\mathbb R}_{t,i}\oplus \operatorname{Coker}\dbar_{v_{i}}$ over $m-i$ copies of ${\mathbb R}^{n+d_A}$ and over the cokernels of the capping operators associated to $v_m,v_{m-1},\dotsc v_{i+1}$. Here we have used an equation similar to \eqref{eq:forenk}, namely that by \ref{c3}, \eqref{eq:dim2} and again \ref{c3} we have that
\begin{equation}\label{eq:forkob}
\dim \operatorname{Coker}\dbar_{b_i,-} \equiv |b_i| \equiv \sum_{j=1}^{m_i}|b_j^i| \equiv \dim \operatorname{Coker}
\operatorname{cap}(v_i,-) \pmod{2}.
\end{equation}
In the last step we switch position of $\operatorname{Coker}\dbar_{b_i,-}$ and $\operatorname{Coker}
\operatorname{cap}(v_i,-)$ for $i=m,\dotsc,1$. Using \eqref{eq:forkob}, we see that this move costs $(-1)^{\sigma_3}$,
\begin{align*}
\sigma_3 &= \dim \left(\operatorname{Coker}\dbar_{b_i,-} \oplus \operatorname{Coker} \operatorname{cap}(v_i,-) \right) \cdot E + \dim \operatorname{Coker}\dbar_{b_i,-} \cdot \dim \operatorname{Coker} \operatorname{cap}(v_i,-) \\
&\equiv 0 \cdot E + \sum_{i=1}^m |b_i|^2 \equiv \sum_{i=1}^m |b_i|.
\end{align*}
Here $E$ is an expression involving dimensions of cokernels of capping operators and of a number of copies of ${\mathbb R}^{n+d_A}$.
Thus, in total we get a permutation sign $(-1)^\sigma$,
\begin{align} \nonumber
\sigma &= \sigma_1 + \sigma_2+ \sigma_3 \equiv m\cdot(n+d_A) + m + \sum_{i=1}^mi\cdot(m_i+1) + \sum_{i=1}^m
|b_i|\\
&\equiv \sum_{i=1}^mi\cdot(m_i+1) + \sum_{i=1}^m(|b_i|+n+d_A+
1) \pmod{2}. \label{eq:sigsig}
\end{align}
(Recall that we assume that $m_i>1$ so that $\operatorname{Ker}\dbar_{v_i}=0$ for $i=1,\dotsc,m$.)
Now we do the similar thing for the $\bar{\partial}_{\tilde T}$-problem, which gives a
total permutation sign $(-1)^{\tilde \sigma}$, where
\begin{equation}\label{eq:tildesigma}
\tilde \sigma = (|c_j| + n +d_A+1) + j\cdot(k+1) + k + l + \sum_{i=1}^{j-1}|c_i|.
\end{equation}
Indeed, in this case we get the following two different gluing sequences
\begin{equation}\label{eq:glu21}
\operatorname{Ker}\dbar_{\tilde T} \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{v_0}\\
\operatorname{Ker}\dbar_{a,+}\\
\operatorname{Ker}\dbar_{c_l,-}\\
\vdots \\
\operatorname{Ker}\dbar_{c_j,-}\\
\vdots \\
\operatorname{Ker}\dbar_{c_1,-}
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{u_j}\\
\operatorname{Ker}\dbar_{c_j,+}\\
\operatorname{Ker} \operatorname{cap}(u_j,-)
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar_{v_0}\\
\operatorname{Coker}\dbar_{a,+}\\
\operatorname{Coker}\dbar_{c_l,-}\\
\vdots \\
\operatorname{Coker}\dbar_{c_j,-}\\
\vdots \\
\operatorname{Coker}\dbar_{c_1,-}
\end{pmatrix}\\
{\mathbb R}_{t}\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{u_j}\\
\operatorname{Coker}\dbar_{c_j,+}\\
\operatorname{Coker} \operatorname{cap}(u_j,-)
\end{pmatrix}
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{\tilde T},
\end{equation}
and
\begin{equation}
\label{eq:glu22}
\operatorname{Ker}\dbar_{\tilde T} \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{v_0}\\
\operatorname{Ker}\dbar_{u_j}\\
\operatorname{Ker}\dbar_{a,+}\\
\operatorname{Ker}\dbar_{c_l,-}\\
\vdots\\
\operatorname{Ker}\dbar_{c_{j+1},-}\\
\operatorname{Ker} \operatorname{cap}(u_{j},-)\\
\operatorname{Ker}\dbar_{c_{j-1},-}\\
\vdots\\
\operatorname{Ker}\dbar_{c_1,-}
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{c_j,+}\\
{\mathbb R}_t\\
\operatorname{Ker}\dbar_{c_j,-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar_{v_0}\\
{\mathbb R}_{t}\\
\operatorname{Coker}\dbar_{u_j}\\
\operatorname{Coker}\dbar_{a,+}\\
\operatorname{Coker}\dbar_{c_l,-}\\
\vdots\\
\operatorname{Coker}\dbar_{c_{j+1},-}\\
\operatorname{Coker} \operatorname{cap}(u_{j},-)\\
\operatorname{Coker}\dbar_{c_{j-1},-}\\
\vdots\\
\operatorname{Coker}\dbar_{c_1,-}
\end{pmatrix}\\
{\mathbb R}_t\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar_{c_j,+}\\
\operatorname{Coker}\dbar_{c_j,-}
\end{pmatrix}
\end{bmatrix}
\to \operatorname{Coker}\dbar_{\tilde T}.
\end{equation}
Here $\operatorname{Ker} \operatorname{cap}(u_j,-) = \hat \oplus_{i=1}^{k} \operatorname{Ker}\dbar_{f_i,-}$, and
similar for the cokernel. Again we have omitted the maps $0 \to \operatorname{Ker}\dbar_{\tilde T}$ and $\operatorname{Coker}\dbar_{\tilde T} \to 0$ to fit the sequences within page margins.
To compare these two sequences we do the usual rearrangements. First consider the sequence \eqref{eq:glu22}, and move the ${\mathbb R}_t$-summand in the second column and the lowest ${\mathbb R}_t$-summand in the third column to the bottom. This costs nothing, just like in the case for the $\bar{\partial}_T$-problem.
Next consider the sequence \eqref{eq:glu21}. Begin by moving $\operatorname{Ker}\dbar_{u_j}$ to the place just below $\operatorname{Ker}\dbar_{v_0}$. Again we assume that both disks have more than one negative puncture, so that $\operatorname{Ker}\dbar_{u_j}$ is $1$-dimensional. Hence this costs $(-1)^{\sigma_0}$,
\begin{equation*}
\sigma_0 = l.
\end{equation*}
Here we have as usual used \ref{c3}.
Then switch places of $\operatorname{Ker}\dbar_{c_j,-}$ and $\operatorname{Ker} \operatorname{cap}(u_j,-)$, which costs $(-1)^{\sigma_1}$,
\begin{equation*}
\sigma_1 = kj + j + 1.
\end{equation*}
This is calculated similar to how $\sigma_1$ was calculated in the $\bar{\partial}_T$-case.
Next we move ${\mathbb R}_t \oplus \operatorname{Coker}\dbar_{u_j}$ to the place just below $\operatorname{Coker}\dbar_{v_0}$. This costs $(-1)^{\sigma_2}$,
\begin{equation*}
\sigma_2 = n + d_A + k+1,
\end{equation*}
and is calculated similar to $\sigma_2$ in the $\bar{\partial}_T$-case.
Finally we switch position of $\operatorname{Coker}\dbar_{c_j,-}$ and $\operatorname{Coker} \operatorname{cap} (u_j,-)$, with a cost of $(-1)^{\sigma_3}$,
\begin{equation*}
\sigma_3 = 1+\sum_{i=1}^{j} |c_i|.
\end{equation*}
The reason for the extra $1$ here compared with the $\bar{\partial}_T$-calculation of $\sigma_3$ is that we now have $\dim \operatorname{Coker}\dbar_{c_j,-} \equiv |c_j| \equiv \sum_{i=1}^{k}|f_i|+1\equiv \dim \operatorname{Coker}
\operatorname{cap}(u_j,-)+1$, modulo $2$.
We see that all this sum up to a total sign $(-1)^{\tilde \sigma}$, given in \eqref{eq:tildesigma}.
Now it remains to calculate the contribution from the orientation of
the space of conformal variations. To that end, let $\mu_i$ be the capping sign of $v_i$, $i=0,\dotsc, m$, and $\epsilon_i$ the capping sign of $u_i$, $i=0,j$. To finish the proof of the theorem we copy the arguments from the proof of Lemma \ref{lma:lma3}. That is, we add together the signs we get from the maps in the commutative diagrams corresponding to the diagram \eqref{eq:twostar}. In the case of the $\bar{\partial}_{u_j\# v_0 }$-problem this diagram now reads
\begin{equation}\label{eq:twostartilde}
\xymatrix{
T\mathcal{C}_{v_0}\oplus T\mathcal{C}_{u_j} \oplus {\mathbb R} \ar[d] \ar[r] & T\mathcal{C}_{u_j\#v_0} \ar[d] \\
\operatorname{Coker}\dbar_{v_0} \oplus \operatorname{Coker}\dbar_{u_j} \ar[r] & \operatorname{Coker}\dbar_{u_j\#v_0}.
}
\end{equation}
We assume that the vector spaces are oriented as usual; the cokernels are given their capping orientations, the spaces of conformal variations are given their fixed orientation from Section \ref{sec:confvar}, and the ${\mathbb R}$-summand is given orientation as the outward normal from the space $T\mathcal{C}_{u_j \# v_0}$.
It follows that the left vertical map restricted to $T\mathcal{C}_{v_0}\oplus T\mathcal{C}_{u_j}$ is an isomorphism of sign
\begin{equation*}
\mu_0\epsilon_j,
\end{equation*}
and from Lemma \ref{lma:confvar} we get that the upper horizontal map is an isomorphism of sign
\begin{equation*}
(-1)^{(k-1)j +1}.
\end{equation*}
It remains to compute the sign of
the lower horizontal map. We claim that this is an isomorphism of sign $(-1)^{\sigma_{\tilde T}}$
\begin{align*}
\sigma_{\tilde T} = \tilde \sigma + \tilde \sigma_0+ k +(|c_j| + n +d_A+1).
\end{align*}
Here $\tilde \sigma_0$ is the sign that corresponds to the sign $\sigma_0$ in \eqref{eq:cokersign} and will be computed below, $k$ comes from Definition \ref{def:capping} applied to the capping orientation of $u_j$, and $(|c_j| + n +d_A+1)$ comes from the fact that the orientation of the glued capped disk at $c_j$ is given by $(-1)^{(|c_j| + n +d_A+1)}$ times the canonical orientation, see Definition \ref{def:capor}.
Now we argue that $\tilde \sigma_0$ is given by
\begin{equation*}
\tilde \sigma_0 = \dim \operatorname{Coker}\dbar_{u_j} = k.
\end{equation*}
Indeed, the gluing sequence that corresponds to \eqref{eq:ny1} now reads
\begin{equation*}
0 \to
\operatorname{Ker}\dbar_{u_j \# v_0}
\to
\begin{bmatrix}
\operatorname{Ker}\dbar_{v_0} \\
\operatorname{Ker}\dbar_{u_j} \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{v_0} \\
{\mathbb R}_t \\
\operatorname{Coker}\dbar_{u_j} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{u_j \# v_0}
\to 0,
\end{equation*}
which moreover can be simplified to
\begin{equation}\label{eq:simpa}
0 \to
{\mathbb R}_t
\to
\begin{bmatrix}
0\\
{\mathbb R}_t
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{v_0} \\
{\mathbb R}_t \\
\operatorname{Coker}\dbar_{u_j} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{u_j \# v_0}
\to 0,
\end{equation}
and which in turn can be reduced to
\begin{equation*}
0 \to
0
\to
0\\
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{v_0} \\
\operatorname{Coker}\dbar_{u_j} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_{u_j \# v_0}
\to 0,
\end{equation*}
with a cost of $(-1)^{\sigma_0}$, using that the leftmost non-trivial map in \eqref{eq:simpa} can be seen as given by $t \mapsto (0,t)$, and the map between the third and fourth column can be seen as given by $s \mapsto (0,s,0)$.
So what we get is that
\begin{align*}
\sigma_{\tilde T} = \tilde \sigma + \tilde \sigma_0+ k +(|c_j| + n +d_A+1) \equiv j\cdot(k+1) + l + k + \sum_{i=1}^{j-1}|c_i| \pmod{2}.
\end{align*}
Hence it follows that the orientation of the boundary component of $\mathcal{M}$ that corresponds to the $\bar{\partial}_{u_j\# v_0 }$-problem is given by
\begin{equation}\label{eq:signtilde}
(-1)^{\sigma_{\tilde T }+ (k-1)j +1}\mu_0\epsilon_j = (-1)^{ 1+ l + k + \sum_{i=1}^{j-1}|c_i|}\mu_0\epsilon_j
\end{equation}
times the outward normal.
The situation for the $\bar{\partial}_T$-problem is slightly more involved. To simplify notation let $\bar{\partial}_h = \bar{\partial}_{v_1\#\dotsm \# v_m \#u_0}$.
To compute the sign that corresponds to $\sigma_0$ in \eqref{eq:cokersign}, one should consider the iterated gluing sequence
\begin{equation*}
0 \to
\operatorname{Ker}\dbar_{h}
\to
\begin{bmatrix}
\operatorname{Ker}\dbar_{u_0} \\
\operatorname{Ker}\dbar_{v_m} \\
\vdots\\
\operatorname{Ker}\dbar_{v_1} \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{u_0} \\
{\mathbb R}_{t,m}\\
\operatorname{Coker}\dbar_{v_m} \\
\vdots\\
{\mathbb R}_{t,1}\\
\operatorname{Coker}\dbar_{v_1} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_h
\to 0,
\end{equation*}
which simplifies to
\begin{equation}\label{eq:simpat}
0 \to
0
\to
\begin{bmatrix}
{\mathbb R}_t \\
0 \\
\vdots\\
0 \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{u_0} \\
{\mathbb R}_{t,m}\\
\operatorname{Coker}\dbar_{v_m} \\
\vdots\\
{\mathbb R}_{t,1}\\
\operatorname{Coker}\dbar_{v_1} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_h
\to 0,
\end{equation}
and which we can reduce to
\begin{equation*}
0 \to
0
\to
\begin{bmatrix}
0 \\
0 \\
\vdots\\
0 \\
\end{bmatrix}
\to
\begin{bmatrix}
\operatorname{Coker}\dbar_{u_0} \\
\operatorname{Coker}\dbar_{v_m} \\
{\mathbb R}_{t,m}\\
\operatorname{Coker}\dbar_{v_{m-1}} \\
\vdots\\
{\mathbb R}_{t,2}\\
\operatorname{Coker}\dbar_{v_1} \\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_h
\to 0,
\end{equation*}
if we compensate by the sign $(-1)^{\nu_0}$,
\begin{equation*}
\nu_0 = 1+\sum_{i=1}^m m_i.
\end{equation*}
Here $\sum_{i=1}^m m_i$ comes from moving the ${\mathbb R}_{t,i}$-summand over $\operatorname{Coker}\dbar_{v_i}$ for $i=1,\dotsc, m$, and the $1$ comes from removing the ${\mathbb R}_t$ and the ${\mathbb R}_{t,1}$ summand together with the fact that the leftmost non-trivial map in \eqref{eq:simpat} can now be seen as given by $t \mapsto (0,-t,0,\dotsc,-t,0)$.
To compute the orientation of the boundary component of $\mathcal{M}$ that corresponds to the $\bar{\partial}_h$-problem we consider the commuting diagram
\begin{equation}\label{eq:trestar}
\xymatrix{
T\mathcal{C}_{u_0} \oplus T\mathcal{C}_{v_m} \oplus {\mathbb R} \oplus \dotsc \oplus T\mathcal{C}_{v_{1}} \oplus {\mathbb R} \ar[d] \ar[r] & T\mathcal{C}_{w\#u_0} \ar[d] \\
\operatorname{Coker}\dbar_{u_0} \oplus \operatorname{Coker}\dbar_{v_m} \oplus {\mathbb R} \oplus \dotsc \oplus \operatorname{Coker}\dbar_{v_2} \oplus {\mathbb R} \oplus \operatorname{Coker}\dbar_{v_{1}} \ar[r] & \operatorname{Coker}\dbar_{w \#u_0}
}
\end{equation}
Here $w =v_1\#\dotsm \# v_m$.
We assume that the vector spaces are oriented as usual; the cokernels are given their capping orientations, the spaces of conformal variations are given their fixed orientation from Section \ref{sec:confvar}, and the ${\mathbb R}$-summand to the right of $T\mathcal{C}_{v_{i}}$ is given orientation as the outward normal from the space $T\mathcal{C}_{v_i\#\dotsm \# v_m \#u_0}$ for $i=1, \dotsc, m$. The ${\mathbb R}$-summands in the lower row are given orientations from ${\mathbb R}_t$.
The left vertical map is given by
\begin{equation*}
(x,y_m,t_m,\dotsc,y_1, t_1) \mapsto (\bar\Psi_0(x), \bar\Psi_m(y_m),t_m, \dotsc, t_{2}, \bar\Psi_1(y_1)).
\end{equation*}
Here $\bar\Psi_i$ is the map \eqref{eq:psibar} associated to $v_i$, $i=1, \dotsc,m$, and to $u_0$ for $i=0$. Using the fact that the gluing cokernel which occurs in the ${\mathbb R}_t$-direction can be interpreted as the outward normal of the space of conformal variations corresponding to the glued problem,
we get that this map restricted to the complement of the last ${\mathbb R}$-summand is an isomorphism of sign
\begin{equation*}
\epsilon_0\mu_1\dotsm\mu_m.
\end{equation*}
The upper horizontal map is an isomorphism of sign $(-1)^{\nu_2}$,
\begin{equation*}
\nu_2 = ((m_m-1) m +1)+((m_{m+1}-1) (m-1) +1) +\dotsc +
((m_1-1) 1 + 1 ) \equiv m + \sum_{i=1}^m i\cdot(m_i+1),
\end{equation*}
modulo 2. To see this, we inductively note that by Lemma \ref{lma:confvar} we have that the inclusion of $T\mathcal{C}_{v_{i+1}\#\dotsm \# v_m \#u_0} \oplus T\mathcal{C}_{v_i} \oplus {\mathbb R} $ into $T\mathcal{C}_{v_i\#\dotsm \# v_m \#u_0}$ is an isomorphism of sign $(-1)^{(m_i-1)i + 1}$.
Finally, we have that the lower horizontal map is an isomorphism of sign $(-1)^{\sigma_T}$,
\begin{equation*}
\sigma_T = \sigma + \nu_0 + m + \sum_{i=1}^m(|b_i|+n+d_A+
1) \equiv \sum_{i=1}^mi\cdot(m_i+1) +1+\sum_{i=1}^m m_i + m.
\end{equation*}
Here $m$ comes from Definition \ref{def:capping} applied to the capping orientation of $u_0$, and $\sum_{i=1}^m(|b_i|+n+d_A+1)$ comes from the fact that the orientation of the glued capped disk at $b_i$ is given by $(-1)^{(|b_i| + n +d_A+1)}$ times the canonical orientation for $i = 1, \dotsc,m$. Recall that $\sigma$ was computed in \eqref{eq:sigsig}.
By the commutativity of the diagram \eqref{eq:trestar},
and since the right vertical map is the one we use to get the orientation of the one-dimensional moduli space, it follows that the orientation of the boundary component of $\mathcal{M}$ that corresponds to the $\bar{\partial}_h$-problem is given by
\begin{align*}\label{eq:signt}
(-1)^{\sigma_{T}+ \nu_2}\epsilon_0\mu_1\dotsm\mu_m & =
(-1)^{ \sum_{i=1}^m i\cdot(m_i+1) +1+\sum_{i=1}^m m_i + m + m + \sum_{i=1}^m i\cdot(m_i+1)} \epsilon_0\mu_1\dotsm\mu_m\\
& = (-1)^{l+k } \epsilon_0\mu_1\dotsm\mu_m
\end{align*}
times the outward normal. Here we have used that $\sum_{i=1}^m m_i = l+k -1$.
Since the orientation of $\mathcal{M}$ at the boundary component that corresponds to the $\bar{\partial}_h$-problem is opposite to the orientation at the boundary component of $\mathcal{M}$ that corresponds to the $\bar{\partial}_{u_j\# v_0 }$-problem, it follows that
\begin{equation*}
(-1)^{l+k } \epsilon_0\mu_1\dotsm\mu_m = - (-1)^{ 1+ l + k + \sum_{i=1}^{j-1}|c_i|}\mu_0\epsilon_j.
\end{equation*}
Since $(-1)^{ \sum_{i=1}^{j-1}|c_i|}$ comes from the Leibniz rule it follows that $ \Phi_L \circ \partial_+ = \partial_- \circ \Phi_L$, and we have proved the theorem.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:differentcappings}]
From Remark \ref{rmk:notv}, Lemma \ref{lma:lma3} and the proof of invariance in [\cite{orientbok}, Section 4.3] we get that there exists an admissible system $\mathcal{S}$
of capping operators satisfying the first statement in the
theorem. Now, if $\mathcal{S}'$ is another admissible system of capping
operators, then lift both systems to the
symplectization, as described in Section \ref{sec:orientcap}.
In this setting, the result about the DGA-isomorphism follows by exactly the same arguments as above, with $L = {\mathbb R}
\times \Lambda$ and where $\Lambda_+ = \Lambda$ is equipped with the
system $\mathcal{S}$ and $\Lambda_- = \Lambda$ is equipped with the
system $\mathcal{S'}$. The DGA-isomorphism is then given by
\begin{equation*}
\Phi_{\mathcal{S}, \mathcal{S'}}(a) = \sigma(u_a)a,
\end{equation*}
where $a$ is a Reeb chord of $\Lambda$, $u_a = {\mathbb R} \times a$, and
$\sigma(u_a)$ is the capping sign of $u_a$ with respect to the induced
system of $L$.
Thus, any admissible system gives rise to a DGA which is isomorphic to
the system in \cite{orientbok}, and the result follows.
\end{proof}
\subsection{Proof of Theorem \ref{thm:functorial}}
In this section we prove that $\Phi$ satisfies the functorial
properties stated in Theorem \ref{thm:functorial}. To that end, if $L
= {\mathbb R} \times \Lambda$ and we wish to prove (\ref{eq:funct2}), note that we shall assume that the capping
system $\mathcal{S}$ of $\Lambda_+ = \Lambda$ equals that of $\Lambda_-=
\Lambda$, which in turn equals the induced system of $L$.
To prove (\ref{eq:funct1}), we need to have the following
setup. Assume that $L_1$ is an exact Lagrangian cobordism from
$\Lambda_0$ to $\Lambda_1$, and that $L_2$ is an exact Lagrangian
cobordism from $\Lambda_1$ to $\Lambda_2$. Assume that $L_1$ and $L_2$
are equipped with spin structures such that the induced spin structure
on $\Lambda_1$ from $L_1$, regarded as the negative boundary of $L_1$,
equals the induced spin structure on $\Lambda_1$ induced from $L_2$, regarded as the positive
boundary of $L_2$. Let $\mathcal{S}_i$ be fixed admissible systems of capping
operators for $\Lambda_i$, $i=0,1,2$, let $\mathcal{S}_{01}$ denote the
induced system on $L_1$, $\mathcal{S}_{12}$ the induced system on $L_2$, and
$\mathcal{S}_{02}$ the induced system on the concatenation $L_1\# L_2$.
\begin{proof}[Proof of Theorem \ref{thm:functorial}]
We must prove that
\begin{equation*}
\Phi_{{\mathbb R} \times \Lambda, \mathcal{S}} = \operatorname{id}, \qquad \Phi_{L_2,\mathcal{S}_{12}}
\circ \Phi_{L_1,\mathcal{S}_{01}} = \Phi_{L_1\#L_2, \mathcal{S}_{02}}.
\end{equation*}
We prove the statement for the identity map. The statement for the
concatenation is similar to the proof of Theorem \ref{thm:kalman} and
details are left to the reader. Indeed, the only thing that has to be checked
in that case is the following. Let $u_0 \in \mathcal{M}_{L_1}(a, b_1\cdots b_m)$, $v_i \in
\mathcal{M}_{L_2}(b_i,b^i_1\cdots b^i_{m_i})
$, $i=1,\cdots, m$, $u_1 \in \mathcal{M}_{L_1\#L_2}(a, b_1^1 \cdots b_{m_1}^1\cdots b_1^m\cdots
b_{m_m}^m)$ be rigid disks. Then one must prove that the gluing
sequence for $\bar{\partial}_{\hat u_0} \#_{i=1}^m \bar{\partial}_{\hat v_i}$ induces the
same orientation on the total glued problem as the gluing sequence we
get when we glue the glued capping disks at $b_1,\dotsc,b_m$, with
respect to the $\mathcal{S}_1$-system, to the
dbar-problem $\bar{\partial}_{\hat u_1}$, capped off with the $\mathcal{S}_{02}$-system. But this follows from similar arguments as
in the proof of Theorem \ref{thm:kalman}.
We turn to the trivial cobordism. Let $a$ be a Reeb chord of
$\Lambda$, let $u = {\mathbb R} \times a$, and assume that $\Phi_{{\mathbb R} \times
\Lambda, \mathcal{S}}(a) = (-1)^\sigma a$. Here the sign $(-1)^\sigma$
satisfies that $\operatorname{Ker} \bar{\partial}_u \simeq {\mathbb R}_t$ is given the capping orientation
$(-1)^\sigma \partial_t$ with respect to the system $\mathcal{S}$. We must prove
that $\sigma \equiv 0 \pmod{2}$.
Consider the situation when we concatenate two trivial
cobordisms. That is, let $L_1 = L_2={\mathbb R}\times \Lambda$, let $u_1$
denote the disk ${\mathbb R} \times a$ of $L_1$, let $u_2$ denote the disk ${\mathbb R}
\times a$ of $L_2$ and let $u$ denote the disk ${\mathbb R} \times a$ in $L_1
\# L_2$. Let $\bar{\partial}^h_{a,+}$ denote the capping operator
associated to $a$, regarded as a Reeb chord of the positive end of
$L_1$, and let $\bar{\partial}^m_{a,-}$ denote the capping operator associated to
$a$, regarded as a Reeb chord of the negative end of $L_1$. Similarly, let
$\bar{\partial}^m_{a,+}, \bar{\partial}^l_{a,-}$ denote the capping operators of $a$
regarded as a Reeb chord of the positive and negative end of $L_2$, respectively.
From the concatenation we get the following exact sequence
\begin{equation}\label{eq:sista0}
0 \to \operatorname{Ker}\dbar_u \to
\begin{bmatrix}
\operatorname{Ker}\dbar_{u_1}\\
\operatorname{Ker}\dbar_{u_2}
\end{bmatrix}
\to
{\mathbb R}_t \to 0 \to 0,
\end{equation}
where all non-trivial spaces are isomorphic to ${\mathbb R}_t$. Notice that this sequence
induces $+{\mathbb R}_t$-orientation of $\operatorname{Ker}\dbar_u$, given that $\operatorname{Ker}\dbar_{u_i}$ is given
the orientation $(-1)^\sigma{\mathbb R}_t$, $i=1,2$.
Now, following our standard arguments, consider the exact gluing
sequence
\begin{equation}\label{eq:sista1}
0\to \operatorname{Ker}\dbar_T \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{u_1}\\
\operatorname{Ker}\dbar^h_{a,+}\\
\operatorname{Ker}\dbar^m_{a,-}
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar_{u_2}\\
\operatorname{Ker}\dbar^m_{a,+}\\
\operatorname{Ker}\dbar^l_{a,-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar^h_{a,+}\\
\operatorname{Coker}\dbar^m_{a,-}
\end{pmatrix}\\
{\mathbb R}_t\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar^m_{a,+}\\
\operatorname{Coker}\dbar^l_{a,-}
\end{pmatrix}\\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_T \to 0.
\end{equation}
By Lemma \ref{lma:kan}, this sequence induces the
canonical orientation of $\det \bar{\partial}_T$, where $T$ is the totally
glued problem, given that the capping operators are given their
capping orientation and $\operatorname{Ker}\dbar_{u_i}$ is given the capping orientation $(-1)^\sigma
{\mathbb R}_t$, $i=1,2$. But, similar to our arguments in the proof of Theorem
\ref{thm:kalman}, the $\bar{\partial}_T$-problem can also be obtained from the
following gluing sequence
\begin{equation}\label{eq:sista2}
0 \to \operatorname{Ker}\dbar_T \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{u_1}\\
\operatorname{Ker}\dbar_{u_2}\\
\operatorname{Ker}\dbar^h_{a,+}\\
\operatorname{Ker}\dbar^l_{a,-}
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar^m_{a,+}\\
{\mathbb R}_t\\
\operatorname{Ker}\dbar^m_{a,-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
{\mathbb R}_t\\
\operatorname{Coker}\dbar^h_{a,+}\\
\operatorname{Coker}\dbar^l_{a,-}
\end{pmatrix}\\
{\mathbb R}_t\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar^m_{a,+}\\
\operatorname{Coker}\dbar^m_{a,-}
\end{pmatrix}\\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_T \to 0.
\end{equation}
We see that the orientation of $\det \bar{\partial}_T$ induced by this
sequence, where we assume that all spaces in columns 2 and 3 are
oriented as in sequence \eqref{eq:sista1}, equals the canonical orientation
times $(-1)^{\sigma_1}$,
\begin{equation}\label{eq:siggi}
\sigma_1 = |a| + n+d_A+1.
\end{equation}
Indeed, this sign we get if we do the
usual rearrangements, comparing \eqref{eq:sista1} with
\eqref{eq:sista2}: we can remove the bottom-most ${\mathbb R}_t$ in columns 2 and
3 in \eqref{eq:sista2} without any cost, then move $\operatorname{Ker}\dbar_{u_2}$ and the
gluing cokernel in \eqref{eq:sista1} to the corresponding position in
\eqref{eq:sista2} with the cost of $(-1)^{1+|a|+|a|+n+d_A+1}$, and then switch
positions of $\operatorname{Ker}\dbar^m_{a,-}$ and $\operatorname{Ker}\dbar^l_{a,-}$, and similar for the
cokernels, with a cost of $(-1)^{1+|a|}$.
Now we return to the original sequence \eqref{eq:sista2}. Using associativity of orientations under gluing (see [\cite{orientbok}, Section 3.2.3]),
we can use the sequence
\eqref{eq:sista0} to replace the pair $(\operatorname{Ker}\dbar_{u_1}\oplus \operatorname{Ker}\dbar_{u_2},
{\mathbb R}_t)$ in \eqref{eq:sista2} by the pair $(\operatorname{Ker}\dbar_u, 0)$. We then get the exact sequence
\begin{equation}\label{eq:sista3}
0 \to \operatorname{Ker}\dbar_T \to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Ker}\dbar_{u}\\
\operatorname{Ker}\dbar^h_{a,+}\\
\operatorname{Ker}\dbar^l_{a,-}
\end{pmatrix}\\
\begin{pmatrix}
\operatorname{Ker}\dbar^m_{a,+}\\
{\mathbb R}_t\\
\operatorname{Ker}\dbar^m_{a,-}
\end{pmatrix}
\end{bmatrix}
\to
\begin{bmatrix}
\begin{pmatrix}
\operatorname{Coker}\dbar^h_{a,+}\\
\operatorname{Coker}\dbar^l_{a,-}
\end{pmatrix}\\
{\mathbb R}_t\\
{\mathbb R}^{n+d_A}\\
\begin{pmatrix}
\operatorname{Coker}\dbar^m_{a,+}\\
\operatorname{Coker}\dbar^m_{a,-}
\end{pmatrix}\\
\end{bmatrix}
\to
\operatorname{Coker}\dbar_T \to 0.
\end{equation}
Moreover, from [\cite{orientbok}, Lemma 3.7] we get that this sequence induces the same orientation on $\det \bar{\partial}_T$ as the sequence \eqref{eq:sista2} does. That is, \eqref{eq:sista3} induces the canonical orientation times $(-1)^{\sigma_1}$, given that the capping operators are given their capping orientation and that $\operatorname{Ker}\dbar_u$ is given the orientation induced by the sequence \eqref{eq:sista0}. Recall that the latter orientation is given by ${\mathbb R}_t$, which differs from the capping orientation of $\operatorname{Ker}\dbar_u$ by a sign $(-1)^\sigma$.
Now we change the orientation of $\operatorname{Ker}\dbar_u$ to its capping orientation $(-1)^\sigma{\mathbb R}_t$, keeping the capping orientation of the capping operators. Then the orientation of $\det \bar{\partial}_T$ induced by \eqref{eq:sista3} should change by a sign $(-1)^\sigma$, to be given by $(-1)^{\sigma_2}$ times the canonical orientation of $\det \bar{\partial}_T$, where
\begin{equation}\label{eq:siggi2}
\sigma_2 = \sigma + \sigma_1 = \sigma+ n + d_A +|a|+1.
\end{equation}
However, by our standard arguments it follows that this orientation should be given by $(-1)^{\sigma_3}$ times the canonical orientation of $\det \bar{\partial}_T$, where
\begin{equation}\label{eq:siggi4}
\sigma_3 = n + d_A +|a|+1.
\end{equation}
This follows from Lemma \ref{lma:kan} together with the fact that
the upper part of \eqref{eq:sista3} corresponds to the capping sequence for $\bar{\partial}_u$ which by assumption is given the canonical orientation, and the lower part of \eqref{eq:sista3} corresponds to the gluing sequence for the capping disk at $a$ which by the assumptions together with Definition \ref{def:capor} is given the canonical orientation times $(-1)^{ n + d_A +|a|+1}$.
Comparing \eqref{eq:siggi2} with \eqref{eq:siggi4} we get that we should have
\begin{equation*}
(-1)^{\sigma+ n + d_A +|a|+1} = (-1)^{ n + d_A +|a|+1},
\end{equation*}
implying that $(-1)^{\sigma} = 1$.
This
concludes the proof of
$\Phi_{{\mathbb R} \times \Lambda, \mathcal{S}} = \operatorname{id}$.
\end{proof}
\subsection{A remark on orientation conventions}\label{sec:clock}
The reason that we choose the convention \eqref{eq:gluconv} instead of the convention
\begin{equation*}
0 \to \operatorname{Ker}\dbar_{A\#B}
\xrightarrow{\alpha}
\begin{bmatrix}
\operatorname{Ker}\dbar_A \\
\operatorname{Ker}\dbar_B
\end{bmatrix}
\xrightarrow{\beta}
\begin{bmatrix}
\operatorname{Coker}\dbar_A \\
\operatorname{Coker}\dbar_B
\end{bmatrix}
\xrightarrow{\gamma}
\operatorname{Coker}\dbar_{A\#B} \to 0
\end{equation*}
from \cite{orientbok}, is that our choice seems to simplify the calculations in the proofs. That is, it is easier to have the ${\mathbb R}_t$-summand of the kernel of the $\bar{\partial}_u$-operator, for $u$ a pseudo-holomorphic disks of ${\mathbb R} \times \Lambda$, as close to the top of the columns in the gluing sequences as possible.
In the same time as we want to use this simplification, we also want to take advantage of the calculations in the proofs of [\cite{orientbok}, Lemma 4.9 and Lemma 4.11]. This forces us to have all exact sequences (except for the ones for the glued capping disks) from \cite{orientbok} ''mirrored'' in the horizontal axis. For example, we have to glue the capping operators in the clockwise direction, starting at the positive puncture, and we also have to choose the orientations of the spaces of conformal variations to be ''opposite'' to the one in \cite{orientbok}.
The reason that we keep the same convention for the gluing sequences for the glued capping disks as in \cite{orientbok} (these are not mirrored!) is to get rid of the sign $(-1)^{(n-1)(|a|+1)}$ in Remark \ref{rmk:dumsign}. This also has to do with the sign in Definition \ref{def:capor}. However, the reason for this sign is mostly to get the calculations in the proof of Theorem \ref{thm:kalman} and Theorem \ref{thm:functorial} to work out well.
\section{Morse flow trees and abstract perturbations}\label{sec:trees}
In \cite{tobhonda} the techniques of abstractly perturbed flow trees are used to give explicit descriptions of DGA-morphisms $\Phi_L$ associated to elementary Legendrian isotopies, with coefficients in ${\mathbb Z}_2$. Here we explain how this can be done also with integer coefficients.
Let $L \subset {\mathbb R} \times J^1(M)$ be an exact Lagrangian cobordism. To this cobordism we associate a \emph{Morse cobordism} $L^{MO}$, which will be exact Lagrangian isotopic to $L$ relative the ends. See [\cite{tobhonda}, Section 2]. The advantage of considering $L^{MO}$ instead of $L$ is that we can use the Morse flow tree techniques from \cite{trees} to define $\Phi_{L^{MO}}$ to be given by a count of rigid flow trees instead of rigid disks.
For a definition of Morse flow trees we refer to [\cite{trees}, Section 2.2] and [\cite{tobhonda}, Section 4]. Briefly, these trees are built out of flow lines of local gradient differences associated to $L^{MO}$, and there is a one-to-one correspondence between the rigid Morse flow trees and the rigid pseudo-holomorphic disks of $L^{MO}$. Thus, if we let $\mathcal{M}_{T, L}(a, {\bf b})$ denote the moduli space of Morse flow trees of $L^{MO}$ with positive puncture $a$ and negative punctures ${\bf b}$, it follows from [\cite{tobhonda}, Lemma 5.12] that $\Phi_L$ can be given by
\begin{equation*}
\Phi_L(a) = \sum_{\dim \mathcal{M}_{T, L}(a, {\bf b}) =0} |\mathcal{M}_{T, L}(a, {\bf b})| {\bf b}.
\end{equation*}
Here $ |\mathcal{M}_{T, L}(a, {\bf b})|$ denotes the algebraic count of
elements in the moduli space, where we have used the oriented
identification of $\mathcal{M}_{T, L}(a, {\bf b})$ and $\mathcal{M}_L(a,{\bf b})$ from
[\cite{orienttrees}, Theorem 1.1]. In summary, the orientation of a tree $\Gamma \in
\mathcal{M}_{T,L}(a,{\bf b})$ is defined by considering the cotangent lift of the
tree, which gives rise to a Lagrangian boundary condition for a corresponding punctured disk, and we get an associated linearized $\bar{\partial}$-operator $\bar{\partial}_\Gamma$. Moreover, we can glue capping operators associated to the punctures of $\Gamma$ to this operator, to get a corresponding fully capped problem $\bar{\partial}_{\hat \Gamma}$ on the non-punctured disk. This gives rise to exactly the same gluing sequences as in the case of true $J$-holomorphic disks, and the \emph{capping orientation of $\Gamma$} is defined completely analogous to how it is done for $J$-holomorphic disks. For details we refer to [\cite{orienttrees}, Section 4].
To get the explicit formulas for the DGA-maps in \cite{tobhonda}, the trees of $L^{MO}$ are perturbed. See [\cite{tobhonda}, Section 6.3]. This is first done by a geometric perturbation, which is a perturbation of $L^{MO}$ together with a perturbation of the Riemannian metric. If we extend the orientation scheme of $\mathcal{M}_{T,L}$ to also be defined for the geometrically perturbed trees, it follows by straightforward arguments that for a generic geometric perturbation, the algebraic count of trees in $\mathcal{M}_{T,L}$ will be equal to the algebraic count of rigid, geometrically perturbed trees. We let $\Phi_{L,g}$ denote the
DGA-map defined by the count of the rigid geometric perturbed trees.
Next we consider an abstract perturbation of the trees, as defined in [\cite{tobkal}, Section 3.3 -- 3.4], compare also with [\cite{tobhonda}, Section 6.3.1]. In [\cite{tobhonda}, Lemma 6.4] it is proven that the induced map $\Phi_{L,a}$, given by a count of rigid, abstractly perturbed flow trees, is a DGA-morphism which is chain homotopic to $\Phi_{L,g}$, given that we are using ${\mathbb Z}_2$-coefficients. We claim that this chain homotopy can be lifted to ${\mathbb Z}$-coefficients.
Indeed, the proof of [\cite{tobhonda}, Lemma 6.4] makes use of a 1-parameter family of abstract perturbations, starting at the geometric perturbation (which we can interpret as an abstract perturbation) and ending at the desired abstract perturbation. The chain homotopy is then defined by a count of certain Morse flow trees of $L^{MO} \times D$, where $D$ is the unit disk and where the trees are induced by the 1-parameter family of perturbations.
From [\cite{tobkal}, Section 3.4] it follows that the boundary conditions of the abstractly perturbed trees are close to being boundary conditions for true trees. Hence we can extend the orientation scheme for flow trees to also include abstractly perturbed trees, both for the trees occurring in the formula for $\Phi_{L,a}$ and also for the flow trees in the chain homotopy just described. Thus it follows that the algebraic count of abstractly perturbed trees given by $\Phi_{L,a}$ is signed chain homotopic to the algebraic count of geometrically perturbed trees given by $\Phi_{L,g}$.
It follows that $\Phi_{L,a}$ is chain homotopic to $\Phi_L$ over ${\mathbb Z}$.
\bibliographystyle{halpha}
|
\section*{Acknowledgements}
The authors thank Allison Bishop for valuable discussions in the early stages of this work.
\newpage
\bibliographystyle{plain}
\section{Introduction}
\ShortOnlyVspace{-2mm}
Communication in the presence of \emph{synchronization errors}, which include both insertions and deletions, is a fundamental problem of practical importance which eluded a strong theoretical foundation for decades. This remained true even while communication in the presence of \emph{half-errors}, which consist of symbol corruptions and erasures, has been the subject of an extensive body of research with many groundbreaking results. Synchronization errors are strictly more general than half-errors, and thus synchronization errors pose additional challenges for robust communication.
In this work, we show that one-way and interactive communication in the presence of synchronization errors can be reduced to the problem of communication in the presence of half-errors. We present a series of efficient channel simulations which allow two parties to communicate over a channel afflicted by synchronization errors as though they were communicating over a half-error channel with only a slightly larger error rate. This allows us to leverage existing coding schemes for robust communication over half-error channels in order to derive strong coding schemes resilient to synchronization errors.
One of the primary tools we use are \emph{synchronization strings}, which were recently introduced by Haeupler and Shahrasbi in order to design essentially optimal error correcting codes (ECCs) robust to synchronization errors \cite{haeupler2017synchronization}. For every $\eps > 0$, synchronization strings allow a sender to index a sequence of messages with an alphabet of size $\eps^{-O(1)}$ in such a way that $k$ synchronization errors are efficiently transformed into $(1+\eps)k$ half-errors for the purpose of designing ECCs. Haeupler and Shahrasbi provide a black-box construction which transforms any ECC into an equally efficient ECC robust to synchronization errors. However, channel simulations and interactive coding in the presence of synchronization errors present a host of additional challenges that cannot be solved by the application of an ECC. Before we describe our results and techniques in detail, we begin with an overview of the well-known interactive communication model.
\begin{comment}
\paragraph{Error correction techniques for symbol corruptions.}
\noindent \textbf{One-way communication.} Error-correcting codes have revolutionized electronic communication and storage, proving to be both practical in application and rich from a theoretical perspective. At their core, error-correcting codes allow a communicating party to add redundancy to her original message so that it may be recovered even if parts of the encoded message have been corrupted. In particular, if the original message consists of $n$ symbols from $\Sigma_{in}$, an error-correcting code transforms this string into a coded message which consists of $\alpha n + o(n)$ symbols from a finite alphabet $\Sigma_{out}$, from which the original message can be recovered even if a $\delta$ fraction of the coded symbols are corrupted, with negligible probability of failure. In this case, we say that the \emph{communication rate} $R$ is $\frac{\log |\Sigma_{in}|}{\alpha \log |\Sigma_{out}|}$, and for a given communication channel, we say that its \emph{channel capacity} $C$ is the supremum of achievable communication rates for large $n$.
Shannon, in his seminal work on error-resilient communication, measured the rate of a binary channel with a $\delta$ fraction of random errors in terms of its \emph{entropy} $H(\delta) = \delta \log \frac{1}{\delta} + (1-\delta)\log \frac{1}{1-\delta}$ \cite{shannon1948mathematical}. Shannon proved that over this type of channel, also known as a \emph{binary symmetric channel}, it is possible to achieve a rate of $R = 1-H(\delta)$. Not long after, Hamming and others showed that for worst-case or adversarial errors, it is still possible to achieve a rate of $R = 1-\Theta(H(\delta))$ \cite{hamming1950error}. In this paper, we will be concerned with this harder worst-case error setting. We remark that determining the constant in the $\Theta(H(\delta))$ factor in this tradeoff between the error fraction and communication rate for a binary code remains a fundamental open problem in coding theory, though many conjecture that the Gilbert-Varshamov bound of $R \leq 1- H(2\delta)$ is essentially tight. After Shannon and Hamming broke ground on error-resilient communication, a wealth of subsequent research culminating in the very active field of coding theory has led to many efficient error correcting codes.
After Shannon and Hamming broke ground on error-resilient communication, a wealth of subsequent research has led to efficient \emph{unique decoding} procedures for many interesting families of codes. In the unique decoding paradigm, if the Hamming distance between any two codewords is at least $d$, then the goal is to be able to decode an arbitrary codeword up to $(d-1)/2$ errors. An alternative model is \emph{list-decoding}, which was introduced by Elias \cite{elias1957list} and Wozencraft \cite{wozencraft1958list}. In this model, the decoding procedure outputs a list of potential codewords, and the procedure is deemed successful if the transmitted string is an element of the list. The first efficient list-decoding procedure was designed for Reed-Solomon codes \cite{sudan1997decoding}, and numerous algorithms have been developed since for many code families.
\end{comment}
\paragraph{Interactive communication.} Throughout this work, we study a scenario where there are two communicating parties, whom we call Alice and Bob. The two begin with some input symbols and wish to compute a function of their input by having a conversation. Their goal is to succeed with high probability while communicating as few symbols as possible. In particular, if their conversation would consist of $n$ symbols in the noise-free setting, then they would like to converse for at most $\alpha n$ symbols, for some small $\alpha$, when in the presence of noise. One might hope that Alice and Bob could correspond using error-correcting codes. However, this approach would lead to poor performance because if a party incorrectly decodes a single message, then the remaining communication is rendered useless. Therefore, only a very small amount of noise could be tolerated, namely less than the amount to corrupt a single message.
There are three major aspects of coding schemes for interactive communication that have been extensively studied in the literature. The first is the coding scheme's \important{maximum tolerable error-fraction} or, in other words, the largest fraction of errors for which the coding scheme can successfully simulate any given error-free protocol.
Another important quality of coding schemes for interactive communication, as with one-way communication, is \important{communication rate}, i.e., the amount of communication overhead in terms of the error fraction.
Finally, the \important{efficiency} of interactive coding schemes have been of concern in the previous work.
Schulman initiated the study of error-resilient interactive communication, showing how to convert an arbitrary two-party interactive protocol to one that is robust to a $\delta=1/240$ fraction of adversarial errors with a constant communication overhead \cite{schulman1992communication,schulman1993deterministic}.
Braverman and Rao increased the bound on the tolerable adversarial error rate to $\delta < 1/4$, also with a constant communication overhead \cite{braverman2014toward}. Brakerski et al.~\cite{brakerski2014} designed the first efficient coding scheme resilient to a constant fraction of adversarial errors with constant communication overhead.
The above-mentioned schemes achieve a constant overhead no matter the level of noise. Kol and Raz were the first to study the trade-off between error fraction and communication rate \cite{kol2013interactive}.
Haeupler then provided a coding scheme with a communication rate of $1-O(\sqrt{\delta \log \log (1/\delta)})$ over an adversarial channel \cite{haeupler2014interactive:FOCS}.
Further prior work has studied coding for interactive communication focusing on communication efficiency and noise resilience~\cite{gelles2015, baverman2014listUnique, haeupler2014optimalII} as well as computational efficiency~\cite{brakerski2012, brakerski2013, brakerski2014, gelles2011, gelles2014, haeupler2014optimalII}. Other works have studied variations of the interactive communication problem~\cite{haeupler2014optimalI, franklin2015, Efremenko2016, agrawal2011, brassard2014}.
The main challenge that \emph{synchronization errors} pose is that they may cause the parties to become ``out of sync.'' For example, suppose the adversary deletes a message from Alice and inserts a message back to her. Neither party will know that Bob is a message behind, and if this corruption remains undetected, the rest of the communication will be useless.
In most state-of-the-art interactive coding schemes for symbol corruptions, the parties communicate normally for a fixed number of rounds and then send back and forth a series of checks to detect any symbol corruptions that may have occurred. One might hope that a synchronization error could be detected during these checks, but the parties may be out of sync while performing the checks, thus rendering them useless as well. Therefore, synchronization errors require us to develop new techniques.
Very little is known regarding coding for interactive communication in the presence of synchronization errors. A 2016 coding scheme by Braverman et al.~\cite{braverman2015coding}, which can be seen as the equivalent of Schulman's seminal result, achieves a small constant communication rate while being robust against a $1/18-\eps$ fraction of errors. The coding scheme relies on edit-distance tree codes, which are a careful adaptation of Schulman's original tree codes \cite{schulman1993deterministic} for edit distance, so the decoding operations are not efficient and require exponential time computations.
A recent work by Sherstov and Wu~\cite{sherstov2017optimal} closed the gap for maximum tolerable error fraction by introducing a coding scheme that is robust against $1/6-\eps$ fraction of errors which is the highest possible fraction of insertions and deletions under which any coding scheme for interactive communication can work.
Both Braverman et al.~\cite{braverman2015coding} and Sherstov and Wu~\cite{sherstov2017optimal} schemes are of constant communication rate, over large enough constant alphabets, and inefficient.
In this work we address the next natural questions which, as arose with ordinary corruption interactive communication, are on finding interactive coding schemes that are computationally efficient or achieve super-constant communication efficiency.
\fullOnly{\subsection{Our results}}
\shortOnly{\paragraph{Our results.}}
We present very general channel simulations which allow two parties communicating over an insertion-deletion channel to follow any protocol designed for a regular symbol corruption channel. The fraction of errors on the simulated symbol corruption channel is only slightly larger than that on the insertion-deletion channel. Our channel simulations are made possible by synchronization strings. Crucially, at the cost of increasing the error rate by a constant factor, synchronization strings can be decoded in a streaming manner which preserves linearity of time.
Note that the similar technique is used in Haeupler and Shahrasbi~\cite{haeupler2017synchronization} to transform synchronization errors into ordinary symbol corruptions as a stepping-stone to obtain insertion-deletion codes from ordinary error correcting codes in a black-box fashion. However, in the context of error correcting codes, there is no requirement for this transformation to happen in real time. In other words, in the study of insertion-deletion codes by Haeupler and Shahrasbi~\cite{haeupler2017synchronization}, the entire message transmission is done and then the receiving party uses the entire message to transform the synchronization errors into symbol corruptions. In the channel simulation problem, this transformation is required to happen on the fly.
Interestingly, we have found out that in the harder problem of channel simulation, the factor by which the number of synchronization errors increase by being transformed into corruption errors cannot be improved to $1+o(1)$, in contrast to what is achievable for error correcting codes. This work exhibits the widespread applicability of synchronization strings and opens up several new use cases, such as coding for interactive communication over insertion-deletion channels. Namely, using synchronization strings, we provide techniques to obtain the following simulations of corruption channels over given insertion-deletion channels with binary and large constant alphabet sizes.
\ShortOnlyVspace{-1mm}
\begin{theorem}(Informal Statement of Theorems~\ref{thm:OnewayCnstSizeAlphaSimul}, \ref{thm:InteractiveCnstSizeAlphaSimul}, \ref{thm:NonObliviousGeneralSimulation}, and \ref{thm:BinaryOneWayCompleteSimulation})
\begin{enumerate
\ShortOnlyVspace{-2mm}
\item[(a)] Suppose that $n$ rounds of a one-way/interactive insertion-deletion channel over an alphabet $\Sigma$ with a $\delta$ fraction of insertions and deletions are given. Using an $\eps$-synchronization string over an alphabet $\Sigma_{syn}$, it is possible to simulate $n\left(1-O_\eps(\delta)\right)$ rounds of a one-way/interactive corruption channel over $\Sigma_{sim}$ with at most $O_\eps\left(n\delta\right)$ symbols corrupted so long as $|\Sigma_{sim}| \times |\Sigma_{syn}| \le |\Sigma|$.
\item[(b)] Suppose that $n$ rounds of a binary one-way/interactive insertion-deletion channel with a $\delta$ fraction of insertions and deletions are given. It is possible to simulate
$n(1-\Theta( \sqrt{\delta\log(1/\delta)}))$
rounds of a binary one-way/interactive corruption channel
with $\Theta(\sqrt{\delta\log(1/\delta)})$ fraction of corruption errors between two parties over the given channel.
\end{enumerate}
\end{theorem}
Based on the channel simulations presented above, we present novel interactive coding schemes which simulate any interactive two-party protocol over an insertion-deletion channel.
We use our large alphabet interactive channel simulation along with constant-rate efficient coding scheme of Ghaffari and Haeupler~\cite{haeupler2014optimalII} for interactive communication over corruption channels to obtain a coding scheme for insertion-deletion channels that is efficient, has a constant communication rate, and tolerates up to $1/44-\eps$ fraction of errors. Note that despite the fact that this coding scheme fails to protect against the optimal $1/6-\eps$ fraction of synchronization errors as the recent work by Sherstov and Wu~\cite{sherstov2017optimal} does, it is an improvement over all previous work
in terms of computational efficiency as it is the first efficient coding scheme for interactive communication over insertion-deletion channels.
\ShortOnlyVspace{-1mm}
\begin{theorem}\label{thm:largeAlphaInteractiveCodingScheme}
For any constant $\eps > 0$ and $n$-round alternating protocol $\Pi$, there is an efficient randomized coding scheme simulating $\Pi$ in presence of $\delta=1/44-\eps$ fraction of edit-corruptions with constant rate (i.e., in $O(n)$ rounds) and in $O(n^5)$ time that works with probability $1-2^{\Theta(n)}$. This scheme requires the alphabet size to be a large enough constant $\Omega_\eps(1)$.
\end{theorem}
\ShortOnlyVspace{-2mm}
Next, we use our small alphabet channel simulation and the corruption channel interactive coding scheme of Haeupler~\cite{haeupler2014interactive:FOCS} to introduce an interactive coding scheme for insertion-deletion channels. This scheme is not only computationally efficient, but also the first with super constant communication rate. In other words, this is the first coding scheme for interactive communication over insertion-deletion channels whose rate approaches one as the error fraction drops to zero.
Our computationally efficient interactive coding scheme achieves a near-optimal communication rate of $1 - O(\sqrt{\delta \log (1/\delta)})$ and tolerates a $\delta$ fraction of errors.
Besides computational efficiency and near-optimal communication rate, this coding scheme improves over all previous work in terms of alphabet size. As opposed to coding schemes provided by the previous work\cite{braverman2015coding, sherstov2017optimal}, our scheme does not require a large enough constant alphabet and works even for binary alphabets.
\ShortOnlyVspace{-1mm}
\begin{theorem}\label{thm:InterFullyAdv}
For sufficiently small $\delta$, there is an efficient interactive coding scheme for fully adversarial binary insertion-deletion channels which is robust against $\delta$ fraction of edit-corruptions, achieves a communication rate of
$1 - \Theta({\sqrt{\delta\log(1/\delta)}})$, and works with probability $1 - 2^{-\Theta(n\delta)}$.
\end{theorem}
We also utilize the channel simulations in one-way settings to provide efficient binary insertion-deletion codes correcting $\delta$-fraction of synchronization errors--for $\delta$ smaller than some constant--with a rate of $1-\Theta(\sqrt{\delta\log(1/\delta)})$. This is an improvement in terms of rate-distance trade-off over the state-of-the-art low-rate binary insertion-deletion codes by Guruswami and Wang~\cite{guruswami2017deletion}. The codes by Guruswami and Wang~\cite{guruswami2017deletion} achieve a rate of $1 - O(\sqrt{\delta} \log(1/\delta))$
Finally, we introduce a slightly improved definition of edit-distance tree codes\footnote{This improved definition is independently observed by Sherstov and Wu~\cite{sherstov2017optimal}.}, a generalization of Schulman's original tree codes defined by Braverman et al.~\cite{braverman2015coding}. We show that under our revised definition, edit-distance tree codes are closely related to synchronization strings. For example, edit-distance tree codes can be constructed by merging a regular tree code and a synchronization string. This transfers, for example, Braverman's sub-exponential time tree code construction~\cite{braverman2012towards} and the candidate construction of Schulman~\cite{Schulman03:Postscript} from tree codes to edit-distance tree codes.
Lastly, as a side note, we will show that with the improved definition, the coding scheme of Braverman et al.~\cite{braverman2015coding} can tolerate $1/10-\eps$ fraction of synchronization errors rather than $1/18-\eps$ fraction that the scheme based on their original definition did.
\fullOnly{\subsection{The Organization of the Paper}
We start by reviewing basic definitions and concepts regarding interactive communication and synchronization strings in Section~\ref{sec:def}. Then we study channel simulations under various assumptions in Section~\ref{sec:simulation}. We use these channel simulations to obtain novel coding schemes for one-way and interactive communication in Sections~\ref{sec:simulation} and \ref{sec:insDelCodes}. Finally, in Section~\ref{sec:SynchAndTreeCodes}, we discuss connections between synchronization strings, tree codes and edit-distance tree codes introduced by Braverman et al.~\cite{braverman2015coding}.}
%
\subsection{Definitions and preliminaries}\label{sec:def}}
\fullOnly{\section{Definitions and preliminaries}\label{sec:def}}
\ShortOnlyVspace{-2mm}
In this section, we define the channel models and communication settings considered in this work. We also provide notation and define synchronization strings.
\paragraph{Error model and communication channels.}
In this work, we study two types of channels, which we call \emph{corruption channels} and \emph{insertion-deletion channels.}
In the corruption channel model, two parties communicate with an alphabet $\Sigma$, and if one party sends a message $c \in \Sigma$ to the other party, then the other party will receive a message $\tilde{c} \in \Sigma$, but it may not be the case that $c = \tilde{c}$.
In the one-way communication setting over an insertion-deletion channel, messages to the listening party may be inserted, and messages sent by the sending party may be deleted. The two-way channel requires a more careful setup. We emphasize that we cannot hope to protect against arbitrary insertions and deletions in the two-way setting because in the message-driven model, a single deletion could cause the protocol execution to ``hang.''
Therefore, following the standard model of Braverman et al.'s work \cite{braverman2015coding}
that is employed in all other previous works on this problem~\cite{sherstov2017optimal}, we restrict our attention to \emph{edit corruptions}, which consist of a single deletion followed by a single insertion, which may be aimed at either party.
Braverman et al. \cite{braverman2015coding} provide a detailed discussion on their model and show that it is strong enough to generalize other natural models one might consider, including models that utilize clock time-outs to overcome the stalling issue.
In both the one- and two-way communication settings, we study \emph{adversarial channels} with \emph{error rate} $\delta$.
Our coding schemes are robust in both the \emph{fully adversarial} and the \emph{oblivious adversary} models. \shortOnly{We provide the definitions of these standard models in Appendix~\ref{app:prelim}.}
\fullOnly{In the fully adversarial model, the adversary may decide at each round whether or not to interfere based on its state, its own randomness, and the symbols communicated by the parties so far. In the oblivious adversary model, the adversary must decide which rounds to corrupt in advance, and therefore independently of the communication history.
A simple example of an oblivious adversary is the \emph{random error channel}, where each round is corrupted with probability $\delta$.
In models we study, there is no pre-shared randomness between the parties.}
\paragraph{Interactive and one-way communication protocols.} In an \emph{interactive protocol} $\Pi$ over a channel with an alphabet $\Sigma$, Alice and Bob begin with two inputs from $\Sigma^*$ and then engage in $n$ \emph{rounds} of communication. In a single round, each party either listens for a message or sends a message, where this choice and the message, if one is generated, depends on the party's state, its input, and the history of the communication thus far. After the $n$ rounds, the parties produce an output. We study \emph{alternating protocols}, where each party sends a message every other round and listens for a message every other round. In this message-driven paradigm, a party ``sleeps'' until a new message comes, at which point the party performs a computation and sends a message to the other party.
\fullOnly{Protocols in the interactive communication literature typically fall into two categories: \emph{message-driven} and \emph{clock-driven}. In the message-driven paradigm, a party ``sleeps'' until a new message comes, at which point the party performs a computation and sends a message to the other party. Meanwhile, in the clock-driven model, each party has a clock, and during a single tick, each party accepts a new message if there is one, performs a computation, and sends a message to the other party if he chooses to do so. In this work, we study the message-driven model, since in the clock-driven model, dealing with insertions and deletions is too easy. After all, an insertion would mean that one symbol is changed to another as in the case of the standard corruption model and a deletion would be detectable, as it would correspond to an erasure.}
In the presence of noise, we say that a protocol $\Pi'$ \emph{robustly simulates} a deterministic protocol $\Pi$ over a channel $C$ if given any inputs for $\Pi$, the parties can decode the transcript of the execution of $\Pi$ on those inputs over a noise-free channel from the transcript of the execution of $\Pi'$ over $C$.
Finally, we also study \emph{one-way communication}, where one party sends all messages and the other party listens. Coding schemes in this setting are known as \emph{error-correcting codes}.
\shortOnly
\paragraph{Synchronization Strings\footnote{See the formal definitions and a review of other technical details in Appendix~\ref{app:prelim}. }.} In short, synchronization strings \cite{haeupler2017synchronization} allow communicating parties to protect against synchronization errors by indexing their messages without blowing up the communication rate.
We describe this technique by introducing two intermediaries, $C_A$ and $C_B$, that conduct the communication over the given insertion-deletion channel. $C_A$ receives all symbols that Alice wishes to send to Bob, $C_A$ sends the symbols to $C_B$, and $C_B$ communicates the symbols to Bob. $C_A$ and $C_B$ handle the synchronization strings and all the extra work that is involved in keeping Alice and Bob in sync by guessing the actual index of symbols received by $C_B$.
In this way, Alice and Bob communicate via $C_A$ and $C_B$ as though they were communicating over a half-error channel.
Unfortunately, trivially attaching the indices $1, 2, \dots, n$ to each message will not allow us to maintain a near optimal communication rate. If $C_A$ attaches an index to each of Alice's messages, it would increase the size of $\Sigma$ by a factor of $n$ and the rate would increase by a factor of $1/\log n$, which is far from optimal. Synchronization strings allow the communicating parties to index their messages using an alphabet size that is independent of the total communication length $n$.
Suppose that with each of Alice's $n$ messages, $C_A$ sends an encoding of her index using a symbol from $\Sigma$. Let $S$ be a ``synchronization string'' consisting of all $n$ encoded indices sent by $C_A$. Further, suppose that the adversary injects a total of $n\delta$ insertions and deletions, thus transforming the string $S$ to the string $S_{\tau}$ (The notation will be justified in Appendix~\ref{app:prelim}). Let some element of $S$ like $S[i]$ pass through the channel without being deleted by the adversary and arrive as $S_{\tau}[j]$. We call $S_{\tau}[j]$ a \emph{successfully transmitted symbol}.
We assume that $C_A$ and $C_B$ know the string $S$ \emph{a priori}. The intermediary $C_B$ will receive a set of transmitted indices $S_{\tau}[1], \dots, S_{\tau}[n]$. Upon receipt of the $j$th transmitted index, $C_B$ guesses the actual index of the received symbol when sent by $C_A$. We call the algorithm that $C_B$ runs to determine this an \emph{($n,\delta)$-indexing algorithm.} The algorithm can also return a symbol $\top$ which represents an ``I don't know'' response. Any successfully transmitted symbols that is decoded incorrectly is called a \emph{misdecoding}. The number of misdecodings that an ($n,\delta)$-indexing algorithm might produced is used as a measure to valuate its quality. An indexing algorithm is \emph{streaming} if its guess for a received symbol only depends on previously arrived symbols.
Haeupler and Shahrasbi defined a family of synchronization strings that admit an $(n,\delta)$-indexing algorithm with strong performance \cite{haeupler2017synchronization}. This family is characterized by a parameter $\eps$ as follows.
\begin{definition}[$\eps$-Synchronization String]\label{def:synCode}
A string $S \in \Sigma^n$ is an $\eps$-synchronization string if for every $1 \leq i < j < k \leq n + 1$ we have that $ED\left(S[i, j),S[j, k)\right) > (1-\eps) (k-i)$.
\end{definition}
Haeupler and Shahrasbi \cite{haeupler2017synchronization, haeupler2017synchronization3} prove the existence and provide several fast constructions for $\eps$-synchronization strings and provide a streaming $(n,\delta)$-indexing algorithm that returns a solution with $\frac{c_i}{1-\eps} + \frac{c_d\eps}{1-\eps}$ misdecodings. The algorithm runs in time $O(n^5)$, spending $O(n^4)$ on each received symbol.
}
\global\def\Prelim{
\medskip\noindent\textbf{String notation and edit distance.} Let $S$ be a string of $n$ symbols from an alphabet $\Sigma$. For $i,j \in \{1, \dots, n\}$, we denote the substring of $S$ from the $i^{th}$ index through and including the $j^{th}$ index as $S[i,j]$. We refer to the substring from the $i^{th}$ index through, but not including, the $j^{th}$ index as $S[i,j)$. The substrings $S(i,j]$ and $S(i,j)$ are similarly defined. Finally, $S[i]$ denotes the $i^{th}$ symbol of $S$ and $|S| = n$ is the length of $S$.
Occasionally, the alphabets we use are the Cartesian product of several alphabets, i.e. $\Sigma = \Sigma_1 \times \cdots \times \Sigma_n$. If $T$ is a string over $\Sigma,$ then we write $T[i] = \left[a_1, \dots, a_n\right]$, where $a_i \in \Sigma_i$. Throughout this work, we rely on the well-known \emph{edit distance} metric, which measures the smallest number of insertions and deletions necessary to transform one string to another.
\begin{definition}[Edit distance]
The \emph{edit distance} $ED(c,c')$ between two strings $c,c' \in \Sigma^*$ is the minimum number of insertions and deletions required to transform $c$ into $c'$.
\end{definition}
It is easy to see that edit distance is a metric on any set of strings and in particular is symmetric and satisfies the triangle inequality property. Furthermore, $ED\left(c,c'\right) = |c| + |c'| - 2\cdot LCS\left(c,c'\right)$, where $LCS\left(c,c'\right)$ is the longest common substring of $c$ and $c'$.
We also use the \emph{string matching} notation from~\cite{braverman2015coding}:
\begin{definition}[String matching]
Suppose that $c$ \allowbreak and $c'$ are two strings in $\Sigma^*$, and suppose that $*$ is a symbol not in $\Sigma$. Next, suppose that there exist two strings $\tau_1$ and $\tau_2$ in $\left(\Sigma \cup \{*\}\right)^*$ such that $|\tau_1| = |\tau_2|$, $del\left(\tau_1\right) = c$, $del(\tau_2) = c'$, and $\tau_1[i] \approx \tau_2[i]$ for all $i \in \left\{1, \dots, |\tau_1|\right\}$. Here, $del$ is a function that deletes every $*$ in the input string and $a \approx b$ if $a = b$ or one of $a$ or $b$ is $*$. Then we say that $\tau = \left(\tau_1, \tau_2\right)$ is a \emph{string matching} between $c$ and $c'$ (denoted $\tau: c \to c'$). We furthermore denote with $sc\left(\tau_i\right)$ the number of $*$'s in $\tau_i$.
\end{definition}
Note that the \emph{edit distance} between strings $c,c' \in \Sigma^*$ is exactly equal to $\min_{\tau: c \to c'}\left\{sc\left(\tau_1\right) + sc\left(\tau_2\right)\right\}$.
\begin{definition}[Suffix distance]
Given any two strings $c,\tilde{c} \in \Sigma^*$, the \emph{suffix distance} between $c$ and $\tilde{c}$ is $SD\left(c,\tilde{c}\right) = \min_{\tau: c \to \tilde{c}}\left\{\max_{i = 1}^{|\tau_1|} \left\{ \frac{sc\left(\tau_1\left[i,|\tau_1|\right]\right) + sc\left(\tau_2\left[i,|\tau_2|\right]\right)}{|\tau_1| - i + 1 - sc\left(\tau_1\left[i,|\tau_1|\right]\right)}\right\}\right\}$.
\end{definition}
\medskip\noindent\textbf{Synchronization Strings.} We now recall the definition of synchronization strings, which were first introduced by Haeupler and Shahrasbi \cite{haeupler2017synchronization} and further studied in~\cite{haeupler2017synchronization3, haeupler2018synchronization4, cheng2018synchronization}, along with some important lemmas from~\cite{haeupler2017synchronization} which we will need here. In short, synchronization strings \cite{haeupler2017synchronization} allow communicating parties to protect against synchronization errors by indexing their messages without blowing up the communication rate.
We describe this technique by introducing two intermediaries, $C_A$ and $C_B$, that conduct the communication over the given insertion-deletion channel. $C_A$ receives all symbols that Alice wishes to send to Bob, $C_A$ sends the symbols to $C_B$, and $C_B$ communicates the symbols to Bob. $C_A$ and $C_B$ handle the synchronization strings and all the extra work that is involved in keeping Alice and Bob in sync by guessing the actual index of symbols received by $C_B$.
In this way, Alice and Bob communicate via $C_A$ and $C_B$ as though they were communicating over a half-error channel.
Unfortunately, trivially adding the indices $1, 2, \dots, n$ to each message will not allow us to maintain a near optimal communication rate. After all, suppose that $C_A$ and $C_B$ are communicating over an alphabet $\Sigma$ of constant size. If $C_A$ adds an index to each of Alice's messages, it would increase the size of $\Sigma$ by a factor of $n$ and the rate would increase by a factor of $1/\log n$, which is far from optimal. Synchronization strings allow the communicating parties to index their messages using an alphabet size that is polynomial in $1/\delta$ and is thus independent of the total communication length $n$. Of course, some accuracy is lost when deviating from the trivial indexing strategy. Therefore, if $C_A$ sends $C_B$ indices attached to each of Alice's $n$ messages, we need a notion of how well $C_B$ is able to determine those indices in the presence of synchronization errors. This is where the notion of a string matching comes in handy.
Suppose that with each of Alice's $n$ messages, $C_A$ sends an encoding of her index using a symbol from $\Sigma$. Let $S$ be a ``synchronization string'' consisting of all $n$ encoded indices sent by $C_A$. Next, suppose that the adversary injects a total of $n\delta$ insertions and deletions, thus transforming the string $S$ to the string $S_{\tau}$. Here, $\tau= (\tau_1, \tau_2)$ is a string matching such that $del(\tau_1) = S$, $del(\tau_2) = S_{\tau}$, and for all $k \in [|\tau_1|] = [|\tau_2|]$, \[(\tau_1[k], \tau_2[k]) = \begin{cases} (S[i], *) &\text{if }S[i] \text{ is deleted}\\
(S[i], S_{\tau}[j]) &\text{if }S_{\tau}[j] \text{ is successfully transmitted and sent as }S[i]\\
(*,S_{\tau}[j]) &\text{if }S_{\tau}[j] \text{ is inserted,}\end{cases}\] where $i$ is the index of $\tau_1[1,k]$ upon deleting the stars in $\tau_1[1,k]$, or in other words, $i = |del(\tau_1[1,k])|$ and similarly $j = |del(\tau_1[1,k])|$. We formally say that a symbol $S_{\tau}[j]$ is \emph{successfully transmitted} if there exists a $k$ such that $|del(\tau_2[1,k])| = j$ and $\tau_1[k] = \tau_2[k]$. It was not inserted or deleted by the adversary.
We assume that $C_A$ and $C_B$ know the string $S$ \emph{a priori}. The intermediary $C_B$ will receive a set of transmitted indices $S_{\tau}[1], \dots, S_{\tau}[n]$. Upon receipt of the $j$th transmitted index, for all $j \in [n]$, $C_B$ approximately matches $S_{\tau}[1,j]$ to a prefix $S[1,i]$ and therefore guesses that $C_A$ has sent $i$ messages. We call the algorithm that $C_B$ runs to determine this matching an \emph{($n,\delta)$-indexing algorithm.} The algorithm can also return a symbol $\top$ which represents an ``I don't know'' response. Formally, we define an ($n,\delta)$-indexing algorithm as follows.
\begin{definition}[$(n,\delta)$-Indexing Algorithm] The pair $(S, \mathcal{D}_S)$ consisting of a string $S \in \Sigma^n$ and an algorithm $\mathcal{D}_S$ is called an $(n,\delta)$-indexing algorithm over alphabet $\Sigma$ if for any set of $n\delta$ insertions and deletions corresponding to the string matching $\tau$ and altering the string $S$ to a string $S_{\tau}$, the algorithm $\mathcal{D}_S(S_{\tau})$ outputs either $\top$ or an index between 1 and $n$ for every symbol in $S_{\tau}$.
\end{definition}
How well can an ($n,\delta)$-indexing algorithm perform? We can only hope to \emph{correctly decode} symbols that are successfully transmitted. Recall that the symbol $S_{\tau}[j]$ is successfully transmitted if there exists an index $k$ such that $|del(\tau_2[1,k])| = j$ and $\tau_1[k] = S[i]$ for some $i \in [n]$. It makes sense, then, to say that the algorithm correctly decodes $S_{\tau}[j]$ if it successfully recovers the index $i$. Indeed, we express this notion formally by saying that an $(n,\delta)$-indexing algorithm $(S, \mathcal{D}_S)$ decodes index $j$ correctly under $\tau = (\tau_1, \tau_2)$ if $\mathcal{D}_S(S_{\tau})$ outputs $i$ and there exists a $k$ such that $i = |del(\tau_1[1,k])|$, $j = |del(\tau_2[1,k])|$, $\tau_1[k] = S[i]$, and $\tau_2[k] = S_{\tau}[j]$. Notice that outputting $\top$ counts as an incorrect decoding.
We now have the language to describe how well an $(n,\delta)$-indexing algorithm performs. An algorithm has at most $k$ misdecodings if for any $\tau$ corresponding to at most $n\delta$ insertions and deletions, there are at most $k$ successfully transmitted, incorrectly decoded indices.
An indexing algorithm is \emph{streaming} if the decoded index
for the $i$th element of the string $S_{\tau}$ only depends on $S_{\tau}[1, i]$.
Haeupler and Shahrasbi defined a family of synchronization strings that admit an $(n,\delta)$-indexing algorithm with strong performance \cite{haeupler2017synchronization}. This family is characterized by a parameter $\eps$ as follows.
\begin{definition}[$\eps$-Synchronization String]\label{def:synCode}
A string $S \in \Sigma^n$ is an $\eps$-synchronization string if for every $1 \leq i < j < k \leq n + 1$ we have that $ED\left(S[i, j),S[j, k)\right) > (1-\eps) (k-i)$.
\end{definition}
Haeupler and Shahrasbi provide an efficient randomized construction of $\eps$-synchronization strings of arbitrary length. \begin{lemma}[From \cite{haeupler2017synchronization}]\label{lemma:FiniteSyncConstruction}
There exists a randomized algorithm which, for any $\eps >0$, constructs a $\eps$-synchronization string of length $n$ over an alphabet of size $O(\eps^{-4})$ in expected time $O(n^5)$.
\end{lemma}
They prove the following useful property which leads them to an indexing algorithm.
\begin{lemma}[From \cite{haeupler2017synchronization}]
Let $S \in \Sigma^n$ be an $\eps$-synchronization string and let $S_{\tau}[1,j]$ be a prefix of $S_{\tau}$. Then there exists at most one index $i \in [n]$ such that the suffix distance between $S_{\tau}[1,j]$ and $S[1,i]$, denoted by $SD(S_{\tau}[1,j],S[1,i])$ is at most $1-\eps$.
\end{lemma}
This lemma suggests a simple $(n,\delta)$-indexing algorithm given an input prefix $S_{\tau}[1,j]$: Use dynamic programming to search over all prefixes of $S$ for the one with the smallest suffix distance from $S_{\tau}[1,j]$. Haeupler and Shahrasbi present a dynamic programming algorithm which is efficient and results in a small number of misdecodings, and described in the following theorem.
\begin{theorem}[From \cite{haeupler2017synchronization}]\label{thm:RSPDmisdecodings}
Let $S \in \Sigma^n$ be an $\eps$-synchronization string that is sent over an insertion-deletion channel with a $\delta$ fraction of insertions and deletions. There exists a streaming $(n,\delta)$-indexing algorithm that returns a solution with $\frac{c_i}{1-\eps} + \frac{c_d\eps}{1-\eps}$ misdecodings. The algorithm runs in time $O(n^5)$, spending $O(n^4)$ on each received symbol.
\end{theorem}
}\fullOnly{\Prelim}
\section{Channel Simulations}\label{sec:simulation}
In this section, we show how $\eps$-synchronization strings can be used as a powerful tool to simulate corruption channels over insertion-deletion channels.
In Section~\ref{sec:codings}, we use these simulations to introduce coding schemes resilient to insertion-deletion errors.
\fullOnly{We study the context where Alice and Bob communicate over an insertion-deletion channel, but via a blackbox channel simulation, they are able to run coding schemes that are designed for half-error channels. As we describe in Section~\ref{sec:def}, we discuss this simulation by introducing two intermediaries, $C_A$ and $C_B$, that conduct the simulation by communicating over the given insertion-deletion channel.}
\subsection{One-way channel simulation over a large alphabet} \label{sec:OneWayLargeAlphaSimulation}
Assume that Alice and Bob have access to $n$ rounds of communication over a one-way insertion-deletion channel where the adversary is allowed to insert or delete up to $n\delta$ symbols. In this situation, we formally define a corruption channel simulation over the given insertion-deletion channel as follows:
\begin{definition}[Corruption Channel Simulation]
Let Alice and Bob have access to $n$ rounds of communication over a one-way insertion-deletion channel with the alphabet $\Sigma$. The adversary may insert or delete up to $n\delta$ symbols.
Intermediaries $C_A$ and $C_B$ \emph{simulate} $n'$ rounds of a corruption channel with alphabet $\Sigma_{sim}$ over the given channel as follows. First, the adversary can insert a number of symbols into the insertion-deletion channel between $C_A$ and $C_B$. Then for $n'$ rounds $i=1,\dots, n'$, the following procedure repeats:
\begin{enumerate
\item Alice gives $X_i\in\Sigma_{sim}$ to $C_A$.
\item Upon receiving $X_i$ from Alice, $C_A$ wakes up and sends a number of symbols (possibly zero) from the alphabet $\Sigma$ to $C_B$ through the given insertion-deletion channel.
The adversary can delete any of these symbols or insert symbols \emph{before}, \emph{among}, or \emph{after} them.
\item Upon receiving symbols from the channel, $C_B$ wakes up and reveals a number of symbols (possibly zero) from the alphabet $\Sigma_{sim}$ to Bob. We say all such symbols are \emph{triggered} by $X_i$.
\end{enumerate}
Throughout this procedure, the adversary can insert or delete up to $n\delta$ symbols. However, $C_B$ is required to reveal exactly $n'$ symbols to Bob regardless of the adversary's actions.
Let $\tilde{X}_1,\cdots, \tilde{X}_{n'}\in\Sigma_{sim}$ be the symbols revealed to Bob by $C_B$.
This procedure successfully simulates $n'$ rounds of a corruption channel with a $\delta'$ fraction of errors if for all but $n'\delta'$ elements $i$ of the set $\{1,\dots,n'\}$, the following conditions hold: 1) $\tilde{X}_i = X_i$; and 2) $\tilde{X}_i$ is triggered by $X_i$.
\end{definition}
When $\tilde{X}_i = X_i$ and $\tilde{X}_i$ is triggered by $X_i$, we call $\tilde{X}_i$ an \emph{uncorrupted symbol}.
The second condition, that $\tilde{X}_i$ is triggered by $X_i$, is crucial to preserving linearity of time, which is the fundamental quality that distinguishes channel simulations
from channel codings.
It forces $C_A$ to communicate each symbol to Alice as soon as it arrives. Studying channel simulations satisfying this condition is especially important in situations where Bob's messages depends on Alice's, and vice versa.
Conditions (1) and (2) also require that $C_B$ conveys at most one uncorrupted symbol each time he wakes up.
As the adversary may delete $n\delta$ symbols from the insertion-deletion channel, $C_B$ will wake up at most $n(1-\delta)$ times.
Therefore, we cannot hope for a corruption channel simulation where Bob receives more than $n(1-\delta)$ uncorrupted symbols. In the following theorem, we prove something slightly stronger: no deterministic one-way channel simulation can guarantee that Bob receives more than $n(1-4\delta/3)$ uncorrupted symbols and if the simulation is randomized, the expected number of uncorrupted transmitted symbols is at most $n(1-7\delta/6)$. This puts channel simulation in contrast to channel coding as one can recover $1-\delta-\eps$ fraction of symbols there (as shown in~\cite{haeupler2017synchronization}).
\shortOnly{The proof is in Appendix~\ref{app:chapter4}.}
\begin{theorem}\label{thm:OnewaySimulLowerBound}
Assume that $n$ uses of a one-way insertion-deletion channel over an arbitrarily large alphabet $\Sigma$ with a $\delta$ fraction of insertions and deletions are given. There is no deterministic simulation of a corruption channel over any alphabet $\Sigma_{sim}$ where the simulated channel guarantees more than $n\left(1-4\delta/3\right)$ uncorrupted transmitted symbols. If the simulation is randomized, the expected number of uncorrupted transmitted symbols is at most $n(1-7\delta/6)$.
\end{theorem}
\global\def\OnewaySimulLowerBoundProof{
\shortOnly{\begin{proof}[Proof of Theorem~\ref{thm:OnewaySimulLowerBound}]}
\fullOnly{\begin{proof}}
Consider a simulation of $n'$ rounds of a corruption channel on an insertion-deletion channel.
Note that for any symbol that $C_A$ receives, she will send some number of symbols to $C_B$. This number can be zero or non-zero and may also depend on the content of the symbol she receives. We start by proving the claim for deterministic simulations. Let $X_1, X_2, \cdots, X_{n'}\in \Sigma_{sim}$ and $X'_1, X'_2, \cdots, X'_{n'} \in \Sigma_{sim}$ be two possible sets of
inputs that Alice may pass to $C_A$ such that for any $1\le i\le n'$, $X_ i\neq X'_i$ and $Y_1,\cdots,Y_{m}\in\Sigma$ and let $Y'_1,\cdots,Y'_{m'}\in\Sigma$ be the symbols that $C_A$ sends to $C_B$ through the insertion-deletion channel as a result of receiving $\{X_i\}$ and $\{X'_i\}$ respectively.
Now, consider $Y_1, \cdots, Y_{n\delta}$. Let $k$ be the number of $C_A$'s input symbols which are required to trigger her to output $Y_1, \cdots, Y_{n\delta}$. We prove Theorem~\ref{thm:OnewaySimulLowerBound} in the following two cases:
\begin{enumerate}
\item $k \le \frac{2n\delta}{3}$: In this case, $Y_1, \cdots, Y_{n\delta}$ will cause $C_B$ to output at most $\frac{2n\delta}{3}$ uncorrupted symbols. If the adversary deletes $n\delta$ arbitrary elements among $Y_{n\delta+1}, \cdots, Y_{m}$, then $C_B$ will receive $m-2n\delta \le n-2n\delta$ symbols afterwards; Therefore, he cannot output more than $n-2n\delta$ uncorrupted symbols as the result of receiving $Y_{n\delta+1}, \cdots, Y_{m}$. Hence, no simulation can guarantee $n(1-2\delta) + k < n\left(1-\frac{4}{3}\delta\right)$ uncorrupted symbols or more.
\item $k > \frac{2n\delta}{3}$: Consider the following two scenarios:
\begin{enumerate}
\item[(a)] Alice tries to convey $X_1, X_2, \cdots, X_{n'}$ to Bob using the simulation. The adversary deletes the first $n\delta$ symbols. Therefore, $C_B$ receives $Y_{n\delta+1}, \cdots, Y_{m}$.
\item[(b)] Alice tries to convey $X'_1, X'_2, \cdots, X'_{n'}$ to Bob using the simulation. The adversary inserts $Y_{n\delta+1}, \cdots, Y_{2n\delta}$ at the very beginning of the communication. Therefore, $C_B$ receives $Y_{n\delta+1}, \cdots, Y_{2n\delta}, Y'_1, Y'_2, \cdots, Y'_{m'}$.
\end{enumerate}
Note that the first $n\delta$ symbols that $C_B$ receives in these two scenarios are the same. Assume that $C_B$ outputs $k'$ symbols as the result of the first $n\delta$ symbols he receives. In the first scenario, the number of uncorrupted symbols $C_B$ outputs as the result of his first $n\delta$ inputs is at most $\max\{0, k'-k\}$. Additionally, at most $m-2n\delta \le n-2n\delta$ uncorrupted messages may be conveyed within the rest of the communication. In the second scenario, the number of uncorrupted communicated symbols is at most $n-k'$.
Now, at least for one of these scenarios the number of guaranteed uncorrupted symbols in the simulation is
\begin{eqnarray*}
\min\left\{n-2n\delta+\max\{0,k'-k\}, n-k'\right\} &\le& \max_{k'} \min\left\{n-2n\delta+\max\{0,k'-k\}, n-k'\right\}\\
&\le& \max_{k'} \min\left\{n-2n\delta+\max\{0,k'-2n\delta/3\}, n-k'\right\}\\
&=& n-4n\delta/3 = n\left(1-\frac{4}{3}\delta\right).
\end{eqnarray*}
\end{enumerate}
This completes the proof for deterministic simulations. Now, we proceed to the case of randomized simulations.
Take an arbitrary input sequence $X_1,\cdots,X_{n'}\in\Sigma^{n'}$. Let $K_X$ be the random variable that represents the number of $C_A$'s input symbols which are required to trigger her to output her first $n\delta$ symbols to $C_B$.
If $\Pr\{K_{X} \le 2n\delta/3\} \ge \frac{1}{2}$, for any sequence $X_1,\cdots,\bar{X}_{n'}$ given to $C_A$ by Alice, the adversary acts as follows.
He lets the first $n\delta$ symbols sent by $C_A$ pass through the insertion-deletion channel and then deletes the next $n\delta$ symbols that $C_A$ sends to $C_B$. As in the deterministic case, if $K_X \le 2n\delta/3$, the number of uncorrupted symbols conveyed to Bob cannot exceed $n(1-4\delta/3)$. Hence, the expected number of uncorrupted symbols in the simulation may be upper-bounded by:
\begin{eqnarray*}
\E[\text{Uncorrupted Symbols}] &\le& p\cdot n(1-4\delta/3) + (1-p)\cdot n(1-\delta)\\
&\le& n\left(1-\frac{3+p}{3}\delta\right)\le n\left(1-\frac{3+1/2}{3}\delta\right) = n\left(1-\frac{7}{6}\delta\right).
\end{eqnarray*}
Now, assume that $\Pr\{K_{X} \le 2n\delta/3\} < \frac{1}{2}$. Take an arbitrary input $X'_1, X'_2, \cdots, X'_{n'} \in \Sigma_{sim}$ such that $X_i \neq X'_i$ for all $1\le i \le n'$. Consider the following scenarios:
\begin{enumerate}
\item[(a)] Alice tries to convey $X_1, X_2, \cdots, X_{n'}$ to Bob using the simulation.
The adversary removes the first $n\delta$ symbols sent by $C_A$. This means that $C_B$ receives is $Y_{n\delta+1},\cdots,Y_{m}$ where $Y_{1},\cdots,Y_{m}$ is a realization of $C_A$'s output distribution given $X_1, X_2, \cdots, X_{n'}$ as input.
\item[(b)] Alice tries to convey $X'_1, X'_2, \cdots, X'_{n'}$ to Bob using the simulation. The adversary mimics $C_A$ and generates a sample of $C_A$'s output distribution given $X_1, X_2, \cdots, X_{n'}$ as input. Let that sample be $Y_{1},\cdots,Y_{m}$. The adversary inserts $Y_{n\delta+1},\cdots,Y_{2n\delta}$ at the beginning and then lets the communication go on without errors.
\end{enumerate}
Note that the distribution of the first $n\delta$ symbols that $C_B$ receives, i.e., $Y_{n\delta+1},\cdots,Y_{2n\delta}$, is the same in both scenarios.
Let $K'_{X'}$ be the random variable that represents the number of symbols in $C_B$'s output given that specific distribution over the symbols $Y_{n\delta+1},\cdots,Y_{2n\delta}$. Now, according to the discussion we had for deterministic simulations, for the first scenario:
\begin{eqnarray*}
\E[\text{Uncorrupted Symbols}] &\le& \E[n-2n\delta+\max\{0,K'_{X'}-K_X\}]\\
&\le& n-2n\delta + p\cdot\E[K'_{X'}] + (1-p)(\E[K'_{X'}]-2n\delta/3)\\
&\le& n-2n\delta + \E[K'_{X'}] - 2(1-p)n\delta/3
\end{eqnarray*}
and for the second one:
$$\E[\text{Uncorrupted Symbols}] \le \E[n-K'_{X'}] \le n-\E[K'_{X'}].$$
Therefore, in one of the above-mentioned scenarios
\begin{eqnarray*}
\E[\text{Uncorrupted Symbols}] &\le& \min\{n-2n\delta + \E[K'_{X'}] - 2(1-p)n\delta/3, n-\E[K'_{X'}]\} \\
&\le& \max_{\gamma}\min\{n-2n\delta + \gamma - 2(1-p)n\delta/3, n-\gamma\}\\
&=& n\left(1-\frac{4-p}{3}\delta\right)\\
&<& n\left(1-\frac{4-1/2}{3}\delta\right)\\
&=& n\left(1-\frac{7}{6}\delta\right).
\end{eqnarray*}
Therefore, for any randomized simulation, there exists an input and a strategy for the adversary where
$$\E[\text{Uncorrupted Symbols}] \le n\left(1-\frac{7}{6}\delta\right).$$
\end{proof}
}
\fullOnly{\OnewaySimulLowerBoundProof}
We now provide a channel simulation using $\eps$-synchronization strings. Every time $C_A$ receives a symbol from Alice (from an alphabet $\Sigma_{sim}$), $C_A$ appends a new symbol from a predetermined $\eps$-synchronization string over an alphabet $\Sigma_{syn}$ to Alice's symbol and sends it as one message through the channel.
On the other side of channel, suppose that $C_B$ has already revealed some number of symbols to Bob. Let $\texttt{I}_B$ be the index of the next symbol $C_B$ expects to receive. \fullOnly{In other words, suppose that $C_B$ has already revealed $\texttt{I}_B - 1$ symbols to Bob.} Upon receiving a new symbol from $C_A$, $C_B$ uses the part of the message coming from the synchronization string to guess the index of the message Alice sent. We will refer to this decoded index as $\tilde{\texttt{I}}_A$ and its actual index as $\texttt{I}_A$. If $\widetilde{\texttt{I}}_A < \texttt{I}_B$, then $C_B$ reveals nothing to Bob and ignores the message he just received. Meanwhile, if $\widetilde{\texttt{I}}_A = \texttt{I}_B$, then $C_B$ reveals Alice's message to Bob. Finally, if $\widetilde{\texttt{I}}_A > \texttt{I}_B$, then $C_B$ sends a dummy symbol to Bob and then sends Alice's message.
\fullOnly{Given that the adversary can insert or delete up to $n\delta$ symbols, if $C_A$ sends $n$ symbols, then $C_B$ may receive between $n -n\delta$ and $n+ n\delta$ symbols. We do not assume the parties have access to a clock, so we must prevent $C_B$ from stalling after $C_A$ has sent all $n$ messages. Therefore, $C_B$ only listens to the first $n(1-\delta)$ symbols it receives.}
The protocols of $C_A$ and $C_B$ are more formally described in Algorithm~\ref{alg:OnewayC_A} \shortOnly{in Appendix~\ref{app:chapter4}}. Theorem~\ref{thm:OnewayCnstSizeAlphaSimul} details the simulation guarantees\shortOnly{ and the proof is in Appendix~\ref{app:chapter4}}.
\fullOnly{
\begin{algorithm}
\caption{Simulation of a one-way constant alphabet channel}
\begin{algorithmic}[1]\label{alg:OnewayC_A}
\STATE Initialize parameters: $S \leftarrow \eps$-synchronization string of length $n$
\IF {$C_A$}
\STATE Reset Status: $\texttt{i}\leftarrow 0$
\FOR {$n$ iterations}
\STATE Get $m$ from Alice, send $(m, S[\texttt{i}])$ to $C_B$, and increment \texttt{i} by 1.
\ENDFOR
\ENDIF
\IF {$C_B$}
\STATE Reset Status: $\texttt{I}_B\leftarrow 0$
\FOR {$n(1-\delta)$ iterations}
\STATE Receive $(\tilde{m}, \tilde{s})$ sent by $C_A$ and set $\tilde{\texttt{I}}_A \leftarrow \mbox{Synchronization string decode}(\tilde{s}, S)$
\IF {$\tilde{\texttt{I}}_A = \texttt{I}_B$}
\STATE Send $\tilde{m}$ to Bob and increment $\texttt{I}_B$.
\ENDIF
\IF {$\tilde{\texttt{I}}_A < \texttt{I}_B$}
\STATE Continue
\ENDIF
\IF {$\tilde{\texttt{I}}_A > \texttt{I}_B$}
\STATE Send a dummy symbol and then $\tilde{m}$ to Bob, then increment $\texttt{I}_B$ by 2.
\ENDIF
\ENDFOR
\ENDIF
\end{algorithmic}
\end{algorithm}}
\begin{theorem}\label{thm:OnewayCnstSizeAlphaSimul}
Assume that $n$ uses of a one-way insertion-deletion channel over an alphabet $\Sigma$ with a $\delta$ fraction of insertions and deletions are given. Using an $\eps$-synchronization string over an alphabet $\Sigma_{syn}$, it is possible to simulate $n(1-\delta)$ rounds of a one-way corruption channel over $\Sigma_{sim}$ with at most $2n\delta(2+(1-\eps)^{-1})$ symbols corrupted so long as $|\Sigma_{sim}| \times |\Sigma_{syn}| \le |\Sigma|$ and $\delta < 1/7$.
\end{theorem}
\global\def\OnewayCnstSizeAlphaSimulProof{
\shortOnly{\begin{proof}[Proof of Theorem~\ref{thm:OnewayCnstSizeAlphaSimul}]}
\fullOnly{\begin{proof}}
Let Alice and Bob use $n$ rounds of an insertion-deletion channel over alphabet $\Sigma$ as sender and receiver respectively. We describe the simulation as being coordinated by two intermediaries $C_A$ and $C_B$, who act according to Algorithm~\ref{alg:OnewayC_A}.
In order to find a lower-bound on the number of rounds of the simulated communication that remain uncorrupted, we upper-bound the number of rounds that can be corrupted.
To this end, let the adversary insert $k_i$ symbols and delete $k_d$ symbols from the communication. Clearly, the $k_d$ deleted symbols do not pass across the channel. Also, each of the $k_i$ inserted symbols may cause $C_B$ to change $I_B$. We call these two cases \emph{error-bad} incidents. Further, $n\delta+k_i-k_d$ symbols at the end of the communication are not conveyed to Bob as we truncate the communication at $n(1-\delta)$. Moreover, according to Theorem~\ref{thm:RSPDmisdecodings}, $\frac{k_i}{1-\eps} + \frac{k_d\eps}{1-\eps}$ successfully transmitted symbols may be misdecoded upon their arrival. We call such incidents \emph{decoding-bad} incidents. Finally, we need to count the number of successfully transmitted symbols whose indexes are decoded correctly ($\tilde{I}_A=I_A$) but do not get conveyed to Bob because $\tilde{I}_A\not\in \{I_B, I_B+1\}$, which we call \emph{zero-bad} incidents. Zero-bad incidents happen only if $|I_A-I_B|\not= 0$. To count the number of zero-bad incidents, we have to analyze how $I_A$ and $I_B$ change in any of the following cases:
\begin{table*}[h]
\centering
\begin{tabular}{|| c | c c c||}
\hline
Cases when $I_A > I_B$ & $I_A$ & $I_B$ & $I_A - I_B$ \\ [0.5ex]
\hline\hline
Deletion by the Adversary & +1 & 0 & +1 \\
Insertion by the Adversary & 0 & 0,+1,+2 & 0, -1, -2 \\
Correctly Transmitted but Misdecoded & +1 & 0,+1,+2 & -1, 0, +1 \\
Correctly Transmitted and Decoded & +1 & +2 & -1 \\
\hline
\hline
Cases when $I_A < I_B$ & $I_A$ & $I_B$ & $I_B-I_A$ \\ [0.5ex]
\hline\hline
Deletion by the Adversary & +1 & 0 & -1 \\
Insertion by the Adversary & 0 & 0,+1,+2 & 0, +1, +2 \\
Correctly Transmitted but Misdecoded & +1 & 0,+1,+2 & -1, 0, +1 \\
Correctly Transmitted and Decoded & +1 & 0 & -1 \\
\hline
\hline
Cases when $I_A = I_B$ & $I_A$ & $I_B$ & $I_B-I_A$ \\ [0.5ex]
\hline\hline
Deletion by the Adversary & +1 & 0 & -1 \\
Insertion by the Adversary & 0 & 0,+1,+2 & 0, +1, +2 \\
Correctly Transmitted but Misdecoded & +1 & 0,+1,+2 & -1, 0, +1 \\
Correctly Transmitted and Decoded & +1 & +1 & 0 \\
\hline
\end{tabular}
\caption{How $I_A$ and $I_B$ change in different scenarios.}\label{tbl:indexingSolutions}
\end{table*}
Note that any insertion may increase $|I_A-I_B|$ by up to 2 units and any misdecoding or deletion may increase $|I_A-I_B|$ by up to 1 unit.
Therefore, $|I_A-I_B|$ may be increased $2k_i + k_d + \frac{k_i}{1-\eps} + \frac{k_d\eps}{1-\eps}$ throughout the algorithm.
However, as any successfully transmitted and correctly decoded symbol decreases this variable by at least one, there are at most $2k_i + k_d + \frac{k_i}{1-\eps} + \frac{k_d\eps}{1-\eps}$ zero-bad incidents, i.e. successfully transmitted and correctly decoded symbols that are not conveyed successfully in the simulated corruption channel.
Hence, the following is an upper-bound on the number of symbols that may not remain uncorrupted in the simulated channel:
\begin{eqnarray*}
&&\#(\text{error-bad}) + \#(\text{decoding-bad}) + \#(\text{zero-bad}) + \#(\text{truncated symbols})\\
&\le&k_d + \left[\frac{k_i}{1-\eps} + \frac{k_d\eps}{1-\eps}\right] + \left[2k_i + k_d + \frac{k_i}{1-\eps} + \frac{k_d\eps}{1-\eps}\right] + \left[n\delta +k_i - k_d\right]\\
&=& n\delta+ k_d\left(1+\frac{2\eps}{1-\eps}\right)+ k_i\left(3+\frac{2}{1-\eps}\right) \le n\delta\left(4+\frac{2}{1-\eps}\right).
\end{eqnarray*}
As error fraction shall not exceed one, the largest $\delta$ for which this simulation works is as follows.
$$\left.\frac{n\delta\left(4+\frac{2}{1-\eps}\right)}{n(1-\delta)}\right|_{\eps=0} = \frac{6\delta}{1-\delta} < 1 \Leftrightarrow \delta < \frac{1}{7}$$
\end{proof}}
\fullOnly{\OnewayCnstSizeAlphaSimulProof}
\ShortOnlyVspace{-4mm}
\subsection{Interactive channel simulation over a large alphabet}
\ShortOnlyVspace{-1mm}
We now turn to channel simulations for interactive channels. As in Section~\ref{sec:OneWayLargeAlphaSimulation}, we formally define a corruption interactive channel simulation over a given insertion-deletion interactive channel. We then use synchronization strings to present one such simulation.
\begin{definition}[Corruption Interactive Channel Simulation]
Let Alice and Bob have access to $n$ rounds of communication over an interactive insertion-deletion channel with alphabet $\Sigma$. The adversary may insert or delete up to $n\delta$ symbols.
\fullOnly{The simulation of an interactive corruption channel is performed by a pair of intermediaries $C_A$ and $C_B$ where Alice communicates with $C_A$, $C_A$ interacts over the given insertion-deletion channel with $C_B$, and $C_B$ communicates with Bob. More precisely,} \shortOnly{The intermediaries} $C_A$ and $C_B$ \emph{simulate} $n'$ rounds of a corruption interactive channel with alphabet $\Sigma_{sim}$ over the given channel as follows.
The communication starts when Alice gives a symbol from $\Sigma_{sim}$ to $C_A$. Then Alice, Bob, $C_A$, and $C_B$ continue the communication as follows:
\begin{enumerate
\ShortOnlyVspace{-2mm}
\item Whenever $C_A$ receives a symbol from Alice or $C_B$,
he either reveals a symbol from $\Sigma_{sim}$ to Alice or sends a symbol from $\Sigma$ through the insertion-deletion channel to $C_B$.
\item Whenever $C_B$ receives a symbol from Bob or $C_A$,
he either reveals a symbol from $\Sigma_{sim}$ to Bob or send a symbols from $\Sigma$ through the insertion-deletion channel to $C_A$.
\item Whenever $C_B$ reveals a symbol to Bob, Bob responds with a new symbol from $\Sigma_{sim}$
\item Whenever $C_A$ reveals a symbol to Alice, Alice responds with a symbol in $\Sigma_{sim}$ except for the $\frac{n'}{2}$th time
\end{enumerate}
Throughout this procedure, the adversary can inject up to $n\delta$ edit corruptions. However, regardless of the adversary's actions, $C_A$ and $C_B$ have to reveal exactly $n'/2$ symbols to Alice and Bob respectively.
Let $X_1,\dots, X_{n'}$ be the symbols Alice gives to $C_A$ and $\tilde{X}_1,\dots, \tilde{X}_{n'}\in\Sigma_{sim}$ be the symbols $C_B$ reveals to Bob. Similarly, Let $Y_1,\dots, Y_{n'}$ be the symbols Bob gives to $C_B$ and $\tilde{Y}_1,\dots, \tilde{Y}_{n'}\in\Sigma_{sim}$ be the symbols $C_A$ reveals to Alice.
We call each pair of tuples $(X_i, \tilde{X}_i)$ and $(Y_i, \tilde{Y}_i)$ a \emph{round} of the simulated communication. We call a round \emph{corrupted} if its elements are not equal.
This procedure successfully simulates $n'$ rounds of a corruption interactive channel with a $\delta'$ fraction of errors if for all but $n'\delta'$ of the rounds are corrupted.
\end{definition}
\fullOnly{\begin{algorithm}[t]
\caption{Simulation of a corruption channel using an insertion-deletion channel with a large alphabet: $C_A$'s procedure}
\begin{algorithmic}[1]\label{alg:Inter_big_A}
\STATE $\Pi \leftarrow n$-round interactive coding scheme over a corruption channel to be simulated
\STATE Initialize parameters: $S \leftarrow \eps$-synchronization string of length $n/2$
\STATE $\texttt{I}_A\leftarrow 0$
\FOR {$n/2 - n\delta\left(1 + \frac{1}{1-\eps}\right)$ iterations}
\STATE Get $m$ from Alice, send $(m, S[\texttt{I}_A])$ to $C_B$, and increment $\texttt{I}_A$ by 1.
\STATE Get $(\tilde{m}, \tilde{S})$ from $C_B$ and send $\tilde{m}$ to Alice.
\ENDFOR
\STATE Commit.
\FOR {$n\delta\left(1 + \frac{1}{1-\eps}\right)$ iterations}
\STATE Send $(0, 0)$ to $C_B$, and increment $\texttt{I}_A$ by 1.
\STATE Get $(\tilde{m}, \tilde{S})$ from $C_B$.
\ENDFOR
\end{algorithmic}
\end{algorithm}}
\fullOnly{\begin{algorithm}[t]
\caption{Simulation of a corruption channel using an insertion-deletion channel with a large alphabet: $C_B$'s procedure}
\begin{algorithmic}[1]\label{alg:Inter_big_B}
\STATE $\Pi \leftarrow n$-round interactive coding scheme over a corruption channel to be simulated
\medskip
\STATE Initialize parameters: $S \leftarrow \eps$-synchronization string of length $n/2$
\medskip
\STATE $\texttt{I}_B\leftarrow 0$
\FOR {$n/2$ iterations}
\STATE Receive $(\tilde{m}, \tilde{s})$ from $C_A$.
\STATE $\tilde{\texttt{I}}_A \leftarrow \mbox{Synchronization string decode}(\tilde{s}, S)$.
\IF {Committed}
\STATE Send a dummy message to $C_A$.
\ELSIF {$\tilde{\texttt{I}}_A = \texttt{I}_B$}
\STATE Send $\tilde{m}$ to Bob and increment $\texttt{I}_B$ by 1.
\STATE Receive $m$ from Bob and send $(m, 0)$ to $C_A$.
\ELSIF {$\tilde{\texttt{I}}_A < \texttt{I}_B$}
\STATE Send a dummy message to $C_A$.
\ELSIF {$\tilde{\texttt{I}}_A > \texttt{I}_B$}
\STATE Send a dummy message to Bob.
\STATE Send $\tilde{m}$ to Bob and increment $\texttt{I}_B$ by 2.\label{step:inc2}
\STATE Receive $m$ from Bob and send $(m, 0)$ to $C_A$.
\ENDIF
\STATE If $\texttt{I}_B =n/2 - n\delta\left(1 + \frac{1}{1-\eps}\right)$, commit.
\ENDFOR
\end{algorithmic}
\end{algorithm}}
\fullOnly{\begin{figure}
\centering{\includegraphics[width=.28\textwidth]{steps}}
\caption{Illustration of five steps where only the first is good.}\label{fig:steps}
\end{figure}}
\shortOnly{The protocol and analysis in this large alphabet setting are similar to the harder case where the alphabet is binary. We cover interactive communication for the binary setting in the next section. Therefore, in the interest of space we defer the pseudo-code and algorithm to Appendix~\ref{app:chapter4}. We prove the following theorem.}
\begin{theorem}\label{thm:InteractiveCnstSizeAlphaSimul}
Assume that $n$ uses of an interactive insertion-deletion channel over an alphabet $\Sigma$ with a $\delta$ fraction of insertions and deletions are given. Using an $\eps$-synchronization string over an alphabet $\Sigma_{syn}$, it is possible to simulate $n-2n\delta(1 + (1-\eps)^{-1})$ uses of an interactive corruption channel over $\Sigma_{sim}$ with at most a $\frac{2\delta(5-3\eps)}{1-\eps +2\eps\delta - 4\delta}$ fraction of symbols corrupted so long as $|\Sigma_{sim}| \times |\Sigma_{syn}| \le |\Sigma|$ and $\delta < 1/14$.
\end{theorem}
\fullOnly{\begin{proof}
Suppose that Alice and Bob want to communicate a total of $n-2n\delta\left(1 + \frac{1}{1-\eps}\right)$ symbols over the simulated corruption channel, and that this channel is simulated by the intermediaries $C_A$ and $C_B$, who are communicating via a total of $n$ uses of an insertion-deletion channel. We will show later on that both parties have the chance to commit before the other has sent $n/2$ messages over the insertion-deletion channel.
We say an intermediary \emph{commits} when it finishes simulating the channel for its corresponding party, i.e., when it sends the last simulated symbol out.
Intermediaries may commit and yet carry on exchanging symbols over the channel so that the other intermediary finishes its simulation as well. An intermediary may stall by waiting for receiving symbols from the channel but the nature of simulation necessitates the intermediaries not to stall before they commit.
To analyze this simulation, we categorize the bad events that could occur as follows. We say that $C_A$ takes a \emph{step} when it sends a message, receives a message, and completes its required communication with Alice. We say that $C_A$'s step is \emph{good} if the message $C_A$ receives is an uncorrupted response to its previous outgoing message and $C_B$ correctly decodes the index that $C_A$ sends. Figure~\ref{fig:steps} \shortOnly{in Appendix~\ref{app:chapter4}} illustrates a sequence of steps where only the first is good.
If $C_A$ has a good step when $\texttt{I}_A=\texttt{I}_B$ and neither party has committed, then Alice and Bob are guaranteed to have an error-free round of communication. We lower bound the total number of Alice and Bob's error-free rounds of communication by lower bounding the number of good steps that $C_A$ takes when $\texttt{I}_A=\texttt{I}_B$. The total number of good steps $C_A$ takes is $S = n/2 - n\delta\left(1 + \frac{1}{1-\eps}\right) - d$, where $d$ is the number of rounds there are before $C_A$ commits but after $C_B$ commits, if $C_B$ commits first.
If a step is not good, then we say it is \emph{bad}. Specifically, we say that it is \emph{commit-bad} if $C_B$ commits before $C_A$. We say that it is \emph{decoding-bad} if neither party has committed, $C_B$ receives an uncorrupted message from $C_A$, but $C_B$ does not properly decode the synchronization string index. Otherwise, we say that a bad step is \emph{error-bad}. Since every good step corresponds to a message sent by both Alice and Bob, we may lower bound the number of error-free messages sent over the corruption channel by \begin{align*}&2\left(\frac{n}{2} - n\delta\left(1 + \frac{1}{1-\eps}\right) - d - \#(\text{error-bad steps}) - \#(\text{decoding-bad steps})\right. \\
- & \#(\text{good steps when }\texttt{I}_A \not=\texttt{I}_B)\bigg).\end{align*} In order to lower bound this quantity, we now upper bound $d$, the number of error-bad and decoding-bad steps, and the number of good steps when $\texttt{I}_A \not=\texttt{I}_B$.
We claim that there are at most $n\delta$ error-bad steps since a single adversarial injection could cause $C_A$'s current step to be bad, but that error will not cause the next step to be bad.
Next, we appeal to Theorem~\ref{thm:RSPDmisdecodings} to bound the number of decoding-bad steps. The cited theorem guarantees that if an $\eps$-synchronization string of length $m$ is sent over an insertion-deletion channel with a $\delta'$ fraction of errors, then the receiver will decode the index of the received symbol correctly for all but $\frac{\delta' m}{1-\eps}$ symbols. In this scenario, $m = n/2$ and $\delta' = n\delta / (n/2) = 2\delta$ since an error over the insertion-deletion channel can cause at most one error in the synchronization string transmission. Therefore, there are at most $n\delta / (1-\eps)$ decoding-bad steps.
Now, a single error, be it an adversarial injection or synchronization strings improper decoding, may cause $|\texttt{I}_B - \texttt{I}_A|$ to increase by at most two since a single error may cause $C_B$ to simulate Step~\ref{step:inc2} of Algorithm~\ref{alg:Inter_big_B} and increment $\texttt{I}_B$ by two, while $\texttt{I}_A$ does not change. Further, if an error does not occur but $C_B$ incorrectly decodes the synchronization symbol $C_A$ sends, then $|\texttt{I}_B - \texttt{I}_A|$ may increase by at most one, since $C_B$ might increase $\texttt{I}_B$ by two or zero while $C_A$ increases $\texttt{I}_A$ by one. Meanwhile, if $|\texttt{I}_A - \texttt{I}_B| \geq 1$ and $C_A$ takes a good step, then this difference will decrease by at least 1. In total, over the course of the computation, since there are at most $n\delta$ adversarial injections and $\frac{n\delta}{1-\eps}$ synchronization string improper decodings, we have that $|\texttt{I}_B - \texttt{I}_A|$ increases by at most $2n\delta + \frac{n\delta}{1-\eps}$, which means that $C_A$ will take at most $2n\delta + \frac{n\delta}{1-\eps}$ good steps where $\texttt{I}_B \not= \texttt{I}_A$. We may bound this number of good steps a bit tighter as follows. Let $\texttt{I}_A^*$ and $\texttt{I}_B^*$ be the values of $\texttt{I}_A$ and $\texttt{I}_B$ when the first of $C_A$ or $C_B$ commits. If $\left|\texttt{I}_A^*-\texttt{I}_B^*\right|>0$, then each party only had at most $2n\delta + \frac{n\delta}{1-\eps} - \left|\texttt{I}_A^*-\texttt{I}_B^*\right|$ good steps where $\texttt{I}_B \not= \texttt{I}_A$.
Finally, we must upper bound the number $d$ of rounds there are before $C_A$ commits but after $C_B$ commits if $C_B$ commits first. Assuming $C_B$ commits first, it must be that $\texttt{I}_B^* = \frac{n}{2} - n\delta\left(1 + \frac{1}{1-\eps}\right)$. Therefore, the number of steps before $C_A$ commits is \[d = \frac{n}{2} - n\delta\left(1 + \frac{1}{1-\eps}\right) - \texttt{I}_A^* = \left|\texttt{I}_A^*-\texttt{I}_B^*\right|.\]
We have shown that the number of error-free messages sent over the corruption channel is at least
\begin{align*}
&2\bigg(\frac{n}{2} - n\delta\left(1 + \frac{1}{1-\eps}\right) - d - \#(\text{error-bad steps}) - \#(\text{decoding-bad steps})\\
- &\#(\text{good steps when }\texttt{I}_A \not=\texttt{I}_B)\bigg)\\
\leq \text{ }&2\left(\frac{n}{2} - n\delta\left(1 + \frac{1}{1-\eps}\right) - |\texttt{I}_B^*-\texttt{I}_A^*| - n\delta - \frac{n\delta}{1-\eps} - \left(2n\delta + \frac{n\delta}{1-\eps} - \left|\texttt{I}_A^*-\texttt{I}_B^*\right|\right)\right)\\
= \text{ }&n - 8n\delta - \frac{6n\delta}{1-\eps}. \end{align*} Since Alice and Bob send a total of $n-2n\delta\left(1 + \frac{1}{1-\eps}\right)$ messages over the simulated channel, the error rate $\delta_s$ over the simulated channel is at most $1-\frac{n - 8n\delta - \frac{6n\delta}{1-\eps}}{n-2n\delta\left(1 + \frac{1}{1-\eps}\right)} = \frac{2\delta(5-3\eps)}{1-\eps - 2\delta (1-\eps) - 2\delta}$. Therefore, if $\delta < 1/14$, then $\delta_s < \left.\frac{2\delta(5-3\eps)}{1-\eps +2\delta\eps - 4\delta}\right|_{\eps = 0} = \frac{10\delta}{1 - 4\delta} < 1$, as is necessary.
The last step is to show that $C_A$ has the chance to commit before $C_B$ has sent $n/2$ messages over the insertion-deletion channel, and vice versa. Recall that $C_A$ (respectively $C_B$) commits when $\texttt{I}_A$ (respectively $\texttt{I}_B$) equals $n/2 - n\delta\left(1 + \frac{1}{1-\eps}\right)$. Let $\texttt{i}_B$ be the number of messages sent by $C_B$ over the insertion-deletion channel. The difference $|\texttt{i}_B - \texttt{I}_A|$ only increases due to an error, and a single error can only increase this difference by at most one. Therefore, when $\texttt{i}_B = n/2$, $\texttt{I}_A \geq n/2 - n\delta$, so $C_A$ has already committed. Next, $\texttt{I}_A$ only grows larger than $\texttt{I}_B$ if there is an error or if $C_B$ improperly decodes a synchronization symbol and erroneously chooses to not increase $\texttt{I}_B$. Therefore, $\texttt{I}_A$ is never more than $n\delta\left(1 + \frac{1}{1-\eps}\right)$ larger than $\texttt{I}_B$. This means that when $\texttt{I}_A = n/2$, it must be that $\texttt{I}_B \geq n/2 - \left(1 + \frac{1}{1-\eps}\right)$, so $C_B$ has committed.
\end{proof}}
\subsection{Binary interactive channel simulation}
We now show that with the help of synchronization strings, a binary interactive insertion-deletion channel can be used to simulate a binary interactive corruption channel, inducing a $\widetilde{O}(\sqrt\delta)$ fraction of bit-flips. In this way, the two communicating parties may interact as though they are communicating over a corruption channel. They therefore can employ corruption channel coding schemes while using the simulator as a black box means of converting the insertion-deletion channel to a corruption channel.
The key difference between this simulation and the one-way, large alphabet simulation is that Alice and Bob communicate through $C_A$ and $C_B$ for \emph{blocks} of $r$ rounds, between which $C_A$ and $C_B$ check if they are in sync.
Due to errors, there may be times when Alice and Bob are in disagreement about which block, and what part of the block, they are in. $C_A$ and $C_B$ ensure that Alice and Bob are in sync most of the time.
When Alice sends $C_A$ a message from a new block of communication, $C_A$ holds that message and alerts $C_B$ that a new block is beginning. $C_A$ does this by sending $C_B$ a header that is a string consisting of a single one followed by $s-1$ zeros ($10^{s-1}$). Then, $C_A$ indicates which block Alice is about to start by sending a synchronization symbol to $C_B$. Meanwhile, when $C_B$ receives a $10^{s-1}$ string, he listens for the synchronization symbol, makes his best guess about which block Alice is in, and then communicates with Bob and $C_A$ accordingly. This might entail sending dummy blocks to Bob or $C_A$ if he believes that they are in different blocks. Algorithms~\ref{alg:InterC_A} and \ref{alg:InterC_B} \shortOnly{in Appendix~\ref{app:chapter4}} detail $C_A$ and $C_B$'s protocol. To describe the guarantee that our simulation provides, we first define \emph{block corruption channels}.
\fullOnly{\begin{algorithm}[t]
\caption{Simulation of a corruption channel using an insertion-deletion channel, at $C_A$'s side}
\begin{algorithmic}[1]\label{alg:InterC_A}
\STATE $\Pi \leftarrow n$-round interactive coding scheme over a corruption channel to be simulated \label{step:orig_scheme}
\medskip
\STATE Initialize parameters: $r \leftarrow \sqrt{\frac{\log (1/\delta)}{\delta}}$; $R_{total} \leftarrow \Bigl\lceil n\sqrt{\frac{\delta}{\log (1/\delta)}}\Bigr\rceil$; $s \leftarrow c \log (1/\delta)$; $S \leftarrow \eps$-synchronization string of length $R_{total}$ \label{step:init}
\STATE Reset Status: $\texttt{i}\leftarrow 0$ \label{step:reset_status}
\medskip
\FOR {$R_{total}$ iterations}
\STATE Send $s$ zeros to $C_B$
\STATE Send $S[\texttt{i}]$ to $C_B$
\STATE For $r$ rounds, relay messages between $C_B$ and Alice
\STATE $\texttt{i}\leftarrow\texttt{i}+1$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{Simulation of a corruption channel using an insertion-deletion channel, at $C_B$'s side}
\begin{algorithmic}[1]\label{alg:InterC_B}
\STATE $\Pi \leftarrow n$-round interactive coding scheme over a corruption channel to be simulated \label{step:orig_scheme}
\medskip
\STATE Initialize parameters: $r \leftarrow \sqrt{\frac{\log (1/\delta)}{\delta}}$; $R_{total} \leftarrow \Bigl\lfloor n(1-\delta)\sqrt{\frac{\delta}{\log (1/\delta)}}\Bigr\rfloor$; $s \leftarrow c \log (1/\delta)$; $S \leftarrow \eps$-synchronization string of length $R_{total}$ \label{step:init}
\STATE Reset Status: $\texttt{i}, \texttt{z}, \texttt{I}_B\leftarrow 0$
\medskip
\FOR {$R_{total}$ iterations}
\WHILE {$\texttt{z} < s$}\label{step:wait}
\STATE Receive $b$ from $C_A$
\IF {$b = 0$}
\STATE $\texttt{z}\leftarrow\texttt{z}+1$
\ELSE
\STATE {$\texttt{z} \leftarrow 0$}
\ENDIF
\STATE Send dummy bit to $C_A$
\ENDWHILE
\STATE $\texttt{z} \leftarrow 0$
\medskip
\STATE Receive $m$, the next $|\Sigma_{syn}|$ bits sent by $C_A$ \label{step:sync_symb}
\STATE $\tilde{\texttt{I}}_A \leftarrow \mbox{Synchronization string decode}(m, S)$
\IF {$\tilde{\texttt{I}}_A = \texttt{I}_B$}
\STATE For $r$ rounds, relay messages between $C_A$ and Bob
\STATE $\texttt{I}_B \leftarrow \texttt{I}_B + 1$
\ENDIF
\IF {$\tilde{\texttt{I}}_A < \texttt{I}_B$}\label{step:lt}
\STATE For $r$ rounds, send dummy messages to $C_A$
\ENDIF
\IF {$\tilde{\texttt{I}}_A > \texttt{I}_B$}\label{step:gt}
\STATE For $r$ rounds, send dummy messages to Bob
\STATE For $r$ rounds, relay messages between $C_A$ and Bob
\STATE $\texttt{I}_B \leftarrow \texttt{I}_B + 2$
\ENDIF
\medskip
\ENDFOR
\end{algorithmic}
\end{algorithm}}
\begin{definition}[Block Corruption Channel]
An $n$-round adversarial corruption channel is called a \emph{$(\delta, r)$-block corruption channel} if the adversary is restricted to corrupt $n\delta$ symbols which are covered by $n\delta / r$ blocks of $r$ consecutively transmitted symbols.
\end{definition}
\begin{theorem}\label{thm:InterSim}
Suppose that $n$ rounds of a binary interactive insertion-deletion channel with a $\delta$ fraction of insertions and deletions are given. For sufficiently small $\delta$, it is possible to deterministically simulate
$n(1-\Theta( \sqrt{\delta\log(1/\delta)}))$
rounds of a binary interactive
$ (\Theta(\sqrt{\delta\log(1/\delta)}), \sqrt{(1/\delta)\log (1 / \delta)})$-block corruption channel between two parties, Alice and Bob, assuming that
all substrings of form $10^{s-1}$ where $s = c\log (1/\delta)$ that Alice sends can be covered by $n\delta$ intervals of $\sqrt{(1/\delta)\log (1 / \delta)}$ consecutive rounds.
The simulation is performed efficiently if the synchronization string is efficient and works.
\end{theorem}
\shortOnly{
\begin{proof}[Proof Sketch
Suppose Alice and Bob communicate via intermediaries $C_A$ and $C_B$ who act according to Algorithms~\ref{alg:InterC_A} and \ref{alg:InterC_B}. In total, Alice and Bob will attempt to communicate $n_s$ bits to one another over the simulated channel, while $C_A$ and $C_B$ communicate a total of $n$ bits to one another. The adversary is allowed to insert or delete up to $n\delta$ symbols and $C_A$ sends $n/2$ bits, so $C_B$ may receive between $n/2 -n\delta$ and $n/2+ n\delta$ symbols. To prevent $C_B$ from stalling indefinitely, $C_B$ only listens to the first $n(1-2\delta)/2$ bits he receives.
For $r = \sqrt{(1/\delta)\log (1 / \delta)}$, we define a \emph{chunk} to be $r_c := (s + |\Sigma_{syn}| + r/2)$ consecutive bits that are sent by $C_A$ to $C_B$. In particular, a chunk corresponds to a section header and synchronization symbol followed by $r/2$ rounds of messages sent from Alice. As $C_B$ cares about the first $n(1-2\delta)/2$ bits it receives, there are $\frac{n(1-2\delta)}{2r_c}$ chunks in total.
Hence,
$n_s = \frac{n(1-2\delta)}{2r_c} \cdot r$
since $C_B$ and $C_A$'s communication is alternating.
Note that if Alice sends a substring of form $10^{s-1}$ in the information part of a chunk, then Bob mistakenly detects a new block. With this in mind, we say a chunk is \emph{good} if:
\ShortOnlyVspace{-2mm}
\begin{enumerate
\item No errors are injected in the chunk or affecting $C_B$'s detection of the chunk's header,
\item $C_B$ correctly decodes the index that $C_A$ sends during the chunk, and
\item $C_A$ does not send a $10^{s-1}$ substring in the information portion of the chunk.
\end{enumerate}
\ShortOnlyVspace{-1mm}
If a chunk is not good, we call it \emph{bad}. If the chunk is bad
because $C_B$ does not decode $C_A$'s index correctly even though they were in sync and no errors were injected, then we call it \emph{decoding-bad}. If it is bad because Alice sends a $10^{s-1}$ substring, we call it \emph{zero-bad} and otherwise, we call it \emph{error-bad}. Throughout the protocol, $C_B$ uses the variable $\texttt{I}_B$ to denote the next index of the synchronization string $C_B$ expects to receive and we use $\texttt{I}_A$ to denote the index of the synchronization string $C_A$ most recently sent. Notice that if a chunk is good and $\texttt{I}_A = \texttt{I}_B$, then all messages are correctly conveyed
We now bound the maximum number of bad chunks that occur over the course of the simulation. Suppose the adversary injects errors into the $i^{th}$ chunk, making that chunk bad. The $(i+1)^{th}$ chunk may also be bad, since Bob may not be listening for $10^{s-1}$ from $C_A$ when $C_A$ sends them, and therefore may miss the block header. However, if the adversary does not inject any errors into the $(i+1)^{th}$ and the $(i+2)^{th}$ chunk, then the $(i+2)^{th}$ chunk will be good.
In effect, a single error may render at most two chunks useless. Since the adversary may inject $n\delta$ errors into the insertion-deletion channel, this means that the number of chunks that are error-bad is at most $2n \delta$. Additionally, by assumption, the number of zero-bad chunks is also at most $n\delta$.
We also must consider the fraction of rounds that are decoding-bad. In order to do this, we appeal to Theorem~\ref{thm:RSPDmisdecodings}, which guarantees that if an $\eps$-synchronization string of length $N$ is sent over an insertion-deletion channel with a $\delta'$ fraction of insertions and deletions, then the receiver will decode the index of the received symbol correctly for all but $2N\delta'/(1-\eps)$ symbols.
In this context, $N$ is the number of chunks, i.e. $N = n(1-2\delta)/(2r_c)$, and the fraction of chunks corrupted by errors is $\delta' = 4n\delta/
$. Therefore, the total number of bad chunks is at most $4 \delta n + 2N\delta'/(1-\eps) = 4 \delta n(3 - \eps)/(1-\eps)$.
In the rest of the proof, which is available in Appendix~\ref{app:chapter4}, we show that all but
$12\frac{3-\eps}{1-\eps} \cdot \delta n$
chunks are good chunks and have $\texttt{I}_A = \texttt{I}_B$ upon their arrival on Bob's side and we conclude that the simulated channel is a
$\left( \frac{3-\eps}{1-\eps} \frac{24\delta}{1 - 2\delta} r_c, r\right)$-block corruption channel. For the asymptotically optimal choice of $r = \sqrt{(1/\delta)\log (1 / \delta)}$, we derive the simulation described in the theorem statement.
\end{proof}
}
\global\def\ProofOfThmInterSim{
\fullOnly{\begin{proof}}
\shortOnly{\begin{proof}[Proof of Theorem~\ref{thm:InterSim}]}
Suppose Alice and Bob communicate via intermediaries $C_A$ and $C_B$ who act according to Algorithms~\ref{alg:InterC_A} and \ref{alg:InterC_B}. In total, Alice and Bob will attempt to communicate $n_s$ bits to one another over the simulated channel, while $C_A$ and $C_B$ communicate a total of $n$ bits to one another. The adversary is allowed to insert or delete up to $n\delta$ symbols and $C_A$ sends $n/2$ bits, so $C_B$ may receive between $n/2 -n\delta$ and $n/2+ n\delta$ symbols. To prevent $C_B$ from stalling indefinitely, $C_B$ only listens to the first $n(1-2\delta)/2$ bits he receives.
For $r = \sqrt{(1/\delta)\log (1 / \delta)}$, we define a \emph{chunk} to be $r_c := (s + |\Sigma_{syn}| + r/2)$ consecutive bits that are sent by $C_A$ to $C_B$. In particular, a chunk corresponds to a section header and synchronization symbol followed by $r/2$ rounds of messages sent from Alice. As $C_B$ cares about the first $n(1-2\delta)/2$ bits it receives, there are $\frac{n(1-2\delta)}{2r_c}$ chunks in total.
Hence,
$n_s = \frac{n(1-2\delta)}{2r_c} \cdot r$
since $C_B$ and $C_A$'s communication is alternating.
Note that if Alice sends a substring of form $10^{s-1}$ in the information part of a chunk, then Bob mistakenly detects a new block. With this in mind, we say a chunk is \emph{good} if:
\begin{enumerate
\item There are no errors injected in the chunk or affecting $C_B$'s detection of the chunk's header,
\item $C_B$ correctly decodes the index that $C_A$ sends during the chunk, and
\item $C_A$ does not send a $10^{s-1}$ substring in the information portion of the chunk.
\end{enumerate}
If a chunk is not good, we call it \emph{bad}. If the chunk is bad
because $C_B$ does not decode $C_A$'s index correctly even though they were in sync and no errors were injected, then we call it \emph{decoding-bad}. If it is bad because Alice sends a $10^{s-1}$ substring, we call it \emph{zero-bad} and otherwise, we call it \emph{error-bad}. Throughout the protocol, $C_B$ uses the variable $\texttt{I}_B$ to denote the next index of the synchronization string $C_B$ expects to receive and we use $\texttt{I}_A$ to denote the index of the synchronization string $C_A$ most recently sent. Notice that if a chunk is good and $\texttt{I}_A = \texttt{I}_B$, then all messages are correctly conveyed
We now bound the maximum number of bad chunks that occur over the course of the simulation. Suppose the adversary injects errors into the $i^{th}$ chunk, making that chunk bad. The $(i+1)^{th}$ chunk may also be bad, since Bob may not be listening for $10^{s-1}$ from $C_A$ when $C_A$ sends them, and therefore may miss the block header. However, if the adversary does not inject any errors into the $(i+1)^{th}$ and the $(i+2)^{th}$ chunk, then the $(i+2)^{th}$ chunk will be good.
In effect, a single error may render at most two chunks useless. Since the adversary may inject $n\delta$ errors into the insertion-deletion channel, this means that the number of chunks that are error-bad is at most $2n \delta$. Additionally, by assumption, the number of zero-bad chunks is also at most $n\delta$.
We also must consider the fraction of rounds that are decoding-bad. In order to do this, we appeal to Theorem~\ref{thm:RSPDmisdecodings}, which guarantees that if an $\eps$-synchronization string of length $N$ is sent over an insertion-deletion channel with a $\delta'$ fraction of insertions and deletions, then the receiver will decode the index of the received symbol correctly for all but $2N\delta'/(1-\eps)$ symbols.
In this context, $N$ is the number of chunks, i.e. $N = n(1-2\delta)/(2r_c)$, and the fraction of chunks corrupted by errors is $\delta' = 4n\delta/
$. Therefore, the total number of bad chunks is at most $4 \delta n + 2N\delta'/(1-\eps) = 4 \delta n(3 - \eps)/(1-\eps)$.
We will now use these bounds on the number of good and bad chunks to calculate how many errors there are for Alice and Bob, communicating over the simulated channel. As noted, so long as the chunk is good and $\texttt{I}_A = \texttt{I}_B$, then all messages are correctly conveyed to Alice from Bob and vice versa. Meanwhile, a single error, be it an adversarial injection or a synchronization string improper decoding, may cause $|\texttt{I}_B - \texttt{I}_A|$ to increase by at most two since a single error may cause Bob to erroneously simulate Step~\ref{step:sync_symb} and therefore increment $\texttt{I}_B$ by two when, in the worst case, $\texttt{I}_A$ does not change. On the other hand, if $|\texttt{I}_B - \texttt{I}_A| \geq 1$ and the chunk is good, this difference will decrease by at least 1, as is clear from Lines~\ref{step:lt} and \ref{step:gt} of Algorithm~\ref{alg:InterC_B}.
In total, we have that over the course of the computation,
$|\texttt{I}_B - \texttt{I}_A|$ increases at most
$\frac{3 - \eps}{1-\eps} \cdot 4 \delta n$
times and each time by at most 2.
Therefore, there will be at most
$\frac{3 - \eps}{1-\eps} \cdot 8 \delta n$
good chunks during which $|\texttt{I}_B - \texttt{I}_A| \geq 1$.
This gives that all but
$\frac{3 - \eps}{1-\eps} \cdot 12 \delta n$
chunks are good chunks and have $\texttt{I}_A = \texttt{I}_B$ upon their arrival on Bob's side.
Remember that the total number of chunks is $\frac{n(1-2\delta)}{2r_c}$, hence, the simulated channel is a
$\left(\frac{12 \delta n(3 - \eps)/(1-\eps)}{n(1-2\delta)/(2r_c)}, r\right) = \left( \frac{24\delta r_c(3-\eps)}{(1-\eps)(1 - 2\delta)}, r\right)$
block corruption channel.
Thus far, we have shown that one can simulate $n(1-2\delta)\frac{r}{2r_c}$ rounds of a $\left( \frac{24\delta r_c(3-\eps)}{(1-\eps)(1 - 2\delta)}, r\right)$-block corruption channel over a given channel as described in the theorem statement. More specifically, over $n(1 - 2\delta)$ rounds of communication over the insertion-deletion channel, a $\frac{2r_c - r}{2r_c} = \frac{s + |\Sigma_{syn}|}{r_c}$ fraction of rounds are used to add headers and synchronization symbols to chunks and a $\frac{24\delta r_c(3-\eps)}{(1-\eps)(1 - 2\delta)}$ fraction can be lost due to the errors injected by the adversary or $10^{s-1}$ strings in Alice's stream of bits. Therefore, the overall fraction of lost bits in this simulation is
$\frac{s + |\Sigma_{syn}|}{r_c} + \frac{24\delta r_c(3-\eps)}{(1-\eps)(1 - 2\delta)}$.
Since $s = c\log\frac{1}{\delta}$, $\eps$ and $|\Sigma_{syn}|$ are constants, and $1-2\delta$ approaches to one for small $\delta$, the optimal asymptotic choice is $r = \sqrt
{(1/\delta)\log 1 / \delta
$. This choice gives a simulated channel with characteristics described in the theorem statement. This simulation is
performed efficiently if the synchronization symbols are efficiently computable.
\end{proof}}
\fullOnly{\ProofOfThmInterSim}
\ShortOnlyVspace{-2mm}
The simulation stated in Theorem~\ref{thm:InterSim} burdens an additional condition on Alice's stream of bits by requiring it to have a limited number of substrings of form $10^{s-1}$. We now introduce a high probability technique to modify a general interactive communication protocol in a way that makes all substrings of form $10^{s-1}$ in Alice's stream of bits fit into $n\delta$ intervals of length $r=\sqrt{(1/\delta)\log(1/\delta)}$.
\begin{lemma}\label{theorem:InterPreCodeOblivious}
Assume that $n$ rounds of a binary interactive insertion-deletion channel with an oblivious adversary who is allowed to inject $n\delta$ errors are given. There is a pre-coding scheme that can be utilized on top of the simulation introduced in Theorem~\ref{thm:InterSim}. It modifies the stream of bits sent by Alice so that with probability $1 - e^{-\frac{c-3}{2}n\delta\log\frac{1}{\delta}(1+o(1))}$, all substrings of form $10^{s-1}$ where $s = c\log(1/\delta)$ in the stream of bits Alice sends over the simulated channel can be covered by $n\delta$ intervals of length $r=\sqrt{(1/\delta)\log(1/\delta)}$.
This pre-coding scheme comes at the cost of a $\Theta(\sqrt{\delta\log(1/\delta)})$ fraction of the bits Alice sends through the simulated channel.
\end{lemma}
\global\def\InterPreCodeOblivious{
\shortOnly{\begin{proof}[Proof of Lemma~\ref{theorem:InterPreCodeOblivious}]}
\fullOnly{\begin{proof}}
Note that in the simulation process, each $\frac{r}{2}$ consecutive bits Alice sends will form one of the chunks $C_A$ sends to $C_B$ alongside some headers. The idea of this pre-coding is simple. Alice uses the first $\frac{s}{2}$ data bits (and not the header) of each chunk to share $\frac{s}{2}$ randomly generated bits with Bob (instead of running the interactive protocol) and then both of them extract a string $S'$ of $\frac{r}{2}$ $\frac{s}{2}$-wise independent random variables. Then, Alice XORs the rest of data bits she passes to $C_A$ with $S'$ and Bob XORs those bits with $S'$ again to retrieve the original data.
We now determine the probability that the data block of a chunk of the simulation contains a substring of form $10^{s-1}$. Note that if a block of size $\frac{r}{2}$ contains a $10^{s-1}$ substring, then one of its substrings of length $\frac{s}{2}$ starting at positions $0, \frac{s}{2}, \frac{2s}{2}, \cdots$ is all zero. Since $P$ is $\frac{s}{2}$-wise independent, the probability of each of these $\frac{s}{2}$ substrings containing only zeros is $2^{-\frac{s}{2}} = \delta^\frac{c}{2}$. Taking a union bound over all these substrings, the probability of a block containing a $10^{s-1}$ substring can be bounded above by
$$p = \frac{r/2}{s/2} \cdot \delta^\frac{c}{2} = \sqrt{\frac{\delta^{c-1}}{\log(1/\delta)}}.$$
Now, we have:
\begin{eqnarray*}
\Pr\left\{\text{Number of blocks containing }10^{s-1} > n\delta \right\} &<& {\frac{n}{s} \choose n\delta} p^{n\delta}\le \left(\frac{ne}{ns\delta}\right)^{n\delta}\left(\sqrt{\frac{\delta^{c-1}}{\log(1/\delta)}}\right)^{n\delta} \\
&<& \left(\frac{\delta^{(c-3)/2}e}{c\log^{3/2}(1/\delta)}\right)^{n\delta} =
e^{-\frac{c-3}{2}n\delta\log\frac{1}{\delta}(1+o(1))}.
\end{eqnarray*}
\end{proof}
}
\fullOnly{\InterPreCodeOblivious}
\shortOnly{\begin{proof}[Proof sketch]
In the simulation process, each $r/2$ consecutive bits Alice sends forms one of the chunks $C_A$ sends to $C_B$ alongside some headers. The idea of this pre-coding scheme is simple. Alice uses the first $s/2$ data bits (and not the header) of each chunk to share $s/2$ randomly generated bits with Bob (instead of running the interactive protocol) and then both of them extract a string $S'$ of $r/2$ $(s/2)$-wise independent random variables. Then, Alice XORs the rest of data bits she passes to $C_A$ with $S'$ and Bob XORs those bits with $S'$ again to retrieve the original data. In Appendix~\ref{app:chapter4}, we show that this pre-coding scheme guarantees the requirements mentioned in the theorem statement.
\end{proof}}
\ShortOnlyVspace{-2mm}
Applying this pre-coding
for $c \ge 3$ on top of the simulation from Theorem~\ref{thm:InterSim} implies the following.
\begin{theorem}\label{thm:ObliviousGeneralSimulation}
Suppose that $n$ rounds of a binary interactive insertion-deletion channel with a $\delta$ fraction of insertions and deletions performed by an oblivious adversary are given.
For sufficiently small $\delta$, it is possible to simulate
$n(1-\Theta( \sqrt{\delta\log(1/\delta)}))$
rounds of a binary interactive
$
(\Theta(\sqrt{\delta\log(1/\delta)}), \sqrt{(1/\delta)\log 1 / \delta})$-block corruption channel between two parties over the given channel.
The simulation works with probability $1 - \exp(-\Theta(n\delta\log (1/\delta)))$ and is efficient if the synchronization string is efficient.
\end{theorem}
\begin{lemma}\label{theorem:InterPreCodeFully}
Suppose that $n$ rounds of a binary, interactive, fully adversarial insertion-deletion channel with a $\delta$ fraction of insertions and deletions are given.
The pre-coding scheme proposed in Lemma~\ref{theorem:InterPreCodeOblivious} ensures that the stream of bits sent by Alice contains fewer than $n\delta$ substrings of form $10^{s-1}$ for $s = c\log(1/\delta)$ and $c>5$ with probability $1 - e^{-\Theta\left(n\delta\log (1/\delta)\right)}$.
\end{lemma}
\global\def\InterPreCodeFullyProof{
\shortOnly{\begin{proof}[Proof of Lemma~\ref{theorem:InterPreCodeFully}]}
\fullOnly{\begin{proof}}
Lemma~\ref{theorem:InterPreCodeOblivious} ensures that for a fixed adversarial error pattern, the stream of bits sent by Alice contains fewer than $n\delta$ substrings of form $10^{s-1}$ upon applying the pre-coding scheme. However, in the fully adversarial setting, the adversary need not fix the error pattern in advance. Since the communication is interactive, the adversary can thus adaptively alter the bits Alice chooses to send. In this proof, we take a union bound over all error patterns with a $\delta$ fraction of errors and show that with high probability, upon applying the pre-coding scheme, the stream of bits sent by Alice contains fewer than $\Theta(n\delta)$ substrings of form $10^{s-1}$.
We claim that the number of error patterns with exactly $k$ insertions or deletions is at most $3^k{n+k\choose k}$. Note that if symbols $s_1, \dots, s_n$ are being sent, each of the $k$ errors can potentially occur within the intervals $[s_1, s_2), [s_2, s_3), \dots, [s_{n-1}, s_n)$, or after $s_n$ is sent. Each error could be a deletion, insertion of ``1'', or insertion of ``0''. This gives the claimed error pattern count. Further, any error pattern with fewer than $k-1$ errors can be thought of as an error pattern with either $k-1$ or $k$ errors where the adversary deletes an arbitrary set of symbols and then inserts the exact same symbols immediately.
Therefore, the number of all possible error patterns with at most $n\delta$ insertions or deletions can be upper-bounded by
$$\sum_{k=n\delta-1}^{n\delta}3^k{n+k\choose k} \le 2\cdot3^{n\delta} {n(1+\delta)\choose n\delta} \le 2\cdot3^{n\delta}\left(\frac{n(1+\delta)e}{n\delta}\right)^{n\delta} < 2\left(\frac{6e}{\delta}\right)^{n\delta} = e^{n\delta\log\frac{1}{\delta}(1+o(1))}.$$
Now, since summation of $\exp\left(n\delta\log(1/\delta)(1+o(1))\right)$ many probabilities any of which smaller than $\exp\left(-\frac{c-3}{2}n\delta\log(1/\delta)(1+o(1))\right)$ is still $\exp\left(-\Theta\left(n\delta\log (1/\delta)\right)\right)$ for $c > 5$, the probability of this pre-coding making more than $n\delta$ disjoint $10^{s-1}$ substrings for $s = c\log\frac{1}{\delta}$ in fully adversarial setting is again $1 - \exp\left(-\Theta\left(n\delta\log (1/\delta)\right)\right)$.
\end{proof}
}
\fullOnly{\InterPreCodeFullyProof}
\global\def\NonObliviousGeneralSimulation{
\begin{theorem}\label{thm:NonObliviousGeneralSimulation}
Suppose that $n$ rounds of a binary interactive insertion-deletion channel with a $\delta$ fraction of insertions and deletions performed by a non-oblivious adversary are given. For a sufficiently small $\delta$, it is possible to simulate
$n\left(1-\Theta\left( \sqrt{\delta\log(1/\delta)}\right)\right)$
rounds of a binary interactive
$
\left(\Theta\left(\sqrt{\delta\log(1/\delta)}\right), \sqrt\frac{\log 1 / \delta}{\delta}\right)$-block corruption channel between two parties, Alice and Bob, over the given channel.
The simulation is efficient if the synchronization string is efficient and works with probability $1 - \exp\left(-\Theta\left(n\delta\log (1/\delta)\right)\right)$.
\end{theorem}
}
\ShortOnlyVspace{-1mm}
Theorem~\ref{thm:InterSim} and Lemma~\ref{theorem:InterPreCodeFully} allow us to conclude that one can perform the simulation stated in Theorem~\ref{thm:InterSim} over any interactive protocol with high probability (see Theorem~\ref{thm:NonObliviousGeneralSimulation}).
\fullOnly{\NonObliviousGeneralSimulation}\shortOnly{Note that one can trivially extend the results of Theorems~\ref{thm:InterSim}~and~\ref{thm:ObliviousGeneralSimulation} to one-way binary communication by ignoring the bits Bob sends. In Appendix~\ref{app:BinaryOneWay}, we discuss the analogies of these theorems in the one-way setting.}
\global\def\BinaryOneWayCommunicationSection{
\subsection{Binary One Way Communication}\label{app:BinaryOneWay}
It is trivial to simplify Algorithms~\ref{alg:InterC_A} and \ref{alg:InterC_B} from Section~\ref{sec:simulation} to prove our simulation guarantees over binary alphabets. Specifically, the messages that Bob sends may be completely ignored and thereby we immediately obtain the following result for one-way insertion-deletion channels:
\begin{theorem}\label{thm:OneWayBinSimul}
Suppose that $n$ rounds of a binary one-way insertion-deletion channel with a $\delta$ fraction of insertions and deletions are given.
For a sufficiently small $\delta$, it is possible to deterministically simulate
$n\left(1-\Theta\left( \sqrt{\delta\log(1/\delta)}\right)\right)$
rounds of a binary
$$\left(\Theta\left(\sqrt{\delta\log\frac{1}{\delta}}\right), \sqrt \frac{\log (1/\delta)}{\delta}\right)$$
one-way block corruption channel between two parties, Alice and Bob, over the given channel assuming that all substrings of form $10^{s-1}$ for $s = c\log \frac{1}{\delta}$ in Alice's stream of bits can be covered by $n\delta$ intervals of $\sqrt{\frac{\log(1/\delta)}{\delta}}$ consecutive rounds.
This simulation is performed efficiently if the synchronization string is efficient.
\end{theorem}
Further, one can use the pre-coding technique introduced in Lemma~\ref{theorem:InterPreCodeOblivious} and Theorem~\ref{thm:OneWayBinSimul} to show that:
\begin{theorem}\label{thm:BinaryOneWayCompleteSimulation}
Suppose that $n$ rounds of a binary one-way insertion-deletion channel with a $\delta$ fraction of insertions and deletions are given.
For sufficiently small $\delta$, it is possible to simulate
$n\left(1-\Theta\left( \sqrt{\delta\log(1/\delta)}\right)\right)$
rounds of a binary
$\left(\Theta\left(\sqrt{\delta\log(1/\delta)}\right), \sqrt\frac{\log 1 / \delta}{\delta}\right)$
block corruption channel between two parties, Alice and Bob, over the given channel.
The simulation works with probability $1 - e^{-\Theta\left(n\delta\log(1/\delta)\right)}$, and is efficient if the synchronization string is efficient.
\end{theorem}
\section{Applications: Binary Insertion-Deletion Codes}\label{sec:insDelCodes}
The binary one-way simulation in Theorem~\ref{thm:BinaryOneWayCompleteSimulation} suggests a natural way to overcome insertion-deletion errors in one-way binary channels. One can simply simulate the corruption channel and use appropriate corruption-channel error correcting codes on top of the simulated channel to make the communication resilient to insertion-deletion errors. However, as the simulation works with high probability, this scheme is not deterministic.
As preserving the streaming quality is not necessary for the sake of designing binary insertion-deletion codes, we can design a deterministic pre-coding and error correcting code that can be used along with the deterministic simulation introduced in Theorem~\ref{thm:OneWayBinSimul} to generate binary insertion-deletion codes.
\begin{lemma}\label{lemma:blockCorCoding}
There exist error correcting codes for
$\left(\Theta\left(\sqrt{\delta\log(1/\delta)}\right), \sqrt {\log (1/\delta)/\delta}\right)$
block corruption channels with rate $1-\Theta(\sqrt{\delta\log(1/\delta)}) - \delta^{c/2}$ whose codewords are guaranteed to be free of substrings of form $10^{s-1}$ for $c\log\frac{1}{\delta}$.
\end{lemma}
\begin{proof
Assume that we are sending $n$ bits over a
$$(p_b, r_b) = \left(\Theta\left(\sqrt{\delta\log\frac{1}{\delta}}\right), \sqrt \frac{\log (1/\delta)}{\delta}\right)$$
block corruption channel. Let us chop the stream of bits into blocks of size $r_b$. Clearly, the fraction of blocks which may contain any errors is at most $2p_b$. Therefore, by looking at each block as a symbol from a large alphabet of size $2^{r_b}$, one can protect them against $2p_b$ fraction of errors by having $\Theta(H_{2^{r_b}}(2p_b) + p_b)$ fraction of redundant blocks.
In the next step, we propose a way to efficiently transform each block of length $r_b$ of the encoded string into a block of $r_b(1+\delta^{c/2})$ bits so that the resulting stream of bits be guaranteed not to contain $10^{s-1}$ substrings.
To do so, we think of each block of length $r_b$ as a number in the range of zero to $2^{r_b}-1$. Then, we can represent this number in $2^{s/2} - 1$ base and then map each of the symbols of this presentation to strings of $\frac{s}{2}$ bits except the $\frac{s}{2}$ all-zero string. This way, one can efficiently code the string into a stream of bits free of $10^{s-1}$ substrings by losing a $2^{-s/2}$-fraction of bits.
On the other side of the channel, Bob has to split the stream he receives into blocks of length $\frac{s}{2}$, undo the map to find out a possibly corrupted version of the originally encoded message and then decode the $2^{r_b}$-sized alphabet error correcting code to extract Alice's message.
This introduces an insertion-deletion code with rate of $1-\Theta(H_{2^{r_b}}(2p_b) + p_b)-2^{-s/2}$. As $s = c \log\frac{1}{\delta}$, $2^{-s/2} = \delta^{c/2}, p_b = \sqrt{\delta\log\frac{1}{\delta}}$, and
$$H_{2^{r_b}}(2p_b) = \Theta(2p_b \log_{2^{r_b}} (1/2p_b)) = \Theta\left(\frac{2p_b \log_2 (1/2p_b)}{r_b}\right) = \Theta\left(\frac{\sqrt{\delta\log \frac{1}{\delta}}\log\frac{1}{\delta}}{\sqrt{\frac{\log(1/\delta)}{\delta}}}\right) = \Theta\left(\delta\log\frac{1}{\delta}\right)$$
the proof is complete.
\end{proof}
One can set $c \ge 1$ and then apply such error correcting codes on top of a simulated channel as described in Theorem~\ref{thm:OneWayBinSimul} to construct a binary insertion-deletion code resilient to $\delta$ fraction of insertions and deletions with rate $1-\Theta\left(\sqrt{\delta\log\frac{1}{\delta}}\right)$ for a sufficiently small $\delta$.
\begin{theorem}
There exists a constant $0<\delta_0 <1$ such that for any $0<\delta<\delta_0$ there is a binary insertion-deletion code with rate $1-\Theta\left(\sqrt{\delta\log\frac{1}{\delta}}\right)$ which is decodable from $\delta$ fraction of insertions and deletions.
\end{theorem}
In comparison to the previously known binary insertion-deletion codes\cite{guruswami2016, schulman1999asymptotically}, this code has an improved rate-distance trade-off up to logarithmic factors.
}
\fullOnly{\BinaryOneWayCommunicationSection}
\section{Applications: New Interactive Coding Schemes}\label{sec:codings}
\noindent\textbf{Efficient Coding Scheme Tolerating $1/44$ Fraction of Errors.}\label{sec:largeAlphabetCodingScheme}
In this section, we will provide an efficient coding scheme for interactive communication over insertion-deletion channels by first making use of large alphabet interactive channel simulation provided in Theorem~\ref{thm:InteractiveCnstSizeAlphaSimul} to effectively transform the given channel into a simple corruption interactive channel and then use the efficient constant-rate coding scheme of Ghaffari and Haeupler~\cite{haeupler2014optimalII} on top of the simulated channel. This will give an efficient constant-rate interactive communication over large enough constant alphabets as described in Theorem~\ref{thm:largeAlphaInteractiveCodingScheme}. We review the following theorem of Ghaffari and Haeupler~\cite{haeupler2014optimalII}
before proving Theorem~\ref{thm:largeAlphaInteractiveCodingScheme}.
\begin{theorem}[Theorem 1.1 from \cite{haeupler2014optimalII}]\label{GhaffariHaeupler2014Scheme}
For any constant $\eps > 0$ and $n$-round protocol $\Pi$ there is a randomized non-adaptive coding scheme that robustly simulates $\Pi$ against an adversarial error rate of $\rho \le 1/4 - \eps$ using $N = O(n)$ rounds, a near-linear $n\log^{O(1)} n$ computational complexity, and failure probability $2^{-\Theta(n)}$.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{thm:largeAlphaInteractiveCodingScheme}]
For a given insertion-deletion interactive channel over alphabet $\Sigma$ suffering from $\delta$ fraction of edit-corruption errors, Theorem~\ref{thm:InteractiveCnstSizeAlphaSimul} enables us to simulate
$n-2n\delta(1+(1-\eps')^{-1})$
rounds of ordinary interactive channel with
$\frac{2\delta(5-3\eps')}{1-\eps' +2\eps'\delta - 4\delta}$
fraction of symbol by designating $\log |\Sigma_{syn}|$ bits of each symbol to index simulated channel's symbols with an $\eps'$-synchronization string over $\Sigma_{syn}$.
One can employ the scheme of Ghaffari and Haeupler~\cite{haeupler2014optimalII} over the simulated channel as long as error fraction is smaller than $1/4$.
Note that
$\left.\frac{2\delta(5-3\eps')}{1-\eps' +2\delta\eps' - 4\delta}\right|_{\eps' = 0} = \frac{10\delta}{1 - 4\delta} < \frac{1}{4} \Leftrightarrow \delta < \frac{1}{44}.$
Hence, as long as $\delta = 1/44-\eps$ for $\eps > 0$, for small enough $\eps' = O_\eps(1)$, the simulated channel has an error fraction that is smaller than $1/4$. Therefore, by running the efficient coding scheme of Theorem~\ref{GhaffariHaeupler2014Scheme} over this simulated channel one gets a constant rate coding scheme for interactive communication that is robust against $1/44-\eps$ fraction of edit-corruptions. Note that this simulation requires the alphabet size to be large enough to contain synchronization symbols (which can come from a polynomially large alphabet in terms of $\eps'$ according to Lemma~\ref{lemma:FiniteSyncConstruction}) and also meet the alphabet size requirements of Theorem~\ref{GhaffariHaeupler2014Scheme}. This requires the alphabet size to be $\Omega_\eps(1)$, i.e., a large enough constant merely depending on $\eps$. The success probability and time complexity are direct consequences of Theorem~\ref{GhaffariHaeupler2014Scheme} and Theorem~\ref{thm:RSPDmisdecodings}.
\end{proof}
\medskip
\noindent\textbf{Efficient Coding Scheme with Near-Optimal Rate over Small Alphabets.}
In this section we present another insertion-deletion interactive coding scheme that achieves near-optimal communication efficiency as well as computation efficiency by employing a similar idea as in Section~\ref{sec:largeAlphabetCodingScheme}.
\shortOnly{In order to derive a rate-efficient interactive communication coding scheme over small alphabet insertion-deletion channels, Algorithms~\ref{alg:InterC_A} and \ref{alg:InterC_B} can be used to simulate a corruption channel and then the rate-efficient interactive coding scheme for corruption channels introduced by Haeupler \cite{haeupler2014interactive:FOCS} can be used on top of the simulated channel.
In Appendix~\ref{app:chapter5-interactive}, we provide a quick review of Haeupler's approach alongside its performance over block corruption channels and the proof of the following results.}
\global\def\HaeuplerFourteenRecap{
In order to derive a rate-efficient interactive communication coding scheme over small alphabet insertion-deletion channels, Algorithms~\ref{alg:InterC_A} and \ref{alg:InterC_B} can be used to simulate a corruption channel and then the rate-efficient interactive coding scheme for corruption channels introduced by Haeupler~\cite{haeupler2014interactive:FOCS} can be used on top of the simulated channel.
We start by a quick review of Theorems 7.1 and 7.2 of Haeupler~\cite{haeupler2014interactive:FOCS}.
\begin{theorem}[Theorem 7.1 from \cite{haeupler2014interactive:FOCS}]\label{thm:H14Thm7.1}
Suppose any $n$-round protocol $\Pi$ using any alphabet $\Sigma$. Algorithm 3 [from \cite{haeupler2014interactive:FOCS}] is a computationally efficient randomized coding scheme which given $\Pi$, with probability $1-2^{-\Theta(n\delta)}$, robustly simulates it over any oblivious adversarial error channel with alphabet $\Sigma$ and error rate $\delta$. The simulation uses $n(1 + \Theta(\sqrt{\delta}))$ rounds and therefore achieves a communication rate of $1-\Theta(\sqrt{\delta})$.
\end{theorem}
\begin{theorem}[Theorem 7.2 from \cite{haeupler2014interactive:FOCS}]
\label{thm:H14Thm7.2}
Suppose any $n$-round protocol $\Pi$ using any alphabet $\Sigma$. Algorithm 4 [from \cite{haeupler2014interactive:FOCS}] is a computationally efficient randomized coding scheme which given $\Pi$, with probability $1-2^{-\Theta(n\delta)}$, robustly simulates it over any fully adversarial error channel with alphabet $\Sigma$ and error rate $\delta$. The simulation uses $n(1 + \Theta\left(\sqrt{\delta\log\log \frac{1}{\delta}}\right)$ rounds and therefore achieves a communication rate of $1-\Theta\left(\sqrt{\delta\log\log\frac{1}{\delta}}\right)$.
\end{theorem}
The interaction between the error rate and rate loss provided in Theorems~\ref{thm:H14Thm7.1} and~\ref{thm:H14Thm7.2} (Theorems 7.1 and 7.2 of \cite{haeupler2014interactive:FOCS}) leads us to the following corollary.
\begin{corollary}\label{obs:interactive}
There are high-probability efficient coding schemes for interactive communication over insertion-deletion channels that are robust against $\delta$ fraction of edit-corruptions for sufficiently small $\delta$ and have communication rate of $1 - \Theta\left({\sqrt[4]{\delta\log\frac{1}{\delta}}}\right)$ against oblivious adversaries and
$1 - \Theta\left({\sqrt[4]{\delta\log\frac{1}{\delta} \log^2\log\frac{1}{\delta} }}\right)$
in fully adversarial setup.
\end{corollary}
However, these results can be improved upon by taking a closer look at the specifics of the interactive communication coding scheme in \cite{haeupler2014interactive:FOCS}.
In a nutshell, the interactive coding scheme proposed in \cite{haeupler2014interactive:FOCS} simulates an interactive protocol $\Pi$ by splitting the communication into iterations.
In each iteration, the coding scheme lets Alice and Bob communicate for a block of $\bar{r}$ rounds, then uses $r_c$ rounds of communication after each block so the parties can verify if they are on the same page and then decide whether to continue the communication or to roll back the communication for some number of iterations. The parties perform this verification by exchanging a fingerprint (hash) of their versions of the transcript. Next, each party checks if the fingerprint he receives matches his own, which in turn identifies whether the parties agree or disagree about the communication transcript. Based on the result of this check, each party decides if he should continue the communication from the point he is already at or if he should roll back the communication for some number of iterations and continue from there. (see Algorithms 3 and 4 of \cite{haeupler2014interactive:FOCS})
The analysis in Section 7 of Heaupler \cite{haeupler2014interactive:FOCS} introduces a potential function $\Phi$ which increases by at least one whenever a round of communication is free of any errors or hash collisions. A \emph{hash collision} occurs if the parties' transcripts do not agree due to errors that occurred previously, yet the two parties' fingerprints erroneously match. The analysis also shows that whenever a round of communication contains errors or hash collisions, regardless of the number of errors or hash collisions happening in a round, the potential drops by at most a fixed constant. (Lemmas 7.3 and 7.4 of \cite{haeupler2014interactive:FOCS})
For an error rate of $\bar\delta$, there can be at most $n\bar\delta$ rounds suffering from an error. Haeupler \cite{haeupler2014interactive:FOCS} shows that the number of hash collisions can also be bounded by $\Theta(n\bar\delta)$ with exponentially high probability. (Lemmas 7.6 and 7.7 and Corollary 7.8 of \cite{haeupler2014interactive:FOCS})
Given that the number of errors and hash collisions is bounded by $\Theta(n\bar\delta)$, Haeupler \cite{haeupler2014interactive:FOCS} shows that if $\Phi > \frac{n}{\bar{r}} + \Theta(n\bar\delta)$, then the two parties will agree on the first $n$ steps of communication and therefore the communication will be simulated thoroughly and correctly. Therefore, after $\frac{n}{\bar{r}} + \Theta(n\bar\delta)$ rounds the simulation is complete, and the rate of this simulation is $1 - \Theta\left(r\bar\delta + \frac{r_c}{\bar{r}}\right)$.
}
\fullOnly{\HaeuplerFourteenRecap}
\begin{theorem}[Interactive Coding against Block Corruption]\label{thm:InteractiveCodeBlock}
By choosing an appropriate block length in the Haeupler \cite{haeupler2014interactive:FOCS} coding scheme for oblivious adversaries (Theorem~\ref{thm:H14Thm7.1}), one obtains a robust efficient interactive coding scheme for $(\delta_b, r_b)$-block corruption channel with communication rate
$1-\Theta(\sqrt{\delta_b\max\left\{\delta_b, 1/r_b\right\}})$
that works with probability $1-2^{-\Theta(n\delta_b/r_b )}$.
\end{theorem}
\global\def\ProofOfThmInteractiveCodeBlock{
\shortOnly{\begin{proof}[Proof of Theorem~\ref{thm:InteractiveCodeBlock}]}
\fullOnly{\begin{proof}}
Let us run Haeupler \cite{haeupler2014interactive:FOCS} scheme with block size $\bar{r}$. This way, each block of corruption that rises in channel, may corrupt
$\max\left\{\frac{r_b}{\bar{r}}, 1\right\}$
blocks of Haeupler \cite{haeupler2014interactive:FOCS} scheme. Therefore, the total number of corrupted blocks in Haeupler \cite{haeupler2014interactive:FOCS} scheme can be:
$$\frac{n\delta_b}{r_b}\max\left\{\frac{r_b}{\bar{r}}, 1\right\}$$
Therefore, the total fraction of Haeupler \cite{haeupler2014interactive:FOCS} scheme's blocks containing errors is at most
$$\bar{\delta} = \delta_b\max\left\{1, \frac{\bar{r}}{r_b}\right\}$$
For oblivious adversaries, Haeupler \cite{haeupler2014interactive:FOCS} suggests a verification process which can be performed in $r_c = \Theta(1)$ steps at the end of each round. We use the exact same procedure. Lemma 7.6 of \cite{haeupler2014interactive:FOCS} directly guarantees that the fraction of hash collisions using this procedure is upper-bounded by $\Theta(\bar\delta)$ with probability $1 - 2^{-\Theta(n\bar{\delta}/\bar{r})}$. As the fraction of blocks suffering from hash collisions or errors is at most $\Theta(\bar{\delta})$, the communication can be shown to be complete in $(1+\bar{\delta})$ multiplicative factor of rounds by the same potential argument as in \cite{haeupler2014interactive:FOCS}.
Therefore, the rate lost in this interactive coding scheme is \[\Theta\left(\bar\delta + \frac{r_c}{\bar r}\right) = \Theta\left(\delta_b\max\left\{1, \frac{\bar{r}}{r_b}\right\} + \frac{1}{\bar r}\right)\]
Now, the only remaining task is to find $\bar r$ that minimizes the rate loss mentioned above.
If we choose $\bar r \le r_b$, the best choice is $\bar r = r_b$ as it reduces the rate to
$\delta_b + \frac{1}{r_b}$.
On the other hand if we choose $\bar r \ge r_b$, the optimal choice is $\bar r = \sqrt{\frac{r_b}{\delta_b}}$ if $\sqrt{\frac{r_b}{\delta_b}} \ge r_b \Leftrightarrow r_b \le \frac{1}{\delta_b}$ or $\bar r = r_b$ otherwise.
Hence, we set
$$
\bar r = \Bigg\{
\begin{tabular}{cc}
$\sqrt{\frac{r_b}{\delta_b}}$ & if $r_b \le \frac{1}{\delta_b}$ \\[2mm]
$r_b$ & $r_b > \frac{1}{\delta_b}$
\end{tabular}\\
$$
Plugging this values for $\bar r$ gives that:
$$
Rate = \Bigg\{
\begin{tabular}{cc}
$1-\Theta\left(\sqrt{\frac{\delta_b}{r_b}}\right)$ & $r_b \le \frac{1}{\delta_b}$ \\[2mm]
$1-\Theta\left(\delta_b\right)$ & $r_b > \frac{1}{\delta_b}$
\end{tabular} = 1-\Theta\left(\sqrt{\delta_b\max\left\{\delta_b, \frac{1}{r_b}\right\}}\right)\\
$$
Also, the probability of this coding working correctly is:
$$
1 - 2^{-\Theta(n\bar{\delta}/\bar{r})} = 1-2^{-\delta_b\max\left\{\frac{1}{\bar{r}} , \frac{1}{r_b}\right\}}= \Bigg\{
\begin{tabular}{cc}
$1-2^{-\delta_b\max\left\{\sqrt\frac{\delta_b}{r_b}, \frac{1}{r_b}\right\}}$ & $r_b \le \frac{1}{\delta_b}$ \\[2mm]
$1-2^{-\delta_b\max\left\{\frac{1}{r_b}, \frac{1}{r_b}\right\}}$ & $r_b > \frac{1}{\delta_b}$
\end{tabular} = 1-2^{-\Theta(\frac{\delta_b}{r_b}n)}\\
$$
\end{proof}
}
\fullOnly{\ProofOfThmInteractiveCodeBlock}
Applying the coding scheme of Theorem~\ref{thm:InteractiveCodeBlock} over the simulation from Theorem~\ref{thm:ObliviousGeneralSimulation} implies the following.
\begin{theorem}\label{thm:InterOblvAdv}
For sufficiently small $\delta$, there is an efficient interactive coding scheme over binary insertion-deletion channels which, is robust against $\delta$ fraction of edit-corruptions by an oblivious adversary, achieves a communication rate of $1 - \Theta({\sqrt{\delta\log(1/\delta)}})$, and works with probability $1 - 2^{-\Theta(n\delta)}$.
\end{theorem}
\begin{comment}
\begin{theorem}\label{thm:InterOblvAdv}
There is a coding scheme for interactive communication over an insertion-deletion channels with rate $1 - \Theta\left({\sqrt{\epsilon\log\frac{1}{\epsilon}}}\right)$ against oblivious adversaries.
\end{theorem}
\begin{proof}
We use the \cite{haeupler2014interactive:FOCS} coding scheme with round size of
$\bar{r} = \sqrt \frac{\log ({1}/{\epsilon})}{\epsilon}$
on top of the simulated channel.
Theorem \ref{thm:InterSim} guarantees that all of the errors that happen on the simulated channel can be covered by $\Theta(\epsilon n)$ intervals of length $r = \sqrt \frac{\log ({1}/{\epsilon})}{\epsilon}$.
Each of these intervals can intersect at most $\lceil\frac{r}{\bar r + r_c}\rceil + 1 = 2$ rounds of the \cite{haeupler2014interactive:FOCS} coding scheme.
Therefore, at most an $\Theta(\epsilon n)$ fraction of the interactive communication coding rounds can be rendered useless by errors.
For oblivious adversaries, \cite{haeupler2014interactive:FOCS} suggests a verification process which can be performed in $r_c = \Theta(1)$ steps at the end of each round. We use the exact same procedure. Lemma 7.6 of \cite{haeupler2014interactive:FOCS} directly guarantees that the number of hash collisions using this procedure is upper-bounded by $\Theta(n\epsilon)$ with probability $1 - 2^{-\Theta(n\epsilon)}$. As the number of rounds suffering from hash collisions or errors is at most $\Theta(n\epsilon)$, the communication can be shown to be complete in $\frac{n}{\bar{r}} + \Theta(n\epsilon)$ rounds by the same potential argument as in \cite{haeupler2014interactive:FOCS}.
Therefore, the rate lost in this interactive coding scheme is $\Theta(r\epsilon) + \frac{r_c}{\bar r} = O\left(\sqrt{\epsilon \log \frac{1}{\epsilon} }\right)$.
\end{proof}
\end{comment}
Moreover, \shortOnly{in Appendix~\ref{app:chapter5-interactive}, }we show that this result is extendable for the fully adversarial setup, as summarized in Theorem~\ref{thm:InterFullyAdv}.
\global\def\ProofOfInterFullyAdv{
\begin{proof}[Proof of Theorem~\ref{thm:InterFullyAdv}]
Similar to the proofs of Theorems~\ref{thm:InteractiveCodeBlock}~and~\ref{thm:InterOblvAdv}, we use the simulation discussed in Theorem~\ref{thm:NonObliviousGeneralSimulation} and coding structure introduced in Haeupler \cite{haeupler2014interactive:FOCS} with rounds of length $\bar{r} = \sqrt \frac{\log ({1}/{\delta})}{\delta}$ on top of that. However, this time, we use another verification strategy with length of $r_c = \log\log\frac{1}{\delta}$ which is used in Haeupler \cite{haeupler2014interactive:FOCS} as a verification procedure in the interactive coding scheme for fully adversarial channels.
The idea of this proof, similar to the proof of Theorem \ref{thm:InteractiveCodeBlock}, is to show that the number of rounds suffering from errors or hash collisions can be bounded by $\Theta(n\delta)$ with high probability and then apply the same potential function argument. All of the steps of this proof are, like their analogs in Theorem \ref{thm:InteractiveCodeBlock}, implications of Haeupler\cite{haeupler2014interactive:FOCS}, except for the fact that the number of hash collisions can be bounded by $\Theta(n\delta)$.
This is because of the fact that the entire Haeupler \cite{haeupler2014interactive:FOCS} analysis, except Lemma 7.7 that bounds hash collisions, are merely based on the fact that all but $O(n\delta)$ rounds are error free.
Therefore, if we show that the number of hash collisions in the fully adversarial case is bounded by $\Theta(n\delta)$, the simulation rate will be
$$1 - \left(\frac{\Theta(n\delta)}{n/r} + \frac{r_c}{\bar r}\right) = 1 - \Theta\left(\sqrt \frac{\log ({1}/{\delta})}{\delta} \cdot \delta + \frac{\log \log \frac{1}{\delta}}{\sqrt \frac{\log ({1}/{\delta})}{\delta}}\right) = 1 - \Theta\left(\sqrt{\delta \log \frac{1}{\delta} }\right)$$
and the proof will be complete.
We now bound the number of hash collisions in our interactive coding scheme. The verification process for fully adversarial setting uses a two level hash function $hash_2(hash_1(\cdot))$, where the seed of $hash_1$ is randomly generated in each round.
Lemma 7.7 of Haeupler \cite{haeupler2014interactive:FOCS} implies that for any oblivious adversary, the number of hash collisions due to $hash_1$ is at most $\Theta(n\delta)$ with probability $1-\delta^{\Theta(n\delta)}$. To find a similar bound for non-oblivious adversaries, we count the number of all possible oblivious adversaries and then use a union bound.
The number of oblivious adversaries is shown to be less than
$\left(\frac{3e}{\delta}\right)^{n\delta} = e^{n\delta\log(1/\delta)(1+o(1))}$
in Lemma~\ref{theorem:InterPreCodeFully}.
Hence, the probability of having more than $\Theta(n\delta)$ $hash_1$ hash collisions in any fully adversarial scenario is at most $2^{-\Theta(n\delta)}$. Now, Corollary 7.8 of Haeupler \cite{haeupler2014interactive:FOCS} can be directly applied to show that the number of $hash_2$ hash collisions can also be bounded by $\Theta(n\delta)$ with probability $1-2^{-\Theta(n\delta)}$. This completes the proof.
\end{proof}
}
\fullOnly{\ProofOfInterFullyAdv}
This insertion-deletion interactive coding scheme is, to the best of our knowledge, the first to be computationally efficient, to have communication rate approaching one, and to work over arbitrarily small alphabets.
\section{Synchronization Strings and Edit-Distance Tree Codes}\label{sec:SynchAndTreeCodes}
We start by providing a new upper bound on the error tolerance of Braverman et al.'s coding scheme for interactive communication with a large alphabet over a insertion-deletion channel \cite{braverman2015coding}. We tweak the definition of edit-distance tree codes, the primary tool that Braverman et al. use in their coding scheme. In doing so, we show that their scheme has an error tolerance of $1/10 - \eps$ rather than $1/18-\eps$, which is the upper bound provided by Braverman et al. In particular, we prove the following theorem, which is a restatement of Theorem 1.4 from Braverman et al.'s work except for the revised error tolerance. The proof can be found in Section~\ref{sec:revisedDef}.
\begin{theorem}\label{thm:improvedBGMO}
For any $\eps > 0$, and for any binary (noiseless) protocol $\Pi$ with communication $CC(\Pi)$, there exists a noise-resilient coding scheme with communication $O_{\eps}(CC(\Pi))$ that succeeds in simulating $\Pi$ as long as the adversarial edit-corruption rate is at most $1/10-\eps$.
\end{theorem}
We then review the definition and key characteristics of edit-distance tree codes and discuss the close relationship between edit-distance tree codes and synchronization strings using the revised definition of edit-distance tree codes.
\subsection{Revised definition of edit-distance tree codes}\label{sec:revisedDef}
Before proceeding to the definition of edit-distance tree codes, we begin with some definitions.
\begin{definition}[Prefix Code, Definition 3.1 of \cite{braverman2015coding}]
A prefix code $C:\Sigma_{in}^{n}\rightarrow\Sigma_{out}^{n}$ is a code such that $C(x)[i]$ only depends on $x[1,i]$. $C$ can also be considered as a $|\Sigma_{in}|$-ary tree of depth $n$ with symbols written on edges of the tree using the alphabet $\Sigma_{out}$. On this tree, each tree path from the root of length $l$ corresponds to a string $x \in \Sigma_{out}^l$ and the symbol written on the deepest edge of this path corresponds to $C(x)[l]$.
\end{definition}
\begin{definition}[$\eps$-bad Lambda, Revision of Definition 3.2 of \cite{braverman2015coding}]\label{def:bad_lambda}
We say that a prefix code $C$ contains an $\eps$-bad lambda if when this prefix code is represented as a tree, there exist four tree nodes $A,B,D,E$ such that $B \not= D$, $B \not= E$, $B$ is $D$ and $E$'s common ancestor, $A$ is $B$'s ancestor or $B$ itself, and $ED(AD, BE) \le (1-\eps) \cdot (|AD| + |BE|)$.\footnote{Braverman et al. say that $A,B,D,E$ form an $\eps$-bad lambda if $ED(AD, BE) \le (1-\eps) \cdot \max(|AD|, |BE|)$.} Here $AD$ and $BE$ are strings of symbols along the tree path from $A$ to $D$ and the tree path from $B$ to $E$. See Figure~\ref{fig:pathtree} for an illustration.
\end{definition}
\begin{definition}[Edit-distance Tree Code, Definition 3.3 of \cite{braverman2015coding}]
We say that a prefix code $C: \Sigma^n_{in} \rightarrow \Sigma^n_{out}$ is an $\eps$-edit-distance tree code if $C$ does not contain an $\eps$-bad lambda.
\end{definition}
Using this revised definition of edit-distance tree code, we prove Theorem~\ref{thm:improvedBGMO} as follows:
\begin{proof}[Proof of Theorem~\ref{thm:improvedBGMO}]
We relax Braverman et al.'s definition of an \emph{edit-distance tree code} \cite{braverman2015coding}, which is an adaptation of Schulman's original tree codes \cite{schulman1993deterministic}. Edit-distance tree codes are parameterized by a real value $\eps$, and Braverman et al. define an $\eps$-edit-distance tree code to be a prefix code $C: \Sigma^n_{in} \rightarrow \Sigma^n_{out}$ that does not contain what they refer to as an \emph{$\eps$-bad lambda.} The code $C$ contains an $\eps$-bad lambda if when one considers $C$ as a tree, there exist four tree nodes $A,B,D,E$ such that $B \not= D$, $B \not= E$, $B$ is $D$ and $E$’s common ancestor, $A$ is $B$’s ancestor or $B$ itself, and $ED(AD, BE) \le (1-\eps) \cdot \max(|AD|, |BE|)$. Here $AD$ and $BE$ are strings of symbols along the tree path from $A$ to $D$ and the tree path from $B$ to $E$. See Figure~\ref{fig:pathtree} for an illustration. We relax the definition of an $\eps$-bad lambda to be any four tree nodes $A,B,D,E$ as above such that $ED(AD, BE) \le (1-\eps) \cdot (|AD|+ |BE|)$.
We use synchronization strings together with Schulman's original tree codes to prove that edit-distance tree codes, under this revised definition, exist. In particular, in Theorem~\ref{thm:syncandtreecodes}, we show that synchronization strings can be concatenated with tree codes to produce edit-distance tree codes. Given that tree codes and synchronization strings exist, this means that edit-distance tree codes according to our revised definition exist.
As we saw in Lemma~\ref{lemma:treecodes_SD}, if $C: \Sigma^n_{in} \rightarrow \Sigma^n_{out}$ is an $\eps$-edit-distance tree code and $\tilde{c} \in \Sigma^m_{out}$, then there exists at most one $c \in \cup_{i = 1}^n C\left(\Sigma_{in}^n\right)[1,i]$ such that $SD(c, \tilde{c}) \leq 1-\eps$. This leads to an improved version of Lemma 4.2 from Braverman et al.'s work, which we describe after we set up notation. In Braverman et al.'s protocol, let $N_A$ and $N_B$ be the number of messages Alice and Bob have sent when one of them reaches the end of the protocol. Let $s_A$ and $r_A$ be the messages Alice has sent and received over the course of the protocol, and let $s_B$ and $r_B$ be the messages Bob has sent and received over the course of the protocol. Let $\tau_A = (\tau_1, \tau_2)$ be the string matching between $s_B[1,N_B]$ and $r_A[1, N_A]$ and let $\tau_B = (\tau_3, \tau_4)$ be the string matching between $s_A[1, N_A]$ and $r_B[1, N_B]$. Braverman et al. call a given round a \emph{good decoding} if Alice correctly decodes the entire set of tree code edges sent by Bob. Lemma~\ref{lemma:treecodes_SD} admits the following improved version of Lemma 4.2 from Braverman et al's work:
\begin{lemma}[Revision of Lemma 4.2 from \cite{braverman2015coding}]\label{lem:improved4.2}
Alice has at least $N_A + \left(1 - \frac{1}{1-\eps}\right)sc(\tau_2) - \left(1 + \frac{1}{1-\eps}\right)sc(\tau_1)$ good decodings. Bob has at least $N_B + \left(1 - \frac{1}{1-\eps}\right)sc(\tau_4) - \left(1 + \frac{1}{1-\eps}\right)sc(\tau_3)$ good decodings.
\end{lemma}
The proof of this revised lemma follows the exact same logic as the proof of Lemma 4.2 in Braverman et al.'s work. The theorem statement now follows the same logic as the proof of Theorem 4.4 in Braverman et al.'s work. We include the revised section of the proof below, where the only changes are the maximum error tolerance $\rho$, the value of $n$ (which is smaller by a factor of $\frac{1}{2}$), and several equations. We set $\rho = \frac{1}{10}-\eps$ and $n = \lceil\frac{T}{8\eps}\rceil$, where $T$ is the length of the original protocol. Let $g_A$ be the number of good decodings of Alice and $b_A = N_A - g_A$ be the number of bad decodings of Alice. Similarly, let $g_B$ be the number of good decodings of Bob and $b_B = N_B - g_B$ be the number of bad decodings of Bob. In Braverman et al.'s protocol, Alice and Bob share an edit-distance tree code $C: \Sigma^n_{in} \rightarrow \Sigma^n_{out}$. By definition of edit distance, $sc(\tau_1) + sc(\tau_3) = sc(\tau_2) + sc(\tau_4) \leq 2\rho n$. By Lemma~\ref{lem:improved4.2}, \begin{align*}
b_A + b_B &\leq \frac{1}{1-\eps}(sc(\tau_1) + sc(\tau_2)) + sc(\tau_1) - sc(\tau_2) + \frac{1}{1-\eps}(sc(\tau_3) + sc(\tau_4)) + sc(\tau_3) - sc(\tau_4)\\
&\leq \frac{4 \rho n}{1-\eps}.
\end{align*} Therefore,
\begin{align*}
g_A &= N_A - b_A \geq N_A - \frac{4\rho n}{1-\eps} = (N_A - n(1-2\rho) + n(1-2\rho) + n(1-2\rho) - \frac{4\rho n}{1-\eps}\\
& \geq b_A + b_B - \frac{4 \rho n}{1-\eps} + (N_A - n(1-2\rho)) + n(1-2\rho) - \frac{4\rho n}{1-\eps}\\
& = b_A + b_B + (N_A - n(1-2\rho)) + n\left(1- \frac{8\rho}{1-\eps} - 2\rho\right)\\
& \geq b_A + b_B + (N_A - n(1-2\rho)) + n\left(1- \frac{8\rho}{1-\eps} - 2\rho\right)\\
& \geq b_A + b_B + (N_A - n(1-2\rho)) + 8\eps n \geq b_A + b_B + (N_A - n(1-2\rho)) + T.
\end{align*} The remainder of the proof follows exactly as is the proof of Theorem 4.4 from Braverman et al.'s work.
\end{proof}
Suppose that Alice uses an $\eps$-edit-distance tree code $C: \Sigma^n_{in} \rightarrow \Sigma^n_{out}$ to send a message $c$ to Bob. In the following theorem, we show that if $SD(c,\tilde{c}) < 1 - \eps$, then Bob can decode Alice's message. Braverman et al. proved a weaker version of this theorem \cite{braverman2015coding}, but our revised definition of an edit-distance tree code easily admits the follow lemma.
\begin{lemma}[Revision of Lemma 3.10 from \cite{braverman2015coding}]\label{lemma:treecodes_SD}
Let $C: \Sigma^n_{in} \rightarrow \Sigma^n_{out}$ be an $\eps$-edit-distance tree code, and let $\tilde{c} \in \Sigma^m_{out}$. There exists at most one $c \in \cup_{i = 1}^n C\left(\Sigma_{in}^n\right)[1,i]$ such that $SD(c, \tilde{c}) \leq 1-\eps$.
\end{lemma}
\begin{proof}
The proof follows exactly the same logic as the proof of Lemma 3.10 from \cite{braverman2015coding}. The proof begins by assuming, for a contradiction, that there exist two messages $c,c' \in \cup_{i = 1}^n C\left(\Sigma_{in}^n\right)[1,i]$ such that both $SD(c, \tilde{c}) \leq (1-\eps)$ and $SD(c', \tilde{c}) \leq (1-\eps)$. Then, the proof reaches its conclusion by the exact same logic.
\end{proof}
In Theorem~\ref{thm:extension}, we will show that one can extend an $\eps$-synchronization string $S$ to a $(2\eps-\eps^2)$-edit-distance tree code. This implies the following corollary of Lemma~\ref{lemma:treecodes_SD}.
\begin{corollary}
Let $S\in\Sigma^n$ be an $\eps$-synchronization string, and $c \in \Sigma^m$. There exists at most one prefix $\tilde{c}$ of $S$ for which $SD(c, \tilde{c})\le (1-\eps)^2$.
\end{corollary}
\subsection{Edit-distance tree codes and synchronization strings}
We prove that under the revised definition, edit-distance tree codes still exist by relating edit-distance tree codes to Schulman's original tree codes. We introduce a method to obtain an edit-distance tree code using ordinary tree codes and synchronization strings. Intuitively, using synchronization strings and tree codes together, one can overcome the synchronization problem using the synchronization string and then overcome the rest of the decoding challenges using tree codes, which are standard tools for overcoming corruption errors. Tree codes are defined as follows.
\begin{definition}[Tree Codes, from \cite{schulman1996Interactive}]
A $d$-ary tree code over an alphabet $\Sigma$ of distance parameter $\alpha$ and depth $n$ is a $d$-ary tree of depth $n$ in which every arc of the tree is labeled with a character from the alphabet $\Sigma$ subject to the following condition. Let $v_1$ and $v_2$ be any two nodes at some common depth $h$ in the tree. Let $h - l$ be the depth of their least common ancestor. Let $W(v_1)$ and $W(v_2)$ be the concatenation of the letters on the arcs leading from the root to $v_1$ and $v_2$ respectively. Then $\Delta(W(v_1), W(v_2)) \ge \alpha l$.
\end{definition}
Schulman \cite{schulman1993deterministic,schulman1996Interactive} proved that for any $\alpha < 1$, there exists a $d$-ary tree code with distance $\alpha$ and infinite depth over an alphabet of size $(cd)^{1/(1-\alpha)}$, for some constant $c < 6$. We now prove that edit-distance tree codes can be obtained using synchronization strings and tree codes.
\begin{theorem}\label{thm:syncandtreecodes}
Let $T$ be a tree code of depth $n$ and distance parameter $\left(1-\alpha\right)$ and let $S$ be an $\eps$-synchronization string of length $n$. Let the tree $T'$ be obtained by concatenating all edges on $i^{th}$ level of $T$ with $S[i]$. Then $T'$ is a $(1-\eps-\alpha)$-edit-distance tree code.
\end{theorem}
\begin{proof}
To prove that $T'$ is a $(1-\eps-\alpha)$-edit-distance tree code, we show that $T'$ does not contain any $(1-\eps-\alpha)$-bad lambdas. Let $A, B, D$, and $E$ be nodes in $T'$ that form a lambda structure.
We refer to the string consisting of edge labels on the path from node $A$ to $D$ by $AD$ and similarly we refer to the string consisting of labels on the path from $B$ to $E$ by $BE$. Let $\tau$ be the string matching characterizing the edit distance of $AD$ and $BE$. We call $i$ a \emph{matching position} if $\tau_1[i] = \tau_2[i]$. Further, we call $i$ a \emph{same-level matching position} if $\tau_1[i]$ and $\tau_2[i]$ correspond to symbols on the same level edges and \emph{disparate-level matching position} otherwise.
By the definition of a tree code, the number of same-level matching positions is at most
\[\alpha \min\left\{|BD|, |BE|\right\} \leq \alpha (|AD| + |BE|).\]
Further, Haeupler and Shahrasbi prove that for any monotone matching between an $\eps$-synchronization string and itself, i.e. a set of pairs $M = \left\{\left(a_1, b_1\right), \dots, \left(a_m, b_m\right)\right\}$ where $a_1 < \cdots < a_m$, $b_1 < \cdots b_m$, and $S\left[a_i\right] = S\left[b_i\right]$, the number of pairs where $a_i \not= b_i$ is at most $\eps |S|$ \cite{haeupler2017synchronization}.
This means that the number of disparate-level matching positions is at most $\eps(|AD|+|BE|)$.
Therefore,
\begin{eqnarray*}
ED(AD, BE) &\ge& |AD|+|BE| - \eps(|AD|+|BE|) - \alpha (|AD|+|BE|)\\
&\ge& (1-\eps-\alpha)(|AD|+|BE|).
\end{eqnarray*}
\end{proof}
Note that Theorem~\ref{thm:syncandtreecodes} suggests a construction for edit-distance tree codes by simply constructing and concatenating an ordinary tree code and a synchronization string. As synchronization strings are efficiently constructible and tree codes can be constructed in sub-exponential time~\cite{Schulman03:Postscript, braverman2012towards}, this construction runs in sub-exponential time which improves over the construction of edit-distance tree codes from Braverman et al.~\cite{braverman2015coding}.
The following theorems discuss further connections between synchronization strings and edit-distance tree codes.
Theorems~\ref{thm:TreeToSynch} and \ref{thm:extension} show that edit-distance tree codes and synchronization strings are essentially similar combinatorial objects.
In Theorem~\ref{thm:TreeToSynch}, we show that the edge labels along a monotonically down-going path in an edit-distance tree code form a synchronization string as though, in a manner, synchronization strings are one dimensional edit-distance tree codes. On the other hand, in Theorem~\ref{thm:extension}, we show that for any synchronization string $S$, there is an edit-distance tree code that has $S$ on one of its monotonically down-going paths.
\begin{lemma}\label{lem:TreeToSynch}
In an $\eps$-edit-distance tree code, for every three vertices $X$, $Y$, and $Z$ where $X$ is an ancestor of $Y$ and $Y$ is an
ancestor of $Z$, we have that $ED(XY,YZ) \ge \left(1 - \eps - \frac{1}{|XZ|}\right) |XZ|$.
\end{lemma}
\begin{proof}
To show this, we set $A=X, B = Y, E=Z$ and denote the child of $Y$ which is not in the path from $Y$ to $Z$ by $D$ (see Figure \ref{fig:pathtree}). Then $A, B, D$, and $E$ form a lambda in the tree (Definition~\ref{def:bad_lambda}). As $\eps$-edit-distance tree codes are $\eps$-bad lambda free, we know that $ED(AD, BE) \ge (1-\eps) \cdot(|AD|+|BE|)$.
\begin{figure}
\centering
\includegraphics[scale=1.2]{Path-Tree.pdf}
\caption{$(A, B, D, E)$ form a Lambda structure.}
\label{fig:pathtree}
\end{figure}
Note that $ED(AD, BE) = ED(XD, YZ) \le 1 + ED(XY, YZ)$ and $|AD| = |XY| + 1 > |XY|$, which means that
$ED(XY,YZ) + 1 \ge (1-\eps) |XZ|.$
Therefore, $ED(XY,YZ) \ge \left(1 - \eps-\frac{1}{|XZ|}\right)\cdot |XZ|$.
\end{proof}
Using this property, one can obtain synchronization strings using monotonically down-going paths in a given edit-distance tree code as follows:
\begin{theorem}\label{thm:TreeToSynch}
Concatenating the symbols on any monotonically down-going path in an $\eps$-edit-distance tree code with a string consisting of repetitions of $1,2,\cdots, l$ gives an $\left(\eps + \frac{1}{l}\right)$-synchronization string.
\end{theorem}
\begin{proof}
Consider three indices $i < j < k$ in such string. If $k - i < l$, the $\left(\eps + \frac{1}{l}\right)$-synchronization property holds as a result of the $1,\cdots, l,\cdots$ portion. Unless, the edit-distance path symbols satisfy $\left(\eps + \frac{1}{l}\right)$-synchronization property as a result of Lemma~\ref{lem:TreeToSynch}.
\end{proof}
Theorem~\ref{thm:TreeToSynch} results into the following corollary:
\begin{corollary}
Existence of synchronization strings over constant-sized alphabets can be implied by Theorem~\ref{thm:TreeToSynch} and the fact that edit-distance tree codes exist from~\cite{braverman2015coding}. However, the alphabet size would be exponentially large in terms of $\frac{1}{\eps}$.
\end{corollary}
\begin{theorem}\label{thm:extension}
Any $\eps$-synchronization string in $\Sigma_1^n$ can be extended to a $(2\eps - \eps^2)$-edit-distance tree code $C:\Sigma_{in}\rightarrow\left(\Sigma_{out}\cup\Sigma_1\right)^n$ using the $\eps$-edit-distance tree code $C':\left(\Sigma_{in}^n\rightarrow\Sigma_{out}^n\right)$ such that a monotonically down-going path on $C$ is labeled as $S$.
\end{theorem}
\begin{proof}
We simply replace the labels of edges on the rightmost path in the tree associated with $C'$ to obtain $C$. We show that $C$ is a valid $(2\eps - \eps^2)$-edit-distance tree code. To prove this, we need to verify that this tree does not contain an $(2\eps - \eps^2)$-bad lambda structure. In this proof, we set $\alpha = 1 - (2\eps - \eps^2) = (1-\eps)^2$ and $\alpha' = 1 - \eps$.
Let $A, B, D, E$ form a lambda structure in the corresponding tree of $C$.
We will prove the claim under two different cases:
\begin{itemize}
\item \textbf{$BE$ does not contain any edges from the rightmost path:}
Assume that $AD$ contains $l$ edges from the rightmost path. In order to turn $AD$ into $BE$, we must remove all $l$ symbols associated with the rightmost path since BE has no symbols from $\Sigma_1$. Hence, $ED(AD, BE)$ is equal to $l$ plus the edit distance of $BE$ and $AD$ after removal of those $l$ edges. Note that removing $l$ elements in $AD$ and then converting the remaining string ($\widetilde{AD}$) to $BE$ is also a way of converting $AD$ to $BE$ in the original tree. Thus,
$$ED_{original}(AD, BE) \le l + ED(\widetilde{AD}, BE) = ED_{modified}(AD+ BE)$$
On the other hand, $ED_{original}(AD, BE) \ge \alpha'\cdot(|AD|+ |BE|).$
Therefore, \[ED_{modified}(AD, BE) \ge \alpha'\cdot(|AD|+ |BE|).\]
\item \textbf{$BE$ contains some edges from the rightmost path:}
It must be that all of $AB$ and a non-empty prefix of $BE$, which we refer to as $BX$, both lie in the rightmost path in tree code (Fig. \ref{fig:distance}).
Since the symbols in the rightmost path are from the alphabet $\Sigma_1$,
$$ED(AD, BE) = ED(AB, BX) + ED(BD, XE).$$
We consider the following two cases, where $c$ is a constant we set later.
\begin{enumerate}
\item $c\cdot(|AB| + |BX|) > |BD| + |XE|$:
In this case,
\begin{eqnarray*}
ED(AD, BE) &=& ED(AB, BX) + ED(BD, XE)\\
&\ge& ED(AB, BX)\\
&\ge& (1-\eps)\cdot (|AB| + |BX|)\\
&\ge& \frac{1-\eps}{c+1}\cdot \left((|AB|+ |BX|)+(|BD|+|EX|)\right)\\
&\ge& \frac{1-\eps}{c+1}\cdot (|AD|+ |BE|).
\end{eqnarray*}
\item $c\cdot(|AB|+ |BX|) \le |BD|+|XE|$:
Since $(A, B, D, E)$ form a lambda structure, \[ED_{original}(AD, BE) \ge \alpha'\cdot(|AD|+ |BE|)\]
and therefore,
\begin{eqnarray*}
ED_{modified}(AD, BE) &\ge& ED_{modified}(BD, EX) + ED(AB, BX)\\
&\ge& \left[ED_{original}(AD, BE) - (|AB|+|BX|)\right] + ED(AB, BX)\\
&\ge& \alpha'\cdot(|AD|+ |BE|) - (|AB|+|BX|) + (1-\eps)\cdot(|AB|+|BX|)\\
&=& \alpha'\cdot(|AD|+ |BE|) - \eps(|AB|+|BX|)\\
&\ge& \alpha'\cdot(|AD|+ |BE|) - \frac{\eps}{c+1}(|AD|+|BE|)\\
&=&\left(\alpha' - \frac{\eps}{c+1}\right)(|AD|+|BE|).
\end{eqnarray*}
\end{enumerate}
\end{itemize}
Hence:
$$ED(AD, BE) \ge \min\left\{\alpha' - \frac{\eps}{c+1}, \frac{1-\eps}{c+1},\alpha' \right\}(|AD|+|BE|)$$
As $\alpha'=1-\eps$, setting
$c = \frac{\eps}{1-\eps}$ gives that
$$ED(AD, BE) \ge (1-\eps)^2\cdot(|AD|+|BE|)$$ which finishes the proof.
\begin{figure}
\centering
\includegraphics[scale=.90]{distance.pdf}
\caption{$AD$ and $BE$ both contain edges from the rightmost path. Straight lines represent the edges in the rightmost path and dashed ones represent other edges.}
\label{fig:distance}
\end{figure}
\end{proof}
|
\section{Introduction}
Internet of Things analytics (IoTA) involves a lot of applications \cite{2014cikmSnehasis} \cite{2013wam} \cite{2012www} \cite{2013WIsnehasis} in health, wellness, sustainability, transportation, smart city administration, and urban health while deploying any industrial sensor systems. Some of such applications in health care like blood pressure classification, Coronary Artery Disease (CAD) classification or machine health classification (into good condition or bad condition) involves classification as the machine learning (ML) task \cite{Bengio:Book}. It is required to identify some characters, known as feature, of the input signal in order to classify it using popular ML methods like Support Vector Machine (SVM) or Random Forest (RF). But the method of identifying a set of suitable features for the classifier is a time and cost absorbing task as it needs mostly domain expertise of the signal processing and IoT expert. Hiring people with such niche skill set is costly. Another problem is that some times the features listed by the domain expert are not relevant and if the data set includes less number of instances with a lot of features, the curse of dimensionality comes into play. This results into the need for feature reduction which can be accomplished into two different ways namely feature dimension reduction \cite{Ullman:book} or reducing the number of features by feature selection \cite{Maji:paperMRMSexp}, \cite{Peng:MRMR}. These two steps of listing features and reducing the number of features is known as Feature Engineering (FE) as a whole. Another approach to get rid of this issue of FE is to use some Deep Learning (DL) techniques like Multi Layer Perceptron (MLP) with fully connected layers or Convolutional Neural Network (CNN) which can automatically identify suitable feature set for the classification. But DL methods makes an assumption of having a huge number of annotated data-set to train the system which is mostly not available for IoT tasks. DL remains as a very popular technique for computer vision and natural language processing (NLP) and speech recognition. Another popular method is using Principal Component Analysis (PCA) to derive features that have high correlation to the class. However, in both cases interpretation of the features is an issue, and the domain expert cannot take advantage to carry out causal analysis for a given problem, once suitable features are identified.
For example, a cardiologist can justify that the PPG (photoplethysmogram) signals should contain a signal at a particular frequency range to classify a CAD signal. But neither DL methods nor the feature dimension reduction techniques can interpret the features as the feature space do not directly relate to physical world. So in this paper authors have proposed a method of FE for classification using ML techniques where a Feature Listing Database is maintained that becomes useful once good features are identified. The contributions of the proposed work is enlisted:\\
1. A survey has been carried out to understand the workflow of a typical IoT System development and identify the stages that demand maximum effort and niche skills, thereby costing more money and time.\\
2. The work proposes a method for automation of FE so that the suitable features for classification can be obtained in less time, less cost and the recommended features are physically interpretable to assist the domain expert in making casual analysis.\\
3. A WIDE architecture for feature listing is proposed that starts to use basic features in the basic layer and extracts more derived features on the higher layer. This layered architecture also helps to reduce the complexity as the features for different layers are computed iteratively. Features of higher layer are not computed once the desired performance is obtained at a lower layer. The performance is measured using different metrics like sensitivity, specificity, F-score.\\
The proposed method is tested against annotated data sets on man and machine predictive analytics, the use cases of which demand interpretation of features. The used data sets contains (i) NASA's bearing data set that includes good and bad bearing data, (ii) Emotion dataset to classify emotion of users, and (iii) MIMIC-II data set to classify high and low blood pressure from PPG signal. The accuracy obtained by applying some classifier on the recommended feature set are compared against the performance obtained by applying the same classifier on the SoA features reported in literature to solve these point problems. Some of these point problems were also tried in our research lab and hence we were able to compare the development time. We have also compared the proposed method against DL techniques like many layered Multi Layer Perceptron (MLP) method as most of the literatures have used MLP for sensor signal processing \cite{voyant:mlpTime}. The proposed feature selection technique is compared against a popular feature dimension reduction technique namely Principal Component Analysis (PCA). Finally, through an example we have also given a physical interpretation of the recommended features so that the recommended features can be validated against the physical phenomenon.
\begin{figure}[t]
\centering{
\includegraphics[width=5cm]{steps_of_signal_processing_new.eps}
\caption{High level workflow for IoT Analytics}}
\label{fig:workflow}
\end{figure}
\begin{figure}[!htb]
\centering{
\includegraphics[width=9cm]{analysis.eps}
\caption{Analysis of pain areas in IoT Analytics}}
\label{fig:analysis}
\vspace{-1.5em}
\end{figure}
Section 2 discusses about the survey conducted to analyse typical IoTA tasks. Section 3 describes the process and corresponding system architecture. Section 4 discusses the various experiments carried out. Finally section 5 concludes the paper.
\section{Background}
\begin{table*}[!t]
\small
\centering
\begin{tabular}{|l|c|c|c|}
\hline
Sensor &Signal Type &Goal &Steps \\
\hline
Camera, mobile, Kinect, &Stochastic &Object and Scene &Sampling$\rightarrow$Calibration$\rightarrow$Object \\
& & Recognition & Classification \\
\hline
Hyper spectral sensors, &NA &Object recognition &Image acquisition$\rightarrow$Pre-Processing $\rightarrow$\\
& & &Calibration$\rightarrow$ Feature listing$\rightarrow$\\
& & &Supervised or Unsupervised learning \\
\hline
Biosensors like ECG, PPG, &NA &Object recognition &Pre-Processing $\rightarrow$Feature Listing\\
EEG, EMG, Camera, & & &$\rightarrow$Classification \\
\hline
Camera &NA &Object Recognition &Pre-Processing $\rightarrow$Feature Listing\\
& & &$\rightarrow$Classification \\
\hline
Microphone, Camera, &Periodic, Stationary, &Physiological &Pre-processing$\rightarrow$Signal Quality Checker \\
IR Camera &Non-stationary &condition estimation &$\rightarrow$Feature Listing$\rightarrow$Feature \\
& & &Selection$\rightarrow$Computation \\
\hline
Thermal camera, &Multiple frequencies &Classification, Estimation &Filtering$\rightarrow$ Denoising $\rightarrow$ \\
LED/PhotoDiode & &calculation, imaging &features detection $\rightarrow$classification \\
& &beamforming, denoising & \\
\hline
EEG, GSR, SpO2, &Aperiodic, Non-stationary & Classification and &Pre-processing$\rightarrow$Noisy window removal \\
Camera, & &clustering, estimation &$\rightarrow$Feature Listing$\rightarrow$Feature Selection \\
Eye tracker & & &$\rightarrow$Computation \\
\hline
\end{tabular}
\caption{Steps involved in IoT sensor signal processing and analytics for different applications}
\label{tab:steps}
\vspace{-1.5em}
\end{table*}
We made a survey over the associates within our organization involved in the IoTA to identify the pain areas of the application developers. In this survey, we have surveyed seven ($7$) projects which include ninety-five ($95$) employees (female: $35$ and male: $60$). The associates under each project form a team. Each team consists of members among whom 30\% having more than 10 years of experience and Master's or higher academic qualification and each team is mostly led by a Ph.D. person. Sixty per cent ($60$\%) associates in each team have 5-10 years of experience and expertise in signal processing and the remaining $10$\% are developers mostly equipped with good coding skill in C/Java/Python but don't have in depth signal processing, IoT or domain knowledge.
Seven projects involved in this survey are (i) Motion capture, object recognition and rendering of articulated objects from Kinect based skeleton data, (ii) Object classification and recognition from Hyper-spectral sensor, multi-spectral sensors (Landsat, digital globe WorldView etc.) (iii) Bio sensors based object recognition, (iv) Outdoor camera based object recognition. (v) Detecting heart rate and blood pressure from PPG, (vi) Thermal imaging based object classification, and (vii) EEG (electroencephalogram) based cognitive load classification. For each project, the project goal, sensors used, and the steps followed to achieve the goal are shown in Table \ref{tab:steps}. These steps can be combined to construct a superset of steps followed in any sensor signal processing based IoTA as shown in Figure 1. The steps used in this figure are briefly described here:
\begin{itemize}
\item{Pre-Processing}: This module aims to filter out the noises in a signal. It also helps to remove the outlier in the signals. Different pre-processing techniques are available in the literature.
\item{Feature Listing}: Any signal is represented by some features. For example, time domain or frequency domain analysis normally give many features for an input signal.
\item{Feature Reduction}: The possible number of features for a sensor signal is very huge and so it is required to reduce the number of features. Feature reduction techniques should be selected with the following points kept in mind: (i) Interaction among the features, (ii) Interaction among the features with machine learning tool, (iii) Goal to achieve which can be obtained from an annotated data set, (iv) Interpretability of the reduced feature.
\item{Classification}: The technique takes the experience which is a representation of the feature set and the label as input and optimizes some parameters like accuracy, sensitivity to accomplish a goal like classification.
\end{itemize}
The analysis as shown in Figure 2 clearly reveals that most of the associates under survey express that feature listing and feature selection or feature reduction requires the maximum domain knowledge and technical knowledge. The other steps like pre-processing (including noise cleaning and outlier removal) takes the highest time to design. As the pre-processing algorithms are well established and even source code for them are available in the web, we propose to provide the developer with a list of source codes for all those different state of the art preprocessing techniques so that the developer can try any of the methods on their data set and select the best one among them. One such method is described in \cite{TCS:Sensys}. But feature selection is a difficult process. The input to our proposed automated system is the labeled data set of the problem to solve. Another aspect of the study was to list down the features, commonly reported in the related literature, that are used in IoTA applications and finally recommend the features to be used for the given use case.\\
\section{Proposed Method}
\begin{figure*}[ht]
\centering
\includegraphics[width=6.0in]{FE.eps}
\caption{Proposed Method of Feature Recommendation}
\label{fig:FeatureEngineering}
\end{figure*}
In this section, we are going to discuss how to automate the feature selection step which is the most time consuming and domain specific expertise dependent step. We plan to use state of the art pattern recognition techniques to achieve this. The key idea behind this is as follows.
{\it Exploration of all possible features}: Finding the right set of features for any pattern classification task is still an unsolved problem. Great amount of research is still concentrating on this problem \cite{TCS:Sensys}. In our approach, we, at first, have studied the existing literature~\cite{Lu:3} - \cite{Coutrix:20} to have a more or less exhaustive list of features which have been used by different researchers for different classification tasks on different sensor data sets.
We have organized these reported features in an hierarchical manner as shown in Figure ~\ref{fig:FeatureEngineering}. The features reported in literature can be mainly classified in three types: (i) time domain features (TD) (ii) Fourier transformation based features (FD) (iii) Discrete Wavelet transformation based features (DWT). A challenge of using features from Wavelet Transform is the appropriate selection of a suitable mother wavelet. This is because more than 100 different types of mother wavelets are reported in literature, and a ready made automatic wavelet selection method or tool do not exist. Hence, considering the the different feature types, it is possible to get a large number (say, $N$) of features (including coefficients of the transform domain) from the sensor signals. This results in $2^N-1$ possible combinations of features.
{\it Feature Selection}:
The value of $N$ could be as high as $6,000,000$ as explained later in this section. Therefore, in order to find the optimal feature set, exploitation of all such $2^N-1$ combinations is practically infeasible as this would require testing of machine performance for each combination. This can be achieved mostly by the domain experts from their experience and understanding about the physical phenomena. Typically feature selection method takes 4-6 months for such problems. In pattern recognition literature, there are several feature reduction techniques available which can give a reduced set of features giving optimum performance \cite{Chattopadhyay:ICIP}, \cite{2013tanuICDAR}, \cite{2011tanuICCE}. Principle Component Analysis (PCA) is one such commonly used technique for feature reduction. Actually, what PCA does is feature extraction, i.e. the resultant features are not interpretable. However, we need interpretable features for our tasks. For instance, in health analytics, if we know that certain features are very strong in predicting CAD (coronary artery disease) patients, the doctors need to know what these features are. Therefore, in order to reduce the number of feature to make the feature recommendation task tractable we follow feature selection instead of feature extraction as feature selection methods do not alter the original features.
In our method, we followed an iterative feature selection where $k$-features are selected at each iteration and system performance (e.g. classification accuracy) is checked for this feature set. If the selected feature set results in {\it expected} performance, that feature set is recommended and process halts. Else, another set of $k$-features is selected in the succeeding iteration and similar steps are carried out. For checking the classification accuracy, we choose SVM-(support vector machines) based classification with different kernels. Different values of $k$ are fitted in progressive increments to get a good result. For a given value of $k$, features are selected using two techniques namely, MRMS \cite{Maji:paperMRMS} and mRMR \cite{Peng:MRMR}. The reason for choosing these two techniques is their impressive efficiency in feature selection as demonstrated in~\cite{Maji:paperMRMSexp}. Details on mRMR and MRMS used for this framework can be found in \cite{2016nipsSnehasis}.
\subsection{The proposed Architecture for Feature Recommendation}
The proposed method is applicable in extracting and recommending features from 1D sensor signals. The sensor signal is subjected to time domain, frequency domain and time frequency domain analysis to extract features which for example counts to almost 12 million features for input data of size 20,000. Let us assume a 1D signal of length n sampled at fs frequency. So after subtracting the mean from the signal we get the same n number of TD features. Now the entire signal is splitted into multiple overlapping windows to extract the FD features. So if we consider a 1sec window with 50\% overlap then we get (n*2)/fs number of windows. Considering the STFT window size to be 256, number of FD features can be obtained is (256*n*2)/fs. Similarly by applying DWT we obtain another n features. So extracting first level features we get n+n+(512*n)/fs $\approx$ 3n number of features. The level two features are derived from each window of the level one features. Thus in level 2 around 20*(3n/fs)*2$\approx$120n features are extracted. Therefore around 3n+120n=123n features are extracted at the end of layer2. Similarly in 3rd layer features are derived from each window of 2nd level features which is around 2*(120n/fs)*2 $\approx$ 480n. So after 3rd layer the number of features extracted is of the order 123n+480n $\approx$ 600n. Now in dataset 1, value of n is nearly 20,000. So number of features extracted after 3rd layer is nearly 600X20,000$\approx$ 12,000,000.
Level two features are applied on top of each of the level one features. Thus after level 2 this number of features extracted is becoming nearly of the order of 4n*4=16*n. Similarly after 3rd layer this number becomes of the order or 30*n. Now in data set 1 this n is nearly 20,000. So after 3rd layer of feature extraction total number of feature extracted from the signal is nearly $30\times20,000 \approx$ 6,00,000.
So, it is not possible to apply any exhaustive search on top of this 12,000,000 features so that one can optimize the performance metrics say sensitivity or specificity. The feature selection architecture used in this paper is shown in Figure~\ref{fig:FeatureEngineering}. The proposed method is as described below: \\
Input: Time domain signal \\
Output: Recommended feature list
\begin{itemize}
\item{} Compute the mean subtracted time domain signal by subtracting the mean from the original signal. This entire signal is considered as the time domain (TD) feature.
\item{} Compute the short term Fourier transform(STFT) for each window where window size is provided as an user input.
\item{} The given signal is subjected to four level Discrete Wavelet Transform (DWT) using the most optimal mother wavelet. The optimal mother wavelet is selected using the property of maximum energy to entropy ratio \cite{ngui2013wavelet}.
\item{} Take the union of the all three TD, FD, and DWT features. Let this set be defined as $FL_{1}$
\item{} Apply minimum redundancy maximum relevance (mRMR) and Maximum Relevance Maximum Significance (MRMS) for $FL_{1}$. Each of these methods recommends a different feature set say x and y.
\item{} Iteratively select the number of features to be recommended by each of these methods so that the recommended features applied on a classifier can exceed the performance metric score defined by the user, say $\tau$.
\item{} Take union of the features recommended by both mRMR and MRMS methods. Let the recommended feature set be z where $z = x \cup y$.
\item{} Features from level 2 ($FL_{2}$) and level 3 ($FL_{3}$) are extracted using the algorithms shown in Figure \ref{fig:FeatureEngineering}.
\item{} In level 2 we have extracted spectral features, statistical features, and peak-trough features. Spectral features used in the proposed method are Spectral Centroid, Spectral Crest Factor, Spectral Decrease, Spectral Flatness, Spectral Flux, Spectral Kurtosis, Spectral Rolloff, Spectral Skewness, Spectral Slope and Spectral Spread which are computed for each window of mean subtracted signal. Statistical Features used here are mean, variance, standard deviation, root mean square, skewness, kurtosis. Average peak amplitude, average trough amplitude, average peak to peak distance and average trough to trough distance are computed as the peak-trough features. $FL_{3}$ includes different ratios and derivatives of the $FL_{2}$ features.
\item{} These $FL_{2}$ and $FL_{3}$ features are also reduced using mRMR and MRMS similarly.
\item{} Once the feature space is reduced, exhaustively generate all possible combinations of features and apply classifier with different parameters.
\item{} Apply Support Vector Machines (SVM) classifiers with different kernels namely (i) linear, (ii) radial basis function (RBF), (iii) sigmoid, and (iv) polynomial and each of the kernels is tested with different parameter values.
\item{} Thus construct the model for classification by selecting the feature list and the SVM kernel for which the recognition accuracy maximizes. SVM is selected as a classifier as the problems at hand are binary classification where SVM is known to excel and converge quickly.
\end{itemize}
\section{Experiments}
\begin{table*}
\caption{Comparison of proposed method against state of the art for different data sets}
\label{Table:comparison}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
Data set &SOA no of &Recommended no &SOA &Recommended & SOA &Recommended \\
&features &of features &accuracy &accuracy &effort &effort \\
\hline
NASA data set &15 &10 &99.38 &100 &6 months &2 days \\
\hline
Emotion data set &16 &11 &82.3 &90.91 &4 months &2 days \\
\hline
BP data set &23 &15 &79.5 &87.8 &6 months &2 days \\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Comparison of proposed method against 5-layered MLP}
\label{Table:comparison}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
Data set &Activation &No. of Epoch &MLP Accuracy &SoA Accuracy &Proposed\\
&Function & & & & Method Accuracy \\
\hline
NASA data set &softmax &5 & 71.00 &99.38 &100 \\
\hline
Emotion data set &relu &15 & 50.0 &82.3 &90.91 \\
\hline
BP data set &softmax &10 & 50.0 &79.5 &87.8 \\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Comparison of proposed method against PCA}
\label{Table:comparison}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
Data set &Dimensionality &No. of Principal &PCA Accuracy &SoA Accuracy &Proposed\\
&Reduction Algorithm &Components & & & Method Accuracy \\
\hline
NASA data set &svd &5 & 94.00 &99.38 &100 \\
\hline
Emotion data set &svd &10 & 50.0 &82.3 &90.91 \\
\hline
BP data set &eig &5 & 62.50 &79.5 &87.8 \\
\hline
\end{tabular}
\end{table*}
\subsection{Dataset}
The experiment is performed on 3 data sets out of which two data set are publicly available i.e. the Bearing dataset from the Prognostics Data Repository and second one is the MIMIC data-set for classifying blood pressure. The third data-set is internal to our organization and this aims to classify the emotion of a person into three classes namely happy, sad, and neutral. As there is the rest two data sets are two class classification problem, we have used our own data set to prove that our proposed system works on multi-class use-cases also. The first data set used as a machine automation case study. The rest two data sets are used in health care and emotion detection. 2nd and 3rd data sets are based on photoplethysmogram (PPG) sensor signal. The bearing data set (named IMS-Rexnord) consists of three datasets describing a test to failure experiment. Each dataset comprises of several files each of which has a record of 1 sec vibration signal snapshot which are recorded at specific intervals. The sampling rate is 20KHz, recording interval is 10 min and there are 20480 data points in each of the file. In our experiment we have used the second dataset. It has 984 files. Each individual file holds the record of 4 channels representing 4 bearings where the first bearing eventually turns faulty due to outer race failure.
The experiment is performed in two ways. In the first case only the bearing 1 data is considered. State of the art \cite{dong2014bearing} shows that the bearing 1 starts degrading after the 700th point where each file of 1 minute readings is denoted as a point. So here we have formulated a two class classification problem where the first 700 files are considered to be healthy and the rest 282 files are considered to be faulty (the last 2 files are discarded due to presence of many noisy signal values).
In the second case we have considered all the bearings to formulate the two class classification problem. The bearing dataset has 984 files. Each file has record of 4 bearings. Therefore the total number of good bearing samples is 3652 (1st 700 samples of 1st bearing and 984 samples of each of the three other bearings) and the number of faulty bearing samples is 282 (last 282 files of 1st bearing). To avoid any biasness we have segregated the data into 5 folds. Each of the dataset consisting of 282 faulty bearing samples and 730 good bearing samples.
The second dataset which is used to classify the blood pressure into high and low, records the PPG signal of 118 subjects from Bangalore and Gujarat. Among the 118 subjects, 15 subjects have high systolic blood pressure and 103 have low systolic blood pressure. To avoid any biasness due to imbalance in the dataset, the dataset is segregated into three datasets. Each of the dataset consists of the PPG signal record of 49 subjects, 15 subjects having high systolic blood pressure and 34 subjects having low systolic blood pressure. The sampling rate is 60 Hz.
The third dataset (used to classify the emotion into happy and sad) records the fingertip pulse oximeter data of 33 healthy subjects (13F and 20M) with average age 27. No two emotion elicitation video was shown to a subject in one single day. The Pulse Oximeter is used to detect and record the PPG signal. We used standard video stimuli which itself served as ground-truth and the rigorous experimentation procedure ensured that the time synchronization error between the stimuli and recorded physiological data is always less than 1sec.
Table 2 shows the performance of proposed method against state of the art results, proving the efficacy of the method. It is to be noted that time and cost is a huge factor when comparing the performance. In many cases, a trade-off fits the solution well when performance gain is not much when automotion is compared to the manual effort.
\subsection{Comparison against MLP}
Learning representations from raw data is an emerging field. Theano has been used as the software platform to carry out experiments.
Different number of layers (based on standard thumb rules of input size) for Multi-Layer Perceptron (MLP) with Dropout feature has been tried out so that automatic feature learning can take place. Different activation functions like tan hyperbolic, softmax, sigmoid, rectified linear unit (relu) etc. has been investigated at different layer levels to get a suitable architecture for classification task for the given problems.
Table 3 lists the configurations obtained for a 5 layered MLP for which the best performance was obtained. It is seen that MLP techniques fail in comparison to proposed method as well as state of the art. Also, MLP has some drawbacks when put into the IoT spectrum - a) need for a lot of data for training and b) non-interpretable feature extraction.
\subsection{Comparison against PCA}
Principal component analysis (PCA) is a statistical procedure that uses an orthogonal transformation to derive principal components representative of the features under consideration. This has two outcomes: a) dimension of feature space can be reduced by selecting most prominent principal components b) derived features is supposed to represent the feature space better.
Gaussian kernel is used for SVM based classification on the principal component features extracted. Various dimension reduction techniques has been used like Alternating Least Squares (als), Eigen Value Decomposition (eig) and the traditional Singular Value Decomposition (svd). Table 4 lists the configurations leading to the best results. It is apparent from the table that PCA based methods do not perform well in comparison to proposed method. Another drawback of PCA is the derived features are not interpretable.
\section{Conclusion}
This paper has carried out a survey to understand the workflow time and cost needs of a typical IoTA task. In the survey, Feature Engineering came out as the most taxing task among all endeavors that comprise IotA. To tackle the challenge, a system was built to automate this sub-task of IoTA. The system has been tested on three datasets and has been found to give good results when compared to state of the art. The proposed method is compared with PCA and MLP, which are two divergent paths of feature engineering. Feature interpretation is another notable aspect of the system. Future work will look into automation of parameter tuning and selection of machine learning models. Automatic window selection \cite{2013cikmSnehasis} for a given dataset is also planned. Integration of knowledge bases and use of reasoning \cite{2013patentDebnath} \cite{2013OTMdebnath} for ease of interpreting features and causality analysis is also planned.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
Urban traffic gridlock is a familiar scene. At the same time, the mean occupancy rate of personal vehicle trips in the United States is only 1.6 persons per vehicle mile. Ridesharing can solve many environmental, congestion, and energy problems. With these advantages, ridesharing is becoming increasingly popular. Many ridesharing services such as Uber and Lyft allow for peer-to-peer ride matching through mobile-phone application \cite{ huang2014large}. However, with current ridesharing systems, a user has little information about their rideshare partner. A user cannot choose or approve a partner before the ride. There are risks and social discomfort that arise due to lack of trust amongst co-passengers. Trust in co-passengers is the most important factor influencing willingness to give or take rides from others \cite{chaube2010leveraging}. Understanding the factors that influence perceptions of trust among strangers has important implications
for research seeking to reduce urban gridlock. Ridesharing has great potential to increase carpooling and reduce urban congestion, but only if the users trust the potential rideshare partner and do not feel they are entering a risky situation.
With the widespread use of social media, it is possible to obtain basic information of users who want to use a ride-sharing service. Social media allows people to create, access, and interact with a wide range of information \cite{jang2015generation}. In this work, we use Instagram to extract trust factors. Instagram allows users to share photos and comments. It has quickly become one of the most popular social media platforms in the United States since its launch in 2010. By the end of January 2017, Instagram had 600 million active users per month with 400 million of these users accessing the platform every single day \cite{Instagram}. Instagram has two types of users, private-profile users and public-profile users. With public-profile users, the profiles are available for viewing even if one does not follow a user. With private-profile users, one must get the permission of the user to follow in order to see their posts. Unlike other social media, each post by an Instagram user is a picture or a video \cite{ hu2014we}.
By combining ridesharing systems with social media, we attempt to provide more information for assigning trust on a rider or driver in a ridesharing system. We provide pictures and useful social information from a user$'$s Instagram. Our analysis identifies three key factors that influence trust formation. By identifying key factors that people use when making assessments of trustworthiness for social media users, understanding is increased regarding what aspects of a profile compel users to increase perceptions of trust and reduce perceptions of risk. This paper makes the following contributions:
\begin{itemize}
\item We propose a system that allows a user to gain more information of a potential co-rider through linking social media and ridesharing system.
\item We identify three crucial factors that influence trust building:\textit{ social proof}, \textit{social approval}, and \textit{self-disclosure} from social media posts. These three factors account for 40\%, 33\%, and 13\% of the variance, respectively. Subsequent factors extracted accounted for substantially less variance. This result expands our understanding of which factors persuade people to trust potential rideshare partners as co-riders. In addition, we analyze the manipulability of the factors and shed light on the robustness of the information that could be used to build trust.
\item As researchers and practitioners gain understanding of factors that influence perceptions of social media trust, increased collaboration can occur among users and rideshare companies to provide the information that users need to view for adequate levels of trust to be reached.
\item Using three key factors, we successfully build a trust prediction model with F measure of 87.4\%, an improvement from 78.6\% on using all factors collected.
\end{itemize}
By examining the factors of user profiles that lead to perceptions of increased trust and decreased risk, these results provide insights for both consumers and organizations on the information needed to increase participation in ridesharing applications through increased trust and reduced perceptions of risk.
\section{Related Work}
Vineeta et al. \cite{chaube2010leveraging} summarized the results of a rideshare need assessment survey conducted within the Virginia Tech University. The purpose of this work is to understand commuter travel patterns, their needs, and to identify their preferences for private vehicles and public transit across a variety of travel contexts. The results show that although people seem to be interested in ridesharing, it remains an unpopular mode of transportation. The most important reasons for its unpopularity are the inconvenience, lack of trust, and incentives. Among these three factors, trust is a major factor that people consider while offering or accepting a ride.
Other systems also find trust as one of the major concerns of rideshare users. Joshua et al. \cite{morse2007carloop} found that people are often hesitant to carpool with unknown drivers. They developed a carpooling system which connects commuters in the same organization with each other, leveraging the common background of their shared employer. Rick et al. \cite{wash2005design} designed the RideNow project to help individuals within a group or organization to coordinate ad hoc shared rides. RideNow service is intended to coordinate ride sharing within groups or small organizations of a few hundred to a few thousand people. RideNow project can reduce the uncertainty of dealing with strangers and has been found to increase conversations and dialogue among riders. These works benefit users within the same community and are not designed to address trust issues in general ridesharing.
Zimride \cite{zimride} is one of the applications that combines ridesharing with social network sites such as Facebook and MySpace. Zimride can deal with trust problem in rideshare systems. For instance, a Myspace user can request a ride or offer a ride to other members of their corporation or university. In this approach, users who will share a ride together have a common ground.
Both Uber and Lyft use a rating system. At the end of each ride, a rider can rate a driver and also give feedback based on the driver$'$s driving skill, behavior, and condition of the vehicle. If the cumulative rating is below the minimum for the driver's area, the driver is suspended \cite{Uber}. On the other hand, a driver also has the option to rate a rider based on cancellations, no-shows, and late arrivals. A user$'$s rating will affect a driver$'$s willingness to pick up the user. A rating-based system suffers from a cold start problem when insufficient ratings are available. Ratings can also be biased, as not all users give ratings, especially when the number of ratings of a driver are small. Often, people who are extremely unhappy tend to give ratings with greater frequency.
\section{Data Collection}
We collected the data on January 20, 2017 from Instagram. We first chose one seed public user and then crawled all the followers of the seed user. We chose the public user UNT as our seed. There were a total 27k followers of UNT by the end of the day. Among those users, we filtered out private users, resulting in 8,436 public users. We then chose 100 users randomly from these public users.
\subsection{Attributes of Dataset}
The dataset of Instagram includes various pieces of a user$'$s information. We summarized the information into three basic types: user information, user account information, and post information. There are several attributes under these three types. The detailed information is listed in Table 1 to Table 3.
\begin{table}[!ht]
\centering
\caption{User information}
\begin{tabular}{|c|c|} \hline
\textbf{\textbf{User information}}&\textbf{}\\ \hline \hline
\texttt{\textbf{User name}} &Instagram account name, user can change it\\
\texttt{}& at anytime.\\
\hline
\texttt{}& A user can type their true name or other\\
\texttt{\textbf{Full name}} & preferred name. If a user did not enter the \\
\texttt{} & full name, it will not display the full name \\
\texttt{} & part of this user \\
\hline
\texttt{\textbf{Biography}}& A user can type a few sentences to describe\\
\texttt{} & himself. \\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering
\caption{User account information}
\begin{tabular}{|c|c|} \hline
\textbf{\textbf{User account information}}&\textbf{}\\ \hline \hline
\texttt{\textbf{Followers}} & Total number of users who are\\
\texttt{}& following this user.\\
\hline
\texttt{\textbf{Following}}& Total number of users that this\\
\texttt{} & user is following. \\
\hline
\texttt{\textbf{Posts}}& Total number of posts that the\\
\texttt{} & user has posted through Instagram.\\
\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{Post information}
\begin{tabular}{|c|c|} \hline
\textbf{\textbf{Post information}}&\textbf{}\\ \hline \hline
\texttt{\textbf{Likes}} & Number of users who like this post.\\
\hline
\texttt{\textbf{Comments}}& Number of comments under this post.\\
\hline
\texttt{\textbf{Create time}}& \hspace{0.01cm} Creattion time of this post, format is in Unix \\
\texttt{} & time. \\
\hline
\texttt{\textbf{Caption}}& User description of this post. \\
\hline
\texttt{\textbf{Hashtag}}& Hashtag of this post. \\
\hline
\end{tabular}
\end{table}
Figure 1 shows one Instagram post consisting of pictures.
\begin{figure}[ht!]
\centering
\subfigure[]{
\includegraphics[ width=0.8\linewidth]{new_instagram_1.pdf}
\label{fig:instagram_1}
}
\subfigure[]{
\includegraphics[width=0.8\linewidth]{new_instagram_2.pdf}
\label{fig:instagram_2}
}
\caption{User profile and a post in Instagram}
\end{figure}
\subsection{Factor Selection}
We first select \textit{followers}, \textit{following}, \textit{posts}, \textit{likes}, and \textit{comments} as factors. Posted pictures can also provide trust information. For each profile, we selected the 10 most recently posted pictures. For \textit{likes} and \textit{comments}, we sum up the number of \textit{likes} and number of \textit{comments} under the 10 posts of each user. We are interested in the influence of having people in the pictures on trust. We selected three picture attributes as factors. For the first factor, we counted the total number of people among these 10 pictures. For second factor, we counted the number of pictures that contained a person. For the third factor, we counted the number of pictures that contained the user themself. Together, we have 8 different factors from our dataset.
\subsection{Trust Labeling}
There is limited availability of labeled data dealing with the relationship between social networks and rideshare systems - the two basic embodiments of our research. Survey is one of the most effective methods for generating ground truth dataset. We hosted a survey online by using Amazon Mechanical Turk. The content of each survey is based on the information from the selected Instagram public users that we collected. In the survey, it shows the profile picture of the user, total number of followers, total number of followings, and total number of posts. A user may have many posts. In our survey, we chose the 10 latest posts for each user. Each post contains detailed information, such as the caption, number of likes, number of comments, and the location of the post.
We asked respondents of the survey six questions:
\begin{enumerate}[1)]
\item If this person were a driver for a rideshare service like Uber, I would trust this person to give me a ride.
Yes : I would trust. No: I would not trust.
\item If this person were the driver for a rideshare service like Uber, I would trust putting a close friend in a car with them.
Yes: I would trust. No: I would not trust.
\item Riding with this person would be more risky or less risky than average.
Yes: More risky. No: Less risky.
\item During the ride, I$'$d probably talk to this person more than average or less than average.
Yes: More than average. No: Less than average.
\item If this person was assigned to me as my rideshare partner, I$'$d feel safe riding in the car with them.
Yes: I$'$d feel safe. No: I$'$d not feel safe.
\item If you find out this person will be your rideshare partner, how likely are you to use a ride service.
Yes: More likely. No: Less likely.
\end{enumerate}
Mechanical Turk allows filtering of turkers according to certain criteria. In our survey, we randomly chose 5 different turkers to view each profile. Our criteria required each turker to have an approval rate above 98 percent and for him to be located in the United States. We analyzed the answers from 5 turkers for each profile entry. Table 4 shows the number of survey entries for each combination of numbers of \textquotedblleft Yes\textquotedblright and \textquotedblleft No\textquotedblright among the 5 turkers for each question. The ground truth trust labeling for each survey entry is chosen by majority vote.
\begin{table*}
\centering
\caption{The distribution of different percentages by answer \textquotedblleft Yes\textquotedblright or \textquotedblleft No\textquotedblright to the six questions}
\begin{tabular}{|c|c|c|c|c|c|c|} \hline
\textbf{}&\textbf{\textbf{Question 1}} &\textbf{\textbf{Question 2}} &\textbf{\textbf{Question 3}} &\textbf{\textbf{Question 4}} &\textbf{\textbf{Question 5}} &\textbf{\textbf{Question 6}}\\ \hline \hline
\texttt{\textbf{\textbf{Yes=0 N0=5}}} & 2 & 2 & 34 & 13 & 3 & 7 \\
\hline
\texttt{\textbf{\textbf{Yes=1 N0=4}}} & 8 & 11 & 19 & 17 & 7 & 13 \\
\hline
\texttt{\textbf{\textbf{Yes=2 N0=3}}} & 17 & 19 & 16 & 18 & 17 & 16 \\
\hline
\texttt{\textbf{\textbf{Yes=3 N0=2}}} & 14 & 13 & 15 & 17 & 13 & 14 \\
\hline
\texttt{\textbf{\textbf{Yes=4 N0=1}}} & 15 & 13 & 12 & 14 & 18 & 18 \\
\hline
\texttt{\textbf{\textbf{Yes=5 N0=0}}} & 44 & 42 & 4 & 21 & 42 & 32 \\
\hline
\end{tabular}
\end{table*}
\thispagestyle{plain}
\section{Factor Analysis}
Factor analysis is a statistical technique which extracts important factors from a variable group by reducing redundancy from a set of correlated variables. It explores the basic structure of observed data and presents the data structure with a few latent variables. These latent variables can reflect the most influential information of the original variables \cite{williams2010exploratory}. There are two main factor analysis methods. Exploratory Factor Analysis (EFA) and Confirmatory Factor Analysis (CFA) \cite{yong2013beginner}. The difference of these two methods is that EFA is used to uncover complex patterns by exploring the dataset and testing predictions; CFA is used to confirm hypotheses. In our case, we do not know the patterns, so we choose EFA as our factor analysis technique. In our analysis, there are eight trust factors. However, some of the factors may have some correlations. Therefore, we use EFA to reduce the factors to key factors. We performed five steps for the Exploratory Factor Analysis \cite{IBM}. First, we check whether the EFA approach suits our problem and dataset. Second, the extraction procedure is selected for the dataset. Third, method is decided to determine the total number of factors to retain. Fourth, we select the rotation method to yield a final solution. Fifth, a proper interpretation of each final factor is labeled. Failure to make a decision about one or more of the mentioned steps will cause erroneous results and limit the utility of EFA \cite{taherdoost2014exploratory}. Figure 2 shows the procedure of exploratory factor analysis. We chose the trust label of question 1 in the survey for factor analysis as it is the most general labeling of trust.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.8\linewidth]{steps_1.pdf}
\caption{5-step Exploratory Factor Analysis}
\label{fig:steps}
\end{figure*}
\subsection{Evaluation of data suitability}
There are several tests that can be conducted. The major tests include Kaiser-Meyer-Olkin (KMO) Measure of Sampling Adequacy and Bartlett$'$s Test of Sphericity. KMO measure varies between 0 and 1, and values closer to 1 are better. Table 5 shows that the KMO test value of our data is 0.714. Based on Tabachnick \cite{henson2006use}, 0.50 is considered suitable for factor analysis and Netemeyer \cite{netemeyer2003scaling} pointed that a KMO correlation above 0.60-0.70 is considered adequate for analyzing EFA output. So the KMO result indicates that our data is suitable for EFA. We use Bartlett$'$s Test of Sphericity to test the hypothesis that the correlation matrix is an identity matrix. The output should be significant (less than 0.05) for factor analysis to be suitable. The significant level of \textit{p} in our data is less than 0.001, which is much smaller than the indicator. It means the variables are correlated and have pattern relationships.
\begin{table}
\centering
\caption{KMO and Bartlett$'$s Test}
\begin{tabular}{|c|c|} \hline
\texttt{Kaiser-Meyer-Olkin Measure of Sampling Adequacy.} & .714 \\
\texttt{Bartlett$'$s Test of Sphericity Approx. Chi-Square}& 664.229\\
\texttt{ df }& 28\\
\texttt{ Sig. }& .0001\\
\hline
\end{tabular}
\end{table}
\subsection{Factor Extraction Methods}
A variety of methods have been developed to extract factors: Principle components method (PCA), Maximum likelihood method, principal axis factoring (PAF), unweighted least-squares method, generalized least squares method, alpha method and image factoring \cite{woolley2010evidence}. We chose PCA, the most commonly used method, as our factor extraction method. PCA is used to form linear combinations of the observed variables. The first component has maximum variance. Successive components explain progressively smaller portions of the variance and are uncorrelated with each other. PCA is used to obtain the initial factor solution \cite{thompson1996factor}. Table 6 provides the communality result. Communality is the proportion of each variable$'$s variance explained by the extracted factor structure. The results show that the community of each common extracted factor is nearly above 80\%, indicating the extracted common factor$'$s explanatory ability on each variable is strong. It indicates most information in the variables can be extracted by factors.
\begin{table}
\centering
\caption{Common factor variance}
\begin{tabular}{|c|c|c|} \hline
\textbf{}&\textbf{\textbf{Initial}}&\textbf{\textbf{Extraction}}\\ \hline \hline
\texttt{post} & 1.000 & .873\\
\texttt{follower}& 1.000 & .884\\
\texttt{following}& 1.000 & .895\\
\texttt{likes}& 1.000 & .861\\
\texttt{comments}& 1.000 & .786\\
\texttt{total\_person}& 1.000 & .763\\
\texttt{pic\_person}& 1.000 & .945\\
\texttt{self}& 1.000 & .918\\
\hline
\end{tabular}
\end{table}
\subsection{Factor Retention Methods}
Factor retention is an important step of exploratory factor analysis. Both under-extraction and over-extraction will have negative effects on the results. There are several criteria to decide how many factors to retain \cite{kaiser1960application}. Factor retention methods include: Kaiser$'$s criteria \cite{cattell1966scree}, Scree test \cite{hair2010multivariate}, and the cumulative percent of variance extracted. The results of these criteria do not always agree. Many researchers use multiple criteria to decide the number of factors to retain. In our factor retention steps, we checked all of these three criteria to make sure the final number of factors is convincible. First, we use Kaiser$'$s criteria, which is a rule of thumb. This criterion suggests to retain factors above the eigenvalue of 1. From Table 7, we can see that the factors are in the descending order based on the most explained variance. The initial eigenvalues show that in our case, there are three key factors with an eigenvalue greater than 1. So based on Kaiser$'$s criteria, we keep three factors. Second, we check the result from scree test. The scree test consists of factors and eigenvalues. For the scree test, we need to find the point at which the curve changes drastically. Figure 3 shows the result of the scree test. From point one to point three the curve is a steep slope suggesting to keep the first three points. It indicates the maximum number of components to retain is three. Finally, we used the cumulative percent of variance method to determine the number of factors to retain. For this method no fixed threshold exists, although certain percentages have been suggested. We select the threshold commonly used in humanities, which has the explained variance as low as 50-60\% \cite{kim1978factor}. In our case, we used rotated results to decide the number of key factors. From Table 7, the rotation sums of squared loadings indicates that the cumulative variance of the first three factors is 86.563\%, which is greater than 60\%. It means the first three factors are sufficient to represent the eight variables.
\begin{table*}[ht!]
\caption{Total Variance}
\centering
\begin{tabular}{|c|r|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{Component}
& \multicolumn{3}{|c|}{\textbf{Initial Eigenvalues}} & \multicolumn{3}{|c|}{\textbf{Extraction Sums of Squared Loadings}} & \multicolumn{3}{|c|}{\textbf{Rotation Sums of Squared Loadings}} \\
\cline{2-10}
& Total & \% of Variance & Cumulative \% & Total & \% of Variance & Cumulative \% & Total & \% of Variance & Cumulative \%\\
\hline
\hline
\texttt{1} & 3.202 & 40.030 & 40.030 & 3.202 & 40.030 & 40.030 & 2.642 & 33.030 & 33.030\\
\texttt{2} & 2.672 & 33.395 & 73.425 & 2.672 & 33.395 & 73.425 & 2.554 & 31.924 & 64.954\\
\texttt{3} & 1.051 & 13.138 & 86.563 & 1.051 & 13.138 & 86.563 & 1.729 & 21.608 & 86.563\\
\texttt{4} & .348 & 4.347 & 90.909 & & & & & & \\
\texttt{5} & .324 & 4.048 & 94.957 & & & & & & \\
\texttt{6} & .222 & 2.775 & 97.733 & & & & & & \\
\texttt{7} & .145 & 1.818 & 99.550 & & & & & & \\
\texttt{8} & .036 & .450 & 100.000 & & & & & & \\
\hline
\end{tabular}%
\label{tab:variance}
\end{table*}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\linewidth]{scree_plot_1.pdf}
\caption{Scree Plot}
\label{fig:overview}
\end{figure}
\subsection{Rotation Methods}
Rotation methods deal with a variable that is related to more than one factor. The key idea of rotation is that it maximizes high-item loadings and minimizes low-item loadings. By doing this, it will be much easier to facilitate interpretation of the factors. Table 8 shows the factor matrix, which is the factor loadings before rotation, and Table 9 shows the rotated factor loadings. In our study we use orthogonal varimax approach for our rotation method. Comparing Table 8 and Table 9 we can find that, after the rotation step, the approach can produce factor structures that are uncorrelated. From Table 9 we can see that rotation and suppressing small coefficients helps with the interpretation. In the rotated matrix, we can easily assign each variable to the corresponding factor without overlapping. In our case, we choose a loading cutoff of 0.36.
\thispagestyle{plain}
\subsection{Interpretation and Labeling}
In this step, we need to examine which variables are attributing to a factor and then give the factor a name. For instance, if you have a factor the name \textquoteleft emotional\textquoteright, and variables that load high on this factor are \textquoteleft laugh\textquoteright, \textquoteleft angry\textquoteright, and \textquoteleft crying\textquoteright, you also need to make sure that an unrelated variable such as \textquoteleft writing\textquoteright does not load onto this factor. To give a meaningful interpretation, usually two or three variables must load on a factor. The labeling of a factor is a theoretical and subjective process. In our case, after varimax rotation, we had three significant factors and assigned each variable to its corresponding factor. From Table 9, we can see that, factor 1 consists of total\_person, pic\_person, and self. Factor 2 contains follower, likes, and comments. The third factor contains following and post.
\begin{description}
\item[Factor 1:] Social Proof
This factor, Social Proof, represents the positive influence on trust of users providing evidence, or proof, that they are social. Pictures of the user and other people offer \textquoteleft proof\textquoteright that the user is sociable and likable. This factor contains the total number of people visible in the 10 pictures, the total number of pictures (out of 10) that contained a person, and the total number of pictures that contained the user personally.
\item[Factor 2:] Social Approval
Social Approval by other people on activity of the user$'$s account, expressed through \textquoteleft like\textquoteright of pictures and comments on the user$'$s profile, increased perceptions of trust and reduced perceptions of risk among potential rideshare partners. This factor illustrates the positive influence of social approval by other people on respondents$'$ perceptions of trust.
\item[Factor 3:] User Self-Disclosure
Self-disclosures by the user increased perceptions of user trust. A higher number of total posts by the user was associated with increased trust, as was the total number of people that the user \textquoteleft follows\textquoteright on their accounts. These user self-disclosures of total posts shared and total number of accounts the user follows, illustrate openness and sociability of the user, factors that increase trust and reduce risk.
\end{description}
\begin{table}[ht!]
\caption{Factor Matrix}
\centering
\begin{tabular}{|c|r|c|c|}
\hline
\multirow{2}{*}{}
& \multicolumn{3}{|c|}{\textbf{Component}} \\
\cline{2-4}
& 1 & 2 & 3 \\
\hline
\hline
\texttt{post} & .760 & -.106 & -.532\\
\texttt{follower}& .847 & -.288 & .288\\
\texttt{following}& .676 & .115 & -.652\\
\texttt{likes}& .763 & -.440 & .293\\
\texttt{comments}& .747 & -.334 & .341\\
\texttt{total\_person}& .251 & .817 & .180\\
\texttt{pic\_person}& .365 & .894 & .111\\
\texttt{self}& .338 & .889 & .110\\
\hline
\end{tabular}%
\label{tab:factor_matrix}
\end{table}
\begin{table}[ht!]
\caption{Factor Matrix after Varimax rotation}
\centering
\begin{tabular}{|c|r|c|c|}
\hline
\multirow{2}{*}{}
& \multicolumn{3}{|c|}{\textbf{Component}} \\
\cline{2-4}
& 1 & 2 & 3 \\
\hline
\hline
\texttt{post} & .018 & .350 & \cellcolor[rgb]{.7,.9,.9}.886\\
\texttt{follower}& .059 & \cellcolor[rgb]{.7,.9,.9}.909 & .234\\
\texttt{following}& .170 & .139 & \cellcolor[rgb]{.7,.9,.9}.920\\
\texttt{likes}& -.107 & \cellcolor[rgb]{.7,.9,.9}.904 & .181\\
\texttt{comments}& -.003 & \cellcolor[rgb]{.7,.9,.9}.876 & .133\\
\texttt{total\_person}& \cellcolor[rgb]{.7,.9,.9}.873 & -.025 & .001\\
\texttt{pic\_person}& \cellcolor[rgb]{.7,.9,.9}.965 & -.002 & .123\\
\texttt{self}& \cellcolor[rgb]{.7,.9,.9}.952 & -.021 & .109\\
\hline
\end{tabular}%
\label{tab:factor_rotation_matrix}
\end{table}
Among the three factors, Social Proof is associated with persons in pictures and influenced trust formation the most. Pictures may be manipulated to increase trustworthiness. However, technical advances in picture analysis using deep learning makes recognizing a person very accurate. It is not difficult to require a real profile picture and identify and label each person in the picture. Social Approval, on the other hand, is controlled by other users and difficult to manipulate. User self-disclosure explains 13\% of the variance and will take sustained efforts over time to arbitrarily boost it up.
\section{Results of Logistic Regression}
In our survey, we have six different questions. The answer to each of our survey questions is binary. Thus, we built a logistic regression model to analyze the relation between factors and the results of surveys. We used precision, recall, and F-measure as the indicators of logistic regression. We first analyzed the original eight factors. Table 10 shows detailed results. Then, we analyzed the three factors produced by factor analysis. The detailed results are shown in table 11.
\begin{table*}
\centering
\caption{Result of eight factors}
\begin{tabular}{|c|c|c|c|c|c|c|} \hline
\textbf{}&\textbf{\textbf{Question 1}} &\textbf{\textbf{Question 2}} &\textbf{\textbf{Question 3}} &\textbf{\textbf{Question 4}} &\textbf{\textbf{Question 5}} &\textbf{\textbf{Question 6}}\\ \hline \hline
\texttt{\textbf{\textbf{Precision}}} & 0.783 & 0.782 & 0.714 & 0.609 & 0.746 & 0.700 \\
\hline
\texttt{\textbf{\textbf{Recall}}} & 0.790 & 0.788 & 0.730 & 0.610 & 0.760 & 0.710 \\
\hline
\texttt{\textbf{\textbf{F-Measure}}} & 0.786 & 0.782 & 0.715 & 0.609 & 0.750 & 0.699 \\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{Result of three factors}
\begin{tabular}{|c|c|c|c|c|c|c|} \hline
\textbf{}&\textbf{\textbf{Question 1}} &\textbf{\textbf{Question 2}} &\textbf{\textbf{Question 3}} &\textbf{\textbf{Question 4}} &\textbf{\textbf{Question 5}} &\textbf{\textbf{Question 6}}\\ \hline \hline
\texttt{\textbf{\textbf{Precision}}} & 0.846 & 0.843 & 0.797 & 0.661 & 0.823 & 0.726 \\
\hline
\texttt{\textbf{\textbf{Recall}}} & 0.904 & 0.868 & 0.913 & 0.750 & 0.890 & 0.828 \\
\hline
\texttt{\textbf{\textbf{F-Measure}}} & 0.874 & 0.855 & 0.851 & 0.703 & 0.855 & 0.774 \\
\hline
\end{tabular}
\end{table*}
From Table 10 and Table 11, we can see that, the three key factors have a better result than the original eight factors.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{new_precision_1.pdf}
\caption{Precision}
\label{fig:precision}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{new_recall_1.pdf}
\caption{Recall}
\label{fig:recall}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{new_fmeasure_1.pdf}
\caption{F-Measure}
\label{fig:fmeasure}
\end{figure}
Figure 4 to Figure 6 also shows the comparison of eight factors and three key factors. We can see obviously that the result of three key factors is better than the result of eight factors in precision, recall, and F-measure, among all of the six questions.
\section{Conclusion}
In this paper, we combined social media and ridesharing systems together to analyze the influence of trust on the potential usage of ridesharing services. In our work, we collected the detailed social information from Instagram. We used Mechanical Turk to conduct a survey to learn the relationship between trust factors and the use of ridesharing. By analyzing different social attributes, we chose eight social attributes as trust factors. We then used EFA to summarize the eight factors into three key factors. Finally, we built a logistic regression model by using the results from the survey with eight factors and three factors separately. The results show that with using eight factors, we can achieve 78.6\% in F-Measure. By using only the three most important factors, the result will be improved to 87.4\%.
|
\section{Introduction}
Interesting new directions in topological quantum computing include its extension from anyons to gapped boundaries and symmetry defects with the hope that anyonic systems with non-universal computational power can be enhanced to achieve universality. Enrichment of topological physics in two spatial dimensions by gapped boundaries has been investigated intensively, but their computing power has not been analyzed in detail yet. One interesting case is gapped boundaries of Dijkgraaf-Witten theories both for their experimental relevance and as theoretical exemplars (see \cite{cong2016topological,cong2017cmp,cong2017defects} and the references therein).
In this paper, we study representations of the braid groups from braiding gapped boundaries of Dijkgraaf-Witten theories and their twisted generalizations, which are (twisted) quantum doubled topological orders in two spatial dimensions.
We show that the resulting braid (pure braid) representations are all monomial with respect to some specific bases, hence all such representation images of the braid groups are finite groups (see also \cite{Rowell}). We give explicit formulas for the monomial matrices and the ground state degeneracy of the Kitaev models that are Hamiltonian realizations of Dijkgraaf-Witten theories. Our results imply that braiding gapped boundaries alone cannot provide universal gate sets for topological quantum computing with gapped boundaries.
For a topological order of the form $\mathcal{C}= \mathcal{Z}(\mathcal{S})$, were $\mathcal{S}$ is some unitary fusion category, gapped boundaries are modelled by Lagrangian algebras (see \cite{cong2016topological} and the references therein). For these models the ground manifolds have the form $\operatorname{Hom}_\mathcal{C}(1,A_1\otimes \cdots \otimes A_n)$, where the $A_i$'s are the Lagrangian algebras modelling the gapped boundaries, see \cite[Section 3]{cong2016topological} for details. Recall that a Lagrangian algebra in any modular (tensor) category is a commutative etále algebra whose quantum dimension is maximal. A group theoretical modular category (GTMC) is a category of the form $\mathcal{C}= \mathcal{Z}(\operatorname{Vec}_G^\omega)$ for some finite group $G$ and some $\omega\in Z^3(G,\mathbb{C}^\times)$, where $\mathcal{Z}$ denotes the Drinfeld center and $\operatorname{Vec}_G^\omega$ is the category of finite dimensional $G$-graded vectors spaces with associativity constraint twisted by $\omega\in H^3(G,\mathbb{C}^\times)$.
Kitaev \cite{kitaev2003fault} proposed Hamiltonian realizations of Dijkgraaf-Witten theories, who\-se topological orders are GTMCs. Moreover, extensions of these Hamiltonian realizations to surfaces with boundaries can be constructed from Lagrangian algebras \cite{bravyi1998quantum,bombin2008family,beigi2011quantum,kitaev2012models}.
Lagrangian algebras in GTMC's are one-one correspondence with indecomposable modular categories of $\operatorname{Vec}_G^\omega$ \cite{davydov2013witt}, which are in bijection with pairs $(H,\gamma)$, where $H$ is a subgroup of $G$ and $\gamma\in C^2(H,\mathbb{C}^\times)$ such that $\delta(\gamma)=\omega|_{H^{\times 3}}$, all up to conjugation \cite{natale2016equivalence}. A more direct description between Lagrangian algebras and pairs $(H,\gamma)$ can be found in \cite{davydov2010modular}.
Recently, a quantum computing scheme to use gapped boundaries to achieve universality is proposed \cite{cong2016topological, cong2017cmp,cong2017defects,cong2017universal}. Braiding gapped boundaries can be either added to braiding anyons as in Kitaev's original proposal or as new computing primitives supplemented with other topological operations. Gapped boundaries lead to additional degeneracy to the topologically protected subspace, which potentially allows the implementation of more powerful gates. More precisely, the new gates come from representation matrices of the braid groups, $\mathcal{B}_n$, on objects of the GTMCs that are tensor products of Lagrangian algebras. But a characterization of the computational power of these new braid representations, mathematically a study of the representation images, was left as an important open problem \cite{cong2016topological,cong2017universal}.
The goal of this paper is to provide such a characterization. We find a canonical monomial structure for Lagrangian algebras in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$, which allows us to compute things more easily. This paper is organized as follows. Section \ref{Monomial} develops the theory of monomial representations. Specifically, it shows how to calculate invariants for a representation of $G$ using the monomial structure. In Section \ref{Section: monomial twisted Yetter-Drinfeld} we introduce the notion of monomial twisted Yetter-Drinfeld. We use the theory developed in Section \ref{Monomial} to give an explicit description and a basis for $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C}, V^{\otimes n})$ if $V$ is a monomial object. Next, we describe the representation of $\mathcal{B}_n$ with respect to this basis. Theorem \ref{thm: YD monomial } states the representation is monomial and Theorem \ref{thm:action in basis} gives an explicit formula for the non-zero entries. In Section \ref{Section: representation Lagrangian} we prove that every Lagrangian algebra in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$ has a canonical monomial structure. Then the results of Section \ref{Section: monomial twisted Yetter-Drinfeld} are applied to Lagrangian algebras in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$. We finish the section developing some examples and applications.
\section{Monomial representations}\label{Monomial}
In this section we recall some basic definitions and results on monomial representations of groups.
\begin{defi}
A \textit{monomial space} is a triple $\mathbf{V}=(V,X,(V_x)_{x\in X})$ where,
\begin{itemize}
\item[(i)] V is a finite dimensional complex vector space.
\item[(ii)] X is a finite set.
\item[(iii)] $(V_x)_{x\in X}$ is a family of one dimensional subspaces of $V$, indexed by $X$, such that $ V=\bigoplus_{x\in X} V_x$.
\end{itemize}
Let $G$ be a group. By a \textit{monomial representation} of $G$ on $\mathbf{V}$ we mean a group homomorphism \[\Gamma: G\to \operatorname{GL}(V),\] such that for every $g\in G,$ $\Gamma(g)$ permutes the $V_x$'s; hence, $\Gamma$ induces an action by permutation of $G$ on $X$. We will denote $\Gamma(g)(v)$ just by $g\triangleright v$.
\end{defi}
If $V$ is a representation of $G$, we denote by $V^G$
the subspace of $G$-invariant vectors, \textit{i.e.},
\[V^G=\{v\in V: g\triangleright v=v, \text{ for all } g\in G\}.\]
For each $x\in X$, we will denote $\operatorname{Sta}_G(x)$ the stabilizer of $x$ and by $\mathcal{O}_G(x)$ the $G$-orbit of $x$. For $G$ finite, and a representation $V$ define
\begin{align*}
\operatorname{Av}_G: V \to V,&&
v\mapsto \frac{1}{|G|} \sum_{g\in G} g\triangleright v.
\end{align*}
It is easy to see that $\operatorname{Av}_G$ is a $G$-linear projection onto $V^G$. We define,\begin{align*}
\operatorname{Av}_G(V_\mathcal{O}):=\operatorname{Av}_G(V_x),&& x\in \mathcal{O}(x),
\end{align*}
since for any $x'\in \mathcal{O}_G(x)$, $\operatorname{Av}_G(V_x)=\operatorname{Av}_G(V_{x'})$.
We say that an
element $x \in X$ is \textit{regular} under the monomial action of $G$ if $\Gamma(g)$ is the identity map on $V_x$, for all $g\in \operatorname{Sta}_G(x)$.
Let us write $X/G$ for the set of orbits of the action of $G$ on $X$ and $\tilde{X}$ for the regular ones.
\begin{prop}\cite[Lemma 9.1]{karpilovsky1985projective}\label{Prop:Lemma Karp}
Let $\mathbf{V}=(V,X,(V_x)_{x\in X})$ be a monomial representation of $G$.
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item $x\in X$ is a regular element if and only if $\operatorname{Av}_G(V_x)\neq 0$.
\item If $x \in X$ is a regular element under the monomial action of $G$, then
so are all elements in the $G$-orbit of $x$.
\item
The triple
\[
\mathbf{V}^G = \left( V^G, \tilde{X}, \big (\operatorname{Av}_G(V_\mathcal{O})\big)_{\mathcal{O} \in \tilde{X}} \right)
\]
is a monomial space.
\item The dimension of $V^
G$ is equal to the number of regular $G$-orbits
under the monomial action of $G$ on $X$.
\end{enumerate}
\end{prop}\qed
Let $\mathbf{V}=(V,X,(V_x)_{x\in X})$ and $\mathbf{V}'=(V',Y,(V'_y)_{y\in Y})$ be monomial spaces. A linear isomorphism $T:V\rightarrow V'$ is called an \textit{isomorphism of monomial spaces} if $T(V_x)=V'_y$ for any $x\in X$.
\begin{prop}\label{iso of invariants}
Let $\mathbf{V}=(V,X,(V_x)_{x\in X})$ and $\mathbf{V}'=(V',Y,(V'_y)_{y\in Y})$ be monomial representations of a finite group $G$. If $T:V\to V'$ is a $G$-linear isomorphism of monomial spaces, then $T|_{V^G}:\mathbf{V}^G\to \mathbf{V}'^G$ is an isomorphism of monomial spaces.
\end{prop}
\begin{proof}
Clearly, $T|_{V^G}:V^G \rightarrow V'^G$ is a linear isomorphism.
Let $x\in X$ be a regular element. Since $T$ is an isomorphism of monomial spaces, there is some $y \in Y$ such that $T(V_x)=V'_y$. In that case:
\[
\operatorname{Av}_G(V'_y)=\operatorname{Av}_G(T(V_x))=T(\operatorname{Av}_G(V_x)).
\]
This implies $y$ is regular, because $\operatorname{Av}_G(V_x)\neq \{0\}$ and $T$ is an isomorphism. It also says $T|_{V^G}(\operatorname{Av}_G(V_{\mathcal{O}(x)}))=\operatorname{Av}_G(V_{\mathcal{O}(y)}')$, which means $T|_{V^G}$ is an isomorphism of monomial spaces.
\end{proof}
\section{Monomial representation of the braid group}\label{Section: monomial twisted Yetter-Drinfeld}
In this section we introduce the notion of monomial twisted Yetter-Drinfeld and prove that the representation of the braid groups $\mathcal{B}_n$ over $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C},V^{\otimes n})$ is monomial if $V$ is monomial.
\subsection{Dijkgraaf-Witten theories}
Let $G$ be a discrete group. A (normalized) 3-cocycle $\omega \in Z^3(G, \mathbb{C}^\times)$ is a map $\omega:G\times G\to \mathbb{C}^\times$ such that
\begin{align*}
\omega(ab,c,d)\omega(a,b,cd)&=
\omega(a,b,c)\omega(a,bc,d)\omega(b,c,d), & \omega(a,1,b)=1,
\end{align*}
for all $a,b,c,d\in G. $
Let us recall the description of the modular category $\mathcal{Z}(\operatorname{Vec}_G^\omega)$, the Drinfeld center of the category $\operatorname{Vec}_G^\omega$ sometimes called the category of twisted Yetter-Drinfeld modules. The category $\mathcal{Z}(\operatorname{Vec}_G^\omega)$ is braided equivalent to the representations of the twisted Drinfeld double defined by Dijkgraaf, Pasquier and Roche in
\cite[Section 3.2]{DPR}.
Given $\omega \in Z^3(G; \mathbb{C}^\times)$, we define
\begin{align*}
\omega(g,g';h)&:=\frac{\omega(g,^{g'}h,g')}{\omega(^{gg'}h,g,g')\omega(g,g',h)},\\
\omega(g;f,h)&:=\omega(g,f,h)\omega(\prescript{g}{}{f},g,h)^{-1}\omega(\prescript{g}{}{f},\prescript{g}{}{h},g),
\end{align*}
for $f,g,g^\prime, h \in G$.
The objects of $\mathcal{Z}(\operatorname{Vec}_G^\omega)$ are $G$-graded vector spaces $V=\bigoplus_{g\in G} V_{g}$ with a linear map
$\triangleright: \mathbb{C}^{\omega}G\otimes V\to V$
such that $1\triangleright v=v$ for all $v\in V$,
\begin{align*}
(gh)\triangleright v &=\omega(g,h;k) (g\triangleright (h\triangleright v)), & g,h,k & \in G, & v &\in V_k,
\end{align*}
satisfying the following compatibility condition:
\begin{align*}
g\triangleright V_h & \subseteq V_{ghg^{-1}}, & g,h & \in G.
\end{align*}
Morphisms in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$ are $G$-linear $G$-homogeneous maps.
The tensor product of $V=\oplus_{g\in G}V$ and $W=\oplus_{g\in G}w$ is
$V\otimes W$ as vector space, with
\[(V\otimes W)_g=\bigoplus_{h\in G}V_h\otimes W_{h^{-1}g},\]
and for all $v\in V_g, w\in W_l$,
\[h\triangleright (v\otimes w)= \omega(h;g,l)(h\triangleright v) \otimes (h\triangleright w).\]
For $V, W, Z\in\ \mathcal{Z}(\operatorname{Vec}_G^\omega)$ the associativity constrain is defined by \begin{align*}
a_{V,W,Z}: (V\otimes W)\otimes Z&\to V\otimes (W\otimes Z)\\
(v_g\otimes w_h)\otimes z_k &\mapsto \omega(g,h,k) v_g\otimes( w_h\otimes z_k)
\end{align*}
for all $g,h,k \in G, v_g\in V_x, w_h\in W_h, z_k\in Z_k$. The category is tensor braided, with braiding $c_{V,W}:V\otimes W\to W\otimes V$, $V,W\in\mathcal{Z}(\operatorname{Vec}_G^\omega)$,
\begin{align*}
c_{V,W}(v\otimes w) &= (g\triangleright w) \otimes v, & g & \in G, & v &\in V_g, & w&\in W.
\end{align*}
\subsection{Braid group representation of twisted Yetter-Drinfeld modules}
Since the braided category $\mathcal{Z}(\operatorname{Vec}_G^\omega)$ is not strict, we must be careful about the way we associate terms when we consider tensor products with more than two objects. For a list of objects $A_1, A_2, \ldots, A_n \in \mathcal{Z}(\operatorname{Vec}_G^\omega)$ we define
\[A_1\otimes \cdots \otimes A_n := (\cdots (A_1\otimes A_2)\otimes \cdots \otimes A_{n-1})\otimes A_n, \]
and an isomorphism by
\begin{align}
\sigma_i' = (a^{-1}_{A_1 \otimes \cdots \otimes A_{i-1},A_{i+1},A_{i}}&\otimes \operatorname{id}_{A_{i+2} \otimes \cdots \otimes A_{n}})\circ \label{generadores}\\
(\operatorname{id}_{A_1\otimes \cdots A_{i-1}}&\otimes c_{A_i,A_{i+1}}\otimes \operatorname{id}_{A_{i+2} \otimes \cdots \otimes A_{n}})\circ \notag\\
(&a_{A_1 \otimes \cdots \otimes A_{i-1},A_i,A_{i+1}}\otimes \operatorname{id}_{A_{i+2} \otimes \cdots \otimes A_{n}}), \notag
\end{align}where $a_{V,W,Z}$ denotes the associativity constrains.
If $A=A_1=\cdots =A_n$, there exists a unique group homomorphism $\rho_n:\mathcal{B}_n\to \operatorname{Aut}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(A^{\otimes n})$ sending the generator $\sigma_i \in \mathcal{B}_n$ to $\sigma'_i$.
In general, the pure braid group $\mathcal{P}_n$ acts on $A_1\otimes \cdots \otimes A_n$, in the sense that there exists group homomorphism $\rho_n:\mathcal{P}_n\to \operatorname{Aut}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(A_1\otimes \cdots \otimes A_n)$.
\subsection{Crossed $G$-sets}
Let $G$ be a group. A (left) \textit{crossed} $G$-\textit{set} is a left $G$-set $X$ and a grading function $|-|:X\to G$ such that
\[|gx|=g|x|g^{-1}\]for all $x\in X, g\in G$.
If $X$ and $Y$ are crossed $G$-sets, a $G$-equivariant map $f:X\to Y$ is a morphism of crossed $G$-sets if $|f(x)|=|x|$ for all $x\in X$.
If $X$ and $Y$ are crossed $G$-sets, the cartesian product $X\times Y$ is a crossed $G$-set with the diagonal action and grading map$|(x,y)|=|x||y|$.
The category of crossed $G$-sets is a braided category with braiding
\begin{align*}
c_{X,Y}:X\times Y&\to Y\times X\\
(x,y)&\mapsto (|x|\triangleright y,x).
\end{align*}
Thus, given a crossed $G$-set $X$ the braid group $\mathcal{B}_n$ acts on $X^{n}$, in the following way
\[\sigma_i':=\operatorname{id}_{X^{ i-1}}\times c_{X,X}\times \operatorname{id}_{X^{n-(i-1)}}.\]
\subsection{Monomial objects of $\mathcal{Z}(\operatorname{Vec}_G^\omega)$}
Let $G$ be a finite group and $\omega \in Z^3(G,\mathbb{C}^\times)$ a 3-cocycle.
\begin{defi}
A monomial Yetter-Drinfeld module is a monomial space $\mathbf{V}=(V,X,(V_x)_{x\in X})$ such that $V\in \mathcal{Z}(\operatorname{Vec}_G^\omega)$, the twisted $G$-action $\triangleright$ permutes the $V_x$'s and each $V_x$ is $G$-homogeneous.
\end{defi}
\begin{rem}
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item If $\mathbf{V}=(V,X,(V_x)_{x\in X})$ is a monomial Yetter-Drinfeld module, the set $X$ is a crossed $G$-set with the induced $G$-action and the grading map.
\item If $\mathbf{V}=(V,X,(V_x)_{x\in X})$ is a monomial Yetter-Drinfeld module, the action of $G$ on $(V_e,X_e,(V_x)_{x\in X_e})$ is monomial, where $X_e:=\{x\in X: |x|=e\}$ and $V_e=\oplus_{x\in X_e}V_x$.
\end{enumerate}
\end{rem}
\begin{thm}\label{thm: YD monomial }
Let $G$ be a finite group, $\omega \in Z^3(G,\mathbb{C}^\times)$. If $\mathbf{V}=(V,X,(V_x)_{x\in X})$ is a monomial Yetter-Drinfeld module in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$, then
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item the action of $\mathcal{B}_n$ on $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C}, V^{\otimes n} )$ is monomial,
\item the dimension of $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C}, V^{\otimes n} )$ is equal to the number of regular $G$-orbits under the monomial action of $G$ on \[(X^{ n})_e:=\{(x_1,\ldots,x_n): |x_1|\cdots |x_n|=e\}.\]
\end{enumerate}
\end{thm}
\begin{proof}
The action of $G$ on $(V^{\otimes n}_e,(X^{n})_e,(V_x)_{x\in X_e})$ is monomial. Hence by Proposition \ref{Prop:Lemma Karp}, the triple
\[
\mathbf{V}_e^G := \left( ((V^{\otimes n}_e)^G, \widetilde{(X^{n})_e}, \big (\operatorname{Av}_G((V^{\otimes n}_e)_\mathcal{O})\big)_{\mathcal{O} \in \widetilde{(X^{n})_e}} \right)
\]
is a monomial space. Since $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C}, V^{\otimes n} )= (V^{\otimes n})_e^G$, and each of the automorphisms $\sigma'$ are morphisms in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$, hence $\sigma'|_{V^{\otimes n}_e}:(V^{\otimes n}_e,(X^{n})_e,(V_x)_{x\in X_e})\to (V^{\otimes n}_e,(X^{n})_e,(V_x)_{x\in X_e})$ is a $G$-linear isomorphism of monomial spaces. It follows from Proposition \ref{iso of invariants} that $\sigma'|_{(V^{\otimes n})_e^G}$ is an isomorphism of monomial spaces. Thus, the linear representation \begin{align*}
\rho_n: \mathcal{B}_n&\to \operatorname{GL}((V^{\otimes n}_e)^G)\\
\sigma &\mapsto \sigma',
\end{align*} is a monomial representation of $\mathcal{B}_n$.
The second part follows immediately from Proposition \ref{Prop:Lemma Karp}.
\end{proof}
\subsection{Monomial matrices of the braid representation}\label{subsection:monomial basis}
In this subsection we obtain concrete formulas for the monomial braid representations associated to a monomial Yetter-Drinfeld module.
Let $G$ be a finite group, $\omega \in Z^3(G,\mathbb{C}^\times)$ and $\mathbf{V}=(V,X,(V_x)_{x\in X})$ be a monomial Yetter-Drinfeld module. If we fix non-zero vectors $\mathcal{S}:=\{v_x\in V_x:x\in X\}$, the twisted $G$-action defines a map \[\lambda_X:G\times X\to \mathbb{C}^\times,\]
by $g \triangleright v_x= \lambda_X(g;x)v_{gx}$, where $g \in G, x\in X$.
For the monomial Yetter-Drinfeld module $\mathbf{V}^{\otimes n}=(V^{\otimes n} ,X^{n},(V_x)_{x\in X^{n}})$ and the basis $\mathcal{S}^{\otimes n}:=\{v_{x_1}\otimes \cdots \otimes v_{x_n}:x_i\in X, \ 1\leq i\leq n\}$, the
action is determined by the map $\lambda_{X^n}:G\times X^n\to \mathbb{C}^\times,$
\begin{align}
\lambda_{X^n}(g;x_1,\ldots,x_n)&:=\prod_{i=1}^n\lambda_X(g;x_i)\omega(g;|x_1||x_2|\cdots |x_{n-1}|,|x_n|)\times \label{definition of lambda x n}\\ &\omega(g;|x_1|\cdots |x_{n-2}|,|x_{n-1}|)
\cdots \omega(g;|x_1|,|x_2|), \notag
\end{align}that is, \[g\rhd (v_{x_1}\otimes \cdots \otimes v_{x_n})= \lambda_{X^n}(g;x_1,\ldots,x_n) (v_{gx_1}\otimes \cdots \otimes v_{gx_n}),\] for all $g\in G, x_1,x_2\ldots , x_n \in X$. Hence an element $(x_1,\ldots, x_n) \in (X^{n})_e$ is regular if and only if \begin{align}\label{condicion regular}
\lambda_{X^n}(g;x_1\ldots, x_n)=1,&& \text{ for all } g \in \bigcap_{i=1}^n \operatorname{Sta}(x_i).
\end{align}
Let $\mathcal{R}\subset X^n_e$ be a set of representatives of the regular orbits of $X_e^{\times n}$. Let $\mathcal{S}_{reg}= \{v_{x_1}\otimes \cdots \otimes v_{x_n}: (x_1,\ldots,x_n)\in R \}$. By Proposition \eqref{Prop:Lemma Karp} the set $\{\operatorname{Av}_G(v): v\in \mathcal{S}_{reg}\}$ is a basis of $(V^{\otimes n})_e^G$.
In order to express the action of the generator $\sigma_i \in \mathcal{B}_n$ in terms of $\{\operatorname{Av}_G(v): v\in \mathcal{S}_{reg}\}$, for each $\mathbf{x}=(x_1,\ldots,x_n)\in \mathcal{R}$ choose $g_{\mathbf{x}}\in G$ such that $g_{\mathbf{x}}\triangleright \sigma_i'(\mathbf{x})=\mathbf{y}$, where $\mathbf{y}\in \mathcal{R}$ and $\sigma_i'(\mathbf{x})= (x_1, \cdots , x_{i-1}, |x_i|x_{i+1}, x_i, \cdots , x_n)$. Hence there is $\beta_{i,\mathbf{x}}\in \mathbb{C}^\times$ such that $g_{\mathbf{x}}\triangleright\sigma_i'(v_{x_1}\otimes \cdots v_{x_n})=\beta_{i,\mathbf{x}}v_{y_1}\otimes \cdots \otimes v_{y_n}$.
Since the action of the generator $\sigma_i \in \mathcal{B}_n$ is given by \begin{align}
\sigma'(v_{x_1}\otimes \cdots v_{x_n})&= \omega^{-1}(|x_1|\cdots |x_{i-1}|,|x_{i+1}|,|x_i|)\times \label{formula generador grupo de trenza}\\
&\lambda_X(|x_i|;x_{i+1})\omega(|x_1|\cdots |x_{i-1}|,|X_i|,|x_{i+1}|)\times \notag\\ & v_{x_1}\otimes \cdots \otimes v_{x_{i-1}}\otimes v_{|x_i|x_{i+1}}\otimes v_{x_i}\otimes \cdots \otimes v_{x_n}.\notag
\end{align}
we have that
\begin{align}
\beta_{i,\mathbf{x}}&= \omega^{-1}(|x_1|\cdots |x_{i-1}|,|x_{i+1}|,|x_i|)\times\label{def:beta} \\
&\lambda_X(|x_i|;x_{i+1})\omega(|x_1|\cdots |x_{i-1}|,|X_i|,|x_{i+1}|)\lambda_{X^n}(g_{\mathbf{x}};\sigma_i'(\mathbf{x})).\notag
\end{align}
\begin{thm}\label{thm:action in basis}
Let $G$ be a finite group, $\omega \in Z^3(G,\mathbb{C}^\times)$ and $\mathbf{V}=(V,X,(V_x)_{x\in X})$ be a monomial Yetter-Drinfeld module. Let $Y$ be the set of all regular elements in $X^n_e$ and $\mathcal{R}\subset Y$ a set of representatives of the $G$-orbits of $Y$.
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item The projection $\pi :Y\to \mathcal{R}$ is map of $\mathcal{B}_n$-sets. The image of $\mathbf{x}\in \mathcal{R}$ by the generator $\sigma_i\in \mathcal{B}_n$ will be denoted by $\sigma_i\triangleright \mathbf{x}$.
\item Let $\mathcal{S}_{reg}= \{v_{x_1}\otimes \cdots \otimes v_{x_n}: (x_1,\ldots,x_n)\in \mathcal{R} \}$. The action of the generator $\sigma_i \in \mathcal{B}_n$ in the basis $\{\operatorname{Av}_G(v_{\mathbf{x}}): \mathbf{x}\in \mathcal{R}\}$ is given by \[\sigma_i(\operatorname{Av}_G(v_{\mathbf{x}}))= \beta_{i,\mathbf{x}}\operatorname{Av}_G(v_{\sigma_i \triangleright \mathbf{x}}),\]where $\beta_{i,\mathbf{x}}$ was defined in \eqref{def:beta}.
\end{enumerate}
\end{thm}
\begin{proof}
The first part is consequence of Theorem \ref{thm: YD monomial }.
For the second part, recall that the number $\beta_{i,\mathbf{x}}$ and the element $g_\mathbf{x}\in G$ are such that \[g_\mathbf{x}\triangleright \sigma(v_{\mathbf{x}})=\beta_{i,\mathbf{x}}v_{\sigma_i\triangleright \mathbf{x}}.\] Hence,
\begin{align*}
\sigma_i(\operatorname{Av}_G(v_{\mathbf{x}}))&= \operatorname{Av}_G(\sigma_i(v_{\mathbf{x}}))\\
&=g_\mathbf{x}\triangleright \operatorname{Av}_G(\sigma_i(v_{\mathbf{x}}))\\
&=\operatorname{Av}_G(g_\mathbf{x}\triangleright\sigma_i(v_{\mathbf{x}}))\\
&=\operatorname{Av}_G(\beta_{i,\mathbf{x}}v_{\sigma_i\triangleright \mathbf{x}})\\
&=\beta_{i,\mathbf{x}}\operatorname{Av}_G(v_{\sigma_i \triangleright \mathbf{x}}).
\end{align*}
\end{proof}
\begin{example}\label{ejemplo permutation crossed g-sets}
Let $G$ be a finite group and $X$ be a left crossed $G$-set. Then the linearization $V_X:=\oplus_{x\in X}\mathbb{C}x$ is a (untwisted) Yetter-Drinfeld module in $\mathcal{Z}(\operatorname{Vec}_G)$. Clearly $\lambda_X\equiv 1$, thus every element in $(X^{n})_e$ is regular. Hence the canonical projection
\[(X^{ n})_e\to (X^{ n})_e//G,\]is an epimorphism of $\mathcal{B}_n$-sets. In other words, the linear representation of $\mathcal{B}_n$ on $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G)}(\mathbb{C}, V_X^{\otimes n} )$ is the linearization of the permutation action of $\mathcal{B}_n$ on $(X^{ n})_e//G$.
\end{example}
\section{Braid groups representations Associated to Lagrangian algebras}\label{Section: representation Lagrangian}
In this section we prove that every Lagrangian algebra $\mathcal{Z}(\operatorname{Vec}_G^\omega)$ as a canonical monomial structure. Then the results of Section \ref{Section: monomial twisted Yetter-Drinfeld} can be applied to Lagrangian algebras in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$.
\subsection{Lagrangian algebras }
Following \cite[Corollary 3.17]{davydov2017lagrangian}, we will describe the Lagrangian algebra on $\mathcal{Z}(G,\ThreeCocycle)$ associated to a pair $(H,\gamma)$, where $H\subseteq G$, is a subgroup and $\gamma:H\times H\to \mathbb{C}^\times$ a map such \begin{align*}
\frac{\gamma(ab,c)\gamma(a,b)}{\gamma(a,bc)\gamma(b,c)}=\omega(a,b,c),&& a,b,c \in H.
\end{align*}
Let $\mathbb{C}_{\gamma}[H]=\oplus_{h\in H}\mathbb{C}e_h$ the group algebra of $H$ with the multiplication \begin{align*}
e_{h_1}e_{h_2}=\gamma(h_1,h_2)e_{h_1h_2},&& h_1,h_2 \in H.
\end{align*} The vector space $\mathbb{C}_{\gamma}[H]=\oplus_{h\in H}\mathbb{C}e_h$, is a commutative algebra in $\mathcal{Z}(\operatorname{Vec}_H^\omega)$, where the $H$-action
is given by
\begin{align*}
h_1\triangleright e_{h_2}=\epsilon(h_1,h_2)e_{h_1h_2h_1^{-1}},&& \epsilon(h_1,h_2):=\frac{\gamma(h_1,h_2)}{\gamma(\prescript{h_1}{}{h_2},h_1)},&& h_1,h_2 \in H,
\end{align*}and grading $|e_h|=h$ for all $h\in H$.
Let $\operatorname{Map}(G, \mathbb{C}_{\gamma}[H])$ be
the vector space of all set-theoretic maps from $G$ to $\mathbb{C}_{\gamma}[H]$. With the grading given by
\begin{align*}
\vert a\vert&=f &\Leftrightarrow &&\forall x\in G \quad \vert a(x)\vert =x^{-1}fx,
\end{align*}
and twisted $G$-action
\begin{align*}\label{G-action}
(g\triangleright a)(x):=\omega(x^{-1},g^{-1}; |a|)^{-1}a(g^{-1}\triangleright x), && g,x\in G.
\end{align*}
$\operatorname{Map}(G, \mathbb{C}_{\gamma}[H])$ is twisted Yetter-Drinfeld module.
The Lagrangian algebra $L(H,\gamma)$ is the Yetter-Drinfeld submodule
\begin{align*}
L(H,\gamma):=\{a\in \textrm{Maps}(G,\mathbb{C}_{\gamma}[H])\vert \, a(xh)=\omega(h^{-1},x^{-1}; |a|)h^{-1}\triangleright a(g) \},
\end{align*}see \cite{davydov2017lagrangian} for more details.
\subsection{Monomial structure of the Lagrangian algebras $L(H,\gamma)$}
In this section we will proved that every Lagrangian algebra of the form $L(H,\gamma)$ has a canonical monomial structure.
Let $G$ be a group and $H\subset G$ a subgroup. We can regard $G\times H$ as a left $H$-set with actions given $h\triangleright (g,h')=(gh^{-1},hh'h^{-1})$. Then we can consider the set of $H$-orbits that we will denote by $G \times_H H$. The set $G \times_H H$ is equipped with a left $G$-action given by left multiplication on the first component.
\begin{defi}
Let $L(H,\gamma)$ be a Lagrangian. For each $g\in G$, $f\in H$, define $\chi_{g,f}\in L(H,\gamma)$ by
\begin{equation}
\chi_{g,f}(x)=
\begin{cases}
0, & x\notin gH \\
\omega(h^{-1},g^{-1}; \prescript{g}{}{f})\epsilon(h^{-1},f)e_{hfh^{-1}}, & x=gh, \text{ where } h\in H.
\end{cases}
\end{equation}
\end{defi}
\begin{rem}
The function $\chi_{g,h}$ can be characterized as the unique map in $L(H,\gamma)$ with support $gH$ and such that $\chi_{g,h}(g)=e_h$.
\end{rem}
\begin{lem}\label{Lemma: equaciones}
Let $L(H,\gamma)$ be a Lagrangian algebra in $\mathcal{Z}(G,\ThreeCocycle)$. Then
\begin{align}\label{equ1: chi}
\chi_{gh,f}=\omega(h,(gh)^{-1}; \prescript{gh}{}{f})\epsilon(h,f)\chi_{g,\prescript{h}{}{f}},&& g\in G, f,h\in H.
\end{align}
\begin{align}\label{equ2: chi}
l\triangleright \chi_{g,f}=\omega((lg)^{-1},l^{-1};\prescript{g}{}{f})\chi_{lg,f},&& g,l \in G, h\in H.
\end{align}
\end{lem}
\begin{proof}
\eqref{equ1: chi}. Since the supports of $\chi_{gh,f}$ and $\chi_{g,^hf}$ are $gH$, and
\begin{align*}
\chi_{gh,f}(g) &= \chi_{gh,f}(ghh^{-1})\\
&= \omega(h,(gh)^{-1}; \prescript{gh}{}{f})\epsilon(h,f)\chi_{g,\prescript{h}{}{f}}(g),
\end{align*} we obtain \eqref{equ1: chi}.
\eqref{equ2: chi}. By the definition of the action of $G$ we have \begin{align*}
l\triangleright \chi_{g,f}(lg)&=\omega((lg)^{-1},l^{-1};\prescript{g}{}{f})\chi_{g,f}(g)\\
&=\omega((lg)^{-1},l^{-1};\prescript{g}{}{f})e_f.
\end{align*}Since $l\rhd \chi_{g,f}$ and $\chi_{gl,f}$ are supported in $glH$, we get \eqref{equ2: chi}.
\end{proof}
It follows from Lemma \ref{Lemma: equaciones} that $\mathbb{C}\chi_{gh,\prescript{h}{}{f}}=\mathbb{C}\chi_{g,f}$. Then for any $(g,h)\in G\times_H H$ the space $\mathbb{C}\chi_{g,f}$ is well defined.
\begin{thm}\label{thm: Lagrangian are monomial}
Let $L(H,\gamma)$ be a Lagrangian algebra in $\mathcal{Z}(G,\ThreeCocycle)$. Then $L(H,\gamma)$ with the decomposition \[L(H,\gamma)=\bigoplus_{(g,h)\in G\times_H H}\mathbb{C}\chi_{g,h}\] is a monomial twisted Yetter-Drinfeld module.
\end{thm}
\begin{proof}
First we will check that in fact the sum $\sum_{(g,h)\in G\times_H H}\mathbb{C}\chi_{g,h},$ is direct. Since $\textrm{supp}(\chi_{g,f})=gH$, we have that $\chi_{g,f}$ and $\chi_{g^\prime,f}$ are linearly independent if $gH\neq g^\prime H$. Hence it is suffices to check linear independence of the collections $\{\chi_{g,f}\}_{f\in H}$, with $g$ fixed. But if $f\neq f^\prime$, $\vert \chi_{r,f}\vert \neq \vert \chi_{r,f^\prime}\vert$. It follows that the sum $\sum_{(g,h)\in G\times_H H}\mathbb{C}\chi_{g,h}$ is direct.
In order to see that $L(H,\gamma)=\sum_{(g,h)\in G\times_H H}\mathbb{C}\chi_{g,h}$, fix $\mathcal{R}\subset G$ a set of representative of the left coset of $H$ in $G$. Let $a\in L(H,\gamma)$. For each $r\in \mathcal{R}$, suppose
\begin{equation}
a(r)=\sum_{f\in H} \lambda_{r,f} e_f.
\end{equation}
Then we have
\begin{equation}
a=\sum_{r\in \mathcal{R},f\in H} \lambda_{r,f}\chi_{r,f}\in \sum_{(g,h)\in G\times_H H}\mathbb{C}\chi_{g,h}.
\end{equation}
By \eqref{equ2: chi} and the fact that $|\chi_{g,f}|=gfg^{-1}$, we obtain that $L(H,\gamma)$ is a monomial twisted Yetter-Drinfeld module.
\end{proof}
\begin{cor}\label{cor: Lagrangian are monomial }
Let $G$ be a finite group, $\omega \in Z^3(G,\mathbb{C}^\times)$. If $L(H,\gamma)$ is a Lagrangian algebra in $\mathcal{Z}(\operatorname{Vec}_G^\omega)$, then
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item the action of $\mathcal{B}_n$ on $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C}, L(H,\gamma)^{\otimes n} )$ is monomial,
\item the dimension of $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G^\omega)}(\mathbb{C}, L(H,\gamma)^{\otimes n} )$ is equal to the number of regular $G$-orbits under the monomial action of $G$ on \[(G\times_H H)^{\times n})_e:=\{((g_1,h_1),\ldots,(g_n,h_n)): g_1h_1g_1^{-1}g_2h_2g_2^{-1}\cdots g_nh_ng_n^{-1}=e\}.\]
\end{enumerate}
\end{cor}
\begin{proof}
Follows from Theorem \ref{thm: Lagrangian are monomial} and Theorem \ref{thm: YD monomial }.
\end{proof}
We will fix a set of representatives of the left cosets of $G$ in $H$, $\mathcal{R} \subset G$. Thus
every element $g \in G$ has a unique factorization $g = rh$, $h \in H, r \in \mathcal{R}$. We assume $e \in \mathcal{R}$. The uniqueness of the factorization $G = \mathcal{R} H$ implies that there are well defined maps
\begin{align*}
\triangleright: G\times \mathcal{R}\rightarrow \mathcal{R},&& \kappa:G\times \mathcal{R}\rightarrow H,
\end{align*}
determined by the condition
\begin{align*}
gr=(g\triangleright r)\kappa(g,h), && g\in G, r\in \mathcal{R}.
\end{align*}
As crossed $G$-set we can identify $G\times_H H$ with $\mathcal{R}\times H$ with action
\begin{align*}
g\triangleright(r,h):= (g\triangleright r,^{\kappa(g,r)}h),&& r\in \mathcal{R},\ h\in H,\ g\in G,
\end{align*}
and grading map
\begin{align*}
|-|:\mathcal{R}\times H&\to G\\
(r,h)&\mapsto rhr^{-1}.
\end{align*}
It follows from Theorem \eqref{thm: Lagrangian are monomial} that $B_\mathcal{R}:=\{\chi_{r,h} | \, r\in \mathcal{R}, \, h\in H\}$ is a basis for $L(H,\gamma)$.
In order to apply the results of Subsection \ref{subsection:monomial basis}, we only need to compute the map $\lambda_{\mathcal{R}\times H}:G\times (\mathcal{R}\times H)\to \mathbb{C}^\times$, such that
\begin{align*}
g \triangleright \chi_{r,h}= \lambda_{\mathcal{R}\times H}(g;r,h) \chi_{g\triangleright r,\prescript{\kappa(g,r)}{}{h}}, && g\in G, r\in \mathcal{R}, h\in H.
\end{align*}
Using Lemma \ref{Lemma: equaciones}
we obtain that
\begin{equation}\label{LambdaCoef}
\lambda_{\mathcal{R}\times H}(g;r,h)=
\omega((g r)^{-1},g^{-1}; \prescript{r}{}{h})
\omega(\kappa(g,r),(gr)^{-1}; \prescript{gr}{}{h})
\epsilon(\kappa(g,r),h),
\end{equation}for all $g\in G, r\in \mathcal{R}, h\in H$.
By \eqref{condicion regular},we have that an element $t=((r_1,h_1),\ldots,(r_n,h_n))\in (\mathcal{R}\times H)^n_e$ is regular if and only if
\begin{align}\label{regular lagrangian}
\lambda_{(\mathcal{R}\times H)^n}(g;(r_1,h_1),\ldots,(r_n,h_n))=1,&& \text{for all } g\in \bigcap_{i=1}^{n} r_i^{-1}C_H(h_i)r_i
\end{align}
where $\lambda_{(\mathcal{R}\times H)^n}$ was defined in \eqref{definition of lambda x n} in function of $\lambda_{\mathcal{R}\times H}$ and $\omega$.
\subsection{Applications and examples}
In this last section we present some application of the results of the previous section.
\subsubsection{Central Subgroups}
\begin{prop}
Let $G$ be a finite group and $L(H,\gamma)$ a Lagrangian algebra in $\mathcal{Z}(\operatorname{Vec}_G)$, where $H\subset G$ is a central subgroup. Then
\[\dim\Big (\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G)}(\mathbb{C},L(H,\gamma)^{\otimes n})\Big ) =|G|^{n-1}.\] Moreover, the representation of $\mathcal{B}_n$ is actually a representation of $S_n$.
\end{prop}
\begin{proof}
Since $H$ is a central subgroup, $g\triangleright (r,h) = (g\triangleright r, h)$ and
\begin{equation*}
|\chi_{r_1,h_1} \otimes \dots \otimes \chi_{r_k,h_k}| = h_1 \cdots h_k,
\end{equation*}
for any $ r_1,\dots,r_k \in R, h_1, \dots, h_n \in H$.
Hence,
\begin{align*}
|(\mathcal{R}\times H)^n_e|&=|(\mathcal{R}^n/G)| |H^{n-1}|\\
&= [G:H]^n|H|^{n-1}\\
&= |G|^{n-1}.
\end{align*}
To determine the number of orbits, notice that $\epsilon: H\times H\to \mathbb{C}^\times$ is a bicharacter such that $\epsilon(h_1,h_2)\epsilon(h_2,h_1)=1$. Then by equation \eqref{regular lagrangian} an element \[\Big( (r_1,h_1),\ldots , (r_n,h_n)\Big ) \in (\mathcal{R}\times H)_e^n\] is regular if and only if
\begin{align*}
\prod_{i=1}^n\epsilon(h,h_i)=1, && \text{for all } h\in H.
\end{align*}
But $\prod_{i=1}^n \epsilon(h,h_i) = \epsilon(h,h_1\cdots h_n) = \epsilon(h', e) = 1.$ Hence every element is regular. By Corollary \ref{cor: Lagrangian are monomial } the dimension of $\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G)}(\mathbb{C},L(H,\gamma)^{\otimes n})$ is $|G|^{n-1}$.
Finally, using equation \eqref{formula generador grupo de trenza}, we see that
\begin{align*}
\sigma'_i\circ\sigma'_i (\chi_{r_1,h_1}\otimes\cdots \otimes \chi_{r_n,h_n})&= \epsilon(h_i,h_{i+1})\epsilon(h_{i+1},h_{i}) (\chi_{r_1,h_1}\otimes\cdots \otimes \chi_{r_n,h_n})\\
&= \chi_{r_1,h_1}\otimes\cdots \otimes \chi_{r_n,h_n}.
\end{align*}
Hence representation of $\mathcal{B}_n$ factors as a representation of $S_n$.
\end{proof}
\subsubsection{Lagrangian algebra of the form $L(H,1)$}
The Lagrangian algebras $L(H,1)$ as an object in $\mathcal{Z}(\operatorname{Vec}_G)$ are completely determined by the crossed $G$-set $G\times_H$, and the monomial representation $\operatorname{Hom}(\mathbb{C}, L(H,1)^{\otimes})$ is a permutation representation, see Example \ref{ejemplo permutation crossed g-sets}. Let us see some extreme cases:
\subsubsection*{Case $H=\{e\}$}
In this case the crossed $G$-set is $G$ with the regular action and grading map the constant map $e$. It is clear that the braiding $c_{G,G}$
is just the flip map \[(g_1,g_2)\mapsto (g_2,g_1)\]hence, really the symmetric group $S_n$ acts on $G^n$.
The set of $G$-orbits is in biyection with $G^{n-1}$, \begin{align*}
\mathcal{O}(G^n)&\to G^{n-1}\\
\mathcal{O}_G(g_1,g_2,\ldots, g_n)&\mapsto (e,g_1^{-1}g_2,\ldots, g_1^{-1}g_n).
\end{align*}
Using the previous map the action of $\S_n$ is given by \[\sigma_1\Big( g_1,\ldots, g_{n-1}\Big)= \Big ( g_1^{-1},g^{-1}_1g_2,\ldots, g_1^{-1}g_{n-1}\Big)\]
and \[\sigma_i(g_1,\ldots, g_i,g_{i+1},\ldots,g_{n-1})=(g_1,\ldots,g_{i+1},g_i,\ldots,g_{n-1}), \quad 1<i<n.\]
It is clear that permutation action of $S_n$ on $G^{n-1}$ is faithful, thus the image is isomorphic to $S_n$.
\subsubsection*{Case $H=G$} In this case the crossed $G$-set is $G$ with the action by conjugation and grading map the identity map. Hence, the braiding is given by
\[c_{G,G}:(x,y)\mapsto (y,y^{-1}xy).\]
Note $c_{G,G}$ is symmetric if and only if $G$ is abelian.
If $G$ is abelian, $G^n_e=\{(g_1,\ldots,g_{n-1},(g_1\ldots g_{n-1})^{-1})\}$ is the set of orbits and as the previous example the group $S_n$ acts faithfully.
\subsubsection{Dihedral group}
Every time we take $H$ to be a normal subgroup of $G$, the following proposition provides a way to simplify the situation.
\begin{prop}\label{decoupledBasis}
Let $G$ be a finite group, $H\trianglelefteq G$, $\mathcal{R}$ a collection of representatives for $G/H$. Define $B_\gamma[H]\in\mathcal{Z}(\operatorname{Vec}_G)$ as
\begin{equation*}
B(H,\gamma) := \operatorname{span} \{ b_{r,h} |\, r\in \mathcal{R}, h\in H \}
\end{equation*}
with grading $|b_{r,h}|=h$ and the $G$-action
\begin{equation}\label{actionOnTheBs}
g\triangleright b_{r,h} = \epsilon(\kappa(g,r)\prescript{r^{-1}}{}{h}) b_{g\triangleright r, \prescript{g}{}{h}}.
\end{equation}
Then, the mapping
\begin{align*}
B(H,\gamma) &\rightarrow L(H,\gamma) \\
b_{r,h} &\mapsto \chi_{r, \prescript{r^{-1}}{}{h}}
\end{align*}
is an isomorphism in $\mathcal{Z}(\operatorname{Vec}_G)$.
\end{prop}
\begin{proof}
We need to show the map preserves the grading and the $G$-representation. We have
\begin{equation*}
|\chi_{r,\prescript{r^{-1}}{}{h}}|=\prescript{r}{}{(\prescript{r^{-1}}{}{h})} = h = |b_{r,h}|
\end{equation*}
Now, since
\begin{equation*}
g\cdot \chi_{r,\prescript{r^{-1}}{}{}h} = \epsilon(\kappa(g,r),\prescript{r^{-1}}{}{h})\chi_{g\triangleright r, \prescript{\kappa(g,r)}{}{(\prescript{r^{-1}}{}{h})} } ,
\end{equation*}
and
\begin{equation*}
\prescript{\kappa(g,r)}{}{(\prescript{r^{-1}}{}{h})} = \prescript{(g\triangleright r)^{-1}}{}{ghg^{-1}},
\end{equation*}
we have that
\begin{equation*}
g\triangleright b_{r,h} = \epsilon(\kappa(g,r),\prescript{r^-1}{}{h}) a_{g\triangleright r, \prescript{(g\triangleright r)^{-1}}{}{(\prescript{g}{}{h})}}.
\end{equation*}
Hence by \eqref{actionOnTheBs} the map is equivariant.
\end{proof}
Proposition \ref{decoupledBasis} works particularly well when $\gamma =1$, since equation \eqref{actionOnTheBs} is just
\begin{equation*}
g\triangleright b_{r,h} = b_{g\triangleright r, \prescript{g}{}{h}}.
\end{equation*}
Thus, the action of $G$ is "decoupled". We use this idea in the following example.
Let $G=D_{2k}$ be the dihedral groups of order $2k$ and $H=\langle r \rangle$. We take $\mathcal{R}=\{e,s\} = \{s^i\}_{i\in \mathbb{Z}/2\mathbb{Z}}$.
Then
\begin{equation*}
|b_{s^{i_1},r^{j_1}}\otimes \cdots \otimes b_{s^{i_n},r^{j_n}}| = r^{\sum_{m=1}^n j_m},
\end{equation*}
and
\begin{equation*}
\dim( B(H,\gamma)^{\otimes n}_e) = 2^n \times k^{n-1}.
\end{equation*}
Since
\begin{equation*}
(s^i r^j )(s^k) = s^{i+k} r^{(-1)^k j},
\end{equation*}
we have that
\begin{align*}
(s^i r^j) \triangleright s^k = s^{i+k},&& \text{ and }&
&\kappa(s^ir^j,s^k) = r^{(-1)^kj}.
\end{align*}
Hence, the action, on the set label is
\begin{equation*}
s^i r^j (s^k,r^l) = (s^{i+k},r^l)
\end{equation*}
It follows that the number of orbits in $(\mathcal{R}\times H)^n_e$ is
\begin{equation*}
2^{n-1}\times k^{n-1} = |G|^{n-1}.
\end{equation*}
Since $\gamma = 1$ all orbits are regular and then $\dim(\operatorname{Hom}_{\mathcal{Z}(\operatorname{Vec}_G)}(\mathbb{C},L(H,1)^{\otimes n}))=|G|^{n-1}$.
|
\subsection{Usage Scenario}
We now describe how an analyst uses \thetool to explore a dataset
containing wellbeing indicators for the OECD member countries. This dataset
contains 25 distinct attributes (indicators) about 35 countries and is included
in our demo as an illustration and for ease of comprehension. \thetool
is intended to facilitate interactive exploration of datasets with data items
of the order of 100K and attributes that number in the hundreds.
The analyst loads the OECD dataset in Foresight and eyeballs various insights
displayed in the carousels corresponding to each insight class (Figure
\ref{fig:insights}). She notes instantly that the indicators {\tt
Working Long Hours} and {\tt Time Devoted To Leisure} have a strong negative
correlation, since this is one of the top-ranked correlation insights
recommended by Foresight. Encouraged by this quick discovery, she brings
this insight \textit{into focus} by clicking on it. \thetool updates its
recommendations by choosing a subset of insights within the neighborhood of the
focused insight. The analyst explores the newly recommended correlations
through multiple ranking metrics such as Pearson correlation coefficient and
Spearman rank correlation and is surprised to learn that {\tt Time Devoted To
Leisure} has no correlation with {\tt Self Reported Health}.
Intrigued with this lack of correlation, she checks the univariate
distributional insight classes. The recommendations within these classes,
which have already been updated based on the previous selections, show that
{\tt Time Devoted To Leisure} has a Normal distribution while {\tt Self Reported
Health} has a left-skewed distribution. Having gained greater familiarity
with the OECD dataset, our analyst wonders about the factors that affect
{\tt Self Reported Health}. She clicks on the distribution of \texttt{Self
Reported Health}, adding this as one of the focal insights. Foresight
recommends a new set of correlated attributes and she finds that {\tt
Life Satisfaction} and {\tt Self Reported Health} are highly correlated.
Satisfied with her preliminary discoveries (and armed with deeper questions
about OECD countries than before), our analyst saves the current
\thetool state to revisit later and to share with her colleagues.
\balance
\subsection{Demo Datasets}
Our demonstration will feature the following two datasets in addition to the
OECD dataset described above.
\noindent\textbf{Parkinson:} Parkinson's Disease (PD) is a progressive
neurodegenerative disorder affecting nearly a million people in
the US alone. Our second use case applies \thetool to gain insights
into a dataset of PD patients with measured clinical descriptors
characterizing the disease progression~\cite{goetz2008movement}. The dataset
has 2K rows and 50 columns and is collected under the Parkinson's
Progression Markers Initiative (PPMI).
\noindent\textbf{IMBD:} Our third use case explores a dataset with 5000
movies (rows) and 28 features (columns). The features range from
the director name to the IMBD score for each movie. Questions that
Foresight users will be able to explore are: What factors correlate
highly with a film's profitability? How are critical responses and
commercial success interrelated?
\subsection{Querying Insight Space}
The input data to \thetool is a matrix $A_{n \times d}$, where
each row represents one of $n$ \emph{data items} and each
column represents one of the $d$ \emph{attributes} of an item.
In this work, we assume the data has been pre-cleaned. In
general, the insights provided by \thetool might reveal
additional, more subtle data problems that require further
cleaning, e.g., a strong correlation that makes no real-world
sense.
We define an \emph{insight} as a strong
manifestation of a distributional property of the data, such as
strong correlation, tight clustering, low dispersion, and so
on. We focus throughout on insights involving the marginal
distribution of one, two, or three attributes (Figure \ref{fig:insights}).
We require that each insight have one or more associated
\emph{insight metrics} that can be used to rank $n$-tuples of
attributes based on the strength of the property that
defines the insight. Similarly, each insight must have one or
more associated \emph{data visualizations}. Corresponding to
each insight is an \emph{insight class} that comprises all
feature tuples whose joint distributions are compatible with
the insight's associated metrics and visualizations. For
example, given a data set with attributes $a_1,a_2,\ldots,a_n$,
the insight class corresponding to the insight $\mathcal{I}=$
``high linear correlation" would contain all pairs $(a_i,a_j)$
with $i<j$ such that $a_i$ and $a_j$ are both real-valued
attributes. Finally, an insight may optionally have one or more
associated \emph{overview visualizations} that display the
values of the insight metric over all tuples in the insight
class. The global visualization for the insight~$\mathcal{I}$
above, for example, is a heat map where the $x$ and $y$
coordinates correspond to the different attribute indices and
the color and size of the circle centered at $(x,y)$ encode
the Pearson correlation coefficient (Figure \ref{fig:overview}).
A basic \emph{insight query} returns the visualizations for the
highest-ranked feature tuples according to the insight
metric selected, e.g., the attribute pairs with the highest
correlations. In general, one or more of the attributes might
be fixed, e.g., instead of ranking the highest correlations
over all $(x,y)$ attribute pairs, we can fix $x=\bar{x}$ and
rank correlations only over pairs of the form $(\bar{x},y)$,
i.e., searching for the attributes most correlated with $\bar{x}$.
Insight queries may also have constraints or filters on the
strength metric, e.g., we might want to rank only correlated
attribute pairs whose correlation coefficient falls in the
range $[0.5,0.8]$ because we want to filter out trivially very
high correlations. In future work, queries will also allow
inclusion of constraints involving metadata about attributes,
e.g., to search for attributes that represent currency or
dates.
As can be seen, our framework imposes some structure on the
space of insights that can be exploited during search. Two
insights can be considered ``similar" if their metric scores
are similar or if the sets of fixed attributes are similar. At
any point during the EDA process, the user can step back and
look at the overview visualization of an insight (Figure~\ref{fig:overview}). This helps
ensure that, in analogy with gradient descent, the EDA process
does not get inadvertently ``trapped" in some local
``neighborhood" of attribute tuples. This capability is
particularly important in cases where many attribute tuples
have similarly high insight-metric scores, so that the
particular set visualized for the user is somewhat arbitrary.
Section~\ref{sec:demo} illustrates the insight-navigation
process in a concrete scenario.
\subsection{\thetool's Insight Classes}
Foresight is designed to be an extensible system where a data scientist
can ``plug in" new insight classes along with their corresponding
ranking measures and visualizations. We now briefly describe some of the specific insights supported by
\thetool. Denote by $\mathcal{B}$ and $\mathcal{C}$ the sets of
attribute columns in $A$ that contain numeric and categorical
values. \thetool supports a variety of distinct visual insights, each
with a preferred ranking metric and visualization method. Denote by
$b=(b_1,\ldots,b_n)^\top\in\mathcal{B}$ a numeric column with
mean $\mu_b$ and standard deviation $\sigma_b$, and by
$c=(c_1,\ldots,c_n)^\top\in\mathcal{C}$ a categorical column.
For each insight, the \textit{ranking metric} is italicized.
\noindent\textbf{1. Dispersion:} Very high or low dispersion of
data values around a population mean is measured by the
\textit{variance} $\sigma^2(b) = n^{-1}\sum_i^n (b_i - \mu_b)$
and is visualized via a histogram.
\noindent\textbf{2. Skew:} Skewness is a measure of asymmetry
in a univariate distribution. It is measured by the
\textit{standardized skewness coefficient} $\gamma_1(b) =
n^{-1}\sum_i^n (b_i - \mu_b)^3/\sigma_b^3$ and visualized via a
histogram.
\noindent\textbf{3. Heavy Tails:} Heavy-tailedness is the
propensity of a distribution towards extreme values. It is
measured by \textit{kurtosis} $\text{Kurt}(b) =n^{-1} \sum_i^n
(b_i - \mu_b)^4/\sigma_b^4$ and visualized via a histogram.
\noindent\textbf{4. Outliers:} The presence and significance of
extreme outliers is measured by applying a user-configurable
outlier-detection algorithm---see, e.g.,
\cite{Aggarwal13}---and computing the \textit{average
standardized distance} of the outliers from the mean, where
standardized distance is measured in standard deviations.
Outliers are visualized using box-and-whisker plots.
\noindent\textbf{5. Heterogeneous Frequencies:} For a
categorical column $c$ (or a discrete numerical column $b$),
high heterogeneity in frequencies implies that a few values
(``heavy hitters'') are highly frequent while others are not.
For a configurable parameter~$k$, heterogeneity strength is
measured by $\text{RelFreq}(k,c)$, the \textit{total relative
frequency of the~$k$ most frequent elements in $c$}. This
insight is visualized via a Pareto chart.
\noindent\textbf{6. Linear Relationship:} The strength of a
linear relationship between two columns $x, y \in \mathcal{B}$
is measured using the magnitude of the \textit{Pearson correlation coefficient} $|\rho(x,y)|$, where
$\rho(x, y) = \sum_{i=1}^n
(x_i-\mu_x)(y_i-\mu_y)/(\sigma_x\sigma_y)$ and visualized via a
scatter plot with the best-fit line superimposed.
\noindent\textbf{Additional Insights:} Other insights include multimodality, nonlinear monotonic relationships, general statistical dependencies, and segmentation. Details are suppressed due to lack of space.
\section{Introduction}
\label{sec:intro}
\input{intro}
\section{Insights}
\label{sec:model}
\input{model}
\section{Sketching}
\label{sec:sketch}
\input{sketching}
\section{Demonstration}
\label{sec:demo}
\input{demo}
\section{Conclusion}
\label{sec:conclusions}
\input{conclusions}
\bibliographystyle{abbrv}
|
\section{Introduction}
This paper belongs to a new direction in dynamical systems theory, which originated in the theory of open dynamical systems \cite{PY,DY}. A standard set up in open systems theory includes a "hole" $H$ which is a positive measure subset of the phase space $M$ of some dynamical system generated by a map $T$ preserving a Borel probability measure $\mu$.
When a trajectory hits the hole $H$ it escapes and is not considered any more. In this setting, it is natural to assume that the map $T$ is ergodic. Otherwise, instead of a given hole one should consider its intersections with ergodic components and take each ergodic component as the phase space of an open system. In what follows we always assume that $T$ is ergodic, and therefore almost all trajectories will eventually escape. Let $P_s(n)$ be the probability that a trajectory does not escape until time $n$ (computed with respect to $\mu$), which is called the survival probability. It is natural to ask what the decay rate of $P_s(n)$ is. This decay rate $\rho(H) = \lim_{n\rightarrow\infty}-\frac{1}{n}\log P_s(n)$ is called the escape rate.
Traditionally, the theory of open dynamical systems dealt with small holes \cite{PY,DY}. Therefore such open systems could be (and were) treated as small perturbations of the corresponding closed systems \cite{DY,KL}. In fact, the first paper about open systems \cite{PY} limited its scope to the dynamics (of billiards) with a \textit{small} hole (in the billiard table). Hence the interest was always related to the limit when the size of the hole tends to zero \cite{PY,DY}, besides some special examples when everything is easily computable (some examples can be found in the beautiful review \cite{DY}).
At the same time, a seemingly natural question on how the process of escape depends on the position of the hole in phase space had been overlooked. The second author raised this question inspired by remarkable experiments with atomic billiards \cite{FKCD,MHCR}. Moreover, this question addresses finite, rather than infinitesimal, holes. Indeed, in real systems "holes" are finite, and it is a challenge to the modern theory of dynamical systems to handle this situation.
Originally this question was formulated as follows: "How does escape rate depend on the position of the hole?" \cite{BY}. So it referred to the escape rate $\rho(H)$. Observe that the definition of the escape rate involves a limit as time goes to infinity. Clearly this question makes sense if the measure $\mu$ is invariant, ergodic, and absolutely continuous. Indeed, if e.g. $\mu$ is sitting on some subset (hole) then the escape rate through this hole is infinite, while it could assume various finite values for subsets not belonging to this hole. (In fact, it was shown that in general the escape rate can even behave locally as a devil staircase \cite{DW}).
A standard and natural approach to attacking a new type of a problem or question is to consider a class of systems for which an answer seems possible. Such class of dynamical systems with the strongest chaotic properties was studied in \cite{BY,B}. These dynamical systems are ergodic with respect to an absolutely continuous measure $\mu$ and have such finite Markov partition that the corresponding symbolic representation is a full Bernoulli shift. Moreover, each element in this partition has the same measure, and therefore all entries of the transition probabilities matrix are also equal each other.(In what follows such Markov partition will be called a basic Markov partition).
Hence the measure of any element of the basic Markov partition and all transition probabilities are equal to $1/q$, where $q$ is a number of elements in the Markov partition $\xi$. Therefore the evolution of such dynamical systems is equivalent to the evolution of independent identically distributed (IID) random variables. All values of such random variables have the same probability. A typical example of random trials generating such IIDs is the throwing of a fair dice (with $q$ faces). Therefore this class of systems was called in \cite{B} fair-dice-like (FDL) systems.
The FDL-systems form a narrow subclass with the most uniform hyperbolicity in the class of chaotic (hyperbolic) dynamical systems \cite{BP}.
One of the simplest examples of a FDL-system is the doubling map $x \rightarrow 2x$(mod1) of the unit interval. The Lebesgue measure (length) is invariant. Consider the Markov partition into intervals $(0,1/2)$ and $(1/2,1)$. Then $q=2$ and the measure of each interval and all four transition probabilities equal 1/2.
It is a standard approach in the theory of chaotic dynamical systems to pose questions about what happens in the limit when time goes to infinity, or after averaging some observables over an infinite time interval. The evolution of strongly chaotic dynamical systems is in many respects similar to the evolution of random (stochastic) processes. Therefore in the metric (ergodic) theory of dynamical systems, the main problems are about mixing (i.e. vanishing of correlations in the limit when time goes to infinity), a rate of mixing (correlation decay), and the central limit theorem (CLT) and other limit theorems which again involve a limit when time goes to infinity.
Likewise, all major characteristics of chaotic dynamical systems involve either taking a limit when time goes to infinity or averaging over an infinite time interval. Indeed, look at the definitions of Lyapunov exponents and various entropies, among others. Observe that the definition of the escape rate also involves a limit when time goes to infinity. Analogously in Nonequilibriun Statistical Mechanics, definitions of transport coefficients involve taking a limit when time goes to infinity and averaging over an infinite time interval. Therefore the main result of \cite{BY} came as a complete surprise. Namely, it was proven that for some subclass of FDL systems not only are the escape rates generally different for different subsets (considered to be holes), but also the relations between corresponding survival probabilities can be established for \textit{all moments of time}. Namely, either all the survival probabilities for two holes of the same measure are equal or they coincide only on a short time interval after which all survival probabilities for one hole exceed the ones for another. Therefore there exists a finite moment in time when the process of escape through one hole becomes more intense than for another hole (and this moment is exactly and easily computable for FDL systems).
This result is of an essentially different nature than the ones we are used to in dynamics. Indeed, it deals with finite times rather than with the infinite time limit. Why are some specific finite moments in time important for dynamics (more precisely in this case, for transport in the phase space)? These FDL systems are the most uniformly hyperbolic (chaotic). Why is their dynamics not uniform? This result was generalized to the entire class of FDL systems in \cite{B}. The paper \cite{BB} contains generalizations for Markov chains. Topological analogs of these results were proved in \cite{AB} where the focus was on applications to computer simulations of real systems and networks. In \cite{AB} long time \textit{estimates} were obtained for survival and first hitting probabilities with respect to (not necessarily invariant) Lebesgue measure, which is typically used as an initial distribution in numerical experiments.
The main result of \cite{BY} gives hope that it might be possible to develop a rigorous theory regarding the finite time evolution of strongly chaotic dynamical systems. The first step was to realize that one can use the ideas developed in open systems theory to study transport in the phase space of closed dynamical systems. Indeed, one can make different holes in the phase space of a closed dynamical systems and "look" through these holes at the dynamics of a closed system (like one looks inside through windows). This really represents a turnabout, because here we use open systems built from a closed system to study the dynamics of the original closed system. This is totally different from a standard approach in the theory of open systems which does the opposite.
The idea of "spying" on closed systems by making holes (windows) proved to be efficient and allowed to obtain various new formulas and results useful for applications (see e.g. \cite{BD,D}). It is notable (although natural) that the main impact in a number of follow up papers was made so far not by the main result of \cite{BY} but by one of its byproducts dealing with infinitesimally small holes. Consider a sequence of shrinking holes converging to some point in the phase space. Again one can place this sequence in neighborhoods of different points pursuing an answer to the same question regarding how such placement impacts the escape rate. This was another non-standard question raised in \cite{BY}. Indeed, the escape rate through a point obviously equals zero because a measure in the open systems theory is always assumed to be absolutely continuous. It was shown in \cite{BY} that by normalizing the escape rates of holes which shrink to a point by measures of these holes, one gets a limiting value. Moreover this value varies over different points. Therefore even local escape depends upon the position (and other dynamical characteristics) of the point in phase space. This result about small holes (as well as the results for large holes) was presented at the workshop in the Boltzmann institute in the summer 2008. Immediately \cite{KL,FP} it was generalized to much larger classes by leading experts in open dynamical systems. Now it is an active area because relevant perturbation techniques were already well developed.
But what to do about this new, unexpected, and strange result on large holes? A natural approach would be to generalize the main result of \cite{BY,B} on survival probabilities to a larger class of chaotic hyperbolic systems. It is always the case that when something new is found for a narrow class of systems, the results are generalized for larger and larger classes. However, the main result of \cite{BY} gave hope that something more ambitious would be possible, namely finite time predictions for strongly chaotic dynamical systems. Observe that this main result \cite{BY}, although giving some exact values in time when the survival probabilities for different subsets of the phase space split, does not allow for finite time predictions for evolution of a system. Indeed, by comparing two subsets of the phase space of a FDL system we can only say that it is more probable (over the \textit{infinite} time interval after a certain moment) that trajectories would enter one subset compared to another. So this result does not allow us to make finite time predictions. Therefore leading experts in open systems (as well as other mathematicians) did not move into this new area of research because it was not clear what to do next.
It is a main goal of our paper to present the first rigorous results in the mathematical theory of the finite time dynamics (FTD) of (strongly) chaotic systems. It is the next needed step in this new area. In fact these new rigorous results allow one to make finite time predictions for transport in the phase space of fair-dice-like dynamical systems. Actually, predictions can even be made for the next moment of time, i.e. for an "immediate future". (Some results of the present paper were announced in \cite{BV}. Here proofs are given of those claims as well as of several other statements).
Recall that the $m$th order refinement of a partition $\xi$ is the partition generated by intersection of all preimages of $\xi$ from orders $0$ to $m-1$, i.e the partition generated by the sets $C_\xi^{(j_0)} \cap T^{-1} C_\xi^{(j_1)} \cap \dots \cap T^{-m+1}C_\xi^{(j_{m-1})}$ where $C_\xi^{(k)}$ is some element of the partition $\xi$. It is a well known fact that any refinement of a Markov partition is also a Markov partition.
Our main result says the following. Let $A$ and $B$ be elements of some refinement of a basic Markov partition of a FDL system. Then either the infinite sequences of their first hitting (first passage) probabilities $P_h(A,n)$ and $P_h(B,n)$ coincide, or the entire infinite interval of (positive) times gets partitioned into three subintervals. In the first very short interval the first hitting probabilities coincide. Then in the second interval (of finite length) the first hitting probabilities for $A$ exceed those for $B$. In the third (infinite) interval the opposite inequalities hold.
If $A$ and $B$ are elements of different refinements of the Markov partition of a FDL system then the first (short) time interval disappears because $A$ and $B$ have different measures and only the second and the third intervals remain where there are hierarchies.
To understand the following results better, imagine that for each element we construct a piece-wise linear curve connecting the values of the corresponding first hitting probabilities $P_h(A,n)$ and $P_h(A, n+1)$ for all $n>0$. It follows from ergodicity that $\sum _{n=1}^\infty P_h(H,n)$=1 for any set $H$ of positive measure. Therefore if the first hitting probability curves for two subsets do not coincide then they must intersect. The main result establishes that (besides possibly a very short initial interval of coincidence) \textit{there is only one point of intersection of these curves}.
These results are much stronger than those found in \cite{BY,B}, which can be easily deduced from the results of the present paper. First of all, the following formula holds
\begin{align*}
P_s(A,n) = \sum_{m=n+1}^\infty P_h(A,m).
\end{align*}
Therefore clearly the results on comparison of elements (first hitting probabilities) of infinite series obtained in this paper are stronger than the results on comparison of the sums of such series (survival probabilities) \cite{BY,B}.
Consider now all elements of some refinement of the basic Markov partition. (All these elements have of course the same measures.) Then the main result of the present paper implies that the evolution of any FDL system consists of three stages. At the first stage, which we refer to as the short times interval, there is a hierarchy of the first hitting probabilities for different elements. At the second stage all these curves intersect. After the last such pairwise intersection, the third stage emerges which occupies what we call the long times interval, having infinite length. The intermediate interval between the first and the last intersections of the first hitting probability curves we refer to as the intermediate times interval. Observe that a standard approach in dynamical systems theory would consider only the infinite time limit whereas this partition into three time intervals is something completely new.
A crucial question about the possible practical applications of our results is what happens to the lengths of the finite short times interval and of the intermediate interval when we consider higher order refinements of the basic Markov partition. Practically speaking, it means that we analyze transport in the phase space at finer and finer scales. We prove that the length of the short times interval increases at least linearly with the order of refinement of the Markov partition. In fact numerical experiments, which we also present below, show that this growth is actually exponential. However, a principal fact is that the length of the short times interval tends to infinity when we consider transport in the phase space at finer and finer scales. Indeed, observe that the hierarchy of the first hitting probabilities in the short times interval is \textit{opposite} to the hierarchy in the (infinite) long times interval. Therefore the traditional approach to the studies of transport in the phase space of chaotic systems, which is based on time-asymptotic analysis, seems to be not quite appropriate for practical use. Indeed any analysis (via experiments and observations) of real systems lasts only a finite time. Therefore it essentially belongs to our short times interval where the dynamics/transport has quite different characteristics than in the infinite long times interval. Hence the strategy for analyzing experimental data should perhaps be reconsidered.
It is worthwhile to mention that numerical experiments with dispersing billiards confirmed the existence of different stages in the transport of chaotic systems \cite{BV}. It should be noted however that these numerical computations can not confirm that the corresponding curves of the first hitting probabilities have only one (or even a finite number) of intersections. In fact we believe that there are more intersections than just one for billiards studied numerically in \cite{BV}. Nevertheless, these numerical simulations suggest that there are rather long intervals with alternating hierarchies of the first hitting probabilities curves.
As a byproduct, our results allow one to determine the best base for towers (see definition in the next section) built for an FDL system. When used as the base for a tower, a choice of any element from a Markov partition of a FDL system ensures exponential decay of the first recurrence probabilities to this base. Our results allow one to chose base(s) with the fastest decay of the first recurrence probabilities. It gives hope that the theory of dynamical systems will be able to be developed to such stage when it would be possible to find numerical estimates of various exponential rates of convergence rather than dealing only with qualitative statements like that a certain rate is exponential.
It is naive to expect that the most broad and important class of nonuniformly hyperbolic dynamical systems \cite{BP} will have the same properties as the FDL-systems. However, numerical experiments with dispersing billiards \cite{BV} demonstrate that there exist time intervals of finite lengths with hierarchies somewhat similar to the ones in the FDL-systems. Surely, one should expect that for general nonuniformly hyperbolic systems \cite{BP} there will be more intervals with alternating hierarchies of the first hitting probabilities than for FDL systems. Actually, this was shown by some numerical results in \cite{BV}. To understand better what is going on, it is necessary to analyze some class of hyperbolic dynamical systems with distortion, i.e. with less uniform hyperbolicity than in the FDL-systems.
We also present in this paper an algorithm which allows accelerate escape from the phase spaces of strongly chaotic dynamical systems. This algorithm readily follows from the results of the present paper. This algorithm can be applied to real systems, particularly to atomic billiards \cite{FKCD,MHCR}. In a nutshell, it says the following: make a hole in a certain (optimal!) subset of the phase space and keep it open till a certain moment of time when this subset ceases to be optimal. Then close ("patch") this hole, and make a new hole in another subset which has become an optimal sink at this moment of time. The process continues by switching to other holes as they become optimal.
The structure of the paper is as follows. In the next section we provide necessary definitions and formulate main results. We also present there an algorithm which allows accelerate the process of escape from a FDL-system. In section 3 we introduce some notations and present several preliminary results. Section 4 provides a proof of the main results when considering subsets of phase space having the same measure, under a technical assumption whose proof is relegated to the appendix. Also included in section 4 is a simple example demonstrating why just two time intervals with different hierarchies of the first hitting probabilities may exist. This main result is surprising and therefore such demonstration is helpful for understanding (and breaking up) long and formal proofs. Section 5 contains the proof of the main results for subsets of the phase space with unequal measures. In section 6 we briefly present a few results of computer simulations for the length of the short times interval. The last section 7 contains some concluding remarks.
\section{Definitions and Main Results}
Let $T: M \rightarrow M$ be a uniformly hyperbolic dynamical system preserving Borel probability measure $\mu$. The following definition \cite{B} singles out a class of dynamical systems analogous to the independent, identically distributed (IID) random variables with uniform invariant distributions on their (finite!) state spaces. Classical examples of such stochastic systems are fair coins and dices, hence the corresponding dynamical systems are called fair dice like (FDL) \cite{B}.
\begin{definition} \label{def FDL}
A uniformly hyperbolic dynamical system preserving Borel probability measure $\mu$ is called fair dice like or FDL if there exists a finite Markov partition $\xi$ of its phase space $M$ such that for any integers $m$ and $j_i$, $1 \leq j_i \leq q$ one has $\mu \left( C_\xi^{(j_0)} \cap T^{-1} C_\xi^{(j_1)} \cap \dots \cap T^{-m+1}C_\xi^{(j_{m-1})} \right) = \frac{1}{q^m}$ where $q$ is the number of elements in the partition $\xi$ and $C_\xi^{(j)}$ is element number $j$ of $\xi$.
\end{definition}
Therefore a FDL-system is a full Bernoulli shift with equal probabilities $\frac{1}{q}$ of states.
In what follows we will call a Markov partition in the definition of the FDL systems a basic Markov partition of the FDL-system under consideration. We will be interested in such partitions of the phase space of FDL systems which are refinements of the basic Markov partition $\xi$ featured in the definition of FDL systems. We will say that the kth order refinement of the partition $\xi$ is the partition generated by the intersection of all elements of the partitions $T^{-i}\xi$ where $i$ varies between 1 and $k-1$. It is easy to see that this refinement has $q^k$ elements coded by the words of the length $k$. Clearly any refinement of a Markov partition is also a Markov partition. Therefore in what follows we often refer to refinements of the basic Markov partition as to Markov partitions.
\begin{example}
Let $Tx=qx$ (mod 1) where $x \in M = [0,1]$ and $q \geq 2$ is an integer, with $\mu$ the Lebesgue measure. The corresponding basic Markov partition is the one into equal intervals $[\frac{i}{q},\frac{i+1}{q}]$, $i=0, 1, \dots, q-1$.
\end{example}
\begin{example}
Consider the tent map $Tx$ = $2x$ if $0<x<1/2$ and $Tx$ = $2(1-x)$ if $1/2<x<1$ of the unit interval with Lebesgue measure. To see that it is a FDL-system take the same Markov partition into the intervals (0,1/2) and (1/2,1) as for the doubling map in the introduction.
\end{example}
\begin{example}
Take now the von Neumann-Ulam map of the unit interval where $Tx=4x(1-x)$. This map preserves the measure $\mu$ with density $\frac{1}{\pi \sqrt{x(1-x)}}$. The von Neumann-Ulam map is metrically conjugate to the tent map via the transformation $y=\sin^2(\frac{\pi x}{2})$. To see that it is also FDL just take the same Markov partition as for the tent and doubling maps.
\end{example}
\begin{example}
To see that FDL-systems could be high-dimensional as well consider the baker's map of the unit square, where $(x,y)\rightarrow(2x, y/2)$ if $0<x<1/2$ or $(x,y)\rightarrow(2x-1, y/2 + 1/2)$ if $1/2<x<1$. This map preserves Lebesgue measure (area). To see that baker's map is a FDL-system just take the Markov partition of the unit square into the strips $0<x<1/2$ and $1/2<x<1$
.
\end{example}
Let $\Omega$ denote a finite alphabet of size $q \geq 2$. We will call any finite sequence composed of characters from the alphabet $\Omega$ a string or a word. For convenience both names will be used in what follows without ambiguity. For a fixed string $w = w_k \dots w_1$, $w_i \in \Omega$ let $a_w(n)$ denote the number of strings of length $n$ which do not contain $w$ as a substring of consecutive characters. The survival probability for a subset of phase space coded by the string $w$ is then $\hat{P}_w(n) = \frac{a_w(n)}{q^n}$.
Denote $h_w(n) = q a_w(n-1)-a_w(n)$ for $n \geq k$. It is easy to see that $h_w(n)$ equals the number of strings which contain the word $w$ as their last $k$ characters and do not have $w$ as a substring of $k$ consecutive characters in any other place. Therefore $\frac{h_w(n+k)}{q^{n+k}}$ is the first hitting probability $P_w(n)$ of the word $w$ at the moment $n$.
J. Conway suggested the notion of autocorrelation of strings (see \cite{GO}). Consider any finite alphabet and denote by $|w|$ the length of the word $w$. Let $|w|=k$. Then the autocorrelation cor$(w)$ of the string $w$ is a binary sequence $b_k b_{k-1} \dots b_1$ where $b_i=1$ if $w_j = w_{k-i+j}$ for $j=1, \dots, i$, that is, if there is an overlap of size $i$ between the word $w$ and its shift to the right on $k-i$ characters. For example, suppose that $w=10100101$ in a two symbols (characters) 0 and 1 alphabet . Then cor$(w)=10000101$.
We can compare (values of) autocorrelations by considering them as numbers written in base 2. For instance the sequence 101 becomes the number 5.
Observe that the autocorrelation of a word $w$ is completely defined by its internal periodicities. Indeed all digits which equal to one in cor$(w)$ are at the positions corresponding to these internal periods \cite{GO,BY}.
Let $k=|w|$, $k'=|w'|$, and denote $h_w(n)=h(n)$ and $h_{w'}(n) = h'(n)$. We define
\begin{align*}
s_w = \textrm{max}_{1 \leq j \leq k-1} \{ j: b_j=1 \}
\end{align*}
whenever this maximum exists and we let $s_w=0$ otherwise. We will always denote $s=s_w$ and $s'=s_{w'}$. In what follows we will generally denote any quantity or function that depends on $w'$ by a superscript $'$.
It follows from ergodicity that $\sum_{n=1}^\infty P_w(n) = 1 = \sum_{n=1}^\infty P_{w'}(n)$. Therefore if $P_w(m)-P_{w'}(m) < 0$ for at least one $m$, there must be at least one $n$ for which $P_w(n)-P_{w'}(n) > 0$. Theorems 2.1 and 2.2 establish a surprising fact that for the FDL-systems there is only one $n$ for which the quantity $P_w(n)-P_{w'}(n)$ changes from being negative or zero to positive.
\begin{theorem}
Consider an FDL-system. Let $w$ and $w'$ be words coding some elements of (possibly different) refinements of the basic Markov partition such that cor$(w)>\textrm{cor}(w')$. Then there exists an $N > k$ such that $h(n)-h'(n) \le 0$ for $n<N$, and $h(n)-h'(n)>0$ for $n>N$.
\end{theorem}
Observe that $2^{k-1} \leq \textrm{cor}(w) \leq 2^k-1$. Therefore the assumption cor$(w)>\textrm{cor}(w')$ implies $k \ge k'$. (We also note that Eriksson's conjecture \cite {E,M} in discrete mathematics is a simple corollary of Theorem 2.1).
One may naturally expect that two discrete curves of survival probabilities intersect infinitely many times unless they coincide. (One gets the simplest example of two words of the same length with identical curves of the first hitting probabilities when all zeros in the first word are substituted by ones and all ones by zeros).
\begin{proposition}
Consider a FDL-system.If two words have the same length and equal autocorrelations then all first hitting probabilities for subsets coded by these words are equal to each other at any moment of time.
\end{proposition}
A proof of this proposition immediately follows from the definition of autocorrelations of words. Indeed the first hitting probability for any subset $A$ of the phase space equals its measure until the moment of time equal the minimal period of all periodic orbits which intersect $A$. At this moment the first hitting probability decreases by jumping to a smaller value. Such jumps occur at any moments of time corresponding to periods of periodic orbits intersecting $A$. It immediately follows from the definition of autocorrelation of words that any two sets coded by words with equal autocorrelations intersect only with such periodic orbits which have the same periods. Moreover the corresponding jumps (decreases) of the first hitting probabilities which occur at the same moments are equal each other for the FDL-systems because all elements have the same measure.
Therefore the sequences of the first hitting probabilities for subsets of the phase space coded by the words with equal autocorrelations coincide.
Consider the points $(n,P_w(n))$ on the plane where $n>0$ are integers. We get the first hitting probabilities curve for $w$
by connecting a point $(n,P_w(n))$ to $(n+1,P_w(n+1)$ by the straight segments for all $n$.
The next theorem establishes that nonidentical first hitting probabilities curves intersect only once.
\begin{theorem} \label{prob version}
With $N$ as given in Theorem 2.1 and under the same conditions, there is an $N>k$ such that $P_w(n)-P_{w'}(n) \le 0$ for $n<N$, and $P_w(n) - P_{w'}(n) > 0$ for $n>N$.
\end{theorem}
According to Theorems 2.1 and 2.2, for two words with different lengths the corresponding first hitting probabilities curves intersect only at one point. This point divides the positive semi-line into a finite short times interval and an infinite long time interval. Before the moment of intersection it is less likely to hit the smaller subset of phase space (coded by the longer word) for the first time, and after the intersection it is less likely to hit the larger subset for the first time. For two elements of the same Markov partition (which have the same measure) there is also a short initial interval where the two corresponding first hitting probability curves coincide (unless these two curves coincide forever). The length of this initial interval does not exceed the (common) length of the code-words for elements of the Markov partition. After this interval there is a short times interval where it is more likely to visit one (say the first) element of the Markov partition for the first time than the other one. The last interval is an infinite long times interval where it is more likely at any moment to visit for the first time the other one (second) element.
Take now all elements of a Markov partition. They have equal measures because we are dealing with FDL systems. Then there is initial time interval of the length equal the (same) length of words coding elements of this refinement of a basic Markov partition. After the initial interval comes a finite interval of short times where there is hierarchy of the first hitting probability curves. Then comes intermediate interval where (all!) curves intersect. And finally there is infinite interval where there is a hierarchy of the first hitting probabilities curves which is opposite to the one in the short times interval. Therefore finite time predictions of dynamics are possible in the short times interval and in the last infinite long times interval.
\begin{figure}[htp]
\centering
\includegraphics[width=10cm]{example.png}
\caption{First hitting curves for the doubling map.}
\label{fig}
\end{figure}
Figure \ref{fig} illustrates the statement of Theorem \ref{prob version}. It depicts the first hitting probability curves for three different subsets within the domain of the doubling map. These three subsets are encoded by the words 1111, 1010, and 1000. Each subset has autocorrelation equal to the word that encodes it.
The next statement provides a lower bound on the length of the short times interval.
\begin{theorem}
Under the same conditions as Theorem 1 and with $N$ as defined there, if $k=k'$ and $s=s'$ then $N \ge 4k$. If $k=k'$ and $s>s'$ then $N \ge 3k-s$. If $k>k'$ then $N>k+1$.
\end{theorem}
Let a word $w$ correspond to a subset $A_w$ of some ergodic dynamical system. Because $\mu(A_w)>0$, almost all orbits return to the set $A_w$. Construct now a tower with base (zero floor) $A_w$. The $nth$ floor of a tower consists of all points of the set $T^{n}A_w$ which did not return to $A_w$ within the first $n$ iterations of $T$. Denote by $R_{A_w}(n)$ probability of the first return to $A_w$ at the moment $n$. Let $P_{A_w}(n)$ be the first hitting (first passage) probability corresponding to the measure $\mu$.
\begin{definition}
Consider an ergodic dynamical system and choose two subsets $A$ and $B$ of positive measure. We say that tower $Q_A$ with base $A$ is better than tower $Q_B$ with base $B$ if there exists $n^*$ such that $\sum_{n>n^*}R_A(n) < \sum_{n>n^*}R_B(n)$ for all $n>n^*$.\\ Let $\hat{\xi}$ be a refinement of the Markov partition $\xi$. We say that an element $C_{\hat{\xi}}$ if the partition $\hat{\xi}$ is an optimal base for a tower out of all elements of $\hat{\xi}$ if there is no tower better than $Q_{C_{\hat{\xi}}}$.
\end{definition}
It is well known that the first hitting probability $P_(A_w)(n)$ of $A_w$ at the moment $n$ equals
\begin{align*}
&P_{A_w}(n) = \sum_{m > n} R_{A_w}(m).
\end{align*}
For a given refinement of the basic Markov partition it is generally possible to have several optimal bases with equivalent towers built over them. In view of above relation between the first hitting and the first return probabilities the following statement about an optimal base for a tower is an immediate corollary of Theorems 2.1 and 2.2.
\begin{theorem}
Consider an FDL-system. Then for any refinement of a basic Markov partition $\xi$ there exists an optimal tower with base from this refinement such that no other of its elements yields a tower better than this one.
\end{theorem}
It is well known that for strongly chaotic (hyperbolic) dynamical systems periodic points are everywhere dense \cite{BP}. In particular it is true for FDL systems. Denote by $Per_{C_{\xi}}$ the minimal period out of all periodic orbits intersecting an element $C_{\xi}$ of some Markov partition $\xi$. A proof of Theorem 2.1 (see Section 4 and Appendix) implies the following lemma on periodic points and optimal towers.
\begin{lemma}
Consider an FDL-system. Let $\xi$ be any refinement of the basic Markov partition. An element $C_{\hat{\xi}}$ such that the tower $Q_{C_{\hat{\xi}}}$ is optimal must have the maximum value of $Per_C$ out of all other elements of this Markov partition.
\end{lemma}
Generally an optimal tower for a given Markov partition is not unique, i.e. several elements can serve as optimal bases.
We conclude this section with presenting an algorithm which allows to speed up escape from open dynamical systems built out of FDL-systems. Consider a FDL-system and some refinement of the basic Markov partition. (Observe that a choice of such refinement determines the scale/precision at which we want to analyze dynamical system under study). Then make first a hole in such element $A$ of this refinement which has a minimal autocorrelation. At the moment when the first hitting probability curve corresponding to $A$ gets intersected by another such curve corresponding to another element $B$ "patch" the hole $A$ and make a hole in $B$. Keep a hole in $B$ until its first hitting probabilities curve gets intersected by a curve corresponding to subset (element of refinement) $C$. Then "patch" the "hole" $B$ and make "hole" in $C$. Continue this process by switching to a new hole after each intersection. It immediately follows from Theorems 2.1 and 2.2 that this process requires only a finite number of switches and it ensures the fastest escape from FDL-systems (for holes of a given size defined by the order of refinement of the basic Markov partition).
\section{Some Results on Pattern Avoidance.}
We establish the convention that $b_0=1$ for every word $w$, ie. each word is augmented by the last symbol (digit) $1$. The purpose of this convention is to simplify statements like the following, which without this convention do not make sense when $(k-s) | k$, for example. By the definition of cor$(w)$
\begin{align} \label{cor per}
\begin{split}
&\textrm{if } b_j=1 \textrm{ for some } j \in \{0, 1, \dots, k-1 \},\\
&\textrm{then } b_{k-(k-j)t=1} \textrm{ for all } t \in \{1, \dots, \lfloor \frac{k}{k-j} \rfloor\},
\end{split}
\end{align}
(see (\cite{M})).
Given $i$ such that $b_i=1$, let $[i] = \textrm{max } \{j: b_j=1 \textrm{ and } i=k-t(k-j) \textrm{ for some } 1 \le t \le \lfloor \frac{k}{k-j} \rfloor \}$. In light of the relation (\ref{cor per}), it is natural to define the set $I = \{[i]: b_i=1\}$.
We will need to distinguish a few other digits of the autocorrelation in addition to $s$. Let
\begin{align*}
&d = \textrm{min}_{j \in I} \{ j: b_j=1 \}\\
&r = \textrm{max}_{j \in I} \{ j: b_j=1 \textrm{ and } b'_j=0 \}
\end{align*}
whenever they exist.
The largest member of $I$ is always $s$. An effect of Propositions \ref{p1} and \ref{p2} below is that $I \subset \{1, \dots, k-s-2\} \cup \{s\}$. A further consequence of Proposition \ref{p1} is that the only member $i$ of $I$ for which $|\{j: [j]=i\}| > 1$ is $s$, hence we define $S = \{i: [i]=s \}$.
Let $H_w(n)$ be the number of strings which end with $w$, begin with $w$, and which do not contain $w$ as a substring of $k$ consecutive characters in any other place. For $n > k$ it is easy to see that $H(n)=qh(n-1)-h(n)$. The probability of first returning to the "hole" given by $w$ is $\frac{H_w(n)}{q^n}$.
While $H_w(n)>0$ for $n>2k$, $H_w(2k)=0$ if and only if there is an $i$ for which $b_i=b_{k-i}=1$. It is easy to see that the condition $b_i=b_{k-i}=1$ implies $b_{k-s}=1$, and this in turn can be used to prove that $(k-s) | k$. We can thus evaluate $H_w(n)$ for $n \leq 2k$ as follows.
\begin{align} \label{H vals}
&H_w(n) = \begin{cases} 0 & \textrm{ if } n<k,\\ -1 & \textrm{ if } n=k,\\ 1 & \textrm{ if } n=2k-i \textrm{ for some } i \in I,\\0 & \textrm{ otherwise if } n<2k\\ 0 & \textrm{ if } n=2k \textrm{ and } (k-s) | k\\ 1 & \textrm{ if } n=2k \textrm{ and } (k-s) \nmid k. \end{cases}
\end{align}
It was proved in \cite{M} that
\begin{align} \label{h ineq}
h_w(n) \geq \begin{cases} (q-1) \sum_{t=1}^{k-s} h_w(n-t) & \textrm{if } 0 < s < k,\\ (q-1) \sum_{t=1}^{k-1} h_w(n-1) & \textrm{if } s=0, \end{cases}
\end{align}
and
\begin{align} \label{h rec}
h_w(n) = q h_w(n-1)-h_w(n-k)+\sum_{t=1}^{k-1} b_t H_w(n-k+t).
\end{align}
The latter formula is derived from the following relation \cite{GO}.
\begin{align} \label{original h}
h_w(n) = \sum_{t=1}^k b_i H_w(n+t).
\end{align}
It is easy to see that $H(n) \le h(n-k)$ for $n>k$, and we will prove below that $(q-1) h(n-k-1) \le H(n)$ for $n > k+1$. This result is the content of Corollary \ref{stuff}.
\section{A Proof of Theorems 2.1 and 2.2 for the Case $\mathbf{k=k'}$.}
We prove in this section several technical results which will be used to deduce Theorems 1 and 2.The corresponding proofs are rather long and formal.It is often difficult to understand why really a statement is true although formally it is justified.What is actually a "mechanism" which ensures that a claim is correct? It is very important to have such idea especially for statements answering questions of a new type (rather than for the incremental ones). Therefore we start with presenting a simple example to demonstrate the process (mechanism) which ensures that the first hitting probability curves for FDL-systems have not more than one intersection.
Assume for simplicity that Markov partition in the definition of FDL-systems has just two elements labeled by the symbols 0 and 1. Consider now the second refinement of this Markov partition. This refinement is also a Markov partition with eight elements. Pick for instance the elements coded as (101) and (001). Clearly $cor(101)$=101=5> is greater than $cor(001)$=100=4. The reason for this inequality is that a periodic point with minimum period which belongs to the element coded by (101) has period two while a point with minimum period in the element coded by (001) has period three. Therefore a number of strings
of length $n$ which do not contain (101) will be larger than a number of strings which do not contain (001) for all $n>4$ \cite{GO,BY}. Increase now the length of all strings by one. Then a number of strings will double and some new strings appear which are ended by one or other of our two words. Such strings will be excluded from a future consideration because they contribute to the first hitting probability at this very moment for the corresponding word. Observe though that in case of the word (101) all such strings are ending by one and therefore more strings of the length $n+1$ will remain which have the last digit $0$.
The situation is the opposite for the word (001). Therefore more than half of the remaining strings of the length $n+1$ will generate the word (001) in two steps (i.e. among the strings of the length $n+3$). On another hand less than half of all survived string at the moment $n+1$ will generate strings of the length $n+3$ ending by (101). Such process will continue forever and therefore the first hitting probability curve for (101) after intersecting the one for (001) will always remain above it.
The words of arbitrarily long lengths may contain many internal periodicities. In fact, periodic orbits are everywhere dense in a phase space of any FDL-system. Therefore each element of any refinement of a Markov partition contains infinitely many periodic points.
This is the reason why the proofs for a general case become long and include consideration of many different cases.
We turn now to the formal proofs of Theorems 1-2. Observe at first that Theorem 2 is equivalent to the claim that an $N>k$ exists such that $h(n)-q^{k-k'}h'(n-k+k') \le 0$ for $n<N$ and $h(n)-q^{k-k'}h'(n-k+k') > 0$ for $n \ge N$.
For any $i \in I$ let $T_w(i)= \textrm{max} \{ t>0: w_{k-i} \dots w_{k-i-t+1} = w_{k-j} \dots w_{k-j-t+1} \textrm{ for some } j \in I, j>i \}$, with the convention that if the latter set is empty then $T_w(i)=0$. Again we will often denote $T(i)=T_w(i)$ when $w$ is fixed.
\begin{example}
Consider the word $HTHTHHHTHTH$ over the alphabet $\Omega = \{ H, T \}$. Then cor$(w)=10000010101$ and $I = \{ 5,3,1 \}$. In this example $k=11$ and $w_{k-5} \dots w_1 = HHTHTH$, $w_{k-3} \dots w_1 = THHHTHTH$, and $w_{k-1} \dots w_1 = THTHHHTHTH$. Then $T(5)=T(3)=0$ and $T(1)=2$ since the first two letters of $w_{k-3} \dots w_1$ agree with the first two letters of $w_{k-1} \dots w_1$. Note that for the word $HTHTHTHTHTH$, cor$(w)=10101010101$ and $I=\{9\}$. None of $5,3$ or $1$ are in $I$ because $5=11-3(k-s)$, $3=11-4(k-s)$, and $1=11-5(k-s)$ where $\lfloor \frac{k}{k-s} \rfloor = 5$.
\end{example}
\begin{proposition} \label{p1}
Let $i \in I-\{s\}$. Then $i+T(i) < k-s$.
\end{proposition}
\begin{proof}
For any $i$ such that $b_i=1$ and for any $t$ satisfying $k-i \le t \le k$ one has $w_t \dots w_{t-(k-i)+1} = w_{t-l(k-i)} \dots w_{t-(l+1)(k-i)+1}$ for $0 \le l \le \lfloor \frac{t}{k-i} \rfloor -1$. Further, if $e=t-\lfloor \frac{t}{k-i} \rfloor (k-i)$ and $e>0$ then $w_t \dots w_{t-e+1} = w_e \dots w_1$. This is a consequence of the structure of the correlation function as described by (\ref{cor per}). Therefore when $b_i=1$ we will say that $w$ contains a $k-i$ period.
Let $i \in I$. Suppose first that $T(i)>0$ and for a contradiction suppose that $i+T(i) \geq k-s$. Let $i' \in I$, $i'>i$ be such that $w_{k-i} \dots w_{k-i-T(i)+1} = w_{k-i'} \dots w_{k-i'-T(i)+1}$.
Since $b_{i'}=1$, $w_{i'} \dots w_1 = w_k \dots w_{k-i'+1}$ which implies $w_{k-(i'-i)} \dots w_{k-i'+1} = w_{i'-(i'-i)} \dots w_1 = w_k \dots w_{k-i+1}$. Since $b_i=1$, similarly one has
\begin{align} \label{000}
&w_{k-(i'-i)} \dots w_{k-i'+1} = w_k \dots w_{k-i+1}.
\end{align}
Since $i-(k-s) \geq -T(i)$ we have $w_{k-i'} \dots w_{k-i'+i-(k-s)+1} = w_{k-i} \dots w_{k-i+i-(k-s)+1} = w_{k-i} \dots w_{s+1}$. Therefore
\begin{align*}
&w_{k-(i'-i)} \dots w_{k-(i'-i)-(k-s)+1}=w_k \dots w_{k-i+1} w_{k-i'} \dots w_{k-(i'-i)-(k-s)+1}\\
&=w_k \dots w_{k-i+1} w_{k-i} \dots w_{s+1} = w_k \dots w_{s+1},
\end{align*}
where we have used (\ref{000}) in the first equality.
Since $w$ contains a $k-s$ period,
\begin{align} \label{k-s periodic}
\begin{split}
&w_{k-(i'-i)-l(k-s)} \dots w_{k-(i'-i)-(l+1)(k-s)+1}\\
&=w_{k-(i'-i)} \dots w_{k-(i'-i)-(k-s)+1}= w_k \dots w_{s+1}\\
&=w_{k-l(k-s)} \dots w_{k-(l+1)(k-s)+1}
\end{split}
\end{align}
for every $0 \leq l \leq \lfloor \frac{k-i}{k-s} \rfloor -1$. Further, for $e = k-(i'-i)-\lfloor \frac{k-(i'-i)}{k-s} \rfloor (k-s)$, if $e>0$ one has
\begin{align} \label{k-s periodic remainder}
\begin{split}
w_e \dots w_1 &= w_{k-(i'-i)} \dots w_{k-(i'-i)-e+1}\\
&=w_{k-\lfloor \frac{k-(i'-i)}{k-s} \rfloor (k-s)} \dots w_{i'-i+1}
\end{split}
\end{align}
Together, \ref{k-s periodic} and \ref{k-s periodic remainder} imply that $w_{k-(i'-i)} \dots w_1 = w_k \dots w_{i'-i+1}$, or that $b_{k-(i'-i)}=1$.\\
Our goal now is to show that there is some index $i^*$ such that $b_{i^*}=1$, $i^*>i$, and $i=k-l(k-i^*)$ for some $l>0$. Doing this would contradict the fact that $i \in I$. Let $d_0=i'-i$. We will construct a strictly decreasing sequence $\{ d_n \}_{n=0}^N$ of positive integers such that $b_{k-d_n}=1$, $d_n=k-l_nd_{n-1}-i$ where $l_n$ is the unique positive integer such that $k-l_n d_{n-1}>i>k-(l_n+1)d_{n-1}$, and $i^*=k-d_N$ has the desired property.
If there is some $t$ for which $k-td_0 = i'$ then $k-(t+1)d_0 = i$, and we may take $N=0$. Similarly $N=0$ if $k-td_0=i$ for some $t$. Otherwise, there exists $t$ for which $i' > k-td_0 > i$. Since $b_{k-d_0}=1$, the word $w$ contains a $d_0$ period. With $d_1 = k-td_0-i$ it is easy to see that $b_{k-d_1}=1$ (a more detailed exposition for general $n$ is below).
For $n>1$ suppose that $d_{n-1}$ is already defined. If $i \ne k-ld_{n-1}$ for some $l>0$ then $N = n-1$. In addition one cannot have $k-l_n d_{n-1} = k-l_{n-1}d_{n-2}$ as this implies $i-(k-l_nd_{n-1})=i-(k-l_{n-1}d_{n-2}) = d_{n-1}$ hence $k-(l_n-1)d_{n-1} = i$, and again $N = n-1$. Otherwise denote $\iota = k-l_{n-1}d_{n-2}$ and observe that there is some $l_n$ such that $\iota > k-l_n d_{n-1} > i$. Since $w$ contains a $d_{n-1}$ period, we will show that $b_{k-l_nd_{n-1} - i}=b_{k-d_n}=1$. Let $\delta=\iota-(k-l_nd_{n-1})$. Observe that with this notation, $d_{n-1} = \delta+d_n$.
For any $0 \le l < \lfloor \frac{k-d_n}{d_{n-1}} \rfloor$ one has
\begin{align*}
w_{k-d_n-ld_{n-1}} \dots w_{k-d_n-(l+1)d_{n-1}+1}&=w_{k-d_n-(l_n-1)d_{n-1}} \dots w_{k-d_n-l_nd_{n-1}+1}\\
&=w_{\iota} \dots w_{i+1}\\
&=w_k \dots w_{k-d_{n-1}}\\
&=w_{k-ld_{n-1}} \dots w_{k-(l+1)d_{n-1}}.
\end{align*}
In the first equality we have used the $d_{n-1}$ periodicity of $w$, in the second we have used the fact that $k-d_n-(l_n-1)d_{n-1} = \iota$, in the third that $b_\iota=1$, and again in the fourth the $d_{n-1}$ periodicity of $w$.
Let $e=k-d_n-\lfloor \frac{k-d_n}{d_{n-1}} \rfloor d_{n-1}$. If $e>0$ and $e<\delta$ then one has
\begin{align*}
w_e \dots w_1 &=w_{k-l_nd_{n-1}-d_n} \dots w_{k-l_nd_{n-1}-d_n-e+1}\\
&=w_{i} \dots w_{i-e+1}\\
&=w_k \dots w_{k-e+1}\\
&=w_{k-\lfloor \frac{k-d_n}{d_{n-1}} \rfloor d_{n-1}} \dots w_{d_n+1}.
\end{align*}
If $e>0$ and $e>\delta$ then
\begin{align*}
w_e \dots w_1 &= w_i \dots w_{i-\delta+1} w_{i-\delta} \dots w_{i-e+1}\\
&=w_k \dots w_{k-\delta+1} w_{k-\delta} \dots w_{k-e+1}\\
&=w_k \dots w_{k-e+1}\\
&=w_{k-\lfloor \frac{k-d_n}{d_{n-1}} \rfloor d_{n-1}} \dots w_{d_n+1}.
\end{align*}
One thus has $w_{k-d_n} \dots w_1 = w_k \dots w_{d_n+1}$ and $b_{k-d_n}=1$. Since $d_n < d_{n-1}$ as long as $[i] \ne k-d_{n-1}$, the sequence $\{d_n\}$ is strictly decreasing. Since $k-d_n$ is bounded below by 1, there must be some $n$ for which $[i]=k-d_n$, and we let $N=n$.\\
If $T(i)=0$, the proof is similar to what we have just done. Supposing $i+T(i) = i > k-s$, there is some $t$ for which $k-t(k-s) > i > k-(t+1)(k-s)$, otherwise $[i]=s$ and $i \notin I$. Since $w$ contains a $k-s$ period, one can show that $b_{k-t(k-s)-i}=1$. Again we can construct a strictly increasing sequence of integers $\{ i_n \}_{n=0}^N$ such that $b_{i_n}=1$ and $[i]=i_N$. We omit the proof due to its redundancy.
\end{proof}
\begin{corollary} \label{c0}
$\{i: b_i=1\} =\{i: [i]=s\} \cup \left( I-\{s\} \right)$.
\end{corollary}
\begin{proof}
For $i \in I-\{s\}$ one has $i<s$ and $i<k-s$, hence $i<\frac{k}{2}$. If $b_j=1$ and $j \ne k-t(k-s)$ for any $t$, then $j = k-l(k-i)$ for some $i \in I-\{s\}$ if and only if $l=1$ since $k-i>\frac{k}{2}$. Thus either $j \in \{i: [i]=s\}$ or $j=i$ for some $i \in I-\{s\}$.
\end{proof}
\begin{corollary} \label{stuff}
$H(n) \ge (q-1)h(n-k-1)$ for $n>k+1$.
\end{corollary}
\begin{proof}
Observe that (\ref{H vals}) implies $H(n) \ge 0 = (q-1)h(n-k-1)$ for $k+1 < n \le 2k$.
Let $2k < n < 3k-s$. For $1 \le i \le s$ one has $k+i < n-k+i < 2k-s+i \le 2k$, so by (\ref{H vals}) we have $H(n-k+i)=1$ if and only if $b_{3k-n-i}=1$ and $3k-n-i \in I$. If $3k-n-i \in I-\{s\}$ then $3k-n-i < k-s$ by Proposition 1 and hence $b_{3k-n-i}=0$ for $n \le 2k+s-i$. In particular, $b_{3k-n-i}=0$ when $n=2k+1$ and $i \in I-\{s\}$. Thus $\sum_{i=1}^s b_i H(k-i+1) \le 1$ and $\sum_{i=1}^s b_i H(n-k+i) \le n-2k$. Since $h(n)=q^{n-k}$ for $k \le n < k-s$ one has
\begin{align*}
H(n) &= h(n-k) - \sum_{i=1}^s b_i H(n-k+i)\\
&=qq^{n-2k-1} - \sum_{i=1}^s b_i H(n-k+i)\\
&= (q-1)q^{n-2k-1} + \left( q^{n-2k-1} - \sum_{i=1}^s b_i H(n-k+i) \right)\\
&\ge (q-1)q^{n-2k-1} + \left( q^{n-2k-1} - (n-2k) \right)\\
&\ge (q-1)q^{n-2k-1} = (q-1)h(n-k-1).
\end{align*}
For $n \ge 3k-s$ observe that $\sum_{i \in S} H(n-k+i) \le \sum_{i=1}^k b_i H(n-2k+s+i) = h(n-2k+s)$. Then one has
\begin{align*}
&H(n) = h(n-k) - \sum_{i=1}^s b_i H(n-k+i)\\
&=h(n-k) - \sum_{i \in S} H(n-k+i) - \sum_{i \in I-\{s\}} H(n-k+i)\\
&\ge h(n-k) - h(n-2k+s) - \sum_{i \in I-\{s\}} h(n-2k+i)\\
&\ge (q-1) \sum_{t=1}^{k-s} h(n-k-t) - h(n-2k+s) - \sum_{i=1}^{k-s-2} h(n-2k+s+i)\\
&\ge (q-1) h(n-k-1)
\end{align*}
where we have used Corollary \ref{c0}.
\end{proof}
\begin{proposition} \label{p2}
Suppose that $s \neq k-1$. Then either $b_{t(k-s)}=0$ for every $1 \leq t \leq \lfloor \frac{k}{k-s} \rfloor$ or $b_{t(k-s)-1}=0$ for every $1 \leq t \leq \lfloor \frac{k+1}{k-s} \rfloor$
\end{proposition}
\begin{proof}
We use the following two statements, the first of which is obvious from Proposition 1.
\begin{align} \label{Proposition -1}
&\textrm{If } b_{k-s}=1 \textrm{ then } [k-s]=s.
\end{align}
\begin{align} \label{Proposition -2}
\begin{split}
&\textrm{If } b_{k-s-1}=1 \textrm{ then } [k-s-1]=s.
\end{split}
\end{align}
We prove (\ref{Proposition -2}). Suppose $b_{k-s-1}=1$. If $b_{k-s}=1$ then $b_t=1$ for every $1 \leq t \leq k-s$ and so $w_t = w_k$ for every $1 \leq t \leq k-s$. Since $w$ contains a $k-s$ period $w=\underbrace{w_k*\dots*w_k}_{\textrm{k times}}$ and $[k-s-1]=s$.
If $b_{k-s}=0$ then either $k-s-1=s$ and the result follows, or there is some $t>0$ such that $k-t(k-s)>k-s$. Let $i=k-s-1$, $\iota=k-t(k-s)$, and $d = \iota-i$. Since $b_{\iota}=1$ and $b_i=1$ it is easy to see that $w_{\iota} \dots w_1$ contains a $d$ period. As a result $b_{\iota-l d}=1$ for every $1 \leq l \leq \lfloor \frac{\iota}{d} \rfloor$. Let $L$ be such that $\iota-Ld > k-(k+1)(k-s)>0$ and $\iota-(L+1)d \leq k-(k+1)(k-s)$. If $d \nmid (k-s)$ then $\iota-Ld - \big( k-(k+1)(k-s) \big) = d' < d$. Since $b_{\iota-Ld}=1$ and $b_{k-(k+1)(k-s)}=1$ it must be that $w_\iota \dots w_1$ contains a $d'$ period, and hence $b_{\iota-d'}=1$. Since $\iota-d'>\iota-d=k-s-1$, with $i'=[\iota-d']$ one has $[i'] \neq s$ and $[i']+T([i']) \geq k-s$, a contradiction to Proposition 1. It follows that $d | (k-s)$ which implies that $w$ itself contains a $d$ period. Then $b_{k-d}=1$ but since $d<k-s$ this contradicts the definition of $s$.
The following statement is a corollary of (\ref{Proposition -1}) and (\ref{Proposition -2}).
\begin{align*}
&\textrm{Suppose that } s \neq k-1.\textrm{ Then } I-\{s\} \subset \{1, \dots, k-s-2 \}.
\end{align*}
If $b_{t(k-s)}=1$ for some $t>1$ then by Proposition (\ref{p1}) it must be that $[t(k-s)]=s$ and hence there is some $l$ such that $k-l(k-s)=t(k-s)$, whence $(k-s) | k$. According to (\ref{cor per}) it must be that $b_{k-s}=1$ as well. Thus, if $b_{k-s}=0$ then $b_{t(k-s)}=0$ for every $t$.
If $b_{k-s}=1$ then $b_{k-s-1}=0$ as otherwise $s=k-1$. If $b_{t(k-s)-1}=1$ for some $t>1$ then $[t(k-s)-1]=s$ and $k=m(k-s)-1$. Since $b_{k-t(k-s)}=1$, one has $b_{(m-t)(k-s)-1}=1$ for every $1 \le t \le m-1$. With $t=m-1$ this implies in particular that $b_{k-s-1}=1$, a contradiction. Thus $b_{t(k-s)-1}=0$ for every $t>1$.
\end{proof}
\begin{lemma} \label{l1}
If $s>0$, $1 \leq l \leq k-1$, and $n \geq 2k+l$ then $H(n) \geq (q-1) \sum_{t=1}^l H(n-t)$.
\end{lemma}
\begin{proof}
Rearranging relation (\ref{original h}) we obtain
\begin{align} \label{H rearranged}
&H(n) = h(n-k)-\sum_{t=1}^{k-1} b_t H(n-k+t) = h(n-k)-\sum_{i=1}^s b_i H(n-k+i).
\end{align}
Using (\ref{H rearranged}) one has
\begin{align} \label{H k-s}
\begin{split}
&H(n) - (q-1) \sum_{t=1}^l H(n-t) =h(n-k) - (q-1) \sum_{t=1}^l h(n-k-t)\\
&-\sum_{i=1}^s b_i H(n-k+i) + (q-1) \sum_{t=1}^l \sum_{i=1}^s b_i H(n-k+i-t)=\\
&h(n-k-l) - \sum_{t=0}^{l-1} H(n-k-t) \\
&- \sum_{i=1}^s b_i \left( H(n-k+i) - (q-1) \sum_{t=1}^l H(n-k+i-t) \right).
\end{split}
\end{align}
For any $i \in I-\{d\}$ let $\tilde{i} = \textrm{max } \{\iota < i: b_\iota = 1\}$. Observe that $|i-\tilde{i}| \leq k-s$ since $b_{k-t(k-s)}=1$ always. It follows that $\tilde{i}=i-\tau$ for some $1 \leq \tau \leq k-s$. We have
\begin{align} \label{cancel}
\begin{split}
&\sum_{i=1}^s b_i \left( H(n-k+i) - (q-1) \sum_{t=1}^l H(n-k+i-t) \right) \leq\\
&\sum_{i \in I-\{d\}} \Big( H(n-k+i) - H(n-k+\tilde{i}) \Big) +H(n-k+d)-(q-1) \sum_{t=1}^l H(n-k+d-t)\\
&\le H(n-l) - \sum_{t=0}^{l-1} H(n-k-t).
\end{split}
\end{align}
where we have used the fact that $n \ge 2k+l$ to ensure that $n-k+i-t > k$ for every $i$, hence $H(n-k+i-t)>H(k)=-1$. Combining (\ref{cancel}) and (\ref{H k-s}) one has
\begin{align} \label{H reduced}
\begin{split}
&H(n) - (q-1) \sum_{t=1}^l H(n-t) \geq\\
&h(n-k-l) - \sum_{t=0}^{l-1} H(n-k-t) - H(n-l) + \sum_{t=0}^{l-1} H(n-k-t) \geq\\
&h(n-k-l) - H(n-l) \geq 0.
\end{split}
\end{align}
\end{proof}
Let $\Delta(n)=h(n)-h'(n)$.
\begin{corollary} \label{c1}
Suppose $w'$ is such that cor$(w) \geq \textrm{cor}(w')$, $k=k'$, and $n \ge 3k$. If $\Delta(n-t) \ge (q-1)\Delta(n-t-1)$ for $1 \leq t \leq k-1$ then $\Delta(n) \ge (q-1)\Delta(n-1)$.
\end{corollary}
\begin{proof}
There are three cases. In the first $1<r<k-1$, in the second $r=k-1$, and in the third $r=1$. In the first case, using (\ref{h rec}) one has
\begin{align} \label{D}
\begin{split}
&\Delta(n)=q \Delta(n-1)-\Delta(n-k)+\sum_{i=r+1}^{k-1} b_i [q \Delta(n-k+i-1)-\Delta(n-k+i) ]\\
&+\sum_{i=1}^r b_i H(n-k+i)-\sum_{i=1}^{r-1} b_i' H'(n-k+i)
\end{split}
\end{align}
Using the equality $H(n)-H'(n) = q\Delta(n-1)-\Delta(n)$ and applying Lemma 1 one has
\begin{align*}
&\sum_{i=1}^r b_i H(n-k+i)-\sum_{i=1}^{r-1} b'_i H'(n-k+i)\\
&\ge \sum_{i=1}^{r-1} H(n-k+i) - \sum_{i=1}^{r-1} H'(n-k+i)\\
&= \sum_{i=1}^{r-1} q \Delta(n-k+i-1) - \Delta(n-k+i)
\end{align*}
The expression (\ref{D}) is thus bounded below by
\begin{align*}
&q \Delta(n-1)-\Delta(n-k) + \sum_{i=r+1}^{k-1} b_i [ q \Delta(n-k+i-1)-\Delta(n-k+i)\\
&+\sum_{i=1}^{r-1} [q \Delta(n-k+i-1)-\Delta(n-k+i)].
\end{align*}
By using the inductive assumption it is easy to see that $\Delta(n-1) + \sum_{i=r+1}^{k-1} b_i[ g \Delta(n-k+i-1)-\Delta(n-k+i) ] \geq (q-1) \Delta(n-k+r)$ and $\Delta(n-k+r) + \sum_{i=1}^{r-1} [ q \Delta(n-k+i-1)-\Delta(n-k+i) ] \geq (q-1) \Delta(n-k)$. (For a more detailed explanation, see \cite{M}). Applying both bounds, we have $\Delta(n) \geq (q-1) \Delta(n-1) - \Delta(n-k)+(q-1) \Delta(n-k) \geq (q-1) \Delta(n-1)$.\\
Suppose now that $r=k-1$, hence $s=r=k-1$ and $b_i=1$ $\forall i$. We have
\begin{align} \label{r=k-1}
\begin{split}
&\Delta(n) = q \Delta(n-1) - \Delta(n-k) + H(n-1)+\sum_{t=1}^{k-2} H(n-k-i)-b'_i H'(n-k-i)\\
&\geq q \Delta(n-1) - \Delta(n-k) + H(n-1) + \sum_{t=1}^{k-2} [q \Delta(n-k+i-1)-\Delta(n-k+i)].
\end{split}
\end{align}
Noting that $H(n-1)>0$ since $n \geq 3k$, by subtracting $H'(n-1)$ from (\ref{r=k-1}) we obtain
\begin{align*}
\begin{split}
&\Delta(n) \geq (q-1) \Delta(n-1)-\Delta(n-k)\\
&+\Delta(n-1) + \sum_{t=1}^{k-2} [q \Delta(n-k+t-1)-\Delta(n-k+t)].
\end{split}
\end{align*}
Using $\Delta(n-1)>\Delta(n-2)$ as before we have that
\begin{align*}
&\Delta(n) > (q-1) \Delta(n-1)-\Delta(n-k)+(q-1) \Delta(n-k) \\
&\geq (q-1) \Delta(n-1).
\end{align*}
Finally suppose that $r=1$. Then
\begin{align*}
&\Delta(n) = q \Delta(n-1) - \Delta(n-k) + \sum_{t=2}^{k=1} b_t [q \Delta(n-k+t-1)-\Delta(n-k+t)] + H(n-k+1)\\
&\geq (q-1) \Delta(n-1) - \Delta(n-k) + (q-1) \Delta(n-k+1)\\
&\geq (q-1) \Delta(n-1).
\end{align*}
\end{proof}
\begin{corollary} \label{c2}
Let $n \ge 4k$. Then $\sum_{t=0}^{k-d-1} H(n-k-t) \ge \sum_{i \in I} \sum_{t=1}^{k-i} b_t H(n-2k+i+t)$
\end{corollary}
\begin{proof}
For $i \in I- \{s\}$, applying Lemma 1 we have
\begin{align*}
&\sum_{t=1}^{k-i} b_t H(n-2k+i+t) = \sum_{t=1}^s b_t H(n-2k+i+t)\\
&\le H(n-2k+i+s+1).
\end{align*}
Note that $I-\{s\} \subset \{1, \dots, k-s-2 \}$. If $b_{k-s-2}=1$ and $b_{k-s}=1$ or $b_{k-s-1}=1$ then $I=\{s\}$ and the statement of the lemma holds. It suffices to assume that either $b_{k-s-2}=0$ or $b_{k-s}=b_{k-s-1}=0$.
If $b_{k-s-2}=0$ then $s \ne k-1$ and for $i \in I$ one has $i+s+1 \le k-s-2$. Since it always true that one of $b_{k-s}$ or $b_{k-s-1}=0$ when $s \ne k-1$ (see Propostion 2) one must have
\begin{align*}
-H(n-k)-H(n-k-1)+\sum_{t=1}^{k-s} b_t H(n-2k+s+t) \le 0.
\end{align*}
Since $i+s+1 \le k-s-2$ for $i \in I-\{s\}$ we thus have
\begin{align*}
&-\sum_{t=0}^{k-d-1} H(n-k-t) + \sum_{i \in I} \sum_{t=1}^{k-i} b_t H(n-2k+i+t)\\
&\le -H(n-k)-H(n-k-1) + \sum_{t=1}^{k-s} b_t H(n-2k+s+t)\\
&\le 0.
\end{align*}
If both $b_{k-s}=0$ and $b_{k-s-1}=0$ then
\begin{align*}
-H(n-k)+\sum_{t=1}^{k-s} b_t H(n-2k+s+t) \le 0.
\end{align*}
For $i \in I$ one has $i+s+1 \le k-s-1$ and
\begin{align*}
&-\sum_{t=0}^{k-d-1} H(n-k-t) + \sum_{i \in I} \sum_{t=1}^{k-i} b_t H(n-2k+i+t)\\
&\le -H(n-k) + \sum_{t=1}^{k-s} b_t H(n-2k+s+t)\\
&\le 0.
\end{align*}
\end{proof}
\begin{corollary} \label{c3}
Let $n > 2k$. Then $H(n) \ge \sum_{i \in I} H(n-k+i)$.
\end{corollary}
\begin{proof}
Let $2k<n<3k-s$. For $i \ne s$ observe that $n-k+i < n-2k+s < k$, hence $H(n-k+i)=0$. It follows that
\begin{align*}
H(n)-\sum_{i \in I} H(n-k+i) = H(n)-H(n-k+s) \ge H(n)-1 \ge 0.
\end{align*}
Note that $|I| \le k-s$ and recall $s \le k-1$. For $n \ge 3k-s$ one has
\begin{align*}
H(n) - \sum_{i \in I} H(n-k+i) \ge H(n)-\sum_{i=1}^{k-s} H(n-i) \ge 0
\end{align*}
by application of Lemma 1.
\end{proof}
\begin{corollary} \label{c4}
Let $n \ge 2k$. Then $h(n) \ge \sum_{t=1}^{k-1} H(n+t)$.
\end{corollary}
\begin{proof}
Applying Lemma 1, one has
\begin{align*}
&h(n) = \sum_{t=1}^k b_t H(n+t) \ge \sum_{t \in S \cup \{k\}} H(n+t) \ge \sum_{t=1}^{k-1} H(n+t).
\end{align*}
\end{proof}
Let $S = \{ \iota : [\iota]=s \}$. If $\iota \in S-\{s\}$ then
\begin{align*}
&h(n-2k+\iota) - \sum_{j=1}^s b_j H(n-2k+\iota+j) - H\big(n-2k+s+\iota+(k-s)) = 0.
\end{align*}
It follows that
\begin{align*}
&\sum_{\iota \in S-\{s\}} \left( h(n-2k+\iota) - \sum_{j=1}^s b_j H(n-2k+\iota+j) \right)\\
&+h(n-2k+s) - \sum_{j=1}^s b_j H(n-2k+s+j)\\
&=\sum_{\iota \in S-\{s\}} \left( h(n-2k+\iota)-\sum_{j=1}^s b_j H(n-2k+\iota+j) - H\big(n-2k+s+\iota+(k-s) \big) \right)\\
&+h(n-2k+s)- \sum_{j=1}^{k-s} b_j H(n-2k+s+j).
\end{align*}
From this one has
\begin{align} \label{cancellation of non-prime periods}
\begin{split}
&\sum_{\iota=1}^{k-1} b_i H(n-k+\iota) = \sum_{\iota \in S} H(n-k+\iota) + \sum_{\iota \in I-\{s\}} H(n-k+\iota)=\\
&\sum_{\iota \in S} \left( h(n-2k+\iota) - \sum_{t=1}^{k-1} b_t H(n-2k+\iota+t) \right)+\\
&\sum_{\iota \in I-\{s\}} \left( h(n-2k+\iota) - \sum_{t=1}^{k-1} b_t H(n-2k+\iota+t) \right)=\\
&h(n-2k+s) - \sum_{t=1}^{k-s} H(n-2k+s+t) \\
&+\sum_{\iota \in I-\{s\}} \left( h(n-2k+\iota) - \sum_{t=1}^{k-\iota} b_t H(n-2k+\iota+t) \right).
\end{split}
\end{align}
\begin{lemma} \label{l2}
If $k=k'$ and $s \ne s'$ then $\Delta(n) \le 0$ for $n < 3k-s$. If $s = s'$ then $\Delta(n) \le 0$ for $n < 4k$.
\end{lemma}
The proof of Lemma \ref{l2} when $s=s'$ is divided into two parts which constitute the appendices below. We include the proof when $s \ne s'$ here.
\begin{proof}
We remark that $\Delta(n) \le 0$ for $n \le 2k$ and $\Delta(n) = 0$ for $n < 2k-r$ no matter the values of $s$ and $s'$.
Using (\ref{H rearranged}) one has
\begin{align} \label{delta H}
\begin{split}
H(n)-H'(n) &= \Delta(n-k) - \sum_{i=1}^s b_i H(n-k+i)- b'_i H'(n-k+i).
\end{split}
\end{align}
Suppose that $r=s$. There are two cases; $b_t-b'_t=-1$ for $1 \le t \le s-1$, or $b_\tau-b'_\tau=0$ for some $\tau<s$. In the first case note that $s'=s-1$, and observe that one must have $s \le k-s$. Otherwise $s > s-(k-s) \ge 1$ and since $s-(k-s)=k-2(k-s)$ one has $b_{k-2(k-s)}-b'_{k-2(k-s)} \ge 0$, a contradiction.
From (\ref{delta H}) for $n<3k-s$ one has
\begin{align*}
H(n)-H'(n) &= -H(n-k+s) + \sum_{t=1}^{s-1} H'(n-k+t) \ge -H(n-k+s).
\end{align*}
It is thus easy to see that
\begin{align*}
&H(n)-H'(n) \ge \begin{cases} 0 & 2k < n < 3k-2s \\ -1 & 3k-2s \le n < 3k-s.\end{cases}
\end{align*}
Since $\Delta(2k) \le -1$, using the relation
\begin{align*}
\Delta(n) = q\Delta(n-1)-H(n)+H'(n)
\end{align*}
one has
\begin{align*}
&\Delta(n) \le \begin{cases} q \Delta(n-1) & 2k < n < 3k-2s \\q \Delta(n-1)+1 & 3k-2s \le n < 3k-s. \end{cases}
\end{align*}
It follows that $\Delta(n) \le 0$ for $n<3k-s$.\\
We now suppose that $b_\tau-b'_\tau=0$ for some $1<\tau<s$. One has
\begin{align} \label{estimate delta 2k}
&\Delta(2k) \le -q^s+\sum_{t=1}^s q^{s-t} - q^{s-\tau} \le -2.
\end{align}
From (\ref{delta H}) one has
\begin{align*}
H(n)-H'(n) &\ge -\sum_{i=1}^s b_i H(n-k+i),
\end{align*}
from which we obtain the upper bound
\begin{align*}
&H(n)-H'(n) \ge -(n-2k).
\end{align*}
It follows that
\begin{align} \label{bound r=s 2}
&\Delta(n) \le q\Delta(n-1)+(n-2k).
\end{align}
The inequality (\ref{bound r=s 2}) together with (\ref{estimate delta 2k}) implies
\begin{align*}
\Delta(n) \le -2-(n-2k) \textrm{ for } 2k<n<3k-s
\end{align*}
and the result follows.
\end{proof}
\begin{lemma} \label{l3}
Let $k=k'$ and $N \ge 4k$. Suppose $h(N) > h'(N)$ and that $h(n) \leq h'(n)$ for $n<N$. Then $H(n)-H'(n) \le 0$ for $N \leq n \leq N+k$; In particular $\Delta(n) > (q-1) \Delta(n-1)$ for $N \leq n \leq N+k$.
\end{lemma}
\begin{proof}
Suppose that $r=s$, and observe that this is equivalent to the condition $s>s'$. By Lemma 4.2 we may assume that $N \ge 3k-s$. From (\ref{H rearranged}), (\ref{cancellation of non-prime periods}), and (\ref{chain}) one has
\begin{align*}
H(n) &\le h(n-k)-\sum_{i \in S} H(n-k+i)\\
&=h(n-k)-h(n-2k+s)+\sum_{t=1}^{k-s} b_t H(n-2k+s+t)\\
&=(q-1)\sum_{t=1}^{k-s} h(n-k-t) - \sum_{t=0}^{k-s-1} H(n-k-t) + \sum_{t=1}^{k-s} b_t H(n-2k+s+t)\\
&\le (q-1)\sum_{t=1}^{k-s} h(n-k-t).
\end{align*}
To summarize,
\begin{align} \label{H upper 1}
&H(n) \le (q-1) \sum_{t=1}^{k-s} h(n-k-t).
\end{align}
For any $L>0$ one has
\begin{align} \label{chain}
&h(n) = (q-1) \sum_{t=1}^L h(n-t) - \sum_{t=0}^{L-1} H(n-t).
\end{align}
Note that $s'<k-1$. From Proposition \ref{p2} there exists $\iota \in \{ 0, 1 \}$ such that $b'_{t(k-s')-\iota}=0$ for every $1 \le t \le U$ where $U = \lfloor \frac{k+\iota}{k-s'} \rfloor$. Using (\ref{h ineq}) and (\ref{chain}) one has
\begin{align*}
H'(n) &\ge h'(n-k) - \sum_{i=1}^{s'} b'_i h'(n-2k+i)\\
&= (q-1) \sum_{t=1}^k h'(n-k-t) - \sum_{t=0}^{k-1} H'(n-k-t) - \sum_{i=1}^{s'} b'_i h'(n-2k+i)\\
&\ge (q-1) \sum_{t=1}^{k-s'-1} h'(n-k-t) - \sum_{t=0}^{k-1} H'(n-k-t) \\
&+ \sum_{t=1}^U h'(n-2k-\iota+t(k-s'))+h'(n-2k)\\
&=(q-1) \sum_{t=1}^{k-s'-1} h'(n-k-t) +\left( h'(n-2k)-\sum_{l=1}^{k-s'-\iota} H'(n-2k+l) \right)\\
&+\sum_{t=1}^{U-1} \left( h'(n-2k-\iota+t(k-s')) - \sum_{l=1}^{k-s'} H'(n-2k-\iota+t(k-s')+l) \right)\\
&+ \left( h'(n-2k-\iota+U(k-s'))-\sum_{l=1}^{k+\iota-U(k-s')} H'(n-2k-\iota+U(k-s')+l) \right)\\
&\ge (q-1) \sum_{t=1}^{k-s'-1} h'(n-k-t).
\end{align*}
Thus
\begin{align} \label{H lower 1}
&H'(n) \ge (q-1)\sum_{t=1}^{k-s'-1} h'(n-k-t).
\end{align}
Using (\ref{H upper 1}) and (\ref{H lower 1}) it follows that
\begin{align*}
H(n)-H'(n) &\le (q-1) \sum_{t=1}^{k-s} h(n-k-t) - (q-1) \sum_{t=1}^{k-s'-1} h'(n-k-t) \le 0
\end{align*}
for $N \le n < N+k$.\\
Suppose that $r \neq s$, equivalently $s=s'$. By Lemma \ref{l2} we may assume that $n \ge 4k$. By use of Corollary \ref{c2} and equality (\ref{cancellation of non-prime periods}) one has
\begin{align*}
&H(n) = h(n-k) - \sum_{i \in I} h(n-2k+i) + \sum_{i \in I} \sum_{t=1}^s b_t H(n-2k+i+t)\\
&=(q-1) \sum_{t=1}^{k-d} h(n-k-t) - \sum_{t=0}^{k-d-1} H(n-k-t) - \sum_{i \in I-\{d\}} h(n-2k+i)\\
&+\sum_{i \in I} \sum_{t=1}^s b_t H(n-2k+i+t)\\
&\le (q-1) \sum_{t=1}^{k-d} h(n-k-t)- \sum_{i \in I-\{d\}} h(n-2k+i).
\end{align*}
It is easy to see that
\begin{align*}
&H'(n) \ge h'(n-k) - \sum_{i \in I'} h'(n-2k+i)\\
&=(q-1)\sum_{t=1}^{k-d'} h'(n-k-t) - \sum_{i \in I'-\{d'\}} h'(n-2k+i) - \sum_{t=0}^{k-d'-1} H'(n-k-t).
\end{align*}
It follows that
\begin{align} \label{this}
\begin{split}
&H(n)-H'(n) \le (q-1) \sum_{t=1}^{k-d} h(n-k-t) - \sum_{i \in I-\{d\}} h(n-2k+i)\\
&-(q-1) \sum_{t=1}^{k-d'} h'(n-k-t) + \sum_{i \in I'-\{d'\}} h'(n-2k+i) + \sum_{t=0}^{k-d'-1} H'(n-k-t).
\end{split}
\end{align}
Suppose $d' \le d$. One has
\begin{align*}
&H(n)-H'(n) \le -h'(n-2k+r) + \sum_{i=d+1}^{r-1} h(n-2k+i)+ \sum_{t=0}^{k-d'-1} H'(n-k-t)\\
&\le -h'(n-2k+r)+\sum_{i=d+1}^{r-1} h'(n-2k+i) + \sum_{t=0}^{k-d'-1} H'(n-k-t)\\
&\le -h'(n-2k+d) + \sum_{t=0}^{k-d'-1} H'(n-k-t) \le 0,
\end{align*}
using Corollary \ref{c4} and the fact that $h'(n-2k+r) - (q-1) \sum_{t=1}^{r-d-1} h'(n-2k+r-t) \ge (q-1) h'(n-2k+d)$ since $r-d-1 < k-s$.
Suppose $d' > d$. If $r=d$ then $I= I' \cup \{d\}$. Inequality (\ref{this}) becomes
\begin{align*}
&H(n)-H'(n) \leq (q-1) \sum_{t=k-d'+1}^{k-d} h(n-k-t) - h'(n-2k+d')+\\
&\sum_{t=0}^{k-d'} H'(n-k-t) \leq -h'(n-2k+d-1) + \sum_{t=0}^{k-d} H'(n-k-t) \leq 0,
\end{align*}
where we have used Corollary \ref{c4}.
Assuming now that $r>d'$ we add and subtract $h'(n-2k+d')$ to inequality (\ref{this}) and use corollary \ref{c4} to obtain
\begin{align*}
&H(n)-H'(n) \leq (q-1) \sum_{t=k-r}^{k-d} h(n-k-t) - h(n-2k+r)-\\
&(q-1) \sum_{t=k-r}^{k-d'} h'(n-k-t) + \sum_{t=r-1}^{d'} h'(n-2k+i) - h'(n-2k+d')+\\
&\sum_{t=0}^{k-d'-1} H'(n-k-t) \leq\\
&\sum_{t=k-d'+1}^{k-d} h(n-k-t) - h'(n-2k+r)+\sum_{t=k-r+1}^{k-d'} h(n-k-t) \leq 0.
\end{align*}
Finally, if $d'>r>d$ then $I' \subset I \cap \{k-1, \dots, r+1 \}$ and we have
\begin{align*}
&H(n)-H'(n) \le (q-1) \sum_{t=k-d'+1}^{k-d} h(n-k-t) - h'(n-2k+d')\\
&-h(n-2k+d)+\sum_{t=0}^{k-d'-1}H'(n-k-t) \le 0.
\end{align*}
\end{proof}
By combining statements proved in this section one can deduce Theorems 1 and 2 for the case $k=k'$. Indeed, let $\Delta(n) = h(n)-h'(n)$. Then $\Delta(n) \le 0$ for $n < 4k$, hence $N \ge 4k$. According to Lemma \ref{l3} one has $\Delta(N+t) \ge q \Delta(N+t-1) \ge (q-1) \Delta(N+t-1)$ for $0 \le t \le k-1$, and by Corollary \ref{c1} this implies $\Delta(N+k) \ge (q-1) \Delta(N+k-1)$. By a simple inductive argument, Corollary \ref{c1} then implies that $\Delta(n) \ge (q-1) \Delta(n-1)$ for any $n \ge N$. Theorem 2 follows from Theorem 1 by observing that $P_w(n) = h(n+k)/q^{n+k}$ and likewise $P_{w'}(n) = h'(n+k)/q^{n+k}$. Finally Theorem 3 is an immediate consequence of Lemma \ref{l2}.
\section{A Proof of Theorems 2.1 and 2.2 for the Case $\mathbf{k>k'}$}
\begin{lemma} \label{5.1}
Let $k>k'$. Then $h(n)-q^{k-k'}h'(n-k+k') \le 0$ for $n \le k+1$.
\end{lemma}
\begin{proof}
Note that $h(n)-q^{k-k'}h'(n-k+k') = 0$ for $k-k' < n < k$ and $h(n)-q^{k-k'}h'(n-k+k') = 1-q^{k-k'} < 0$ when $n=k$. When $n=k+1$, for any $0 \le m \le k-k'$ one has
\begin{align*}
h(k+1)-q^mh'(k'+1) &\le q-q^{m+1}+q^{m}H'(k'+1)\\
&\le -q^{m+1}+2q^{m} \le 0.
\end{align*}
\end{proof}
\begin{lemma} \label{5.2}
Let $k>k'$ and $N > k+1$. If $h(n) - q^{k-k'}h'(n-k+k') \le 0$ for $n<N$, then $H(n)-q^{k-k'}H'(n-k+k') \le 0$ for $N \leq n \leq N+k$.
\end{lemma}
\begin{proof}
For $0 \le m \le k-k'$ one has
\begin{align*}
H(n)-q^m H'(n-m) &\le h(n-k)-q^m(q-1)h'(n-k'-1-m)\\
&\le h(n-k)-q^m h'(n-k-m),
\end{align*}
and the result follows.
\end{proof}
Note that in Lemma \ref{5.2} the inequality $H(n)-q^{k-k'}H'(n-k+k') \le 0$ is equivalent to $q \left( h(n-1)-q^{k-k'} h'(n-k+k'-1) \right) \le h(n)-q^{n-k+k'} h'(n-k+k')$.
\begin{lemma} \label{5.3}
Let $k>k'$ and $n \ge 2k$. If $h(n-t)-q^{k-k'}h'(n-k+k'-t) \ge (q-1) \left( h(n-t-1)-q^{k-k'} h'(n-k+k'-t-1) \right)$ for $1 \leq t \leq k-1$ then $h(n)-q^{k-k'}h'(n-k+k'-t) \ge (q-1) \left( h(n-1)-q^{k-k'}h'(n-k+k') \right)$.
\end{lemma}
\begin{proof}
For any $0 \le m \le k-k'$ we have
\begin{align*}
h(n)-q^m h'(n-m)=&qh(n-1)-H(n)-q^{m+1}h'(n-m-1)+q^m H'(n-m)\\
=&q\left( h(n-1)-q^m h'(n-m-1) \right) +q^m H'(n-m)-H(n)\\
\ge &q\left( h(n-1)-q^m h'(n-m-1) \right)\\
&+q^m (q-1) h'(n-k'-m-1)-H(n)\\
\ge &q\left( h(n-1)-q^m h'(n-m-1) \right)\\
&+q^m h'(n-k'-m-1)-h(n-k)\\
\ge &q\left( h(n-1)-q^m h'(n-m-1) \right)+q^m h'(n-k-m)-h(n-k).
\end{align*}
Denote $K=k-k'$. With $m=K$ we apply the inductive assumption to obtain
\begin{align*}
&h(n)-q^m h'(n-m)\\
\ge &q\left( h(n-1)-q^K h'(n-K-1) \right)+q^K h'(n-k-K)-h(n-k)\\
\ge &q\left( h(n-1)-q^K h'(n-K-1) \right)+q^K h'(n-1-K)-h(n-1)\\
=&(q-1)\left( h(n-1)-q^K h'(n-K-1) \right).
\end{align*}
\end{proof}
The lemmas of this section combine to prove the main theorems when $k>k'$ in the following way. Let $\Delta(n) = h(n)-q^{k-k'}h'(n-k+k')$. Then according to Lemma \ref{5.3} one has $\Delta(n) \le 0$ for $n \le k+1$ hence $N > k+1$, which is the statement of Theorem 3. From Lemma \ref{5.2} one has $\Delta(N+t) \ge q \Delta(N+t-1) \ge (q-1) \Delta(N+t-1)$ for $0 \le t \le k-1$, and by Lemma \ref{5.3} this implies $\Delta(N+k) \ge (q-1) \Delta(N+k-1)$. By a simple inductive argument, Lemma \ref{5.3} then implies that $\Delta(n) \ge (q-1) \Delta(n-1)$ for any $n \ge N$. Dividing $\Delta(n)$ by $q^{n-k}$ then yields Theorem 2.
Observe that Lemmas \ref{5.1}, \ref{5.2}, and \ref{5.3} hold if everywhere in their statements we replace $k-k'$ with any $m$ satisfying $0 \le m \le k-k'$. Theorem 1 is a consequence of setting $m=0$.
\section{Numerical Results on Lengths of Short and Intermediate Times Intervals}
If we think about the applicability of our results to finite time predictions of dynamics then one is led to the following key question: How long is the short times interval where there exists a hierarchy of the first hitting probabilities curves, within which predictions are possible for any moment of time? Another time interval where such predictions can be made is the last (third) infinite times interval. Therefore it is of great importance for applications to estimate the lengths of two finite intervals, the short times interval where finite time predictions are possible and the second, intermediate interval.
Clearly these lengths depend on $k$, i.e. the lengths of the words which are coding the subsets of phase space which we consider.
Theorem 2.3 gives a linear estimate on the length of the short times interval. However, numerical simulations show that the lengths of both of these intervals grow exponentially (asymptotically as the common length $k$ of the words increases). If this is indeed the case then, practically speaking, only the short times interval is of interest because experiments and observations are usually not terribly long. In particular, if one is making observations about small subsets of the phase space, the short times interval may become very long, and thus covers the entire time of a reasonable (practically possible) experiment.
The following table presents the beginning and ending moments of the intermediate interval for the doubling map of $[0,1]$, i.e. the moments of time when the first and the last pairs of the first hitting probability curves intersect, respectively. Recall that the intermediate interval starts at the moment when the short times interval ends.
Notably, the length of the short times interval was in our computer experiments always larger than the length of the intermediate interval. It also appears that the ratio of lengths of these intervals converges to one in the limit when $k$ tends to infinity.
\begin{center}
\begin{tabular}{||c c c||}
\hline
$k$ & Beginning of interval & End of interval \\ [0.5ex]
\hline\hline
$4$ & 20 & 26 \\
\hline
$5$ & 37 & 52 \\
\hline
$6$ & 70 & 103 \\
\hline
$7$ & 135 & 208 \\
\hline
$8$ & 264 & 415 \\[1ex]
\hline
\end{tabular}
\end{center}
\begin{figure}[htp]
\centering
\includegraphics[width=10cm]{ex2.png}
\caption{First hitting curves for the doubling map.}
\label{fig2}
\end{figure}
In figure \ref{fig2} we present three first hitting probability curves for the doubling map. These curves correspond to the words $10101010$, $10010010$, and $11100111$ of length eight. (Recall that Figure \ref{fig} presents similar results for words of length four).
Recall that as the number of elements in a refinement of the Markov partition increases so does $k$ too, which is the length of the word coding each element. Therefore when we consider dynamics at finer scales, the length of the short times interval (on which predictions about the dynamics can be made) seems to grow exponentially. It is a very encouraging (although numerical) result which suggests that finite time predictions of dynamics could be made on very long time scales if we consider partition of the phase space with a sufficiently large number of elements.
\section{Concluding Remarks}
Our results demonstrate that interesting and important finite time predictions for the dynamics of systems with the strongest chaotic properties and for the most random stochastic processes are possible. They also indicate how such predictions can be practically made. Numerical simulations \cite{BV} suggest that some finite time predictions for nonuniformly hyperbolic systems are also possible.
However finite time predictions in this case will be not as simple as for FDL-systems which are the uniformly hyperbolic systems with the maximal possible uniformity. It seems natural to expect that for more general classes of chaotic dynamical systems there will be more than two time intervals with different hierarchies of the first hitting probabilities. Some (although too vague for formulating exact conjectures) indications of that can be extracted from computer simulations of dispersing billiards \cite{BV}
Although the theory of finite time dynamics of chaotic systems is in infancy, it is rather clear what to do next and to which classes of dynamical systems these results should be generalized.
For FDL-systems a remaining problem is to prove better estimates of the length of the short times interval.
A natural next step would be to consider IID-like but not FDL dynamical systems. Consider for instance a skewed tent map of the unit interval, i.e. $f(x)=ax$ if $0<x<1/a$ and $f(x)=ax/(1-a)$ if $1/a<x<1$, where $a>1$. This map is not an FDL-system (unless $a=2$) because the absolute values of the derivatives differ at different points of the phase space. (Therefore it is a system with distortion). Then one should follow a standard path in developing dynamical systems theory by trying to obtain finite time dynamics results for more and more non-uniformly hyperbolic dynamical systems.
For instance a natural question is whether it would be possible to find optimal Young towers \cite{LY} for some interesting classes of nonuniformly hyperbolic dynamical systems and estimate corresponding exponents in decay of the first recurrences probabilities.
A significant problem is to develop relevant mathematical approaches and techniques, more dynamical than combinatorial in spirit, to handle the questions arising in the studies of finite time dynamics.
It is also worthwhile to mention that in the Equilibrium Statistical Mechanics the main problem is about phase transitions, i.e. existence or non-existence of several equilibrium distributions. If there are no such transitions then usually problems of the Nonequilibrium Statistical Mechanics come under attack, i.e. what is a rate of approach to the equilibrium etc. The FDL-systems demonstrate that even when a system is in unique equilibrium (i.e. there are no phase transitions) still its evolution could be nonhomogeneous and have interesting properties pertaining to the transport in the phase space.
\section*{Appendix 1. An Upper Bound for $\mathbf{\Delta(n)}$ on the Interval $\mathbf{2k < n \le 4k}$ when $\mathbf{s=s'}$}
Viewing $\Delta(n)$ as a function of the values $b_i$ and $b'_i$ for $i<r$, we will show that if $b_i=0$ then $\Delta(b_i+1, n)-\Delta(b_i,n) \le 0$. One can also show that if $b_i=1$ then $\Delta(b_i-1,n)-\Delta(b_i,n) \ge 0$. In addition, if $b'_i = 0$ or $1$ then $\Delta(b_i'+1, n)-\Delta(b'_i, n) \ge 0$ or $\Delta(b'_i-1,n) - \Delta (b'_i,n) \le 0$, respectively. Because of the almost complete redundancy in all these calculations, we will only display the former case.
Let
\begin{align*}
\tilde{b}_t = \begin{cases} 1 & b_t=1 \textrm{ and } t \in I\\ 0 & \textrm{ otherwise} \end{cases}
\end{align*}
and
\begin{align*}
\delta (n)= \sum_{t=1}^{3k-n-1} b_t \tilde{b}_{3k-n-t} - b'_t \tilde{b}'_{3k-n-t} .
\end{align*}
For $2k \le n<3k-s$ one has
\begin{align} \label{first app 1}
&H(n)-H'(n) = -\delta(n).
\end{align}
To see this, observe that a word of length $n$ beginning and ending with $w$ contains a copy of $w$ beginning in position $n-k+t$ if and only if $b_t=1$ and $H(t+(n-2k)+k)=1$. Therefore we have the equality $H(n) = q^{n-2k}-\sum_{t=1}^k b_t H(n-k+t)$. Since $n -k+t < 2k-s+t \le 2k$ we may apply the identity $H(n-k+t) = \tilde{b}_{3k-n-t}$, which proves (\ref{first app 1}).
We establish the convention that $\delta(n)=0$ when $n < 2k$. When $n=3k-s$ one has
\begin{align*}
H(n)-H'(n) &= -\delta(n) - \sum_{t=3k-n}^s b_t \left( H(n-k+t)- H'(n-k+t) \right) \\
&= -\delta(n)+\delta(n-k+s).
\end{align*}
We now prove inductively that
\begin{align*}
H(n)-H'(n) = -\delta(n)+\delta(n-k+s)
\end{align*}
for $n < 3k-r$.
First we observe that $H(n)-H'(n) = 0$ when $n<3k-s-r$. Note that $3k-n-t \ge 3k-n-s > r$, hence $\tilde{b}_{3k-n-t} = \tilde{b}'_{3k-n-t}$ (note that $I \cap \{i>r\} = I' \cap \{i>r\}$). For $t \ge r$ one therefore has $b_t \tilde{b}_{3k-n-t}-b'_t \tilde{b}'_{3k-n-t} =0$. When $t<r$, observe that $3k-n-t \ge s+r-t > s$, hence $b_t \tilde{b}_{3k-n-t}-b'_t \tilde{b}'_{3k-n-t} =0$. It follows that $H(n)-H'(n)=-\delta(n)=0$ for $n<3k-s-r$.
If $b_t=1$ and $t \notin S$, one has $n-k+t < 3k-s-r$, hence $H(n-k+t)-H'(n-k+t) =0$. Let $p=\lfloor \frac{n-2k}{k-s} \rfloor$. One then has
\begin{align*}
&H(n)-H'(n) = -\delta(n) - \sum_{i=3k-n}^s b_i (H(n-k+i)-H'(n-k+i))\\
&=-\delta(n) - \sum_{t=1}^p \big( H(n-t(k-s))-H'(n-t(k-s)) \big) \\
&- \sum_{i \notin S } b_i \big( H(n-k+i)-H'(n-k+i) \big)\\
&=-\delta(n) - \sum_{l=1}^p \Big( -\delta(n-l(k-s))+\delta \big( n-(l+1)(k-s) \big) \Big)\\
&=-\delta(n) + \delta(n-k+s),
\end{align*}
where we observe that $n-(p+1)(k-s) \le 2k$. It follows that for $n < 3k-r$ one has
\begin{align} \label{H less than k-r}
\begin{split}
H(n)-H'(n) = -\delta(n) +\delta(n-k+s).
\end{split}
\end{align}
For $2k<n<3k-r$ by using (\ref{H less than k-r}) one thus has
\begin{align} \label{Delta < 3k-r}
\begin{split}
\Delta(n) &= -q^{n-2k}\sum_{t=1}^r q^t ( \tilde{b}_t - \tilde{b}'_t ) + \sum_{t=2k}^n q^{n-t} \delta(t) - \sum_{t=3k-s}^n q^{n-t} \delta(t-k+s),
\end{split}
\end{align}
where we have used the fact that
\begin{align*}
\Delta(2k) = -\sum_{t=1}^r q^t ( \tilde{b}_t-\tilde{b}'_t )
\end{align*}
when $s=s'$.
We now fix an index $i$ with $1 \le i \le k$ such that $b_i=0$. For $n \ge k$ let us denote by $h^*(n)$ the solution to the recurrence relation $h^*(n) = \sum_{t=1}^k b^*_t (qh^*(n+t-1)-h^*(n+t))$ where
\begin{align*}
b^*_t = \begin{cases} b_t & t \ne i\\ 1 & t=i, \end{cases}
\end{align*}
subject to the initial conditions $h^*(n)=0$ for $n<k$ and $h^*(k)=1$. Let $H^*(n) = qh^*(n-1)-h^*(n)$.
Denote
\begin{align*}
\delta^*(n) = \sum_{t=1}^{3k-n-i} b^*_t \tilde{b}^*_{3k-n-t}-b_t \tilde{b}_{3k-n-t} = \tilde{b}^*_{3k-n-i}+b^*_{3k-n-i}.
\end{align*}
We will also denote
\begin{align*}
\Delta^*(n) = h^*(n)-h(n).
\end{align*}
It is easy to see that
\begin{align} \label{delta star vals}
\begin{split}
&\delta^*(t) \begin{cases} =0, & t<3k-i-s\\ =2, & t=3k-i-s\\ =0, & t=3k-i-s+1, \end{cases}\\
&0 \le \delta^*(t) \le 2 \textrm{ for } 3k-i-s+1 < t \le 3k-1.
\end{split}
\end{align}
For $n<3k-s$ equality (\ref{Delta < 3k-r}) becomes
\begin{align*}
\Delta^*(n) &= -q^{n-2k+i} + \sum_{t=2k}^n q^{n-t} \left( \tilde{b}^*_{3k-t-i} + b^*_{3k-t-i} \right) \\
& = -q^{n-2k+i} + \sum_{t=3k-s-i}^n q^{n-t} \left( \tilde{b}^*_{3k-t-i} + b^*_{3k-t-i} \right) < 0,
\end{align*}
where we have used the fact that $i \le k-s-2$. If $3k-s \le n < 3k-i$ then
\begin{align} \label{delta < 3k-r}
\begin{split}
\Delta^*(n) &= -q^{n-2k+i} + \sum_{t=2k}^n q^{n-t} \left( \tilde{b}^*_{3k-t-i} + b^*_{3k-t-i} \right) \\
&- \sum_{t=3k-s}^n q^{n-t} \left( \tilde{b}^*_{3k-t-i+k-s} + b^*_{3k-t-i+k-s} \right) < 0.
\end{split}
\end{align}
Let $3k-i \le n \le 3k$. One has
\begin{align} \label{asdfg}
\begin{split}
&\Delta^*(n) \le -q^{n-2k+i} + \sum_{t=2k}^{3k-i-1} q^{n-t} \left( \tilde{b}^*_{3k-t-i} + b^*_{3k-t-i} \right) \\
&- \sum_{t=3k-i}^n q^{n-t} \left(H^*(t)-H( t) \right) = -q^{n-2k+i} + \sum_{t=2k}^{3k-i-1} q^{n-t} \left( \tilde{b}^*_{3k-t-i} + b^*_{3k-t-i} \right) \\
&- \sum_{t=3k-i}^n q^{n-t} \Delta^*(t-k) + \sum_{t=3k-i}^n q^{n-t} H^*(t-k+i)\\
&+ \sum_{t=3k-i}^n q^{n-t} \sum_{j=1}^s b_j \left( H^*( t-k+j)-H(t-k+j) \right).
\end{split}
\end{align}
Note that $t-k+j \le 2k+s < 3k-i$ hence $H^*(t-k+j)-H(t-k+j) = -\delta^*(t-k+j)+\delta^*(t-2k+j+s)$ when $t-k+j \ge 2k$.
One has
\begin{align*}
&\sum_{t=3k-j}^n q^{n-t} b^*_{4k-t-j-i} - \sum_{t=3k-j+(k-s)}^n q^{n-t} b^*_{4k-t-j-i+(k-s)} \\
&=\sum_{t=3k-j}^n q^{n-t} b^*_{4k-t-j-i} - \sum_{t=3k-j}^{n-k+s} q^{n-t-k+s} b^*_{4k-t-j-i} \ge 0
\end{align*}
and similarly
\begin{align*}
&\sum_{t=3k-j}^n q^{n-t} \tilde{b}^*_{4k-t-j-i} - \sum_{t=3k-j+(k-s)}^n q^{n-t} \tilde{b}^*_{4k-t-j-i+(k-s)} \ge 0.
\end{align*}
It follows that
\begin{align} \label{LP -2}
\begin{split}
&\sum_{t=3k-j}^n q^{n-t} (H^*(t-k+j)-H(t-k+j)) \\
&= \sum_{t=3k-j}^n q^{n-t} \delta^*(t-k+j)-\sum_{t=3k-j+(k-s)}^n q^{n-t} \delta^*(t-k+j-(k-s)) \ge 0.
\end{split}
\end{align}
Using (\ref{LP -2}) one has
\begin{align} \label{LP -1}
\begin{split}
&\sum_{t=3k-i}^n q^{n-t} \sum_{j=1}^s b_j \left( H^*(t-k+j)-H(t-k+j) \right)\\
&=\sum_{j=1}^{i-1} b_j \sum_{t=3k-i}^{ 3k-j-1} q^{n-t} \left( \tilde{b}^*_{3k-t-j} - \tilde{b}_{3k-t-j} \right) - \sum_{j=1}^s \sum_{t=3k-j}^n q^{n-t} (H^*(t-k+j)-H(t-k+j))\\
&\le \sum_{j=1}^{i-1} b_j \sum_{t=3k-i}^{ 3k-j-1} q^{n-t} \left( \tilde{b}^*_{3k-t-j} - \tilde{b}_{3k-t-j} \right)
\end{split}
\end{align}
for $n \le 3k$.
For $3k-i \le t < 3k$ note that
\begin{align} \label{LP -2.5}
\Delta^*(t-k) = - \sum_{l=3k-t}^i q^{t-3k+l} \left( b^*_l-b_l \right)
\end{align}
and recall
\begin{align*}
\Delta^*(2k) = -\sum_{l=1}^i q^l (b^*_l - b_l),
\end{align*}
whence $\Delta^*(t-k) = -q^{t-3k+i}$ for $3k-i \le t \le 3k$. Using $H^*(t-k+i) \le q^{t-3k+i}$ for $t-k+i \ge 2k$ and (\ref{LP -1}), equality (\ref{asdfg}) becomes
\begin{align} \label{LP 0}
\begin{split}
&\Delta^*(n) \le -q^{n-2k+i} + \sum_{t=2k}^{3k-i-1} q^{n-t} \left( \tilde{b}^*_{3k-t-i} + b^*_{3k-t-i} \right) \\
&+\sum_{t=3k-i}^n q^{n-3k+i} + \sum_{t=3k-i}^n q^{n-3k+i} +\sum_{j=1}^{i-1} b_j \sum_{t=3k-i}^{ 3k-j-1} q^{n-t} \left( \tilde{b}^*_{3k-t-j} - \tilde{b}_{3k-t-j} \right)\\
&\le -q^{n-2k+i} + 2 \sum_{t=3k-s-i}^{3k-i-1} q^{n-t} + 2(i+1) q^{n-3k+i} + q^{n-3k+2i}\\
&\le -q^{n-2k+i} + q^{n-3k+s+i+2} + q^{n-3k+2i+1}+q^{n-3k+2i} \le 0
\end{split}
\end{align}
where we observe that $\tilde{b}^*_{3k-t-j}-\tilde{b}_{3k-t-j}=1$ if and only if $t=3k-i-j$ and where we have used the facts $i < r \le k-s-2$ and $2(i+1) \le q^{i+1}$.\\
Let $3k < n < 4k$. Using (\ref{cancellation of non-prime periods}) one has
\begin{align} \label{LP 1}
\begin{split}
&-\sum_{t=3k+1}^n q^{n-t} \left( H^*(t)-H(t) \right) =-\sum_{t=3k+1}^n q^{n-t} \Delta^*(t-k) + \sum_{t=3k+1}^n \sum_{j \in I} \Delta^*(t-2k+j)\\
&-\sum_{t=3k+1}^n q^{n-t} \sum_{j \in I} \sum_{l=1}^{k-j} b_l \left( H^*(t-2k+j+l) - H(t-2k+j+l) \right)\\
&+\sum_{t=3k+1}^n q^{n-t} H^*(t-k+i) - \sum_{t=3k+1}^n \sum_{j \in I} q^{n-t} H^*(t-2k+j+i).
\end{split}
\end{align}
For fixed $j$ and $u \le 4k-j-l-1$ one has
\begin{align} \label{LP -2.625}
\begin{split}
&- \sum_{l=1}^{k-j} b_l \sum_{t=3k+1}^u q^{n-t} \left( H^*(t-2k+j+l)-H(t-2k+j+l) \right)\\
&= - \sum_{l=1}^{k-j} b_l \sum_{t=3k+1}^u q^{n-t} \left( \tilde{b}^*_{4k-t-j-l}- \tilde{b}_{4k-t-j-l} \right) \le 0,
\end{split}
\end{align}
observing that $\tilde{b}^*_\iota \ge \tilde{b}_\iota$. For $4k-j-l-1 \le u \le 5k-j-l-i-1$ one also has
\begin{align} \label{LP -2.75}
\begin{split}
&-\sum_{l=1}^{k-j} \sum_{t=4k-j-l-1}^u q^{n-t} \left( H^*(t-2k+j+l)-H(t-2k+j+l) \right)\\
&=\sum_{l=1}^{k-j} \sum_{t=4k-j-l-1}^u q^{n-t} \left( \delta^*(t-2k+j+l)-\delta^*(t-2k+j+l-k+s) \right)\\
&\le \sum_{l=1}^{k-j} \sum_{t=5k-j-l-s-i}^u q^{n-t} \delta^*(t-2k+j+l) \le 2q^{n-4k+i+s+2}.
\end{split}
\end{align}
For $3k-i \le n \le 3k$ one has
\begin{align*}
&H^*(n)-H(n) = \Delta^*(n-k)-\sum_{t=1}^s b_t \left( H^*(n-k+t)- H(n-k+t) \right) - H^*(n-k+i)\\
&\ge \Delta^*(n-k) - \sum_{t=1}^s \delta^*(n-2k+t+s) - H^*(n-k+i)\\
&\ge -q^{n-3k+i}-2s-H^*(n-k+i).
\end{align*}
For $n \ge 4k-i$ one thus has
\begin{align} \label{LP -3}
\begin{split}
&-\sum_{l=1}^{k-j} \sum_{t=5k-j-l-i}^n q^{n-t} \left( H^*(t-2k+j+l)-H(t-2k+j+l) \right)\\
&\le -\sum_{l=1}^{k-j} \sum_{t=5k-j-l-i}^n q^{n-t} \left( -q^{t-5k+j+l+i}-2s-H^*(t-3k+j+l+i) \right)\\
&\le q^{n-4k+2i} + q^{n-4k+s+i+2} + \sum_{l=1}^{k-j} \left( \sum_{t=5k-j-l-i}^{5k-j-l-i-1} q^{n-t} + \sum_{t=5k-j-l-i}^n q^{n-5k+j+l+i} \right)\\
&\le q^{n-4k+2i} + q^{n-4k+s+i+2} + q^{n-4k+i+2} + q^{n-4k+2i}\\
&\le 2q^{n-3k-4}+ q^{n-3k-1}+q^{n-3k-3} \le q^{n-3k-1}+q^{n-3k-2} + q^{n-3k-3},
\end{split}
\end{align}
where we have applied (\ref{LP -2.5}) and used the inequalities
\begin{align} \label{LP 1.5}
&i \le q^{i-1}, \hspace{7mm} i<r, \hspace{7mm} r \le k-s-2, \hspace{7mm} r \le \frac{k}{2}-1, \hspace{7mm} s \le k-3.
\end{align}
For $n< 4k-i$
\begin{align} \label{LP -3 E}
&-\sum_{l=1}^{k-j} \sum_{t=5k-j-l-i}^n q^{n-t} \left( H^*(t-2k+j+l)-H(t-2k+j+l) \right) = 0
\end{align}
since the former sum is empty. Using (\ref{LP -2.625}), (\ref{LP -2.75}), and (\ref{LP -3 E}), for $3k < n < 4k-i$ one has
\begin{align} \label{LP -3.25}
&-\sum_{t=3k+1}^n q^{n-t} \sum_{j \in I} \sum_{l=1}^{k-j} b_l \left( H^*(t-2k+j+l) - H(t-2k+j+l) \right) \le \sum_{j \in I} 2 q^{n-4k+s+i+2}.
\end{align}
For any $3k < n < 4k$ note that
\begin{align} \label{LP -3.5}
\begin{split}
&\sum_{j \in I} \sum_{t=3k+1}^n q^{n-t} H^*(t-2k+j+i)\\
&=\sum_{j \in I} \left( \sum_{t=4k-j-i-s}^{4k-j-i-1} q^{n-t} \tilde{b}^*_{4k-t-j-i} + \sum_{t=4k-j-i}^n q^{n-t} H^*(t-2k+j+i) \right)\\
&\le \sum_{j \in I} \left( q^{n-4k+j+i+s+1} + (n-3k) q^{n-4k+j+i} \right) \le q^{n-3k+s-2} + (n-3k)q^{n-3k-2}.
\end{split}
\end{align}
For $3k < n < 4k-i$ one has $\Delta^*(t-k) \ge -q^{t-3k+i}$. Applying inequalities (\ref{LP -3.25}) and (\ref{LP -3.5}) to (\ref{LP 1}) we have
\begin{align*}
&-\sum_{t=3k+1}^n q^{n-t} \left( H^*(t)-H(t) \right) \le -\sum_{t=3k+1}^n q^{n-t} \Delta^*(t-k) + (k-s) 2 q^{n-4k+s+i+2} \\
&+ (n-3k) q^{n-3k+i} + (n-3k)q^{n-3k-2} + q^{n-3k+s-2}\\
&\le (n-3k-1) q^{n-3k+i} + (k-s) q^{n-4k+s+i+3} + (n-3k) q^{n-3k+i}\\
&+ (n-3k)q^{n-3k-2}+q^{n-3k+s-2}\\
&\le q^{n-3k+i+k/2} + q^{n-3k+i+2} + q^{n-3k+i+k/2} + q^{n-3k+k/2-2}+q^{n-3k+s-2}\\
&\le 4q^{n-2k-3} + q^{n-2k-5},
\end{align*}
where we have used the inequality $n-3k \le k-1 \le q^{k/2}$, inequalities (\ref{LP 1.5}), and the fact that $k \ge 5$ when $s=s'$.
Similar to (\ref{LP 0}) one has
\begin{align*}
&q^{n-2k}\Delta^*(2k) - \sum_{t=2k}^{3k-1} q^{n-t} (H^*(t)-H(t))\\
&\le -q^{n-2k+i}+q^{n-3k+s+i+2}+q^{n-3k+2i+1}+q^{n-3k+2i}.
\end{align*}
It follows that for $3k < n < 4k-i$ we have
\begin{align*}
&\Delta^*(n) \le q^{n-2k}\left( -q^i + q^{-1}+q^{-5}+q^{-6}+4q^{-3}+q^{-5} \right) \le 0,
\end{align*}
where we have again used the inequalities (\ref{LP 1.5}).
For $3k-i \le n \le 3k$ it is easy to see that
\begin{align*}
\Delta^*(n) &\ge -q^{n-2k+i} - \sum_{t=3k-i}^n q^{n-t} (H^*(t)-H(t))\\
&\ge -q^{n-2k+i} -(n-3k+i)q^{n-2k}\\
&\ge -q^{n-2k+i}-q^{n-2k+i}.
\end{align*}
For $4k-i \le n < 4k$ one thus has
\begin{align*}
&-\sum_{t=3k+1}^n q^{n-t} \left( H^*(t)-H(t) \right) \le 4q^{n-2k-3} + q^{n-2k-5} + \sum_{t=4k-i}^n q^{n-3k+i}\\
&+q^{n-3k-1}+q^{n-3k-2} + q^{n-3k-3}\\
&\le 4q^{n-2k-3} + q^{n-2k-5} + q^{n-3k+2i-1}+q^{n-3k}\\
&\le 4q^{n-2k-3} + 3q^{n-2k-5}
\end{align*}
where we have used (\ref{LP -3}).
For $4k-i \le n < 4k$ one thus has
\begin{align*}
&\Delta^*(n) \le q^{n-2k}\left( -q^i + q^{-1}+q^{-5}+q^{-6}+4q^{-3}+3q^{-5} \right) \le 0.
\end{align*}
Given $w$ and $\iota \in I-\{s\}$, let $h_\iota(n)$ be the solution to the recurrence relation defined by
\begin{align*}
&h_\iota(n) = \sum_{t=1}^k c_t (qh_\iota(n-t-1)-h_\iota(n-t)), \hspace{5mm} c_t = \begin{cases} b_t & t>\iota\\ 0 & t=\iota\\ 1 & t<\iota \end{cases}.
\end{align*}
and $h^\iota(n)$ the solution to
\begin{align*}
&h^\iota(n) = \sum_{t=1}^k c_t (qh^\iota(n-t-1)-h^\iota(n-t)), \hspace{5mm} c_t = \begin{cases} b_t & t>\iota\\ 1 & t=\iota\\ 0 & t<\iota \end{cases}.
\end{align*}
We have shown that $\Delta(b_i+1, n) \le \Delta(b_i,n)$ for $2k \le n < 4k$. As we remarked, with minimal alterations to the calculations of this section one can show the inequality $\Delta(b_i-1,n) \ge \Delta(b_i,n)$ if $b_i=1$. Thus $\Delta(n) \le \Delta^*(n)$ for $2k \le n < 4k$ where $h^*(n)$ is the result of setting $b_r=0$.
\section*{Appendix 2. The Upper Bound is Negative}
Let $h^r(n)$ and $h_r(n)$ be as defined in Appendix 1 above. We will show that $h^r(n)-h_r(n) \le 0$ for $2k \le n < 4k$. Throughout this section we will let $h(n) = h^r(n)$ and $h'(n) = h_r(n)$ to avoid further burdening the notation.
For $t<3k-r$ one has
\begin{align*}
&\delta(t) = \tilde{b}_{3k-t-r}-\sum_{l=1}^{r-1} \tilde{b}'_{3k-t-l} + \sum_{l=r+1}^{3k-t-1} b_l \left( \tilde{b}_{3k-t-l} - \tilde{b}'_{3k-t-l} \right)\\
&=\tilde{b}_{3k-t-r}-\sum_{l=1}^{r-1} \tilde{b}'_{3k-t-l} + \sum_{l=r+1}^{3k-t-r-1} b_l \left( \tilde{b}_{3k-t-l} - \tilde{b}'_{3k-t-l} \right)+b_{3k-t-r} - \sum_{l=3k-t-r+1}^{3k-t-1} b_l\\
&=\tilde{b}_{3k-t-r}-\sum_{l=1}^{r-1} \tilde{b}'_{3k-t-l} + b_{3k-t-r} - \sum_{l=3k-t-r+1}^{3k-t-1} b_l\\
&=\tilde{b}_{3k-t-r}-\sum_{l=1}^{r-1} \tilde{b}'_{3k-t-l} + b_{3k-t-r} - \sum_{l=1}^{r-1} b_{3k-t-l}.
\end{align*}
It is thus easy to see that
\begin{align} \label{LP2 -1}
&\delta(t) \begin{cases} =0, & t < 3k-s-r\\ =2, & t=3k-s-r\\ = -2, & t=3k-s-r+1 \end{cases} \delta(t) \begin{cases} \le 2, & t < 3k-r\\ \ge -2, & t<2k+s-r\\ \ge -2(r-1), & 2k+s-r \le t < 3k-r \end{cases}.
\end{align}
Using (\ref{LP2 -1}), for $n<3k-r$ one thus has
\begin{align*}
-\sum_{t=3k-s}^n q^{n-t} \delta(t-k+s) \le 0
\end{align*}
and
\begin{align*}
\sum_{t=2k}^n q^{n-t} \delta(t) \le 2 q^{n-3k+s+r}.
\end{align*}
Using (\ref{Delta < 3k-r}) and the equality $\Delta(2k)=-2$, for $2k \le n < 3k-r$ one thus has
\begin{align*}
&\Delta(n) \le -q^{n-2k} \Delta(2k) + 2q^{n-3k+s+r} \le -2\left( q^{n-2k} - q^{n-3k+s+r} \right) < 0.
\end{align*}
Let $3k-r \le n \le 3k$. One has
\begin{align} \label{LP2 1}
\begin{split}
&\Delta(n) = q^{n-3k+r+1} \Delta(3k-r-1) - \sum_{t=3k-r}^n q^{n-t} (H(t)-H'(t))\\
&= q^{n-3k+r+1} \Delta(3k-r-1) - \sum_{t=3k-r}^n q^{n-t} \Delta(t-k) \\
&+ \sum_{t=3k-r}^n q^{n-t} \sum_{j=1}^s \left( b_j H(t-k+j)-b'_j H'(t-k+j) \right).
\end{split}
\end{align}
Note that
\begin{align*}
&\sum_{t=3k-r}^n q^{n-t} \sum_{j=1}^s \left( b_j H(t-k+j)-b'_j H'(t-k+j) \right)\\
&=\sum_{t=3k-r}^n \left( q^{n-t} H(t-k+r) -\sum_{j=1}^{r-1} q^{n-t} H'(t-k+j) \right)\\
&+ \sum_{j=r+1}^s b_j \sum_{t=3k-r}^n q^{n-t} \left(H(t-k+j)-H'(t-k+j) \right)\\
&\le \sum_{t=3k-r}^n q^{n-t} H(t-k+r)+ \sum_{j=r+1}^s b_j \sum_{t=3k-r}^n q^{n-t} \left(H(t-k+j)-H'(t-k+j) \right).
\end{align*}
For $r < j \le s$ and $3k-r \le t \le n$ one has $2k < t-k+j < 3k-r$, and again using (\ref{LP2 -1}) we have
\begin{align*}
&\sum_{j=r+1}^s b_j \sum_{t=3k-r}^n q^{n-t} \left(H(t-k+j)-H'(t-k+j) \right)\\
&= -\sum_{j=r+1}^s b_j \sum_{t=3k-r}^n q^{n-t} \left( \delta(t-k+j)-\delta(t-2k+j+s) \right)\\
&\le 2(r-1) \sum_{j=r+1}^s b_j q^{n-3k+r+1} + 2 \sum_{j=r+1}^s b_j q^{n-5k+2s+r+j}\\
&\le 2(r-1) q^{n-3k+s} + q^{n-5k+3s+r+2} \le q^{n-3k+s+r-1} + q^{n-4k+2s}\\
&\le q^{n-3k+s+r-1} + q^{n-3k+s-3},
\end{align*}
where we have applied the inequalities $r+s \le k-2$ and $s \le k-3$. Inequality (\ref{LP2 1}) thus becomes
\begin{align} \label{up to 3k}
\begin{split}
&\Delta(n) = q^{n-3k+r+1} \Delta(3k-r-1) - \sum_{t=3k-r}^n q^{n-t} (H(t)-H'(t))\\
&\le q^{n-3k+r+1} \Delta(3k-r-1) - \sum_{t=3k-r}^n q^{n-t} \Delta(t-k)+ q^{n-3k}((r+1)q^r + q^{s+r-1}+q^{s-3}\\
&\le -2q^{n-2k}+2q^{n-3k+r+s}+\sum_{t=3k-r}^n q^{n-t} + q^{n-3k}(q^{2r}+q^{s+r-1}+q^{s-3})\\
&\le -2q^{n-2k}+q^{n-3k}(q^{s+r+1}+q^{r+1}+q^{2r}+q^{s+r-1}+q^{s-3})\\
&\le -q^{n-2k},
\end{split}
\end{align}
where we have used the inequality $\Delta(3k-r-1) \le -2\left( q^{k-r-1}-q^{s-1} \right)$ and the fact that $\Delta(n)=-1$ for $2k-r \le n \le 2k$.\\
Let $3k < n < 4k$. Denote $I_t = I \cap \{t, \dots, s\}$. One has
\begin{align*}
&-\sum_{t=3k+1}^n q^{n-t} (H(t)-H'(t)) = -\sum_{t=3k+1}^n q^{n-t} \Delta(t-k) + \sum_{t=3k+1}^n q^{n-t} \sum_{j \in I_{r+1}} \Delta(t-2k+j) \\
&- \sum_{t=3k+1}^n q^{n-t} \sum_{j \in I_{r+1}} \sum_{l=1}^{k-j} \left( b_l H(t-2k+j+l)-b'_l H'(t-2k+j+l) \right)\\
&+\sum_{t=3k+1}^n q^{n-t} H(t-k+r)-\sum_{j=1}^{r-1} \sum_{t=3k+1}^n q^{n-t} H'(t-k+j),
\end{align*}
hence
\begin{align} \label{last case}
\begin{split}
&-\sum_{t=3k+1}^n q^{n-t} (H(t)-H'(t)) \le -\sum_{t=3k+1}^n q^{n-t} \Delta(t-k)+\sum_{t=3k+1}^n q^{n-t} H(t-k+r)\\
&- \sum_{t=3k+1}^n q^{n-t} \sum_{j \in I_{r+1}} \sum_{l=1}^{k-j} \left( b_l H(t-2k+j+l)-b'_l H'(t-2k+j+l) \right).
\end{split}
\end{align}
Let $n<4k-r$. Note that $t-2k+j+l < 3k-r$. For fixed $j$, if $r<l \le k-j$ one has
\begin{align*}
&-\sum_{t=3k+1}^n q^{n-t} \left( b_l H(t-2k+j+l)-b'_l H'(t-2k+j+l) \right)\\
&= -b_l \sum_{t=3k+1}^n q^{n-t} \left( H(t-2k+j+l)- H'(t-2k+j+l) \right)\\
&\le -b_l \sum_{t=3k+1}^{4k-j-l-1} q^{n-t} \left( \tilde{b}_{4k-t-j-l} - \tilde{b}'_{4k-t-j-l} \right) + b_l \sum_{t=4k-j-l}^n q^{n-t} \delta(t-2k+j+l)\\
&-b_l \sum_{t=4k-j-l}^n q^{n-t} \delta(t-3k+j+l+s).
\end{align*}
For any $u \le 4k-j-l-1$ observe that
\begin{align*}
&-\sum_{j \in I_{r+1}} \sum_{l=r+1}^{k-j} b_l \sum_{t=3k+1}^{u} q^{n-t} \left( \tilde{b}_{4k-t-j-l} - \tilde{b}'_{4k-t-j-l} \right)\\
&=-\sum_{j \in I_{r+1}} \sum_{l=r+1}^{k-j} b_l \sum_{t=4k-j-l-r}^{u} q^{n-t} \left( \tilde{b}_{4k-t-j-l} - \tilde{b}'_{4k-t-j-l} \right)\\
&=-\sum_{j \in I_{r+1}} \sum_{l=r+1}^{k-j} b_l \sum_{t=4k-j-l-u}^r q^{n-4k+j+l+t} \left( \tilde{b}_t-\tilde{b}'_t \right) \le 0.
\end{align*}
Similarly, for any $u \le 5k-j-l-r-1$ one has
\begin{align*}
&-\sum_{j \in I_{r+1}} \sum_{l=1}^{k-j} b_l \sum_{t=4k-j-l}^u q^{n-t} \delta(t-3k+j+l+s) \le 0.
\end{align*}
It follows that
\begin{align} \label{stuff 1}
\begin{split}
&-\sum_{t=3k+1}^n q^{n-t} \sum_{j \in I_{r+1}} \sum_{l=r+1}^{k-j} \left( b_l H(t-2k+j+l)-b'_l H'(t-2k+j+l) \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{l=r+1}^{k-j} b_l \sum_{t=4k-j-l}^n q^{n-t} \delta(t-2k+j+l).
\end{split}
\end{align}
Noting that $r < k-s \le k-j$ for any $j \in I$, one has
\begin{align*}
&-\sum_{t=3k+1}^n q^{n-t} \sum_{j \in I_{r+1}} \sum_{l=1}^r \left( b_l H(t-2k+j+l) - b'_l H'(t-2k+j+l) \right)\\
&=-\sum_{j \in I_{r+1}} \left( \sum_{t=3k+1}^{4k-j-r-1} q^{n-t} \tilde{b}_{4k-t-j-r} - \sum_{l=1}^{r-1} \sum_{t=3k+1}^{4k-j-l-1} q^{n-t} \tilde{b}'_{4k-t-j-l} \right)\\
&-\sum_{j \in I_{r+1}} \left( \sum_{t=4k-j-r}^n q^{n-t} H(t-2k+j+r) - \sum_{l=1}^{r-1} \sum_{t=4k-j-l}^n q^{n-t} H'(t-2k+j+l) \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{l=1}^{r-1} \sum_{t=3k+1}^{4k-j-l-1} q^{n-t} \tilde{b}'_{4k-t-j-l} \\
&-\sum_{j \in I_{r+1}} \sum_{t=4k-j-r}^n q^{n-t} \left( H(t-2k+j+r)-H'(t-2k+j+r) \right).
\end{align*}
It is easy to see that
\begin{align*}
&-\sum_{j \in I_{r+1}} \sum_{t=4k-j-r}^n q^{n-t} \left( H(t-2k+j+r)-H'(t-2k+j+r) \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{t=4k-j-r}^n q^{n-t} \delta(t-2k+j+r).
\end{align*}
We thus have
\begin{align} \label{stuff 2}
\begin{split}
&-\sum_{t=3k+1}^n q^{n-t} \sum_{j \in I_{r+1}} \sum_{l=1}^r \left( b_l H(t-2k+j+l) - b'_l H'(t-2k+j+l) \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{l=1}^{r-1} \sum_{t=3k+1}^{4k-j-l-1} q^{n-t} \tilde{b}'_{4k-t-j-l} + \sum_{j \in I_{r+1}} \sum_{t=4k-j-r}^n q^{n-t} \delta(t-2k+j+r).
\end{split}
\end{align}
Using (\ref{stuff 1}) and (\ref{stuff 2}) we have
\begin{align} \label{stuff 3}
\begin{split}
&-\sum_{t=3k+1}^n \sum_{j \in I_{r+1}} \sum_{l=1}^{k-j} q^{n-t} \left( b_l H(t-2k+j+l)-b'_l H'(t-2k+j+l) \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{l=r}^{k-j} b_l \sum_{t=5k-j-l-s-r}^n q^{n-t} \delta(t-2k+j+l) + \sum_j \sum_{l=1}^{r-1} \sum_{t=4k-j-l-s}^{4k-j-l-1} q^{n-t} \tilde{b}'_{4k-t-j-l}\\
&\le 2 \sum_j \sum_{l=1}^{k-j} q^{n-5k+s+r+j+l} + \sum_j \sum_{l=1}^{r-1} q^{n-4k+s+j+l+1}\\
&\le 2 \sum_j q^{n-4k+s+r+1} + \sum_j q^{n-4k+s+j+r+1} \le q^{k-s-r-1} q^{n-4k+s+r+1}+q^{n-4k+2s+r+2}\\
&\le q^{n-3k}+q^{n-3k+s},
\end{split}
\end{align}
using the fact that $2|I_{r+1}| \le 2(k-s-r-1) \le q^{k-s-r-1}$.
For $2k < n < 3k-r$, using equality (\ref{Delta < 3k-r}) we easily obtain the lower bound
\begin{align*}
\Delta(n) &= -q^{n-2k}\Delta(2k) + \sum_{t=2k}^n q^{n-t} \delta(t) - \sum_{t=2k}^{n-k+s} q^{n-t-k+s} \delta(t)\\
&\ge -2q^{n-2k} + \sum_{t=n-k+s+1}^n q^{n-t} \delta(t).
\end{align*}
If $3k-s-r > n-k+s$ then $\sum_{t=n-k+s+1}^n q^{n-t} \delta(t) \ge 0$. If $3k-s-r \le n-k+s$ then $\sum_{t=2k}^n q^{n-t} \delta(t) \ge 0$ and $-\sum_{t=2k}^{n-k+s} q^{n-t-k+s} \delta(t) \ge -2q^{n-4k+2s+r}$. We thus obtain the lower bound
\begin{align} \label{Delta < 2k}
\Delta(n) \ge -2q^{n-2k}-2q^{n-4k+2s+r}.
\end{align}
Applying (\ref{stuff 3}) and (\ref{Delta < 2k}) to (\ref{last case}) one has
\begin{align*}
&-\sum_{t=3k+1}^n q^{n-t} (H(t)-H'(t)) \le -\sum_{t=3k+1}^n q^{n-t} \Delta(t-k)+\sum_{t=3k+1}^n q^{n-3k+r}+q^{n-3k}+q^{n-3k+s}\\
&\le 2\sum_{t=3k+1}^n q^{n-3k} +2\sum_{t=3k+1}^nq^{n-5k+2s+r}+ (k-r-1) q^{n-3k+r} +q^{n-3k}+q^{n-3k+s}\\
&\le 2(k-r-1)q^{n-3k}+2(k-r-1)q^{n-5k+2s+r} + q^{n-2k-2} + q^{n-3k}+q^{n-3k+s}\\
&\le q^{n-2k-r-1}+q^{n-4k+2s-1} + q^{n-2k-2} + q^{n-3k}+q^{n-3k+s}\\
&\le 2q^{n-2k-2} + q^{n-2k-7}+q^{n-3k}+q^{n-2k-3}.
\end{align*}
One thus has
\begin{align*}
\Delta(n) \le q^{n-2k}(-1+q^{-1}+q^{-7}+q^{-k}+q^{-3}) \le 0.
\end{align*}
Let $4k-r \le n < 4k$. By calculations similar to those above, one has
\begin{align*}
&-\sum_{j \in I_{r+1}} \sum_{l=1}^{k-j} \sum_{t=3k+1}^n q^{n-t} \left( b_l H(t-2k+j+l)-b'_l H'(t-2k+j+l) \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{l=1}^{r-1} \sum_{t=3k+1}^{4k-j-l-1} q^{n-t} \tilde{b}'_{4k-t-j-l} + \sum_{j \in I_{r+1}} \sum_{l=r}^{k-j} b_l \sum_{t=4k-j-l}^{5k-r-j-l-1} q^{n-t} \delta(t-2k+j+l)\\
&-\sum_{j \in I_{r+1}} \sum_{l=r}^{k-j} b_l \sum_{t=5k-r-j-l}^n q^{n-t} \left( H(t-2k+j+l)-H'(t-2k+j+l) \right).
\end{align*}
For $3k-r \le n < 3k$ observe that
\begin{align*}
&H(n)-H'(n) \ge \Delta(n-k)-H(n-k+r)+\sum_{t=r+1}^s ( \delta(n-k+t)-\delta(n-2k+t+s) )\\
&\ge -2 - q^{n-3k+r} + \sum_{t=r+1}^s (\delta(n-k+t)- \delta(n-2k+t+s)).
\end{align*}
It follows that
\begin{align} \label{INE 1}
\begin{split}
&-\sum_{j \in I_{r+1}} \sum_{l=r}^{k-j} b_l \sum_{t=5k-r-j-l}^n q^{n-t} \left( H(t-2k+j+l)-H'(t-2k+j+l \right)\\
&\le \sum_{j \in I_{r+1}} \sum_{l=r}^{k-j} b_l \left( q^{n-5k+r+j+l+2}+(n-5k+r+j+l+1)q^{n-5k+j+l+r} \right.\\
&\left. +2(r-1)q^{n-5k+r+j+l+2}+2q^{n-7k+3s+j+l+r+1} \right)\\
&\le q^{n-3k-s+1}+q^{n-3k+r-s-2}+q^{n-3k+r-s-1}+q^{n-5k+2s+1} \le q^{n-3k}.
\end{split}
\end{align}
One has
\begin{align} \label{INE 2}
\begin{split}
&\sum_{j \in I_{r+1}} \sum_{l=r}^{k-j} b_l \sum_{t=4k-j-l}^{5k-r-j-l-1} q^{n-t} \delta(t-2k+j+l) \le 2\sum_{j \in I_{r+1}} \sum_{l=1}^{k-j} b_l q^{n-5k+j+l+s+r}\\
& \le 2(k-s-r-1) q^{n-4k+s+r+1} \le q^{n-3k}
\end{split}
\end{align}
and
\begin{align} \label{INE 3}
\begin{split}
&\sum_{j \in I_{r+1}} \sum_{l=1}^{r-1} \sum_{t=3k+1}^{4k-j-l-1} q^{n-t} \tilde{b}'_{4k-t-j-l} \le \sum_{j \in I_{r+1}} \sum_{l=1}^{r-1} \sum_{t=4k-s-j-l}^{4k-j-l-1} q^{n-t} b'_{4k-t-j-l}\\
&\le \sum_j \sum_{l=1}^{r-1} q^{n-4k+s+j+l+1} \le \sum_{j \in I_{r+1}} q^{n-4k+s+j+r+1} \le q^{n-4k+2s+r+2} \le q^{n-3k+s}.
\end{split}
\end{align}
Finally, for $3k-r \le n < 3k$, using (\ref{LP2 1}) and (\ref{Delta < 2k}) one has the lower bound
\begin{align} \label{INE 4}
\begin{split}
&\Delta(n) \ge -2 (q^{n-2k}+q^{n-4k+2s+r})\\
&+ \sum_{t=3k-r}^n q^{n-t}\left( H(t-k+r) - \sum_{l=1}^{r-1} H'(t-k+l) - \sum_{l=r+1}^s b_l \delta(t-k+l) \right)\\
&\ge -2 (q^{n-2k}+q^{n-4k+2s+r}) - \sum_{l=1}^{r-1} \sum_{t=3k-r}^{3k-l-1} q^{n-t} \tilde{b}'_{3k-t-l} - \sum_{l=r}^s \sum_{t=3k-r}^n b_l q^{n-t} \delta(t-k+l)\\
&\ge -2 (q^{n-2k}+q^{n-4k+2s+r}) - \sum_{l=1}^{r-1} q^{n-3k+r+1} - 2\sum_{l=r}^s q^{n-4k+s+r+l}\\
&\ge -2 (q^{n-2k}+q^{n-4k+2s+r}) - q^{n-3k+2r-1} - q^{n-4k+2s+r+2}.
\end{split}
\end{align}
Applying (\ref{INE 1}) through (\ref{INE 4}) to (\ref{last case}), for $4k-r \le n < 4k$ one thus has
\begin{align*}
&\Delta(n) \le -q^{n-2k} + 2 \sum_{t=3k+1}^n (q^{n-3k} + q^{n-5k+2s+r}) + \sum_{t=4k-r}^n (q^{n-4k+2r-1} + q^{n-5k+2s+r+2})\\
&+\sum_{t=3k+1}^n q^{n-t} H(t-k+r)+2q^{n-3k}+q^{n-3k+s}\\
&\le -q^{n-2k} + 2(k-1)(q^{n-3k}+q^{n-4k+s-2})+(r-1)(q^{n-3k-5}+q^{n-4k+s})\\
&+(k-1)q^{n-3k+r} + 2q^{n-3k}+q^{n-3k+s}\\
&\le q^{n-2k}(-1+q^{-2}+q^{-7}+q^{-11}+q^{-k-4}+q^{-1}+2q^{-k}+q^{-3}) < 0
\end{align*}
where we have used (\ref{LP 1.5}) and the inequalities $k \ge 5$, $(k-1) \le q^{k-3}$, and $k-1 \le q^{k/2}$.
It follows that $\tilde{\Delta}(n) < 0$ for $n < 4k$. In combination with our results from Section 6, one has
\begin{align*}
&\Delta(n)
< 0 \textrm{ for } n<4k \textrm{ when } s=s' \textrm{ and } k=k'.
\end{align*}
\section*{Acknowledgments}
This work was partially supported by the NSF grant DMS-1600568.
\section*{References}
|
\section{Introduction}\label{s-intro}
\subsection{The vertex enumeraion problem}
The well-known Minkowski-Weyl theorem states that any convex polyhedron $\mathcal P\subseteq\mathbb R^n$
can be represented as the Minkowski sum of the convex hull
of the set $\mathcal V(\mathcal P)$ of its extreme points and the conic hull of the set $\mathcal D(\mathcal P)$
of its extreme directions (see e.g. \cite{S86}).
Given a polyhedron $\mathcal P$ by its linear description
as the intersection of finitely many halfspaces, obtaining the set $\mathcal V(\mathcal P)\cup\mathcal D(\mathcal P)$,
required by the other representation,
is a well-known problem, called \emph{Vertex Enumeration} (VE) (see,. e.g.,\cite{Dye83,BFM98}), which have been extensively
studied in the literature in different (but polynomially equivalent) forms, e.g.,
, the \emph{facet enumeration} problem \cite{BFM98}
or the \emph{polytope-polyhedron problem} \cite{Lov92}.
Clearly, the size of the extreme set $\mathcal V(\mathcal P)\cup\mathcal D(\mathcal P)$ can be (and typically is)
exponential in the dimension $n$ and the number of linear inequalities $m$, and thus when
considering the computational complexity of the vertex enumeration problem,
one is usually interested in \emph{output-sensitive} algorithms \cite{Sei86}, i.e.,
those whose running time depends not only on $n$ and $m$, but also on $|\mathcal V(\mathcal P)\cup\mathcal D(\mathcal P)|$.
Alternatively, we may consider the following, polynomially equivalent, decision variant of the problem:
\begin{description}
\item [\VD{\mathcal L}{\mathcal C(\mathcal P)}:] Given a polyhedron $\mathcal P$, represented by a system of linear inequalities $\mathcal L$,
and a subset $\mathcal X\subseteq\mathcal C(\mathcal P)$, is $\mathcal X=\mathcal C(\mathcal P)$?
\end{description}
In this description, $\mathcal C(\mathcal P)$ could be either $\mathcal V(\mathcal P)$, $\mathcal D(\mathcal P)$, or $\mathcal V(P)\cup\mathcal D(P)$.
The problem of enumerating the elements of $\mathcal C(\mathcal P)$ is said to be solvable in {\it incremental polynomial} time if problem \VD{\mathcal L}{\mathcal C(\mathcal P)} can be solved in time polynomial in the size of the description of $\mathcal L$ and $\mathcal X$.\footnote{Note that if the answer to the decision problem is ``NO" then a new element in $\mathcal C(\mathcal P)\setminus\mathcal X$ can be found by a polynomial number of calls to the decision problem.}
It is
well-known that if the decision problem is NP-hard, then no {\it output} (or {\it total}) {\it polynomial-time} algorithm
can generate the elements of $\mathcal C(\mathcal P)$ unless P=NP (see e.g. \cite{BEGM09}).
Vertex enumeration is an outstanding open problem in computational geometry and polyhedral combinatorics
(see, e.g., \cite{DP77,Lov92,Pfe02}), and has numerous
applications. For example, understanding the structure
of the vertices helps in designing approximation algorithms for
combinatorial optimization problems \cite{V01}; finding all vertices
can be used for computing Nash equilibria for bimatrix games
\cite{ARSS10}.
Numerous algorithmic ideas for vertex or facet enumeration have been introduced in the literature, see, e.g.,
\cite{Abd03,ABS97,AF92,AF96,BFM98,BL98,Chv83,DP77,Dye83,FP96,Pro94,RS14,Sei86}.
\medskip
The main result in \cite{KBBEG08} established that problem
\VD{\mathcal L}{\mathcal V(\mathcal P)} is NP-hard for unbounded polyhedra,
more precisely, when $|\mathcal D(\mathcal P)|$ is exponentially large in the input
size. This negative result holds, even when restricted to $0/1$-{\it
polyhedra} \cite{BEGT11}, that is, when
$\mathcal V(\mathcal P)\subseteq\{0,1\}^n$, and comes in contrast with the fact that
the VE problem for $0/1$-polytopes (i.e., bounded polyhedra) is known to be solvable with
{\it polynomial delay} (that is, the vertices are generated such that the
delay between any successive outputs is polynomial {\it only} in the {\it input}
size) and {\it polynomial space} (that is, the total space used for enumerating all the vertices is polynomial in the {\it input}
size).
\subsection{VE for $0/1$-polyhedra associated with $0/1$-totally unimodular matrices}
\medskip
Let $A\in\{0,1\}^{m\times n}$ be an $m \times n$ $0/1$-matrix such that the
polyhedron
\begin{equation}\label{P}
\mathcal P(A,\b1)=\{x\in\mathbb R^n \mid Ax\geq \b1,~x\geq \b0\}
\end{equation}
has only integral vertices, where $\b1$ (resp., $\b0$) denotes the vector of all ones (resp., zeros) of appropriate dimension.
Then $\mathcal P(A,\b1)$ has only $n$ extreme directions (namely the $n$ unit vectors in $\mathbb R^n$), while the vertices of $\mathcal P(A,\b1)$ are in one-to-one correspondence with the minimal transversals of the hypergraph $\mathcal H[A] \subseteq 2^{[n]}$, whose characteristic vectors of hyperedges are the rows of $A$. One of the most important examples is when the matrix $A$ is \emph{totally unimodular}: in this case, the polyhedron $\mathcal P(A,\b1)$ has integral vertices, and VE is equivalent to finding all minimal transversals\footnote{Note that, it is not possible to reduce the problem of enumerating the vertices of $\mathcal P(A,\b1)$ to that of enumerating the vertices of the $0/1$ polytope $\mathcal P'=\{x\in\mathbb R^n \mid Ax\geq \b1,~\b0\leq x\leq\b1\}$, as $\mathcal P'$ can have exponentially more vertcies than those of $\mathcal P$ (namely, the vertices of $\mathcal P'$ are the (not necessarily minimal) transversals of $\mathcal H[A]$).} of a \emph{unimodular hypergraph} $\mathcal H[A]$. Consequently, it follows from the well-known result in \cite{FK96} that all vetrices of such polyhedra can be enumerated in {\it quasi-polynomial} time, and hence the VE problem in this case is unlikely to be NP-hard. Polynomial time algorithms for special cases of this problem are known; for example, enumerating minimal vertex/edge covers in bipartite graphs \cite{EG95,MP97}, enumerating minimal hitting sets/set covers of interval hypergraphs~\cite{BEGM09}, and enumerating minimal path covers/cut conjunctions in directed trees \cite{BEGM09}. However, the complexity of the VE problem for \raf{P} remains open, even for the totally unimodular matrices $A$. In this paper, we settle the complexity of the VE problem in the latter case.
\begin{theorem}\label{t-main}
Let $A\in\{0,1\}^{m\times n}$ be a totally unimodular matrix. Then the vertices of $\mathcal P(A,\b1)$ can be enumerated in incremental polynomial time.
\end{theorem}
A celebrated result of Seymour \cite{S80} shows that any totally unimodular matrix (with $0$, $\pm1$-entries) arises from (essentially) the so-called {\it network matrices}, by a small set of simple operations.
Similar results for $0/1$-totally unimodular matrices are derived in \cite[Chapter 11]{T92}, with the main building blocks replaced by $0/1$-network matrices.
On the other hand, it has been shown in \cite{BEGM09} that for any polyhedron $\mathcal P(A,\b1)$, with a $0/1$-network matrix $A$ or its transpose, the VE problem can be solved in incremental polynomial time. To prove Theorem \ref{t-main}, we show that the above mentioned decomposition of totally unimodular matrices yields a corresponding decomposition for the hypergraph transversal problem, that can be leveraged into a polynomial time algorithm for the enumeration problem. One of the natural ways to use such decomposition is to recursively partition the input polyhedron into two smaller polyhedra and then combine the outputs from the two subproblems. While such approach works for the simple cases of the decomposition (so-called $1$- and $2$-sum decompositions), it does not work for the more complicated case (so-called $3$-sum decomposition). The main reason is that the number of vertices of the resulting polyhedra in either of the two subproblems may be exponentially larger than that in the original problem. To overcome this difficulty, we need to use the decomposition in a more sophisticated way, utilizing structural properties of the unimodular hypergraph $\mathcal H[A]$. One technical hurdle which arises is that the total input/output size of the resulting subproblems might exceed the input/output size of the original problem, which may eventually lead to an exponential blow-up in the overall running time of the algorithm in terms of the input {\it and} output sizes. To deal with this issue, we introduce a {\it volume} measure as the product of the input and output sizes, and show in each case of our decomposition that the total measure of the subproblems obtained is smaller than the measure of the original problem.
\section{Notation and preliminaries}\label{prelim}
\subsection{Hypergraphs and transversals}
Let $V$ be a finite set. A hypergraph $\mathcal H \subseteq 2^V$ is a family of subsets of $V$. A hypergraph is called {\it Sperner} ({\it simple} or a {\it clutter}), if it has the property that no hyperedge contains another.
For a hypergraph $\mathcal H\subseteq 2^V$, we denote by $\operatorname{Tr}(\mathcal H)$ the family of {\it minimal transversals} of $\mathcal H$, i.e., (inclusion-wise) minimal subsets of $V$ which have a nonempty intersection with each hyperedge of $\mathcal H$; $\operatorname{Tr}(\mathcal H)$ is also called the {\it dual} of $\mathcal H$.
We say that the hypergraph $\mathcal H$ is {\it trivial} if $\mathcal H=\emptyset$ or $\mathcal H=\{\emptyset\}$, and is {\it irredundant} if every $v\in V$ belongs to some $H\in\mathcal H$.
As usual, we assume $\operatorname{Tr}(\{\emptyset\})=\emptyset$ and $\operatorname{Tr}(\emptyset)=\{\emptyset\}$.
Given two hypergraphs $\mathcal H_1$ and $\mathcal H_2$ with vertex set $V$, denote by
\begin{align*}
\mathcal H_1 \wedge \mathcal H_2 &= \operatorname{Min} \{ H_1 \cup H_2~ |~ H_1 \in \mathcal H_1 \text{ and } H_2 \in \mathcal H_2 \},\\
\mathcal H_1 \vee \mathcal H_2 &= \operatorname{Min} (\mathcal H_1 \cup \mathcal H_2),
\end{align*}
the {\it conjunction} and {\it disjunction} of $\mathcal H_1$ and $\mathcal H_2$ respectively, where for hypergraph $\mathcal H$, $\operatorname{Min}(\mathcal H)$ denotes the family of (inclusion-wise) minimal sets in $\mathcal H$. We denote by $\mathcal H_1\dot{\cup}\mathcal H_2$ the {\it disjoint} union of $\mathcal H_1$ and $\mathcal H_2$. For two hypergraphs $\mathcal H_1\subseteq 2^{V_1}$ and $\mathcal H_2\subseteq 2^{V_2}$, we denote by $\mathcal H_1\dot{\wedge}\mathcal H_2$ the conjunction of $\mathcal H_1$ and $\mathcal H_2$ when $V_1$ and $V_2$ are {\it disjoint}. By definition, $|\mathcal H_1\dot{\cup}\mathcal H_2|=|\mathcal H_1|+|\mathcal H_2|$ and $|\mathcal H_1\dot{\wedge}\mathcal H_2|=|\mathcal H_1|\cdot|\mathcal H_2|$.
For a hypergraph $\mathcal H\subseteq 2^V$ and a set $S\subseteq V$, we denote by $\mathcal H_S=\{H\in\mathcal H~ |~ H\subseteq S\}$ and $\mathcal H^S=\operatorname{Min}\{H\cap S~ |~ H\in\mathcal H\}$
the subhypergraph of $\mathcal H$ {\it induced by} $S$, and the {\it projection} of $\mathcal H$ on $S$, respectively. For $W,S\subseteq V$, we write $\mathcal H(W,S)=\{H\in\mathcal H\mid H\cap W=S\}$. Two vertices of $\mathcal H$ are said to be {\it identical} if they belong to exactly the same hyperedges, i.e., the corresponding columns in the hyperedge-vertex incidence matrix are identical.
The following propositions are straightforward (see e.g. \cite{B89,EGM03,LLK80}).
\begin{proposition}\label{prop1}
Given a hypergraph $\mathcal H\subseteq 2^V$ and a set $S\subseteq V$, the following statements hold:
\begin{itemize}
\item[(i)] $\operatorname{Tr}(\operatorname{Tr}(\mathcal H))=\operatorname{Min}(\mathcal H)$,
\item[(ii)] $ \operatorname{Tr}(\mathcal H_S) = \operatorname{Tr}(\mathcal H)^S $ (and hence, $ \operatorname{Tr}(\mathcal H^S) = \operatorname{Tr}(\mathcal H)_S $) and
\item[(iii)] $ |\operatorname{Tr}(\mathcal H_S)| \leq |\operatorname{Tr}(\mathcal H)| $.
\end{itemize}
\end{proposition}
\begin{proposition}\label{prop2}
Given hypergraphs $\mathcal H_1,\ldots,\mathcal H_k \subseteq 2^V$,
\[ \operatorname{Tr}\left(\bigvee_{i=1}^r\mathcal H_i\right)= \bigwedge_{i=1}^r\operatorname{Tr}(\mathcal H_i). \]
\end{proposition}
As a corollary of Proposition~\ref{prop2} we have the following.
\begin{proposition}\label{p2}
Let $\mathcal H\subseteq 2^V$ be a hypergraph and $S_1,\ldots,S_r\subseteq V$
be subsets such that for every hyperhedge $H\in\mathcal H$ there exists an
$i\in\{1,\ldots,r\}$ with $H\subseteq S_i$. Then
\begin{equation*}\label{conj}
\operatorname{Tr}(\mathcal H)=\bigwedge_{i=1}^r\operatorname{Tr}(\mathcal H_{S_i}).
\end{equation*}
\end{proposition}
Throughout the paper, we use the notation: $n=n(\mathcal H)=|V|$, $m=m(\mathcal H)=|\mathcal H|$ and $k=k(\mathcal H)=|\operatorname{Tr}(\mathcal H)|$.
\subsection{Polyhedra}
A convex polyhedron $P\subseteq\mathbb R^n$ is the intersection of finitely many halfspaces, determined by the \emph{facets} of the polyhedron.
A \emph{vertex} or an \emph{extreme point} of $P$ is a point $v\in\mathbb R^n$ which cannot be represented as a convex combination of two other points of $P$, i.e., there exists no $\lambda\in(0,1)$ and $v_1,v_2\in P$ such that $v=\lambda v_1+(1-\lambda)v_2$. A \emph{(recession) direction} of $P$ is a vector $d\in\mathbb R^n$ such that $x_0+\mu d\in P$ whenever $x_0\in P$ and $\mu\ge 0$. An \emph{extreme direction} of $P$ is a direction $d$ that cannot be written as a conic combination of two other directions, i.e.,
there exist no positive real numbers $\mu_1,\mu_2\in\mathbb R_+$ and directions $d_1,d_2$ of $P$ such that $d=\mu_1 d_1+\mu_2 d_2$.
Denote respectively by $\mathcal V(P)$ and $\mathcal D(P)$ the sets of extreme points and extreme directions of polyhedron $P$.
A bounded polyhedron, i.e., one for which $\mathcal D(P)=\emptyset$ is called a \emph{polytope}.
\subsection{Totally unimodular matrices}
A matrix $A\in\{0,1\}^{m\times n}$ is {\it totally unimodular} if every square subdeterminant of it has value in $\{-1,0,1\}$.
We denote by $\mathcal U^{m\times n}$ the set of $m\times n$ $0/1$-totally unimodular matrices. For a matrix $A\in\{0,1\}^{m\times n}$ we denote by $\mathcal H[A] \subseteq 2^{[n]}$ the hypergraph whose characteristic vectors of hyperedges are the rows of $A$.
A hypergraph $\mathcal H$ is said to be {\it unimodular} \cite{B89} if $\mathcal H=\mathcal H[A]$ for a totally unimodular matrix $A$. Note by definition that if $\mathcal H\subseteq 2^V$ is unimodular then for any set $S\subseteq V$ and any subhypergraph $\mathcal H'\subseteq\mathcal H$, the hypergraph $(\mathcal H')^S$ is unimodular.
A $0/1$ matrix is said to be {\it ideal} (see, e.g., \cite{C01}) if the
polyhedron $P=P(A,\b1)$ has only integral vertices. It is well-known that every totally unimodular matrix $A\in\{0,1\}^{m\times n}$ is ideal. Furthermore, the following correspondence holds.
\begin{proposition}[\cite{L65}]\label{L}
Let $A$ be an $m \times n$ ideal matrix.
Then the vertices of the polyhedron $\mathcal P(A,\b1)$ are in one-to-one correspondence with the minimal transversals of the hypergraph $\mathcal H[A]$.
\end{proposition}
As a corollary of Theorem~\ref{t-main}, we obtain the following result.
\begin{corollary}\label{cor1}
Let $A\in\{0,1\}^{m\times n}$ be a totally unimodular matrix and, $A'$ be a $0/1$ matrix whose rows are the characteristic vectors of the vertices of $\mathcal P(A,\b1)$. Then the vertices of $\mathcal P(A',\b1)$ can be enumerated in incremental polynomial time.
\end{corollary}
\noindent {\bf Proof}.~~
By a result of Lehman \cite{L65}, the polyhedron $\mathcal P(A',\b1)$ also is $0/1$, whenever $A'$ is a $0/1$ matrix whose rows are the characteristic vectors of the vertices of $\mathcal P(A,\b1)$ for an deal matrix $A$. It follows that the vertices of $\mathcal P(A',\b1)$ are in one-to-one correspondence with the minimal transversals of $\mathcal H[A']=\operatorname{Tr}(\mathcal H[A])$. By the polynomial equivalence of the enumeration and decision problems~\cite{BI95,GK99}, it is enough to check, for a given sublist $\mathcal X\subseteq\operatorname{Tr}(\mathcal H[A'])$, whether $\mathcal X=\operatorname{Tr}(\mathcal H[A'])$. The latter condition is equivalent to $\operatorname{Tr}(\mathcal X)=\mathcal H[A']$, which can be checked in polynomial time by Theorem~\ref{t-main}, since $\mathcal X$ is a unimodular hypergrpah by assumption.
\hfill$\square$\bigskip
\subsection{$0/1$-Network matrices}
A matrix $A\in\mathcal U^{m\times n}$ is said to be a {\it network} matrix if
there exists a directed tree\footnote{We say that a directed graph $G$ is a {\em directed tree} if
the underlying graph of $G$ (i.e., the undirected graph obtained from $G$ by ignoring
orientation of arcs) is a tree.} $T$ such that the
rows of $A$ one-to-one correspond to the arcs in $T$, and each
column of $A$ is the characteristic vector of a directed path in
$T$. Checking if a given matrix $A$ is a network matrix and finding the corresponding tree representation can be done in polynomial
time (see e.g., \cite{S86}). We call a hypergraph $\mathcal H$ a {\it network hypergraph} if $\mathcal H=\mathcal H[A]$ for some network matrix $A$ or its transpose. It is known that network hypergraphs can be dualized in incremental polynomial time and polynomial space:
\begin{theorem}[\cite{BEGM09}]\label{t-network}
Let $A\in\{0,1\}^{m\times n}$ be a network matrix. Then
\begin{description}
\item[\hspace*{.21cm}{\rm (i)}] all the
vertices of $\mathcal P(A,\b1)$ can be enumerated in incremental polynomial time
using polynomial space;
\item[\hspace*{.21cm}{\rm (ii)}] all the vertices of
$\mathcal P(A^T,\b1)$ can be enumerated in incremental polynomial time
using polynomial space.
\end{description}
\end{theorem}
\subsection{Decomposition of $0/1$-totally unimodular matrices}
Seymour \cite{S80} gave a decomposition theorem that allows one to decompose (in polynomial time) any $0/1$-totally unimodular matrix by repeatedly applying certain operations (called {\it $i$-sums}, for $i=1,2,3$) until simple building blocks are obtained; the building blocks consist of $0/1$-network matrices, their transposes and a specific $5\times 5$ $0/1$-matrix.
For our purposes this theorem can be stated as follows.
\begin{theorem}[\cite{S80,S86,T92}]\label{t-decomp}
Let $A\in\{0,1\}^{m\times n}$ be a totally unimodular matrix. Then one of the following conditions holds:
\begin{itemize}
\item[(I)] $A$ or its transpose is a $0/1$-network matrix;
\item [(II)] $A$, possibly after permuting the rows and columns, is the matrix:
\begin{equation*}
A_0=\left[\begin{array}{lllll}
1 & 0 & 0 & 1 & 1\\
1 & 1 & 0 & 0 & 1\\
0 & 1 & 1 & 0 & 1\\
0 & 0 & 1 & 1 & 1\\
1 & 1 & 1 & 1 & 1
\end{array}\right];
\end{equation*}
\item[(III)] $A$ has a row or a column with at most one ``{\rm 1"}, or $A$ has two identical rows or columns;
\item [(IV)] $A$ can be decomposed as follows:
\begin{equation*}
A=\left[\begin{array}{c|c}
A^1 & Q\\
\hline
R & A^2
\end{array}\right],
\end{equation*}
where $A^1\in\mathcal U^{m_1\times n_1}$, $A^2\in\mathcal U^{m_2\times n_2},$ $Q\in\mathcal U^{m_1\times n_2}$, $R\in\mathcal U^{m_2\times n_1}$, for some positive integers $n_1,m_1,n_2,m_2$ with $m_1+m_2=m$, $n_1+n_2=n$, $n_1+m_1\ge 4$ and $n_2+m_2\ge 4$:
\begin{itemize}
\item $1$-sum decomposition: $R=\mathbf 0$ and $Q=\mathbf 0$;
\item $2$-sum decomposition: $R$ has rank $1$ and $Q=\mathbf 0$;
\item $3$-sum decomposition -- case 1: $R$ and $Q$ have rank $1$;
\item $3$-sum decomposition -- case 2: $R$ has rank $2$, $Q=\mathbf 0$.
\end{itemize}
\end{itemize}
Deciding if $A$ is a network matrix or its transpose, or (up to permutations of rows and columns) $A_0$, and if not, finding a decomposition as above can be done in polynomial time.
\end{theorem}
\begin{corollary}[Decomposition of unimodular hyeprgraphs]\label{c-decomp}
Let $\mathcal H\subseteq 2^V$ be a unimodular (nontrivial) irredundant Sperner hypergraph. Then $\mathcal H$ is a network hypergraph, (isomorphic to) the hypergraph $\mathcal H_0=\{\{1,4,5\},\{1,2,5\},\{2,3,5\},\{3,4,5\}\}$, has two identical vertices, has a hyperedge consisting of a singleton, has a vertex with degree $1$, or there exists a nontrivial partition $V_1\dot\cup V_2=V$ such that $\mathcal H$ can be decomposed as follows
\begin{itemize}
\item $1$-sum decomposition:
\begin{itemize}
\item[(i)] $\mathcal H_{V_1}\neq\emptyset$, $\mathcal H_{V_2}\neq\emptyset$;
\item[(ii)] for all $H\in\mathcal H$: either $H\subseteq V_1$ or $H\subseteq V_2$;
\end{itemize}
\item $2$-sum decomposition: there exists a set nonempty $S\subseteq V_1$ such that
\begin{itemize}
\item[(i)] $\mathcal H_{V_1}\neq\emptyset$, $\mathcal H(V_1,S)\neq\emptyset$, $\mathcal H(V_1,S)^{V_2}\neq\{\emptyset\}$;
\item[(ii)] for all $H\in\mathcal H$ with $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$: $H\cap V_1=S$;
\end{itemize}
\item $3$-sum decomposition -- case 1: there exist two nonempty sets $S_1\subseteq V_1$ and $S_2\subseteq V_2$, such that
\begin{itemize}
\item[(i)] $\mathcal H(V_1,S_1)\neq\emptyset$, $\mathcal H(V_1,S_1)^{V_2}\neq\{\emptyset\}$, $\mathcal H(V_2,S_2)\neq\emptyset$, $\mathcal H(V_2,S_2)^{V_1}\neq\{\emptyset\}$;
\item[(ii)] $|V_1|+|\mathcal H_{V_1}\cup \mathcal H(V_2,S_2)|\ge 4$, $|V_2|+|\mathcal H_{V_2}\cup \mathcal H(V_1,S_1)|\ge 4$;
\item[(iii)] for all $H\in\mathcal H$ with $H\cap V_1\not=\emptyset$ and $H\cap V_2\not=\emptyset$: either $H\cap V_1= S_1$ or $H\cap V_2=S_2$;
\end{itemize}
\item $3$-sum decomposition -- case 2: there exist three nonempty disjoint sets $S_0,S_1,S_2\subseteq V_1$, such that
\begin{itemize}
\item[(i)] $\mathcal H_{V_1}\neq\emptyset$, $\mathcal H(V_1,S_0\cup S_1)\neq\emptyset$, $ \mathcal H(V_1,S_0\cup S_1)^{V_2}\neq\{\emptyset\}$, $\mathcal H(V_1,S_0\cup S_2)\neq\emptyset$, $ \mathcal H(V_1,S_0\cup S_2)^{V_2}\neq\{\emptyset\}$;
\item[(ii)] for all $H\in\mathcal H$ with $H\cap V_1\neq\emptyset$ and $H\cap V_2\not=\emptyset$: either $H\cap V_1=S_0\cup S_1$, or $H\cap V_1= S_0\cup S_2$;
\end{itemize}
\item $3$-sum decomposition -- case 3: there exist two nonempty disjoint sets $S_1,S_2\subseteq V_1$, such that
\begin{itemize}
\item [(i)] $\mathcal H_{V_1}\neq\emptyset$ and at least two of the following three conditions hold: (1) $\mathcal H(V_1,S_1)\neq\emptyset$ and $\mathcal H(V_1,S_1)^{V_2}\neq\{\emptyset\}$, (2) $\mathcal H(V_1,S_2)\neq\emptyset$, $\mathcal H(V_1,S_2)^{V_2}\neq\{\emptyset\}$, (3)
$\mathcal H(V_1,S_1\cup S_2)\neq\emptyset$, $\mathcal H(V_1,S_1\cup S_2)^{V_2}\neq\{\emptyset\}$;
\item[(ii)] $|V_1|+|\mathcal H_{V_1}|\ge4$, $|V_2|+|\mathcal H_{V_2}|\cup \mathcal H(V_1,S_1)\cup\mathcal H(V_1,S_2)\cup \mathcal H(V_1,S_1\cup S_2)|\geq 4$;
\item [(iii)] for all $H\in\mathcal H$ with $H\cap V_1\neq\emptyset$ and $H\cap V_2\not=\emptyset$: either $H\cap V_1=S_1$, $H\cap V_1=S_2$, or $H\cap V_1= S_1\cup S_2$.
\end{itemize}
\end{itemize}
Discovering if $\mathcal H$ is a network hypergraph, or isomorphic to $\mathcal H_0$, and if not finding a decomposition as above can be done in polynomial time.
\end{corollary}
\noindent {\bf Proof}.~~
We consider only the $2$-sum and $3$-sum decomposition rules as
the other cases are immediate from the corresponding cases in Theorem~\ref{t-decomp}. We define $V_1$ and $V_2$ to be the sets of columns of the matrices $A^1$ and $A^2$ in Theorem~\ref{t-decomp}. Consider the $2$-sum case. Since $R$ has rank $1$, each nonzero row in $R$ is a copy of some row vector $e\in\{0,1\}^{V_1}$. Let $S=\{i\in V_1\mid e_i=1\}$. Then by the decomposition of the matrix $A$ in this case, all hyperedges $H\in\mathcal H[A]$ such that $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$, have $H\cap V_1=S$.
Consider next the $3$-sum--case 1 in Theorem~\ref{t-decomp}. Since $R$ and $Q$ have rank 1, each nonzero row in $R$ (resp., $Q$) is a copy of some row vector $e\in\{0,1\}^{V_1}$ (resp. $f\in\{0,1\}^{V_2}$). Let $S_1=\{i\in V_1\mid e_i=1\}$ and $S_2=\{i\in V_2\mid f_i=1\}$. Then by the decomposition of the matrix $A$ in this case, all hyperedges $H\in\mathcal H[A]$ such that $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$, have either $H\cap V_1=S_1$ or $H\cap V_2=S_2$.
Consider finally the $3$-sum--case 2 in Theorem~\ref{t-decomp}. Since $R$ has rank 2, there exist two rows $a,b\in\{0,1\}^{V_1}$ of $R$ that are linearly independent, such that any other row $c\in\{0,1\}^{V_1}$ of $R$ can be written as $c=\lambda_1a+\lambda_2b$, where $\lambda_1,\lambda_2\in\mathbb R$. We consider three subcases:
\begin{itemize}
\item [(I)] $a^Tb\ne 0$ and neither $a\ge b$ nor $b\ge a$: then $(\lambda_1,\lambda_2)\in\{(0,0),(1,0),(0,1)\}$. Let $S_0=\{i\in V_1\mid a_i=1\text{ and } b_i=1\}$, $S_1=\{i\in V_1\mid a_i=1\text{ and } b_i=0\}$, and $S_2=\{i\in V_1\mid a_i=0\text{ and } b_i=1\}$. Then by the decomposition of the matrix $A$ in this case, all hyperedges $H\in\mathcal H[A]$ such that $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$, have either $H\cap V_1=S_0\cup S_1$, or $H\cap V_1=S_0\cup S_2$. This corresponds to case 2 of the $3$-sum decomposition in the corollary.
\item [(II)] $a^Tb\ne 0$ and $a\ge b$ (resp., $b\ge a$): then $(\lambda_1,\lambda_2)\in\{(0,0),(1,0),(0,1),(1,-1)\}$ (resp., $(\lambda_1,\lambda_2)\in\{(0,0),(1,0),(0,1),(-1,1)\}$). Let $S_1=\{i\in V_1\mid a_i=1\text{ and } b_i=1\}$, and $S_2=\{i\in V_1\mid a_i=1\text{ and } b_i=0\}$ (resp., $S_2=\{i\in V_1\mid a_i=0\text{ and } b_i=1\}$). Then by the decomposition of the matrix $A$ in this case, all hyperedges $H\in\mathcal H[A]$ such that $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$, have either $H\cap V_1=S_1$, $H\cap V_1=S_2$, or $H\cap V_1=S_1\cup S_2$. This corresponds to case 3 of the $3$-sum decomposition in the corollary.
\item [(III)] $a^Tb=0$: then $\lambda_1,\lambda_2\in\{0,1\}$. Let $S_1=\{i\in V_1\mid a_i=1\text{ and } b_i=0\}$, and $S_2=\{i\in V_1\mid a_i=0\text{ and } b_i=1\}$. Then by the decomposition of the matrix $A$ in this case, all hyperedges $H\in\mathcal H[A]$ such that $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$, have either $H\cap V_1=S_1$, $H\cap V_1=S_2$, or $H\cap V_1=S_1\cup S_2$. This corresponds to case 3 of the $3$-sum decomposition in the corollary.
\end{itemize}
Finally, we verify the boundary conditions. For the $1$-sum case, $\mathcal H_{V_1},\mathcal H_{V_2}$ are nonempty because of the requirement that $m_1,m_2\ge 1$ in Theorem~\ref{t-decomp}. Consider next the $2$-sum case. $\mathcal H_{V_1}\neq\emptyset$ follows from the requirement that $m_1\ge 1$ in Theorem~\ref{t-decomp}. If $\mathcal H(V_1,S)^{V_2}=\{\emptyset\}$ then, by the simplicity of $\mathcal H$, we would have $\mathcal H(V_1,S)=\{S\}\in\mathcal H_{V_1}$, which would imply that we are also in the $1$-sum case. Consider next the $3$-sum case 1. $\mathcal H(V_1,S_1)\ne\emptyset$ and $\mathcal H(V_2,S_2)\ne\emptyset$ follow, respectively, from the requirements that $m_1\ge 1$ and $m_2\ge 1$ in Theorem~\ref{t-decomp}. If either $\mathcal H(V_1,S_1)^{V_2}=\{\emptyset\}$ or $\mathcal H(V_2,S_2)^{V_1}=\{\emptyset\}$ then, again by the simplicity of $\mathcal H$, we are in the $1$-sum or the $2$-sum cases.
Next consider the $3$-sum--case 2. Note that $\mathcal H_{V_1}\neq\emptyset$, $\mathcal H(V_1,S_0\cup S_1)\neq\emptyset$ and $\mathcal H(V_1,S_0\cup S_2)\neq\emptyset$ in this case. If either $\mathcal H(V_1,S_0\cup S_1)^{V_2}=\{\emptyset\}$ (resp., $\mathcal H(V_1,S_0\cup S_2)^{V_2}=\{\emptyset\}$) then by the simplicity of $\mathcal H$, we would have $\mathcal H(V_1,S_0\cup S_1)=\{S_0\cup S_1\}\in\mathcal H_{V_1}$ (resp., $\mathcal H(V_1,S_0\cup S_2)=\{S_0\cup S_2\}\in\mathcal H_{V_1}$), implying that we are in either the $1$-sum or the $2-$sum cases. A similar argument applies for $3$-sum--case 3.
\hfill$\square$\bigskip
A schematic illustration of these decomposition rules is given in Figures~\ref{f1} and~\ref{f2}.
\begin{remark}\label{r1}
We note that the boundary condition (ii) in the $3$-sum--case 1 is essential, since without insisting on this condition, any hypergraph can be decomposed according to the $3$-sum--case 1 rule (take any $v\in V$ and $H\in\mathcal H$ such that $v\in H$, and let $V_1=S_1=\{v\}$, $V_2=V\setminus\{v\}$ and $S_2=H\setminus\{v\}$). Similarly, our analysis in the $3$-sum--case 3 uses condition (ii). However, a similar condition is not needed for all other cases.
\end{remark}
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.5]{1-2-sum}
\caption{Decomposing a unimodular hypergraph: $1$ and $2$-sums.}
\label{f1}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.5]{3-sum}
\caption{Decomposing a unimodular hypergraph: $3$-sum.}
\label{f2}
\end{figure}
\section{Decomposition of the hypergraph transversal problem}
In the following, we show how to decompose the hypergraph transversal problem for a unimodular hypergraph $\mathcal H$, given the decomposition of $\mathcal H$ as in Corollary~\ref{c-decomp}. Such a decomposition yields naturally a recursive algorithm: each non-leaf node of the recursion tree is responsible for computing the dual of a unimodular hypergraph, while leaves involve the computation of the dual of a network hypergraph or the hypergraph $\mathcal H_0$.
To ensure that the overall running time is polynomial, we need to bound the number of nodes of the recursion tree and the local computation time at each node, which consists of the time required for computing the decomposition and the time for combining the outputs from the recursive calls into the final output at the node. We will measure the ``volume" of each subproblem to compute $\operatorname{Tr}(\mathcal H)$ by $\mu(\mathcal H)=nmk=n(\mathcal H)m(\mathcal H)k(\mathcal H)$. We let $T(\mu)$ be the number of nodes of the recursion subtree rooted at a node of volume $\mu$, and let $L_1(\mu)$ and $L_2(\mu)$ be respectively the local computation time for the decomposition and combining the outputs at a node of volume $\mu$. We
stop the recursion when either $m=m(\mathcal H)$, $n=n(\mathcal H)$ or $k=k(\mathcal H)$ drops below some constant $C$, in which case the hypergraph transversal problem can be solved in $\operatorname{poly}(n,m)$ time using a simple procedure, such as {\it Berge Multiplication}~\cite[Chapter 2]{B89} for $n(\mathcal H)\le C$ or $n(\mathcal H)\le C$, and the methods in \cite{BI95,GK99} for $k(\mathcal H)\le C$, which also show that the condition $k(\mathcal H)\le C$ can be checked in $\operatorname{poly}(n,m)$ time.
We will show by induction (on $\mu\ge 1$) that $T(\mu) \le \mu$. We also show that $L_2(\mu)= O(\mu^{c})$ for some constant $c\ge 1$. Since $L_1(\mu)=\operatorname{poly}(n,m)$ \cite[Chapter 20]{S86}, it would follow then that the total time to compute $\operatorname{Tr}(\mathcal H)$ is at most $O(\mu^{1+c})+\operatorname{poly}(\mu)$, which is polynomial in $n$, $m$, and $k$. This would give a {\it total} polynomial-time algorithm for computing $\operatorname{Tr}(\mathcal H)$ which can be converted into an {\it incremental} polynomial-time algorithm by standard methods \cite{KBEGM-SIDMA05,BEGM09}. Thus, we shall assume in the sequel that $n,m,k$ are larger than any desired constant $C$.
Without loss of generality we assume that the input hypergraph is Sperner and irredundant, and this assumption is maintained for all hypergraphs arising as inputs to the recursive subproblems.
We may also assume that $\mathcal H$ has neither a singleton hyperedge nor a vertex of degree 1 (i.e., contained in exactly one hyperedge). Indeed, if $\mathcal H$ contains a singleton hyperedge $H=\{v\}$, then by the Sperner property, no other hyperedge of $\mathcal H$ contains $v$. In this case, and also in the case when $\mathcal H$ has a vertex $v$ contained exactly in one hyperedge $H\in\mathcal H$, $\operatorname{Tr}(\mathcal H)$ can be computed as follows:
\begin{equation}
\label{e10-}
\operatorname{Tr}(\mathcal H)=\operatorname{Tr}((\mathcal H\setminus\{H\})\ \dot{\cup}\ \{H\})=\operatorname{Tr}((\mathcal H\setminus\{H\}){\wedge} \operatorname{Tr}(\{H\}),
\end{equation}
where $\operatorname{Tr}(\{H\}=\{\{w\} \mid w \in H\}$.
By Proposition~\ref{prop1} (iii), $|\operatorname{Tr}((\mathcal H\setminus\{H\})|\le k(\mathcal H)$ and thus, $\mu(\mathcal H')\le(n-1)(m-1)k\leq\mu(\mathcal H)-1$. Thus, we get by induction that $T(\mu(\mathcal H))\le 1+T(\mu(\mathcal H'))\le \mu(\mathcal H)$.
Moreover, by \raf{e10-}, $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}((\mathcal H\setminus\{H\})$ in $L_2(\mu)=\operatorname{poly}(n,m,k)$ time.
Finally, we may also assume that $\mathcal H$ does not have two identical vertices. Indeed, if it has two such vertices $v,v'$ then we can reduce the problem by calling the algorithm on the hypergraph $\mathcal H'=\{H\setminus\{v'\}~|~H\in\mathcal H\}$ instead of $\mathcal H$. Then the dual of $\mathcal H$ can be obtained as follows:
\begin{equation}
\label{e11-}
\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H')\dot{\cup}\{(T\setminus\{v\}\cup\{v'\})~|~T\in\operatorname{Tr}(\mathcal H'),~v\in T\}.
\end{equation}
Note that \raf{e11-} implies that $k(\mathcal H')\le k(\mathcal H)$ and hence
$\mu(\mathcal H')\le (n-1)mk\le \mu(\mathcal H)-1$. Thus, in this case, we get the recurrence $T(\mu)\le 1+T(\mu-1)$, which again gives by induction on $\mu\ge 1$ that
$T(\mu)\le 1+(\mu-1)\le \mu.$
Moreover, by \raf{e11-}, $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}(\mathcal H')$ in $\operatorname{poly}(n,m,k)$ time.
\medskip
We will use the following simple facts in our analysis of the running time of the algorithm.
\begin{fact}\label{fact1}
Let $\alpha,\beta,N, M$ be positive integers such that $\alpha\le N/2$ and $\beta\le M/2$. Consider the maximization problem:
\begin{align*}\label{m1}
z^*=\max& \quad x_1y_1+x_2y_2\\
{\rm s.t.}&\quad x_1+x_2 \le N,\\
&\quad y_1+y_2 \le M,\\
&\quad x_1,x_2\ge \alpha,\\
&\quad y_1,y_2\ge \beta,\\
&\quad x_1,x_2,y_1,y_2\in\mathbb Z.
\end{align*}
Then $z^*=\alpha\beta+(N-\alpha)(M-\beta)$.
\end{fact}
\noindent {\bf Proof}.~~
Let $(x^*_1,x_2^*,y^*_1,y_2^*)$ be an optimal solution. Clearly, $x_2^*=N-x_1^*$ and $y_2^*=M-y_1^*$. Without loss of generality assume that $x_1^*\ge\frac{N}{2}$. If $y_1^*<M-\beta$, then $(x^*_1,N-x_1^*,y_1^*+1,M-y_1^*-1)$ is also an optimal solution since $$x^*_1(y^*_1+1)+(N-x^*_1)(M-(y^*_1+1))=x^*_1y^*_1+(N-x^*_1)(M-y^*_1)+2x^*_1-N\ge x^*_1y^*_1+(N-x^*_1)(M-y^*_1).$$
Thus we conclude in this case that $(x^*_1,N-x_1^*,M-\beta,\beta)$ is also an optimal solution.
A symmetric argument shows that $(N-\alpha,\alpha,M-\beta,\beta)$ is an optimal solution of the maximization problem.
\hfill$\square$\bigskip
\begin{fact}\label{fact2}
Let $x_i,y_i$, for $i=1,\ldots,h$, and $M$ be positive integers such that $\sum_{i=1}^hx_iy_i\le M$. Then
$\sum_{i=1}^h(x_i+y_i)\le M+h.$
\end{fact}
\noindent {\bf Proof}.~~
For $i=1,\ldots,h$, let $\alpha_i=x_iy_i$. Note that the function $f(x_i)=x_i+\frac{\alpha}{x_i}$ is convex in $x_i>0$, and hence $\max \{x_i+y_i\mid x_iy_i=\alpha_i,~x_i\ge1,y_i\ge1\}$ is achieved at the boundary $(x_i,y_i)=(1,\alpha_i)$ or $(x_i,y_i)=(\alpha_i,1)$. The claim follows by summing the inequality $x_i+y_i\le \alpha_i+1$ over $i=1,\ldots,h$.
\hfill$\square$\bigskip
\subsection{$1$-sum decomposition}
Given a nontrivial partition $V_1\dot\cup V_2=V$ such that for all $H\in\mathcal H$: either $H\subseteq V_1$ or $H\subseteq V_2$, we have the following decomposition of the dual hypergraph by Proposition~\ref{prop2}:
\begin{equation}\label{e1}
\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H_{V_1}\dot\cup\mathcal H_{V_2})=\operatorname{Tr}(\mathcal H_{V_1})\dot\wedge\operatorname{Tr}(\mathcal H_{V_2}).
\end{equation}
Note that both $\mathcal H_{V_1}$ and $\mathcal H_{V_2}$ are unimodular. Thus in this case we get the recurrence:
\begin{align}\label{rec1}
T(\mu)\le 1+T(\mu_1)+T(\mu_2),
\end{align}
where $\mu=\mu(\mathcal H)$, $\mu_1=\mu(\mathcal H_{V_1})$, and $\mu_2=\mu(\mathcal H_{V_2})$.
Let $n_1=|V_1|$, $m_1=|\mathcal H_{V_1}|$, $k_1=|\operatorname{Tr}(\mathcal H_{V_1})|$, $n_2=|V_2|$, $m_2=|\mathcal H_{V_2}|$, and $k_2=|\operatorname{Tr}(\mathcal H_{V_2})|$. Note that $n_1,n_2,m_1,m_2\ge1$, $n_1+n_2=n$ and $m_1+m_2=m$ by the assumptions of the $1$-sum case (Corollary~\ref{c-decomp}) and hence $\mu_1,\mu_2\ge 1$. By Proposition~\ref{prop1}(iii), $k_1\le k$ and $k_2\le k$ (in fact, $k_1k_2=k$). Thus, it follows by Fact~\ref{fact1} and $n\ge 2$, $m\ge 2$ and $k\ge 1$ that
\begin{align}\label{v1}
\mu_1+\mu_2&=n_1m_1k_1 +n_2m_2k_2\le (n_1m_1+n_2m_2)k\nonumber\\
&\le [1+(n-1)(m-1)]k=[2+nm-n-m]k\le (nm-2)k\le nmk-2=\mu-2.
\end{align}
It follows by induction from \raf{rec1} that
\begin{align}\label{t1}
T(\mu)\le 1+\mu_1+\mu_2\le 1+(\mu-2)\leq\mu.
\end{align}
Note also that $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}(\mathcal H_{V_1})$ and $\operatorname{Tr}(\mathcal H_{V_2})$ using~\raf{e1} in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
\subsection{$2$-sum decomposition}
Given a nontrivial partition $V_1\dot\cup V_2=V$ and a nonempty set $S\subseteq V_1$ such that for all $H\in\mathcal H$ with $H\cap V_1\neq\emptyset$ and $H\cap V_2\neq\emptyset$: $H\cap V_1=S$, we have the following decomposition of the dual hypergraph by Proposition~\ref{prop2}:
\begin{equation}\label{e2}
\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H_{V_1}\dot\cup\mathcal H_{V_2\cup S})=\operatorname{Tr}(\mathcal H_{V_1})\wedge\operatorname{Tr}(\mathcal H_{V_2\cup S}),
\end{equation}
as $\mathcal H=\mathcal H_{V_1}\dot{\cup}\mathcal H_{V_2\cup S}$
(note that $\mathcal H_{V_1}\cap \mathcal H_{V_2\cup S}=\emptyset$ since $\mathcal H$ is Sperner and that both $\mathcal H_{V_1}$ and $\mathcal H_{V_2\cup S}$ are unimodular).
Thus in this case we get the recurrence:
\begin{align}\label{rec1-}
T(\mu)\le 1+T(\mu_1)+T(\mu_2),
\end{align}
where $\mu=\mu(\mathcal H)$, $\mu_1=\mu(\mathcal H_{V_1})$, and $\mu_2=\mu(\mathcal H_{V_2\cup S})$.
Let $n_1=n(\mathcal H_{V_1})=|V_1|$, $m_1=|\mathcal H_{V_1}|$, $k_1=|\operatorname{Tr}(\mathcal H_{V_1})|$, $n_2=n(\mathcal H_{V_2\cup S})=|V_2|+|S|$, $m_2=|\mathcal H_{V_2\cup S}|$, and $k_2=|\operatorname{Tr}(\mathcal H_{V_2\cup S})|$. Note that $n_1,n_2,m_1,m_2\ge1$ by the assumptions of the $2$-sum case (Corollary~\ref{c-decomp}) and hence $\mu_1,\mu_2\ge 1$. Then
\begin{align}\label{v2}
\mu_1+\mu_2&=n_1m_1k_1 +n_2m_2k_2\le (n-1)(m_1+m_2)k=(n-1)mk\le \mu-1,
\end{align}
where $k_1\le k$ and $k_2\le k$ by Proposition~\ref{prop1} (iii). It follows by induction from \raf{rec1-} that
\begin{align}\label{t2}
T(\mu)\le 1+\mu_1+\mu_2\le 1+(\mu-1)=\mu.
\end{align}
Note that $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}(\mathcal H_{V_1})$ and $\operatorname{Tr}(\mathcal H_{V_2\cup S})$ using~\raf{e2} in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
\subsection{$3$-sum decomposition - case 1}
Assume we are given a nontrivial partition $V_1\dot\cup V_2=V$ and two nonempty sets $S_1\subseteq V_1$ and $S_2\subseteq V_2$, such that for all $H\in\mathcal H$ with $H\cap V_1\not=\emptyset$ and $H\cap V_2\not=\emptyset$: either $H\cap V_1= S_1$ or $H\cap V_2=S_2$. Let $n_1=|V_1|$, $n_2=|V_2|$, $m_1=|\mathcal H_{V_1}\cup \mathcal H(V_2,S_2)|$ and $m_2=|\mathcal H_{V_2}\cup \mathcal H(V_1,S_1)|$. It is also assumed in this case that $n_1,n_2,m_1,m_2\ge 1$, $n_1+n_2=n$, $m_1+m_2=m$, $n_1+m_1\ge 4$ and $n_2+m_2\ge 4$, and that $\mathcal H(V_1,S_1)$ and $\mathcal H(V_2,S_2)$ are not empty
\medskip
We consider two cases:
\medskip
\noindent {\bf Case I}: there is no hyperedge $H\in\mathcal H$ such that $H\subseteq S_1\cup S_2$. Note that this, together with assumption (i) of the $3$-sum--case 1 in Corollary~\ref{c-decomp}, implies that $S_1\subset V_1$ and $S_2\subset V_2$. In this case, we have the following decomposition of the dual hypergraph:
\begin{equation}\label{e3-1}
\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H_{V_1\cup S_2}\dot{\cup}\mathcal H_{V_2\cup S_1})=\operatorname{Tr}(\mathcal H_{V_1\cup S_2})\wedge\operatorname{Tr}(\mathcal H_{V_2\cup S_1}),
\end{equation}
(Note by assumption that $\mathcal H_{V_1\cup S_2}\cap \mathcal H_{V_2\cup S_1}=\emptyset$.)
Thus in this case we get the recurrence:
\begin{align}\label{rec1--}
T(\mu)\le 1+T(\mu_1)+T(\mu_2),
\end{align}
where $\mu=\mu(\mathcal H)$, $\mu_1=\mu(\mathcal H_{V_1\cup S_2})$, and $\mu_2=\mu(\mathcal H_{V_2\cup S_1})$.
Let $n_1'=n(\mathcal H_{V_1\cup S_2})=|V_1|+|S_2|$, $m_1=|\mathcal H_{V_1\cup S_2}|$, $k_1=|\operatorname{Tr}(\mathcal H_{V_1\cup S_2})|$, $n_2'=n(\mathcal H_{V_2\cup S_1})=|V_2|+|S_1|$, $m_2=|\mathcal H_{V_2\cup S_1}|$, and $k_2=|\operatorname{Tr}(\mathcal H_{V_2\cup S_1})|$.
Then
\begin{align}\label{v3}
\mu_1+\mu_2&=n_1'm_1k_1 +n_2'm_2k_2\le (n-1)(m_1+m_2)k\le\mu-1,
\end{align}
where $k_1\le k$ and $k_2\le k$ by Proposition~\ref{prop1} (iii). It follows by induction from \raf{rec1--} that
\begin{align}\label{t3}
T(\mu)\le 1+\mu_1+\mu_2\le \mu.
\end{align}
Note that $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}(\mathcal H_{V_1\cup S_2})$ and $\operatorname{Tr}(\mathcal H_{V_2\cup S_1})$ using ~\raf{e3-1} in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
\medskip
\noindent {\bf Case II}: there is a hyperedge $H_0\in\mathcal H$ such that $H_0\subseteq S_1\cup S_2$. Note that $H_0\cap S_1\neq\emptyset$ and $H_0\cap S_2\neq\emptyset$ since otherwise by the simplicity of $\mathcal H$ we are in the $2$-sum case. Without loss of generality, assume that $H_0\cap V_1=S_1$ and $H_0\cap V_2\subseteq S_2$. We assume that $\mathcal H(V_1,S_1)$ and $\mathcal H(V_2,S_2)$ are not empty; otherwise, we are in the $1$-sum or $2$-sum cases. Given these assumptions, we use the following decomposition of the dual hypergraph:
\begin{equation}\label{e3-2}
\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H_1{\cup}\mathcal H_2)=\operatorname{Tr}(\mathcal H_1)\wedge\operatorname{Tr}(\mathcal H_2),
\end{equation}
where $\mathcal H_1=\mathcal H_{V_1}\cup\mathcal H(V_2,S_2)\cup\{H_0\}$ and $\mathcal H_2=\mathcal H_{V_2}\dot \cup\mathcal H(V_1,S_1)$.
Note that $H_0$ is the the only hyperedge that belongs to both $\mathcal H_1$ and $\mathcal H_2$. Note also that neither $\mathcal H_1$ nor $\mathcal H_{2}$ may be a projection of $\mathcal H$ (i.e., of the form $\mathcal H_S$ for some $S\subseteq V$) since there are hyperedges $H\subseteq S_1\cup S_2$ that may not be included in $\mathcal H_1$ and $\mathcal H_2$. Hence, Proposition~\ref{prop1} {\it cannot} be used to bound the sizes of $\operatorname{Tr}(\mathcal H_1)$ and $\operatorname{Tr}(\mathcal H_2)$. Nevertheless, due to the special structure of the decomposition in this case, we can use the bounds given in Lemma~\ref{cl1} below instead. Let $\bar \mathcal H_1\subseteq 2^{V_1\cup \{v_2\}}$ (resp., $\bar \mathcal H_2\subseteq 2^{V_2\cup \{v_1\}}$) be the hypergraph obtained from $\mathcal H_1$ (resp., $\mathcal H_2$) by replacing $S_2$ (resp., $S_1$) by a {\it new} single vertex $v_2$ (resp., $v_1$), that is,
\begin{align*}
\bar\mathcal H_1=\mathcal H_{V_1}\cup \bar \mathcal H(V_2,S_2)\cup\{\bar H_0\},\qquad
\bar\mathcal H_2=\mathcal H_{V_2}\cup\bar \mathcal H(V_1,S_1),
\end{align*}
where $\bar \mathcal H(V_2,S_2)=\{(H\setminus S_2)\cup\{v_2\}\mid H\in\mathcal H(V_2,S_2)\}$, $\bar H_0=(H_0\setminus S_2)\cup\{v_2\}$, and $\bar \mathcal H(V_1,S_1)=\{(H\setminus S_1)\cup\{v_1\}\mid H\in\mathcal H(V_1,S_1)\}$.
\begin{lemma}\label{cl0}
If $\mathcal H$ is unimodular, then both $\bar \mathcal H_1$ and $\bar \mathcal H_2$ are unimodular.
\end{lemma}
\noindent {\bf Proof}.~~
Let $v$ be an arbitrary vertex in $H_0\cap S_2$. Then the (hyperedge-vertex) incidence matrix of the hypergraph $\bar\mathcal H_1$ is a submatrix of that of $\mathcal H$, with rows restricted to $\mathcal H_{V_1}\cup\mathcal H(V_2,S_2)\cup\{H_0\}$, and columns restricted to $V_1\cup\{v\}$. This shows that this submatrix is totally unimodular. A similar argument shows that $\bar \mathcal H_2$ is also unimodular.
\hfill$\square$\bigskip
\begin{lemma} \label{cl1}
$|\operatorname{Tr}(\bar\mathcal H_1)|\le|\operatorname{Tr}(\mathcal H)|$ and $|\operatorname{Tr}(\bar\mathcal H_2)|\le|\operatorname{Tr}(\mathcal H)|$.
\end{lemma}
\noindent {\bf Proof}.~~
We prove the claim that $|\operatorname{Tr}(\bar\mathcal H_1)|\le|\operatorname{Tr}(\mathcal H)|$; the other claim can be proved similarly.
It is enough to show that for every minimal transversal $T\in \operatorname{Tr}(\bar\mathcal H_1)$, there is a minimal transversal $T'\in\operatorname{Tr}(\mathcal H)$ such that for any distinct $T_1,T_2\in\operatorname{Tr}(\bar\mathcal H_1)$, $T_1'$ and $T_2'$ are distinct.
Let $\mathcal T_1=\{T\in\operatorname{Tr}(\bar \mathcal H_1): v_2\not\in T\}$ and $\mathcal T_2=\operatorname{Tr}(\bar \mathcal H_1)\setminus\mathcal T_1$. Consider first $T\in\mathcal T_1$. By assumption $T\cap S_1\neq\emptyset$ since $T$ has a nonempty intersection with $\bar H_0$. It follows that the only hyperedges of $\mathcal H$ having empty intersection with $T$ are those in $\mathcal H_{V_2}$. Note that none of these hyperedges are contained in $S_2$ since $\mathcal H$ is Sperner. This implies that $\mathcal H_{V_2}^{V_2\setminus S_2}\neq\{\emptyset\}$ and therefore $\operatorname{Tr}(\mathcal H_{V_2}^{V_2\setminus S_2})\neq\emptyset$. Let $T''$ be an arbitrary minimal transversal in $\operatorname{Tr}(\mathcal H_{V_2}^{V_2\setminus S_2})$. Then it is easy to see that $T'=T\cup T''$ is in $\operatorname{Tr}(\mathcal H)$.
Consider now $T\in\mathcal T_2$. By the minimality of $T$, there is a hyperedge $H\in\bar\mathcal H(V_2,S_2)\cup\{\bar H_0\}$ such that $H\cap T=\{v_2\}$. Furthermore, for every $v\in T\setminus\{v_0\}$, there is an $H\in\mathcal H_{V_1}$ such that $T\cap H=\{v\}$. Let $\mathcal H(T)=\{H\in\mathcal H\mid H\cap T\setminus\{v_2\}=\emptyset\}$ and note that $\mathcal H(T)^{V_2}$ is nontrivial. Pick $T''\in\operatorname{Tr}(\mathcal H(T)^{V_2})$ arbitrarily.
Then it is easy to see that $T'=T\cup T''$ is in $\operatorname{Tr}(\mathcal H)$.
Finally, note that for any distinct $T_1,T_2\in\mathcal T_1$ (resp., $T_1,T_2\in\mathcal T_2$), the constructed minimal transversals $T_1',T_2'\inT(\mathcal H)$ are distinct. Moreover, for $T_1\in\mathcal T_1$ and $T_2\in\mathcal T_2$, $T_1'$ and $T'_2$ are distinct because $T'_1\cap S_2=\emptyset$ while $T_2'\cap S_2\not=\emptyset$.
\hfill$\square$\bigskip
To compute~\raf{e3-2}, we find $\operatorname{Tr}(\bar\mathcal H_1)$ and $\operatorname{Tr}(\bar \mathcal H_2)$, recursively. Then $\operatorname{Tr}(\mathcal H_1)$ and $\operatorname{Tr}(\mathcal H_2)$ are given by the following claim.
\begin{lemma}\label{cl2}
Let $\mathcal T_1=\{T\in\operatorname{Tr}(\bar \mathcal H_1)\mid v_2\not\in T\}$, $\mathcal T_2=\{T\in\operatorname{Tr}(\bar \mathcal H_1)\mid v_2\in T,~S_1\cap T=\emptyset\}$, $\mathcal T_3=\operatorname{Tr}(\bar \mathcal H_1)\setminus(\mathcal T_1\cup\mathcal T_2)$,
$\mathcal T_1'=\{T\in\operatorname{Tr}(\bar \mathcal H_2)\mid v_1\not\in T\}$ and $\mathcal T_2'=\operatorname{Tr}(\bar \mathcal H_2)\setminus\mathcal T_1'$. Then
\begin{align}\label{e3-3}
\operatorname{Tr}(\mathcal H_1)&=\mathcal T_1\dot{\cup}\{(T\setminus \{v_2\})\cup\{v\}\mid v\in H_0\cap S_2\text{ and }T\in\mathcal T_2\}\dot\cup \{(T\setminus \{v_2\})\cup \{v\}\mid v\in S_2\text{ and }T\in\mathcal T_3\},\\
\operatorname{Tr}(\mathcal H_2)&=\mathcal T_1'\dot\cup \{(T\setminus \{v_1\})\cup\{v\}\mid v\in S_1\text{ and }T\in\mathcal T_2'\}.\label{e3-4}
\end{align}
\end{lemma}
\noindent {\bf Proof}.~~
Let us prove \raf{e3-3}, since the proof of \raf{e3-4} is similar. Suppose $T\in\operatorname{Tr}(\mathcal H_1)$. If $T\cap S_2=\emptyset$ then (it is easy to see that) $T\in\mathcal T_1$.
If $T\cap S_2\neq\emptyset$ then by minimality of $T$, $|T\cap S_2|=1$; let $T\cap S_2=\{v\}$. If $T\cap S_1=\emptyset$ then necessarily $v\in H_0$, in which case $(T\setminus\{v\})\cup\{v_2\}\in\mathcal T_2$; otherwise $v$ can be any element in $S_2$, and hence, $(T\setminus\{v\})\cup\{v_2\}\in\mathcal T_3$. On the other direction, if $T\in\mathcal T_1$ then clearly $T\in\operatorname{Tr}(\mathcal H_1)$; if $T\in\mathcal T_2$ then
$T\cap\bar H_0=\{v_2\}$ which implies that $(T\setminus \{v_2\})\cup\{v\}\in \operatorname{Tr}(\mathcal H_1)$ for every $v\in H_0\cap S_2$; finally, if $T\in\mathcal T_3$ then
there is a hyperedge $H\in\bar\mathcal H(V_2,S_2)$ such that $H\cap T=\{v_2\}$, which implies in turn that $(T\setminus \{v_2\})\cup\{v\}\in \operatorname{Tr}(\mathcal H_1)$.
\hfill$\square$\bigskip
Note that $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}(\bar\mathcal H_1)$ and $\operatorname{Tr}(\bar\mathcal H_2)$ using ~\raf{e3-2} and Lemma~\ref{cl2} in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
Let $n_1'=n(\bar{\mathcal H_1})=n_1+1$, $m_1'=|\bar\mathcal H_{1}|\in\{m_1, m_1+1\}$, $k_1=|\operatorname{Tr}(\bar\mathcal H_{1})|$, $n_2'=n(\bar{\mathcal H_2})=n_2+1$, $m_2=|\bar\mathcal H_{2}|$, and $k_2=|\operatorname{Tr}(\bar\mathcal H_{2})|$. By the decomposition, $n_1'+n_2'=n+2$ and $m_1'+m_2= m+1$, and by Lemma~\ref{cl1}, $k_1\le k$ and $k_2\le k$. Note that $n_1',n_2'\ge 2$, $m_1',m_2\ge1$, $n_1+m_1 \ge 4$, and $n_2+m_2\ge 4$, by the assumptions of the $3$-sum case 1 in Corollary~\ref{c-decomp}.
\medskip
We consider 3 subcases.
\medskip
\noindent {\bf Case II-I}: $2\le n_1'\le 3$. Then a simple procedure will be used to compute $\operatorname{Tr}(\bar\mathcal H_1)$, and hence we need only to recurse on $\bar\mathcal H_2$, giving the simpler recurrence: $T(\mu)\le 2+T(\mu_2)$. Note that $m_2\le m-2$ since $n_1\le 2$ implies $m_1\ge 2$ and hence $m_2=m-m_1\le m-2$. Since $\mu_2=n_2'm_2k_2\le n(m-2)k\le\mu-2$, we get by induction that
\begin{align}\label{t4}
T(\mu)\le 2+\mu_2\le \mu.
\end{align}
\medskip
\noindent {\bf Case II-II}: $n_2'=2$. Then a simple procedure will be used to compute $\operatorname{Tr}(\bar\mathcal H_2)$, and hence we need only to recurse on $\bar\mathcal H_1$, giving the simpler recurrence: $T(\mu)\le 2+T(\mu_1)$. As above, $m_1\le m-3$ implying that $\mu_1=n_1'm_1'k_2\le n(m-2)k\le\mu-2$, and giving by induction again that $T(\mu)\le \mu.$
\medskip
\noindent {\bf Case II-III}: $n_1'\ge 4$ and $n_2'\ge 3$. We first note that $m_1',m_2\ge 2$. Indeed, if $m_1'=1$ (resp., $m_2=1$), then $\mathcal H_{V_1}=\emptyset$ and $\mathcal H(V_2,S_2)=\{H_0\}$ (resp., $\mathcal H_{V_2}=\emptyset$ and $\mathcal H(V_1,S_1)=\{H_0\}$). Since we assume that $\mathcal H$ does not have identical vertices, we must have $n_1=1$ (resp., $n_2=1$). In either case we get a contradiction to the boundary assumtpions (ii) of the $3$-sum--case 1 in Corollary~\ref{c-decomp}.
Lemmas~\ref{cl0} and~\ref{cl2} imply that, in this case, we get the recurrence:
\begin{align}\label{rec1---}
T(\mu)\le 1+T(\mu_1)+T(\mu_2),
\end{align}
where $\mu=\mu(\mathcal H)$, $\mu_1=\mu(\bar\mathcal H_{1})$, and $\mu_2=\mu(\bar\mathcal H_{2})$.
Then by Fact~\ref{fact1}, applied with $x_1=n_1'$, $y_1=m_1'$, $x_2=n_2'$, $y_2=m_2$, $N=n+2$, $M=m+1$, $\alpha=3$ and $\beta=2$, we get (as $n\ge 5$ and $m\ge 3$)
\begin{align}\label{v5}
\mu_1+\mu_2&=n_1'm_1'k_1 +n_2'm_2k_2\le(n_1'm_1'+n_2'm_2)k\le((n-1)(m-1)+6)k
\nonumber\\
&= nmk-(n+m-7)k \le\mu-1.
\end{align}
It follows by induction from \raf{rec1---} that
\begin{align}\label{t5}
T(\mu)\le 1+\mu_1+\mu_2\le\mu.
\end{align}
\subsection{$3$-sum decomposition - case 2}
\label{section-3-sum-case2}
Let
$\mathcal H_1=\mathcal H_{V_1}$ and $\mathcal H_2=\mathcal H_{V_2}$.
By Corollary \ref{c-decomp}, we have three nonempty disjoint sets $S_0,S_1,S_2$ in $V_2$, and
the following two families are nonempty:
\begin{eqnarray}
\mathcal F_1&=&\{H \in \mathcal H \mid H \cap V_1=S_0 \cup S_2, H \cap V_2\not=\emptyset\},\\
\mathcal F_2&=&\{H \in \mathcal H \mid H \cap V_1=S_0 \cup S_1, H \cap V_2\not=\emptyset\}.
\end{eqnarray}
Note that $V_1,V_2\not=\emptyset$, $\mathcal H_1\not=\emptyset$, and
$\mathcal H$ can be partitioned in the following way.
\begin{equation}
\label{eq--aed1}
\mathcal H=\mathcal H_1 \dot{\cup} \mathcal H_2 \dot{\cup} \mathcal F_1 \dot{\cup} \mathcal F_2,
\end{equation}
where $\dot{\cup}$ denotes the disjoint union.
For $i=0,1,2$, let
\begin{equation}
\mathcal T_i=\{T \in \operatorname{Tr}(\mathcal H_1) \mid T\cap S_i \not=\emptyset, T \cap S_j=\emptyset\,\, (j\not=i)\},
\end{equation}
and let
\begin{equation}
\mathcal T=\operatorname{Tr}(\mathcal H_1) \setminus (\mathcal T_0 \cup \mathcal T_1 \cup \mathcal T_2).
\end{equation}
By definition, we have
\begin{equation}
\operatorname{Tr}(\mathcal H_1)=\mathcal T \dot{\cup} \mathcal T_0 \dot{\cup} \mathcal T_1 \dot{\cup} \mathcal T_2.
\end{equation}
Let
\begin{equation}
\mathcal P=\mathcal H_{S_0 \cup S_1\cup S_2}\,\,(=\{H \in \mathcal H \mid H \subseteq S_0\cup S_1 \cup S_2\}).
\end{equation}
We separately consider the following 4 cases.
\medskip
\noindent
{\bf Case I}: $\mathcal P=\emptyset$.
\smallskip
\noindent
{\bf Case II}: $\mathcal P=\{S_0\cup S_1 \cup S_2\}$.
\smallskip
\noindent
{\bf Case III}: $\mathcal P\not= \emptyset, \{S_0\cup S_1 \cup S_2\}$ and $\mathcal T\not=\emptyset$.
\smallskip
\noindent
{\bf Case IV}: $\mathcal P\not= \emptyset, \{S_0\cup S_1 \cup S_2\}$ and $\mathcal T=\emptyset$.
\noindent
\subsubsection*{Case I}
$\mathcal H$ can be partitioned into $\mathcal H_{1}$ and $\mathcal H_{S_0\cup S_1\cup S_2\cup V_2}$, i.e.,
$\mathcal H=\mathcal H_{1} \dot{\cup}\mathcal H_{S_0\cup S_1\cup S_2\cup V_2}$ and $\mathcal H_{1}, \mathcal H_{S_0\cup S_1\cup S_2\cup V_2}\not=\emptyset$.
Since $\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H_{1}) \wedge \operatorname{Tr}(\mathcal H_{S_0\cup S_1\cup S_2\cup V_2})$, we obtain $\operatorname{Tr}(\mathcal H)$ by computing $\operatorname{Tr}(\mathcal H_{1})$ and $\operatorname{Tr}(\mathcal H_{S_0\cup S_1\cup S_2\cup V_2})$.
Let $n_1=|V_1|$, $m_1=|\mathcal H_{1}|$, $k_1=|\operatorname{Tr}(\mathcal H_{1})|$,
$n'_2=|S_0\cup S_1\cup S_2\cup V_2|$,
$m'_2=|\mathcal H_{S_0\cup S_1\cup S_2\cup V_2}|$, $k_2=|\operatorname{Tr}(\mathcal H_{S_0\cup S_1\cup S_2\cup V_2})|$.
Similar to the $2$-sum decomposition case, we can show that $T(\mu)\le \mu$ and the computation of $\operatorname{Tr}(\mathcal H)$ can be done in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
\subsubsection*{Case II}
We consider two cases: {\bf II-I}: $|\mathcal H_{1}| \geq 2$ and {\bf II-II}: $|\mathcal H_{1}|=1$.
\medskip
\noindent
{\bf Case II-I}: $|\mathcal H_{1}| \geq 2$.
Let $\mathcal G$ be a hypergraph obtained from $\mathcal H_{S_0\cup S_1\cup S_2\cup V_2}$ by replacing $S_0, S_1$, and $S_2$ by new vertices $v_0$, $v_1$ and $v_2$, respectively.
For any hyperedge $H \in \mathcal H_{S_0\cup S_1\cup S_2\cup V_2}$,
$H \cap S_i\not=\emptyset$ implies that $S_i \subseteq H$.
Thus $\mathcal G$ is well-defined.
Note that $\operatorname{Tr}(\mathcal H_{S_0\cup S_1\cup S_2\cup V_2})$ can be obtained from $\operatorname{Tr}(\mathcal G)$ in polynomial time by replacing $v_i$ with any element in $S_i$.
Since $\mathcal H=\mathcal H_{1} \cup\mathcal H_{S_0\cup S_1\cup S_2\cup V_2}$, we have $\operatorname{Tr}(\mathcal H)=\operatorname{Tr}(\mathcal H_{1}) \wedge \operatorname{Tr}(\mathcal H_{S_0\cup S_1\cup S_2\cup V_2})$. We thus decompose $\mathcal H$ into $\mathcal H_1$ and $\mathcal G$. Namely we compute $\operatorname{Tr}(\mathcal H)$ from $\operatorname{Tr}(\mathcal H_1)$ and $\operatorname{Tr}(\mathcal G)$.
Since $|\operatorname{Tr}(\mathcal H_1)|, |\operatorname{Tr}(\mathcal G)|\leq |\operatorname{Tr}(\mathcal H)|\,(=k)$,
this can be done in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
Let us next show that $T(\mu)\leq \mu$.
Let $n_1=|V_1|$, $m_1=|\mathcal H_{1}|$, $k_1=|\operatorname{Tr}(\mathcal H_{1})|$,
$n'_2=|V_2|$+3,
$m'_2=|\mathcal G|$, and $k_2=|\operatorname{Tr}(\mathcal G)|$.
Note that $\mathcal H=\mathcal H_{1} \cup\mathcal H_{S_0\cup S_1\cup S_2\cup V_2}$, $\mathcal H_{1}\cap \mathcal H_{S_0\cup S_1\cup S_2\cup V_2}=\{S_0\cup S_1\cup S_2\}$.
This, together with definition and the discussion above, implies
that
\begin{equation*}
1\leq n_1,n_2' \leq n, \ n_1 +n'_2=n+3, \
2\leq m_1, m'_2 \leq m-1, \ m_1 +m'_2=m+1, \
k_1,k_2 \leq k.
\end{equation*}
Thus we have
\begin{equation}
\label{eq3d8ge1-eq}
n_1m_1k_1 + n_2'm_2'k_2 \leq (n_1m_1+ n_2'm_2')k\leq (n(m-1)+6)k \leq nmk-1
\end{equation}
where Fact \ref{fact1} is used for the second inequality,
and the third ineuqality is obtained by assuming that $n$ is at least $7$.
It follows from \raf{eq3d8ge1-eq} that $T(\mu) \leq \mu$.
We recall that $\operatorname{Tr}(\mathcal H)$ is directly computed from $\mathcal H$
if at least one of $n$, $m$,
and $k$ is bounded by some constant $C$.
Thus in case $n< 7$, we have $T(\mu)=1$, which also satisfies $T(\mu) \leq \mu$.
\medskip
\noindent
{\bf Case II-II}: $|\mathcal H_{1}| =1$. In this case, we have $H_1=\{S_0 \cup S_1\cup S_2\}$.
Therefore, the following lemma is satisfied.
\begin{lemma}
\label{lemma-sum3-2-2-1}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1}
and $\mathcal H_1=\{S_0 \cup S_1\cup S_2\}$.
Then we have $\operatorname{Tr}(\mathcal H)=\{\{v\} \mid v\in S_0\} \dot\wedge \operatorname{Tr}(\mathcal H_2) \ \dot\cup \ \operatorname{Tr}(H^{V\setminus S_0})$.
\end{lemma}
\noindent
\noindent {\bf Proof}.~~
From the definition, it is not difficult to see that $\operatorname{Tr}(\mathcal H)\supseteq \{\{v\} \mid v\in S_0\} \dot\wedge \operatorname{Tr}(\mathcal H_2) \ \dot\cup \ \operatorname{Tr}(H^{V\setminus S_0})$.
For the converse inclusion, let $T\in \operatorname{Tr}(\mathcal H)$.
If $T \cap S_0=\emptyset$, then $T$ is contained in $\operatorname{Tr}(H^{V\setminus S_0})$.
Assume next that $T \cap S_0 \not=\emptyset$.
For any $i=0,1,2$ and any hyperedge $H \in \mathcal H$,
$H \cap S_i\not=\emptyset$ inplies that $S_0 \subseteq H$.
This means that $|T\cap S_0|=1$ and $T \cap S_i=\emptyset$ for $i=1,2$. Moreover, we have $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$, which completes the converse inclusion.
\hfill$\square$\bigskip
Note that $\mathcal H_2,\mathcal H^{V\setminus S_0}\not=\{\emptyset\}$, and hence
$\operatorname{Tr}(\mathcal H_2),\operatorname{Tr}(H^{V\setminus S_0})\not=\emptyset$.
Based on Lemma \ref{lemma-sum3-2-2-1}, we decompose $\mathcal H$ into $\mathcal H_2$ and $\mathcal H^{V\setminus S_0}$. Namely, we compute $\operatorname{Tr}(\mathcal H)$ from
$\operatorname{Tr}(\mathcal H_2)$ and $\operatorname{Tr}(H^{V\setminus S_0})$ in time $L_2(\mu)=\operatorname{poly}(n,m,k)$.
Let $n'_1=|V_2|$, $m'_1=|\mathcal H_2|$, $k'_1=|\operatorname{Tr}(\mathcal H_2)|$,
$n'_2=n-|S_0|$, $m'_2=|\mathcal H^{V\setminus S_0}|$, and $k'_2=|\operatorname{Tr}(\mathcal H^{V\setminus S_0})|$.
Then we have $n'_1, n'_2\leq n-1$, $m'_1, m'_2\leq m$, and
$k'_1, k_2'\leq k-1$ and $k_1'+k_2'\leq k$.
Thus we have
$n_1'm_1'k_1'+n_2'm_2'k_2'\leq (n-1)(m-1)k\leq nmk-1$,
where the last inequality is obtained from $n\geq 3$.
This implies that $T(\mu)\leq \mu$.
\subsubsection*{Case III}
For a set $W \subseteq V$, define $\eta(W)\subseteq \{0,1,2\}$ by
\[
\eta(W)=\{i\mid W \cap S_i\not=\emptyset\}.
\]
For $i=1,2$, let
\begin{eqnarray}
\mathcal F_i'&=&\operatorname{Min}(\mathcal F_i^{V_2} \cup \mathcal H_2).
\end{eqnarray}
\begin{lemma}
\label{lemma-maxk12}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1} and $\mathcal P\not=\emptyset$.
For a minimal transversal $T \in \operatorname{Tr}(\mathcal H)$, the following statements are satisfied.
\begin{description}
\item[(i)] If $\eta(T)=\{0\}$, then $T\cap V_1 \in \mathcal T_0$ and $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$.
\item[(ii)] If $\eta(T)=\{i\}$ for $i=1,2$, then $T\cap V_1 \in \mathcal T_i$ and $T\cap V_2 \in \operatorname{Tr}(\mathcal F_i')$.
\item[(iii)] If $\eta(T)=\{i,j\}$, then $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$, and moreover,
exactly one of the following two conditions holds.
{\bf (iii-1)} $T\cap V_1 \in \mathcal T$.
{\bf (iii-2)} There exists a vertex $v$ such that
either $\bigl(T\cap S_i=\{v\}$ and $(T\setminus \{v\})\cap V_1 \in \mathcal T_j\bigr)$ or $\bigl(T\cap S_j=\{v\}$ and $(T\setminus \{v\})\cap V_1 \in \mathcal T_i\bigr)$.
\item[(iv)] If $\eta(T)=\{0,1,2\}$, then $T\cap V_1 \in \mathcal T$ and $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$.
\end{description}
\end{lemma}
\noindent
\noindent {\bf Proof}.~~
Let $T$ be a minimal transversal of $\mathcal H$.
Since $\mathcal P\not=\emptyset$, we have $\eta(T)\not=\emptyset$.
{\bf (i)} and {\bf (ii)}. Let $\eta(T)=\{i\}$ for some $i$.
We first show that $T\cap V_1$ is a minimal transveral of $\mathcal H_1$, implying that
$T\cap V_1 \in \mathcal T_i$.
By definition, $T\cap V_1$ is a transveral of $\mathcal H_1$. Since $T \in \operatorname{Tr}(\mathcal H)$, for each $w \in T \cap V_1$, there exists a hyperedge $H_w \in \mathcal H$ such that $H_w \cap T =\{w\}$.
For $w \in T \cap (V_1 \setminus S_i)$, we have
$H_w \in \mathcal H_1$, implying that $(T\setminus \{w\})\cap V_1$ is not a transveral of $\mathcal H_1$.
For $w \in T \cap S_i$, if $T \cap S_i=\{w\}$ holds, then we have $(T\setminus w) \cap P=\emptyset$ for a $P \in \mathcal P\,(\subseteq \mathcal H_1)$.
Thus $T\cap V_1$ is a minimal transveral of $\mathcal H_1$.
On the other hand, if $|T \cap S_i|\geq 2$, we have $H_w \in \mathcal H_1$, since any $H \in F_j$ ($j\not=i$) contains $S_i$.
This implies that $T\cap V_1$ is a minimal transveral of $\operatorname{Tr}(\mathcal H_1)$.
If $\eta(T)=\{0\}$, for any $F \in \mathcal F_1 \cup \mathcal F_2$, we have $F \cap T\not=\emptyset$.
This implies that $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$.
We next show that $T\cap V_2 \in \operatorname{Tr}(\mathcal F_i')$ if $\eta(T)=\{i\}$ for $i=1,2$.
Since $T\cap (\cup_{j\not=i}S_j)=\emptyset$, $T\cap V_2$ is a transversal of $\mathcal F_i$.
Since $T$ is a minimal transversal of $\mathcal H$, for each $w \in T\cap V_2$, there exists a hyperedge $H_w \in \mathcal H$ such that $H_w \cap T=\{w\}$. Since $w \in V_2$ and $T \cap S_j=\emptyset$ for $j\,(\not=i)$, $H_w$ is contained in $\mathcal F_i\cup \mathcal H_2$.
Thus $T\cap V_2$ is a minimal transversal of $\mathcal F_i \cup \mathcal H_2$, which implies that $T\cap V_2 \in \operatorname{Tr}(\mathcal F_i')$.
{\bf (iii)}. Let $\eta(T)=\{i,j\}$.
By definition, we have $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$.
Let us assume that $T\cap V_1 \not\in \mathcal T$, and show that {\bf (iii-2)} holds.
Since $T \in \operatorname{Tr}(\mathcal H)$, for each $w \in T \cap V_1$, there exists a hyperedge $H_w \in \mathcal H$ such that $H_w \cap T =\{w\}$.
For $w \in T \cap (V_1 \setminus (S_i\cup S_j))$, we have
$H_w \in \mathcal H_1$, implying that $(T\setminus \{w\})\cap V_1$ is not a transveral of $\mathcal H_1$.
However,
since $T\cap V_1 \not\in \mathcal T$, there exists a $v \in T\cap V_1$ such that $T\setminus \{v\}\cap V_1$ is a transversal of $\mathcal H_1$. Thus $v$ is contained in $S_i \cup S_j$.
Let us assume that $v \in S_i$, since otherwise we exchange $i$ with $j$.
For this $v$, we have $H_v \in \mathcal F_1 \cup \mathcal F_2$.
This means that $j\not=0$ and $H_v \in \mathcal F_j$.
Since $S_i \subseteq H_v$, $T\cap S_i=\{v\}$ holds.
We next show that $(T\setminus \{v\})\cap V_1 \in \mathcal T_j$.
If $|T\cap S_j|=1$, say $T\cap S_j=\{u\}$ for some $u$, then
$(T\setminus \{v,u\}) \cap V_1$ is not a transversal of $\mathcal H_1$, since $\mathcal P\not=\emptyset$.
Thus in this case $(T\setminus \{v\})\cap V_1$ is a minimal transversal of $\mathcal H_1$, and is contained in $\mathcal T_j$.
On the other hand, if $|T\cap S_j| \geq 2$, then we have $H_u \in \mathcal H_1$ for any $u \in T\cap S_j$.
Thus it holds that $(T\setminus \{v\})\cap V_1 \in \mathcal T_j$.
{\bf (iv)}. By definition, we have $T\cap V_2 \in \operatorname{Tr}(\mathcal H_2)$.
Since $T \in \operatorname{Tr}(\mathcal H)$, for each $w \in T \cap V_1$, there exists a hyperedge $H_w \in \mathcal H$ such that $H_w \cap T =\{w\}$.
For $w \in T \cap (V_1 \setminus (S_0\cup S_1 \cup S_2))$, we have
$H_w \in \mathcal H_1$, implying that $(T\setminus \{w\})\cap V_1$ is not a transveral of $\mathcal H_1$.
For $w \in T \cap (S_0\cup S_1 \cup S_2)$, we again have $H_w \in \mathcal H_1$, implying that $(T\setminus \{w\})\cap V_1$ is not a transveral of $\mathcal H_1$.
Therefore, $T\cap V_1 \in \mathcal T$ holds.
\hfill$\square$\bigskip
By Lemma \ref{lemma-maxk12}, we can compute in polynomial time $\operatorname{Tr}(\mathcal H)$ from $\operatorname{Tr}(\mathcal H_1)\,(=\mathcal T \cup \mathcal T_0 \cup \mathcal T_1 \cup \mathcal T_2)$ and
$\operatorname{Tr}(\mathcal H_2)$, and $\operatorname{Tr}(\mathcal F'_i)$ ($i=1,2)$ if $\mathcal P\not=\emptyset$.
\begin{lemma}
\label{lemma-maxk12edg}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1} and $\mathcal P\not=\emptyset$.
Then we have
\[
\operatorname{Tr}(\mathcal H) \subseteq \Bigl(\mathcal T_0 \cup \mathcal T \cup \bigcup_{i=1,2} \{ T \cup \{v\} \mid T \in \mathcal T_i, v \in S_0 \cup S_{3-i} \}\Bigr) \wedge \operatorname{Tr}(\mathcal H_2) \cup \bigcup_{i=1,2} (\mathcal T_i \wedge \operatorname{Tr}(\mathcal F_i')).
\]
\end{lemma}
The following two lemmas show that $|\operatorname{Tr}(\mathcal H_1)|$, $|\operatorname{Tr}(\mathcal H_2)|$, and $|\operatorname{Tr}(\mathcal F'_i)|$ ($i=1,2)$
are bounded by the size of $\operatorname{Tr}(\mathcal H)$, which immediately implies that the sum of their volume is bounded by the volume of $\mathcal H$.
\begin{lemma}
\label{lemma-maxk13}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1}.
Then the following statements holds.
\begin{description}
\item[(i)] If $\mathcal T_0\not=\emptyset$, then we have $\mathcal T_0\dot{\wedge}\operatorname{Tr}(\mathcal H_2) \in \operatorname{Tr}(\mathcal H)$.
\item[(ii)] For $i=1,2$, we have $\mathcal T_i\not=\emptyset$ and $\mathcal T_i\dot{\wedge}\operatorname{Tr}(\mathcal F_i') \in \operatorname{Tr}(\mathcal H)$.
\item[(iii)] If $\mathcal T \not=\emptyset$ and $\mathcal P\not=\emptyset$, then $\mathcal T \dot{\wedge}\operatorname{Tr}(\mathcal H_2) \in \operatorname{Tr}(\mathcal H)$.
\end{description}
\end{lemma}
\noindent
\noindent {\bf Proof}.~~
{\bf (i)}. For each $T_1 \in \mathcal T_0$ and $T_2\in \operatorname{Tr}(\mathcal H_2)$, $T_1 \cup T_2$ is a transversal of $\mathcal F_j$ for any $j$.
Thus $T_1 \cup T_2$ is a minimal transversal of $\mathcal H$.
{\bf (ii)}. For $i=1,2$, $\mathcal T_i=\emptyset$ if and only if there exists a hyperedge $H$ such that $H \subseteq \cup_{j\not=i}S_j$. Since $\mathcal H$ is Sperner, the latter implies that $\mathcal F_i=\emptyset$.
Thus $\mathcal F_i\not=\emptyset$ implies $\mathcal T_i\not=\emptyset$.
Morever, for each $T_1 \in \mathcal T_i$ and $T_2\in \operatorname{Tr}(\mathcal F_i')$, $T_1 \cup T_2$ is a transversal of $\mathcal F_j$ for $j \,(\not=i)$.
Thus $T_1 \cup T_2$ is a minimal transversal of $\mathcal H$.
{\bf (iii)}. For each $T_1 \in \mathcal T$ and $T_2\in \operatorname{Tr}(\mathcal H_2)$, $T_1 \cup T_2$ is a transversal of $\mathcal F_j$ for any $j$,
since $\eta(T_1)\geq 2$.
Thus $T_1 \cup T_2$ is a minimal transversal of $\mathcal H$.
\hfill$\square$\bigskip
As a corollary of Lemma \ref{lemma-maxk13}, we have the following result.
\begin{lemma}
\label{lemma-maxk14}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1}, $\mathcal P\not=\emptyset$ and $\mathcal T\not=\emptyset$.
Then we have
\begin{equation}
(\mathcal T \dot{\cup} \mathcal T_0) \dot{\wedge}\operatorname{Tr}(\mathcal H_2) \,\ \dot{\cup} \ \ \mathcal T_1\dot{\wedge}\operatorname{Tr}(\mathcal F_1') \,\ \dot{\cup} \ \ \mathcal T_2\dot{\wedge}\operatorname{Tr}(\mathcal F_2')
\ \subseteq \ \operatorname{Tr}(\mathcal H).
\end{equation}
\end{lemma}
\noindent
\noindent {\bf Proof}.~~
Immediate from Lemma \ref{lemma-maxk13}.
\hfill$\square$\bigskip
We remark that $\mathcal T$, $\operatorname{Tr}(\mathcal H_2)$, $\mathcal T_i$ $(i=1,2)$, and $\operatorname{Tr}(\mathcal F_i')$ $(i=1,2)$ in Lemma \ref{lemma-maxk13} are all nonemtpy. However, $\mathcal T_0$ might be empty.
Based on Lemmas \ref{lemma-maxk12edg} and \ref{lemma-maxk14}, in {\bf Case III},
our procedure decomposes $\mathcal H$ into
$4$ hypergraphs $\mathcal H_1$, $\mathcal H_2$, and $\mathcal F_i'$ $(i=1,2)$.
\medskip
Let $n_1=|V_1|$, $n_2=|V_2|$, $m_1=|\mathcal H_1|$, $k_1=|\operatorname{Tr}(\mathcal H_1)|$, $m_2'=|\mathcal H_2|$, $k_2=|\operatorname{Tr}(\mathcal H_2)|$, $m_3=|\mathcal F_1'|$, $k_3=|\operatorname{Tr}(\mathcal F_1')|$, $m_4=|\mathcal F_2'|$ and $k_4=|\operatorname{Tr}(\mathcal F_2')|$. Note that $n(\mathcal H_1)=n_1$, $n(\mathcal H_2)=n(\mathcal F_1')=n(\mathcal F_2')=n_2$. By the assumptions in this case, $n_1+n_2=n$, $n_1,n_2\ge 1$, and $m_1,m_3,m_4\ge1$.
Let $k_{10}=|\mathcal T_0|+|\mathcal T|$, $k_{11}=|\mathcal T_1|$ and $k_{11}=|\mathcal T_2|$. Then $k_1=k_{10}+k_{11}+k_{12}$ and by the assumptions of this case and Lemma~\ref{lemma-maxk13}, we have $k_{10},k_{11},k_{12},k_2,k_3,k_4\ge 1$. Moreover, Lemma~\ref{lemma-maxk14} implies that
\begin{equation}
\label{max-e1}
k_{10}k_2+k_{11}k_3+k_{12}k_4\le k.
\end{equation}
Let $\mu_1=\mu(\mathcal H_1)$, $\mu_1=\mu(\mathcal H_2)$, $\mu_3=\mu(\mathcal F_1')$ and $\mu_4=\mu(\mathcal F_2')$.
If $m_2'=0$ then $\operatorname{Tr}(\mathcal H_2)=\{\emptyset\}$ and hence we do not recurse on $\mathcal H_2$, giving thus the recurrence:
\begin{align}\label{rec1-1--}
T(\mu)\le 1+T(\mu_1)+T(\mu_3)+T(\mu_4).
\end{align}
In the general case, we get the recurrence
\begin{align}\label{rec1-1-}
T(\mu)\le 1+T(\mu_1)+T(\mu_2)+T(\mu_3)+T(\mu_4).
\end{align}
Note that \raf{max-e1} implies by Fact~\ref{fact2} that $\sum_{i=1}^4k_i\le k+3$. Using $n_1,n_2\ge 1$ and $k_1,k_2+k_3+k_4\ge 3$, we get by Fact~\ref{fact1} that
\begin{align}\label{e-r}
\mu_1+\mu_2+\mu_3+\mu_4&=n_1m_1k_1+n_2m_2'k_2+n_2m_3k_3+n_2m_4k_4\nonumber\\
&\le m[n_1k_1+n_2(k_2+k_3+k_4)]\le m[3+(n-1)k]=\nonumber\\&m[nk-(k-3)]\le\mu-1,
\end{align}
for $k\ge 4$. Using \raf{e-r} in \raf{rec1-1--} and \raf{rec1-1--}, we get by induction that $T(\mu)\le\mu$ in both cases.
\subsubsection*{Case IV}
By Lemma \ref{lemma-maxk12}, we can compute in polynomial time $\operatorname{Tr}(\mathcal H)$ from $\operatorname{Tr}(\mathcal H_1)$ and
$\operatorname{Tr}(\mathcal H_2)$, and $\operatorname{Tr}(\mathcal F_i)$ ($i=1,2)$.
However, in Case {\bf IV}, $\mathcal T\cup \mathcal T_0$ might be empty. Hence,
it is not clear from Lemma \ref{lemma-maxk14} that the sum of their volume $\sum_{i=1}^2(\mu(\mathcal H_i)+\mu(\mathcal F_i'))$ is bounded by $\mu(\mathcal H)$.
We therefore deompose $\mathcal H$ in the way different from {\bf Case III}, if $\mathcal T_0$ is empty.
Since $\mathcal P\not=\emptyset, \{S_0\cup S_1 \cup S_2\}$, there exist a vertex $v^* \in S_0\cup S_1 \cup S_2$ and a
hyperedge $P \in \mathcal P$ such that $v^* \not\in P$.
Let us assume without loss of generality that $v^* \in S_0 \cup S_2$, and define
\begin{eqnarray}
\mathcal F_1^*&=&\{F \cup \{v^*\} \mid F \in \mathcal F_1^{V_2}\}\cup \mathcal H_2.
\end{eqnarray}
In other word, any hyperedge $H^* \in \mathcal F_1^* \setminus \mathcal H_2$ is obtained from $H \in \mathcal F_1$ by replacing $S_0 \cup S_2$ by vertex $v^*$.
Let $\mathcal T^0$ be a family of minimal transversals $T$ of $\mathcal F_1^*$ with $v^*\not\in T$, and
let $\mathcal T^1$ be a family of minimal transversals $T$ of $\mathcal F_1^*$ with $v^*\in T$.
\begin{lemma}
\label{lemma-kz46-1}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1}.
Then
we have $\operatorname{Tr}(\mathcal H_2) \subseteq (\mathcal T^1)^{V_2} \cup \mathcal T^0$ and $\mathcal T^0= \operatorname{Tr}(\mathcal F_1')$.
\end{lemma}
By Lemmas \ref{lemma-maxk12} and \ref{lemma-kz46-1}, we have the following inclusion.
\begin{lemma}
\label{lemma-maxk12edg1}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1}. Assume that $\mathcal T,\mathcal T_0=\emptyset$ and $\mathcal P\not=\emptyset$.
Then we have
\[
\operatorname{Tr}(\mathcal H) \subseteq \Bigl(\bigcup_{i=1,2} \{ T \cup \{v\} \mid T \in \mathcal T_i, v \in S_0 \cup S_{3-i} \}\Bigr) \wedge
\Bigl((\mathcal T^1)^{V_2} \cup \mathcal T^0\Bigr)
\cup \mathcal T_1 \wedge \mathcal T^0 \cup \mathcal T_2 \wedge \operatorname{Tr}(\mathcal F_2').
\]
\end{lemma}
By this lemma, if $\mathcal T,\mathcal T_0=\emptyset$ and $\mathcal P\not=\emptyset$,
we can compute in polynomial time $\operatorname{Tr}(\mathcal H)$ from $\operatorname{Tr}(\mathcal H_1)\,(=\mathcal T_1 \cup \mathcal T_2)$ and
$\operatorname{Tr}(\mathcal F_1^*)\,(=\mathcal T^0 \cup \mathcal T^1)$, and $\operatorname{Tr}(\mathcal F'_2)$
We next bound the size of three transversal hypergraphs $\operatorname{Tr}(\mathcal H_1)$ and
$\operatorname{Tr}(\mathcal F_1^*)$, and $\operatorname{Tr}(\mathcal F'_2)$.
\begin{lemma}
\label{lemma-kz46-2}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1} and $\mathcal T=\emptyset$.
Assume that $v^* \in S_0 \cup S_2$ and $P \in \mathcal P$ satisfy $v^* \not\in P$.
Then for any $T \in \mathcal T_1$ and any $U \in \mathcal T^1$,
$T \cup U$ is a minimal transversal of $\mathcal H$.
\end{lemma}
\noindent {\bf Proof}.~~
By definition, $T \cup U$ is a transversal of $\mathcal H$.
Since $U \in \mathcal T^1$, for any $u \in U$, $T \cup (U\setminus \{u\})$ is not a transversal of $\mathcal H$.
Assuming that $T' \cup U$ is a minimal transversal of $\mathcal H$ such that $T' \subsetneq T$, we derive a contradiction.
Since $T$ is a minimal transversal of $\mathcal H_1$, $T'$ is not a transversal of $\mathcal H_1$.
However, by our assumption, $T' \cup \{v^*\}$ is a transversal of $\mathcal H_1$, implying that there exist a minimal transversal $T'' \cup \{v^*\}$ of $\mathcal H_1$ such that $T'' \subseteq T'$.
Note that $T'' \cap S_1 \not=\emptyset$, since we have a hyperedge $P\,(\in \mathcal P)$ with $v^*\not\in P$.
This implies that $T'' \cup \{v^*\} \in \mathcal T$, a contradiction. \hfill$\square$\bigskip
\begin{lemma}
\label{lemma-maxk46-36}
Let $\mathcal H$ be a hypergraph that satisfies \raf{eq--aed1} and $\mathcal T=\emptyset$.
Let $v^* \in S_0 \cup S_2$ and $P \in \mathcal P$ with $v^* \not\in P$.
Then we have
\begin{equation}
\mathcal T_1 \dot{\wedge} \operatorname{Tr}(\mathcal F_1^*)\,\ \dot{\cup} \,\ \mathcal T_2\dot{\wedge}\operatorname{Tr}(\mathcal F_2') \ \subseteq \ \operatorname{Tr}(\mathcal H).
\end{equation}
\end{lemma}
\noindent
\noindent {\bf Proof}.~~
By definition and Lemma \ref{lemma-kz46-1},
we have $\operatorname{Tr}(\mathcal F_1^*)=\mathcal T^0 \dot{\cup} \mathcal T^1=\operatorname{Tr}(\mathcal F_1')\dot{\cup} \mathcal T^1$.
Thus,
\[\mathcal T_1 \dot{\wedge} \operatorname{Tr}(\mathcal F_1^*)\,\ \dot{\cup} \,\ \mathcal T_2\dot{\wedge}\operatorname{Tr}(\mathcal F_2') =
\mathcal T_1\dot{\wedge} \mathcal T^1 \,\ \dot{\cup} \,\ \mathcal T_1\dot{\wedge}\operatorname{Tr}(\mathcal F_1') \,\ \dot{\cup} \,\ \mathcal T_2\dot{\wedge}\operatorname{Tr}(\mathcal F_2')
\]
By Lemmas \ref{lemma-maxk13} and \ref{lemma-kz46-2}, it is contained in $\operatorname{Tr}(\mathcal H)$.
\hfill$\square$\bigskip
In {\bf Case IV},
if $\mathcal T_0$ is nonempty, then we decompose $\mathcal H$ in the way described in {\bf Case III}.
Otherwise,
based on Lemmas \ref{lemma-maxk12edg1} and \ref{lemma-maxk46-36},
we decomposes $\mathcal H$ into
$3$ hypergraphs $\mathcal H_1$, $\mathcal F_1^*$, and $\mathcal F_2'$,
where we assume that $v^* \in S_0 \cup S_2$.
We note that $\operatorname{Tr}(\mathcal H)$ can be computed from $\operatorname{Tr}(\mathcal H_1)$, $\operatorname{Tr}(\mathcal F_1^*)$, and $\operatorname{Tr}(\mathcal F_2')$
in $L_2(\mu)=\operatorname{poly}(n,m,k)$ time.
In order to analyze $T(\mu)$, let
$n_1=|V_1|$, $m_1= |\mathcal H_1|$, $k_1=\operatorname{Tr}(\mathcal H_1)$,
$n'_2=|V_2|+1$, $m'_2= |\mathcal F_1^*|$, $k_2=\operatorname{Tr}(\mathcal F_1^*)$,
$n_3=|V_2|$, $m_3= |\mathcal F_2'|$, and $k_3=\operatorname{Tr}(\mathcal F_2')$.
Note that all are positive, $n_1,n'_2\leq n-1$, $n_1+n_2=n+1$, and $m_1,m_2',m_3'\leq m$.
Since $|\mathcal T_1|, |\mathcal T_2|, k_2, k_3 \geq 1$, it follows from Fact \ref{fact2} and Lemma \ref{lemma-maxk46-36} that $k_1,k_2\leq k$ and
$k_1+k_2+k_3\leq k+2$. Thus,
\begin{eqnarray}
n_1m_1k_1+n'_2m'_2k_2+n_3m_3k_3 &\leq& (n_1k_1+n'_2k_2+n_3k_3)m \nonumber\\
&\leq& (n_1k_1+n'_2(k_2+k_3))m \nonumber\\
&\leq& ((n-1)k+ 4)m \label{eq-89edes}\\
&\leq& nmk-1, \label{eq-89edes1}
\end{eqnarray}
where \raf{eq-89edes} is obtained from Fact \ref{fact1} and
\raf{eq-89edes1} is obtained from $k\geq 5$.
This implies that $T(\mu) \leq \mu$.
\subsection{$3$-sum decomposition - case 3}
Let
$\mathcal H_1=\mathcal H_{V_1}$ and $\mathcal H_2=\mathcal H_{V_2}$.
By Corollary \ref{c-decomp}, we have two nonempty disjoint sets $S_1$ and $S_2$ in $V_2$.
Define
\begin{eqnarray}
\mathcal F_0&=&\{H \in \mathcal H \mid H \cap V_1= S_1 \cup S_2, H \cap V_2\not=\emptyset\},\\
\mathcal F_1&=&\{H \in \mathcal H \mid H \cap V_1= S_2, H \cap V_2\not=\emptyset\},\\
\mathcal F_2&=&\{H \in \mathcal H \mid H \cap V_1= S_1, H \cap V_2\not=\emptyset\}.
\end{eqnarray}
Note that $V_1,V_2\not=\emptyset$, $\mathcal H_1\not=\emptyset$, and at least two of $\mathcal F_i$ are nonempy.
$\mathcal H$ can be partitioned in the following way.
\begin{equation}
\label{eq--aed1-a}
\mathcal H=\mathcal H_1 \dot{\cup} \mathcal H_2\dot{\cup} \mathcal F_0 \dot{\cup} \mathcal F_1 \dot{\cup} \mathcal F_2,
\end{equation}
where $\dot{\cup}$ denotes the disjoint union.
For $i=1,2$, let
\begin{equation}
\mathcal T_i=\{T \in \operatorname{Tr}(\mathcal H_1) \mid T\cap S_i \not=\emptyset, T \cap S_j=\emptyset\,\, (j\not=i)\},
\end{equation}
and let
\begin{equation}
\mathcal T=\operatorname{Tr}(\mathcal H_1) \setminus (\mathcal T_1 \cup \mathcal T_2).
\end{equation}
By definition, we have
\begin{equation}
\operatorname{Tr}(\mathcal H_1)=\mathcal T \dot{\cup} \mathcal T_1 \dot{\cup} \mathcal T_2.
\end{equation}
Let
\begin{equation}
\mathcal P=\mathcal H_{ S_1\cup S_2}\,\,(=\{H \in \mathcal H \mid H \subseteq S_1 \cup S_2\}).
\end{equation}
We separately consider the following 4 cases, similar to the case analysis in Section
\ref{section-3-sum-case2}
\medskip
\noindent
{\bf Case I}: $\mathcal P=\emptyset$.
\smallskip
\noindent
{\bf Case II}: $\mathcal P=\{S_1 \cup S_2\}$.
\smallskip
\noindent
{\bf Case III}: $\mathcal P\not= \emptyset, \{S_1 \cup S_2\}$ and $\mathcal T\not=\emptyset$.
\smallskip
\noindent
{\bf Case IV}: $\mathcal P\not= \emptyset, \{S_1 \cup S_2\}$ and $\mathcal T=\emptyset$.
\medskip
\noindent
In {\bf Case I}, we can decompose $\mathcal H$ into $\mathcal H_{1}$ and $\mathcal H_{S_1\cup S_2\cup V_2}$, where the proof is similar to {\bf Case I} in Section \ref{section-3-sum-case2}.
In {\bf Case II}, we note that $|\mathcal H_1|\geq 2$ holds.
To see this, recall from Corollary \ref{c-decomp} that $\mathcal H$ is irredundant Sperner and
has neither identical vertices nor vertices of degree $1$.
Moreover, we have $|\mathcal H_1|+|V_1|\geq 4$ from $3$-sum--case 3 (ii) in Corollary \ref{c-decomp}, which implies $|\mathcal H_1|\geq 2$.
We therefore decompose $\mathcal H$ into $\mathcal H_{1}$ and $\mathcal H_{S_1\cup S_2\cup V_2}$ in a way similar to {\bf Case II-I} in Section \ref{section-3-sum-case2}, where the proof is also similar to {\bf Case II-I} in Section \ref{section-3-sum-case2}.
For the other cases, we note that $\mathcal F_0=\emptyset$ since $\mathcal P\not=\emptyset$ and $\mathcal H$ is Sperner.
Therefore, decompositions similar to the corresponding cases in Section \ref{section-3-sum-case2} work properly.
\section*{Acknowledgements}
We thank Endre Boros and Vladimir Gurvich for helpful discussions.
The research of the second author is supported by Grant-in-Aid for Scientific Research 24106002, 25280004, 26280001 and JST CREST Grant Number JPMJCR1402, Japan.
\newcommand{\etalchar}[1]{$^{#1}$}
|
\section{Introduction}
Although all the elementary particles in the standard model (SM) have now been discovered, with the discovery of the Higgs boson,
there still exist phenomena which cannot be explained in the framework of the SM.
One of these is the neutrino oscillation phenomenon,
which implies two non-zero neutrino mass squared differences and two large lepton mixing angles.
In order to explain this, many authors propose a neutrino flavor model with non-Abelian discrete flavor symmetry
in the lepton sector (for reviews see~\cite{Ishimori:2010au,Ishimori:2012zz,Ishimori:2013woa,King:2014nza}).
Even before the discovery of the non-zero $\theta _{13}$~\cite{An:2012eh,Ahn:2012nd,Abe:2014lus},
a few authors suggested a tiny mixing angle $\theta _{13}$ based on non-Abelian discrete flavor symmetry~\cite{Shimizu:2011xg}.
Recent results from the T2K and NO$\nu $A experiments~\cite{Abe:2013hdq,Adamson:2016tbq} imply CP violation through the Dirac CP phase.
They studied electron neutrino appearance in a muon neutrino beam.
The Majorana phases are also sources of the CP violating phases if neutrinos are Majorana particles.
The KamLAND-Zen experiment~\cite{KamLAND-Zen:2016pfg} is searching for neutrinoless double beta ($0\nu \beta \beta $) decay to check the Majorana nature of neutrinos.
Therefore, it is important to predict not only mixing angles but also CP phases with the non-Abelian discrete flavor model.
The non-Abelian discrete flavor symmetry can easily explain large lepton mixing angles,
e. g. tri-bimaximal mixing (TBM)~\cite{Harrison:2002er,Harrison:2002kp}, which is a simple framework for the lepton mixing angles.
Indeed, Altarelli and Feruglio (AF) proposed a simple flavor model and predicted TBM by using $A_4$ discrete flavor symmetry~\cite{Altarelli:2005yp,Altarelli:2005yx}.
They introduced $SU(2)$ gauge singlet scalar fields, so-called ``flavons'', and derived the TBM in the lepton sector.
The non-zero $\theta_{13}$ can be realized by another $A_4$ non-trivial singlet flavon~\cite{Shimizu:2011xg} in addition to the flavons introduced by AF.
The origin of non-vanishing $\theta_{13}$ is related to a new contribution to the mass matrices.
Matrices which have the same structure as that in Ref.~\cite{Shimizu:2011xg} also appear in extra-dimensional models with the $S_3$ and $S_4$ flavor symmetries~\cite{Haba:2006dz,Ishimori:2010fs}.
The $\Delta (27)$ model also includes these matrices~\cite{Grimus:2008tt}.
In this paper, we study phenomenological aspects of a supersymmetric model with $A_4\times Z_3$ symmetries.
The three generations of the left-handed leptons are expressed as the $A_4$ triplet, $l=(l_e,l_\mu,l_\tau )$,
while the right-handed charged leptons $e_R$, $\mu_R$, and $\tau_R$ are $A_4$ singlets denoted as ${\bf 1}$, ${\bf 1''}$, and ${\bf 1'}$ respectively.
Three right-handed neutrinos are also described as the triplet of $A_4$.
We introduce the $SU(2)$ gauge singlet flavons of $A_4$ triplets, $\phi_T=(\phi _{T1},\phi _{T2},\phi _{T3})$ and $\phi_S=(\phi _{S1},\phi _{S2},\phi _{S3})$.
In addition, $\xi $ and $\xi '$ are also introduced as the $SU(2)$ gauge singlet flavons with the two kinds of singlet representations of $A_4$, ${\bf 1}$ and ${\bf 1'}$ respectively.
We focus on the vacuum structure of the flavor model.
The scalar sectors of this model consist of many flavons in addition to the SM Higgs boson.
Then, we analyze the scalar potential and show the 24 different sets of VEVs
which come from 24 combinations of 4 (6) possible VEVs of the flavon $\phi_T$ ($\phi_S$).
The 24 different vacua are classified into two types which are not related to each other under the transformations $A_4$.
Therefore, we expect that the two types of vacua have different expressions for the physical observables in terms of the model parameters such as Yukawa couplings.
We ask the following question: whether these different vacua are physically distinct from each other.
The purpose of this paper is to clarify the differences and relations among the VEVs and their physical consequences.
In particular, we investigate the mixing angles, CP violating phase, and effective mass for neutrinoless double beta ($0\nu\beta\beta$) decay.
This paper is organized as follows.
In Section~\ref{sec:model}, we introduce the supersymmetric model with $A_4\times Z_3$ symmetry.
In Section~\ref{sec:2types}, we study the classification of vacua and derive the formulae for the mixing angles and CP phases.
In Section~\ref{sec:num}, we discuss the phenomenological aspects for mixing angles and CP violating phases.
The numerical analyses for the effective mass of $0\nu\beta\beta$ decay are presented.
Section~\ref{sec:con} is devoted to a summary.
In Appendix~\ref{sec:multiplication-rule}, we show the multiplication rule of the $A_4$ group.
\section{Supersymmetric Model with $A_4\times Z_3$ Symmetry}
\label{sec:model}
In this section, we introduce a supersymmetric model with $A_4\times Z_3$ symmetry.
We analyze the scalar potential and derive the mass matrices of the lepton sector.
\subsection{Model}
We introduce three heavy right-handed Majorana neutrinos.
The leptons and scalars in our model are listed in Table~\ref{tb:fields}.
\begin{table}[b]
\caption{\small{
The representations of $SU(2)_L$ and $A_4$, and the charge assignment of $Z_3$ and $U(1)_R$ for leptons and scalars:
~$l_{e,\mu,\tau}$, $\{e,\mu,\tau\}_R$, $\{\nu_e,\nu_\mu,\mu_\tau\}_R$, and $h_{u,d}$ denote left-handed leptons,
right-handed charged leptons, right-handed neutrinos, and Higgs fields, respectively.
The other scalars are gauge singlet flavons and denoted as $\phi_T$, $\phi_S$, $\xi$, and $\xi'$.
$\omega$ is the $Z_3$ charge and stands for $e^{2\pi i/3}$.}}
\hspace{-0.5cm}
\begin{tabular}{|c||c|c|c|c|c||c|c|c|c|c|} \hline
&$l= \begin{pmatrix}
l_e\\
l_\mu\\
l_\tau
\end{pmatrix}$
&$e_R$
&$\mu_R$
&$\tau_R$
&$\nu_R=\begin{pmatrix}
\nu_{eR}\\
\nu_{\mu R}\\
\nu_{\tau R}
\end{pmatrix}$
&$h_{u,d}$
&$\phi_T=\begin{pmatrix}
\phi_{T1}\\
\phi_{T2}\\
\phi_{T3}
\end{pmatrix}$
&$\phi_S=\begin{pmatrix}
\phi_{S1}\\
\phi_{S2}\\
\phi_{S3}
\end{pmatrix}$
&$\xi$
&$\xi'$\\ \hline \hline
$SU(2)_L$&2&1&1&1&1&2&1&1&1&1\\
$A_4$&3&1&$1''$&$1'$&3&1&3&3&1&$1'$\\
$Z_3$&$\omega$&$\omega^2$&$\omega^2$&$\omega^2$&$\omega^2$&1&1&$\omega^2$&$\omega^2$&$\omega^2$\\
$U(1)_R$&1&1&1&1&1&0&0&0&0&0\\ \hline
\end{tabular}\label{tb:fields}
\end{table}
The superpotential of Yukawa interactions is
\begin{align}
w_Y=w_l+w_D+w_R,\label{eq:wY}
\end{align}
where $w_l,w_D$ and $w_R$ are Yukawa interactions for charged lepton, Dirac neutrino and Majorana neutrino sectors respectively:
\begin{align}
&w_l=
y_e(\phi_Tl)_{\bf 1}e_Rh_d/\Lambda+
y_\mu(\phi_Tl)_{\bf 1'}\mu_Rh_d/\Lambda+
y_\tau(\phi_Tl)_{\bf 1''}\tau_Rh_d/\Lambda+h.c., \label{eq:wl} \\
&w_D=y_D(l\nu_R)_{\bf 1}h_u+h.c., \label{eq:wD} \\
&w_R=y_{\phi_S}\phi_S(\nu_R\nu_R)_{\bf 3}+
y_\xi\xi(\nu_R\nu_R)_{\bf 1}+
y_{\xi'}\xi'(\nu_R\nu_R)_{\bf 1''}+h.c., \label{eq:wR}
\end{align}
where the lower indices denote $A_4$ representations.
Moreover, the $y$'s and $\Lambda$ denote the Yukawa coupling constants and cut-off scale respectively.
The multiplication rule for $A_4$ representations is shown in Appendix~\ref{sec:multiplication-rule}.
In order to obtain the mass matrices of these leptons, we analyze the following superpotential of the scalar fields:
\begin{align}
w_d\equiv w_d^T+w_d^S, \label{eq:wd}
\end{align}
where
\begin{align}
&w_d^T=-M(\phi_0^T\phi_T)_{\bf 1}+g\phi_0^T(\phi_T\phi_T)_{\bf 3}, \label{eq:wdT} \\
&w_d^S=
g_1\phi_0^S(\phi_S\phi_S)_{\bf 3}+
g_2(\phi_0^S\phi_S)_{\bf 1}\xi+g_2'(\phi_0^S\phi_S)_{\bf 1''}\xi '+
g_3(\phi_S\phi_S)_{\bf 1}\xi_0-
g_4\xi_0\xi\xi \label{eq:wdS}.
\end{align}
We have introduced the additional $SU(2)$ gauge singlet fields, $\phi_0^T$, $\phi_0^S$ and $\xi_0$, which are called {\it ``driving fields''}.
The charge assignments of these fields are summarized in Table \ref{tb:driving}.
\begin{table}[h!]
\caption{The driving fields and their representations and charge assignment.}
\centering
\begin{tabular}{|c||c|c|c|} \hline
& $\phi_0^T=\begin{pmatrix}
\phi_{01}^T\\
\phi_{02}^T\\
\phi_{03}^T
\end{pmatrix}$
& $\phi_0^S=\begin{pmatrix}
\phi_{01}^S\\
\phi_{02}^S\\
\phi_{03}^S
\end{pmatrix}$
& $\xi_0$ \\ \hline \hline
SU(2) & $1$ & $1$ & $1$ \\
$A_4$ & $3$ & $3$ & $1$ \\
$Z_3$ & $1$ & $\omega^2$ & $\omega^2$ \\
$U(1)_R$ & $2$ & $2$ & $2$ \\ \hline
\end{tabular}\label{tb:driving}
\end{table}
\subsection{Potential Analysis}
In this subsection, we derive the VEVs for the scalar fields $\phi_T,\phi_S,\xi,\xi',\phi^T_0,\phi^S_0,\xi_0$.
One can derive the scalar potential from the superpotentials in Eqs.~\eqref{eq:wdT} and \eqref{eq:wdS} as
\begin{align}
V=V_T+V_S,
\end{align}
where
\begin{align}
V_T=\sum_{X}\left|\frac{\partial w_d^T}{\partial X}\right|^2
=&\left| -M\phi_{T1}+\frac{2}{3}g(\phi_{T1}^2-\phi_{T2}\phi_{T3})\right|^2 \nonumber \\
+&\left| -M\phi_{T3}+\frac{2}{3}g(\phi_{T2}^2-\phi_{T3}\phi_{T1})\right|^2 \nonumber \\
+&\left| -M\phi_{T2}+\frac{2}{3}g(\phi_{T3}^2-\phi_{T1}\phi_{T2})\right|^2 \nonumber \\
+&\left| -M\phi_{01}^T+\frac{2}{3}g(2\phi_{01}^T\phi_{T1}-\phi_{03}^T\phi_{T2}-\phi_{02}^T\phi_{T3})\right|^2 \nonumber \\
+&\left| -M\phi_{03}^T+\frac{2}{3}g(2\phi_{02}^T\phi_{T2}-\phi_{01}^T\phi_{T3}-\phi_{03}^T\phi_{T1})\right|^2 \nonumber \\
+&\left| -M\phi_{02}^T+\frac{2}{3}g(2\phi_{03}^T\phi_{T3}-\phi_{02}^T\phi_{T1}-\phi_{01}^T\phi_{T2})\right|^2 \label{eq:VT},
\end{align}
and
\begin{align}
V_S
=\sum_{Y}\left|\frac{\partial w_d^S}{\partial Y}\right|^2
=&\left|\frac{2}{3}g_1(\phi_{S1}^2-\phi_{S2}\phi_{S3})-g_2\phi_{S1}\xi+g_2'\phi_{S3}\xi'\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(\phi_{S2}^2-\phi_{S3}\phi_{S1})-g_2\phi_{S3}\xi+g_2'\phi_{S2}\xi'\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(\phi_{S3}^2-\phi_{S1}\phi_{S2})-g_2\phi_{S2}\xi+g_2'\phi_{S1}\xi'\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(2\phi_{01}^S\phi_{S1}-\phi_{03}^S\phi_{S2}-\phi_{02}^S\phi_{S3})
-g_2\phi_{01}^S\xi+g_2'\phi_{03}^S\xi'+2g_3\phi_{S1}\xi_0\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(2\phi_{02}^S\phi_{S2}-\phi_{01}^S\phi_{S3}-\phi_{03}^S\phi_{S1})
-g_2\phi_{03}^S\xi+g_2'\phi_{02}^S\xi'+2g_3\phi_{S3}\xi_0\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(2\phi_{03}^S\phi_{S3}-\phi_{02}^S\phi_{S1}-\phi_{01}^S\phi_{S2})
-g_2\phi_{02}^S\xi+g_2'\phi_{01}^S\xi'+2g_3\phi_{S2}\xi_0\right|^2 \nonumber \\
+&\left|-g_2(\phi_{01}^S\phi_{S1}+\phi_{03}^S\phi_{S2}+\phi_{02}^S\phi_{S3})-2g_4\xi\xi_0\right|^2 \nonumber \\
+&\left|g_2'(\phi_{02}^S\phi_{S2}+\phi_{01}^S\phi_{S3}+\phi_{03}^S\phi_{S1})\right|^2 \nonumber \\
+&\left|g_3(\phi_{s1}^2+2\phi_{S2}\phi_{S3})-g_4\xi^2\right|^2 \label{eq:VS}.
\end{align}
The sum for $X,Y$ runs over all the scalar fields:
$$
X=\{\phi_{T1}, \phi_{T2}, \phi_{T3}, \phi_{01}^T, \phi_{02}^T, \phi_{03}^T\}\ , \quad
Y=\{\phi_{S1}, \phi_{S2}, \phi_{S3}, \phi_{01}^S, \phi_{02}^S, \phi_{03}^S, \xi, \xi ', \xi_0\}.
$$
The scalar potential $V$ is minimized at $V=V_T=V_S=0$.
There are several solutions for the minimization condition.
We obtain sets of solutions denoted as $\eta_m$ and $\lambda_n^\pm$ ($m=1$-$4$, $n=1$-$3$),
where $\eta_m$ and $\lambda_n^\pm$ are the solutions of $V_T=0$ and $V_S=0$ respectively.
Hereafter, we call them the set of VEV alignments and show them explicitly as follows:
\begin{align}
&\eta_1\equiv\left\{\langle\phi_T\rangle=v_T\begin{pmatrix}1\\0\\0\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta1}\\
&\eta_2\equiv\left\{\langle\phi_T\rangle=\frac{v_T}{3}\begin{pmatrix}-1\\2\\2\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta2}\\
&\eta_3\equiv\left\{\langle\phi_T\rangle=\frac{v_T}{3}\begin{pmatrix}-1\\2\omega\\2\omega^2\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta3}\\
&\eta_4\equiv\left\{\langle\phi_T\rangle=\frac{v_T}{3}\begin{pmatrix}-1\\2\omega^2\\2\omega\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta4}\\
&\lambda_1^\pm\equiv\left\{\langle\phi_S\rangle=\pm v_S\begin{pmatrix}1\\1\\1\end{pmatrix},\
\langle\xi'\rangle=u',\ \langle\phi_0^S\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:lam1}\\
&\lambda_2^\pm\equiv\left\{\langle\phi_S\rangle=\pm v_S\begin{pmatrix}1\\\omega\\\omega^2\end{pmatrix},\
\langle\xi'\rangle=\omega u',\ \langle\phi_0^S\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:lam2}\\
&\lambda_3^\pm\equiv\left\{\langle\phi_S\rangle=\pm v_S\begin{pmatrix}1\\\omega^2\\\omega\end{pmatrix},\
\langle\xi'\rangle=\omega^2u',\ \langle\phi_0^S\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:lam3}
\end{align}
where $v_T=\frac{3M}{2g}$, $v_S=\sqrt{\frac{g_4}{3g_3}}u$, $u'=\frac{g_2}{g_2'}u$ and $u$ is the VEV of $\xi$, $\langle\xi\rangle=u$
\footnote{
There are still other solutions for $V=0$, including the trivial solution which makes all the VEVs vanish.
It leads to the vanishing of all the lepton masses and mixing angles.
In addition to the trivial solution, there are solutions with non-zero VEVs of the driving fields.
This case leads to the breakdown of $U(1)_R$ symmetry.
In this paper, we only discuss the vacua where $U(1)_R$ symmetry is conserved.}.
The superscript of $\lambda^\pm$ denotes the overall sign of the VEV $\langle\phi_S\rangle$.
In total, we obtain 24 sets of vacua, since there are four sets of alignment for $\eta_m$ and six sets for $\lambda^\pm_n$.
\subsection{Mass Matrix for Charged Leptons and Neutrinos}
We derive charged lepton mass matrices and neutrino mass matrices from the Yukawa interactions in Eqs.~(\ref{eq:wl}),(\ref{eq:wD}), and (\ref{eq:wR}).
These matrices are expressed in various forms corresponding to the VEV alignments.
The charged lepton mass matrices ${M_l}^{(m)}$ for Eqs.~\eqref{eq:eta1}-\eqref{eq:eta4} are
\begin{align}
&{M_l}^{(1)}=\frac{v_dv_T}{\Lambda}
\begin{pmatrix}
y_e&0&0\\
0&y_\mu&0\\
0&0&y_\tau
\end{pmatrix},
\label{eq:Ml1} \\[0.3cm]
&{M_l}^{(2)}=\frac{v_dv_T}{3\Lambda}
\begin{pmatrix}
-y_e&2y_\mu&2y_\tau\\
2y_e&-y_\mu&2y_\tau\\
2y_e&2y_\mu&-y_\tau
\end{pmatrix}
=S{M_l}^{(1)}, \label{eq:Ml2} \\[0.3cm]
&{M_l}^{(3)}=\frac{v_dv_T}{3\Lambda}
\begin{pmatrix}
-y_e&2\omega y_\mu&2\omega^2y_\tau\\
2\omega^2y_e&-y_\mu&2\omega y_\tau\\
2\omega y_e&2\omega^2y_\mu&-y_\tau
\end{pmatrix}
=T^\dag ST{M_l}^{(1)}, \label{eq:Ml3} \\[0.3cm]
&{M_l}^{(4)}=\frac{v_dv_T}{3\Lambda}
\begin{pmatrix}
-y_e&2\omega^2y_\mu&2\omega y_\tau\\
2\omega y_e&-y_\mu&2\omega^2y_\tau\\
2\omega^2y_e&2\omega y_\mu&-y_\tau
\end{pmatrix}
=TST^\dag{M_l}^{(1)}, \label{eq:Ml4}
\end{align}
respectively, where the matrices $S$ and $T$ are
\begin{align}
S=\frac{1}{3}\begin{pmatrix}
-1&2&2\\
2&-1&2\\
2&2&-1
\end{pmatrix}\ ,\quad
T=\begin{pmatrix}
1&0&0\\
0&\omega&0\\
0&0&\omega^2
\end{pmatrix}.
\label{eq:ST}
\end{align}
The Dirac mass matrix for neutrinos obtained from Eq.~\eqref{eq:wD} is
\begin{align}
M_D=y_Dv_u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}.
\label{eq:MD}
\end{align}
It is noted that the Dirac mass matrix is determined independently of the VEV alignments.
The Majorana mass matrices ${M_R^{(n)}}^\pm$ for the corresponding set of solutions Eqs.~\eqref{eq:lam1},\eqref{eq:lam2},\eqref{eq:lam3} are given as follows:
\begin{align}
{M_R^{(1)}}^\pm&=
\pm\frac{1}{3}y_{{\phi_S}}v_S\begin{pmatrix}
2&-1&-1\\
-1&2&-1\\
-1&-1&2
\end{pmatrix}+y_\xi u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+y_{\xi'}u'\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}, \label{eq:MR1} \\
{M_R^{(2)}}^\pm&=
\pm\frac{1}{3}y_{{\phi_S}}v_S\begin{pmatrix}
2&-\omega^2&-\omega\\
-\omega^2&2\omega&-1\\
-\omega&-1&2\omega^2
\end{pmatrix}+y_\xi u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+\omega y_{\xi'}u'\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}\nonumber\\
&=T^\dag{M_R^{(1)}}^\pm T^\dag, \label{eq:MR2} \\
{M_R^{(3)}}^\pm&=
\pm\frac{1}{3}y_{{\phi_S}}v_S\begin{pmatrix}
2&-\omega&-\omega^2\\
-\omega&2\omega^2&-1\\
-\omega^2&-1&2\omega
\end{pmatrix}+y_\xi u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+\omega^2y_{\xi'}u'\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}\nonumber\\
&=T{M_R^{(1)}}^\pm T. \label{eq:MR3}
\end{align}
In order to generate the light neutrino mass matrices, we adopt the seesaw mechanism~\cite{Minkowski,Yanagida:1979as,GellMann:1980vs}.
The effective neutrino mass matrices are given by the well-known formula, $M_\nu=-M_DM_R^{-1}M_D^\mathrm{T}$,
through the seesaw mechanism.
We obtain the 6 different effective neutrino mass matrices from Eqs.~\eqref{eq:MD}-\eqref{eq:MR3} as follows:
\begin{align}
&{M_\nu^{(1)}}^\pm=\pm a\begin{pmatrix}
1&0&0\\
0&1&0\\
0&0&1
\end{pmatrix}+b^\pm\begin{pmatrix}
1&1&1\\
1&1&1\\
1&1&1
\end{pmatrix}+c\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+d\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}, \label{eq:Mv1}\\
&{M_\nu^{(2)}}^\pm=T^\dag{M_\nu^{(1)}}^\pm T^\dag, \label{eq:Mv2}\\
&{M_\nu^{(3)}}^\pm=T{M_\nu^{(1)}}^\pm T, \label{eq:Mv3}
\end{align}
where
\begin{align}
&a=ky_{{\phi_S}}v_S, \nonumber\\
&c=k(y_{\xi'}u'-y_\xi u), \nonumber\\
&d=ky_{\xi'}u', \nonumber\\
&b^\pm=\mp\frac a3+\frac{a^2}{2d-c}\left(\frac13-\frac{d^2}{a^2}\right), \nonumber\\
&k=\frac{{y_D}^2{v_u}^2}{y_\xi^2u^2+y_{\xi'}^2{u'}^2-(y_{\phi_S}^2v_S^2+y_\xi uy_\xi'u')}. \nonumber
\end{align}
\section{Classification of Vacua and PMNS Mixing Matrix}\label{sec:2types}
In this section, we classify the 24 different vacua and derive the lepton mixing matrix $U_\mathrm{PMNS}$, called the {\it Pontecorvo-Maki-Nakagawa-Sakata} (PMNS) mixing matrix.
In order to classify the vacua, we discuss the relations among the VEV alignments with the transformations of $A_4$.
We show that the 24 vacua are classified into two types in the following subsection.
Then, one finds the two different PMNS matrices with diagonalizing matrices for the charged lepton and effective neutrino mass matrices Eqs.~(\ref{eq:Ml1})-(\ref{eq:Ml4}), and (\ref{eq:Mv1}),(\ref{eq:Mv2}),(\ref{eq:Mv3}).
\subsection{Relations among Sets of VEV Alignments}
The generators of $A_4$ are expressed as the following forms for the representations ${\bf 1},{\bf 1'},{\bf 1''}$ and ${\bf 3}$,
\begin {align}
&S({\bf 1})=S({\bf 1'})=S({\bf 1''})=1\ ,\quad S({\bf 3})
=\frac{1}{3}\begin{pmatrix}
-1&2&2\\
2&-1&2\\
2&2&-1
\end{pmatrix},
\label{eq:Strans}\\
&T({\bf 1})=1,\ T({\bf 1'})=\omega,\ T({\bf 1''})=\omega^2\ ,\quad T({\bf 3})=
\begin{pmatrix}
1&0&0\\
0&\omega&0\\
0&0&\omega^2
\label{eq:Ttrans}
\end{pmatrix}.
\end{align}
The sets of VEV alignment $\eta_m,\lambda^\pm_n$ are associated through the transformations of these generators.
As an example, we show the $T$ transformation on $\lambda_1^+$:
\begin{align}
T\left[\lambda_1^+\right]&\equiv
\left\{\langle\phi_S\rangle=T({\bf 3})v_S
\begin{pmatrix}
1\\
1\\
1
\end{pmatrix},
\langle\xi'\rangle=T({\bf 1'})u',
\langle\phi_0^S\rangle=T({\bf 3})
\begin{pmatrix}
0\\
0\\
0
\end{pmatrix}\right\}\nonumber \\
&=\left\{\langle\phi_S\rangle=v_S
\begin{pmatrix}
1\\
\omega\\
\omega^2
\end{pmatrix},
\langle\xi'\rangle=\omega u',\
\langle\phi_0^S\rangle=
\begin{pmatrix}
0\\
0\\
0
\end{pmatrix}\right\}=\lambda_2^+.
\end{align}
The $S$ and $T$ transformations on all the sets of the VEV alignment are summarized in Fig.~\ref{fig:map}.
Some transformations preserve the VEVs of either $\eta_m$ or $\lambda^\pm_n$.
These vacua have $Z_3$ or $Z_2$ symmetries as the residual symmetries of $A_4$ respectively.
For the VEVs described as $\eta_m$, they are invariant under the following transformation,
\begin{align}
T\left[\eta_1\right]=T^{-1}\left[\eta_1\right]=\eta_1\ ,\quad
TST\left[\eta_2\right]=(TST)^{-1}\left[\eta_2\right]=\eta_2 \ ,\nonumber\\
ST\left[\eta_3\right]=(ST)^{-1}\left[\eta_3\right]=\eta_3\ ,\quad
TS\left[\eta_4\right]=(TS)^{-1}\left[\eta_4\right]=\eta_4\ .
\end{align}
It is easy to confirm that such transformations correspond to $Z_3$ symmetries:
\begin{align}
T^3=(TST)^3=(ST)^3=(TS)^3=1.
\end{align}
Each $\lambda^\pm_n$ has $Z_2$ symmetry as follows:
\begin{align}
S\left[\lambda^\pm_1\right]=\lambda^\pm_1\ ,\quad TST^2\left[\lambda^\pm_2\right]
=\lambda^\pm_2\ ,\quad T^2ST\left[\lambda^\pm_3\right]=\lambda^\pm_3,
\end{align}
where
\begin{align}
S^2=(TST^2)^2=(T^2ST)^2=1.
\end{align}
\begin{figure}[h]
\begin{center}
\includegraphics[width=7.0cm]{vevtransition.pdf}
\caption{
Map of the transitions among the VEV alignments under the transformations $S$ and $T$:
The solid arrow corresponds to the transition due to $T$ transformation
and the dashed two headed arrow shows the transition due to $S$ transformation.
In the map, $ \eta_1$ is invariant under $T$ transformation
while $\lambda_1^{\pm}$ are invariant under $S$ transformation. }
\label{fig:map}
\end{center}
\end{figure}
\subsection{Classification of 24 Vacua}
In this subsection, we show the relations among the 24 different Lagrangians derived from the 24 different combinations of VEV alignments in Eqs.~\eqref{eq:eta1}-\eqref{eq:lam3}.
We find the two sets of 12 equivalent Lagrangians with the appropriate field redefinitions.
Then, the 24 Lagrangians are classified into two types.
For simplicity, we write the Lagrangian of this model in a short form:
\begin{align}
\mathcal{L}(\psi,\phi_1,\phi_2)\label{eq:L},
\end{align}
where $\psi$ represents the fermion fields such as $l$ and $\nu_R$.
$\phi_1$ and $\phi_2$ represent the scalar fields,
which should have their VEVs written as $\eta_m$ and $\lambda_n^\pm$ respectively.
We write the Lagrangian in the broken phase for the VEV alignment ($\eta_m$, $\lambda^\pm_n$) with fluctuations $h_1$ and $h_2$ as
\begin{align}
\mathcal{L}_{mn}^\pm(\psi,h_1,h_2)\equiv\mathcal{L}(\psi,\eta_m+h_1,\lambda_n^\pm+h_2).
\end{align}
Then, we prove the following equation:
\begin{align}
\mathcal{L}(\psi',\eta_m+h_1',\lambda^\pm_n+h_2')=
\mathcal{L}(\psi,G^{-1}\eta_m+h_1,G^{-1}\lambda_n^\pm+h_2),\label{eq:eqre}
\end{align}
where $G$ denotes the transformation composed of $S$ and $T$ in Eqs.~\eqref{eq:Strans} and \eqref{eq:Ttrans}.
There are 12 independent transformations including the identity element:
\begin{align}
G;\{e,\ T,\ T^2,\ S,\ TS,\ T^2S,\ ST,\ ST^2,\ T^2ST,\ TST,\ TST^2,\ T^2ST^2\}.
\label{eq:allG}
\end{align}
The redefined fields are written as follows,
\begin{align}
\psi'=G\psi\ ,\quad h_i'=Gh_i\quad(i=1,2).\label{eq:redef}
\end{align}
The right-hand side of Eq.~\eqref{eq:eqre} corresponds to the Lagrangian for the vacuum ($G^{-1}\eta_m,G^{-1}\lambda^\pm_n$) while the left-hand side is the Lagrangian for the vacuum ($\eta_m,\lambda_n^\pm$) in terms of the redefined fields.
In the symmetric phase, the Lagrangian $\mathcal{L}(\psi,\phi_1,\phi_2)$ is invariant under the $G$ transformation,
\begin{align}
\mathcal{L}(G\psi,G\phi_1,G\phi_2)=\mathcal{L}(\psi,\phi_1,\phi_2).\label{eq:INV}
\end{align}
One obtains the following equation from Eq.~\eqref{eq:INV} for the vacuum ($G^{-1}\eta_m,G^{-1}\lambda^\pm_n$),
\begin{align}
\mathcal{L}(G\psi,\eta_m+Gh_1,\lambda_n^\pm+Gh_2)
=&\mathcal{L}(\psi,G^{-1}\eta_m+h_1,G^{-1}\lambda_n^\pm+h_2).
\label{eq:aeqre}
\end{align}
Finally, one obtains the relation Eq.~\eqref{eq:eqre} by applying the field definition Eq.~\eqref{eq:redef} to the left-hand side of Eq.~\eqref{eq:aeqre}.
The relation Eq.~\eqref{eq:eqre} implies the equality of the Lagrangians for the two vacua ($\eta_m,\lambda^\pm_n$) and ($G^{-1}\eta_m,G^{-1}\lambda^\pm_n$).
Here, we briefly show how to find the equivalent vacua with Fig.~\ref{fig:map}.
For example, let us consider the $T$ transformation in terms of the vacuum of $(\eta_1,\lambda^+_1)$.
One finds that $\eta_1$ is invariant and $\lambda^+_1$ transfers to $\lambda^+_2$ under the $T$ transformation.
Therefore, $\mathcal{L}^+_{11}$ and $\mathcal{L}^+_{12}$ are equivalent.
One can find 12 equivalent vacua by applying 12 independent transformations in Eq.~\eqref{eq:allG} to the vacuum $(\eta_1,\lambda^+_1)$.
Then, we classify the 24 Lagrangians into two types:
\begin{align}
&\mathrm{Type\ I};
\{\mathcal{L}^+_{11},\mathcal{L}^+_{12},\mathcal{L}^+_{13},\mathcal{L}^+_{21},\mathcal{L}^+_{32},\mathcal{L}^+_{43},
\mathcal{L}^-_{22},\mathcal{L}^-_{23},\mathcal{L}_{31}^-,\mathcal{L}^-_{33},\mathcal{L}^-_{41},\mathcal{L}^-_{42}\},\\
&\mathrm{Type\ I\hspace{-.1em}I};
\{\mathcal{L}^-_{11},\mathcal{L}^-_{12},\mathcal{L}^-_{13},\mathcal{L}^-_{21},\mathcal{L}^-_{32},\mathcal{L}^-_{43},
\mathcal{L}^+_{22},\mathcal{L}^+_{23},\mathcal{L}^+_{31},\mathcal{L}^+_{33},\mathcal{L}^+_{41},\mathcal{L}^+_{42}\}.
\end{align}
Type I and type I\hspace{-.1em}I are disconnected because of the absence of a transformation which relates one type to the other.
Since all the Lagrangians which belong to the same type lead to the same physical consequences,
we consider only $\mathcal{L}^+_{11}$ and $\mathcal{L}^-_{11}$ as the representatives of their types:
\begin{align}
\mathcal{L}^\mathrm{I}\equiv\mathcal{L}^+_{11}\ ,\ \ \mathcal{L}^{\mathrm{I\hspace{-.1em}I}}\equiv\mathcal{L}^-_{11}.
\end{align}
We also define the representative mass matrices for charged leptons and neutrinos as
\begin{align}
M_l\equiv M_l^{(1)}\ ,\quad
M_\nu^\mathrm{I}\equiv {M_\nu^{(1)}}^+\ ,\quad
M_\nu^\mathrm{I\hspace{-.1em}I}\equiv {M_\nu^{(1)}}^-.
\end{align}
It is noted that the charged lepton mass matrix $M_l^{(1)}$ is diagonal.
\subsection{PMNS Matrices for Two Types}\label{sec:PMNS}
In this subsection, we construct the PMNS matrices for the two types,
$\mathcal{L}^\mathrm{I}$ and $\mathcal{L}^{\mathrm{I\hspace{-.1em}I}}$.
Since the charged lepton mass matrix $M_l$ is diagonal,
the PMNS matrix is determined so that it diagonalizes the neutrino mass matrices in Eq.~(\ref{eq:Mv1}):
\begin{align}
(U_\mathrm{PMNS}^\mathrm{I})^\dag M_\nu^\mathrm{I}(U_\mathrm{PMNS}^\mathrm{I})^*=
(U_\mathrm{PMNS}^\mathrm{I\hspace{-.1em}I})^\dag M_\nu^\mathrm{I\hspace{-.1em}I}
(U_\mathrm{PMNS}^\mathrm{I\hspace{-.1em}I})^*=
\begin{pmatrix}
m_1&&\\
&m_2&\\
&&m_3
\end{pmatrix},
\label{eq:diag. cond.}
\end{align}
where the left-handed neutrino masses $m_1,m_2$ and $m_3$ are positive.
The PMNS matrices are expressed as the following forms for the two types:
\begin{align}
U_\mathrm{PMNS}^\mathrm{I}&=U_\mathrm{TBM}U_{13}(\theta,\sigma)
\begin{pmatrix}
e^{i\phi_1}&& \\
&e^{i\phi_2}& \\
&&e^{i\phi_3}
\end{pmatrix},
\label{eq:PMNSI}\\
U_\mathrm{PMNS}^\mathrm{I\hspace{-.1em}I}&=U_\mathrm{TBM}\begin{pmatrix}
&&-i\\
&1&\\
i&&
\end{pmatrix}U_{13}(\theta,\sigma)\begin{pmatrix}
e^{i\phi_1}&& \\
&e^{i\phi_2}& \\
&&e^{i\phi_3}
\end{pmatrix}, \nonumber\\
&=U_\mathrm{TBM}U_{13}^*(\theta+\frac{\pi}{2},\sigma)\begin{pmatrix}
-ie^{i(\phi_1+\sigma)}&& \\
&e^{i\phi_2}& \\
&&-ie^{i(\phi_3-\sigma)}
\end{pmatrix}.
\label{eq:PMNSII}
\end{align}
The unitary matrix $U_\mathrm{TBM}$ is the tri-bimaximal mixing matrix and $U_{13}(\theta,\sigma)$ denotes the unitary rotation matrix:
\begin{align}
&U_\mathrm{TBM}=
\begin{pmatrix}
2/\sqrt{6}&1/\sqrt{3}&0\\
-1/\sqrt{6}&1/\sqrt{3}&-1/\sqrt{2}\\
-1/\sqrt{6}&1/\sqrt{3}&1/\sqrt{2}
\end{pmatrix},\label{eq:TBM}\\
&U_{13}(\theta,\sigma)=
\begin{pmatrix}
\mathrm{cos}\theta&0&e^{-i\sigma}\mathrm{sin}\theta\\
0&1&0\\
-e^{i\sigma}\mathrm{sin}\theta&0&\mathrm{cos}\theta
\end{pmatrix}.\label{eq:R13}
\end{align}
We have introduced the parameters $\theta,\sigma$ and $\phi_i$ ($i=1,2,3$).
They are written in terms of the complex parameters of the neutrino mass matrix, $a,b,c$ and $d$, in Eq.~\eqref{eq:Mv1}~\footnote{
There are six real parameters since $b$ is written by using $a,c,d$.}.
In the rest of this subsection, we derive the explicit forms of the parameters $\theta,\sigma$ and $\phi_i$ in terms of the model parameters $a,b,c$ and $d$.
In the first step, one rotates $M_\nu M_\nu^\dag$ with the tri-bimaximal mixing matrix.
\begin{align}
&U_\mathrm{TBM}^\dag M_\nu^\mathrm{I}(M_\nu^\mathrm{I})^\dag U_\mathrm{TBM}=
\begin{pmatrix}
A&0&B\\
0&C&0\\
B^*&0&D
\end{pmatrix},
\label{eq:MMdagTBM1}\\
&U_\mathrm{TBM}^\dag M_\nu^\mathrm{I\hspace{-.1em}I}(M_\nu^\mathrm{I\hspace{-.1em}I})^\dag U_\mathrm{TBM}=
\begin{pmatrix}
D&0&-B^*\\
0&C&0\\
-B&0&A
\end{pmatrix}.
\label{eq:MMdagTBM2}
\end{align}
where
\begin{align}
&A=\left|a+c-\frac{d}{2}\right|^2+\left|\frac{\sqrt{3}}{2}d\right|^2, \label{eq:A} \\
&B=\left(a+c-\frac{d}{2}\right)\frac{\sqrt{3}}{2}d^*+\frac{\sqrt{3}}{2}d\left(a-c+\frac{d}{2}\right)^* \label{eq:B}
\equiv|B|e^{i\varphi_B}, \\
&C=\left|\frac{a^2-(c^2-cd+d^2)}{2d-c}\right|^2, \label{eq:C} \\
&D=\left|a-c+\frac{d}{2}\right|^2+\left|\frac{\sqrt{3}}{2}d\right|^2. \label{eq:D}
\end{align}
The mass eigenvalues are determined as
\begin{align}
m_1^2&=\frac{A+D}{2}\mp\frac{1}{2}\sqrt{(A-D)^2+4|B|^2},\\
m_2^2&=C,\\
m_3^2&=\frac{A+D}{2}\pm\frac{1}{2}\sqrt{(A-D)^2+4|B|^2},
\end{align}
where the upper and lower signs in these mass eigenvalues correspond to the normal hierarchy (NH) and the inverted hierarchy (IH).
Next, we diagonalize the rotated mass matrices, Eqs.~\eqref{eq:MMdagTBM1} and \eqref{eq:MMdagTBM2},
with $U_{13}(\theta,\sigma)$ and $U_{13}^*(\theta+\frac{\pi}{2},\sigma)$ respectively:
\begin{align}
&U_{13}(\theta,\sigma)^\dag
\begin{pmatrix}
A&0&B\\
0&C&0\\
B^*&0&D
\end{pmatrix}
U_{13}(\theta,\sigma)
=\begin{pmatrix}
m_1^2&&\\
&m_2^2&\\
&&m_3^2
\end{pmatrix}, \\
&U_{13}(\theta+\frac{\pi}{2},\sigma)^\mathrm{T}
\begin{pmatrix}
D&0&-B^*\\
0&C&0\\
-B&0&A
\end{pmatrix}
U_{13}(\theta+\frac{\pi}{2},\sigma)^*
=\begin{pmatrix}
m_1^2&&\\
&m_2^2&\\
&&m_3^2
\end{pmatrix},
\end{align}
where $\theta$ and $\sigma$ are determined as,
\begin{align}
\mathrm{tan}2\theta=\frac{2|B|}{D-A}\ ,\ \sigma=-\varphi_B.
\end{align}
Finally, the other parameters $\phi_i$ are determined as follows,
\begin{align}
&\phi_1=\nonumber\\
&\frac{1}{2}\left[\mathrm{tan}^{-1}\left[\frac
{\left(\mathrm{Im}\left[a\right]+\mathrm{Im}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
+\left(\mathrm{Re}\left[a\right]\mathrm{cos}2\theta+\mathrm{Re}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
-\frac{\sqrt{3}}{2}\mathrm{Im}\left[d\right]\mathrm{sin}2\theta}
{\left(\mathrm{Re}\left[a\right]+\mathrm{Re}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
-\left(\mathrm{Im}\left[a\right]\mathrm{cos}2\theta+\mathrm{Im}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
-\frac{\sqrt{3}}{2}\mathrm{Re}\left[d\right]\mathrm{sin}2\theta}\right]-\sigma\right],\\
&\phi_2=\frac{1}{2}\mathrm{tan}^{-1}\left[\frac
{\mathrm{Im}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Re}\left[2d-c\right]
-\mathrm{Re}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Im}\left[2d-c\right]}
{\mathrm{Re}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Re}\left[2d-c\right]
+\mathrm{Im}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Im}\left[2d-c\right]}\right],\\
&\phi_3=\nonumber\\&\frac{1}{2}\left[\mathrm{tan}^{-1}\left[\frac
{\left(\mathrm{Im}\left[a\right]-\mathrm{Im}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
-\left(\mathrm{Re}\left[a\right]\mathrm{cos}2\theta-\mathrm{Re}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
+\frac{\sqrt{3}}{2}\mathrm{Im}\left[d\right]\mathrm{sin}2\theta}
{\left(\mathrm{Re}\left[a\right]-\mathrm{Re}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
+\left(\mathrm{Im}\left[a\right]\mathrm{cos}2\theta-\mathrm{Im}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
+\frac{\sqrt{3}}{2}\mathrm{Re}\left[d\right]\mathrm{sin}2\theta}\right]+\sigma\right].
\end{align}
We briefly explain the derivation of $\phi_i$ for the mass matrix $M_\nu^{\mathrm{I}}$.
We first diagonalize $M_\nu^{\mathrm{I}}$ with the unitary matrices $U_\mathrm{TBM}$ and $U_{13}(\theta,\sigma)$ according to Eq.~\eqref{eq:diag. cond.}.
However, the diagonalized neutrino mass matrix consists of complex elements.
Then, the parameters $\phi_i$ are determined so that all the elements of the diagonalized matrix are real and positive.
\section{Phenomenological Aspects}\label{sec:num}
We study the phenomenological aspects of this model and show the differences between the two types of vacua.
The observables, such as mixing angles and CP violating phases, are described with the model parameters in different forms for the two types.
In the following subsections, we discuss the relation between the observables and model parameters.
The numerical analyses are also shown in this section.
\subsection{Mixing Angles and CP Violating Phases}
In this subsection, we discuss the lepton mixing angles, CP violating phases and the effective mass for $0\nu\beta\beta$ decay.
At first, we introduce the PDG parametrization of the PMNS matrix:
\begin{align}
U_{\mathrm{PMNS}}^{\mathrm{PDG}}=
\begin{pmatrix}
c_{12}c_{13} & s_{12}c_{13} & s_{13}e^{-i\delta_{CP}} \\
-s_{12}c_{23}-c_{12}s_{23}s_{13}e^{i\delta_{CP}} & c_{12}c_{23}-s_{12}s_{23}s_{13}e^{i\delta_{CP}} & s_{23}c_{13} \\
s_{12}c_{23}-c_{12}c_{23}s_{13}e^{i\delta_{CP}} & -c_{12}s_{23}-s_{12}c_{23}s_{13}e^{i\delta_{CP}} & c_{23}c_{13}
\end{pmatrix}
\begin{pmatrix}
e^{i\alpha}&&\\
&e^{i\beta}&\\
&&1
\end{pmatrix},
\label{eq:PDG}
\end{align}
where $s_{ij}$ and $c_{ij}$ denote the lepton mixing angles $\sin\theta_{ij}$ and $\cos\theta_{ij}$, respectively.
They are written in terms of the PMNS matrix elements:
\begin{align}
\sin^2\theta_{12}=\frac{|U_{e2}|^2}{1-|U_{e3}|^2}\ ,\quad
\sin^2\theta_{23}=\frac{|U_{\mu3}|^2}{1-|U_{e3}|^2}\ ,\quad
\sin^2\theta_{13}=|U_{e3}|^2,
\label{eq:moduli}
\end{align}
where $U_{\alpha i}$ denote the PMNS matrix elements.
The Dirac CP violating phase $\delta_{CP}$ can be obtained with the Jarlskog invariant
\begin{align}
\sin\delta_{CP}&=\frac{J_{CP}}{s_{23}c_{23}s_{12}c_{12}s_{13}c_{13}^2},\label{eq:cp}\\
J_{CP}&=\mathrm{Im}\left[U_{e1}U_{\mu2}U_{\mu1}^*U_{e2}^*\right].
\label{eq:jcp}
\end{align}
In order to obtain these parameters from our model, we substitute the PMNS matrix elements in Eqs.~\eqref{eq:PMNSI} and \eqref{eq:PMNSII}.
For the type I case, the matrix elements are given as follows:
\begin{align}
U_{e1}&=\frac{2}{\sqrt{6}}e^{i\phi_1}\cos\theta,\\
U_{e2}&=U_{\mu2}=\frac{1}{\sqrt{3}} e^{i\phi_2},\\
U_{e3}&=\frac{2}{\sqrt{6}}e^{-i(\sigma-\phi_3)}\sin\theta,\\
U_{\mu1}&=\left(-\frac{1}{\sqrt{6}}\cos\theta+\frac{1}{\sqrt{2}}e^{i\sigma}\sin\theta\right)e^{i\phi_1},\\
U_{\mu3}&=\left(-\frac{1}{\sqrt{6}}e^{-i\sigma}\sin\theta-\frac{1}{\sqrt{2}}\cos\theta\right)e^{i\phi_3}.
\end{align}
The mixing angles, Dirac CP violating phase and Majorana phases for both types are listed in Table~\ref{tb:angles}.
\begin{table}[h]
\caption{Mixing angles, Dirac CP phase and Majorana phases for the two types of vacua}
\centering
\begin{tabular}{|c||c|c|} \hline
&$\mathrm{Type~I}$&$\mathrm{Type~I\hspace{-.1em}I}$\\ \hline \hline
$\mathrm{sin}^2\theta_{12}$&
${\displaystyle \frac{1}{2+\mathrm{cos}2\theta}}$&
${\displaystyle \frac{1}{2-\mathrm{cos}2\theta}}$\\[0.3cm]
$\mathrm{sin}^2\theta_{23}$&
${\displaystyle \frac{1}{2}(1+\frac{\sqrt{3}\mathrm{sin}2\theta}{2+\mathrm{cos}2\theta}\mathrm{cos}\sigma)}$&
${\displaystyle \frac{1}{2}(1-\frac{\sqrt{3}\mathrm{sin}2\theta}{2-\mathrm{cos}2\theta}\mathrm{cos}\sigma)}$\\[0.3cm]
$\mathrm{sin}^2\theta_{13}$&
${\displaystyle \frac{1}{3}(1-\mathrm{cos}2\theta)}$&
${\displaystyle \frac{1}{3}(1+\mathrm{cos}2\theta)}$\\[0.3cm]
$\mathrm{sin}\delta_\mathrm{CP}$&
${\displaystyle -\frac{\sin2\theta}{|\sin2\theta|}\frac{(2+\mathrm{cos}2\theta)\mathrm{sin}\sigma}
{\sqrt{(2+\mathrm{cos}2\theta)^2-3\mathrm{sin}^22\theta\mathrm{cos}^2\sigma}}}$&
${\displaystyle -\frac{\sin2\theta}{|\sin2\theta|}\frac{(2-\mathrm{cos}2\theta)\mathrm{sin}\sigma}
{\sqrt{(2-\mathrm{cos}2\theta)^2-3\mathrm{sin}^22\theta\mathrm{cos}^2\sigma}}}$ \\[0.3cm]
$\alpha+\delta_{\mathrm{CP}}$ & $\phi_1-\phi_3+\sigma$ & $\phi_1-\phi_3+\sigma$ \\
$\beta+\delta_{\mathrm{CP}}$ & $\phi_2-\phi_3+\sigma$ & ${\displaystyle \phi_2-\phi_3+\frac{\pi}{2}}$ \\ \hline
\end{tabular}
\label{tb:angles}
\end{table}
One can adopt either of the two types to predict the mixing angles and the Dirac CP violating phases,
since both types give the same predictions.
However, we note the following two facts.
First, if one fixes $\cos2\theta\simeq1$ to obtain small $\sin^2\theta_{13}$ in type I,
$\sin^2\theta_{13}$ in type I\hspace{-.1em}I reaches $2/3$, which is disfavored in the experiments.
Second, as shown in Subsection \ref{sec:PMNS}, the model parameters $\theta$, $\sigma$ and $\phi_i$ are expressed in the same forms for the two types with $a$, $b$, $c$ and $d$ of Eq.~\eqref{eq:Mv1}.
Therefore, those parameters have common values for both types.
Hence, the two types can not realize the experimental results simultaneously.
Next, we discuss the effective mass for $0\nu\beta\beta$ decay, $m_{ee}=\sum_im_iU_{ei}^2$,
and the Majorana phases, $\alpha$ and $\beta$.
The effective mass is given as follows:
\begin{align}
\left|m_{ee}^\mathrm{I}\right|&=
\frac{1}{3}\left|m_1(1+\mathrm{cos}2\theta)e^{2i\phi_1}
+m_2e^{2i\phi_2}+m_3(1-\mathrm{cos}2\theta)e^{2i(\phi_3-\sigma)}\right|,
\label{eq:Mee1}\\
\left|m_{ee}^\mathrm{I\hspace{-.1em}I}\right|&=
\frac{1}{3}\left|m_1(1-\mathrm{cos}2\theta)e^{2i(\phi_1+\sigma)}
-m_2e^{2i\phi_2}+m_3(1+\mathrm{cos}2\theta)e^{2i\phi_3}\right|,
\label{eq:Mee2}
\end{align}
where the superscripts I and I\hspace{-.1em}I denote the types of vacuum.
Equivalently, one can rewrite Eqs.~\eqref{eq:Mee1} and \eqref{eq:Mee2} as
\begin{align}
\left|m_{ee}^\mathrm{I}\right|&=
\frac{1}{3}\left|m_1(1+\mathrm{cos}2\theta)e^{2i(\phi_1-\phi_3+\sigma)}
+m_2e^{2i(\phi_2-\phi_3+\sigma)}+m_3(1-\mathrm{cos}2\theta)\right|,
\label{eq:Mee3}\\
\left|m_{ee}^\mathrm{I\hspace{-.1em}I}\right|&=
\frac{1}{3}\left|m_1(1-\mathrm{cos}2\theta)e^{2i(\phi_1-\phi_3+\sigma)}
-m_2e^{2i(\phi_2-\phi_3)}+m_3(1+\mathrm{cos}2\theta)\right|,
\label{eq:Mee4}
\end{align}
On the other hand, the effective mass in the PDG parametrization is written as
\begin{align}
\left|m_{ee}\right|=\left|
m_1c_{13}^2c_{12}^2e^{2i(\alpha+\delta_{CP})}+
m_2c_{13}^2s_{12}^2e^{2i(\beta+\delta_{CP})}+
m_3s_{13}^2\right|.
\label{eq:MeePDG}
\end{align}
One can obtain the Majorana CP violating phases $\alpha$ and $\beta$ by comparing Eqs~\eqref{eq:Mee3}-\eqref{eq:MeePDG},
\begin{align}
(\mathrm{Type~I})\quad
&\alpha+\delta_{CP}=\phi_1-\phi_3+\sigma\ ,\quad
\beta+\delta_{CP}=\phi_2-\phi_3+\sigma,
\label{eq:phase1}\\[0.3cm]
(\mathrm{Type~I\hspace{-.1em}I})\quad
&\alpha+\delta_{CP}=\phi_1-\phi_3+\sigma\ ,\quad
\beta+\delta_{CP}=\phi_2-\phi_3+\frac{\pi}{2}.
\label{eq:phase2}
\end{align}
\subsection{Numerical Analysis}
In this subsection, we show numerical analysis to find the difference between two types.
We use recent experimental results with $3\sigma$ range~\cite{Esteban:2016qun}, as summarized in Table~\ref{tb:exp}.
\begin{table}[h]
\caption{
The experimental data for the mass squared differences and mixing angles with $3\sigma$ range~\cite{Esteban:2016qun}}
\centering
\begin{tabular}{|c||c|c|} \hline
& Normal Hierarchy (NH) & Inverted Hierarchy (IH) \\ \hline
$\Delta m_{21}^2~\left[\mathrm{eV}^2\right]$ & $(7.03\sim8.09)\times10^{-5}$ & $(7.03\sim8.09)\times10^{-5}$ \\
$\Delta m_{31}^2~\left[\mathrm{eV}^2\right]$ & $(2.407\sim2.643)\times10^{-3}$ & $-(2.565\sim2.318)\times10^{-3}$ \\
$\mathrm{sin}^2\theta_{12}$ & $0.271\sim0.345$ & $0.271\sim0.345$ \\
$\mathrm{sin}^2\theta_{23}$ & $0.385\sim0.635$ & $0.393\sim0.640$ \\
$\mathrm{sin}^2\theta_{13}$ & $0.01934\sim0.02392$ & $0.01953\sim0.02408$ \\ \hline
\end{tabular}
\label{tb:exp}
\end{table}
As we have shown in the previous subsection, the mixing angles and Dirac CP phase are expressed in terms of the model parameters $\theta$ and $\sigma$ in different forms for the two types.
The experimental data for $\sin^2\theta_{13}$ in Table~\ref{tb:exp} is realized by the following value of $\theta$ with NH or IH:
\begin{align}
\mathrm{Type~I}\ ;\quad
9.81^\circ\leq |\theta| \leq10.9^\circ\ (\mathrm{NH})\ ,\
9.86^\circ\leq |\theta| \leq11.0^\circ\ (\mathrm{IH}),\\
\mathrm{Type~I\hspace{-.1em}I}\ ;\quad
79.1^\circ\leq |\theta| \leq80.2^\circ\ (\mathrm{NH})\ ,\
79.0^\circ\leq |\theta| \leq80.1^\circ\ (\mathrm{IH}).
\end{align}
The value of $\sigma$ is allowed in $-180^\circ\leq\sigma\leq180^\circ$ for both of the two types,
since the error of $\sin^2\theta_{23}$ from the experiments is large.
Next, we discuss the parameters $\phi_i$ in the expressions of the Majorana phases of Eqs.~\eqref{eq:phase1} and \eqref{eq:phase2}.
The effective mass $|m_{ee}|$ in Eq.~\eqref{eq:MeePDG} depends on the two combinations of Dirac and Majorana phases, $2(\alpha+\delta_{CP})$ and $2(\beta+\delta_{CP})$.
If we determine both $|m_{ee}|$ and the lightest neutrino mass, we obtain the constraints on these two combinations.
In order to find how the numerical constraints on $\phi_i$ are different in the two types,
we consider a specific situation.
As an example, we assume that $|m_{ee}|$ is predicted in the region as shown in Fig.~\ref{fig:meeA1}.
We note that the lightest neutrino mass is constrained from the cosmological upper bound for the neutrino mass sum, $\sum_im_i<0.16~\mathrm{eV}$~\cite{Alam:2016hwk}.
This plot is obtained when the Dirac and Majorana phases are randomly chosen from the region A1 in Fig.~\ref{fig:block},
\begin{align}
0 < \alpha + \delta_{CP} < \pi/4\ ,\ 0< \beta + \delta_{CP} < \pi/4.
\end{align}
In this situation, the phase differences $\phi_1 - \phi_3$ and $\phi_2 - \phi_3$ for one type can be distinguished from those for the other type.
The constraints on the phase differences are shown in Fig.~\ref{fig:1323}.
For type I, the phase difference $\phi_2-\phi_3$ is proportional to $\phi_1-\phi_3$.
However, for type I\hspace{-.1em}I, $\phi_2-\phi_3$ is independent of the value of $\phi_1-\phi_3$ because $\sigma$ is absent in the expression of $\phi_2-\phi_3$ in Eq.~\eqref{eq:phase2}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=7.0cm]{16.png}
\caption{16 divided regions for $2(\alpha+\delta_{CP})$ and $2(\beta+\delta_{CP})$.}
\label{fig:block}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\linewidth]{A1.pdf}
\caption{
The prediction of effective mass for $0\nu\beta\beta$ decay in region A1 of Fig.~\ref{fig:block}.
The upper region corresponds to the IH case , while the lower one corresponds to the NH case.}
\label{fig:meeA1}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{tabular}{cc}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=\linewidth]{1213nh1.pdf}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=\linewidth]{1213nh2.pdf}
\end{center}
\end{minipage}
\end{tabular}
\caption{
The allowed regions of the model parameters, $\phi_1-\phi_3$ and $\phi_2-\phi_3$ for both types of vacua.
These plots correspond to region A1 of Fig.~\ref{fig:block}.}
\label{fig:1323}
\end{figure}
\newpage
\section{Summary}\label{sec:con}
We have studied phenomenological aspects of a supersymmetric model with $A_4\times Z_3$ symmetry.
We found 24 degenerate vacua at the 24 minima of the scalar potential.
Then, we discussed the relations among the 24 different vacua and classified them into two types.
Both types consist of 12 vacua which are related to each other by transformations of $A_4$.
We proved that the 12 vacua are equivalent and lead to the same physical consequences.
However, we found that we obtain different physical consequences from the vacua of different types.
Therefore, we analyzed the two types of vacua to find the different phenomenological consequences of the two types.
In particular, we investigated observables such as mixing angles, Dirac CP phase, Majorana phases and effective mass for $0\nu\beta\beta$ decay.
These observables are expressed in terms of the model parameters $\theta$, $\sigma$ and $\phi_i$.
The angle $\theta$ and phase $\sigma$ are determined by the deviation from the tri-bimaximal mixing matrix.
The two types lead to different expressions for the mixing angles and Dirac CP violating phase in terms of $\theta$ and $\sigma$.
Therefore, one should take different model parameters in each type in order to realize the experimental results.
Although one can adopt both of the two types to predict the observable parameters,
the two types cannot realize the current experimental data simultaneously.
The Majorana phases $\alpha$ and $\beta$ are parametrized in the different expressions for each type by the model parameters $\phi_i$ in addition to $\theta$ and $\sigma$.
In order to find numerical differences between the two types of Majorana phase,
we considered the specific situation where the lightest mass and effective mass for the $0\nu\beta\beta$ decay are determined in a certain region.
We showed the allowed regions of the phase differences, $\phi_1-\phi_3$ and $\phi_2-\phi_3$.
The regions are quite different for the two types:
the phase differences for type I are proportional to each other,
while those for type I\hspace{-.1em}I are not.
The VEVs $\eta_m$ and $\lambda^\pm_n$ transfer to the different VEVs by transformations of $A_4$.
However, the transformations for $\eta_m$ and $\lambda^\pm_n$ are closed differently since they have the $Z_3$ and $Z_2$ residual symmetries from $A_4$ respectively.
We have pointed out that some combinations of the VEVs can lead to different physical consequences.
When we consider models with two or more flavons, we should take account of the combination of VEVs.
\vspace{0.5cm}
\noindent
{\bf Acknowledgement}
\noindent
This work is supported by JSPS KAKENHI Grant Number JP17K05418 (T.M.).
This work is also supported in part by Grants-in-Aid for Scientific Research [No.~16J05332 (Y.S.), Nos. 24540272, 26247038, 15H01037, 16H00871, and 16H02189 (H.U.)] from the Ministry of Education,
Culture, Sports, Science and Technology in Japan.
H.O. is also supported by Hiroshima Univ. Alumni Association.
\newpage
\section{Introduction}
Although all the elementary particles in the standard model (SM) have now been discovered, with the discovery of the Higgs boson,
there still exist phenomena which cannot be explained in the framework of the SM.
One of these is the neutrino oscillation phenomenon,
which implies two non-zero neutrino mass squared differences and two large lepton mixing angles.
In order to explain this, many authors propose a neutrino flavor model with non-Abelian discrete flavor symmetry
in the lepton sector (for reviews see~\cite{Ishimori:2010au,Ishimori:2012zz,Ishimori:2013woa,King:2014nza}).
Even before the discovery of the non-zero $\theta _{13}$~\cite{An:2012eh,Ahn:2012nd,Abe:2014lus},
a few authors suggested a tiny mixing angle $\theta _{13}$ based on non-Abelian discrete flavor symmetry~\cite{Shimizu:2011xg}.
Recent results from the T2K and NO$\nu $A experiments~\cite{Abe:2013hdq,Adamson:2016tbq} imply CP violation through the Dirac CP phase.
They studied electron neutrino appearance in a muon neutrino beam.
The Majorana phases are also sources of the CP violating phases if neutrinos are Majorana particles.
The KamLAND-Zen experiment~\cite{KamLAND-Zen:2016pfg} is searching for neutrinoless double beta ($0\nu \beta \beta $) decay to check the Majorana nature of neutrinos.
Therefore, it is important to predict not only mixing angles but also CP phases with the non-Abelian discrete flavor model.
The non-Abelian discrete flavor symmetry can easily explain large lepton mixing angles,
e. g. tri-bimaximal mixing (TBM)~\cite{Harrison:2002er,Harrison:2002kp}, which is a simple framework for the lepton mixing angles.
Indeed, Altarelli and Feruglio (AF) proposed a simple flavor model and predicted TBM by using $A_4$ discrete flavor symmetry~\cite{Altarelli:2005yp,Altarelli:2005yx}.
They introduced $SU(2)$ gauge singlet scalar fields, so-called ``flavons'', and derived the TBM in the lepton sector.
The non-zero $\theta_{13}$ can be realized by another $A_4$ non-trivial singlet flavon~\cite{Shimizu:2011xg} in addition to the flavons introduced by AF.
The origin of non-vanishing $\theta_{13}$ is related to a new contribution to the mass matrices.
Matrices which have the same structure as that in Ref.~\cite{Shimizu:2011xg} also appear in extra-dimensional models with the $S_3$ and $S_4$ flavor symmetries~\cite{Haba:2006dz,Ishimori:2010fs}.
The $\Delta (27)$ model also includes these matrices~\cite{Grimus:2008tt}.
In this paper, we study phenomenological aspects of a supersymmetric model with $A_4\times Z_3$ symmetries.
The three generations of the left-handed leptons are expressed as the $A_4$ triplet, $l=(l_e,l_\mu,l_\tau )$,
while the right-handed charged leptons $e_R$, $\mu_R$, and $\tau_R$ are $A_4$ singlets denoted as ${\bf 1}$, ${\bf 1''}$, and ${\bf 1'}$ respectively.
Three right-handed neutrinos are also described as the triplet of $A_4$.
We introduce the $SU(2)$ gauge singlet flavons of $A_4$ triplets, $\phi_T=(\phi _{T1},\phi _{T2},\phi _{T3})$ and $\phi_S=(\phi _{S1},\phi _{S2},\phi _{S3})$.
In addition, $\xi $ and $\xi '$ are also introduced as the $SU(2)$ gauge singlet flavons with the two kinds of singlet representations of $A_4$, ${\bf 1}$ and ${\bf 1'}$ respectively.
We focus on the vacuum structure of the flavor model.
The scalar sectors of this model consist of many flavons in addition to the SM Higgs boson.
Then, we analyze the scalar potential and show the 24 different sets of VEVs
which come from 24 combinations of 4 (6) possible VEVs of the flavon $\phi_T$ ($\phi_S$).
The 24 different vacua are classified into two types which are not related to each other under the transformations $A_4$.
Therefore, we expect that the two types of vacua have different expressions for the physical observables in terms of the model parameters such as Yukawa couplings.
We ask the following question: whether these different vacua are physically distinct from each other.
The purpose of this paper is to clarify the differences and relations among the VEVs and their physical consequences.
In particular, we investigate the mixing angles, CP violating phase, and effective mass for neutrinoless double beta ($0\nu\beta\beta$) decay.
This paper is organized as follows.
In Section~\ref{sec:model}, we introduce the supersymmetric model with $A_4\times Z_3$ symmetry.
In Section~\ref{sec:2types}, we study the classification of vacua and derive the formulae for the mixing angles and CP phases.
In Section~\ref{sec:num}, we discuss the phenomenological aspects for mixing angles and CP violating phases.
The numerical analyses for the effective mass of $0\nu\beta\beta$ decay are presented.
Section~\ref{sec:con} is devoted to a summary.
In Appendix~\ref{sec:multiplication-rule}, we show the multiplication rule of the $A_4$ group.
\section{Supersymmetric Model with $A_4\times Z_3$ Symmetry}
\label{sec:model}
In this section, we introduce a supersymmetric model with $A_4\times Z_3$ symmetry.
We analyze the scalar potential and derive the mass matrices of the lepton sector.
\subsection{Model}
We introduce three heavy right-handed Majorana neutrinos.
The leptons and scalars in our model are listed in Table~\ref{tb:fields}.
\begin{table}[b]
\caption{\small{
The representations of $SU(2)_L$ and $A_4$, and the charge assignment of $Z_3$ and $U(1)_R$ for leptons and scalars:
~$l_{e,\mu,\tau}$, $\{e,\mu,\tau\}_R$, $\{\nu_e,\nu_\mu,\mu_\tau\}_R$, and $h_{u,d}$ denote left-handed leptons,
right-handed charged leptons, right-handed neutrinos, and Higgs fields, respectively.
The other scalars are gauge singlet flavons and denoted as $\phi_T$, $\phi_S$, $\xi$, and $\xi'$.
$\omega$ is the $Z_3$ charge and stands for $e^{2\pi i/3}$.}}
\hspace{-0.5cm}
\begin{tabular}{|c||c|c|c|c|c||c|c|c|c|c|} \hline
&$l= \begin{pmatrix}
l_e\\
l_\mu\\
l_\tau
\end{pmatrix}$
&$e_R$
&$\mu_R$
&$\tau_R$
&$\nu_R=\begin{pmatrix}
\nu_{eR}\\
\nu_{\mu R}\\
\nu_{\tau R}
\end{pmatrix}$
&$h_{u,d}$
&$\phi_T=\begin{pmatrix}
\phi_{T1}\\
\phi_{T2}\\
\phi_{T3}
\end{pmatrix}$
&$\phi_S=\begin{pmatrix}
\phi_{S1}\\
\phi_{S2}\\
\phi_{S3}
\end{pmatrix}$
&$\xi$
&$\xi'$\\ \hline \hline
$SU(2)_L$&2&1&1&1&1&2&1&1&1&1\\
$A_4$&3&1&$1''$&$1'$&3&1&3&3&1&$1'$\\
$Z_3$&$\omega$&$\omega^2$&$\omega^2$&$\omega^2$&$\omega^2$&1&1&$\omega^2$&$\omega^2$&$\omega^2$\\
$U(1)_R$&1&1&1&1&1&0&0&0&0&0\\ \hline
\end{tabular}\label{tb:fields}
\end{table}
The superpotential of Yukawa interactions is
\begin{align}
w_Y=w_l+w_D+w_R,\label{eq:wY}
\end{align}
where $w_l,w_D$ and $w_R$ are Yukawa interactions for charged lepton, Dirac neutrino and Majorana neutrino sectors respectively:
\begin{align}
&w_l=
y_e(\phi_Tl)_{\bf 1}e_Rh_d/\Lambda+
y_\mu(\phi_Tl)_{\bf 1'}\mu_Rh_d/\Lambda+
y_\tau(\phi_Tl)_{\bf 1''}\tau_Rh_d/\Lambda+h.c., \label{eq:wl} \\
&w_D=y_D(l\nu_R)_{\bf 1}h_u+h.c., \label{eq:wD} \\
&w_R=y_{\phi_S}\phi_S(\nu_R\nu_R)_{\bf 3}+
y_\xi\xi(\nu_R\nu_R)_{\bf 1}+
y_{\xi'}\xi'(\nu_R\nu_R)_{\bf 1''}+h.c., \label{eq:wR}
\end{align}
where the lower indices denote $A_4$ representations.
Moreover, the $y$'s and $\Lambda$ denote the Yukawa coupling constants and cut-off scale respectively.
The multiplication rule for $A_4$ representations is shown in Appendix~\ref{sec:multiplication-rule}.
In order to obtain the mass matrices of these leptons, we analyze the following superpotential of the scalar fields:
\begin{align}
w_d\equiv w_d^T+w_d^S, \label{eq:wd}
\end{align}
where
\begin{align}
&w_d^T=-M(\phi_0^T\phi_T)_{\bf 1}+g\phi_0^T(\phi_T\phi_T)_{\bf 3}, \label{eq:wdT} \\
&w_d^S=
g_1\phi_0^S(\phi_S\phi_S)_{\bf 3}+
g_2(\phi_0^S\phi_S)_{\bf 1}\xi+g_2'(\phi_0^S\phi_S)_{\bf 1''}\xi '+
g_3(\phi_S\phi_S)_{\bf 1}\xi_0-
g_4\xi_0\xi\xi \label{eq:wdS}.
\end{align}
We have introduced the additional $SU(2)$ gauge singlet fields, $\phi_0^T$, $\phi_0^S$ and $\xi_0$, which are called {\it ``driving fields''}.
The charge assignments of these fields are summarized in Table \ref{tb:driving}.
\begin{table}[h!]
\caption{The driving fields and their representations and charge assignment.}
\centering
\begin{tabular}{|c||c|c|c|} \hline
& $\phi_0^T=\begin{pmatrix}
\phi_{01}^T\\
\phi_{02}^T\\
\phi_{03}^T
\end{pmatrix}$
& $\phi_0^S=\begin{pmatrix}
\phi_{01}^S\\
\phi_{02}^S\\
\phi_{03}^S
\end{pmatrix}$
& $\xi_0$ \\ \hline \hline
SU(2) & $1$ & $1$ & $1$ \\
$A_4$ & $3$ & $3$ & $1$ \\
$Z_3$ & $1$ & $\omega^2$ & $\omega^2$ \\
$U(1)_R$ & $2$ & $2$ & $2$ \\ \hline
\end{tabular}\label{tb:driving}
\end{table}
\subsection{Potential Analysis}
In this subsection, we derive the VEVs for the scalar fields $\phi_T,\phi_S,\xi,\xi',\phi^T_0,\phi^S_0,\xi_0$.
One can derive the scalar potential from the superpotentials in Eqs.~\eqref{eq:wdT} and \eqref{eq:wdS} as
\begin{align}
V=V_T+V_S,
\end{align}
where
\begin{align}
V_T=\sum_{X}\left|\frac{\partial w_d^T}{\partial X}\right|^2
=&\left| -M\phi_{T1}+\frac{2}{3}g(\phi_{T1}^2-\phi_{T2}\phi_{T3})\right|^2 \nonumber \\
+&\left| -M\phi_{T3}+\frac{2}{3}g(\phi_{T2}^2-\phi_{T3}\phi_{T1})\right|^2 \nonumber \\
+&\left| -M\phi_{T2}+\frac{2}{3}g(\phi_{T3}^2-\phi_{T1}\phi_{T2})\right|^2 \nonumber \\
+&\left| -M\phi_{01}^T+\frac{2}{3}g(2\phi_{01}^T\phi_{T1}-\phi_{03}^T\phi_{T2}-\phi_{02}^T\phi_{T3})\right|^2 \nonumber \\
+&\left| -M\phi_{03}^T+\frac{2}{3}g(2\phi_{02}^T\phi_{T2}-\phi_{01}^T\phi_{T3}-\phi_{03}^T\phi_{T1})\right|^2 \nonumber \\
+&\left| -M\phi_{02}^T+\frac{2}{3}g(2\phi_{03}^T\phi_{T3}-\phi_{02}^T\phi_{T1}-\phi_{01}^T\phi_{T2})\right|^2 \label{eq:VT},
\end{align}
and
\begin{align}
V_S
=\sum_{Y}\left|\frac{\partial w_d^S}{\partial Y}\right|^2
=&\left|\frac{2}{3}g_1(\phi_{S1}^2-\phi_{S2}\phi_{S3})-g_2\phi_{S1}\xi+g_2'\phi_{S3}\xi'\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(\phi_{S2}^2-\phi_{S3}\phi_{S1})-g_2\phi_{S3}\xi+g_2'\phi_{S2}\xi'\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(\phi_{S3}^2-\phi_{S1}\phi_{S2})-g_2\phi_{S2}\xi+g_2'\phi_{S1}\xi'\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(2\phi_{01}^S\phi_{S1}-\phi_{03}^S\phi_{S2}-\phi_{02}^S\phi_{S3})
-g_2\phi_{01}^S\xi+g_2'\phi_{03}^S\xi'+2g_3\phi_{S1}\xi_0\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(2\phi_{02}^S\phi_{S2}-\phi_{01}^S\phi_{S3}-\phi_{03}^S\phi_{S1})
-g_2\phi_{03}^S\xi+g_2'\phi_{02}^S\xi'+2g_3\phi_{S3}\xi_0\right|^2 \nonumber \\
+&\left|\frac{2}{3}g_1(2\phi_{03}^S\phi_{S3}-\phi_{02}^S\phi_{S1}-\phi_{01}^S\phi_{S2})
-g_2\phi_{02}^S\xi+g_2'\phi_{01}^S\xi'+2g_3\phi_{S2}\xi_0\right|^2 \nonumber \\
+&\left|-g_2(\phi_{01}^S\phi_{S1}+\phi_{03}^S\phi_{S2}+\phi_{02}^S\phi_{S3})-2g_4\xi\xi_0\right|^2 \nonumber \\
+&\left|g_2'(\phi_{02}^S\phi_{S2}+\phi_{01}^S\phi_{S3}+\phi_{03}^S\phi_{S1})\right|^2 \nonumber \\
+&\left|g_3(\phi_{s1}^2+2\phi_{S2}\phi_{S3})-g_4\xi^2\right|^2 \label{eq:VS}.
\end{align}
The sum for $X,Y$ runs over all the scalar fields:
$$
X=\{\phi_{T1}, \phi_{T2}, \phi_{T3}, \phi_{01}^T, \phi_{02}^T, \phi_{03}^T\}\ , \quad
Y=\{\phi_{S1}, \phi_{S2}, \phi_{S3}, \phi_{01}^S, \phi_{02}^S, \phi_{03}^S, \xi, \xi ', \xi_0\}.
$$
The scalar potential $V$ is minimized at $V=V_T=V_S=0$.
There are several solutions for the minimization condition.
We obtain sets of solutions denoted as $\eta_m$ and $\lambda_n^\pm$ ($m=1$-$4$, $n=1$-$3$),
where $\eta_m$ and $\lambda_n^\pm$ are the solutions of $V_T=0$ and $V_S=0$ respectively.
Hereafter, we call them the set of VEV alignments and show them explicitly as follows:
\begin{align}
&\eta_1\equiv\left\{\langle\phi_T\rangle=v_T\begin{pmatrix}1\\0\\0\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta1}\\
&\eta_2\equiv\left\{\langle\phi_T\rangle=\frac{v_T}{3}\begin{pmatrix}-1\\2\\2\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta2}\\
&\eta_3\equiv\left\{\langle\phi_T\rangle=\frac{v_T}{3}\begin{pmatrix}-1\\2\omega\\2\omega^2\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta3}\\
&\eta_4\equiv\left\{\langle\phi_T\rangle=\frac{v_T}{3}\begin{pmatrix}-1\\2\omega^2\\2\omega\end{pmatrix},\
\langle\phi_0^T\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:eta4}\\
&\lambda_1^\pm\equiv\left\{\langle\phi_S\rangle=\pm v_S\begin{pmatrix}1\\1\\1\end{pmatrix},\
\langle\xi'\rangle=u',\ \langle\phi_0^S\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:lam1}\\
&\lambda_2^\pm\equiv\left\{\langle\phi_S\rangle=\pm v_S\begin{pmatrix}1\\\omega\\\omega^2\end{pmatrix},\
\langle\xi'\rangle=\omega u',\ \langle\phi_0^S\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:lam2}\\
&\lambda_3^\pm\equiv\left\{\langle\phi_S\rangle=\pm v_S\begin{pmatrix}1\\\omega^2\\\omega\end{pmatrix},\
\langle\xi'\rangle=\omega^2u',\ \langle\phi_0^S\rangle=\begin{pmatrix}0\\0\\0\end{pmatrix}\right\},\label{eq:lam3}
\end{align}
where $v_T=\frac{3M}{2g}$, $v_S=\sqrt{\frac{g_4}{3g_3}}u$, $u'=\frac{g_2}{g_2'}u$ and $u$ is the VEV of $\xi$, $\langle\xi\rangle=u$
\footnote{
There are still other solutions for $V=0$, including the trivial solution which makes all the VEVs vanish.
It leads to the vanishing of all the lepton masses and mixing angles.
In addition to the trivial solution, there are solutions with non-zero VEVs of the driving fields.
This case leads to the breakdown of $U(1)_R$ symmetry.
In this paper, we only discuss the vacua where $U(1)_R$ symmetry is conserved.}.
The superscript of $\lambda^\pm$ denotes the overall sign of the VEV $\langle\phi_S\rangle$.
In total, we obtain 24 sets of vacua, since there are four sets of alignment for $\eta_m$ and six sets for $\lambda^\pm_n$.
\subsection{Mass Matrix for Charged Leptons and Neutrinos}
We derive charged lepton mass matrices and neutrino mass matrices from the Yukawa interactions in Eqs.~(\ref{eq:wl}),(\ref{eq:wD}), and (\ref{eq:wR}).
These matrices are expressed in various forms corresponding to the VEV alignments.
The charged lepton mass matrices ${M_l}^{(m)}$ for Eqs.~\eqref{eq:eta1}-\eqref{eq:eta4} are
\begin{align}
&{M_l}^{(1)}=\frac{v_dv_T}{\Lambda}
\begin{pmatrix}
y_e&0&0\\
0&y_\mu&0\\
0&0&y_\tau
\end{pmatrix},
\label{eq:Ml1} \\[0.3cm]
&{M_l}^{(2)}=\frac{v_dv_T}{3\Lambda}
\begin{pmatrix}
-y_e&2y_\mu&2y_\tau\\
2y_e&-y_\mu&2y_\tau\\
2y_e&2y_\mu&-y_\tau
\end{pmatrix}
=S{M_l}^{(1)}, \label{eq:Ml2} \\[0.3cm]
&{M_l}^{(3)}=\frac{v_dv_T}{3\Lambda}
\begin{pmatrix}
-y_e&2\omega y_\mu&2\omega^2y_\tau\\
2\omega^2y_e&-y_\mu&2\omega y_\tau\\
2\omega y_e&2\omega^2y_\mu&-y_\tau
\end{pmatrix}
=T^\dag ST{M_l}^{(1)}, \label{eq:Ml3} \\[0.3cm]
&{M_l}^{(4)}=\frac{v_dv_T}{3\Lambda}
\begin{pmatrix}
-y_e&2\omega^2y_\mu&2\omega y_\tau\\
2\omega y_e&-y_\mu&2\omega^2y_\tau\\
2\omega^2y_e&2\omega y_\mu&-y_\tau
\end{pmatrix}
=TST^\dag{M_l}^{(1)}, \label{eq:Ml4}
\end{align}
respectively, where the matrices $S$ and $T$ are
\begin{align}
S=\frac{1}{3}\begin{pmatrix}
-1&2&2\\
2&-1&2\\
2&2&-1
\end{pmatrix}\ ,\quad
T=\begin{pmatrix}
1&0&0\\
0&\omega&0\\
0&0&\omega^2
\end{pmatrix}.
\label{eq:ST}
\end{align}
The Dirac mass matrix for neutrinos obtained from Eq.~\eqref{eq:wD} is
\begin{align}
M_D=y_Dv_u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}.
\label{eq:MD}
\end{align}
It is noted that the Dirac mass matrix is determined independently of the VEV alignments.
The Majorana mass matrices ${M_R^{(n)}}^\pm$ for the corresponding set of solutions Eqs.~\eqref{eq:lam1},\eqref{eq:lam2},\eqref{eq:lam3} are given as follows:
\begin{align}
{M_R^{(1)}}^\pm&=
\pm\frac{1}{3}y_{{\phi_S}}v_S\begin{pmatrix}
2&-1&-1\\
-1&2&-1\\
-1&-1&2
\end{pmatrix}+y_\xi u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+y_{\xi'}u'\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}, \label{eq:MR1} \\
{M_R^{(2)}}^\pm&=
\pm\frac{1}{3}y_{{\phi_S}}v_S\begin{pmatrix}
2&-\omega^2&-\omega\\
-\omega^2&2\omega&-1\\
-\omega&-1&2\omega^2
\end{pmatrix}+y_\xi u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+\omega y_{\xi'}u'\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}\nonumber\\
&=T^\dag{M_R^{(1)}}^\pm T^\dag, \label{eq:MR2} \\
{M_R^{(3)}}^\pm&=
\pm\frac{1}{3}y_{{\phi_S}}v_S\begin{pmatrix}
2&-\omega&-\omega^2\\
-\omega&2\omega^2&-1\\
-\omega^2&-1&2\omega
\end{pmatrix}+y_\xi u\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+\omega^2y_{\xi'}u'\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}\nonumber\\
&=T{M_R^{(1)}}^\pm T. \label{eq:MR3}
\end{align}
In order to generate the light neutrino mass matrices, we adopt the seesaw mechanism~\cite{Minkowski,Yanagida:1979as,GellMann:1980vs}.
The effective neutrino mass matrices are given by the well-known formula, $M_\nu=-M_DM_R^{-1}M_D^\mathrm{T}$,
through the seesaw mechanism.
We obtain the 6 different effective neutrino mass matrices from Eqs.~\eqref{eq:MD}-\eqref{eq:MR3} as follows:
\begin{align}
&{M_\nu^{(1)}}^\pm=\pm a\begin{pmatrix}
1&0&0\\
0&1&0\\
0&0&1
\end{pmatrix}+b^\pm\begin{pmatrix}
1&1&1\\
1&1&1\\
1&1&1
\end{pmatrix}+c\begin{pmatrix}
1&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}+d\begin{pmatrix}
0&0&1\\
0&1&0\\
1&0&0
\end{pmatrix}, \label{eq:Mv1}\\
&{M_\nu^{(2)}}^\pm=T^\dag{M_\nu^{(1)}}^\pm T^\dag, \label{eq:Mv2}\\
&{M_\nu^{(3)}}^\pm=T{M_\nu^{(1)}}^\pm T, \label{eq:Mv3}
\end{align}
where
\begin{align}
&a=ky_{{\phi_S}}v_S, \nonumber\\
&c=k(y_{\xi'}u'-y_\xi u), \nonumber\\
&d=ky_{\xi'}u', \nonumber\\
&b^\pm=\mp\frac a3+\frac{a^2}{2d-c}\left(\frac13-\frac{d^2}{a^2}\right), \nonumber\\
&k=\frac{{y_D}^2{v_u}^2}{y_\xi^2u^2+y_{\xi'}^2{u'}^2-(y_{\phi_S}^2v_S^2+y_\xi uy_\xi'u')}. \nonumber
\end{align}
\section{Classification of Vacua and PMNS Mixing Matrix}\label{sec:2types}
In this section, we classify the 24 different vacua and derive the lepton mixing matrix $U_\mathrm{PMNS}$, called the {\it Pontecorvo-Maki-Nakagawa-Sakata} (PMNS) mixing matrix.
In order to classify the vacua, we discuss the relations among the VEV alignments with the transformations of $A_4$.
We show that the 24 vacua are classified into two types in the following subsection.
Then, one finds the two different PMNS matrices with diagonalizing matrices for the charged lepton and effective neutrino mass matrices Eqs.~(\ref{eq:Ml1})-(\ref{eq:Ml4}), and (\ref{eq:Mv1}),(\ref{eq:Mv2}),(\ref{eq:Mv3}).
\subsection{Relations among Sets of VEV Alignments}
The generators of $A_4$ are expressed as the following forms for the representations ${\bf 1},{\bf 1'},{\bf 1''}$ and ${\bf 3}$,
\begin {align}
&S({\bf 1})=S({\bf 1'})=S({\bf 1''})=1\ ,\quad S({\bf 3})
=\frac{1}{3}\begin{pmatrix}
-1&2&2\\
2&-1&2\\
2&2&-1
\end{pmatrix},
\label{eq:Strans}\\
&T({\bf 1})=1,\ T({\bf 1'})=\omega,\ T({\bf 1''})=\omega^2\ ,\quad T({\bf 3})=
\begin{pmatrix}
1&0&0\\
0&\omega&0\\
0&0&\omega^2
\label{eq:Ttrans}
\end{pmatrix}.
\end{align}
The sets of VEV alignment $\eta_m,\lambda^\pm_n$ are associated through the transformations of these generators.
As an example, we show the $T$ transformation on $\lambda_1^+$:
\begin{align}
T\left[\lambda_1^+\right]&\equiv
\left\{\langle\phi_S\rangle=T({\bf 3})v_S
\begin{pmatrix}
1\\
1\\
1
\end{pmatrix},
\langle\xi'\rangle=T({\bf 1'})u',
\langle\phi_0^S\rangle=T({\bf 3})
\begin{pmatrix}
0\\
0\\
0
\end{pmatrix}\right\}\nonumber \\
&=\left\{\langle\phi_S\rangle=v_S
\begin{pmatrix}
1\\
\omega\\
\omega^2
\end{pmatrix},
\langle\xi'\rangle=\omega u',\
\langle\phi_0^S\rangle=
\begin{pmatrix}
0\\
0\\
0
\end{pmatrix}\right\}=\lambda_2^+.
\end{align}
The $S$ and $T$ transformations on all the sets of the VEV alignment are summarized in Fig.~\ref{fig:map}.
Some transformations preserve the VEVs of either $\eta_m$ or $\lambda^\pm_n$.
These vacua have $Z_3$ or $Z_2$ symmetries as the residual symmetries of $A_4$ respectively.
For the VEVs described as $\eta_m$, they are invariant under the following transformation,
\begin{align}
T\left[\eta_1\right]=T^{-1}\left[\eta_1\right]=\eta_1\ ,\quad
TST\left[\eta_2\right]=(TST)^{-1}\left[\eta_2\right]=\eta_2 \ ,\nonumber\\
ST\left[\eta_3\right]=(ST)^{-1}\left[\eta_3\right]=\eta_3\ ,\quad
TS\left[\eta_4\right]=(TS)^{-1}\left[\eta_4\right]=\eta_4\ .
\end{align}
It is easy to confirm that such transformations correspond to $Z_3$ symmetries:
\begin{align}
T^3=(TST)^3=(ST)^3=(TS)^3=1.
\end{align}
Each $\lambda^\pm_n$ has $Z_2$ symmetry as follows:
\begin{align}
S\left[\lambda^\pm_1\right]=\lambda^\pm_1\ ,\quad TST^2\left[\lambda^\pm_2\right]
=\lambda^\pm_2\ ,\quad T^2ST\left[\lambda^\pm_3\right]=\lambda^\pm_3,
\end{align}
where
\begin{align}
S^2=(TST^2)^2=(T^2ST)^2=1.
\end{align}
\begin{figure}[h]
\begin{center}
\includegraphics[width=7.0cm]{vevtransition.pdf}
\caption{
Map of the transitions among the VEV alignments under the transformations $S$ and $T$:
The solid arrow corresponds to the transition due to $T$ transformation
and the dashed two headed arrow shows the transition due to $S$ transformation.
In the map, $ \eta_1$ is invariant under $T$ transformation
while $\lambda_1^{\pm}$ are invariant under $S$ transformation. }
\label{fig:map}
\end{center}
\end{figure}
\subsection{Classification of 24 Vacua}
In this subsection, we show the relations among the 24 different Lagrangians derived from the 24 different combinations of VEV alignments in Eqs.~\eqref{eq:eta1}-\eqref{eq:lam3}.
We find the two sets of 12 equivalent Lagrangians with the appropriate field redefinitions.
Then, the 24 Lagrangians are classified into two types.
For simplicity, we write the Lagrangian of this model in a short form:
\begin{align}
\mathcal{L}(\psi,\phi_1,\phi_2)\label{eq:L},
\end{align}
where $\psi$ represents the fermion fields such as $l$ and $\nu_R$.
$\phi_1$ and $\phi_2$ represent the scalar fields,
which should have their VEVs written as $\eta_m$ and $\lambda_n^\pm$ respectively.
We write the Lagrangian in the broken phase for the VEV alignment ($\eta_m$, $\lambda^\pm_n$) with fluctuations $h_1$ and $h_2$ as
\begin{align}
\mathcal{L}_{mn}^\pm(\psi,h_1,h_2)\equiv\mathcal{L}(\psi,\eta_m+h_1,\lambda_n^\pm+h_2).
\end{align}
Then, we prove the following equation:
\begin{align}
\mathcal{L}(\psi',\eta_m+h_1',\lambda^\pm_n+h_2')=
\mathcal{L}(\psi,G^{-1}\eta_m+h_1,G^{-1}\lambda_n^\pm+h_2),\label{eq:eqre}
\end{align}
where $G$ denotes the transformation composed of $S$ and $T$ in Eqs.~\eqref{eq:Strans} and \eqref{eq:Ttrans}.
There are 12 independent transformations including the identity element:
\begin{align}
G;\{e,\ T,\ T^2,\ S,\ TS,\ T^2S,\ ST,\ ST^2,\ T^2ST,\ TST,\ TST^2,\ T^2ST^2\}.
\label{eq:allG}
\end{align}
The redefined fields are written as follows,
\begin{align}
\psi'=G\psi\ ,\quad h_i'=Gh_i\quad(i=1,2).\label{eq:redef}
\end{align}
The right-hand side of Eq.~\eqref{eq:eqre} corresponds to the Lagrangian for the vacuum ($G^{-1}\eta_m,G^{-1}\lambda^\pm_n$) while the left-hand side is the Lagrangian for the vacuum ($\eta_m,\lambda_n^\pm$) in terms of the redefined fields.
In the symmetric phase, the Lagrangian $\mathcal{L}(\psi,\phi_1,\phi_2)$ is invariant under the $G$ transformation,
\begin{align}
\mathcal{L}(G\psi,G\phi_1,G\phi_2)=\mathcal{L}(\psi,\phi_1,\phi_2).\label{eq:INV}
\end{align}
One obtains the following equation from Eq.~\eqref{eq:INV} for the vacuum ($G^{-1}\eta_m,G^{-1}\lambda^\pm_n$),
\begin{align}
\mathcal{L}(G\psi,\eta_m+Gh_1,\lambda_n^\pm+Gh_2)
=&\mathcal{L}(\psi,G^{-1}\eta_m+h_1,G^{-1}\lambda_n^\pm+h_2).
\label{eq:aeqre}
\end{align}
Finally, one obtains the relation Eq.~\eqref{eq:eqre} by applying the field definition Eq.~\eqref{eq:redef} to the left-hand side of Eq.~\eqref{eq:aeqre}.
The relation Eq.~\eqref{eq:eqre} implies the equality of the Lagrangians for the two vacua ($\eta_m,\lambda^\pm_n$) and ($G^{-1}\eta_m,G^{-1}\lambda^\pm_n$).
Here, we briefly show how to find the equivalent vacua with Fig.~\ref{fig:map}.
For example, let us consider the $T$ transformation in terms of the vacuum of $(\eta_1,\lambda^+_1)$.
One finds that $\eta_1$ is invariant and $\lambda^+_1$ transfers to $\lambda^+_2$ under the $T$ transformation.
Therefore, $\mathcal{L}^+_{11}$ and $\mathcal{L}^+_{12}$ are equivalent.
One can find 12 equivalent vacua by applying 12 independent transformations in Eq.~\eqref{eq:allG} to the vacuum $(\eta_1,\lambda^+_1)$.
Then, we classify the 24 Lagrangians into two types:
\begin{align}
&\mathrm{Type\ I};
\{\mathcal{L}^+_{11},\mathcal{L}^+_{12},\mathcal{L}^+_{13},\mathcal{L}^+_{21},\mathcal{L}^+_{32},\mathcal{L}^+_{43},
\mathcal{L}^-_{22},\mathcal{L}^-_{23},\mathcal{L}_{31}^-,\mathcal{L}^-_{33},\mathcal{L}^-_{41},\mathcal{L}^-_{42}\},\\
&\mathrm{Type\ I\hspace{-.1em}I};
\{\mathcal{L}^-_{11},\mathcal{L}^-_{12},\mathcal{L}^-_{13},\mathcal{L}^-_{21},\mathcal{L}^-_{32},\mathcal{L}^-_{43},
\mathcal{L}^+_{22},\mathcal{L}^+_{23},\mathcal{L}^+_{31},\mathcal{L}^+_{33},\mathcal{L}^+_{41},\mathcal{L}^+_{42}\}.
\end{align}
Type I and type I\hspace{-.1em}I are disconnected because of the absence of a transformation which relates one type to the other.
Since all the Lagrangians which belong to the same type lead to the same physical consequences,
we consider only $\mathcal{L}^+_{11}$ and $\mathcal{L}^-_{11}$ as the representatives of their types:
\begin{align}
\mathcal{L}^\mathrm{I}\equiv\mathcal{L}^+_{11}\ ,\ \ \mathcal{L}^{\mathrm{I\hspace{-.1em}I}}\equiv\mathcal{L}^-_{11}.
\end{align}
We also define the representative mass matrices for charged leptons and neutrinos as
\begin{align}
M_l\equiv M_l^{(1)}\ ,\quad
M_\nu^\mathrm{I}\equiv {M_\nu^{(1)}}^+\ ,\quad
M_\nu^\mathrm{I\hspace{-.1em}I}\equiv {M_\nu^{(1)}}^-.
\end{align}
It is noted that the charged lepton mass matrix $M_l^{(1)}$ is diagonal.
\subsection{PMNS Matrices for Two Types}\label{sec:PMNS}
In this subsection, we construct the PMNS matrices for the two types,
$\mathcal{L}^\mathrm{I}$ and $\mathcal{L}^{\mathrm{I\hspace{-.1em}I}}$.
Since the charged lepton mass matrix $M_l$ is diagonal,
the PMNS matrix is determined so that it diagonalizes the neutrino mass matrices in Eq.~(\ref{eq:Mv1}):
\begin{align}
(U_\mathrm{PMNS}^\mathrm{I})^\dag M_\nu^\mathrm{I}(U_\mathrm{PMNS}^\mathrm{I})^*=
(U_\mathrm{PMNS}^\mathrm{I\hspace{-.1em}I})^\dag M_\nu^\mathrm{I\hspace{-.1em}I}
(U_\mathrm{PMNS}^\mathrm{I\hspace{-.1em}I})^*=
\begin{pmatrix}
m_1&&\\
&m_2&\\
&&m_3
\end{pmatrix},
\label{eq:diag. cond.}
\end{align}
where the left-handed neutrino masses $m_1,m_2$ and $m_3$ are positive.
The PMNS matrices are expressed as the following forms for the two types:
\begin{align}
U_\mathrm{PMNS}^\mathrm{I}&=U_\mathrm{TBM}U_{13}(\theta,\sigma)
\begin{pmatrix}
e^{i\phi_1}&& \\
&e^{i\phi_2}& \\
&&e^{i\phi_3}
\end{pmatrix},
\label{eq:PMNSI}\\
U_\mathrm{PMNS}^\mathrm{I\hspace{-.1em}I}&=U_\mathrm{TBM}\begin{pmatrix}
&&-i\\
&1&\\
i&&
\end{pmatrix}U_{13}(\theta,\sigma)\begin{pmatrix}
e^{i\phi_1}&& \\
&e^{i\phi_2}& \\
&&e^{i\phi_3}
\end{pmatrix}, \nonumber\\
&=U_\mathrm{TBM}U_{13}^*(\theta+\frac{\pi}{2},\sigma)\begin{pmatrix}
-ie^{i(\phi_1+\sigma)}&& \\
&e^{i\phi_2}& \\
&&-ie^{i(\phi_3-\sigma)}
\end{pmatrix}.
\label{eq:PMNSII}
\end{align}
The unitary matrix $U_\mathrm{TBM}$ is the tri-bimaximal mixing matrix and $U_{13}(\theta,\sigma)$ denotes the unitary rotation matrix:
\begin{align}
&U_\mathrm{TBM}=
\begin{pmatrix}
2/\sqrt{6}&1/\sqrt{3}&0\\
-1/\sqrt{6}&1/\sqrt{3}&-1/\sqrt{2}\\
-1/\sqrt{6}&1/\sqrt{3}&1/\sqrt{2}
\end{pmatrix},\label{eq:TBM}\\
&U_{13}(\theta,\sigma)=
\begin{pmatrix}
\mathrm{cos}\theta&0&e^{-i\sigma}\mathrm{sin}\theta\\
0&1&0\\
-e^{i\sigma}\mathrm{sin}\theta&0&\mathrm{cos}\theta
\end{pmatrix}.\label{eq:R13}
\end{align}
We have introduced the parameters $\theta,\sigma$ and $\phi_i$ ($i=1,2,3$).
They are written in terms of the complex parameters of the neutrino mass matrix, $a,b,c$ and $d$, in Eq.~\eqref{eq:Mv1}~\footnote{
There are six real parameters since $b$ is written by using $a,c,d$.}.
In the rest of this subsection, we derive the explicit forms of the parameters $\theta,\sigma$ and $\phi_i$ in terms of the model parameters $a,b,c$ and $d$.
In the first step, one rotates $M_\nu M_\nu^\dag$ with the tri-bimaximal mixing matrix.
\begin{align}
&U_\mathrm{TBM}^\dag M_\nu^\mathrm{I}(M_\nu^\mathrm{I})^\dag U_\mathrm{TBM}=
\begin{pmatrix}
A&0&B\\
0&C&0\\
B^*&0&D
\end{pmatrix},
\label{eq:MMdagTBM1}\\
&U_\mathrm{TBM}^\dag M_\nu^\mathrm{I\hspace{-.1em}I}(M_\nu^\mathrm{I\hspace{-.1em}I})^\dag U_\mathrm{TBM}=
\begin{pmatrix}
D&0&-B^*\\
0&C&0\\
-B&0&A
\end{pmatrix}.
\label{eq:MMdagTBM2}
\end{align}
where
\begin{align}
&A=\left|a+c-\frac{d}{2}\right|^2+\left|\frac{\sqrt{3}}{2}d\right|^2, \label{eq:A} \\
&B=\left(a+c-\frac{d}{2}\right)\frac{\sqrt{3}}{2}d^*+\frac{\sqrt{3}}{2}d\left(a-c+\frac{d}{2}\right)^* \label{eq:B}
\equiv|B|e^{i\varphi_B}, \\
&C=\left|\frac{a^2-(c^2-cd+d^2)}{2d-c}\right|^2, \label{eq:C} \\
&D=\left|a-c+\frac{d}{2}\right|^2+\left|\frac{\sqrt{3}}{2}d\right|^2. \label{eq:D}
\end{align}
The mass eigenvalues are determined as
\begin{align}
m_1^2&=\frac{A+D}{2}\mp\frac{1}{2}\sqrt{(A-D)^2+4|B|^2},\\
m_2^2&=C,\\
m_3^2&=\frac{A+D}{2}\pm\frac{1}{2}\sqrt{(A-D)^2+4|B|^2},
\end{align}
where the upper and lower signs in these mass eigenvalues correspond to the normal hierarchy (NH) and the inverted hierarchy (IH).
Next, we diagonalize the rotated mass matrices, Eqs.~\eqref{eq:MMdagTBM1} and \eqref{eq:MMdagTBM2},
with $U_{13}(\theta,\sigma)$ and $U_{13}^*(\theta+\frac{\pi}{2},\sigma)$ respectively:
\begin{align}
&U_{13}(\theta,\sigma)^\dag
\begin{pmatrix}
A&0&B\\
0&C&0\\
B^*&0&D
\end{pmatrix}
U_{13}(\theta,\sigma)
=\begin{pmatrix}
m_1^2&&\\
&m_2^2&\\
&&m_3^2
\end{pmatrix}, \\
&U_{13}(\theta+\frac{\pi}{2},\sigma)^\mathrm{T}
\begin{pmatrix}
D&0&-B^*\\
0&C&0\\
-B&0&A
\end{pmatrix}
U_{13}(\theta+\frac{\pi}{2},\sigma)^*
=\begin{pmatrix}
m_1^2&&\\
&m_2^2&\\
&&m_3^2
\end{pmatrix},
\end{align}
where $\theta$ and $\sigma$ are determined as,
\begin{align}
\mathrm{tan}2\theta=\frac{2|B|}{D-A}\ ,\ \sigma=-\varphi_B.
\end{align}
Finally, the other parameters $\phi_i$ are determined as follows,
\begin{align}
&\phi_1=\nonumber\\
&\frac{1}{2}\left[\mathrm{tan}^{-1}\left[\frac
{\left(\mathrm{Im}\left[a\right]+\mathrm{Im}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
+\left(\mathrm{Re}\left[a\right]\mathrm{cos}2\theta+\mathrm{Re}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
-\frac{\sqrt{3}}{2}\mathrm{Im}\left[d\right]\mathrm{sin}2\theta}
{\left(\mathrm{Re}\left[a\right]+\mathrm{Re}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
-\left(\mathrm{Im}\left[a\right]\mathrm{cos}2\theta+\mathrm{Im}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
-\frac{\sqrt{3}}{2}\mathrm{Re}\left[d\right]\mathrm{sin}2\theta}\right]-\sigma\right],\\
&\phi_2=\frac{1}{2}\mathrm{tan}^{-1}\left[\frac
{\mathrm{Im}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Re}\left[2d-c\right]
-\mathrm{Re}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Im}\left[2d-c\right]}
{\mathrm{Re}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Re}\left[2d-c\right]
+\mathrm{Im}\left[a^2-(c^2-cd+d^2)\right]\mathrm{Im}\left[2d-c\right]}\right],\\
&\phi_3=\nonumber\\&\frac{1}{2}\left[\mathrm{tan}^{-1}\left[\frac
{\left(\mathrm{Im}\left[a\right]-\mathrm{Im}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
-\left(\mathrm{Re}\left[a\right]\mathrm{cos}2\theta-\mathrm{Re}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
+\frac{\sqrt{3}}{2}\mathrm{Im}\left[d\right]\mathrm{sin}2\theta}
{\left(\mathrm{Re}\left[a\right]-\mathrm{Re}\left[c-\frac{d}{2}\right]\mathrm{cos}2\theta\right)\mathrm{cos}\sigma
+\left(\mathrm{Im}\left[a\right]\mathrm{cos}2\theta-\mathrm{Im}\left[c-\frac{d}{2}\right]\right)\mathrm{sin}\sigma
+\frac{\sqrt{3}}{2}\mathrm{Re}\left[d\right]\mathrm{sin}2\theta}\right]+\sigma\right].
\end{align}
We briefly explain the derivation of $\phi_i$ for the mass matrix $M_\nu^{\mathrm{I}}$.
We first diagonalize $M_\nu^{\mathrm{I}}$ with the unitary matrices $U_\mathrm{TBM}$ and $U_{13}(\theta,\sigma)$ according to Eq.~\eqref{eq:diag. cond.}.
However, the diagonalized neutrino mass matrix consists of complex elements.
Then, the parameters $\phi_i$ are determined so that all the elements of the diagonalized matrix are real and positive.
\section{Phenomenological Aspects}\label{sec:num}
We study the phenomenological aspects of this model and show the differences between the two types of vacua.
The observables, such as mixing angles and CP violating phases, are described with the model parameters in different forms for the two types.
In the following subsections, we discuss the relation between the observables and model parameters.
The numerical analyses are also shown in this section.
\subsection{Mixing Angles and CP Violating Phases}
In this subsection, we discuss the lepton mixing angles, CP violating phases and the effective mass for $0\nu\beta\beta$ decay.
At first, we introduce the PDG parametrization of the PMNS matrix:
\begin{align}
U_{\mathrm{PMNS}}^{\mathrm{PDG}}=
\begin{pmatrix}
c_{12}c_{13} & s_{12}c_{13} & s_{13}e^{-i\delta_{CP}} \\
-s_{12}c_{23}-c_{12}s_{23}s_{13}e^{i\delta_{CP}} & c_{12}c_{23}-s_{12}s_{23}s_{13}e^{i\delta_{CP}} & s_{23}c_{13} \\
s_{12}c_{23}-c_{12}c_{23}s_{13}e^{i\delta_{CP}} & -c_{12}s_{23}-s_{12}c_{23}s_{13}e^{i\delta_{CP}} & c_{23}c_{13}
\end{pmatrix}
\begin{pmatrix}
e^{i\alpha}&&\\
&e^{i\beta}&\\
&&1
\end{pmatrix},
\label{eq:PDG}
\end{align}
where $s_{ij}$ and $c_{ij}$ denote the lepton mixing angles $\sin\theta_{ij}$ and $\cos\theta_{ij}$, respectively.
They are written in terms of the PMNS matrix elements:
\begin{align}
\sin^2\theta_{12}=\frac{|U_{e2}|^2}{1-|U_{e3}|^2}\ ,\quad
\sin^2\theta_{23}=\frac{|U_{\mu3}|^2}{1-|U_{e3}|^2}\ ,\quad
\sin^2\theta_{13}=|U_{e3}|^2,
\label{eq:moduli}
\end{align}
where $U_{\alpha i}$ denote the PMNS matrix elements.
The Dirac CP violating phase $\delta_{CP}$ can be obtained with the Jarlskog invariant
\begin{align}
\sin\delta_{CP}&=\frac{J_{CP}}{s_{23}c_{23}s_{12}c_{12}s_{13}c_{13}^2},\label{eq:cp}\\
J_{CP}&=\mathrm{Im}\left[U_{e1}U_{\mu2}U_{\mu1}^*U_{e2}^*\right].
\label{eq:jcp}
\end{align}
In order to obtain these parameters from our model, we substitute the PMNS matrix elements in Eqs.~\eqref{eq:PMNSI} and \eqref{eq:PMNSII}.
For the type I case, the matrix elements are given as follows:
\begin{align}
U_{e1}&=\frac{2}{\sqrt{6}}e^{i\phi_1}\cos\theta,\\
U_{e2}&=U_{\mu2}=\frac{1}{\sqrt{3}} e^{i\phi_2},\\
U_{e3}&=\frac{2}{\sqrt{6}}e^{-i(\sigma-\phi_3)}\sin\theta,\\
U_{\mu1}&=\left(-\frac{1}{\sqrt{6}}\cos\theta+\frac{1}{\sqrt{2}}e^{i\sigma}\sin\theta\right)e^{i\phi_1},\\
U_{\mu3}&=\left(-\frac{1}{\sqrt{6}}e^{-i\sigma}\sin\theta-\frac{1}{\sqrt{2}}\cos\theta\right)e^{i\phi_3}.
\end{align}
The mixing angles, Dirac CP violating phase and Majorana phases for both types are listed in Table~\ref{tb:angles}.
\begin{table}[h]
\caption{Mixing angles, Dirac CP phase and Majorana phases for the two types of vacua}
\centering
\begin{tabular}{|c||c|c|} \hline
&$\mathrm{Type~I}$&$\mathrm{Type~I\hspace{-.1em}I}$\\ \hline \hline
$\mathrm{sin}^2\theta_{12}$&
${\displaystyle \frac{1}{2+\mathrm{cos}2\theta}}$&
${\displaystyle \frac{1}{2-\mathrm{cos}2\theta}}$\\[0.3cm]
$\mathrm{sin}^2\theta_{23}$&
${\displaystyle \frac{1}{2}(1+\frac{\sqrt{3}\mathrm{sin}2\theta}{2+\mathrm{cos}2\theta}\mathrm{cos}\sigma)}$&
${\displaystyle \frac{1}{2}(1-\frac{\sqrt{3}\mathrm{sin}2\theta}{2-\mathrm{cos}2\theta}\mathrm{cos}\sigma)}$\\[0.3cm]
$\mathrm{sin}^2\theta_{13}$&
${\displaystyle \frac{1}{3}(1-\mathrm{cos}2\theta)}$&
${\displaystyle \frac{1}{3}(1+\mathrm{cos}2\theta)}$\\[0.3cm]
$\mathrm{sin}\delta_\mathrm{CP}$&
${\displaystyle -\frac{\sin2\theta}{|\sin2\theta|}\frac{(2+\mathrm{cos}2\theta)\mathrm{sin}\sigma}
{\sqrt{(2+\mathrm{cos}2\theta)^2-3\mathrm{sin}^22\theta\mathrm{cos}^2\sigma}}}$&
${\displaystyle -\frac{\sin2\theta}{|\sin2\theta|}\frac{(2-\mathrm{cos}2\theta)\mathrm{sin}\sigma}
{\sqrt{(2-\mathrm{cos}2\theta)^2-3\mathrm{sin}^22\theta\mathrm{cos}^2\sigma}}}$ \\[0.3cm]
$\alpha+\delta_{\mathrm{CP}}$ & $\phi_1-\phi_3+\sigma$ & $\phi_1-\phi_3+\sigma$ \\
$\beta+\delta_{\mathrm{CP}}$ & $\phi_2-\phi_3+\sigma$ & ${\displaystyle \phi_2-\phi_3+\frac{\pi}{2}}$ \\ \hline
\end{tabular}
\label{tb:angles}
\end{table}
One can adopt either of the two types to predict the mixing angles and the Dirac CP violating phases,
since both types give the same predictions.
However, we note the following two facts.
First, if one fixes $\cos2\theta\simeq1$ to obtain small $\sin^2\theta_{13}$ in type I,
$\sin^2\theta_{13}$ in type I\hspace{-.1em}I reaches $2/3$, which is disfavored in the experiments.
Second, as shown in Subsection \ref{sec:PMNS}, the model parameters $\theta$, $\sigma$ and $\phi_i$ are expressed in the same forms for the two types with $a$, $b$, $c$ and $d$ of Eq.~\eqref{eq:Mv1}.
Therefore, those parameters have common values for both types.
Hence, the two types can not realize the experimental results simultaneously.
Next, we discuss the effective mass for $0\nu\beta\beta$ decay, $m_{ee}=\sum_im_iU_{ei}^2$,
and the Majorana phases, $\alpha$ and $\beta$.
The effective mass is given as follows:
\begin{align}
\left|m_{ee}^\mathrm{I}\right|&=
\frac{1}{3}\left|m_1(1+\mathrm{cos}2\theta)e^{2i\phi_1}
+m_2e^{2i\phi_2}+m_3(1-\mathrm{cos}2\theta)e^{2i(\phi_3-\sigma)}\right|,
\label{eq:Mee1}\\
\left|m_{ee}^\mathrm{I\hspace{-.1em}I}\right|&=
\frac{1}{3}\left|m_1(1-\mathrm{cos}2\theta)e^{2i(\phi_1+\sigma)}
-m_2e^{2i\phi_2}+m_3(1+\mathrm{cos}2\theta)e^{2i\phi_3}\right|,
\label{eq:Mee2}
\end{align}
where the superscripts I and I\hspace{-.1em}I denote the types of vacuum.
Equivalently, one can rewrite Eqs.~\eqref{eq:Mee1} and \eqref{eq:Mee2} as
\begin{align}
\left|m_{ee}^\mathrm{I}\right|&=
\frac{1}{3}\left|m_1(1+\mathrm{cos}2\theta)e^{2i(\phi_1-\phi_3+\sigma)}
+m_2e^{2i(\phi_2-\phi_3+\sigma)}+m_3(1-\mathrm{cos}2\theta)\right|,
\label{eq:Mee3}\\
\left|m_{ee}^\mathrm{I\hspace{-.1em}I}\right|&=
\frac{1}{3}\left|m_1(1-\mathrm{cos}2\theta)e^{2i(\phi_1-\phi_3+\sigma)}
-m_2e^{2i(\phi_2-\phi_3)}+m_3(1+\mathrm{cos}2\theta)\right|,
\label{eq:Mee4}
\end{align}
On the other hand, the effective mass in the PDG parametrization is written as
\begin{align}
\left|m_{ee}\right|=\left|
m_1c_{13}^2c_{12}^2e^{2i(\alpha+\delta_{CP})}+
m_2c_{13}^2s_{12}^2e^{2i(\beta+\delta_{CP})}+
m_3s_{13}^2\right|.
\label{eq:MeePDG}
\end{align}
One can obtain the Majorana CP violating phases $\alpha$ and $\beta$ by comparing Eqs~\eqref{eq:Mee3}-\eqref{eq:MeePDG},
\begin{align}
(\mathrm{Type~I})\quad
&\alpha+\delta_{CP}=\phi_1-\phi_3+\sigma\ ,\quad
\beta+\delta_{CP}=\phi_2-\phi_3+\sigma,
\label{eq:phase1}\\[0.3cm]
(\mathrm{Type~I\hspace{-.1em}I})\quad
&\alpha+\delta_{CP}=\phi_1-\phi_3+\sigma\ ,\quad
\beta+\delta_{CP}=\phi_2-\phi_3+\frac{\pi}{2}.
\label{eq:phase2}
\end{align}
\subsection{Numerical Analysis}
In this subsection, we show numerical analysis to find the difference between two types.
We use recent experimental results with $3\sigma$ range~\cite{Esteban:2016qun}, as summarized in Table~\ref{tb:exp}.
\begin{table}[h]
\caption{
The experimental data for the mass squared differences and mixing angles with $3\sigma$ range~\cite{Esteban:2016qun}}
\centering
\begin{tabular}{|c||c|c|} \hline
& Normal Hierarchy (NH) & Inverted Hierarchy (IH) \\ \hline
$\Delta m_{21}^2~\left[\mathrm{eV}^2\right]$ & $(7.03\sim8.09)\times10^{-5}$ & $(7.03\sim8.09)\times10^{-5}$ \\
$\Delta m_{31}^2~\left[\mathrm{eV}^2\right]$ & $(2.407\sim2.643)\times10^{-3}$ & $-(2.565\sim2.318)\times10^{-3}$ \\
$\mathrm{sin}^2\theta_{12}$ & $0.271\sim0.345$ & $0.271\sim0.345$ \\
$\mathrm{sin}^2\theta_{23}$ & $0.385\sim0.635$ & $0.393\sim0.640$ \\
$\mathrm{sin}^2\theta_{13}$ & $0.01934\sim0.02392$ & $0.01953\sim0.02408$ \\ \hline
\end{tabular}
\label{tb:exp}
\end{table}
As we have shown in the previous subsection, the mixing angles and Dirac CP phase are expressed in terms of the model parameters $\theta$ and $\sigma$ in different forms for the two types.
The experimental data for $\sin^2\theta_{13}$ in Table~\ref{tb:exp} is realized by the following value of $\theta$ with NH or IH:
\begin{align}
\mathrm{Type~I}\ ;\quad
9.81^\circ\leq |\theta| \leq10.9^\circ\ (\mathrm{NH})\ ,\
9.86^\circ\leq |\theta| \leq11.0^\circ\ (\mathrm{IH}),\\
\mathrm{Type~I\hspace{-.1em}I}\ ;\quad
79.1^\circ\leq |\theta| \leq80.2^\circ\ (\mathrm{NH})\ ,\
79.0^\circ\leq |\theta| \leq80.1^\circ\ (\mathrm{IH}).
\end{align}
The value of $\sigma$ is allowed in $-180^\circ\leq\sigma\leq180^\circ$ for both of the two types,
since the error of $\sin^2\theta_{23}$ from the experiments is large.
Next, we discuss the parameters $\phi_i$ in the expressions of the Majorana phases of Eqs.~\eqref{eq:phase1} and \eqref{eq:phase2}.
The effective mass $|m_{ee}|$ in Eq.~\eqref{eq:MeePDG} depends on the two combinations of Dirac and Majorana phases, $2(\alpha+\delta_{CP})$ and $2(\beta+\delta_{CP})$.
If we determine both $|m_{ee}|$ and the lightest neutrino mass, we obtain the constraints on these two combinations.
In order to find how the numerical constraints on $\phi_i$ are different in the two types,
we consider a specific situation.
As an example, we assume that $|m_{ee}|$ is predicted in the region as shown in Fig.~\ref{fig:meeA1}.
We note that the lightest neutrino mass is constrained from the cosmological upper bound for the neutrino mass sum, $\sum_im_i<0.16~\mathrm{eV}$~\cite{Alam:2016hwk}.
This plot is obtained when the Dirac and Majorana phases are randomly chosen from the region A1 in Fig.~\ref{fig:block},
\begin{align}
0 < \alpha + \delta_{CP} < \pi/4\ ,\ 0< \beta + \delta_{CP} < \pi/4.
\end{align}
In this situation, the phase differences $\phi_1 - \phi_3$ and $\phi_2 - \phi_3$ for one type can be distinguished from those for the other type.
The constraints on the phase differences are shown in Fig.~\ref{fig:1323}.
For type I, the phase difference $\phi_2-\phi_3$ is proportional to $\phi_1-\phi_3$.
However, for type I\hspace{-.1em}I, $\phi_2-\phi_3$ is independent of the value of $\phi_1-\phi_3$ because $\sigma$ is absent in the expression of $\phi_2-\phi_3$ in Eq.~\eqref{eq:phase2}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=7.0cm]{16.png}
\caption{16 divided regions for $2(\alpha+\delta_{CP})$ and $2(\beta+\delta_{CP})$.}
\label{fig:block}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\linewidth]{A1.pdf}
\caption{
The prediction of effective mass for $0\nu\beta\beta$ decay in region A1 of Fig.~\ref{fig:block}.
The upper region corresponds to the IH case , while the lower one corresponds to the NH case.}
\label{fig:meeA1}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{tabular}{cc}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=\linewidth]{1213nh1.pdf}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=\linewidth]{1213nh2.pdf}
\end{center}
\end{minipage}
\end{tabular}
\caption{
The allowed regions of the model parameters, $\phi_1-\phi_3$ and $\phi_2-\phi_3$ for both types of vacua.
These plots correspond to region A1 of Fig.~\ref{fig:block}.}
\label{fig:1323}
\end{figure}
\newpage
\section{Summary}\label{sec:con}
We have studied phenomenological aspects of a supersymmetric model with $A_4\times Z_3$ symmetry.
We found 24 degenerate vacua at the 24 minima of the scalar potential.
Then, we discussed the relations among the 24 different vacua and classified them into two types.
Both types consist of 12 vacua which are related to each other by transformations of $A_4$.
We proved that the 12 vacua are equivalent and lead to the same physical consequences.
However, we found that we obtain different physical consequences from the vacua of different types.
Therefore, we analyzed the two types of vacua to find the different phenomenological consequences of the two types.
In particular, we investigated observables such as mixing angles, Dirac CP phase, Majorana phases and effective mass for $0\nu\beta\beta$ decay.
These observables are expressed in terms of the model parameters $\theta$, $\sigma$ and $\phi_i$.
The angle $\theta$ and phase $\sigma$ are determined by the deviation from the tri-bimaximal mixing matrix.
The two types lead to different expressions for the mixing angles and Dirac CP violating phase in terms of $\theta$ and $\sigma$.
Therefore, one should take different model parameters in each type in order to realize the experimental results.
Although one can adopt both of the two types to predict the observable parameters,
the two types cannot realize the current experimental data simultaneously.
The Majorana phases $\alpha$ and $\beta$ are parametrized in the different expressions for each type by the model parameters $\phi_i$ in addition to $\theta$ and $\sigma$.
In order to find numerical differences between the two types of Majorana phase,
we considered the specific situation where the lightest mass and effective mass for the $0\nu\beta\beta$ decay are determined in a certain region.
We showed the allowed regions of the phase differences, $\phi_1-\phi_3$ and $\phi_2-\phi_3$.
The regions are quite different for the two types:
the phase differences for type I are proportional to each other,
while those for type I\hspace{-.1em}I are not.
The VEVs $\eta_m$ and $\lambda^\pm_n$ transfer to the different VEVs by transformations of $A_4$.
However, the transformations for $\eta_m$ and $\lambda^\pm_n$ are closed differently since they have the $Z_3$ and $Z_2$ residual symmetries from $A_4$ respectively.
We have pointed out that some combinations of the VEVs can lead to different physical consequences.
When we consider models with two or more flavons, we should take account of the combination of VEVs.
\vspace{0.5cm}
\noindent
{\bf Acknowledgement}
\noindent
This work is supported by JSPS KAKENHI Grant Number JP17K05418 (T.M.).
This work is also supported in part by Grants-in-Aid for Scientific Research [No.~16J05332 (Y.S.), Nos. 24540272, 26247038, 15H01037, 16H00871, and 16H02189 (H.U.)] from the Ministry of Education,
Culture, Sports, Science and Technology in Japan.
H.O. is also supported by Hiroshima Univ. Alumni Association.
\newpage
|
\section{Analysis}
We analyze the statistical differences between reformulation causes prior to the experiments and exploit the findings for engineering features.
\subsection{Analysis of Correction Types} \label{subsec:correction}
\begin{table}[t]
\begin{center}
\begin{tabular}{ll}
\textbf{Type} & \textbf{Example} \\ \hline
ADD & What's the weather in Nara today?\\
OMIT & What's the weather? \\
PAR & How about the weather today? \\
OTHER & How about the humidity today? \\
\end{tabular}
\caption{Example corrections of ``What's the weather today?'' in each correction type.}
\label{tab:correction_examples}
\end{center}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{lllll}
& \textbf{ADD} & \textbf{OMIT} & \textbf{PAR} & \textbf{OTHER} \\ \hline
No error & 27.1 & 7.2 & 58.1 & 7.4 \\ \hdashline
ASR error & 7.5 & 6.0 & 74.2 & 12.3 \\
NLU error & 27.9 & 19.8 & 41.9 & 10.5 \\
LG error & 23.5 & 19.6 & 47.1 & 9.8 \\
\end{tabular}
\caption{Distribution of correction types (\%).}
\label{tab:correction_types}
\end{center}
\end{table}
Here, we analyze the correction types when an IA user is trying to correct a previous utterance.
We simplify the definition of \cite{swerts00} and define four correction types: $ADD$, $OMIT$, $PAR$, and $OTHER$.
Their definitions are as follows.
\begin{description}
\item[ADD] A sequence of words is added to $U_1$.
\item[OMIT] A sequence of words in $U_1$ is omitted in $U_2$.
\item[PAR] A sequence of words in $U_1$ changes into another sequence of words in $U_2$.
\item[OTHER] Match none of preceding correction types.
\end{description}
Note that $U_1$ and $U_2$ have at least one word in common in $ADD$, $OMIT$, and $PAR$.
Table \ref{tab:correction_examples} shows examples of each correction type.
Table \ref{tab:correction_types} presents the distribution of correction types. As shown in Table \ref{tab:correction_types}, the percentage of $PAR$ in \textit{ASR error} is higher than those of other correction types. In \textit{No error}, the percentages of $ADD$ and $PAR$ are large. We can also see that \textit{NLU error} and \textit{LG error} have similar distributions and that $OMIT$ is more frequent than the other correction types. This is because when users encounter \textit{NLU error} or \textit{LG error}, they use different types of corrections depending on the situation: adding words to clarify their intent, omitting words that seem to be the noise for the system, and so on. We expect correction types to be useful for detecting reformulation causes as the distribution of correction types differs for different error types.
\subsection{Analysis of Input Types} \label{subsec:analysis_input_types}
\label{subsec:input_type}
\begin{table}[t]
\begin{center}
\begin{tabular}{lllll}
& \textbf{V2V} & \textbf{T2T} & \textbf{V2T} & \textbf{T2V} \\ \hline
No error & 75.2 & 23.5 & 1.3 & 0.0 \\
ASR error & 94.6 & 0.0 & 5.4 & 0.0 \\
NLU error & 70.5 & 23.8 & 1.7 & 4.6 \\
LG error & 76.5 & 23.5 & 0.0 & 0.0 \\
\end{tabular}
\caption{Distribution of input type switches (\%). For example, V2T means input type of $U_1$ is voice input and that of $U_2$ is text input.}
\label{tab:input_types}
\end{center}
\end{table}
Here, we analyze the correlation between error causes and input types. Input types sometimes differ between $U_1$ and $U_2$. For example, \newcite{Jiang13} have reported that users switch from voice inputs to text inputs when they encounter ASR errors. Table \ref{tab:input_types} presents the distribution of input type switches.
As shown in Table \ref{tab:input_types}, distribution of input type switches differs slightly among the error causes. As \newcite{Jiang13} have shown, voice-to-text switches are more frequent for \textit{ASR error} than in other causes. We can also see that both text-to-voice and voice-to-text switches are observed for \textit{NLU error}. Compared with these causes, input type switches are rarely observed for \textit{No error}. Interestingly, text-to-voice switches are not observed for \textit{No error} either. On the other hand, small number of voice-to-text switches are observed for \textit{No error}. We guess that some users switch input from voice to text when they submit similar query by copy and paste the part of ASR result of their previous utterances. These findings suggest that users tend to keep using the same input type while their intents are correctly recognized. Unlike the other causes, no input type switches are observed for \textit{LG error}. We expect this is due to insufficient data.
\section{Complete List of Features}
\section{Conclusion}
This paper attempted to predict reformulation causes in intelligent assistants (IAs). Prior to the prediction, we first analyzed the cause of reformulation in IAs using user logs obtained from a commercial IA. Based on the analysis, we defined reformulation cause prediction as four-class classification problem of classifying user utterances into \textit{No error}, \textit{ASR error}, \textit{NLU error}, or \textit{LG error}. Features are divided into five categories mainly on the basis of the relations with the components in the IA.
The experiments demonstrated that the proposed method, which combines all feature sets, outperforms the baseline which uses component-independent features such as session information and reformulation related information.
\section{Corpus}
We consisted two datasets of user logs of a commercial intelligent assistant\footnote{Because the intelligent assistant supports only Japanese, all utterances are made in Japanese. In this paper, we present English translations rather than the original Japanese to facilitate non-Japanese readers’ understanding.} for reformulation and positive/negative based on analysis shown in Section \ref{sec:analysis}. We randomly sampled 4,722 utterance pairs of ($U_1$, $U_2$) and corresponding information for reformulation dataset and 4,762 pairs for positive/negative dataset with the following five conditions.
First, $U_1$ and $U_2$ are text or voice inputs (\textit{i.e.} Quasi inputs such as autosuggest are excluded) .
Second, interval time between $U_1$ and $U_2$ is equal to or less than 30 minutes (followed previous research such as \citet{Jiang15} and \citet{Sano16}) .
In reformulation dataset, data where normalized Levenshtein edit distance \cite{Li07} between $U_1$ and $U_2$ equal to zero (\textit{i.e.} $U_1$ and $U_2$ are exactly same utterances) or more than 0.5 are excluded.
In positive/negative dataset, data where $U_2$ does not contain positive nor negative phrases in the categories listed in Table \ref{tab:posi_nega_type} are excluded.
Finally, data which lacks some information (\textit{e.g.} Voice data) or contains identical information is excluded.
As our focus is to predict user satisfaction from user utterances, we made efforts to exclude data where $U_2$ does not relate whether or not $U_1$ is SAT. First four conditions are filtering intended to make datasets high precision with reasonable recall.
Crowdsourcing workers are shown the triplets of texts of ($U_1$, $R$, $U_2$) and asked to annotate whether or not a user is satisfied with the system response $R$.
They are also instructed to estimate the intent of the user prior to the annotation. Note that voice inputs are converted into texts with automatic speech recognition.
To improve annotation quality, following two conditions are applied.
First, the labels are determined by the majority votes of five workers.
Second, workers who wrong the answer of the check problems are excluded.
The annotation results of reformulation dataset are 1,364 SAT data and 3,408 DSAT data, while those of positive/negative dataset are 2,378 SAT data and 2,384 DSAT data.
\section{Experiments}
\subsection{Experimental Settings}
The experimental settings of reformulation cause prediction are as follows.
\begin{itemize}
\item The dataset described in Section \ref{subsec:corpus} is used for evaluation. It contains 928 samples.
\item We evaluate the performance of the model with 10-fold cross validation.
\item We train the model using a linear SVM classifier with the features described in section \ref{sec:features}.
\item We optimize hyper parameters of the classifier with an additional 5-fold cross validation using only training sets (9-folds used for a training set is combined and split into 5 new folds in each validation).
\item The baseline model is trained in the same conditions except that the model uses only Session and Reformulation features. Comparison between the proposed and the baseline methods enables us to evaluate the effect of the features related to the components of IA.
\end{itemize}
We choose linear SVM because it has scalability and has outperformed RBF-kernel SVM.
\subsection{Results} \label{subsec:results}
\begin{table}[t]
\begin{center}
\begin{tabular}{llll}
& \textbf{Precision} & \textbf{Recall} & \textbf{F$_1$} \\ \hline
Baseline (B.) & 0.51 & 0.53 & 0.51 \\
Proposed & \textbf{0.67} & \textbf{0.66} & \textbf{0.67}$^+$ \\ \hdashline
B. + ASR & 0.56 & 0.59 & 0.57$^+$ \\
B. + NLU & 0.63 & 0.60 & 0.61$^\star$ \\
B. + LG & 0.50 & 0.50 & 0.49 \\
\end{tabular}
\caption{The results of the reformulation cause prediction. }
\label{tab:exp_overall}
\end{center}
\end{table}
Table \ref{tab:exp_overall} presents the results for the reformulation cause prediction. The first row compares the proposed method with the baseline. The proposed method obtains a 0.67-point F$_1$-measure and outperforms the baseline. This result shows the effectiveness of the features related to the components of IA. The second row illustrates the performance when one feature set is added to the baseline. We can see that ASR and NLU features improve the performance of the baseline. In F$_1$-measure, statistical significant differences from the baseline detected by the paired t-test are denoted by $+$ ($p < 0.01$) and $\star$ ($p < 0.05$).
\begin{table}[t]
\begin{center}
\begin{tabular}{lcccc}
Gold \textbackslash \ Predict & No & ASR & NLU & LG \\ \hline
No error & \textbf{216} & 132 & 29 & 10\\
ASR error & 67 & \textbf{219} & 21 & 10 \\
NLU error & 61 & 54 & \textbf{52} & 6 \\
LG error & 19 & 17 & 14 & \textbf{1} \\
\end{tabular} \\
(a) Baseline \\
\begin{tabular}{lcccc}
Gold \textbackslash \ Predict & No & ASR & NLU & LG \\ \hline
No error & \textbf{284} & 55 & 27 & 21 \\
ASR error & 38 & \textbf{230} & 37 & 12 \\
NLU error & 44 & 29 & \textbf{81} & 19 \\
LG error & 8 & 12 & 11 & \textbf{20} \\
\end{tabular} \\
(b) Proposed \\
\caption{Confusion matrix of reformulation cause prediction of (a) baseline and (b) proposed methods.}
\label{tab:exp_overall_cm}
\end{center}
\end{table}
Table \ref{tab:exp_overall_cm} presents the confusion matrix of the proposed and baseline methods. The results going diagonally show agreement between the gold labels and the predicted labels. As shown in Table \ref{tab:exp_overall_cm}, the proposed method outperforms the baseline regardless of the gold labels. Again, these results indicate the effectiveness of the features related to the components of IA.
Table \ref{tab:exp_components} presents F$_1$-measures of each gold label. Again, the proposed method outperforms the other methods. Focusing on individual labels, F$_1$-measures of the proposed method are better for \textit{No error} and \textit{ASR error} than for \textit{NLU error} and \textit{LG error}. In other words, F$_1$-measures for \textit{NLU error} and \textit{LG error} are not high. As the performances of individual labels are in the order of the number of samples belonging to the label, we expect that these low performances are mainly due to insufficient data and will improve given sufficient data. Focusing on individual feature sets, ASR and NLU features are useful for distinguishing \textit{No error} from other labels. Note that statistical significant differences from the baseline detected by the paired t-test are denoted by $+$ ($p < 0.01$) and $\star$ ($p < 0.05$).
\begin{table}[t]
\begin{center}
\begin{tabular}{lllll}
& \textbf{No} & \textbf{ASR} & \textbf{NLU} & \textbf{LG} \\ \hline
Baseline (B.) & 0.58 & 0.59 & 0.36 & 0.03 \\
Proposed & \textbf{0.75}$^+$ & \textbf{0.72}$^+$ & \textbf{0.49}$^\star$ & \textbf{0.33}$^+$ \\ \hdashline
B. + ASR & 0.66$^+$ & 0.67$^+$ & 0.35 & 0.16 \\
B. + NLU & 0.71$^+$ & 0.65 & 0.43 & 0.25$^\star$ \\
B. + LG & 0.55 & 0.57 & 0.32 & 0.08 \\
\end{tabular}
\caption{F$_1$-measure in reformulation cause prediction for each label.}
\label{tab:exp_components}
\end{center}
\end{table}
\subsection{Results by Input Types}
Here we analyze the result of Table \ref{tab:exp_overall} by input types.
Table \ref{tab:exp_input_types} presents the results of reformulation cause prediction by input types.
As shown in Table \ref{tab:exp_input_types}, Both the F$_1$-measures of the proposed method in voice inputs and text inputs are 0.66. These results suggest that the proposed method is robust for both input types. On the other hand, the F$_1$-measure of the baseline method in voice inputs is lower than that in text inputs. Particularly, the F$_1$ measure of \textit{No error} in voice inputs is lower than that in text inputs. Table \ref{tab:exp_input_types_filtered} presents the distribution of predicted labels with the following two conditions. First, $U_1$ and $U_2$ are voice inputs. Second, gold labels of all samples are \textit{No error}. As shown in Table \ref{tab:exp_input_types_filtered}, misclassification rate of the proposed method in \textit{No error} as \textit{ASR error} is less than that of the baseline method. In other words, the proposed method distinguishes between \textit{No error} and \textit{ASR error} more accurately compared to the baseline method. These results suggest that ASR features contribute to the performance improvement of the proposed method.
\begin{table}[t]
\begin{center}
\begin{tabular}{lccccc}
& No & ASR & NLU & LG & total \\ \hline
Baseline & 0.49 & 0.58 & 0.33 & 0.03 & 0.47 \\
Proposed & \textbf{0.73} & \textbf{0.71} & \textbf{0.49} & \textbf{0.30} & \textbf{0.66} \\ \hdashline
\# samples & 291 & 300 & 122 & 39 & 752 \\
\end{tabular} \\
(a) $U_1$ and $U_2$ are voice inputs. \\
\begin{tabular}{lccccc}
& No & ASR & NLU & LG & total \\ \hline
Baseline & 0.80 & N.A. & 0.32 & 0.00 & 0.60 \\
Proposed & \textbf{0.81} & N.A. & \textbf{0.42} & \textbf{0.39} & \textbf{0.66} \\ \hdashline
\# samples & 91 & N.A. & 40 & 12 & 143 \\
\end{tabular} \\
(b) $U_1$ and $U_2$ are text inputs. \\
\caption{F$_1$-measure in reformulation cause prediction of each label between input types. }
\label{tab:exp_input_types}
\end{center}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{lcccc}
Predicted label & No & ASR & NLU & LG \\ \hline
Baseline & 129 & 129 & 25 & 8 \\
Proposed & 205 & 53 & 16 & 17 \\
\end{tabular}
\caption{ The number of predicted samples in reformulation cause prediction in following two conditions. First, $U_1$ and $U_2$ are voice inputs. Second, gold labels of all samples are \textit{No error}. }
\label{tab:exp_input_types_filtered}
\end{center}
\end{table}
\subsection{Investigation of Feature Weights}
\begin{table}[t]
\footnotesize
\begin{center}
\begin{tabular}{llc}
Label & Feature & Weight \\ \hline
No error & ASRConf* & 1.07 \\
& IntentType(SingSong)* & 1.05 \\ \hdashline
& Voice2Text & -1.01 \\
& IntentType(DeviceControl)* & -1.10 \\ \hline
ASR error & Voice2Text & 1.27 \\
& IntentType(Search)* & 0.90 \\ \hdashline
& ASRConf* & -1.51 \\
& InputType* & -1.77 \\ \hline
NLU error & Text2Voice & 1.43 \\
& InputType* & 0.94 \\ \hdashline
& SameIntent & -0.61 \\
& IntentType(Dictionary)* & -0.77 \\ \hline
LG error & IntentType(DeviceControl)* & 1.08 \\
& DialogAct(IDU)* & 0.81 \\ \hdashline
& DialogAct(Praise)* & -0.95 \\
& Voice2Text & -1.04 \\
\end{tabular}
\caption{ Top and Bottom two feature weights of the proposed method. Features marked with ``*'' were computed for $U_1$. }
\label{tab:feature_weights}
\end{center}
\end{table}
We investigate weights of the features learned by the linear-kernel SVM to clarify what features contribute to the reformulation cause prediction.
Table \ref{tab:feature_weights} presents top and bottom feature weights in each label. The median value of the 10 models which are obtained with cross validation are used for weights in Table \ref{tab:feature_weights}.
Features calculated from $U_1$ appear in Table \ref{tab:feature_weights} but that calculated from $U_2$ do not appear. This result is not surprising because information related to $U_1$ has more relationship to reformulation causes compared to that related to $U_2$.
Next, we focus on the features in individual labels. Features related to their corresponding components appear in Table \ref{tab:feature_weights} such as \textit{ASRConf} in \textit{ASR error}, \textit{SameIntent} in \textit{NLU error}, and \textit{DialogAct} in \textit{LG error}. These results indicate that features designed to detect reformulation causes related to their corresponding components work as designed.
Finally, we focus on the individual features. We observe that features of input type switches are useful for predicting reformulation causes. In particular, \textit{Voice2Text} is useful for detecting \textit{ASR error} and \textit{Text2Voice} is useful for detecting \textit{NLU error}. These results are consistent with findings in section \ref{subsec:analysis_input_types}.
\section{Features} \label{sec:features}
\begin{table*}[t]
\begin{center}
\begin{tabular}{lll}
\textbf{Category} & \textbf{Name} & \textbf{Definition}\\ \hline
Session & \textbf{CharLen*} & Number of characters in utterance. \\
& \textbf{WordLen*} & Number of words in utterance. \\
& \textbf{InputType*} & 1 if utterance is text input else 0 \\
& \textbf{Interval} & Time between $U_1$ and $U_2$ \\ \hdashline
Reformulation & \textbf{EditDistance} & Normalized Levenshtein edit distance between $U_1$ and $U_2$ \\
& \textbf{Correction($t$)} & 1 if $U_2$ is correction type $t$ of $U_1$ \\
& \textbf{CommonWords} & Number of words appearing in both $U_1$ and $U_2$ \\
& \textbf{Voice2Text} & 1 if $U_1$ is voice input and $U_2$ is text input \\
& \textbf{Text2Voice} & 1 if $U_1$ is text input and $U_2$ is voice input \\ \hdashline
ASR & \textbf{ASRConf*} & Speech recognition confidence \\
& \textbf{VoiceLen*} & Speech recognition time \\ \hdashline
NLU & \textbf{SameIntent} & 1 if recognized intents between $U_1$ and $U_2$ are same \\
& \textbf{DifferentIntent} & 1 if recognized intents between $U_1$ and $U_2$ are different \\
& \textbf{DifferentSlot} & 1 if some slots in $U_2$ are different from those in $U_1$ \\
& \textbf{IntentType($t$)*} & 1 if recognized intent of utterance is $t$ \\ \hdashline
LG & \textbf{DialogAct($t$)*} & 1 if utterance contains phrases in dialogue act $t$ \\
\end{tabular}
\caption{List of features. Features marked with ``*'' were computed for both $U_1$ and $U_2$.}
\label{tab:feature}
\end{center}
\end{table*}
We divide the features into five categories by their functions. Session features and reformulation features, which are useful for reformulation prediction \cite{Hassan15} and system evaluation \cite{Jiang15}, are designed to distinguish between errors or non-errors. Though these features are also useful for reformulation cause detection, these are not specialized for detecting reformulation causes. ASR, NLU, and LG features are designed to detect reformulation causes related to their corresponding components from causes related to the other components. Table \ref{tab:feature} lists the features.
\subsection{Session Features}
Features related to session information belong to this category. In \textit{InputType}, 1 if an utterance is text input and 0 if an utterance is voice input \cite{Hassan15}. In a web search system, the interval time between inputs is a useful indicator of search success \cite{Huang09}. Therefore, \textit{Interval} is useful for distinguishing \textit{No error} from the other labels. If \textit{CharLen} or \textit{WordLen} of the utterance is long or short, the utterance possibly contains noise information or lacks information for the system.
\subsection{Reformulation Features}
Features related to reformulation belong to this category. Features in this category are widely used in previous methods such as query reformulation detection \cite{Hassan15}, error detection \cite{Litman06}, and system performance evaluation \cite{Jiang15}. The correction type $t$ of \textit{Correction($t$)} is one of ADD, OMIT, PAR, or OTHER described in Section \ref{subsec:correction}. As shown in Section \ref{subsec:correction}, the distribution of correction types differs among annotation labels. Therefore, \textit{Correction($t$)} is useful information for predicting reformulation causes. \textit{Voice2Text} and \textit{Text2Voice} are designed to distinguish \textit{ASR error} and \textit{NLU error} from other errors on the basis of analysis in Section \ref{subsec:input_type}.
\subsection{ASR Features}
Features related to the ASR component belong to this category. Low \textit{ASRConf} indicates speech recognition errors \cite{Hassan15}. The probability of misrecognition increases as the recognized voice length increases \cite{Hassan15}. Therefore, long \textit{VoiceLen} may be one signal related to ASR error. Note that these features are calculated only when the input type of the utterance is voice input.
\subsection{NLU Features}
Features related to the NLU component belong to this category. When a user's intent in $U_1$ is misunderstood and that in $U_2$ is correctly understood, recognized intents or filled slots between the utterances are different. On the other hand, when a user's intent in $U_1$ is correctly understood by the system, the user sometimes uses similar functions subsequently (e.g., requesting weather information for Osaka after requesting it for Tokyo.). Therefore, \textit{DifferentIntent} and \textit{DifferentSlot} are useful to distinguish \textit{NLU error} from the other errors, and \textit{SameIntent} is useful to distinguish \textit{No error} from the other errors. Note that the intents used in these features are the intents recognized by the IA system such as weather information, web search, application launch, and chat.
\subsection{LG Features}
\begin{table}[t]
\begin{center}
\begin{tabular}{ll}
\textbf{Type} & \textbf{Examples} \\ \hline
Praise & Wow!; Great. \\
Thanking & Thanks.; Thank you. \\
Backchannel & I see.; Yeah. \\
Accept & Yes.; Exactly. \\
Abuse & Shit.; Shut up. \\
Reject & No.; Not like that. \\
IDU & What do you mean? \\
\end{tabular}
\caption{List of dialog acts.}
\label{tab:dialog_act}
\end{center}
\end{table}
Features related to the LG component belong to this category. If users expect the system to chat, their utterances may contain phrases commonly used in chatting. \textit{DialogAct($t$)} is designed to detect these phrases in the utterance. User utterances of chat intent in IAs have unique characterictics (e.g., some users curse at the intelligent assistants \cite{Akasaki17}). We defined seven types of dialogue acts that are common in chats between users and IAs, as listed in Table \ref{tab:dialog_act}. Frequently occurring phrases in the user log of the commercial IA are used for the phrases of each dialogue act.
\section{Future Work}
While the proposed method has outperformed the baseline, there is room for improvement on the performance.
As mentioned in section \ref{subsec:results}, the performances of individual labels are in the order of the number of samples belonging to the labels. Therefore, we expect that the performance improves as the dataset size increases.
The performance will also improve if some features used in previous studies are added to our features. For example, linguistic features such as word n-gram and language model score are used in previous studies \cite{Hassan15, Meena15} but not used in ours.
\section{Introduction}
Intelligent assistants (IAs) such as Apple's Siri and Cortana have gained considerable attention as mobile devices have become prevalent in our daily lives. They are hybrids of search and dialogue systems that conversationally interact with users and execute a wide range of actions (e.g., searching the Web, setting alarms, making phone calls and chatting). IAs can support these actions through the combination of various components such as automatic speech recognition (ASR), natural language understanding (NLU), and language generation (LG).
One major concern in the development of commercial IAs is how to speed up the cycle of the system performance enhancement. The enhancement process is often performed by manually investigating user logs, finding erroneous data, detecting the component responsible for that error, and updating the component. As IAs are composed of various components, the error cause detection becomes an obstacle in the manual process. In this paper, we attempt to automate the error cause detection to overcome this obstacle.
One approach to do this is to utilize user feedback. In this work, we focus on reformulation, i.e., when a user modifies the previous input. In web search and dialogue systems, reformulation is known as an implicit feedback signal that the user could not receive a desired response to the previous input due to one or more system components failing.
In IAs, ASR error is a major cause of reformulation and has been extensively studied \cite{Hassan15,Schmitt15}.
Besides correcting ASR errors, users of IAs reformulate their previous utterances when they encounter NLU errors, LG errors, and so on. For example, when a user utters ``alarm'', the NLU component may mistakenly conclude that s/he wants to perform a Web search, and consequently the system shows the search results for ``alarm.'' \newcite{Sarikaya17} reported that only 12\% of the errors in an IA system are related to ASR components, which is the smallest percentage across six components. They also reported that the NLU component is the biggest source of errors (24\%). Therefore, the errors related to the other components should not be ignored to improve system performance. However, previous work mainly focused on reformulation caused by ASR error, and reformulation caused by the other components has received little attention.
In this work, we propose a method to predict reformulation causes, i.e., detect the component responsible for causing the reformulation in IAs. Features are divided into several categories mainly on the basis of their relations with the components in an IA. The experiments demonstrate that these features can improve the error detection performances of corresponding components. The proposed method which combines all features sets, outperforms the baseline, which uses component-independent features such as session information and reformulation related information.
Our work makes the following contributions. First, we investigate the reformulation causes among the components in IAs from real data of a commercial IA. Second, we create dataset of human annotated data obtained from a commercial IA. Finally, we develop the method to predict reformulation causes in IAs.
\section{Related Work}
Three research areas are related to our work, and the most closely related is reformulation (also called correction). As reformulation is frequently caused by system errors, the second related area is error analysis and error detection in search or dialogue systems. The third area is system evaluation in search or dialogue systems. Reformulation is a useful feature for system evaluation.
\subsection{Reformulation and Correction}
Users of search or dialogue systems often reformulate their previous inputs when trying to obtain better results \cite{Hassan13} or correct errors (e.g., ASR errors) \cite{Jiang13, Hassan15}. Our research focuses on the latter category of reformulation, which relates to correction. Studies on correction have mainly focused on automatic detection of correction \cite{Levow98,Hirschberg01a,Litman06}.
Some studies have also tried to improve system performance beyond correction detection. \newcite{Shokouhi16} constructed a large-scale dataset of correction pairs from search logs and showed that the database enables ASR performance to be improved.
The research most related to ours is that of \newcite{Hassan15}.
In addition to reformulation detection, they proposed a method to detect whether a reformulation is caused by ASR error or not. We extend their study to determine which component (e.g., ASR, NLU, and LG) is responsible for the error.
\subsection{Error Analysis and Error Detection}
Besides reformulation, researchers have studied system errors in search or dialogue systems. For example, \newcite{Meena15} and \newcite{Hirst94} focused on miscommunication in spoken dialogue systems and \newcite{Feild10} focused on user frustrations in search systems. In this paper, we focus on predicting the cause of errors among the different components in IAs. In spoken dialogue systems, \newcite{Georgiladakis16} reported that ASR error is the most frequent cause of errors among seven components (65.9\% in Let's Go datasets \cite{Raux05}. On the other hand, \newcite{Sarikaya17} reported that ASR error is the least frequent cause of errors among six components (12\% in an IA). These results indicate that errors in IA have various causes and that the causes other than ASR error also should not be ignored. In this paper, we focus on reformulation and propose a method to automatically detect the reformulation causes in IAs.
\subsection{System Evaluation}
Users of the search or conversational systems often reformulate their inputs when they are dissatisfied with the system responses to their previous inputs. Therefore, reformulation is a useful signal of user dissatisfaction and information related to reformulation has been widely used as a feature to automatically evaluate system performance in web search \cite{Hassan13}, dialogue \cite{Schmitt15}, and IA systems \cite{Jiang15, Kiseleva16a, Sano16}. However, these studies paid little attention to detecting causes of user dissatisfaction. Information of these causes is beneficial to the developers for both improving system design and engineering feature of automatic evaluation methods. Thus, we propose a method to automatically detect the reformulation causes in IAs.
\section{Reformulation Cause Prediction} \label{sec:analysis}
In this section, we describe the task of reformulation cause prediction in IAs.
\begin{table*}[t]
\begin{center}
\footnotesize
\begin{tabular}{llll}
\textbf{Label} & \textbf{U1} & \textbf{R} & \textbf{U2} \\ \hline
No error & What's the weather? & It will be sunny today.& What's the weather tomorrow? \\
ASR error & What's the. & Sorry? & What's the weather? \\
NLU error & Alarm. & Here are the search results for ``Alarm''.& Open alarm. \\
LG error & What's your name? & I'm twenty years old. & Tell me your name. \\
Unsupported action & Play videos of cats. & Sorry, I can't support that action. & Search for videos of cats. \\
Endpoint error & Search Obama's age.& No results found in for ``Obama's age''. & Search Obama. \\
Uninterpretable input & Aaaa. & Sorry? & Aaa. \\
\end{tabular}
\caption{List of annotation labels. U1, R, and U2 are example conversations. }
\label{tab:annotation}
\end{center}
\end{table*}
\begin{table}[t]
\begin{center}
\begin{tabular}{lll}
\textbf{Label} & \textbf{Rate} & \textbf{Error Rate}\\ \hline
No error & 38.7\% & N.A.\\ \hdashline
ASR error & 31.7\% & 57.2\% \\
NLU error & 17.3\% & 31.2\%\\
LG error & 5.1\% & 9.2\% \\
Unsupported action & 0.8\% & 1.4\% \\
Endpoint error & 0.5\% & 0.9\% \\ \hdashline
Uninterpretable input & 5.9\% & N.A.\\
\end{tabular}
\caption{Percentage of annotation labels in our dataset. Error rate is calculated using labels related to reformulation causes.}
\label{tab:label_percentage}
\end{center}
\end{table}
\subsection{Definition}
First, we define notations used in this paper.
\begin{description}
\item[\bf $U_1$]: The user utterance
\item[$R$]: The corresponding system response to $U_1$
\item[$U_2$]: The next utterance to $U_1$
\item[Reformulation]: A pair of ($U_1$, $U_2$) is a \textit{reformulation} if $U_2$ is uttered to modify $U_1$ in order to satisfy the same intent as in $U_1$
\end{description}
To define the \textit{reformulation}, we referred to the definition in the work of \newcite{Hassan15} that is used for voice search systems.
\subsection{Corpus} \label{subsec:corpus}
We constructed a dataset of user logs of a commercial IA\footnote{Because the IA supports only Japanese, all utterances are made in Japanese. In this paper, we present English translations rather than the original Japanese to facilitate non-Japanese readers’ understanding.} for analyzing and predicting reformulation causes. We randomly sampled 1,000 utterance pairs of ($U_1$, $U_2$) and corresponding information with the following conditions.
\begin{itemize}
\item $U_1$ and $U_2$ are text or voice inputs
\item Interval time between $U_1$ and $U_2$ is equal to or less than 30 minutes (the same as in previous research \cite{Jiang15,Sano16}.)
\item Samples where normalized Levenshtein edit-distance \cite{Li07} between $U_1$ and $U_2$ is equal to 0 (i.e., $U_1$ and $U_2$ are identical utterances) or more than 0.5 are excluded
\item $U_1$ and $U_2$ were both uttered in June, 2016
\end{itemize}
With the first three conditions, we can exclude utterances that are not reformulations and can focus on reformulation.
We calculated character-based, rather than word-based, edit-distance (white spaces are ignored), because we found word-based edit-distances sometimes fail to identify reformulation pairs such as "what's up" and "whatsapp."
\begin{figure*}[t]
\includegraphics[width=0.9\linewidth]{components.pdf}
\caption{Components and processes of typical IA system.}
\label{fig:components}
\end{figure*}
All samples in the dataset are manually annotated with the label of the reformulation causes between different components in IAs. Table \ref{tab:annotation} lists the annotation labels. Here, we explain them. In this paper, we assume an IA system that has the components shown in Figure \ref{fig:components}. \textit{ASR error} means that the ASR component misrecognizes $U_1$. \textit{NLU error} means that ASR is correct but the NLU component misunderstands the intent of $U_1$ or fails to fill one or more slots correctly. \textit{LG error} means that ASR and NLU are correct but the LG component fails to generate appropriate response. This error mainly caused by response generation failure in chat intent. Note that the NLU component only determines whether or not an utterance is chat intent and the LG component generates appropriate response of the utterance in our system. Therefore, we consider that the case shown in Table \ref{tab:annotation} is LG error rather than NLU error. \textit{Endpoint error} means that endpoint API (application program interface) fails to respond with correct information. \textit{Unsupported action} means that the system cannot support the action that the user expects and so cannot generate a correct response. \textit{No error} means that a sample contains no error. Submitting similar utterances to obtain better results in search intents (e.g., $U_1$ is ``Search for image of strawberry wallpaper'' and $U_2$ is ``Search for image of strawberry'') and using similar functions (e.g., $U_1$ is ``Turn on Wi-Fi'' and $U_2$ is ``Turn off Wi-Fi'') are typical utterances in this label. Finally, \textit{Uninterpretable input} means that $U_1$ is uninterpretable.
As expert knowledge about the components of IAs is required for the annotation, annotation is performed by an expert developer of the commercial IA.
Figure \ref{fig:annotation_flow} shows the annotation flowchart. First, we listen to the voice of $U_1$ and read the texts of $U_1$, $R$, and $U_2$. Text information is used to support guessing the user intent of $U_1$.
\textit{Uninterpretable input} such as one-word utterances and misrecognized input of background noises is distinguished in this phase. Next, \textit{ASR error} samples are distinguished using transcription of $U_1$.
Afterwards, \textit{No error} samples are distinguished if $R$ of the samples correctly satisfies the intent of $U_1$.
Finally, one of the other four error causes is annotated to the remaining samples.
The annotation results are shown in Table \ref{tab:label_percentage}. We ignored the cases where the latter components could recover from the errors generated by the former components because these cases were rarely observed. For example, a percentage that the NLU component could recover from ASR errors is 1.2\% in our dataset.
\subsection{Discussion on Annotated Labels}
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{emnlp_annotation_flow_2.pdf}
\caption{Annotation flowchart.}
\label{fig:annotation_flow}
\end{figure}
Here, we discuss the annotation results and which labels should be included or excluded in reformulation cause prediction.
To come to the point, we do not use \textit{Uninterpretable input}, \textit{Unsupported action}, or \textit{Endpoint error} for reformulation cause prediction. We will explain why these labels are excluded.
As shown in Table \ref{tab:label_percentage}, the most frequent cause is \textit{ASR error}. This result differs from that of \newcite{Sarikaya17} in which ASR error is the least frequent cause. The reason for the difference is that \newcite{Sarikaya17} used whole samples, whereas we use only reformulation samples. These results indicate that the reformulation tendency when a user encounters an error differs depending on the cause of errors. For example, the percentage of unsupported actions is 1.4\%, which is much smaller than that reported by \newcite{Sarikaya17}, 14\%. This finding indicate that when users encounter a response that notifies them that the action they expected was unsupported, they would rather give up than reformulate their previous utterances. The same is true for endpoint error.
Next, we discuss which labels should be included or excluded in the task. First, \textit{Uninterpretable input} should be excluded because these utterances have no appropriate responses and become noise for the task. We also exclude \textit{Unsupported action} and \textit{Endpoint error}. An IA do not require user feedback for these errors because no components in the IA is responsible for the errors and the IA is aware the error causes in these labels. In addition, the benefits of detecting these errors are limited because these errors are rarely observed in reformulation as shown in Table \ref{tab:label_percentage}. In conclusion, we use \textit{No error}, \textit{ASR error}, \textit{NLU error}, and \textit{LG error} for the task.
\subsection{Task Definition}
Here, we describe the task of reformulation cause prediction.
Our goal is to predict the reformulation cause of $U_1$ between different components using the information from $U_1$ and $U_2$. Specifically, we predict one of the four labels in Table \ref{tab:annotation}, \textit{No error}, \textit{ASR error}, \textit{NLU error}, and \textit{LG error}. For example, we want to predict as \textit{NLU error} from the conversation logs of ($U_1$:``Alarm.'', $R$:``Here are the search results for Alarm.'', $U_2$:``Open alarm.''). These labels are useful for IA developers to improve system performance.
|
\section*{Appendix I. Mean-field solution of the model Hamiltonian}
Considering that the boundary condition is open in the $x$ direction and periodic in the $y$ direction, the mean-field decoupling of the Hubbard model can be transferred into the half-momentum space ($k$ for the $y$ direction) as
\begin{eqnarray} \label{Sq2}
H &=&-t\sum_{i,k\sigma }\left( \gamma _{k}a_{i,k\sigma }^{+}b_{i,k\sigma
}+a_{i,k\sigma }^{+}b_{i-1,k\sigma }\right) +\text{h.c.} \nonumber \\
&&+\sum_{i,k\sigma }\left(\mu _{i,\sigma}^{A}a_{i,k\sigma }^{+}a_{i,k\sigma
}+\mu _{i,\sigma}^{B}b_{i,k\sigma }^{+}b_{i,k\sigma }\right)+E_{0}\text{,}
\end{eqnarray}%
where $a$ and $b$ are the electron annihilation operators in the $A$ and $B$ sublattices, respectively. $%
\gamma _{k}=\left( 1+e^{-ik}\right) $ with the distance between the
nearest-neighbor atoms in the same sublattice being the unit. $%
E_{0}=-U\sum_{i\eta }\left( \left\langle n_{i}^{\eta }\right\rangle
^{2}/4-\left( m_{i}^{\eta }\right) ^{2}\right)$ with $\eta=A, B$. The effective chemical potential $\mu_{i,\sigma}^{\eta}=-\langle n_{i}^{\eta}\rangle/2+\sigma m_{i}^{\eta}-\mu$. The
charge density and the spin polarization at a given site are defined as $%
\left\langle n_{i}^{\eta }\right\rangle =\sum_{\sigma}\langle
n_{i\sigma}^{\eta }\rangle$ and $m_{i}^{\eta }=\sum_{\sigma}\sigma\langle
n_{i\sigma }^{\eta }\rangle/2$. The effective mean-field Hamiltonian has
been shown to yield results in good agreement with first-principles
calculations \cite{Yazyev-PRL2008,Jung-PRB2009,Rossier-PRB2008}.
\begin{figure}[tbp]
\vspace{-0.3in} \hspace{-0.0in} \center
\includegraphics[width=0.5\columnwidth]{SI-Eanalysis.eps} \vspace{-0.0in}
\caption{Analysis of energy difference between the AFCE and FMCE states for $\delta=0.012$ with $U=2.0t$. (a) Difference in the free energy $\delta F$, total energy $\delta E$, and the entropy part $T\delta S$. (b) Individual contributions from the kinetic, magnetic, and charge sectors ($\delta E_K$, $\delta E_m$, and $\delta E_n$, respectively). The light gray lines result from Landau theory presented below with fitting parameter $x_{0}=0.5$, $\gamma=1.1$, $\lambda=12.3$, $\alpha=-2.8$, and $\beta=49.1$.}
\label{S1}
\end{figure}
The free energy of the system is given by
\begin{equation}\label{Sq3}
F=-\frac{k_{B}T}{N_{y}}\sum_{k\sigma }\sum_{\nu =1}^{2N}\ln \left(
1+e^{-E_{k\sigma }^{\nu}/k_{B}T}\right)+\mu\sum_{\eta }\sum_{i =1}^{N}\langle n_{i }^{\eta }\rangle-E_{0},
\end{equation}
where $E^{\nu}_{k\sigma}$ is the eigenvalue of Eq.~(\ref{Sq2}) and $N_{y}$ is the number of $k$ points. The temperature $T$ is fixed at $0.01t$. Therefore, the magnetic properties discussed in present paper can be applied at the room temperature. The total energy $E$ is
\begin{equation}\label{Sq4}
E=\langle H \rangle=-\frac{\partial}{\partial \beta}\ln Z=\frac{1}{N_{y}}\sum_{k\sigma}\sum_{\nu=1}^{2N}E_{k\sigma}^{\nu}f_{k\sigma}^{\nu},
\end{equation}
where $Z$ is the partition function, $\beta=1/k_{B}T$, and $f_{k\sigma}^{\nu}=1/(1+e^{\beta E_{k\sigma}^{\nu}})$ is the Fermi distribution.
The total energy can be further divided into three parts, i.e. $E=E_{K}+E_{m}+E_{n}$ with $E_{K}=-t\langle \sum_{\left\langle i\text{,}j\right\rangle \sigma }c_{i\sigma }^{\dag
}c^{}_{j\sigma } \rangle$, $E_{m}=-U\sum_{\eta}\sum_{i=1}^{N} (m_{i}^{\eta})^{2}$, and $E_{n}=\frac{U}{4}\sum_{\eta }\sum_{i =1}^{N}\left\langle n_{i}^{\eta }\right\rangle
^{2}$ being the energy from kinetic energy, magnetization, and charge density, respectively. In Fig.~\ref{S1} we show the relative energy for $\delta=0.012$, whose oscillatory features look qualitatively similar to those of Fig.~3 in the main text for $\delta=0.004$. The AFCE-FMCE phase transition determined by the total free energy tracks well with the energy from magnetization.
\begin{figure}[tbp]
\vspace{-0.3in} \hspace{-0.0in} \center
\includegraphics[width=0.8\columnwidth]{SI-Phase.eps} \vspace{-0.0in}
\caption{(a) Magnetic phase diagram as a function of the ribbon width and doping level $\delta$ with $U=1.5t$. (b) Bandgap $\Delta$, measuring the energy difference between the lowest upper band and highest lower band, as a function of $N$ for $\delta=0.004$. (c) Strength of the edge magnetization in both the AFCE and FMCE phases for $\delta=0.004$.}
\label{S2}
\end{figure}
The magnetic phase oscillations are robust against the on-site Coulomb repulsion $U$ adopted in various literatures. We plot the phase diagram with a weaker on-site repulsion $U=1.5t$ in Fig.~\ref{S2}. Compared with the stronger repulsion with $U=2.0t$ presented in the main text, the magnetic phase oscillations remain but with the expanded paramagnetic phase. Especially, the resultant phase transitions change little for lower dopings below $\delta=0.008$ though the period of oscillation increases slightly. In Fig.~\ref{S3} we further show the difference in the free energy between the AFCE and FMCE states for different on-site repulsion $U$ at the fixed doping $\delta=0.004$. The phase oscillation remains robust even with weak enough $U=1.2t$.
\begin{figure}[!bp]
\vspace{-0.2in} \hspace{-0.0in} \center
\includegraphics[width=0.6\columnwidth]{SI-diffU.eps} \vspace{-0.0in}
\caption{Difference in the free energy between the AFCE and FMCE states for $\delta=0.004$ with various on-site repulsion $U$. The data for $U=1.5$ and $U=1.2$ have been enlarged by a factor for comparison.}
\label{S3}
\end{figure}
In Fig.~\ref{S4}, we show the distributions of magnetization inside the nanoribbon for both AFCE and FMCE states for more values of $N$ than in Fig.~2 of the main text. Evident oscillation of magnetization can be found in the bulk. The oscillatory period is about $22$ ($33$a with a the carbon-carbon bond length) for $\delta=0.004$ and $U=2.0t$ as shown in the main text. To gain positive coherence from the opposite edge, the magnetic correlation between the two opposite edges has to alternate between AFCE and FMCE when the ribbon width changes. This is particularly evident for large widths. For example $N=36$, the edge magnetization acquires the positive coherence from another edge in FMCE state while negative coherence from another edge in AFCE state.
\begin{figure}[tbp]
\vspace{-0.3in} \hspace{-0.0in} \center
\includegraphics[width=0.8\columnwidth]{SI-Mdist.eps} \vspace{-0.0in}
\caption{Distributions of magnetization of the \emph{i}th zigzag chain of nanoribbon for (a) AFCE and (b) FMCE states at fixed $\delta=0.004$ and $U=2.0t$. The data has been renormalized by the magnitude along the edge to highlight the oscillations inside the bulk. The magnetization of \emph{i}th zigzag chain is defined in the main text.}
\label{S4}
\end{figure}
\section*{Appendix II. Landau theory description for AFCE-FMCE phase transition}
As mentioned in the main text, the stability of the magnetic phase is dominated by the distribution magnetization, our starting point is the
Landau theory with the single order parameter $M(x)$ as
\begin{equation}
F=\int dx\left[ f_{0}+\frac{1}{2}\alpha M^{2}\left( x\right) +\frac{1}{4}%
\beta M^{4}\left( x\right) -h M\left( x\right) +\cdots \right] \text{,}
\end{equation}
where $\alpha <0$, $\beta>0$ are fitting parameters, $h$ is the effective external magnetic field. $M(x)$ is the magnetization of the \emph{x}th zigzag chain, which is defined as $M_{i}=m_{i}^{A}-m_{i}^{B}$ for $x=i$ ($i=1$, $2$, $\cdots$, $N$) to account for the antiferromagnetic correlations between the nearest neighbor sublattice $A$ and $B$\cite{Fallarino-PRB2015}. The magnetic momentum follows an oscillatory decay from the boundaries and can be well simulated by%
\begin{eqnarray}
M^{\pm }\left( x\right) &=&M_{L}^{\pm }e^{-\gamma \sqrt{x-1}}\cos \left( \kappa
(x-1-x_{0})\right) \nonumber \\
&&\pm M_{R}^{\pm }e^{-\gamma \sqrt{N-x}}\cos \left(
\kappa \left(N-x-x_{0}\right)\right) \text{,}
\end{eqnarray}%
where $M^{\pm}_{L}\cos(\kappa x_{0})$ and $M^{\pm}_{R}\cos(\kappa x_{0})$ is the
order parameter at the left ($x=1$) and right edge ($x=N$) at sufficiently large size ($N\gg 1$), respectively. $\kappa x_{0}$ is
an initial phase introduced to well simulate our numerical data. $\kappa
=2\pi /\lambda $ with $\lambda $ the oscillatory period. The superscript $%
+/-$ corresponds to the AFCE/FMCE state.
The Euler-Lagrange equation at the two edges ($x=1$ and $x=N$) satisfies
\begin{equation}
\alpha M(x)+\beta M^{3}(x)-h=0\text{,}
\end{equation}
where the superscript $\pm $ has been neglected.
\bigskip
\textit{In the absence of the external magnetic field}---The Euler-Lagrange
equation at the two edges reduces to $M^{\pm }\left( 1\right) =\sqrt{%
\left\vert \alpha /\beta \right\vert }$ and $M^{\pm }\left( N\right) =\pm
\sqrt{\left\vert \alpha /\beta \right\vert }$, which requires $M_{L}^{\pm
}=M_{R}^{\pm }=M_{1}^{\pm }$. The value of $M_{1}^{\pm }$ can be evaluated at the limit of $N\rightarrow
\infty $, yielding
\begin{equation}
M_{1}^{+}=M_{1}^{-}=M_{1}=\sqrt{\left\vert \alpha /\beta \right\vert }/\cos
(\kappa x_{0})\text{,}
\end{equation}
in which the two edges are no longer correlated with each other and
therefore the difference between AFCE and FMCE disappears. In the present
case, since $M^{2}\left( x\right) \ll 1$, the dominating contribution comes
from the first term in the Landau free energy, especially
the magnetization near the edge due to exponentially decay departing from the edges.
To minimizing the total free energy, strong edge spin polarization $m(1)$ and $m(N)$ are expected.
At the edge ($x=1$)
\begin{equation}
M^{\pm}(1)=\sqrt{\left\vert\frac{\alpha}{\beta}\right\vert}\left[1\pm e^{-\gamma \sqrt{N-1}}\cos \left(
\kappa (N-1-x_{0})\right)\right].
\end{equation}
When $2n\pi-\pi/2<\kappa (N-1-x_{0})<2n\pi+\pi/2$, $M^{+}(1)>M^{-}(1)$, otherwise, $M^{+}(1)<M^{-}(1)$. The amplitude of the edge magnetization $M^{+}(1)$ and $M^{-}(1)$ is alternatively dominant upon the numbers of the zigzag chains $N$. The edge magnetization acquires the positive coherence by adjusting the edge magnetic correlations, which naturally generates the AFCE-FMCE phase oscillations. The simple Landau theory description is in good agreement with the numerical calculations, manifesting its validity. Therefore, the physics behind the phase oscillations is the quantum confinement (finite size $N$) and quantum interference (positive coherence).
\bigskip
\textit{In the presence of the external magnetic field}---The inversion antisymmetry in AFCE state is broken under the external magnetic field, leading to $M_{L}\neq M_{R}$. We consider the infinite case ($N\rightarrow
\infty $) to show this discrepancy. The induced Euler-Lagrange equation in
this situation at the two edges is (the superscript $+$ has been neglected
for simplicity) $\alpha M_{L}\cos(\kappa x_{0})+\beta \left( M_{L}\cos
(\kappa x_{0})\right)^{3}-h=0$ at left edge ($x=1$), and $\alpha \left(
M_{R}\cos (\kappa x_{0})\right) +\beta \left( M_{R}\cos(\kappa x_{0})\right)
^{3}+h=0$ at right edge ($x=N$). Using the Cardano formula for one variable
cubic equation, we have%
\begin{equation}
x_{2}=\omega \left( -q+\left( q^{2}+p^{3}\right) ^{\frac{1}{2}}\right) ^{\frac{1}{3}}+\omega
^{\ast }\left( -q-\left( q^{2}+p^{3}\right) ^{\frac{1}{2}}\right) ^{\frac{1}{3}}\text{,}
\end{equation}
where $p=\beta /2\alpha $, $q=\pm h/3\alpha $ ($+$ for the right, and $-$ for
the left edge), and $\omega =(-1+\sqrt{3}i)/2$. Usually, the external magnetic field
is $3\sim4$ orders weaker than the spontaneous magnetization, we expand the above
equation to the first order of the external magnetic field $h$, yielding
\begin{equation}
M_{L/R}\cos (\kappa x_{0})=\sqrt{\left\vert \frac{\beta }{\alpha }\right\vert } \pm%
\frac{h}{2\beta }\text{.}
\end{equation}
Therefore, the edge magnetization is enhanced at one edge and weakened at another edge due to the breaking of the inversion antisymmetry in the AFCE states under the external magnetic field.
In comparison, the inversion symmetry is preserved in the FMCE states under the external magnetic field, which means $M_{L}=M_{R}=M$. Similar
relation $M\cos(\kappa x_{0})=\sqrt{\left\vert \frac{\beta }{\alpha }%
\right\vert }+\frac{h}{2\beta }$ is subsequently obtained according to the Euler-Lagrange equation. The edge magnetization is enhanced at both edges.
Due to the different response of the edge magnetization in the AFCE and FMCE state, the FMCE is more energetically favorable under the external
magnetic field.
\section*{Appendix III. Disorder effects}
\begin{figure}[hbp]
\vspace{-0.0in} \hspace{-0.0in} \center
\includegraphics[width=0.5\columnwidth]{SI-disorders.eps} \vspace{-0.0in}
\caption{Free energy difference between the AFCE and FMCE state $\Delta F=F_{AFCE}-F_{FMCE}$ with disorders. The black dotted lines is the phase boundary of the AFCE and FMCE states ($\Delta F=0$). Red circles are for the moderate strength of disorders with $W=100 meV$. The blue circles are for the critical strength of the disorders, where the occasional AFCE-FMCE instability occurs. The solid, and hollow circles energetically favor the AFCE, and FMCE state, respectively. The solid lines (red/blue) are the averages for the respective disorders. The solid black lines are for $W=0$. $\delta=0.012$ and $U=2.0t$ and $t=2.6$ eV is adoped. (a) The $1$\emph{st} AFCE state with $N=4$; (b) The $1$\emph{st} FMCE state with $N=8$; (c) The $2$\emph{nd} AFCE state with $N=16$.}
\label{S5}
\end{figure}
We consider the non-magnetic disorders along the edges $H_{W}=\sum_{\sigma i\in edges}W_{i}n_{i\sigma}$, where $W_{i}$ is the strength of the disorder at the given site randomly distributed in the interval $[-W/2,W/2]$ with $W$ the strength of the disorders. We solve the Hamiltonian in the real-space with the periodic boundary condition $N_{y}=24$ along the infinite $y$-direction due to the broken translation symmetry along $y$-direction. In fact, the results presented here are insensitive to the large enough $N_{y}$. For simplicity, only the disorders at the edges with inversion symmetry are considered.
The influence of the edge disorders on the antiferromagnetic correlated edges (AFCE) and ferromagnetic correlated edges (FMCE) states is displayed in Fig.~\ref{S5} with $50$ times random disorders. At the doping $\delta=0.012$, the red circles with moderate strength of the edge disorders $W=100$ meV ($t=2.6$ eV \cite{Jung-PRB2009}) well locate in the respective phase for the $1$-\emph{st} AFCE ($N=4$), $1$-\emph{st} FMCE ($N=8$), and $2$-\emph{nd} AFCE state ($N=16$). Therefore, the magnetic phases are quite robust at least for the low-\emph{th} magnetic states. When the strength of the disorders is enhanced to a critical value, the magnetic phase may be unstable. For the $1$-\emph{st} AFCE state, the FMCE occasionally has lower energy for strong enough disorder $W_{c}=1100$ meV (blue hollow circles in Fig.~\ref{S5}(a)). This critical strength depends on the number of the zigzag chains $N$, it is about $300$ meV, $210$ meV, for the $1$-\emph{st} FMCE ($N=8$), and $2$-\emph{nd} AFCE ($N=16$) state, respectively. Interestingly, the average effect of the disorders seems to stabilize the AFCE states while weaken the FMCE states. This is probably due to that the scattering of the edge disorders will open a gap in the metallic FMCE state.
|
\section{Introduction}
\label{S:intro}
The existence of closed geodesics in a Riemannian manifold is one of the most fundamental questions in geometry that has been studied extensively since the time of Poincar\'{e} \cite{Poincare04}. The critical point theories developed by Morse and Lusternik-Schnirelmann have played an essential role in this problem in the early 20th century (see \cite{Klingenberg} for a detailed exposition up to 1978). Although there does not exist closed geodesics in $\mathbb{R}^n$, it is natural to look for geodesics contained in a bounded domain $\Omega \subset \mathbb{R}^n$ which meets $\partial \Omega$ orthogonally at its end points. These are called \emph{orthogonal geodesic chords} (see Definition \ref{D:OGC} for a precise definition). In \cite{Lusternik-Schnirelmann}, Lusternik and Schnirelmann proved the following celebrated result:
\begin{theorem}[Lusternik-Schnirelmann]
\label{T:LS}
Any bounded domain in $\mathbb{R}^n$ with smooth convex boundary contains at least $n$ distinct orthogonal geodesic chords.
\end{theorem}
Kuiper \cite{Kuiper64} showed that the same conclusion holds if the boundary is only $C^{1,1}$. For our convenience, we will assume that all the submanifolds and maps are $C^\infty$. Recall that the boundary of a domain $\Omega \subset \mathbb{R}^n$ is said to be (locally) \emph{convex} if the second fundamental form $A$ of $\partial \Omega$ with respect to the unit normal $\nu$ (pointing into $\Omega$) is positive semi-definite, i.e. for all $p \in \partial \Omega$, $u \in T_p \partial \Omega$, we have
\begin{equation}
\label{E:convex}
A(u,u):=\langle D_u u ,\nu \rangle \geq 0,
\end{equation}
where $D$ is the standard flat connection in $\mathbb{R}^n$. Notice that Theorem \ref{T:LS} gives an optimal lower bound as seen in the example of the convex region bounded by the ellipsoid given by
\[ \Omega:=\left\{(x_1,\cdots,x_n) \in \mathbb{R}^n \; : \; \sum_{i=1}^n \frac{x_i^2}{a_i^2} \leq 1 \right\} \]
where $a_1,\cdots,a_n$ are distinct positive real numbers.
In \cite{Bos63}, Bos generalized Lusternik-Schnirelmann's result to the setting of Riemannian (or even Finsler) manifolds.
\begin{theorem}[Bos]
\label{T:Bos}
A compact Riemannian manifold $(M^n,g)$ which is homeomorphic to the closed unit ball in $\mathbb{R}^n$ with locally convex boundary contains at least $n$ orthogonal geodesic chords.
\end{theorem}
Moreover, he showed that the convexity assumption cannot be dropped even in $\mathbb{R}^2$ (see Figure \ref{F:Bos}).
\begin{figure}[h]
\centering
\includegraphics[height=4cm]{Bos.png}
\caption{Bos' example of a non-convex domain $\Omega$ in $\mathbb{R}^2$ which does not have any orthogonal geodesic chord contained in $\Omega$.}
\label{F:Bos}
\end{figure}
Nonetheless, one can still ask for the existence of orthogonal geodesic chords, by allowing them to go \emph{outside} the domain. This problem was first introduced by Riede \cite{Riede68}, where he studied the variational calculus of the space $\Gamma$ consisting of piecewise smooth curves in a complete Riemannian manifold $(M^n,g)$ with end points lying on a compact submanifold $\Sigma^k \subset M$. In particular, he estimated the minimum number of critical points, which are orthogonal geodesic chords, in terms of certain topological invariant called the ``cup-length'' of the equivariant cohomology of $\Gamma$ with respect to the $\mathbb{Z}_2$-action reversing the orientation of a curve. In \cite{Hayashi82}, Hayashi computed the cup-length when $\Sigma$ is a compact submanifold in $\mathbb{R}^n$ and hence proved the following result.
\begin{theorem}[Riede-Hayashi]
\label{T:double-normals}
Any $k$-dimensional compact submanifold $\Sigma$ in $\mathbb{R}^n$ admits at least $k+1$ orthogonal geodesic chords.
\end{theorem}
Note that Theorem \ref{T:double-normals} generalizes Theorem \ref{T:LS} by taking $\Sigma$ to be the boundary of a bounded convex domain. However, we emphasize that if $\Sigma=\partial \Omega$ is the boundary of a non-convex domain $\Omega \subset \mathbb{R}^n$, then the orthogonal geodesic chords obtained in Theorem \ref{T:double-normals} are not necessarily contained in $\Omega$ (recall Figure \ref{F:Bos}).
The original proof of Theorem \ref{T:LS}, \ref{T:Bos} and \ref{T:double-normals} all used a discrete curve shortening process similar to the one introduced by Birkhoff \cite{Birkhoff17} in the study of existence of closed geodesics in Riemannian manifolds. A description of the process can be found in \cite{Gluck-Ziller83} (see also a modified version in \cite{Zhou16}). The curve shortening process, denoted by $\Psi$, take a piecewise smooth curve $c:[0,1] \to M$ with end points lying on $\Sigma$ to a piecewise geodesic curve $\Psi(c):[0,1] \to M$ which meets $\Sigma$ orthogonally at its end points. The most important properties of $\Psi$ are summarized below:
\begin{itemize}
\item[(1)] $\operatorname{Length}(\Psi(c)) \leq \operatorname{Length}(c)$ and equality holds if and only if $c$ is an orthogonal geodesic chord, in which case $\Psi(c)=c$.
\item[(2)] $\Psi(c)$ depends continuously on $c$, with respect to the $C^0$ topology.
\item[(3)] $c$ and $\Psi(c)$ are homotopic in $M$ relative to $\Sigma$, i.e. there exists a continuous family $c_t:[0,1] \to M$, $t \in [0,1]$, with end points on $\Sigma$ such that $c_0=c$ and $c_1=\Psi(c)$. Moreover, the family $c_t$ depends continuously on $c$.
\end{itemize}
The curve shortening process $\Psi$ involves subdividing the curves and connecting points on the curve by minimizing geodesic segments (additional care has to be taken at the end points). The construction depends on some fixed parameter (which depends on the geometry of $M$, $\Sigma$ and $\operatorname{Length}(c)$). However, it can be shown that for curves with uniformly bounded length, the parameters can be chosen uniformly to make (1) - (3) above hold. In fact (1) and (3) follows easily from the constructions, but (2) requires some convexity estimates (see \cite[Lemma 3.2]{Zhou16}). Using (1) - (3), it is not difficult to see that the sequence $\{\Psi^i(c)\}_{i=1}^\infty$ either converges to a point on $\Sigma$ or has a subsequence converging to an orthogonal geodesic chord. Theorem \ref{T:LS}, \ref{T:Bos} and \ref{T:double-normals} then follows from the abstract Lusternik-Schnirelmann theory applied to families of curves with end points on $\Sigma$ which represent a non-trivial homology class relative to point curves on $\Sigma$. Interested readers can refer to \cite{Gluck-Ziller83} or \cite{Giannoni-Majer97} for more details (for Theorem \ref{T:LS} there is a more elementary proof - see \cite{Kuiper64} for example).
In this paper, we introduce a new curve shortening process called the \emph{chord shortening flow} (see Definition \ref{D:CSF}), which evolves a geodesic chord according to the ``contact angle'' that the chord makes with $\Sigma$ at its end points. It is the negative gradient flow for the length functional on the space of chords. We study the fundamental properties including the short-time existence and uniqueness and long-time convergence of the flow when the ambient space is $\mathbb{R}^n$. Note that the flow still makes sense in certain Riemannian manifolds but for simplicity we postpone the details to another forthcoming paper. The chord shortening flow, as a negative gradient flow, clearly satisfies all the properties (1) - (3) above; hence provide the most natural curve shortening process required in the proof of Theorem \ref{T:LS} and \ref{T:double-normals} (but not Theorem \ref{T:Bos} in its full generality).
\begin{remark}
We would like to mention that Lusternik and Schnirelmann used the same ideas to prove the \emph{Theorem of Three Geodesics} which asserts that any Riemannian sphere $(S^2,g)$ contains at least three \emph{geometrically distinct} closed embedded geodesics. Unfortunately, the original proof by Lusternik-Schnirelmann \cite{Lusternik-Schnirelmann} contains a serious gap and various attempts have been made to fix it (see \cite{Taimanov92}). The fundamental issue there is \emph{multiplicity}, that one of the geodesics obtained may just be a multiple cover of another geodesic. It is extremely technical (and many false proofs were given) to rule out this situation by modifying the method of Lusternik-Schnirelmann. In \cite{Grayson89}, Grayson gave a rigorous proof of the Theorem of Three Geodesics by a careful analysis of the curve shortening flow on Riemannian surfaces. He proved that under the curve shortening flow, any embedded curve remains embedded and would either converge to a point in finite time or an embedded closed geodesic as time goes to infinity. As a curve which is initially embedded stays embedded throughout the flow, this prevents the multiplicity problem encountered by Lusternik-Schnirelmann's approach using a discrete curve shortening process of Birkhoff \cite{Birkhoff17}. On the other hand, the situation in Theorem \ref{T:LS} and \ref{T:double-normals} are simpler as multiplicity cannot occur (see \cite[Remark 3.2]{Giannoni-Majer97}).
\end{remark}
We show that the convergence behavior for the chord shortening flow is similar to that for the curve shortening flow on a closed Riemannian surface \cite{Grayson89}. In particular, we prove that under the chord shortening flow, any chord would either converge to a point in finite time or to an orthogonal geodesic chord as time goes to infinity. Unlike \cite{Grayson89}, this dichotomy holds in any dimension and codimension, in contrast with the curve shortening flow where an embedded curve may develop self-intersections or singularities after some time when codimension is greater than one \cite{Altschuler91}. In the special case that $\Sigma=\partial \Omega$ where $\Omega \subset \mathbb{R}^2$ is a compact convex planar domain, we give a sufficient condition for an initial chord to converge to a point in finite time. In fact, any ``\emph{convex}'' chord in $\Omega$ which is not an orthogonal geodesic chord would converge to a point on $\partial \Omega$ in finite time. This can be compared to the famous result of Huisken \cite{Huisken84} which asserts that any compact embedded convex hypersurface in $\mathbb{R}^n$ converges to a point in finite time under the mean curvature flow.
The chord shortening flow is also of independenet interest from the analytic point of view. Since any chord in $\mathbb{R}^n$ is determined uniquely by its end points, we can regard the chord shortening flow as an evolution equation for the two end points lying on $\Sigma$. As a result, the flow is a \emph{non-local} evolution of a pair of points on $\Sigma$ as it depends on the chord joining them. In fact, the chord shortening flow can be regarded as the heat equation for the half-Laplacian (or the \emph{Dirichlet-to-Neumann map}).
The organization of this paper is as follows. In Section 2, we introduce the chord shortening flow, give a few examples, and prove the short time existence and uniqueness of the flow. In Section 3, we derive the evolution equations for some geometric quantities under the chord shortening flow. In Section 4, we prove the long time existence to the flow provided that it does not shrink the chord to a point in finite time. In Section 5, we prove that an initial convex chord inside a compact convex domain in $\mathbb{R}^2$ would shrink to a point in finite time under the flow, provided that the initial chord is not an orthogonal geodesic chord.
\vspace{.3cm}
\noindent \textbf{Acknowledgement.} The author would like to express his gratitude to Prof. Richard Schoen for his interest in this work. He also want to thank Mario Micallef and Luen-Fai Tam for helpful comments and discussions. These work are partially supported by a research grant from the Research Grants Council of the Hong Kong Special Administrative Region, China [Project No.: CUHK 14323516] and CUHK Direct Grant [Project Code: 3132705].
\vspace{.3cm}
\noindent \textit{Notations.} Throughout this paper, we will denote $I:=[0,1]$ with $\partial I=\{0,1\}$. The Euclidean space $\mathbb{R}^n$ is always equipped with the standard inner product $\langle \cdot, \cdot \rangle$ and norm $|\cdot |$. For any subset $S \subset \mathbb{R}^n$, we use $d(\cdot, S)$ to denote the distance function from $S$.
\section{Chord Shortening Flow}
\label{S:CSF}
In this section, we introduce a new geometric flow called \emph{chord shortening flow}. This flow has some similarities with the classical curve shortening flow.
The main result in this section is the short-time existence and uniqueness theorem for the chord shortening flow (Theorem \ref{T:short-time-existence}). We also study some basic examples of such a flow.
Let $\Sigma$ be a $k$-dimensional smooth submanifold \footnote{In fact all the following discussions make sense for \emph{immersed} submanifolds. However, for simplicity, we will assume that all submanifolds are \emph{embedded}.} in $\mathbb{R}^n$. Note that $\Sigma$ can be disconnected in general. For any two points $p,q \in \Sigma$, we can consider the extrinsic chord distance between them in $\mathbb{R}^n$.
\begin{definition}
\label{D:chord-distance}
The \emph{chord distance function} $d: \Sigma \times \Sigma \to \mathbb{R}_{\geq 0}$ is defined to be
\[ d(p,q):=\textrm{dist}_{\mathbb{R}^n} (p,q)=|p-q|. \]
\end{definition}
Since any two distinct points in $\mathbb{R}^n$ are connected by a unique line segment realizing their distance, the chord distance function $d$ is smooth away from the diagonal $\{(p,p) \in \Sigma \times \Sigma \, : \, p \in \Sigma\}$.
\begin{definition}
\label{D:Cpq}
For any two distinct points $p,q$ on $\Sigma$, we will use $C_{p,q}$ to denote the unique oriented chord from $p$ to $q$. The outward unit conormal, denoted by $\eta$, is the unit vector at $\partial C_{p,q}$ tangent to $C_{p,q}$ pointing out of $C_{p,q}$. Note that $\eta(p)=-\eta(q)$. (see Figure \ref{F:Cpq})
\end{definition}
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{Cpq.png}
\caption{A chord $C_{p,q}$ joining $p$ and $q$, the outward unit conormals $\eta$ and their (negative) tangential components along $\Sigma=\partial \Omega$}
\label{F:Cpq}
\end{figure}
Let $C(t)=C_{p_t,q_t}$ be a smooth family of chords with distinct end points $p_t,q_t \in \Sigma$. If $\ell(t)=d(p_t,q_t)$ is the length of the chord $C(t)$, the first variation formula for arc length (see for example \cite[(1.5)]{Cheeger-Ebin}) implies that
\begin{equation}
\label{E:1st-var-a}
\frac{d \ell}{dt} =\langle \frac{dp_t}{dt}, \eta(p_t) \rangle + \langle \frac{dq_t}{dt}, \eta(q_t) \rangle.
\end{equation}
Note that the interior integral term vanishes as $C(t)$ is a geodesic for every $t$. Since $p_t$ and $q_t$ lies on $\Sigma$ for all $t$, both $dp_t/dt$ and $dq_t/dt$ are tangential to $\Sigma$. Therefore, we can express (\ref{E:1st-var-a}) as
\begin{equation}
\label{E:1st-var-b}
\frac{d \ell}{dt} = \langle \frac{dp_t}{dt}, \eta^T(p_t) \rangle + \langle \frac{dq_t}{dt}, \eta^T(q_t) \rangle
\end{equation}
where $(\cdot)^T$ denotes the tangential component of a vector relative to $\Sigma$. More precisely, if $\pi_x:\mathbb{R}^n \to T_x\Sigma$ is the orthogonal projection onto the tangent space $T_x \Sigma \subset \mathbb{R}^n$, then $v^T=\pi_x(v)$ for any vector $v \in T_x \mathbb{R}^n \cong \mathbb{R}^n$.
It is natural to consider the (negative) gradient flow to the chord length functional, which leads to the following definition.
\begin{definition}[Chord Shortening Flow]
\label{D:CSF}
A smooth family of curves
\[ C(u,t):I \times [0,T) \to \mathbb{R}^n\]
is a solution to the \emph{chord shortening flow} (relative to $\Sigma$) if for all $t \in [0,T)$, we have
\begin{itemize}
\item[(a)] $p_t:=C(0,t)$ and $q_t:=C(1,t)$ lies on $\Sigma$,
\item[(b)] $C(t):=C(\cdot,t):I \to \mathbb{R}^n$ is a constant speed parametrization of $C_{p_t,q_t}$,
\item[(c)]
\[ \frac{\partial C}{\partial t} (0,t) = - \eta^T(C(0,t)) \qquad \text{and} \qquad \frac{\partial C}{\partial t} (1,t) = - \eta^T(C(1,t)).\]
\end{itemize}
\end{definition}
Let us begin with some basic examples of the chord shortening flow as defined in Definition \ref{D:CSF}.
\begin{example}
\label{ex:flat}
Let $\Sigma$ be an affine $k$-dimensional subspace in $\mathbb{R}^n$. The chord shortening flow with respect to $\Sigma$ will contract any initial chord $C(0)=C_{p,q}$ to a point in finite time. The end points would move towards each other with unit speed along the chord $C(0)$ until they meet at the mid-point of $C(0)$ at the ``blow-up'' time $T=d(p,q)/2$.
\end{example}
\begin{example}
\label{ex:annulus}
Let $\Sigma$ be a union of two disjoint circles in $\mathbb{R}^2$. We will see (from Theorem \ref{T:convergence}) that any chord joining two distinct connected components of $\Sigma$ would evolve under the chord shortening flow to a limit chord $C_\infty$ orthogonal to $\Sigma$ as $t \to \infty$. The same phenomenon holds for any $\Sigma \subset \mathbb{R}^n$ which is disconnected.
\end{example}
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{annulus.png}
\caption{A limit chord $C_\infty$ which meets $\partial \Omega$ orthogonally but not lying inside $\Omega$.}
\label{F:annulus}
\end{figure}
\begin{example}
Let $\Sigma$ be the ellipse $\{(x,y) \in \mathbb{R}^2 \, : \, x^2+4y^2=1\}$ in $\mathbb{R}^2$. By symmetry it is not difficult to see that for any initial chord passing through the origin (with the exception of the major axis), it would evolve under the chord shortening flow to the minor axis of the ellipse, which is a chord orthogonal to $\Sigma$ and contained inside the region enclosed by the ellipse. See Figure \ref{F:ellipse}. This example shows that the number of distinct orthogonal chords guaranteed by the Lyusternik-Schnirelmann Theorem is optimal. If we start with an initial chord that lies completely on one side of the major or minor axis, then the chord will shrink to a point in finite time (by Theorem \ref{T:contraction}).
\end{example}
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{ellipse.png}
\caption{Any initial chord $C_0$ through the origin (other than the major axis) would converge under the chord shortening flow to the minor axis.}
\label{F:ellipse}
\end{figure}
We end this section with a fundamental result on the short time existence and uniqueness for the chord shortening flow.
\begin{proposition}[Short-time existence and uniqueness]
\label{T:short-time-existence}
For any initial chord $C_0:I \to \mathbb{R}^n$ with $C_0(\partial I) \subset \Sigma$, there exists an $\epsilon>0$ and a smooth solution $C(u,t):I \times [0,\epsilon) \to \mathbb{R}^n$ to the chord shortening flow relative to $\Sigma$ as in Definition \ref{D:CSF} with initial condition $C(\cdot,0)=C_0$. Moreover, the solution is unique.
\end{proposition}
\begin{proof}
Note that for any given $p \neq q \in \Sigma$, the outward unit conormal $\eta$ at the end points $p,q$ of the chord $C_{p,q}$ is given by
\[ \eta(p)=\frac{p-q}{|p-q|} =-\eta(q). \]
Therefore, Definition \ref{D:CSF} (c) is equivalent to the following system of nonlinear system of first order ODEs:
\begin{equation}
\label{E:CSF'}
\left\{ \begin{array}{l}
\frac{dp}{dt}=- \frac{\pi_p(p-q)}{|p-q|} \\
\frac{dq}{dt}=- \frac{\pi_q(q-p)}{|q-p|}
\end{array} \right.
\end{equation}
where $\pi_x:\mathbb{R}^n \to \mathbb{R}^n$ is the orthogonal projection onto $T_x\Sigma$ (which depends smoothly on $x$). Since the right hand side of (\ref{E:CSF'}) is a Lipschitz function in $p$ and $q$ as long as $|p-q|$ is bounded away from $0$. Therefore, the existence and uniqueness to the initial value problem follows from the fundamental local existence and uniqueness theorem for first order ODE systems (see for example \cite[Theorem 2.1]{Taylor}). Hence, (\ref{E:CSF'}) is uniquely solvable on some interval $t \in [0,\epsilon)$ for any initial data $p(0)=p_0$ and $q(0)=q_0$ such that $p_0 \neq q_0 \in \Sigma$. Finally we get a solution $C(u,t):I \times [0,\epsilon) \to \mathbb{R}^n$ to the chord shortening flow by defining $C(\cdot,t):I \to \mathbb{R}^n$ to be the constant speed parametrization of the chord $C_{p_t,q_t}$.
\end{proof}
\section{Evolution equations}
\label{S:evolution}
In this section, we derive the evolution of some geometric quantities under the chord shortening flow relative to any $k$-dimensional submanifold $\Sigma$ in $\mathbb{R}^n$.
\begin{definition}
Let $C:I=[0,1] \to \mathbb{R}^n$ be a chord joining $p$ to $q$. For any (vector-valued) function $f:\partial I =\{ 0,1\} \to \mathbb{R}^m$, we define the \emph{$L^2$-norm} $\|f\|_{L^2}$ and the \emph{sum} $\overline{f}$ of $f$ to be\begin{equation}
\label{E:norm}
\|f\|_{L^2}:=( |f(0)|^2+|f(1)|^2)^{\frac{1}{2}} \qquad \text{and} \qquad \overline{f}:=f(0)+f(1).
\end{equation}
Also, we define the $\frac{1}{2}$-Laplacian of $f$ relative to the chord $C$ to be the vector-valued function $\Delta^{\frac{1}{2}} f:\partial I=\{0,1\} \to \mathbb{R}^m$ defined by
\begin{equation}
\label{E:DN}
(\Delta^{\frac{1}{2}} f)(0)= \frac{f(0)-f(1)}{\ell} =- (\Delta^{\frac{1}{2}} f)(1),
\end{equation}
where $\ell=|p-q|$ is the length of the chord $C$.
\end{definition}
\begin{lemma}
\label{L:DN}
Given any $f:\partial I \to \mathbb{R}^m$, we have $\overline{\Delta^{\frac{1}{2}} f}=0$ and $\overline{\langle f, \Delta^{\frac{1}{2}} f \rangle}=\frac{\ell}{2} \|\Delta^{\frac{1}{2}} f\|^2_{L^2} \leq \frac{2}{\ell} \|f\|^2_{L^2}$.
\end{lemma}
\begin{proof}
It follows directly from (\ref{E:norm}) and (\ref{E:DN}).
\end{proof}
\begin{definition}
Let $C=C_{p,q}:I \to \mathbb{R}^n$ be a chord joining two distinct points $p,q$ on $\Sigma$. We define the tangential outward conormal $\eta^T:\partial I=\{0,1\} \to \mathbb{R}^n$ to be the tangential component (relative to $\Sigma$) of the outward unit conormal of $C$, i.e. (recall (\ref{E:1st-var-b}) and Definition \ref{D:Cpq})
\begin{equation}
\label{E:eta-T}
\eta^T(u)=\pi_{C(u)} \eta \qquad \text{for $u=0,1$}.
\end{equation}
\end{definition}
\begin{lemma}[Evolution of chord length]
\label{L:length-evolution}
Suppose $C(u,t):I \times [0,T) \to \mathbb{R}^n$ is a solution to the chord shortening flow relative to $\Sigma$ as in Definition \ref{D:CSF}. If we denote the length of the chord $C(t)$ at time $t$ by
\[ \ell(t):=d(C(0,t),C(1,t)),\]
then $\ell$ is a non-increasing function of $t$ and (recall (\ref{E:norm}) and (\ref{E:eta-T}))
\begin{equation}
\label{E:ell-evolution}
\frac{d \ell}{dt} = - \|\eta^T\|_{L^2}^2 \leq 0 .
\end{equation}
\end{lemma}
\begin{proof}
It follows directly from the first variation formula (\ref{E:1st-var-b}).
\end{proof}
\begin{theorem}
\label{T:eta-T-evolution}
Suppose $C(u,t):I \times [0,T) \to \mathbb{R}^n$ is a solution to the chord shortening flow relative to $\Sigma$ as in Definition \ref{D:CSF}. Then the tangential outward conormal $\eta^T$ of the chord $C(t)$ satisfies the following evolution equation:
\begin{eqnarray}
\label{E:eta-T-evolution}
\frac{\partial}{\partial t} \eta^T &=& - \Delta^{\frac{1}{2}} \eta^T + \frac{1}{\ell} \|\eta^T\|^2_{L^2} \eta^T - \sum_{i=1}^k \langle A(\eta^T,e_i),\eta^N \rangle e_i \\
&& \phantom{aaaaaaaaa} - \frac{1}{\ell} (\overline{\eta^T}-\eta^T)^N - A(\eta^T,\eta^T), \nonumber
\end{eqnarray}
where $\{e_i\}_{i=1}^k$ is an orthonormal basis of $T\Sigma$ at the end points of $C(t)$. Here, $(\cdot)^N$ denotes the normal component of a vector relative to $\Sigma$ and $A:T\Sigma \times T\Sigma \to N\Sigma$ is the second fundamental form of $\Sigma$ defined by $A(u,v):=(D_u v)^N$.
\end{theorem}
\begin{proof}
Let $C(u,t):I \times [0,T) \to \mathbb{R}^n$ be a solution to the chord shortening flow relative to $\Sigma$. Since $C(t)=C( \cdot, t):I \to \mathbb{R}^n$ is a family of chords which are parametrized proportional to arc length, $\frac{\partial}{\partial t}$ is a Jacobi field (not necessarily normal) along each chord which can be explicitly expressed as
\[ \frac{\partial }{\partial t} = -(1-u) \, \eta^T(0) - u \, \eta^T(1),\]
where $\eta$ is the outward unit conormal for $C(t)$. Since $[\frac{\partial}{\partial u}, \frac{\partial }{\partial t}]=0$, we have
\begin{equation}
\label{E:commutator}
D_{\frac{\partial}{\partial t}} \frac{\partial }{\partial u}=D_{\frac{\partial}{\partial u}} \frac{\partial }{\partial t} = \eta^T(0)-\eta^T(1).
\end{equation}
Moreover, as $C(t)$ is parametrized with constant speed, we have $\| \frac{\partial}{\partial u}\|= \ell$, thus
\[ -\eta(0)= \frac{1}{\ell} \left.\frac{\partial }{\partial u} \right|_{u=0} \quad \text{and} \quad \eta(1)=\frac{1}{\ell} \left. \frac{\partial }{\partial u} \right|_{u=1}.\]
Fix $u=0$. Let $p=C(0,t) \in \Sigma$ and $\{e_1,\cdots, e_k\}$ be an orthonormal basis of $T_p \Sigma$ such that $(D_{e_i} e_j(p))^T=0$ for $i,j=1,\cdots,k$. Therefore, we have
\begin{equation}
\label{E:A}
D_{\frac{\partial}{\partial t}} e_i = -A(\eta^T,e_i).
\end{equation}
Using Lemma \ref{L:length-evolution}, (\ref{E:commutator}) and (\ref{E:A}), we have:
\begin{eqnarray*}
\frac{\partial \eta^T}{\partial t} &=& \frac{\partial}{\partial t} \left( -\frac{1}{\ell} \right) \sum_{i=1}^k \langle \frac{\partial }{\partial u}, e_i \rangle e_i- \frac{1}{\ell} \sum_{i=1}^k \frac{\partial }{\partial t} \left( \langle \frac{\partial }{\partial u},e_i \rangle e_i \right)\\
&=& \frac{1}{\ell} \|\eta^T\|^2_{L^2} \eta^T - \frac{1}{\ell} \sum_{i=1}^k \left( \langle D_{\frac{\partial}{\partial u}} \frac{\partial}{\partial t}, e_i \rangle e_i+ \langle \frac{\partial}{\partial u}, D_{\frac{\partial}{\partial t}} e_i \rangle e_i + \langle \frac{\partial }{\partial u},e_i \rangle D_{\frac{\partial}{\partial t}} e_i \right) \\
&=& \frac{1}{\ell} \|\eta^T\|^2_{L^2} \eta^T - \frac{\eta^T}{\ell} - A(\eta^T,\eta^T) - \frac{1}{\ell} \sum_{i=1}^k \left( \langle -\eta^T(1),e_i \rangle e_i + \ell \langle \eta^N, A(\eta^T,e_i) \rangle e_i \right) \\
&=& - \Delta^{\frac{1}{2}} \eta^T + \frac{1}{\ell} \|\eta^T\|^2_{L^2} \eta^T - \sum_{i=1}^k \langle A(\eta^T,e_i),\eta^N \rangle e_i - \frac{1}{\ell} (\overline{\eta^T}-\eta^T)^N - A(\eta^T,\eta^T).
\end{eqnarray*}
A similar calculation yields (\ref{E:eta-T-evolution}) at $u=1$. This proves the proposition.
\end{proof}
\begin{corollary}
Under the same assumptions as in Theorem \ref{T:eta-T-evolution}, we have
\begin{equation}
\label{E:eta-norm-evolution}
\frac{1}{2} \frac{d}{dt} \|\eta^T\|^2_{L^2}= - \frac{\ell}{2} \| \Delta^{\frac{1}{2}} \eta^T \|^2_{L^2} + \frac{1}{\ell} \|\eta^T\|^4_{L^2} -\overline{\langle A(\eta^T,\eta^T), \eta \rangle}.
\end{equation}
\end{corollary}
\begin{proof}
Using (\ref{E:eta-T-evolution}) and Lemma \ref{L:DN}, noting that the last two terms of (\ref{E:eta-T-evolution}) are normal to $\Sigma$, we have
\[
\frac{1}{2} \frac{d}{dt} \|\eta^T\|^2_{L^2} = \overline{\langle \eta^T, \frac{\partial \eta^T}{\partial t} \rangle}
= - \frac{\ell}{2} \| \Delta^{\frac{1}{2}} \eta^T \|^2_{L^2} + \frac{1}{\ell} \|\eta^T\|^4_{L^2} -\overline{\langle A(\eta^T,\eta^T), \eta^N \rangle}.
\]
\end{proof}
\begin{example}
\label{ex:flat2}
In the case of Example \ref{ex:flat}, we have $\eta^T(0)=-\eta^T(1)$ equals to a constant unit vector independent of $t$ and hence both sides are identically zero in (\ref{E:eta-T-evolution}) and (\ref{E:eta-norm-evolution}).
\end{example}
\begin{example}
\label{ex:strip}
Consider the vertical strip $\Omega:=\{(x,y) \in \mathbb{R}^2 : 0 \leq x \leq 1\}$ with boundary $\Sigma=\partial \Omega$ consists of two parallel vertical lines. Let $p_0=(0,-h/2)$ and $q_0=(1,h/2)$ for some $h>0$. It is easy to check that the solution to the chord shortening flow with initial chord $C_{p_0,q_0}$ is given by $p_t=(0,-h(t)/2)$, $q_t=(1,h(t)/2)$ where $h(t)$ is the unique solution to the ODE
\[ h'(t)=-\frac{2h(t)}{\sqrt{1+h^2(t)}} \]
with initial condition $h(0)=h$. From this we can see that the solution $h(t)$ exists for all $t \geq 0$. Moreover, $-h'(t) \leq 2h(t)$ implies $h(t) \leq h e^{-2t}$ and thus $h(t) \to 0$ exponentially as $t \to +\infty$. Therefore, the chord converges to a chord meeting $\partial \Omega$ orthogonally. In this case, we have
\[ -\eta^T(0)= \frac{1}{\sqrt{1+h^2(t)}} (0,h(t)) = \eta^T(1), \]
which satisfies the evolution equation (\ref{E:eta-T-evolution}) and $\eta^T \to 0$ as $t \to +\infty$. See Figure \ref{F:strip}.
\end{example}
\begin{figure}[h]
\centering
\includegraphics[height=6cm]{strip.png}
\caption{A chord converging to a limit chord orthogonal to $\partial \Omega$.}
\label{F:strip}
\end{figure}
\section{Long time existence}
\label{S:convergence}
In this section, we prove our main convergence result which says that the only two possible convergence scenarios are given in Example \ref{ex:flat2} and \ref{ex:strip}. One should compare this convergence result with a similar result of Grayson \cite[Theorem 0.1]{Grayson89} for curve shortening flow on surfaces. For simplicity, we assume that $\Sigma$ is compact. However, the same result holds for non-compact $\Sigma$ which satisfies some convexity condition at infinity as in \cite{Grayson89}.
\begin{theorem}[Long time convergence]
\label{T:convergence}
Let $\Sigma \subset \mathbb{R}^n$ be a compact $k$-dimensional smooth submanifold without boundary. Suppose $C(0):I \to \mathbb{R}^n$ is a chord with distinct end points on $\Sigma$. Then there exists a maximally defined smooth family of chords $C(t):I \to \mathbb{R}^n$ for $t \in [0,T)$ with distinct end points on $\Sigma$, and $C(t)=C(\cdot,t)$ where $C(u,t):I \times [0,T) \to \mathbb{R}^n$ is the unique solution to the chord shortening flow (relative to $\Sigma$) as in Definition \ref{D:CSF}.
Moreover, if $T < + \infty$, then $C(t)$ converges to a point on $\Sigma$ as $t \to T$. If $T$ is infinite, then $C(t)$ converges to an orthogonal geodesic chord with end points on $\Sigma$ as $t \to \infty$.
\end{theorem}
By the short time existence and uniqueness theorem (Theorem \ref{T:short-time-existence}), the chord shortening flow continues to exist and is unique as long as $\ell >0$. Therefore, $C(t)$ is uniquely defined for $t \in [0,T)$ where either $T < +\infty$ or $T=+\infty$.
\begin{lemma}
\label{L:ell-limit}
Let $C(t)$, $t \in [0,T)$, be a maximally defined chord shortening flow. Then one of the following holds:
\begin{itemize}
\item[(a)] $T < +\infty$ and $C(t)$ shrinks to a point on $\Sigma$ as $t \to T$;
\item[(b)] $T=+\infty$ and $\ell(t) \to \ell_\infty >0$ as $t \to +\infty$.
\end{itemize}
\end{lemma}
\begin{proof}
As $\ell(t)$ is a non-increasing function of $t$ by Lemma \ref{L:length-evolution}, it either converges to $0$ or to some positive number $\ell_\infty>0$ as $t \to T$. By short time existence (Theorem \ref{T:short-time-existence}), it cannot converge to $\ell_\infty>0$ in finite time. So when $T<+\infty$, $C(t)$ must converge to a point on $\Sigma$ by compactness of $\Sigma$. It remains to show that $\ell(t)$ cannot converge to $0$ if $T=+\infty$. We will prove this by a contradiction argument.
Suppose, on the contrary, that $T=+\infty$ and $\ell(t) \to 0$ as $t \to +\infty$. Since $\Sigma$ is compact, there exists some constant $\epsilon_0>0$ such that for any two points $p,q \in \Sigma$ with $d(p,q)<\epsilon_0$, the chord $C_{p,q}$ joining them has $\|\eta^T\|^2_{L^2}$ be bounded from below by a universal positive constant (see, for example, \cite[Lemma 5.2]{Colding-Minicozzi}). By Lemma \ref{L:length-evolution}, $\ell(t)$ must decrease to zero in finite time, which is a contradiction.
\end{proof}
Next, we claim that if the flow exists for all time (i.e. $T=+\infty$), then it must converge to an orthogonal geodesic chord to $\Sigma$ as $t \to \infty$. Since $|\eta^T| \leq \|\eta^T\|_{L^2}$, it suffices to prove the following lemma. Theorem \ref{T:convergence} clearly follows from Lemma \ref{L:ell-limit} and \ref{L:kto0}.
\begin{lemma}
\label{L:kto0}
Under the same assumption as Lemma \ref{L:ell-limit} and suppose $T=+\infty$, then $\|\eta^T\|_{L^2} \to 0$ as $t \to +\infty$.
\end{lemma}
\begin{proof}
Write $\ell_t=\ell(t)$ for $t \in [0,+\infty]$. By Lemma \ref{L:length-evolution} and \ref{L:ell-limit}, we have
\begin{equation}
\label{E:ell-monotone}
\ell_0 \geq \ell_t \geq \ell_\infty>0 \quad \text{ for all $t$}.
\end{equation}
Moreover, integrating the inequality in Lemma \ref{L:length-evolution} we obtain
\[ \ell_t- \ell_\infty = \int_{t}^\infty \|\eta^T\|^2_{L^2} \; d\tau \geq 0. \]
As a result, $\int_t^\infty \|\eta^T\|^2_{L^2} \; d\tau \to 0$ as $t \to \infty$. In other words, $\|\eta^T\|^2_{L^2}$ is $L^2$-integrable on $t \in [0,+\infty)$. If we can control the time derivative of $\|\eta^T\|_{L^2}^2$, then we can conclude that $\|\eta^T\|_{L^2} \to 0$ as $t \to \infty$. Using (\ref{E:eta-norm-evolution}), (\ref{E:ell-monotone}), Lemma \ref{L:DN} and $\|\eta^T \|^2_{L^2} \leq 2$, we have the following differential inequality
\begin{equation}
\label{E:norm-ineq}
\frac{1}{2}\frac{d}{dt} \|\eta^T\|^2_{L^2} \leq \left( C+\frac{4}{\ell_\infty} \right) \|\eta^T\|^2_{L^2}
\end{equation}
where $C=\sup_{\Sigma} |A| > 0$ is a constant depending only on the compact submanifold $\Sigma$. We now combine (\ref{E:norm-ineq}) with the fact that $\int_t^\infty \|\eta^T\|^2_{L^2} \; d\tau \to 0$ as $t \to \infty$ to conclude that $\|\eta^T\|^2_{L^2} \to 0$ as $t \to \infty$.
To simplify notation, let $f(t):=\|\eta^T\|^2_{L^2}$ and $c:=C+\frac{4}{\ell_\infty}$. Then $\int_t^\infty f \to 0$ as $t \to \infty$ and $f' \leq cf$. We argue that $f(t) \to 0$ as $t \to \infty$. Suppose not, then there exists an increasing sequence $t_n \to +\infty$ such that
\begin{equation}
\label{E:conv}
f(t_n) > \frac{1}{n} \qquad \text{and} \qquad \int_{t_n/2}^\infty f \leq \frac{1}{n^3}.
\end{equation}
We claim that there exists $t_n^* \in (t_n -\frac{1}{n},t_n+\frac{1}{n})$ such that $f(t_n^*) \leq 1/n^2$. If not, then by (\ref{E:conv})
\[ \frac{2}{n^3} \leq \int_{t_n-\frac{1}{n}}^{t_n+\frac{1}{n}} f \leq \int_{t_n/2}^\infty f \leq \frac{1}{n^3}, \]
which is a contradiction. Using that $f' \leq c f$, we see that by (\ref{E:conv})
\[ \frac{1}{n} < f(t_n) \leq f(t_n^*) e^{\frac{c}{n}} \leq \frac{1}{n^2} e^{\frac{c}{n}}. \]
As a result, there is a contradiction when $n$ is sufficiently large. We have thus proved that $f(t) \to 0$ as $t \to \infty$, as claimed.
\end{proof}
\section{Existence of Orthogonal Geodesic Chords}
\label{S:applications}
In this section, we give several geometric applications of the chord shortening flow concerning the existence of multiple orthogonal geodesic chords. We first give the precise definition.
\begin{definition}
\label{D:OGC}
Let $\Sigma \subset \mathbb{R}^n$ be a smooth $k$-dimensional submanifold without boundary. An \emph{orthogonal geodesic chord for $\Sigma$} is a geodesic $c:[0,1] \to \mathbb{R}^n$ with endpoint $c(0)$ and $c(1)$ lying on $\Sigma$ such that $c'(0)$ and $c'(1)$ are normal to $\Sigma$ at $c(0)$ and $c(1)$ respectively.
\end{definition}
An \emph{orthogonal geodesic chord} is also called a \emph{free boundary geodesic} \cite{Zhou16} or a \emph{double normal} \cite{Kuiper64} in the literature. Note that in case $\Sigma \subset \mathbb{R}^n$ is an embedded hypersurface which bounds a domain $\Omega$ in $\mathbb{R}^n$, our definition of orthogonal geodesic chords does not require the chord to be contained inside $\overline{\Omega}$ as for example in \cite{Giambo-Giannoni-Piccione14}. The problem of the existence of multiple orthogonal geodesic chord for submanifolds in $\mathbb{R}^n$ was first treated by Riede \cite{Riede68} as follows. Let $\mathcal{C}_\Sigma$ be the space of all piecewise smooth curve $c:[0,1] \to \mathbb{R}^n$ with end points on $\Sigma$, endowed with the compact open topology. There exists a $\mathbb{Z}_2$-action on $\mathcal{C}_\Sigma$ by $c(t) \mapsto c(1-t)$ whose fixed point set is denoted by $\Delta'$. Denote by $H^{\mathbb{Z}_2}_*(\mathcal{C}_\Sigma,\Delta')$ and $H^*_{\mathbb{Z}_2}(\mathcal{C}_\Sigma)$ the $\mathbb{Z}_2$-equivariant homology groups (relative to $\Delta'$) and cohomology groups respectively. The following result is taken from \cite[Satz (5.5)]{Riede68}.
\begin{lemma}
\label{L:Riede}
If there exists $\beta \in H^{\mathbb{Z}_2}_*(\mathcal{C}_\Sigma,\Delta')$ and $\alpha_1,\cdots,\alpha_s \in H^*_{\mathbb{Z}_2}(\mathcal{C}_\Sigma)$ (not necessarily distinct) with deg $\alpha_i>0$ for all $i$ such that $(\alpha_1 \cup \cdots \cup \alpha_s) \cap \beta \neq 0$, then there exists at least $s+1$ orthogonal geodesic chords for $\Sigma$.
\end{lemma}
The proof of Lemma \ref{L:Riede} involves a discrete curve shortening process $\Psi$ on $\mathcal{C}_\Sigma$ which satisfies properties (1) - (3) as described in the introduction. Since any curve $c \in \mathcal{C}_\Sigma$ can be continuously deformed into the unique chord joining the same end points, we can restrict $\mathcal{C}_\Sigma$ to the subset $\mathcal{C}^0_\Sigma$ consisting of all the chords with end points on $\Sigma$. The chord shortening flow is then a curve shortening process on $\mathcal{C}^0_\Sigma$ which satisfies all the required properties. Moreover, the space of chords $\mathcal{C}^0_\Sigma$ can also be described as the orbit space of $\Sigma \times \Sigma$ under the $\mathbb{Z}_2$-action $(p,q) \mapsto (q,p)$. As before, if we let $\Delta \subset \Sigma \times \Sigma$ be the fixed point set of the $\mathbb{Z}_2$-action, and $H^{\mathbb{Z}_2}_*(\Sigma \times \Sigma, \Delta)$, $H^*_{\mathbb{Z}_2}(\Sigma \times \Sigma)$ be the $\mathbb{Z}_2$-equivariant homology and cohomology respectively, we have by naturality
\begin{equation}
\label{E:isomorphism}
H^{\mathbb{Z}_2}_*(\Sigma \times \Sigma,\Delta) \cong H^{\mathbb{Z}_2}_*(\mathcal{C}_\Sigma,\Delta') \qquad \text{and} \qquad H^*_{\mathbb{Z}_2}(\Sigma \times \Sigma) \cong H^*_{\mathbb{Z}_2}(\mathcal{C}_\Sigma).
\end{equation}
In \cite{Hayashi82}, Hayashi studied the equivariant (co)homology of $\Sigma \times \Sigma$ and obtained the following result in \cite[Theorem 2]{Hayashi82}.
\begin{lemma}
\label{L:Hayashi}
There exists $\beta \in H^{\mathbb{Z}_2}_{2k}(\Sigma \times \Sigma,\Delta)$ and $\alpha \in H^1_{\mathbb{Z}_2}(\Sigma \times \Sigma)$ such that $\alpha^k \cap \beta \neq 0$ in $H^{\mathbb{Z}_2}_k(\Sigma \times \Sigma, \Delta)$, where $\alpha^k=\alpha \cup \cdots \cup \alpha$ is the $k$-th power of cup products of $\alpha$ and $k=\dim \Sigma$.
\end{lemma}
Combining Lemma \ref{L:Hayashi}, \ref{L:Riede} and (\ref{E:isomorphism}), we have proved Theorem \ref{T:double-normals}, which clearly implies Lusternik-Schnirelmann's theorem (Theorem \ref{T:LS}) as a special case since the orthogonal geodesic chords must be contained inside the convex domain by convexity of the domain $\Omega \subset \mathbb{R}^n$.
\section{Shrinking convex chord to a point}
\label{S:convex-chord}
In this section, we study the evolution of chords inside a convex connected planar domain in $\mathbb{R}^2$. In particular, we prove that if an initial chord is \emph{convex}, then it will shrink to a point in finite time under the chord shortening flow. In order to make precise the concept of \emph{convexity}, we need to be consistent with the orientation of a curve in $\mathbb{R}^2$. For this reason, we restrict our attention to plane curves which bounds a domain in $\mathbb{R}^2$.
\begin{definition}[Boundary orientation]
\label{D:boundary-orientation}
For any smooth domain $\Omega \subset \mathbb{R}^2$, we always orient the boundary $\partial \Omega$ as the boundary of $\Omega$ with the standard orientation inherited from $\mathbb{R}^2$. The orientation determines uniquely a global unit tangent vector field, called the \emph{orientation field}, $\xi:\partial \Omega \to T(\partial \Omega)$ such that $\nu:=J \xi$ is the inward pointing normal of $\partial \Omega$ relative to $\Omega$. Here, $J:\mathbb{R}^2 \to \mathbb{R}^2$ is the counterclockwise rotation by $\pi/2$.
\end{definition}
Using Definition \ref{D:boundary-orientation}, we can define the \emph{boundary angle} $\Theta$ which measures the contact angle between a chord $C$ and the boundary $\partial \Omega$.
\begin{definition}[Boundary angle]
\label{D:Theta}
For any (oriented) chord $C_{p,q}$ joining $p$ to $q$ with $p \neq q \in \partial \Omega$, we define the \emph{boundary angle} $\Theta:\{p,q\} \to \mathbb{R}$ by
\[ \Theta(p):=\langle \eta(p), \xi(p) \rangle, \quad \text{and} \quad \Theta(q):=- \langle \eta(q), \xi(q) \rangle,\]
where $\xi$ is the orientation field on $\partial \Omega$ as in Definition \ref{D:boundary-orientation}.
\end{definition}
\begin{definition}
\label{D:convex-chord}
An oriented chord $C_{p,q}$ is \emph{convex} if $\Theta \geq 0$ at both end points.
\end{definition}
\begin{remark}
If we change the orientation of the chord from $C_{p,q}$ to $C_{q,p}$, the boundary angle $\Theta$ changes sign. Since the orientation field $\xi$ is always tangent to $\partial \Omega$, we have $\Theta(p)=\Theta(q)=0$ if and only if $C_{p,q}$ meets $\partial \Omega$ orthogonally at its end points $p$ and $q$.
\end{remark}
If we define the ``unit normal'' $N$ of $\partial C_{p,q}=\{p,q\}$ inside $\partial \Omega$ by setting
\[ N(p)=-\xi(p) \quad \text{and} \quad N(q)=\xi(q),\]
then a solution to the chord shortening flow (\ref{E:CSF'}) can be consider as a smooth $1$-parameter family of pair of points on $\partial \Omega$ given by $\gamma:\{0,1\} \times [0,T) \to \partial \Omega$ such that
\begin{equation}
\label{E:CSF''}
\frac{\partial \gamma}{dt} (u,t) =\Theta (\gamma(u,t)) N(\gamma(u,t)),
\end{equation}
where $\Theta$ is the boundary angle for the oriented chord from $\gamma(0,t)$ to $\gamma(1,t)$. Since the value of $\Theta$ at $u=0$ depends also on the other end point $\gamma(1,t)$, this is a non-local function. Therefore, the chord shortening flow can be thought of as a non-local curve shortening flow driven by the boundary angle $\Theta$.
We are now ready to state the main theorem of this section. The readers can compare Theorem \ref{T:contraction} with the famous result of Huisken \cite{Huisken84} which says that any compact embedded convex hypersurface in $\mathbb{R}^n$ would contract to a point in finite time under the mean curvature flow.
\begin{theorem}
\label{T:contraction}
Let $\Omega \subset \mathbb{R}^2$ be a compact connected domain with smooth convex boundary. Any convex chord which is not an orthogonal geodesic chord would converge to a point in finite time under the chord shortening flow.
\end{theorem}
To prove Theorem \ref{T:contraction} we need to establish a few propositions, which are of geometric interest. We first state the evolution of the boundary angle $\Theta$ under the chord shortening flow. Note that we always have $|\Theta| \leq 1$ by definition.
\begin{proposition} [Evolution of boundary angle]
\label{P:Theta-evolution}
Suppose $C(u,t):I \times [0,T) \to \mathbb{R}^2$ is a solution to the chord shortening flow as in Definition \ref{D:CSF}. Then, the boundary angle $\Theta(u,t): \{0,1\} \times [0,T) \to \mathbb{R}$ satisfies the following equation (recall (\ref{E:norm}) and (\ref{E:DN}))
\begin{equation}
\label{E:Theta-evolution}
\frac{\partial }{\partial t} \Theta = -\Delta^{\frac{1}{2}} \Theta+ \frac{1}{\ell} \Big(\|\Theta\|^2_{L^2}+ \ell k \langle -\eta,\nu \rangle \Big) \Theta + \frac{1}{\ell} \Big( 1+ \langle \xi(p),\xi(q)\rangle \Big) (\Theta-\overline{\Theta}),
\end{equation}
where $k:=\langle \nabla_\xi \xi,\nu \rangle$ is the curvature of $\partial \Omega$ with respect to $\nu$ (recall Definition \ref{D:boundary-orientation}), $\ell=\ell(t)$ is the length of the chord $C(\cdot,t):I \to \mathbb{R}^2$ with outward unit conormal $\eta$.
\end{proposition}
\begin{proof}
It follow directly from Theorem \ref{T:eta-T-evolution} and Definition \ref{D:Theta}
\end{proof}
Using (\ref{E:Theta-evolution}), we immediately have the following evolution equations.
\begin{corollary}
Under the same hypothesis as Proposition \ref{P:Theta-evolution}, we have:
\begin{eqnarray}
\label{E:Theta-avg-evolution}
\frac{d}{dt} \overline{\Theta} &=& \frac{1}{\ell} \Big( \|\Theta\|^2_{L^2} -1-\langle \xi(p),\xi(q) \rangle \Big) \overline{\Theta} + \overline{k \langle -\eta, \nu \rangle \Theta}, \\
\label{E:Theta-norm-evolution}
\frac{1}{2} \frac{d}{dt} \|\Theta\|^2_{L^2} &=& \frac{\ell}{2} \langle \xi(p),\xi(q) \rangle \|\Delta^{\frac{1}{2}} \Theta\|^2_{L^2} + \overline{k \langle -\eta,\nu \rangle \Theta^2} \\
&& \hspace{2cm} + \frac{1}{\ell} \Big( \|\Theta\|^2_{L^2}-1-\langle \xi(p),\xi(q) \rangle \Big) \|\Theta\|^2_{L^2}. \nonumber
\end{eqnarray}
\end{corollary}
\begin{proof}
Both equation follows from (\ref{E:Theta-evolution}) and Lemma \ref{L:DN}.
\end{proof}
Our first lemma is that convexity is preserved under the chord shortening flow. From now on, we will use $C(t)$ to denote the unique solution to the chord shortening flow with initial chord $C(0)$ defined on the maximal time interval $t \in [0,T)$ (where $T$ could be infinite).
\begin{lemma}
\label{L:convex}
Let $C(0)$ be a convex chord inside a compact domain $\Omega \subset \mathbb{R}^2$ with convex boundary $\partial \Omega$. Then, $C(t)$ remains convex for all $t \in [0,T)$.
\end{lemma}
\begin{proof}
Let $\Theta_{min}$ and $\Theta_{max}$ be the minimum and maximum of $\Theta$, both of which is a Lipschitz function of $t$. By (\ref{E:Theta-evolution}), we have the following equality
\begin{equation}
\label{E:Theta-min-1}
\frac{d}{dt} \Theta_{min} = \frac{1}{\ell} \Big( (\|\Theta\|^2_{L^2}-1) \Theta_{min} + \ell k \langle -\eta,\nu \rangle \Theta_{min} - \langle \xi(p),\xi(q)\rangle \Theta_{max} \Big)
\end{equation}
As $\partial \Omega$ is convex, we have $k \geq 0$ and $\langle -\eta,\nu \rangle \geq 0$. Moreover, if the chord is convex, then $\Theta_{min} \geq 0$. Therefore, (\ref{E:Theta-min-1}) implies the following differential inequality
\begin{equation}
\label{E:Theta-min-2}
\frac{d}{dt} \Theta_{min} \geq \frac{1}{\ell} \Big( (\|\Theta\|^2_{L^2}-1) \Theta_{min} - \langle \xi(p),\xi(q)\rangle \Theta_{max} \Big).
\end{equation}
By elementary geometry (see Figure \ref{F:xi}), we can express the term involving the orientation field as
\begin{equation}
\label{E:xi}
\langle \xi(p),\xi(q) \rangle = \Theta_p \Theta_q - \sqrt{(1-\Theta_p^2)(1-\Theta_q^2)}.
\end{equation}
Combining (\ref{E:Theta-min-2}) with (\ref{E:ell-evolution}), noting that $\|\eta^T\|^2_{L^2}=\|\Theta\|^2_{L^2}$ and using (\ref{E:xi}),
\begin{eqnarray*}
\frac{d}{dt} \left( \frac{\Theta_{min}}{\ell} \right) &\geq& \frac{1}{\ell^2} \Big( (2\|\Theta\|^2_{L^2}-1) \Theta_{min} - \langle \xi(p),\xi(q)\rangle \Theta_{max} \Big) \\
&=& \frac{1}{\ell^2} \Big( 2\Theta^3_{min} -(1-\Theta_{max}^2)\Theta_{min} + \sqrt{(1-\Theta_{min}^2)(1-\Theta_{max}^2)} \Theta_{max} \Big) \\
& \geq & \frac{1}{\ell^2} \Big( 2 \Theta_{min}^3 + (1-\Theta_{max}^2) (\Theta_{max} -\Theta_{min}) \Big) \geq 0.
\end{eqnarray*}
Therefore, if $\Theta_{min} \geq 0$ at $t=0$, then $\Theta_{min} /\ell$ is a non-decreasing function of $t$, hence is non-negative for all $t \in [0,T)$. This proves that $C(t)$ remains convex for all $t \in [0,T)$.
\end{proof}
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{xi.png}
\caption{The convex region cut out by a convex chord in $\Omega$. Note that $\langle \xi(p) ,\xi(q) \rangle =\cos (\theta_p+\theta_q)$.}
\label{F:xi}
\end{figure}
We are now ready to prove the main result of this section.
\begin{proof}[Proof of Theorem \ref{T:contraction}]
By Theorem \ref{T:convergence}, it suffices to show that the chord shortening flow $C(t)$ exists only on a maximal time interval $t \in [0,T)$ with $T<+\infty$. First of all, $\Theta \geq 0$ for all $t \in [0,T)$ by Lemma \ref{L:convex}. Using (\ref{E:Theta-avg-evolution}) and (\ref{E:ell-evolution}), notice that $2\|\Theta\|^2_{L^2} \geq \overline{\Theta}^2$, a similar argument as in the proof of Lemma \ref{L:convex} gives\begin{eqnarray*}
\frac{d}{dt} \left( \frac{\overline{\Theta}}{\ell} \right) &\geq & \frac{1}{\ell^2} \Big(\overline{\Theta}^2-1-\langle \xi, \xi \rangle \Big) \overline{\Theta} \\
& \geq & \frac{1}{\ell^2} \Big( \Theta_{min}^2 + \Theta_{min} \Theta_{max} \Big) \overline{\Theta} \geq 0.
\end{eqnarray*}
Therefore, $\overline{\Theta}/\ell$ is a non-decreasing function of $t$. Since $\overline{\Theta}/\ell >0$ at $t=0$, it remains bounded away from zero for all $t \in [0,T)$. Therefore, if $T=+\infty$, by Theorem \ref{T:convergence} we must have $C(t)$ converges to an orthogonal geodesic chord and thus $\overline{\Theta}/\ell \to 0$, which is a contradiction.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction}\label{Sec:introduction}
Let $ G=(V_{G},E_{G}) $ be a finite simple graph.
A \textbf{proper coloring} of $ G $ is a function $ \kappa \colon V_{G} \to \mathbb{N} = \{1, 2, \dots\} $ such that $ \{u,v\} \in E_{G} $ implies $ \kappa(u) \neq \kappa(v) $.
Every proper coloring of $ G $ can be regarded as a graph homomorphism from $ G $ to $ K_{\mathbb{N}} $, the complete graph on $ \mathbb{N} $.
Let $ \Hom(G,K_{\mathbb{N}}) $ denote the set of proper colorings of $ G $.
Stanley \cite{stanley1995symmetric-aim} defined the \textbf{chromatic symmetric function} of $ G $ as follows:
\begin{align*}
X(G,\boldsymbol{x}) \coloneqq \sum_{\kappa \in \Hom(G,K_{\mathbb{N}})} \prod_{v \in V_{G}}x_{\kappa(v)},
\end{align*}
where $ \boldsymbol{x} $ denotes infinitely many indeterminates $ (x_{1}, x_{2}, \dots) $.
By definition, the chromatic symmetric function is homogeneous of degree $ |V_{G}| $.
Stanley conjectured in \cite{stanley1995symmetric-aim} that the chromatic symmetric function distinguishes trees.
Namely, if two trees $ T_{1}, T_{2} $ have the same chromatic symmetric function, then $ T_{1} $ and $ T_{2} $ are isomorphic.
A finite poset $ P $ admits the order quasisymmetric functions, which are kinds of $ P $-partition generating functions studied by Gessel \cite{gessel1984multipartite-cm}.
The order quasisymmetric functions are considered to be analogs of the chromatic symmetric function.
A recent study \cite{hasebe2017order-joac} by Hasebe and the author showed that the order quasisymmetric functions distinguish rooted trees (with the natural poset structures).
The proof is based on algebraic structures of the ring of quasisymmetric functions.
In this paper, we will focus on algebraic structures of the ring of symmetric functions and consider the similar problem for trivially perfect graphs.
We will define classes of graphs which are treated in this paper.
Let $ G,H $ be simple graphs.
The \textbf{disjoint union} $ G \sqcup H $ is defined by $ V_{G \sqcup H} \coloneqq V_{G} \sqcup V_{H} $ and $ E_{G\sqcup H} \coloneqq E_{G} \sqcup E_{H} $ (the set theoretical disjoint unions).
The \textbf{join} $ G + H $ is defined by $ V_{G + H} \coloneqq V_{G} \sqcup V_{H} $ and $ E_{G+H} \coloneqq E_{G} \sqcup E_{H} \sqcup \Set{\{u,v\} | u \in V_{G}, v \in V_{H}} $.
Note that some authors use the symbol $ ``+" $ for disjoint unions.
See Figure \ref{Fig:disjoint union and join} for examples.
\begin{figure}[t]
\centering
\begin{tikzpicture}
\draw (0,-1) node[v](g1){};
\draw (0, 0) node[v](g2){};
\draw (0, 1) node[v](g3){};
\draw (g1)--(g2)--(g3);
\draw (0.7,0) node(){$ \sqcup $};
\draw (1.4,-0.5) node[v](h1){};
\draw (1.4, 0.5) node[v](h2){};
\draw (h1)--(h2);
\draw (2.1,0) node(){$ = $};
\draw (2.8,-1) node[v](gg1){};
\draw (2.8, 0) node[v](gg2){};
\draw (2.8, 1) node[v](gg3){};
\draw (gg1)--(gg2)--(gg3);
\draw (3.8,-0.5) node[v](hh1){};
\draw (3.8, 0.5) node[v](hh2){};
\draw (hh1)--(hh2);
\end{tikzpicture}
\hspace{25mm}
\begin{tikzpicture}
\draw (0,-1) node[v](g1){};
\draw (0, 0) node[v](g2){};
\draw (0, 1) node[v](g3){};
\draw (g1)--(g2)--(g3);
\draw (0.7,0) node(){$ + $};
\draw (1.4,-0.5) node[v](h1){};
\draw (1.4, 0.5) node[v](h2){};
\draw (h1)--(h2);
\draw (2.1,0) node(){$ = $};
\draw (2.8,-1) node[v](gg1){};
\draw (2.8, 0) node[v](gg2){};
\draw (2.8, 1) node[v](gg3){};
\draw (gg1)--(gg2)--(gg3);
\draw (3.8,-0.5) node[v](hh1){};
\draw (3.8, 0.5) node[v](hh2){};
\draw (hh1)--(hh2);
\draw (hh1)--(gg1)--(hh2);
\draw (hh1)--(gg2)--(hh2);
\draw (hh1)--(gg3)--(hh2);
\end{tikzpicture}
\caption{Examples of the disjoint union and the join}\label{Fig:disjoint union and join}
\end{figure}
Some classes $ \mathcal{C} $ of simple graphs can be generated by graph operations.
We consider the following rules.
\begin{enumerate}[(1)]
\item \label{rule 1} $ K_{1} \in \mathcal{C} $.
\item \label{rule 2} If $ G \in \mathcal{C} $, then $ G \sqcup K_{1} \in \mathcal{C} $.
\item \label{rule 3} If $ G \in \mathcal{C} $, then $ G + K_{1} \in \mathcal{C} $.
\item \label{rule 4} If $ G,H \in \mathcal{C} $, then $ G \sqcup H \in \mathcal{C} $.
\item \label{rule 5} If $ G,H \in \mathcal{C} $, then $ G+H \in \mathcal{C} $.
\item \label{rule 6} If $ G \in \mathcal{C} $, then $ \overline{G} \in \mathcal{C} $.
\end{enumerate}
Note that $ K_{n} $ denotes the complete graph on $ n $ vertices and $ \overline{G} $ denotes the complement of $ G $.
A member of the class generated by rules (\ref{rule 1},\ref{rule 2},\ref{rule 3}) is called a \textbf{threshold graph}.
Threshold graphs were introduced by Chv{\'a}tal and Hammer \cite{chvatal1977aggregation-aodm} by a different definition and they gave several characterizations.
Our definition of threshold graphs is equivalent to the original definition by \cite[Theorem 1]{chvatal1977aggregation-aodm}.
A member of the class generated by rules (\ref{rule 1},\ref{rule 3},\ref{rule 4}) is called a \textbf{trivially perfect graph} (or a \textbf{quasi-threshold graph}).
Trivially perfect graphs were introduced by Wolk \cite{wolk1962comparability-potams,wolk1965note-potams} as a comparability graph of an order-theoretic tree.
A number of characterizations for trivially perfect graphs are known.
Our definition of trivially perfect graphs is equivalent to the original definition by \cite[Theorem 3]{jing-ho1996quasi-threshold-dam}.
A member of the class generated by rules (\ref{rule 1},\ref{rule 4},\ref{rule 6}) is called a \textbf{cograph} (short for \textbf{complement reducible graph}).
Cographs were discovered independently by several researchers and many characterizations are known.
In the definition, we can replace the rule (\ref{rule 6}) by (\ref{rule 5}) since we have the formula $ G + H = \overline{\overline{G} \sqcup \overline{H}} $.
Obviously, we have the inclusions
\begin{align*}
\{\text{threshold graphs}\} \subseteq \{\text{trivially perfect graphs}\} \subseteq \{\text{cographs}\}.
\end{align*}
For a class $ \mathcal{F} $ of simple graphs, a simple graph is said to be \textbf{$ \mathcal{F} $-free} if it has no induced subgraphs isomorphic to a member of $ \mathcal{F} $.
The three classes above have forbidden induced subgraph characterizations.
\begin{theorem}[{\cite[Theorem 3]{chvatal1977aggregation-aodm}},\ {\cite[Theorem 2]{golumbic1978trivially-dm}}, \ {\cite[Theorem 2]{corneil1981complement-dam}}]\label{FISC}
Let $ G $ be a simple graph.
\begin{enumerate}[(1)]
\item \label{FISC1} $ G $ is threshold if and only if $ G $ is $ \{2K_{2},C_{4},P_{4}\} $-free.
\item \label{FISC2} $ G $ is trivially perfect if and only if $ G $ is $ \{C_{4}, P_{4}\} $-free.
\item \label{FISC3} $ G $ is a cograph if and only if $ G $ is $ P_{4} $-free.
\end{enumerate}
Here, $ 2K_{2} = K_{2} \sqcup K_{2} $, $ C_{4} $ is a cycle of length four, and $ P_{4} $ is a path on four vertices (see Figure \ref{Fig:forbidden}).
\end{theorem}
\begin{figure}[t]
\centering
\begin{tikzpicture}
\draw (0,1) node[v](1){};
\draw (0,0) node[v](2){};
\draw (1,0) node[v](3){};
\draw (1,1) node[v](4){};
\draw (1)--(2);
\draw (3)--(4);
\draw (0.5,-0.5) node(){$ 2K_{2} $};
\end{tikzpicture}
\qquad
\begin{tikzpicture}
\draw (0,1) node[v](1){};
\draw (0,0) node[v](2){};
\draw (1,0) node[v](3){};
\draw (1,1) node[v](4){};
\draw (1)--(2)--(3)--(4)--(1);
\draw (0.5,-0.5) node(){$ C_{4} $};
\end{tikzpicture}
\qquad
\begin{tikzpicture}
\draw (0,1) node[v](1){};
\draw (0,0) node[v](2){};
\draw (1,0) node[v](3){};
\draw (1,1) node[v](4){};
\draw (1)--(2)--(3)--(4);
\draw (0.5,-0.5) node(){$ P_{4} $};
\end{tikzpicture}
\caption{The forbidden graphs}\label{Fig:forbidden}
\end{figure}
One of two main theorems of this paper is as follows.
\begin{theorem}\label{main theorem 1}
The chromatic symmetric function distinguishes trivially perfect graphs.
Namely, if two trivially perfect graphs $ G,H $ have the same chromatic symmetric function, then $ G $ and $ H $ are isomorphic.
\end{theorem}
\begin{corollary}
The chromatic symmetric function distinguishes threshold graphs.
\end{corollary}
However, the chromatic symmetric function cannot distinguish cographs.
We will give the smallest counter example (see Subsection \ref{Subsec:CSF cannot distinguish cographs}).
To state the other main theorem, we will define $ e $-positivity of graphs.
An \textbf{integer partition} $ \lambda $ is a finite multiset consisting of positive integers.
We write an integer partition as $ \langle 1^{r_{1}} \, 2^{r_{2}}, \dots \rangle $, where $ r_{i} $ is the multiplicity of $ i $.
If $ \lambda \neq \varnothing $ (the empty set), we may write $ \lambda $ as a non-increasing sequence $ (\lambda_{1}, \dotsm \lambda_{\ell}) $ of positive integers.
We call $ \ell $ the \textbf{length} of $ \lambda $.
For a positive integer $ k $, we define the elementary symmetric function $ e_{k} $ to be
\begin{align*}
e_{k} \coloneqq \sum_{i_{1}< \dots < i_{k}}x_{i_{1}} \cdots x_{i_{k}}.
\end{align*}
Moreover, given an integer partition $ \lambda = (\lambda_{1}, \dots, \lambda_{\ell}) $, define $ e_{\lambda} $ to be
\begin{align*}
e_{\lambda} \coloneqq e_{\lambda_{1}} \cdots e_{\lambda_{\ell}}
\end{align*}
and $ e_{\varnothing} \coloneqq 1 $.
It is well known that $ \{e_{\lambda}\}_{\lambda} $ forms a basis for the vector space of symmetric functions over $ \mathbb{Q} $.
There is another well-known basis $ \{s_{\lambda}\}_{\lambda} $, where $ s_{\lambda} $ denotes the Schur function (we omit the definition in this paper).
A simple graph is called \textbf{$ e $-positive} (resp. \textbf{$ s $-positive}) if its chromatic symmetric function can be written as non-negative linear combination of elementary symmetric functions (resp. Schur functions).
It is known that $ e $-positivity implies $ s $-positivity.
Stanley and Stembridge (\cite[Conjecture 5.5]{stanley1993immanants-joctsa} and \cite[Conjecture 5.1]{stanley1995symmetric-aim}) have conjectured that the incomparability graph of $ (\boldsymbol{3}+\boldsymbol{1}) $-free poset is $ e $-positive.
Gasharov \cite[Theorem 2]{gasharov1996incomparability-dm} gave a weaker result: the incomparability graph of $ (\boldsymbol{3}+\boldsymbol{1}) $-free poset is $ s $-positive.
The \textbf{claw graph} is a complete bipartite graph $ K_{1,3} $ (see Figure \ref{Fig:claw}).
\begin{figure}[t]
\centering
\begin{tikzpicture}
\draw (0,0) node[v](0){};
\draw (-1,1) node[v](1){};
\draw ( 0,1) node[v](2){};
\draw ( 1,1) node[v](3){};
\draw (0)--(1);
\draw (0)--(2);
\draw (0)--(3);
\end{tikzpicture}
\caption{The claw graph $ K_{1,3} $}\label{Fig:claw}
\end{figure}
A $ K_{1,3} $-free graph is called \textbf{claw-free}.
Note that every incomparability graph of $ (\boldsymbol{3}+\boldsymbol{1}) $-free poset is claw-free.
Gasharov has conjectured that every claw-free graph is $ s $-positive, which is stated in Stanley's paper \cite[Conjecture 1.4]{stanley1998graph-dm}.
The complete graph $ K_{n} $ is $ e $-positive since $ X(K_{n},\boldsymbol{x})=n!e_{n} $.
The edgeless graph $ \overline{K}_{n} $ is also $ e $-positive since $ X(\overline{K}_{n})=e_{1}^{n} $.
Path graphs and cycle graphs are also known to be $ e $-positive (\cite[Proposition 5.3, Proposition 5.4]{stanley1995symmetric-aim}).
To prove our second theorem, we need the following lemma.
\begin{lemma}[{\cite[Excercise 7.47j]{stanley1999enumerative}}]
If the complement of a simple graph $ G $ is $ K_{3} $-free, then $ G $ is $ e $-positive.
\end{lemma}
Our second main theorem is as follows.
\begin{theorem}\label{main theorem 2}
Let $ G $ be a claw-free cograph, that is, a $ \{K_{1,3},P_{4}\} $-free graph.
Then the complement $ \overline{G} $ is $ K_{3} $-free and hence $ G $ is $ e $-positive.
\end{theorem}
This paper is organized as follows.
In Section \ref{Sec:preliminaries}, we review a few basic concepts of the ring of symmetric functions and investigate properties of chromatic symmetric functions.
In Section \ref{Sec:discrimination}, we give a proof of Theorem \ref{main theorem 1} and the counter example for cographs.
In Section \ref{Sec:e-positivity}, we prove Theorem \ref{main theorem 2}.
\section{Preliminaries}\label{Sec:preliminaries}
\subsection{The ring of symmetric functions}
In this subsection, we review some basic concepts on the theory of symmetric functions.
Our standard reference is \cite{macdonald1995symmetric}.
Recall that $ \boldsymbol{x} = (x_{1},x_{2}, \dots) $ denotes infinitely many indeterminates.
A formal series $ f \in \mathbb{Q}[[\boldsymbol{x}]] $ is called a \textbf{symmetric function} if the following conditions are satisfied.
\begin{enumerate}[(i)]
\item The degrees of the monomials of $ f $ are bounded.
\item $ f $ is invariant under any permutation of the indeterminates.
\end{enumerate}
Let $ \Sym_{\mathbb{Q}} $ denote the subset of the symmetric functions.
It is well known that $ \Sym_{\mathbb{Q}} $ is a subring of $ \mathbb{Q}[[\boldsymbol{x}]] $, which is called the \textbf{ring of symmetric functions}.
For every integer partition $ \lambda $, we associate it with the \textbf{monomial symmetric function} $ m_{\lambda} $, defined by
\begin{align*}
m_{\lambda} \coloneqq \sum_{\alpha} \prod_{i=1}^{\infty}x_{i}^{\alpha_{i}},
\end{align*}
where $ \alpha=(\alpha_{1},\alpha_{2}, \dots) $ runs over all distinct rearrangements of $ \lambda $ considered as a sequence $ (\lambda_{1}, \dots, \lambda_{\ell}, 0, \dots) $ of non-negative integers.
Moreover, we define the \textbf{augmented monomial symmetric function} $ \tilde{m}_{\lambda} $ to be
\begin{align*}
\tilde{m}_{\lambda} \coloneqq \left(\prod_{i=1}^{\infty}r_{i}!\right)m_{\lambda},
\end{align*}
where $ r_{i} $ denotes the multiplicity of $ i $ in $ \lambda $, that is $ \lambda=\langle 1^{r_{1}}, 2^{r_{2}}, \dots \rangle $.
For the empty partition, define $ \tilde{m}_{\varnothing} \coloneqq 1 $.
It is easy to show that the set $ \{\tilde{m}_{\lambda}\}_{\lambda} $ forms a linear basis for $ \Sym_{\mathbb{Q}} $ over $ \mathbb{Q} $.
As with the case of symmetric polynomials (in finite indeterminates), the ring of symmetric function $ \Sym_{\mathbb{Q}} $ is a free commutative algebra, that is, there exists a system of symmetric functions $ \{f_{k}\}_{k \in \mathbb{N}} $ which is algebraically independent over $ \mathbb{Q} $ such that $ \Sym_{\mathbb{Q}}=\mathbb{Q}[f_{k} \mid k \in \mathbb{N}] $.
One of those systems is the system $ \{e_{k}\}_{k \in \mathbb{N}} $ of the elementary symmetric functions.
Another well-known system is the system
$ \{p_{k}\}_{k \in \mathbb{N}} $ of \textbf{power sum symmetric functions}, defined by
\begin{align*}
p_{k} \coloneqq \tilde{m}_{k} = \sum_{i=1}^{\infty}x_{i}^{k}.
\end{align*}
We also define $ p_{\lambda} \coloneqq p_{\lambda_{1}} \cdots p_{\lambda_{\ell}} $ for an integer partition $ \lambda=(\lambda_{1},\dots,\lambda_{\ell}) $ and $ p_{\varnothing} \coloneqq 1 $.
Note that the set $ \{p_{\lambda}\}_{\lambda} $ forms a $ \mathbb{Q} $-basis for $ \Sym_{\mathbb{Q}} $.
\subsection{Chromatic symmetric functions}
In this subsection, we review some properties of chromatic symmetric functions and prepare to prove our main theorem.
For each simple graph $ G $, it is well known that there exists a polynomial $ \chi(G,t) \in \mathbb{Z}[t] $ such that
\begin{align*}
\chi(G,n) = |\Hom(G,K_{n})| \text{ for all } n \in \mathbb{N}.
\end{align*}
The polynomial $ \chi(G,t) $ is called the \textbf{chromatic polynomial} of $ G $.
From the definition of the chromatic symmetric function, we have
\begin{align*}
X(G,\boldsymbol{1}^{n}) = \chi(G,n) \text{ for all } n \in \mathbb{N}, \text{ where } \boldsymbol{1}^{n} \coloneqq (\underbrace{1, \dots, 1}_{n}, 0, \dots).
\end{align*}
Recall that every symmetric function is represented by a polynomial in the power sum symmetric functions.
Define a ring homomorphism $ \varepsilon_{p} \colon \Sym_{\mathbb{Q}} \to \mathbb{Q}[t] $ by the extension of $ \varepsilon_{p}(p_{k}) \coloneqq t $.
\begin{proposition}
Given a simple graph $ G $, we have
\begin{align*}
\varepsilon_{p}(X(G,\boldsymbol{x})) = \chi(G,t).
\end{align*}
\end{proposition}
\begin{proof}
This follows by $ p_{k}(\boldsymbol{1}^{n})=n $ and the discussion above.
\end{proof}
Every simple graph $ G $ has a decomposition $ G=G_{1}\sqcup \dots \sqcup G_{s} $ into the connected components.
The chromatic symmetric function $ X(G,\boldsymbol{x}) $ is determined by the connected components of $ G $.
\begin{proposition}[{\cite[Proposition 2.3]{stanley1995symmetric-aim}}]\label{Stanley CSF disjoint}
Let $ G,H $ be simple graphs.
Then
\begin{align*}
X(G \sqcup H, \boldsymbol{x}) = X(G, \boldsymbol{x})X(H, \boldsymbol{x}).
\end{align*}
\end{proposition}
Cho and van Willigenburg made generators of $ \Sym_{\mathbb{Q}} $ consisting of chromatic symmetric functions.
\begin{theorem}[{\cite[Theorem 5]{cho2016chromatic-tejoc}}]\label{CvW}
Let $ \{G_{k}\}_{k \in \mathbb{N}} $ be a set of connected simple graphs $ G_{k} $ on $ k $ vertices.
Then $ \Sym_{\mathbb{Q}}=\mathbb{Q}[X(G_{k},\boldsymbol{x}) \mid k \in \mathbb{N}] $ and $ \{X(G_{k},\boldsymbol{x})\}_{k \in \mathbb{N}} $ is algebraically independent over $ \mathbb{Q} $.
\end{theorem}
In this paper, the following corollary is required.
\begin{corollary}\label{CvW cor}
Let $ G $ be a simple graph.
Then $ G $ is connected if and only if $ X(G,\boldsymbol{x}) $ is irreducible in $ \Sym_{\mathbb{Q}} $.
\end{corollary}
\begin{proof}
If $ X(G,\boldsymbol{x}) $ is irreducible, then $ G $ is connected by Proposition \ref{Stanley CSF disjoint}.
To show the converse, suppose that $ G $ is a connected graph on $ n $ vertices.
Define a collection of graphs $ \{G_{k}\}_{k \in \mathbb{N}} $ by $ G_{n} \coloneqq G $ and $ G_{k} \coloneqq K_{k} $ for any $ k \neq n $.
By Theorem \ref{CvW}, the set $ \{G_{k}\}_{k \in \mathbb{N}} $ is algebraically independent over $ \mathbb{Q} $ and generates $ \Sym_{\mathbb{Q}} $.
Assume that $ X(G,\boldsymbol{x}) $ is reducible.
Then $ X(G,\boldsymbol{x}) $ can be represented as a polynomial in $ \{X(G_{k},\boldsymbol{x})\}_{k < n} $, which is a contradiction.
Therefore $ X(G,\boldsymbol{x}) $ is irreducible.
\end{proof}
A set partition of the vertex set $ V_{G} $ of a simple graph $ G $ is a collection $ \pi = \{B_{1}, \dots, B_{\ell}\} $ of non-empty subsets of $ V_{G} $ such that $ B_{1} \sqcup \dots \sqcup B_{s} = V_{G} $.
Every $ B_{i} $ is called a block.
the \textbf{type} of a partition $ \pi $ is the integer partition $ \{|B_{1}|, \dots, |B_{\ell}|\} $, denoted by $ \type(\pi) $.
A set partition is called \textbf{stable} if every block induces an edgeless subgraph of $ G $.
Let $ \St_{\lambda}(G) $ denote the set of stable partitions of $ G $ of type an integer partition $ \lambda $.
The chromatic symmetric function can be represented in terms of stable partitions.
\begin{proposition}[{\cite[Proposition 2.4]{stanley1995symmetric-aim}}]\label{Stanley CSF stable partition}
Given a simple graph $ G $, we have
\begin{align*}
X(G,\boldsymbol{x}) = \sum_{\lambda} |\St_{\lambda}(G)| \tilde{m}_{\lambda},
\end{align*}
where $ \lambda $ runs over all integer partitions.
\end{proposition}
This proposition may be considered as a generalization of the following proposition.
\begin{proposition}[{\cite[Theorem 15]{read1968introduction-joct}}]\label{Read coeff falling factorial}
Given a simple graph $ G $, we have
\begin{align*}
\chi(G,t) = \sum_{\ell=1}^{|V_{G}|}|\St_{\ell}(G)|(t)_{\ell},
\end{align*}
where $ \St_{\ell}(G) $ denotes the set of stable partitions of $ G $ consisting of $ \ell $ blocks and $ (t)_{\ell} \in \mathbb{Q}[t] $ denotes the falling factorial.
Namely $ (t)_{\ell} \coloneqq t(t-1)\cdots(t-\ell+1) $.
\end{proposition}
Define a map $ \varepsilon_{\tilde{m}} \colon \Sym_{\mathbb{Q}} \to \mathbb{Q}[t] $ by the linear extension of $ \varepsilon_{\tilde{m}}(\tilde{m}_{\lambda}) \coloneqq (t)_{\ell} $, where $ \ell $ is the length of $ \lambda $.
\begin{proposition}\label{map m_tilde}
Given a simple graph $ G $, we have
\begin{align*}
\varepsilon_{\tilde{m}}(X(G,\boldsymbol{x})) = \chi(G,t).
\end{align*}
\end{proposition}
\begin{proof}
This follows immediately by Propositions \ref{Stanley CSF stable partition} and \ref{Read coeff falling factorial}.
\end{proof}
Note that the maps $ \varepsilon_{p} $ and $ \varepsilon_{\tilde{m}} $ are different since $ \varepsilon_{\tilde{m}} $ is not a ring homomorphism from $ \Sym_{\mathbb{Q}} $ to $ \mathbb{Q}[t] $.
However, if we restricts the domain to the set of chromatic symmetric functions, then $ \varepsilon_{p} $ and $ \varepsilon_{\tilde{m}} $ coincide.
We will introduce multiplications on $ \Sym_{\mathbb{Q}} $ and $ \mathbb{Q}[t] $ such that the map $ \varepsilon_{\tilde{m}} $ becomes a ring homomorphism.
For integer partitions $ \lambda $ and $ \mu $, let $ \lambda \uplus \mu $ denote the union as multisets.
For example, $ (3,2,2,1) \uplus (4,2,1) = (4,3,2,2,2,1) $.
Define a multiplication $ \odot $ on $ \Sym_{\mathbb{Q}} $ by the linear extension of $ \tilde{m}_{\lambda} \odot \tilde{m}_{\mu} \coloneqq \tilde{m}_{\lambda \uplus \mu} $.
Let $ (\Sym_{\mathbb{Q}}, \odot) $ denote the $ \mathbb{Q} $-algebra equipped with the usual addition and the multiplication $ \odot $.
Since $ \{\tilde{m}_{\lambda}\}_{\lambda} $ is a $ \mathbb{Q} $-basis for $ \Sym_{\mathbb{Q}} $, the algebra $ (\Sym_{\mathbb{Q}}, \odot) $ is a free commutative algebra generated by $ \{\tilde{m}_{k}\}_{k \in \mathbb{N}} $.
Moreover, define a multiplication $ \odot $ on $ \mathbb{Q}[t] $ by the linear extension of $ (t)_{\ell} \odot (t)_{m} \coloneqq (t)_{\ell+m} $.
Let $ (\mathbb{Q}[t],\odot) $ be the $ \mathbb{Q} $-algebra equipped with the usual addition and the multiplication $ \odot $.
Then $ (\mathbb{Q}[t],\odot) $ is a free commutative algebra generated by $ (t)_{1} $.
It is easy to verify that the map $ \varepsilon_{\tilde{m}} $ is a ring homomorphism from $ (\Sym_{\mathbb{Q}}, \odot) $ to $ (\mathbb{Q}[t], \odot) $.
We will see that the chromatic symmetric function of the join $ G+H $ is a product of the chromatic symmetric functions of $ G $ and $ H $ with respect to the multiplication $ \odot $.
The following proposition is required, which is an analogy of \cite[Proposition 3.11]{hasebe2017order-joac}.
\begin{proposition}\label{stable partitions of join}
Let $ G $ and $ H $ be simple graphs.
For every integer partition $ \lambda $, there exists a bijection
\begin{align*}
\St_{\lambda}(G+H) \simeq \bigsqcup_{\mu \uplus \nu = \lambda} \left(\St_{\mu}(G) \times \St_{\nu}(H)\right).
\end{align*}
\end{proposition}
\begin{proof}
Every block of a stable partition $ \pi \in \St_{\lambda}(G+H) $ consists of either vertices in $ G $ or vertices in $ H $ since each vertex of $ G $ is adjacent to the vertices of $ H $.
Let $ \pi_{G}, \pi_{H} $ denote the collection of blocks consisting of vertices in $ G, H $, respectively.
Then we have that $ \pi = \pi_{G} \sqcup \pi_{H} $.
Hence the mapping $ \pi \mapsto (\pi_{G},\pi_{H}) $ is a desired bijection.
\end{proof}
The following proposition is an analogy of \cite[Proposition 3.12]{hasebe2017order-joac}.
\begin{lemma}\label{CSF join}
Let $ G $ and $ H $ be simple graphs.
Then
\begin{align*}
X(G+H,\boldsymbol{x}) = X(G,\boldsymbol{x}) \odot X(H,\boldsymbol{x}).
\end{align*}
\end{lemma}
\begin{proof}
By Propositions \ref{Stanley CSF stable partition} and \ref{stable partitions of join}, we have
\begin{align*}
X(G+H,\boldsymbol{x}) &= \sum_{\lambda} |\St_{\lambda}(G+H)| \tilde{m}_{\lambda} \\
&= \sum_{\lambda} \sum_{\mu \uplus \nu = \lambda}|\St_{\mu}(G)||\St_{\nu}(H)|\tilde{m}_{\mu \uplus \nu} \\
&= \sum_{\mu,\nu} |\St_{\mu}(G)||\St_{\nu}(H)|\tilde{m}_{\mu} \odot \tilde{m}_{\nu} \\
&= \left(\sum_{\mu}|\St_{\mu}(G)|\tilde{m}_{\mu}\right)\odot\left(\sum_{\nu}|\St_{\nu}(H)|\tilde{m}_{\nu}\right) \\
&= X(G,\boldsymbol{x})\odot X(H,\boldsymbol{x}).
\end{align*}
\end{proof}
Using Proposition \ref{map m_tilde} and Lemma \ref{CSF join}, we can recover the following result of Read.
\begin{proposition}[{\cite[Theorem 4]{read1968introduction-joct}}]
Let $ G,H $ be simple graphs.
Then
\begin{align*}
\chi(G + H, t) = \chi(G,t) \odot \chi(H,t).
\end{align*}
\end{proposition}
\begin{remark}
There is no unary operation on $ \Sym_{\mathbb{Q}} $ which is compatible with taking the complement.
Stanley's example shows that the graphs $ G $ and $ H $ in Figure \ref{Fig:stanley's example} have the same chromatic symmetric function:
\begin{align*}
X(G,\boldsymbol{x}) = X(H,\boldsymbol{x}) = \tilde{m}_{11111}+4\tilde{m}_{2111}+2\tilde{m}_{221}.
\end{align*}
However, the chromatic symmetric functions of their complements are distinct:
\begin{align*}
X(\overline{G},\boldsymbol{x}) &= \tilde{m}_{11111}+6\tilde{m}_{2111}+5\tilde{m}_{221}+2\tilde{m}_{311}+2\tilde{m}_{32}, \\
X(\overline{H},\boldsymbol{x}) &= \tilde{m}_{11111}+6\tilde{m}_{2111}+5\tilde{m}_{221}+2\tilde{m}_{311}+\tilde{m}_{32}.
\end{align*}
\end{remark}
\begin{figure}[t]
\centering
\begin{tikzpicture}
\draw (-0.86, 0.5) node[v](1){};
\draw (-0.86,-0.5) node[v](2){};
\draw ( 0, 0) node[v](3){};
\draw ( 0.86, 0.5) node[v](4){};
\draw ( 0.86,-0.5) node[v](5){};
\draw (3)--(1)--(2)--(3)--(4)--(5)--(3);
\draw (0,-1) node(){$ G $};
\end{tikzpicture}
\hspace{15mm}
\begin{tikzpicture}
\draw (0,1) node[v](1){};
\draw (0,0) node[v](2){};
\draw (1,0) node[v](3){};
\draw (1,1) node[v](4){};
\draw (2,0) node[v](5){};
\draw (1)--(2)--(3)--(4)--(1)--(3);
\draw (4)--(5);
\draw (1,-0.5) node(){$ H $};
\end{tikzpicture}
\hspace{15mm}
\begin{tikzpicture}
\draw (0,1) node[v](1){};
\draw (0,0) node[v](2){};
\draw (1,0) node[v](3){};
\draw (1,1) node[v](4){};
\draw (0.5,0.5) node[v](5){};
\draw (1)--(2)--(3)--(4)--(1);
\draw (0.5,-0.5) node(){$ \overline{G} $};
\end{tikzpicture}
\hspace{15mm}
\begin{tikzpicture}
\draw (0,1) node[v](1){};
\draw (0,0) node[v](2){};
\draw (1,0) node[v](3){};
\draw (1,1) node[v](4){};
\draw (2,0) node[v](5){};
\draw (1)--(2);
\draw (1)--(3)--(5);
\draw (1)--(4);
\draw (1,-0.5) node(){$ \overline{H} $};
\end{tikzpicture}
\caption{Stanley's examples and their complements}\label{Fig:stanley's example}
\end{figure}
\section{Discrimination}\label{Sec:discrimination}
\subsection{Discrimination for trivially perfect graphs}
We now ready to prove Theorem \ref{main theorem 1}.
The following proof is almost as same as the proof of \cite[Theorem 1.3]{hasebe2017order-joac}.
\begin{proof}[Proof of Theorem \ref{main theorem 1}]
We proceed by induction on $ |V_{G}| $.
When $ |V_{G}|=1 $, we have $ G=H=K_{1} $.
Suppose that $ |V_{G}| \geq 2 $.
Decompose $ G $ and $ H $ into their connected components:
\begin{align*}
G=\bigsqcup_{i=1}^{n}G_{i}, \qquad H=\bigsqcup_{i=1}^{m}H_{i}.
\end{align*}
By the assumption $ X(G,\boldsymbol{x}) = X(H,\boldsymbol{x}) $ and Proposition \ref{Stanley CSF disjoint}, we have
\begin{align*}
\prod_{i=1}^{n}X(G_{i},\boldsymbol{x}) = \prod_{i=1}^{m}X(H_{i},\boldsymbol{x}).
\end{align*}
The ring of symmetric functions $ \Sym_{\mathbb{Q}} $ is a free commutative algebra and hence it is a unique factorization domain.
Using Corollary \ref{CvW cor}, we have that $ n=m $ and $ X(G_{i},\boldsymbol{x}) = X(H_{i}, \boldsymbol{x}) $ for each $ i $ after a suitable renumbering.
Assume that $ n \geq 2 $.
The induced subgraphs $ G_{i},H_{i} $ are also trivially perfect by Theorem \ref{FISC}(\ref{FISC2}) and the number of vertices of $ G_{i} $ is less than $ |V_{G}| $.
Therefore, by our induction hypothesis, we have that $ G_{i} $ is isomorphic to $ H_{i} $.
Hence $ G $ and $ H $ are isomorphic.
Now consider the case $ n=1 $, that is, $ G $ and $ H $ are connected.
By the definition of trivially perfect graphs, there are trivially perfect graphs $ G^{\prime}, H^{\prime} $ such that $ G=G^{\prime}+K_{1} $ and $ H=H^{\prime}+K_{1} $.
Since $ X(K_{1},\boldsymbol{x}) = \tilde{m}_{1} $, using Lemma \ref{CSF join}, we have
\begin{align*}
X(G^{\prime}, \boldsymbol{x}) \odot \tilde{m}_{1} = X(H^{\prime}, \boldsymbol{x}) \odot \tilde{m}_{1}.
\end{align*}
Since the algebra $ (\Sym_{\mathbb{Q}},\odot) $ is an integral domain, we have $ X(G^{\prime},\boldsymbol{x})=X(H^{\prime},\boldsymbol{x}) $.
Our induction hypothesis forces that $ G^{\prime} $ is isomorphic to $ H^{\prime} $.
Thus $ G $ and $ H $ are isomorphic.
\end{proof}
\subsection{Discrimination for cographs}\label{Subsec:CSF cannot distinguish cographs}
As mentioned in Section \ref{Sec:introduction}, the chromatic symmetric function cannot distinguish cographs.
We will raise an example.
A simple graph is called \textbf{coconnected} if its complement is connected.
Consider a simple graph $ G $ and a decomposition $ \overline{G} = \overline{G}_{1} \sqcup \dots \sqcup \overline{G}_{n} $, where $ \overline{G}_{i} $ is a connected component of $ \overline{G} $.
Taking complements of the both sides, we obtain $ G = G_{1} + \dots + G_{n} $.
Every $ G_{i} $ is called a \textbf{coconnected component}.
Since the connected components of a simple graph are uniquely determined, hence coconnected components are also uniquely determined.
The isomorphic classes of cographs is closed under taking the disjoint union $ \sqcup $ and taking the join $ + $.
Let $ \Cograph $ denote the algebraic system equipped with two commutative and associative operations $ \sqcup $ and $ + $ whose underlying set consists of the isomorphic classes of cographs.
\begin{proposition}\label{cograph free}
The algebraic system $ \Cograph $ is free and generated by $ K_{1} $.
\end{proposition}
\begin{proof}
Let $ G $ be a cograph.
We proceed by induction on $ |V_{G}| $.
If $ |V_{G}|=1 $, then $ G=K_{1} $ and there are no other representations.
Assume that $ |V_{G}| \geq 2 $.
By the definition of cographs, $ G $ is either a disjoint union or a join of some cographs.
By the induction hypothesis, the connected components or the coconnected components of $ G $ are represented uniquely by using $ K_{1} $.
Therefore $ G $ also has a unique representation by using $ K_{1} $.
Thus $ \Cograph $ is a free algebraic system.
\end{proof}
\begin{remark}
One can construct an algebraic system called a commutative De Morgan bisemigroup from $ \Cograph $.
A generalized result of Proposition \ref{cograph free} was proven by \cite{esik2003free-ac}.
In \cite{corneil1981complement-dam}, it was shown that every cograph admits a unique cotree representation, which is equivalent to Proposition \ref{cograph free}.
\end{remark}
For the proof of Theorem \ref{main theorem 1}, it plays an important role that a simple graph is connected if and only if its chromatic symmetric function is irreducible in $ \Sym_{\mathbb{Q}} $ (Corollary \ref{CvW cor}).
However, there is no reason why the chromatic symmetric function of a coconnected cograph is irreducible in $ (\Sym_{\mathbb{Q}}, \odot) $.
In fact, we have the following equalities by using Proposition \ref{Stanley CSF stable partition}.
\begin{align*}
X(K_{2} \sqcup K_{1}, \boldsymbol{x}) &= \tilde{m}_{111}+2\tilde{m}_{21} = \tilde{m}_{1} \odot (\tilde{m}_{11}+2\tilde{m}_{2}), \\
X(K_{6}\sqcup K_{1}, \boldsymbol{x}) &= \tilde{m}_{1111111}+6\tilde{m}_{211111} = \tilde{m}_{11111}\odot(\tilde{m}_{11}+6\tilde{m}_{2}), \\
X(K_{4}\sqcup K_{2},\boldsymbol{x}) &= \tilde{m}_{111111}+8\tilde{m}_{21111}+12\tilde{m}_{2211} = \tilde{m}_{11}\odot(\tilde{m}_{11}+2\tilde{m}_{2})\odot(\tilde{m}_{11}+6\tilde{m}_{2}), \\
X(K_{4},\boldsymbol{x}) &= \tilde{m}_{1111}.
\end{align*}
By Lemma \ref{CSF join}, these equalities yield that both of the cographs $ (K_{2}\sqcup K_{1})+(K_{6}\sqcup K_{1}) $ and $ (K_{4}\sqcup K_{2})+K_{4} $ have the same chromatic symmetric function
\begin{align*}
\tilde{m}_{111111}\odot(\tilde{m}_{11}+2\tilde{m}_{2})\odot(\tilde{m}_{1}+6\tilde{m}_{2}).
\end{align*}
Furthermore, by Proposition \ref{cograph free}, we have that these graphs are not isomorphic (Figure \ref{Fig:example cographs}).
\begin{figure}[t]
\centering
\begin{tikzpicture}[baseline=0]
\draw (0,0) node[v](l3){};
\draw (0,2) node[v](l2){};
\draw (0,4) node[v](l1){};
\draw (l1)--(l2);
\draw (3,-1) node[v](r7){};
\draw (3, 0) node[v](r6){};
\draw (3, 1) node[v](r5){};
\draw (3, 2) node[v](r4){};
\draw (3, 3) node[v](r3){};
\draw (3, 4) node[v](r2){};
\draw (3, 5) node[v](r1){};
\draw (r1)--(r2)--(r3)--(r4)--(r5)--(r6);
\draw[bend left] (r1) to (r3);
\draw[bend left] (r1) to (r4);
\draw[bend left] (r1) to (r5);
\draw[bend left] (r1) to (r6);
\draw[bend left] (r2) to (r4);
\draw[bend left] (r2) to (r5);
\draw[bend left] (r2) to (r6);
\draw[bend left] (r3) to (r5);
\draw[bend left] (r3) to (r6);
\draw[bend left] (r4) to (r6);
\draw (l1)--(r1);
\draw (l1)--(r2);
\draw (l1)--(r3);
\draw (l1)--(r4);
\draw (l1)--(r5);
\draw (l1)--(r6);
\draw (l1)--(r7);
\draw (l2)--(r1);
\draw (l2)--(r2);
\draw (l2)--(r3);
\draw (l2)--(r4);
\draw (l2)--(r5);
\draw (l2)--(r6);
\draw (l2)--(r7);
\draw (l3)--(r1);
\draw (l3)--(r2);
\draw (l3)--(r3);
\draw (l3)--(r4);
\draw (l3)--(r5);
\draw (l3)--(r6);
\draw (l3)--(r7);
\end{tikzpicture}
\hspace{25mm}
\begin{tikzpicture}[baseline=-5mm]
\draw (0,4) node[v](l1){};
\draw (0,3) node[v](l2){};
\draw (0,2) node[v](l3){};
\draw (0,1) node[v](l4){};
\draw (0,0) node[v](l5){};
\draw (0,-1) node[v](l6){};
\draw (l1)--(l2)--(l3)--(l4);
\draw[bend right] (l1) to (l3);
\draw[bend right] (l1) to (l4);
\draw[bend right] (l2) to (l4);
\draw (l5)--(l6);
\draw (3,3) node[v](r1){};
\draw (3,2) node[v](r2){};
\draw (3,1) node[v](r3){};
\draw (3,0) node[v](r4){};
\draw (r1)--(r2)--(r3)--(r4);
\draw[bend left] (r1) to (r3);
\draw[bend left] (r1) to (r4);
\draw[bend left] (r2) to (r4);
\draw (l1)--(r1);
\draw (l1)--(r2);
\draw (l1)--(r3);
\draw (l1)--(r4);
\draw (l2)--(r1);
\draw (l2)--(r2);
\draw (l2)--(r3);
\draw (l2)--(r4);
\draw (l3)--(r1);
\draw (l3)--(r2);
\draw (l3)--(r3);
\draw (l3)--(r4);
\draw (l4)--(r1);
\draw (l4)--(r2);
\draw (l4)--(r3);
\draw (l4)--(r4);
\draw (l5)--(r1);
\draw (l5)--(r2);
\draw (l5)--(r3);
\draw (l5)--(r4);
\draw (l6)--(r1);
\draw (l6)--(r2);
\draw (l6)--(r3);
\draw (l6)--(r4);
\end{tikzpicture}
\caption{The smallest example of two non-isomorphic cographs which have the same chromatic symmetric function}\label{Fig:example cographs}
\end{figure}
\section{$ e $-positivity of claw-free cographs}\label{Sec:e-positivity}
In this section, we will prove Theorem \ref{main theorem 2} and conclude that every claw-free cograph is $ e $-positive.
\begin{lemma}\label{coconnected components}
Every coconnected component of a connected claw-free cograph is $ K_{1} $ or a disjoint union of two complete graphs.
\end{lemma}
\begin{proof}
Let $ G $ be a connected claw-free graph.
If $ G $ is complete, then the assertion holds since $ G $ is the join of some single-vertex graphs.
Suppose that $ G $ is non-complete.
The connectivity of $ G $ shows that $ G $ has at least two coconnected components.
Assume that there is a coconnected component $ G_{1} $ such that it consists of at least three connected components.
Take vertices $ a,b,c $ from distinct connected components of $ G_{1} $ and take a vertex $ d $ from a coconnected component distinct from $ G_{1} $.
Then the subgraph of $ G $ induced by $ \{a,b,c,d\} $ is isomorphic to the claw graph, which is a contradiction.
Therefore the number of connected components of every coconnected component of $ G $ is at most two.
\end{proof}
Now we ready to prove Theorem \ref{main theorem 2}.
\begin{proof}[Proof of Theorem \ref{main theorem 2}]
Let $ G $ be a claw-free graph.
Without loss of generality we may assume that $ G $ is connected and non-complete.
By Lemma \ref{coconnected components}, our graph $ G $ is one of the following form:
\begin{align*}
&(G_{1} \sqcup G_{1}^{\prime}) + \dots + (G_{m} \sqcup G_{m}^{\prime}), \\
&(G_{1} \sqcup G_{1}^{\prime}) + \dots + (G_{m} \sqcup G_{m}^{\prime}) + G_{m+1},
\end{align*}
where $ G_{i},G_{i}^{\prime} $ are complete graphs on some vertices.
In order to show that $ \overline{G} $ is $ K_{3} $-free, it suffices to show that any subgraph of $ G $ induced by three vertices $ \{a,b,c\} $ has at least one edge.
If $ a $ belongs to $ G_{m+1} $, then $ a $ is adjacent to any other vertices.
In particular, we obtain edges $ \{a,b\} $ and $ \{a,c\} $.
Suppose that two of $ \{a,b,c\} $ belong to distinct coconnected components.
Then there is an edge connecting these two vertices.
Hence we may assume that $ a,b,c $ belong to $ G_{i} \sqcup G_{i}^{\prime} $ for some $ i $.
In this case, at least two of $ \{a,b,c\} $ belong to the same component and hence we have an edge.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{INTRODUCTION}
\label{introduction}
The {\it initial binary population} (IBP) corresponds to initial
properties of binaries in star clusters, which follow particular
distributions of their parameters, i.e.primary mass, period,
eccentricity, and mass ratio. The concept of an IBP is important
because the full numerical solution to the problem of cloud collapse
leading to star or binary formation is not viable \citep[e.g.][]{Kroupa_2011}, but a
description of initial populations is nevertheless needed for a wide
variety of astrophysical problems -- e.g. star cluster modelling,
stellar population synthesis and dynamical population synthesis
\citep[e.g.][]{Kouwenhoven_2009}.
It has been proposed that the field population of stars comes from the
dissolution of star clusters \citep[e.g.][]{Lada_2003} after the expulsion
of the residual gas in the star formation process. Clustered star formation
and dissolution of these embedded clusters via gas expulsion is the dominant
process that populates the field with stars
\citep[e.g.][]{Lada_2003,Bressert_2010,Lada_2010,Kroupa_2011,Marks_2011b}.
Additionally, it has been argued that the outcome of star-formation processes
depends only weakly on the physical conditions of the molecular cloud
\citep[e.g.][and references therein]{Kroupa_2011}. This is intrinsically
associated with the {\it IBP universality hypothesis} which corresponds to
an environment-independent star formation process
\citep[][see their section 5.2.1]{Marks_2015}.
Alternatively, some authors suggest that the star-formation process is not
universal \citep[e.g.][]{King_2012,Parker_2014}.
For example, \citet[][]{King_2012},
on the basis of the comparison between the properties of close binary systems in seven
young regions and in the field, conclude that the origin of multiplicity is not universal.
In a similar way, \citet{Parker_2014}
argue that the binary population in the field is indicative of the primordial
binary population in star-forming regions, at least for systems with primary
masses in the range 0.02 -- 3.0 M$_\odot$, based on comparisons between $N$-body
simulations and binary properties in the Galactic field.
Concerning these two competing scenarios, we emphasize that both have
the potential of explaining the observed data. However, as discussed
in \citet[][]{Marks_2015},
the hypothesis of the cluster origin of the galactic binary population
in connection with the universality hypothesis is the only currently
available approach that allows binary populations to be predicted
in the Milky Way and other galaxies \citep{Marks_2011b}, and for such a
reason this scenario is assumed in this work.
A promising candidate for the universal IBP has been inferred from
observational data \citep{Kroupa_1995b,Kroupa_2013}, hereafter called
the Kroupa IBP. It has been tested against observations, and has
successfully explained the observational features of young clusters,
associations and Galactic field late-type binaries
\citep[e.g.][]{Kroupa_2011,Marks_2012}. In addition,
\citet{Belloni_2017a} found that models that follow the Kroupa IBP
show good agreement with the observed Galactic cataclysmic variable
white dwarf mass distribution. Direct imaging studies indeed reveal
that most stars form in wide binaries \citep{Sadavoy_2017}.
Regarding the binary fraction in GCs, an efficient way of detecting
the binary content is the search for binary main-sequences (MS) in
colour-magnitude diagrams (CMDs) \citep[e.g.][]{Rubenstein_1997}.
\citet{Sollima_2007} were the first to investigate binaries in a
reasonably large sample of 13 GCs, and found
that all the analysed globular clusters contain a minimum binary fraction
larger than 6 per cent within the core radius and that the global binary fractions
lie between 10 to 50 per cent (depending on the cluster).
\citet{Milone_2012}, who extended the sample
and analysed MS binary populations in a sample of 59 GC CMDs,
found for the first time an anti-correlation
between the binary fraction and the total cluster mass.
Considering that, \citet{Leigh_2015} investigated
whether or not the observed present-day distribution of Galactic GC
binary fractions can be reproduced assuming the Kroupa IBP
(i.e. with a significant fraction of soft binaries).
These authors showed that high initial binary fractions with a
significant soft component (Kroupa IBP) combined with high initial densities can
reproduce the observed anti-correlation between the binary fraction
(both inside and outside the half-mass radius) and the total cluster
mass, which corroborates the idea that an environmental independent
universal IBP might exist associated with the binary population in the Galactic field,
in associations, in young clusters and also in GCs.
Concerning the observed MS binary mass ratios in Galactic GCs,
\citet{Milone_2012} found that the distribution of the mass ratio
is generally flat (for $q>0.5$), with the exception of only
four GCs (E3, Terzan 7, NGC6366 and NGC6496). Since the MS binary mass ratio
distribution is associated with the colour distribution on the red side
of the fiducial MS, we should expect that such a
distribution is also typically flat.
So far, no attempt to compare predicted and observed
GC MS binary distributions has been carried out, taking into account the Kroupa IBP.
On this regard, we will show in this work (Section \ref{problem})
that, even though initial models following the Kroupa IBP predict present-day
binary fractions supported by observations, there is
a problem related to present-day MS binary distributions.
In its original formalism, the Kroupa IBP contains a large
fraction of binaries with approximately equal masses (i.e. $q \approx
1$, where $q$ is the mass ratio). This causes a clearly visible binary
sequence close and above the MS turn-off in present-day GC CMD, and also colour
distributions (below the turn-off) characterized by a strong increase towards
the right edge of the distribution. Such features are, however, not commonly
observed in GCs \citep{Milone_2012}
Rather than providing recipes for simulating present-day binary populations in GCs,
the objective of this paper is first to show that the Kroupa IBP
predicts present-day MS binary distributions (via analysis
of present-day CMD colour distributions) which disagree with the observed ones,
and second to develop a refined Kroupa IBP such that it
can solve the above-mentioned problem with respect to GC MS binary
distributions. Our modifications not only
solve the problem, but also provide
{
similar good
}
agreement with Galactic field
late-type binaries. In addition, this corroborates even more the universality
hypothesis and provides a step forward in better descriptions for energy and
angular momentum redistribution processes during star formation.
The paper is structured as follows. In Section \ref{assumptions} we define
the main concepts and assumptions used in this investigation. In Section
\ref{original}, we introduce the Kroupa IBP and in Section \ref{problem},
we present the problem associated with the original prescription.
The codes utilised in this investigation are described in Section \ref{codes},
and a refined version which solves the problem with MS binary distributions
is developed in Section \ref{modified}. We show the main results
with respect to binaries in the Galactic field and in GCs in
Section \ref{results}. A summary of the main points of this investigation
is provided in Section \ref{discussion}.
\section{DEFINITIONS AND ASSUMPTIONS}
\label{assumptions}
In order to clarify the terminology adopted in this work,
in this section we define the main concepts explored and
the main assumptions adopted. Note that the concept of IBP
was already defined in Section \ref{introduction} and it will be
skipped in this section.
\citet{Kroupa_1995b} developed a simple model
for the redistribution of energy and angular momentum in proto-binary
systems such that it directly leads to binary properties
(mass ratio, eccentricity and period) at the moment the star cluster
dynamical evolution becomes effective. In practice, this process
creates the IBP that should be used at the starting point of star cluster
simulations. This process of converting birth binaries into initial binaries
is usually called {\it pre-main-sequence eigenevolution}, since
it occurs during the pre-main-sequence phase of binaries.
The first distinction we make is between the terms {\it birth} and
{\it initial} binary populations. The term {\it birth} here is applied, as
in \citet{Kroupa_1995b}, to all protostars that are embedded in
circumprotostellar material. On the other hand, the term {\it initial}
corresponds to pre-main-sequence stars not embedded anymore in such
a material, it being accreted or expelled during the process of
redistribution of energy and angular momentum. We use the
subscripts `ini' and `bir' to refer to the initial and birth population,
respectively.
The birth population can be viewed as a theoretical devise
to allow calculations, and it is not observable because proto-binaries
evolve rapidly on a time-scale of $10^5$ yr in a deeply embedded phase.
This is similar to the concept of initial mass function (i.e. the
distribution of birth-stellar masses), which is not an observable distribution
function and also does not exist, but it is an important tool for computations
\citep{Kroupa_2013}. With such a tool, predictions can be made for observable
populations, which is why such theoretical distribution functions are needed.
The second distinction we consider is related to {\it low-mass} binaries
and {\it high-mass} binaries. Low-mass binaries in this work correspond
to all binaries whose primary mass is smaller than $5$ M$_\odot$. Binaries
with primary masses greater than this are called high-mass binaries. Note
that we always assume that primaries are more massive than secondaries,
and that mass ratios are always smaller than or equal to unity
(i.e. $q=M_2/M_1 \leq 1$).
The third distinction concerns {\it short-period} binaries and {\it long-period}
binaries. Short-period binaries have $P < 10^3$ days while long-period binaries
have $P \geq 10^3$ days.
Another distinction we make is related to the {\it original} Kroupa IBP and
the {\it modified} Kroupa IBP. The original Kroupa IBP corresponds to
the IBP generated through pre-main-sequence eigenevolution as described in \citet{Kroupa_1995b},
for low-mass binaries and to the IBP described in \citet{Kroupa_2013}, for high-mass
binaries. As we will see in Section \ref{problem}, an inconsistency appears
in simulated GC CMDs while comparing with real GC CMDs when the
original Kroupa IBP is adopted. In order to overcome this
problem, in this paper we propose a revised pre-main-sequence eigenevolution formulation, which leads in turn to
a modified Kroupa IBP. Additionally, the modified Kroupa IBP, for high-mass
binaries, assumes observed distributions for binaries whose primaries are O
and B-dwarfs.
As a final concept, we use the term {\it stimulated evolution} to refer
to the process of cluster dynamical evolution such that it converts
the IBP to binaries whose distributions resemble those observed in
the Galactic field. Usually, stimulated evolution is relatively
fast and lasts for just a few Myr. In addition, residual gas removal due to
the evolution of the most massive stars leading to significant cluster expansion
\citep[e.g.][]{Brinkmann_2016} is assumed to stop stimulated evolution.
As usually accepted, we assume that the Galactic field stellar population has its origin
in clustered star formation (i.e. dynamical processing of binaries
takes place before the dissolved cluster becomes part of the Galactic
field. In addition, we assume that such clusters contain a high binary
fraction \citep[e.g.][]{Duchene_1999,Kroupa_INITIAL}.
Actually, we assume $\approx$ 100
per cent of binaries, which
is consistent with resolving the angular momentum problem in star formation
and the result that triples and higher order systems are rarely
the outcome of late-type star formation \citep{Goodwin_2005},
such that it leads to an appropriate binary fraction and binary properties
for different spectral type stars after stimulated evolution.
\section{ORIGINAL KROUPA IBP}
\label{original}
In this section we describe the features of the original Kroupa IBP.
These include properties of low-mass binaries (Section \ref{latetype1})
and high-mass binaries (Section \ref{earlytype1}). In the end, in Section
\ref{problem}, we motivate the necessity of changing slightly the original
prescriptions based on mock observations of present-day GC CMDs.
\subsection{Low-mass binaries (primary star mass < 5 M$_\odot$)}
\label{latetype1}
Note that the pre-main-sequence eigenevolution was developed
in order to explain observational correlations found for G, K and M-dwarf
binaries in the Galactic field \citep{DM_1991,M_1992,FM_1992}, and was
ultimately confirmed when data of all late-type binary systems near
the Sun became available \citep{RG_1997}. Therefore,
pre-main-sequence eigenevolution is a process associated closely with late-type or
low-mass binaries.
In what follows, we describe pre-main-sequence eigenevolution
in its original formulation, i.e. \citet{Kroupa_1995b}.
\subsubsection{Birth population}
\label{birth1}
In order to pass though pre-main-sequence eigenevolution, birth
binaries are born with specific distributions, assumed here as:
\begin{description}
\item[i)] {\it Primary mass}: randomly chosen from the Kroupa canonical initial mass function (IMF) \citep{Kroupa_1991}.
\item[ii)] {\it Secondary mass}: secondary is randomly chosen from the same IMF (i.e. the mass
ratio distribution is such that the binary components are randomly paired).
Note that the primary and secondary
are evident only after both stars are chosen independently from the canonical IMF.
\item[iii)]{\it Eccentricity}: it follows a thermal distribution, i.e.
\begin{equation}
f_e \ = \ 2e \; . \label{EQ2}
\end{equation}
Here ${\rm d}N = f_e {\rm d}e$ is the fraction of orbits with
eccentricity in the range $e$ to $e+{\rm d}e$, amongst all orbits.
\item[iv)] {\it Period}: it follows Eq. 8 in \citet{Kroupa_1995b}, i.e.
\begin{equation}
f_P \ = \ 2.5 \ \frac{\log_{10}(P/{\rm days}) - 1}{45 + \left[ \ \log_{10}(P/{\rm days}) \ - \ 1 \ \right]^2} . \label{EQ3}
\end{equation}
Here ${\rm d}N = f_P {\rm d} \log_{10} P$ is the fraction of orbits with period in the range
$\log_{10} P$ to $\log_{10} P + {\rm d}\log_{10} P$ such that the integral of $f_P$ over
all $\log_{10} P$ values equals the binary fraction of the stellar population, being one in this case.
\end{description}
During the pre-main-sequence eigenevolution phase, significant changes
in the binary properties occur when the less massive object (secondary) is
at the pericentre ($R_{p}$), i.e., the distributions above (i--iv) applied
for the birth population can be drastically changed during subsequent
passages through the pericentre. A convenient way to calibrate
the strength of the pre-main-sequence eigenevolution is by means of a function of $R_{p}$.
This is expressed in the function $\rho$, which is defined as
\begin{equation}
\rho \ = \ \left( \frac{\lambda \, {\rm R}_\odot}{R_{p}} \right)^{\chi}, \label{EQ4}
\end{equation}
where $\lambda = 28$ and $\chi = 3/4$ \citep{Kroupa_1995b}, and $R_{p}$ is in units of R$_\odot$.
Note that the larger the pericentre distance, the smaller is the value of
$\rho$.
\subsubsection{Initial population}
\label{initial1}
During the pre-main-sequence eigenevolution which has a duration of about $10^5$ yr, the birth population is converted to the
initial population. The change in the eccentricity due to pre-main-sequence eigenevolution
is given by
\begin{equation}
\ln(e_{\rm ini}) \ = \ - \ \rho \ + \ \ln(e_{\rm bir}) . \label{EQ5}
\end{equation}
As during passages through the pericentre the secondary might accrete
matter from the circumstellar disc around the primary, the changes in the
mass ratio due to the pre-main-sequence eigenevolution is
\begin{equation}
q_{\rm ini} = \left\{
\begin{array}{rcl}
q_{\rm bir} + \rho (1 - q_{\rm bir}),& \mbox{if} & \rho \leq 1 ,\\
1 ,& \mbox{if} & \rho > 1 . \\
\end{array}
\right. \label{EQ6}
\end{equation}
And the change of the secondary mass is given by
\begin{equation}
M_{2,\rm ini} \ = \ q_{\rm ini} \, M_{1,\rm bir} .
\end{equation}
Note that the primary mass remains unchanged during pre-main-sequence eigenevolution, i.e.
$M_{1,\rm ini} = M_{1,\rm bir}$ since the secondary unlikely has an appreciable
circumstellar disk \citep{Bonnell_1992}.
Finally, the period after the pre-main-sequence eigenevolution is given by
\begin{equation}
P_{\rm ini} \ = \ P_{\rm bir} \ \left( \frac{M_{1,\rm bir} + M_{2,\rm bir}}{M_{1,\rm ini} + M_{2,\rm ini}} \right)^{1/2} \ \left( \frac{1 - e_{\rm bir}}{1 - e_{\rm ini}} \right)^{3/2} . \label{EQ7}
\end{equation}
We must comment a few things at this point before proceeding further.
First, concerning mergers during the process of pre-main-sequence eigenevolution, binaries
with pericentre distances greater than $1.1 \times (R_1 + R_2)$ survive, and
they merge otherwise, where $R_1$ and $R_2$ are the primary and secondary radii,
respectively.
Second, the pre-main-sequence eigenevolution changes mainly short-period binaries
since for these systems the function $\rho$ can be truly large.
For long-period binaries, the effect of pre-main-sequence eigenevolution is inexpressive because
the $\rho$ is extremely small (due to large pericentre distance) which leads to
initial properties similar to birth properties.
Third, given the large amount of late-type pre-main-sequence objects generated
according to the adopted IMF, many short-period low-mass binaries might have equal
masses and low eccentricity due to pre-main-sequence eigenevolution, since they more easily have
their properties changed. This has a huge impact in a star cluster population
evolved over a Hubble time (Section \ref{problem}) as these binaries are likely
to survive for several Gyr of cluster dynamical evolution.
In Fig. \ref{FIG01} we plot the main distributions associated with the original Kroupa
IBP regarding low-mass binaries, i.e. primary mass (top left-hand panel), mass ratio
(top right-hand panel), period (bottom left-hand panel), and eccentricity (bottom
right-hand panel). Note that we included as well the modified Kroupa IBP (Section
\ref{latetype2}) for comparison.
\begin{figure}
\begin{center}
\includegraphics[width=0.99\linewidth]{./FIG01.eps}
\end{center}
\caption{Primary mass (top left-hand panel), mass ratio (top right-hand panel),
period (bottom left-hand panel), and eccentricity (bottom right-hand panel) distributions
for all binaries such that $M_1 < 5$ M$_\odot$ in the original Kroupa IBP (Section
\ref{latetype1}, red solid line) and in the modified Kroupa IBP (Section \ref{latetype2},
blue dashed line).}
\label{FIG01}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.99\linewidth]{./FIG02.eps}
\end{center}
\caption{Primary mass (top left-hand panel), mass ratio (top right-hand panel),
period (bottom left-hand panel), and eccentricity (bottom right-hand panel) distributions
for all binaries such that $M_1 > 5$ M$_\odot$ in the original Kroupa IBP (Section
\ref{earlytype1}, red solid line) and in the modified Kroupa IBP (Section \ref{earlytype2},
blue dashed line).}
\label{FIG02}
\end{figure}
\subsection{High-mass binaries (primary star mass > 5 M$_\odot$)}
\label{earlytype1}
Since the pre-main-sequence eigenevolution theory was developed mainly to explain
observed properties of Galactic field low-mass binaries (A, F, G, K and M-dwarfs),
the question concerning how high-mass binaries (O and B-dwarfs)\footnote{Based
on our definition for low-mass and high-mass binaries, binaries whose
primaries are A-dwarfs are low-mass binaries, even though they are early-type binaries.
We follow here the same threshold as in \citet{Kroupa_2013}.} evolve remains open.
According to \citet{Kroupa_2013}, the initial properties of high-mass binaries follow
precisely birth distributions, since pre-main-sequence eigenevolution is not applied in this
range of masses.
This is a reasonable assumption since for this range of masses, the timescale
of pre-main-sequence evolution is short enough that it is safe to neglect it
\citep{Railton_2014}.
Then, for high-mass binaries, birth and initial binaries
follow distributions (i--iv) in Section \ref{latetype1}, with one exception,
namely the secondary mass (i.e. mass ratio). Until very recently, observations
of high-mass binaries have commonly revealed systems with nearly equal masses
\citep[e.g.][]{Pinsonneault_2006,Mayer_2017}.
In order to account for this feature, in the original Kroupa IBP, a different
pairing is usually adopted for high-mass binaries, namely {\it ordered} pairing
\citep{Kroupa_2013}.
The procedure can be summarized as follows. Once all stars (twice the
number of binaries) are generated according to the canonical Kroupa IMF, those with
masses greater than $5$ M$_\odot$ are subsequently ordered. The most massive star in
the array is paired with the second most massive stars, and so on. This procedure
guarantees that the IMF is preserved and generates high-mass binaries with similar masses,
as previously thought from observational results.
Fig. \ref{FIG02} exhibits the main distributions associated with the original Kroupa
IBP with respect to high-mass binaries, i.e. primary mass (top left-hand panel), mass
ratio (top right-hand panel), period (bottom left-hand panel), and eccentricity
(bottom right-hand panel). Note that we included as well the modified Kroupa IBP
(Section \ref{earlytype2}) for comparison, as in Fig. \ref{FIG01}.
\section{NUMERICAL TOOLS}
\label{codes}
We describe is this section the two codes used in this investigation
for star cluster simulations ({\sc mocca}, Section \ref{mocca}) and
for the photometric reduction of mock observations of the simulated star clusters ({\sc cocoa},
Section \ref{cocoa}).
\subsection{The {\sc mocca} code}
\label{mocca}
The {\sc mocca} code \citep[][and references therein]{Hypki_2013,Giersz_2013}
is based on the orbit-averaged Monte Carlo technique
for cluster evolution developed by \citet{Henon_1971}, which
was further improved by \citet{Stodolkiewicz_1986}. It also includes the
{\sc fewbody} code, developed by \citet{Fregeau_2004}, to perform
numerical scattering experiments of gravitational interactions.
To model the Galactic potential, {\sc mocca} assumes a point-mass with
total mass equal to the enclosed Galaxy mass at the Galactocentric
radius. The description of the escape processes in tidally limited
clusters follows the procedure derived by
\citet{Fukushige_2000}. Stellar evolution is implemented via
the {\sc sse} code developed by \citet{Hurley_2000} for single stars and
the {\sc bse} code developed by \citet{Hurley_2002} for binary evolution.
{\sc mocca} was extensively tested against $N$-body codes.
For instance, \citet{Giersz_2013} concluded that {\sc mocca} is
capable of reproducing $N$-body results with reasonable precision, not
only for the rate of cluster evolution and the cluster mass distribution,
but also for the detailed distributions of mass and binding energy of
binaries. Additionally, \citet{Wang_2016} also compared {\sc mocca} with
the state-of-the-art {\sc nbody6++gpu} and showed good agreement
between the two codes. Finally, Madrid et al. (in prep.) showed that
the {\sc mocca} code is able to reproduce the escape rate from tidally
limited clusters when compared with $N$-body codes, provided that
they are not closer than a few kpc from the Galactic center.
\subsection{The {\sc cocoa} code}
\label{cocoa}
The {\sc cocoa} code \citep[][]{Askar_2017b} can create idealised mock observational data
in the form of FITS files using numerical simulation snapshots at a specific time
(provided by codes such as {\sc mocca}/{\sc nbody6}/{\sc nbody6++gpu}). The {\sc cocoa} code
has been developed to extend results of numerical simulations of star clusters for the purpose of
direct comparisons with observations. The input parameters in {\sc cocoa} can be adjusted to create
synthetic observations from virtually any optical telescope and the code can also carry out PSF photometry on
the mock observations to create a catalogue of all observed stars in the cluster. These results can be used to
observationally determine cluster parameters and create observational CMDs of a simulated star cluster model.
The {\sc cocoa} code has already been used for creating mock observations using 12 Gyr simulation snapshots
in investigations with the {\sc nbody6++gpu} code \citep{Wang_2016} and the {\sc mocca} code \citep{Askar_2017a}.
{\sc cocoa} has many different applications \citep{Askar_2017b} and can check if there are any systematics or
biases associated with actual observational data and techniques used to determine cluster properties.
\section{PROBLEM WITH THE ORIGINAL KROUPA IBP}
\label{problem}
The original Kroupa IBP (Section \ref{original})
has been tested against both numerical simulations and observations,
and has successfully explained the observational features of young clusters,
associations, Galactic field late-type binaries, and even binaries in old GCs
\citep[e.g.][and references therein]{Kroupa_2011,Marks_2012,Leigh_2015}.
Even though the original prescription gives good results, there is
at least one problem with the way pre-main-sequence eigenevolution changes the birth population.
As noted in Section \ref{latetype1}, the pre-main-sequence eigenevolution might be very
strong for short-period G, K and M-dwarfs, in the sense that their
initial properties can be extremely different from their birth properties.
For example, substantial number of these short-period binaries will have equal
masses and very low-eccentricities. This makes them very dynamically hard
binaries which in turn allows them to survive in star clusters during
long-term dynamical evolution (see Fig. \ref{FIG03.3}).
The main implication of the above-mentioned result of the pre-main-sequence eigenevolution
in low-mass short-period binaries in GCs are MS binary mass ratio distributions characterized
by a significant fraction of binaries with $q \approx 1$,
which are not typical in observed GCs \citep{Milone_2012}.
First, in Section \ref{problem_model}, we evolve 6 GC models (which follow
initially the Kroupa IBP) with the {\sc mocca}
code and perform the photometric analysis in those models with the {\sc cocoa} code.
Then we show that the colour distributions in the synthetic CMDs, for all
models, have conspicuous peaks towards the right edge of the distributions.
Such peaks are associated with binaries
whose mass ratios are $\approx$ 1, as discussed above.
Second, in Section \ref{problem_real}, we show CMDs of two real GCs and also
their colour distributions, following the same procedure employed in the
synthetic CMDs. We end this section by showing that the Kroupa IBP predicts
colour distributions (and in turn MS binary mass ratio distributions) that
are unlikely to exist in real GCs.
\subsection{GC models, synthetic CMDs and colour distributions}
\label{problem_model}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\linewidth]{./CMD1.eps}
\includegraphics[width=0.48\linewidth]{./CMD2.eps}
\includegraphics[width=0.48\linewidth]{./CMD3.eps}
\includegraphics[width=0.48\linewidth]{./CMD4.eps}
\includegraphics[width=0.48\linewidth]{./CMD5.eps}
\includegraphics[width=0.48\linewidth]{./CMD6.eps}
\end{center}
\caption{Present-day CMDs from simulated observations of the six GC models
(Table 1) evolved with the original Kroupa IBP (Section \ref{original}).
Regions used to generate colour distributions in Fig. \ref{FIG03.2} are demarcated by gray areas.
Note that there is a pronounced binary sequence (particularly above the MS turn-off ) due
to short-period low-mass binaries with mass ratios of unity in the model that has been
indicated by the black arrows in the figure.
For more detail see Section \ref{problem_model}.
}
\label{FIG03.1}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\linewidth]{./C1ori.eps}
\includegraphics[width=0.49\linewidth]{./C2ori.eps}
\includegraphics[width=0.49\linewidth]{./C3ori.eps}
\includegraphics[width=0.49\linewidth]{./C4ori.eps}
\includegraphics[width=0.49\linewidth]{./C5ori.eps}
\includegraphics[width=0.49\linewidth]{./C6ori.eps}
\includegraphics[width=0.98\linewidth]{./cumori.eps}
\end{center}
\caption{Colour distributions derived from the CMDs shown in Fig. \ref{FIG03.1},
following the procedure described in Section \ref{problem_model}.
In the first six panels, we show histograms together with horizontal lines which are
average numbers, assuming flat distributions. Vertical lines in the histograms correspond
to Poisson errors. In the last panel,
we compare cumulative distributions, and display in the key $p$-values of
one-sample Kolmogorov-Smirnov tests for uniformity, for each model.
Only stars with $(B-V)'>0.05$ are used to compute the above distributions.
Note that in all colour distributions a conspicuous peak is observed towards
the right edge the distributions, which is associated with the binary sequences visible in the CMDs
exhibited in Fig. \ref{FIG03.1}.
With respect to the cumulative distributions,
it is clear that the colour distributions in the six models following the original
Kroupa IBP are not uniform. This is also supported by the statistical test, which allows us to
reject the hypothesis that they are uniform with more than 99 \% of confidence (see $p$-values).
For more detail see Sections \ref{problem_model}.
}
\label{FIG03.2}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{./per_ini_pd.eps}
\end{center}
\caption{Initial and present-day period distributions of all MS binaries
whose $q>0.99$, in all six models described in Table 1.
Note that both distributions are practically the same, which strongly supports
the idea that the binary sequence for binaries with $q \approx 1$ comes from
primordial binaries practically not affected by dynamics. The reason is because
such binaries are very-short-period, having the majority of binaries with
$q \approx 1$ periods shorter than 100 days. This suggests that in order
to avoid the prediction of such binary sequences, the IPB has to be changed.
For more detail see Sections \ref{problem_model}.
}
\label{FIG03.3}
\end{figure}
\begin{sidewaystable*}
\centering
\begin{adjustbox}{max width=\textwidth}
\label{Tab01}
\noindent
\begin{threeparttable}
\noindent
\setlength\tabcolsep{1.5pt}
\begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c}
\caption{GC models, initial parameters that define them and present-day properties (i.e. after 12 Gyr of evolution).
The binary fraction, the metallicity and the the King model parameter are
95 per cent, 0.001 and 6.0, respectively, for all models,
In addition, for all models, we adopt the canonical \citet{Kroupa_1991}. IMF (i.e. with two segments),
and the original Kroupa IBP (Section \ref{original}).
The first column presents the model name and the following 9
columns give the initial values for the total
mass (in M$_\odot$), the total number of stars, the central density (in M$_\odot$ pc$^{-3}$),
the core radius (in pc), the half-mass radius (in pc),
the tidal radius (in pc), the Galactocentric
distance (in kpc), the half-mass radius relaxation time (in Myr),
and the central velocity dispersion (in km s$^{-1}$). The remaining 11 columns provide
present-day values for the total
mass (in M$_\odot$), the total number of stars, the central density (in M$_\odot$ pc$^{-3}$),
the core radius (in pc), the half-mass radius (in pc),
the tidal radius (in pc), the half-mass radius relaxation time (in Myr),
the non-mass-weighted central velocity dispersion (in km s$^{-1}$), the cluster type \tnote{a} and
the total cluster binary fraction.
See Section \ref{mocca}, for more details.} \label{Tab01} \\
\hline
& \multicolumn{10}{c}{initial} & & \multicolumn{9}{c}{present-day} \\
\hline
Model &
\specialcell{M\\$\left[ \times \, 10^5 \, {\rm M}_\odot \right]$} &
\specialcell{N\\$\left[ \times \, 10^5 \right]$} &
\specialcell{$\rho_c$\\$\left[ {\rm M}_\odot \; {\rm pc}^{-3} \right]$} &
\specialcell{$R_{\rm c}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm h}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm t}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm G}$\\$[{\rm kpc}]$} &
\specialcell{$T_{\rm r,h}$\\$\left[ {\rm Myr} \right]$} &
\specialcell{$\sigma_c$\\${\rm \left[ km \, s^{-1} \right]}$} &
&
\specialcell{M\\$\left[ \times \, 10^5 \, {\rm M}_\odot \right]$} &
\specialcell{N\\$\left[ \times \, 10^5 \right]$} &
\specialcell{$\rho_c$\\$\left[ {\rm M}_\odot \; {\rm pc}^{-3} \right]$} &
\specialcell{$R_{\rm c}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm h}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm t}$\\$[{\rm pc}]$} &
\specialcell{$T_{\rm r,h}$\\$\left[ {\rm Myr} \right]$} &
\specialcell{$\sigma_c$\\${\rm \left[ km \, s^{-1} \right]}$} &
Type \tnote{a} &
\specialcell{Total binary\\fraction} \\
\hline\hline
1 & 3.70 & 7.8 & 2.43 $\times 10^4$ & 0.71 & 2.4 &60& 4.4 & 5.60 $\times 10^2$ & 25.2 &
& 2.07 & 6.2 & 3.6 $\times 10^7$ & 0.38 & 5.6 & 49.4 & 3.5 $\times 10^3$ & 34.5 & cIMBH & 0.12 \\ \hline
2 & 3.70 & 7.8 & 4.87 $\times 10^2$ & 2.62 & 8.8 &60& 4.4 & 3.95 $\times 10^3$ & 13.2 &
& 1.06 & 3.0 & 1.9 $\times 10^2$ & 1.81 & 8.1 & 39.2 & 4.2 $\times 10^3$ & 6.1 & pc & 0.24 \\ \hline
3 & 3.70 & 7.8 & 3.89 $\times 10^3$ & 1.31 & 4.4 &30& 1.6 & 1.40 $\times 10^3$ & 18.6 &
& 0.09 & 0.2 & 2.4 $\times 10^5$ & 0.02 & 1.4 & 8.9 & 7.2 $\times 10^1$ & 4.5 & c & 0.30 \\ \hline
4 & 9.25 & 19.5 & 1.05 $\times 10^5$ & 0.74 & 2.4 &60& 2.8 & 7.99 $\times 10^2$ & 30.7 &
& 5.55 & 16.9 & 5.8 $\times 10^8$ & 0.58 & 4.9 & 50.6 & 4.6 $\times 10^3$ & 45.7 & cIMBH & 0.09 \\ \hline
5 & 9.25 & 19.5 & 2.11 $\times 10^3$ & 2.73 & 8.8 &60& 2.8 & 5.64 $\times 10^3$ & 16.0 &
& 2.91 & 8.5 & 3.6 $\times 10^2$ & 2.53 & 8.6 & 40.8 & 7.1 $\times 10^3$ & 10.1 & pc & 0.20 \\ \hline
6 & 9.25 & 19.5 & 1.69 $\times 10^4$ & 1.37 & 4.4 &30& 1.0 & 1.99 $\times 10^3$ & 22.7 &
& 1.15 & 2.5 & 3.3 $\times 10^4$ & 0.19 & 2.4 & 14.9 & 5.8 $\times 10^2$ & 11.0 & pc & 0.16 \\ \hline
\hline
\label{Tab01}
\end {tabular}
\label{Tab01}
\begin{tablenotes}
\item[a] The cluster present-day type can be: post-core collapse (c), post-core collapse with intermediate-mass black hole (cIMBH) and pre-core collapse (pc).
\end{tablenotes}
\label{Tab01}
\end{threeparttable}
\label{Tab01}
\end{adjustbox}
\qquad
\vspace{2.0cm}
\begin{adjustbox}{max width=\textwidth}
\noindent
\begin{threeparttable}
\noindent
\setlength\tabcolsep{1.5pt}
\begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c}
\label{Tab02}
\caption{GC models (whose initial conditions are exactly the same as those listed in Table 1),
having initial binaries following the modified Kroupa IBP
(Section \ref{modified}). Columns are the same as in Table 1.
Notice that model 4, which contains a very massive IMBH (which dominates the dynamics of the core),
has a significantly larger central velocity dispersion, when compared with the same model in Table 1.
This is because the process of IMBH formation is stochastic \citep{Giersz_2015}, so the
masses of IMBHs at 12 Gyr are different, which causes differences in non-mass-weighted central velocity dispersions.
See Section \ref{mocca}, for more details.}\\
\hline
& \multicolumn{10}{c}{initial} & & \multicolumn{9}{c}{present-day} \\
\hline
Model &
\specialcell{M\\$\left[ \times \, 10^5 \, {\rm M}_\odot \right]$} &
\specialcell{N\\$\left[ \times \, 10^5 \right]$} &
\specialcell{$\rho_c$\\$\left[ {\rm M}_\odot \; {\rm pc}^{-3} \right]$} &
\specialcell{$R_{\rm c}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm h}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm t}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm G}$\\$[{\rm kpc}]$} &
\specialcell{$T_{\rm r,h}$\\$\left[ {\rm Myr} \right]$} &
\specialcell{$\sigma_c$\\${\rm \left[ km \, s^{-1} \right]}$} &
&
\specialcell{M\\$\left[ \times \, 10^5 \, {\rm M}_\odot \right]$} &
\specialcell{N\\$\left[ \times \, 10^5 \right]$} &
\specialcell{$\rho_c$\\$\left[ {\rm M}_\odot \; {\rm pc}^{-3} \right]$} &
\specialcell{$R_{\rm c}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm h}$\\$[{\rm pc}]$} &
\specialcell{$R_{\rm t}$\\$[{\rm pc}]$} &
\specialcell{$T_{\rm r,h}$\\$\left[ {\rm Myr} \right]$} &
\specialcell{$\sigma_c$\\${\rm \left[ km \, s^{-1} \right]}$} &
Type \tnote{a} &
\specialcell{Total binary\\fraction}\\
\hline\hline
1 & 3.70 & 7.8 & 2.43 $\times 10^4$ & 0.71 & 2.4 &60& 4.4 & 5.60 $\times 10^2$ & 25.2 &
& 2.05 & 6.24 & 5.4 $\times 10^7$ & 0.21 & 5.5 & 49.4 & 3.5 $\times 10^3$ & 33.8 & cIMBH & 0.12 \\ \hline
2 & 3.70 & 7.8 & 4.87 $\times 10^2$ & 2.62 & 8.8 &60& 4.4 & 3.95 $\times 10^3$ & 13.2 &
& 1.04 & 2.99 & 6.8 $\times 10^2$ & 3.62 & 8.1 & 39.5 & 4.2 $\times 10^3$ & 6.0 & pc & 0.23 \\ \hline
3 & 3.70 & 7.8 & 3.89 $\times 10^3$ & 1.31 & 4.4 &30& 1.6 & 1.40 $\times 10^3$ & 18.6 &
& 0.09 & 0.16 & 4.3 $\times 10^5$ & 0.05 & 1.3 & 8.9 & 6.9 $\times 10^1$ & 4.7 & c & 0.26 \\ \hline
4 & 9.25 & 19.5 & 1.05 $\times 10^5$ & 0.74 & 2.4 &60& 2.8 & 7.99 $\times 10^2$ & 30.7 &
& 5.48 & 16.98 & 8.6 $\times 10^8$ & 0.36 & 4.9 & 50.6 & 4.6 $\times 10^3$ & 66.9 & cIMBH & 0.08 \\ \hline
5 & 9.25 & 19.5 & 2.11 $\times 10^3$ & 2.73 & 8.8 &60& 2.8 & 5.64 $\times 10^3$ & 16.0 &
& 2.76 & 8.22 & 2.5 $\times 10^3$ & 3.28 & 8.5 & 40.3 & 7.0 $\times 10^3$ & 9.6 & pc & 0.20 \\ \hline
6 & 9.25 & 19.5 & 1.69 $\times 10^4$ & 1.37 & 4.4 &30& 1.0 & 1.99 $\times 10^3$ & 22.7 &
& 1.04 & 2.25 & 2.5 $\times 10^4$ & 0.32 & 2.3 & 14.5 & 5.3 $\times 10^2$ & 10.8 & pc & 0.16 \\ \hline
\hline
\end {tabular}
\begin{tablenotes}
\item[a] The cluster present-day type can be: post-core collapse (c), post-core collapse with intermediate-mass black hole (cIMBH) and pre-core collapse (pc).
\end{tablenotes}
\end{threeparttable}
\end{adjustbox}
\end{sidewaystable*}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\linewidth]{./M4.eps}
\includegraphics[width=0.49\linewidth]{./NGC6397.eps}
\includegraphics[width=0.95\linewidth]{./ACS.eps}
\includegraphics[width=0.95\linewidth]{./cumGC.eps}
\end{center}
\caption{CMD (top panels) and colour distributions (middle and bottom panels)
of M4 (left-hand column) and NGC6397 (right-hand column).
The CMDs were generated with the HST Globular Cluster Treasury catalogue
\citep{Sarajedini_2007}, which is based on HST ACS/WFC data, and the
colour distributions were generated following the prescription in Section \ref{problem_model}.
Only stars with $(m_{\rm F606W}-m_{\rm F814W})'>0.05$ are used to compute the above distributions.
The gray areas in the CMDs indicate the region from where the colour distributions
were derived, and the horizontal lines in the histograms are average numbers,
assuming flat distributions.
Notice that in these two GCs, taken as examples because of their proximities,
pronounced sequences caused by binaries with $q \approx 1$ are not visible, especially in the
region above and close to the turn-off (compare with Fig. \ref{FIG03.1}).
Additionally, their colour distributions are consistent with uniform distributions. This is
is indicated in the last panel which shows the colour cumulative distributions and the
$p$-values of one-sample Kolmogorov-Smirnov tests for uniformity. The test does not allow us to
reject the hypothesis that they are uniform. For more detail see Sections \ref{problem_real}.
}
\label{FIG04}
\end{figure}
In order to compare different synthetic CMDs with real CMDs, we evolved
6 models with different initial conditions for 12 Gyr. We adopt for all models the
canonical IMF that follows the broken power law $\xi(m) \propto m^{-\alpha}$, defined by
\citet{Kroupa_2001}, where $\alpha = 1.3$ for $0.08 \leq m/{\rm M_\odot} \leq 0.5$ and
$\alpha = 2.3$ for $0.5 \leq m/{\rm M_\odot} \leq m_{\rm max}/{\rm M_\odot}$,
and the star mass in this study lies between $0.08 {\rm M_\odot}$ and $100 {\rm M_\odot}$.
Additionally, all models have 95 per cent of primordial binaries\footnote{
We set the initial binary fraction
different from 100 per cent in order to avoid computational problems
that arise in MOCCA if there is no single star in the initial
model.}. The metallicity equals $Z=0.001$,
which is typical for GCs, and their binary properties are given by the
original Kroupa IBP (Section \ref{original}).
We assume that all stars are on the zero-age main sequence
when the simulation begins and that any residual gas from the star
formation process has already been removed from the cluster. Additionally,
all models are initially at virial equilibrium, and have neither
rotation nor mass segregation. Moreover, all models are evolved
for 12 Gyr which is associated with the present-day in this investigation.
In Table 1, we summarize the main parameters of the initial and present-day
models. Notice that we have a small set of GC models and, even being
so, we still have very different initial and present-day properties,
which guarantees that a considerable region of the parameter space is covered,
allowing us in turn to access different CMD morphologies arising from
different choices of initial conditions.
In order to investigate the distribution of binaries in GC models
simulated with the original Kroupa IBP (Section \ref{original}),
CMDs were obtained by imaging and carrying out photometry on the inner
parts of the clusters (within their half-light radii) with the {\sc cocoa} code,
assuming that the clusters are at a heliocentric distance of 5 kpc. The observations
were simulated with an HST type telescope with a pixel scale of 0.05''/pix,
a seeing value\footnote{We emphasize that with HST-like observations we should have no seeing,
a priori. However, in order to perform photometry, we need to achieve a good full width at
half maximum (FWHM) value, which implies that a definition of seeing is required.}
in both the filters (B and V) of 0.15'', and a Gaussian PSF was used.
While these idealized mock observations do not model the exact HST PSF, the combination
of high spatial resolution images with extremely low seeing values can reproduce FWHM
values for HST images in optical filters. The PSF photometry obtained from
these synthetic images are comparable to the results from HST photometry particularly
for the magnitude regime that we are interested in.
From the synthetic CMDs, we obtained the fiducial MS in the following way,
which is similar to the procedure adopted in \citet{Milone_2012}.
First we select all stars in the magnitude interval of interest,
which is adopted here as $[5.0,5.5]$. Second we divide such an interval
into 40 bins and obtain the median (in colour and magnitude) in the 2D
region defined by each interval. Finally the fiducial MS, in the interval, has been
derived by fitting these median points with a line. From such a fiducial MS,
we have the line properties, which allows us to rotate and
translate this region such that the MS becomes a vertical
line centred at the origin of the x axis (the colour axis). From this rotated
and translated part of each CMD, we compute the rotated colour $(B-V)'$ distribution,
starting from the value $(B-V)' = 0.05$. This lower limit guarantees that the systems,
for $(B-V)' > 0.05$, are likely binaries with $q > 0.5$. System with $(B-V)' < 0.05$
are either single stars belonging to the MS or binaries close to the MS
(i.e. binaries with $q<0.5$).
Fig. \ref{FIG03.1} exhibits the synthetic CMDs and Fig. \ref{FIG03.2} the colour
distributions. Notice that the colour distribution is an increasing function of the colour
and that the peaks shown in the right edge of each distribution
are directly connected with the abundant presence of MS binaries whose mass ratios
are $\approx$ 1. The large width of the peak is connected with observational errors, which
are larger for less luminous stars.
From the histograms in Fig. \ref{FIG03.2} we can expect that the distributions are not
uniform, since the dashed horizontal lines (average number, assuming a uniform
colour distribution) are unlikely good fits for the histograms. In order to confirm
this, we applied a one-sample Kolmogorov-Smirnov test for uniformity to the six
models. All models exhibit rather small $p$-values (see keys in bottom panel
of Fig. \ref{FIG03.2}). The result of this test allows rejection of the null hypothesis
that the colour distributions are uniform with more than 99 per cent of confidence.
This indicates that, if MS binary mass ratio distributions (and in turn colour distributions)
in real GCs are consistent with uniformity, then a potential problem with
the Kroupa IBP in matching binary distributions in GCs seems to exist.
In order to show that the peaks in the colour distribution in Fig. \ref{FIG03.2} are instrinsically connected with the
Kroupa IBP, we show in in Fig. \ref{FIG03.3} the initial and present-day period distributions of
all present-day MS binaries whose $q>0.99$ in all six models. We note that both distributions
are basically the same, which supports that they are
primordial binaries not strongly affected by dynamics. In addition, the majority
of these binaries have periods shorter than 100 days, which makes them very
dynamically hard and not easily subject to disruption via dynamical interactions.
In the following section, we show, as examples, two real GC CMDs and
their colour distributions, compare them with results obtained for the
original Kroupa IBP, and conclude that the original Kroupa IBP is probably
not good in matching MS binary distributions in real GCs.
\subsection{Real GC CMDs and the problem with the original Kroupa IBP}
\label{problem_real}
In order to show that GC models set with the
original Kroupa IBP are not good to reconcile observed
CMD morphologies and MS binary colour distributions, we analyse here
two real GCs, namely M4 and NGC6397. These two clusters are
the nearest GCs \citep[][2010 edition]{Harris} and are also therefore well-studied.
In addition to the quality of data concerning these two GCs, M4 can be described
by a classic King profile, while NGC6397 is a post-core collapse cluster,
which allows us to investigate CMDs and colour distributions in two
different `dynamical regimes'.
Fig. \ref{FIG04} exhibits the CMDs and the colour distributions of M4
and NGC6397. The CMDs were generated with the HST Globular Cluster Treasury catalogue
\citep{Sarajedini_2007}, which is based on HST ACS/WFC data, and the
colour distributions were generated following the prescription in Section
\ref{problem_model}, being the magnitude interval $[17.5,18.5]$ in both
GCs.
The first incompatibility we notice while comparing Figs. \ref{FIG03.2}
and \ref{FIG04} is the absence of a pronounced binary sequence above the
MS turn-off in the CMDs of M4 and NGC6397. This sort of sequence
is present in all CMDs of Fig. \ref{FIG03.2}, which is caused by
binaries which are very dynamically hard and have $q \approx 1$, as illustrated
in Figs. \ref{FIG03.2} and \ref{FIG03.3}. Additionally, if
this feature is present in all CMDs, and if the original Kroupa IBP
should correspond to initial GC binary properties, then such a
behaviour should be also visible in real GCs. Even in the nearest
GCs this is not the case, which implies that the original Kroupa
IBP seems not to be adequate to reproduce binary distributions in real
GCs.
Similarly, contrary to what is predicted when models are set with the
original Kroupa IBP, the colour distributions do seem to be flat in real
GCs, since the dashed horizontal lines in the histograms appear to be a good
parent model for the distributions. Again, in order to test this in
a more coherent way, we applied one-sample Kolmogorov-Smirnov tests
for uniformity. The $p$-values are 0.305 and 0.46, for M4 and NGC6397,
respectively. This test does not allow us to reject the hypothesis that
their colour distributions are uniform, which implies that they are consistent
with uniform distributions.
We stress here that we decided to perform the analysis with respect to only two GCs
because several investigations have already been carried out with regards
to CMDs obtained with HST data \citep[e.g.][]{Piotto_2002,Sarajedini_2007,Milone_2012}.
None of them is consistent with a mass ratio distribution peaked at $q \approx 1$.
In addition, since \citet{Milone_2012} have already concluded that the observed MS mass ratio
distributions in GCs are consistent with uniform distributions, we expect that
our procedure concerning the colour distributions would reveal the same trends found for M4 and
NGC6397 (i.e. colour distribution consistent with the hypothesis of uniformity), if more
GCs were included.
Additionally, the HST F606W and F814W photometric filter bands approximately correspond to the V and I
band filters, respectively. So our use of the observational CMDs in terms of
(V-I) vs I are a very good approximation to HST data sufficient for the present
purpose. \citet{Askar_2017b} have shown that such a binary sequence is also very
pronounced (see their figs. 11, 20 and 21) in CMDs generated with the V, U and I
bands. Therefore, there should not be any serious biases in the compared data sets,
in particular, since we are not aiming at a detailed modelling of a particular cluster.
Summarising, binary distributions found in present-day GC models set with the original Kroupa IBP
are unlikely to match those derived from observations and the evidence for that is twofold. First
in part of the CMD above the MS turn-off, a pronounced binary sequence is clearly predicted, although
not observed. Second, the observed colour distributions below the turn-off are consistent with uniform distributions,
and predicted distributions are characterized by a peak in the right edge of the
distributions. As both discrepancies are mainly caused by the presence of a significant fraction of short-period
low-mass binaries whose component masses are similar, a way to avoid such a non-observed
effect is to reduce the fraction of these systems in the IBP. In what follows, we
describe how we did this by changing slightly the pre-main-sequence eigenevolution
prescription as well as what distributions we adopt for high-mass
binaries, which differ drastically from those of the original Kroupa
IBP.
\section{MODIFIED KROUPA IBP}
\label{modified}
In order to prevent the
discrepancies pointed out in the previous section with respect
to MS binary distributions
(due to binaries whose components
have equal masses) in present-day GC CMDs, we can
change the pre-main-sequence eigenevolution prescription. In what follows,
we describe the procedure to find a revised pre-main-sequence eigenevolution process
as well as what sort of distributions for high-mass binaries we should
adopt in order to respect observational constraints.
\subsection{Looking for a revised pre-main-sequence eigenevolution}
\label{method}
The first step we adopted in changing the pre-main-sequence eigenevolution is the
strength with which the secondary accretes matter from the circumstellar
disk around the primary. This is intrinsically associated with one
parameter in Section \ref{latetype1}, namely $\lambda$ (Eq. \ref{EQ4})
and with the efficiency of the accretion expressed by $\rho$ (Eq. \ref{EQ6}).
In order to reduce such a strength of increasing the secondary mass,
we assumed a mass-depend $\lambda$ such that the lower the primary mass,
the weaker the redistribution of energy and angular momentum within their
circumstellar material. This dependence with the primary mass
would take into account differences in binaries with different spectral types.
Thus, we assumed that $\lambda(M_1)$ is a monotonic
function of the primary mass up to a critical primary mass $M_{\rm crit}$ and a constant
after that, i.e.
\begin{equation}
\lambda(M_1) = \left\{
\begin{array}{rcl}
\lambda_0 \times M_1^\eta,& \mbox{if} & M_1 \leq M_{\rm crit} , \\
\lambda_0 ,& \mbox{if} & M_{\rm crit} < M_1 < 5 \; {\rm M}_\odot ,\\
\end{array}
\right. \label{EQ8}
\end{equation}
where $\lambda_0$, $\eta$, and $M_{\rm crit}$ are parameters to be determined.
In addition to a mass-dependent $\lambda(M_1)$, we also added a parameter
$\varepsilon \in (0,1)$ in Eq. \ref{EQ6} (associated with the efficiency at which the
secondary mass accretes matter from the primary circumstellar disk).
We also added in the expression a stochastic component in the accretion process of
the secondary for low-mass binaries whose primary masses are smaller than $M_{\rm crit}$.
The modified secondary accretion process, i.e. modified initial mass ratio is given by
\begin{equation}
q_{\rm ini} = \left\{
\begin{array}{rcl}
q_{\rm bir} + \rho_0 (1 - q_{\rm bir}) , & \mbox{if} & \rho_0 < 1 ,\\
0.9 + 0.1 \times U(0,1) , & \mbox{if} & \rho_0 \geq 1 \ {\rm and} \ M_1 < M_{\rm crit} ,\\
1 , & \mbox{if} & \rho_0 \geq 1 \ {\rm and} \ M_1 \geq M_{\rm crit} ,\\
\end{array}
\right. \label{EQ9}
\end{equation}
where $ U(0,1)$ is the standard uniform distribution and
\begin{equation}
\rho_0 = \left\{
\begin{array}{rcl}
\varepsilon \rho , & \mbox{if} & M_1 < M_{\rm crit} ,\\
\rho , & \mbox{if} & M_1 \geq M_{\rm crit} .\\
\end{array}
\right. \label{EQ9.2}
\end{equation}
Note that in the original Kroupa IBP, $\varepsilon = 1$ by construction. Since we aim to reduce
the fraction of short-period low-mass binaries with equal mass components, we allowed $\varepsilon$
(efficiency of secondary mass increase) to be smaller than unity, its value heeding to be determined
as for $\lambda_0$, $\eta$, and $M_{\rm crit}$ . This is a reasonable assumption if one realises
that the change in the eccentricity need not be coupled with the change in the secondary mass. They are
in principle two distinct physical process. In fact, as we will see, the efficiency for changing the
eccentricity is different from that associated with the changes in the secondary mass.
Note also that we introduced a random initial mass ratio in the range of $(0.9,1.0)$, when
$\rho_0$ is greater than unity and $M_1$ is less than $M_{\rm crit}$.
This was done to account for
a stochastic component in the accretion process of the secondary.
Again, this improvement was motivated by present-day CMDs which exhibit
pronounced high-q binary sequences
caused by G-dwarfs and K-dwarfs with mass ratio
near 1.0. With such a modification, short-period low-mass binaries unlikely have
components with nearly the same mass.
Finally, note that in this procedure, for convenience, we kept $\chi$ in Eq. \ref{EQ4} fixed
and given by $3/4$ as in the original Kroupa IBP. We calculated more than 200 models varying
conveniently the above-mentioned parameters with the following values
\begin{itemize}
\item $\lambda_0$: \hspace{0.02cm} 20, 28, 36 ,
\item $M_{\rm crit}$: \hspace{0.02cm} 1.0, 1.5, 2.0 ,
\item $\eta$: \hspace{0.02cm} 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, 2.00 ,
\item $\varepsilon$: \hspace{0.02cm} 0.25, 0.30, 0.50.
\end{itemize}
Given each model, we used properties of late-type binaries that are not strongly
affected by dynamics, mainly mass-ratio and eccentricity distribution of short-period binaries,
to compare with observational data of G-dwarf binaries in the Galaxy \citep[][]{DM_1991},
using the minimum $\chi^2$ method of model-fitting. Our best-fit model has the following values:
\begin{equation}
\left\{
\begin{array}{lcl}
\lambda_0 & = & 28 , \\
M_{\rm crit} & = & 1 \ {\rm M}_\odot , \\
\eta & = & 1/4 , \\
\varepsilon & = & 1/2 . \\
\end{array}
\right. \label{EQ10}
\end{equation}
Note that $M_{\rm crit} = 1 \ {\rm M}_\odot$ and $\lambda_0 = 28$
which implies that our modifications are effective only for G, K and M-dwarfs.
For the rest of the low-mass binaries, the modified Kroupa IBP is exactly
the same as the original Kroupa IBP.
Note also that $\varepsilon = 1/2$ which implies that the efficiency in the change
of the mass ratio is half in comparison with the change in the eccentricity, for
binaries whose primary mass is smaller than $M_{\rm crit} = 1 \ {\rm M}_\odot$.
Finally, since $\lambda_0 = 28$ and $\eta = 1/4$,
the $\rho$ as defined in Eqs. \ref{EQ4} and \ref{EQ8} is a smooth, continuous and monotonically
non-decreasing function of the primary mass $M_1$, in the range where pre-main-sequence eigenevolution
should be applied.
In Section \ref{results} we show the results for the modified Kroupa
IBP that will be described in the following sections.
\subsection{Low-mass binaries}
\label{latetype2}
To obtain the initial population for low-mass binaries in our revised
pre-main-sequence eigenevolution procedure, we follow exactly the same approach as that
to generate the original Kroupa IBP. The only difference in the modified
Kroupa IBP is the different equations that are used.
We first generate the birth population with distributions (i--iv) in
Section \ref{birth1}. Then we convert the birth population to the
initial population using Eqs. \ref{EQ8}, \ref{EQ4}, \ref{EQ5},
\ref{EQ9}, and \ref{EQ7}, with the parameter values given by Eq. \ref{EQ10}.
In Fig. \ref{FIG01} we plot the main distributions associated with the modified Kroupa
IBP for low-mass binaries, i.e. primary mass (top left-hand panel), mass ratio
(top right-hand panel), period (bottom left-hand panel), and eccentricity (bottom
right-hand panel). Note that the original Kroupa IBP (Section \ref{latetype1})
was included for comparison.
Note that the main difference between the original and the modified Kroupa IBPs
is the mass ratio distribution.
Indeed, in the modified Kroupa IBP the fraction
of binaries in the range of $q \in [0.9,1.0]$ is reduced due to the revised $\lambda$
(Eq. \ref{EQ8}) and the revised secondary accretion process (Eq. \ref{EQ9}).
\subsection{High-mass binaries}
\label{earlytype2}
In contrast with the original Kroupa IBP, for high-mass binaries,
we adopt observational distributions derived for O and B-dwarfs
to generate the birth population. Similarly to the original Kroupa
IBP, no pre-main-sequence eigenevolution is applied in our modified version.
We adopt the distributions derived by \citet{Sana_2012} who
analysed the O star population of six nearby Galactic open
stellar clusters. Note that the binarity found by \citet{Sana_2012}
is not 100 per cent and, since we assume, as usual, that star clusters
are formed with a dominant population of binaries, i.e. with $\approx$ 100
per cent of binaries, we should extend those distributions and normalise them
in order to take this into account.
Following \citet{Oh_2015}, the period distribution is given by
\begin{equation}
f_P \ = \ 0.23 \times \log_{10}(P/{\rm days})^{-0.55} , \label{EQ11}
\end{equation}
where $\log_{10}(P/{\rm days}) \in [0.15,6.7]$. This distribution and
range ensure that the cumulative binary fraction becomes unity.
In an analogous way, we extend the eccentricity distribution up
to unity which results in the following distribution
\begin{equation}
f_e \ = \ 0.55 \times e^{-0.45} , \label{EQ12}
\end{equation}
where $e \in [0.0,1.0]$.
Regarding the mass ratio, \citet{Sana_2012} found a uniform distribution, i.e.
\begin{equation}
f_q \ = \ 1.0 , \label{EQ13}
\end{equation}
where $q \in [0.1,1.0]$.
We emphasize that the correct way of pairing birth low-mass and high-mass binaries and preserving
the IMF is as follows. First, we generate an array of all stars (twice the number of binaries)
from the IMF. Second, we select the more massive star (primary) from the array of stars, a mass
ratio from the uniform distribution, and then compute the `ideal' secondary mass. After that,
the star that has the closest mass to the `ideal' secondary mass is chosen from the array to be paired with
the primary. Finally, both stars are removed from the array. Then, we proceed with the binary
generation until the primary mass is smaller than 5 M$_\odot$. After that point, the above procedure
remains the same with the exception that the secondary is chosen randomly in the array.
Finally, as done by \citet{Kroupa_1995b} for low-mass binaries,
the upper envelope and circularization period in the plane $e$ vs.
$\log_{10}(P)$ for high-mass binaries have to be consistent with
to observational data. Here, we applied the criterion derived in
\citet{Moe_2016} who considered observational data of O and B-dwarfs.
This guarantees that the binary components do not fill their Roche-lobe
by a factor greater than 70 per cent at the pericentre and it is achieved
by the following restriction
\begin{equation}
e_{\rm max} \ = \ 1 \ - \ \left( \frac{P}{P_{\rm circ}} \right)^{-2/3} , \label{EQ14}
\end{equation}
where $P > P_{\rm circ}$, and the circularization period is given by
$P_{\rm circ} = 2$ days.
In Fig. \ref{FIG02}, we show the main distributions associated with the modified Kroupa
IBP with respect to high-mass binaries, i.e. primary mass (top left-hand panel), mass
ratio (top right-hand panel), period (bottom left-hand panel), and eccentricity
(bottom right-hand panel). We also include the distributions of the original Kroupa IBP
(Section \ref{earlytype1}) for comparison, as in Fig. \ref{FIG01}.
We emphasize here that a way to determine the birth high-mass binary
distributions would be to investigate if the assumed hypothesis leads
to the observed number of runaway massive stars through dynamical mass
segregation to the cluster core and partner exchanges through
dynamical encounters there between the massive stars \citep{Kroupa_INITIAL}.
This was partially done by \citet{Oh_2016} who showed that the birth mass
ratio distribution for O-star primaries must be near uniform for mass
ratios greater than $0.1$, consistent with observational results of
\citet{Sana_2012}. Finally, as massive stars require their own process of
converting the birth population into the initial population,
we assume here that such a process has already taken place and
we adopt directly observational distributions, which is a reasonable approach
considering the lack of understanding in this process. This is
also consistent with the fact that this process should be very fast for massive
stars \citep{Railton_2014}.
\section{RESULTS}
\label{results}
In this section, we present the main results associated with the
modified Kroupa IBP in a comparative way with the original Kroupa
IBP. We show results for Galactic field late-type binaries and
GCs. We end this section by showing that our modified
Kroupa IBP solves the problem associated with present-day GC CMDs.
\subsection{Stimulated evolution}
\label{stimulated}
In order to compare our results with Galactic
field binaries, we assume as usual that Galactic field binaries
come from the dissolution of dense star clusters, after stimulated evolution \citep{Kroupa_1995b}.
Such a dissolution is assumed to be caused mainly by the residual
gas removal due the evolution of massive stars (winds and, eventually, supernovas).
For the embedded clusters, we assume that initial clusters follow
the Marks-Kroupa relation between half-mass radius and embedded mass of stars
in the embedded cluster \citep{Marks_2012}, i.e.
\begin{equation}
\left( \frac{R_h}{{\rm pc}} \right) \ = \ 0.1 \, \left( \frac{M_{\rm ecl}}{{\rm M}_\odot} \right)^{0.13} , \label{EQ15}
\end{equation}
where $R_h$ is the cluster initial half-mass radius and $M_{\rm ecl}$
is the stellar mass in the embedded cluster.
Even though this is the most accurate approach, this is not doable
yet with the current version of the {\sc mocca} code, because the expansion
of the embedded cluster due to residual gas removal, which takes place after few $10^5$ yr
of cluster evolution \citep[e.g.][]{Banerjee_2017}, is not yet implemented in {\sc mocca}.
In order to mimic this expansion and residual gas removal,
we performed many numerical experiments to look for an initial expansion
factor in the Marks-Kroupa relation to provide good agreement with observations.
In order to find the best expansion factor, we assumed five different values,
namely 1, 2, 5, 7, and 10, that should be multiplied by the half-mass radius in
Eq. \ref{EQ15} to provide appropriate input for {\sc mocca} models. In addition,
since the stimulated evolution varies with the cluster mass, we also verify which
expansion factor gives better results for combinations of stimulated evolution
timescales and cluster mass. We assumed three different timescales, namely 1, 5 and 10 Myr,
and four different cluster masses, which is given in {\sc mocca} by the number of objects
(single stars + binaries),
namely 10k, 20k, 50k and 100k. Note that these different timescales and
cluster masses allow us to infer the strength of the dynamical processing in
shaping the distributions through stimulated evolution.
All these models were evolved and compared with the observational mass-dependent binary
fraction \citep[e.g.][their Fig. 2]{Dorval_2017} and properties of Galactic field late-type
binaries (Section \ref{gflatetype}).
The best agreement with observations was
achieved for an expansion factor of $\approx$ 2 -- 4, for a stimulated evolution
timescale of $\approx$ 5 Myr. This is also consistent with the results found by \citet{Pelupessy_2012}
and \citet{Brinkmann_2016} who showed, using direct $N$-body calculations, that embedded clusters expand by approximately
this factor, if the star-formation efficiency is $\approx$ 1/3, i.e. if $\approx$ 2/3 of the initial gas
is lost in the star cluster formation process.
The properties of this MOCCA model are given in Table \ref{TabMOCCA}.
\begin{table}
\caption{Initial conditions of the best-fit model, which was used in the stimulated evolution,
obtained by averaging over all 10 models. See Section \ref{stimulated}, for more details.}
\label{TabMOCCA}
\begin{adjustbox}{max width=530px}
\noindent
\begin{threeparttable}
\noindent
\begin{tabular}{l|l}
\hline\hline
Mass [${\rm M}_\odot$] & $(9.23 \pm 0.16) \times 10^3$ \\
Number of objects & $10^4$ \\
Binary fraction\tnote{a} \, [\%] & $95$ \\
Metallicity [Z$_\odot$] & $0.02$ \\
Central density [${\rm M}_\odot \; {\rm pc}^{-3}$] & $(1.5 \pm 0.5) \times 10^4$ \\
Tidal radius [pc] & $26.0$ \\
Half-mass radius [pc] & $0.86 \pm 0.01$ \\
Core radius [pc] & $0.28 \pm 0.01$ \\
Galactocentric distance [kpc] & $8.01 \pm 0.07$ \\
Half-mass radius relaxation time [Myr] & $32.5 \pm 0.7$ \\
Central velocity dispersion [km s$^{-1}$] & $5.1 \pm 0.7$ \\
Duration of stimulated evolution [Myr] & $4.9 \pm 0.5$ \\
\hline\hline
\end {tabular}
\begin{tablenotes}
\item[a] We set the initial binary fraction different from 100 per cent in order \\
to avoid computational problems that arise in MOCCA if there is no \\
single star in the initial model.
\end{tablenotes}
\end{threeparttable}
\end{adjustbox}
\end{table}
Assuming an expansion factor of $\approx$ 3, the initial half-mass radius of {\sc MOCCA} models should be
\begin{equation}
\left( \frac{R_{h,\rm MOCCA}}{{\rm pc}} \right) \ \approx \ 0.3 \, \left( \frac{M_{\rm ecl}}{{\rm M}_\odot} \right)^{0.13} , \label{EQ16}
\end{equation}
Note that \citet{Marks_2011} found that 50 per cent of all Galactic field binaries
originate from clusters with $M_{\rm ecl} \lesssim$ 300 M$_\odot$. This in principle
might be a problem due to the fact that {\sc mocca} models should contain at least
10k objects (single stars + binaries) which gives, in general, masses of one order of magnitude higher than
that. However, in our experiments, models with 10k and 20k give similar results
and the 10k model might be used as an upper limit associated with the influence of dynamics
during stimulated evolution. Said that, we decided to show the results for
a cluster of 10k which gives not only such an upper limit, but also is the
closest possible {\sc mocca} model to a low-mass embedded cluster from which the majority
of Galactic field binaries originate.
The validity of this assertion is demonstrated in the Appendix, in which we show that
during the phase of the cluster expansion connected with residual gas
removal, binaries with periods longer than $\approx 10^{7}$ days are modified by
dynamical interactions. Therefore, only the tail of the observed period distribution can be
modified by dynamics before the binaries become a part of the field population.
{\it This {\sc mocca} model is then dynamically equivalent for the binary population to the embedded clusters which have an expansion phase.}
We define `{\it dynamical equivalence}' as follows: if two clusters with different masses and
different initial radii dynamically evolve an identical initial binary population to
similar distribution functions of binaries, then these two clusters are `dynamically
equivalent'. Cluster evolution need not be over the same time scales for dynamical
equivalence to be reached.
For instance, when comparing the results for the models in
\citet{Kroupa_1995b}, composed initially of 200 binaries and computed for a few hundred Myr, with the models in
\citet{Kroupa_2001} composed of 5,000 binaries and integrated for a Myr, we notice that both clusters
are dynamically equivalent. The corresponding binary distribution functions for these two clusters look
very similar. They started with the same binary-star distributions and were evolved in very different
cluster environments with and without stellar evolution (compare, for instance, fig. 7 in \citet{Kroupa_1995b}
with figs. 10 and 11 in \citet{Kroupa_2001}). In addition, \citet{Giersz_2016} showed
that {\sc mocca} and N-body model binary distributions are evolved in a remarkably similar way, not only their
period distributions (see their Fig. 2 which shows a comparison between {\sc mocca} model and {\sc BiPoS} code
\citep{Marks_2011}), but also for binding energy, mass ratio and eccentricity distributions.
We emphasize
here that dynamically equivalent models do not have exactly the same binary populations (i.e. same binary
distributions) after stimulated evolution. However, they are very similar and with this concept it is
possible to find general solutions to the class of dynamically equivalent clusters which lead to a very
similar final binary population after finding one N-body solution (sec. 4.3 and sec. 6.4 in \citet{Kroupa_1995a}
and sec. 4 in \citet{Kroupa_1995c})\footnote{
We would like to emphasize that using dynamical equivalence does not imply that a cluster evolves dynamically in the same way as another cluster. It merely means that, over a certain time scale, the dynamical evolution processes the binary population similarly.}.
In order to further illustrate this important concept, we show in the
Appendix \ref{ap2}, as an example, that three different initial {\sc mocca} models are dynamically
equivalent over different stimulated evolution timescales by showing their resulting binary distributions
(Fig. \ref{FIG_AP})\footnote{
We note that a demonstration of dynamical equivalence as defined here in the form of low-N dynamical simulations for the specific case studied in this work will be useful to further test this conjecture, whereby the small-N computations presented in \citet[][]{Kroupa_1995a,Kroupa_1995b,Kroupa_1995c} are the basis of formulating it. There, the scaling from one dynamically-equivalent solution to another is also discussed.}.
All the above implies that our approach is reasonable and that
{\sc mocca} models represent clusters which have a dynamically equivalent history to the real embedded clusters
which have undergone an expansion phase.
The stimulated evolution was performed as follows.
We evolved for $\approx$ 5 Myr two star cluster models with $M_{\rm ecl}$
$\approx 10^4$ M$_\odot$ and 95 per cent of binaries that follow the original
Kroupa IBP and the modified Kroupa IBP. For each cluster model, we performed
10 realizations in order to reduce fluctuation that arise from the small number
of binaries in the clusters ($\approx$ 10k).
In what follows, we show the results for our averaged (over 10 realizations)
above-described models following Eq. \ref{EQ16}.
\subsection{Galactic field late-type star distributions}
\label{gflatetype}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\linewidth]{./FIG05_1.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_2.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_3.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_4.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_5.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_6.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_7.eps}
\includegraphics[width=0.49\linewidth]{./FIG05_8.eps}
\end{center}
\caption{Comparison with Galactic field late-type binaries, after stimulated evolution,
for an averaged model evolved with the original Kroupa IBP (Section \ref{original}, solid line) and
an averaged model evolved with the modified Kroupa IBP (Section \ref{modified}, dashed line).
In panels (a) and (b), we plot the period distributions of G and M-dwarfs respectively.
In panels (c) and (d) we show the mass ratio distributions of short-period and long-period
G-dwarfs, respectively. Mass ratio distributions for M-dwarfs and all dwarfs such that
$M_1$ is smaller than 2 M$_\odot$ are shown in panels (e) and (f), respectively. Finally,
in panels (g) and (h), we exhibit the eccentricity distributions for short-period and
long-period G-dwarfs, respectively. Observational data is plotted with filled circles and
were extracted from \citet{DM_1991} (DM91), \citet{M_1992} (M1992), \citet{FM_1992} (FM92),
\citet{RG_1997} (RG97), and \citet{B_2010} (B2010). Vertical lines for the IBPs correspond to
the entire range of all 10 realizations outcomes, and for the data to Poisson errors.
For more details, see Section \ref{gflatetype}.}
\label{FIG05}
\end{figure}
One of our goals, apart from the solution of the problems presented in Section \ref{problem} related
to GC CMDs, is to not worsen the pre-main-sequence eigenevolution process with respect to late-type binaries in the Solar
neighbourhood, i.e. eccentricity, period and mass ratio correlations, and for short-period binaries, a
bell-shaped eccentricity distribution and a mass ratio distribution which rises towards unity.
These observational features were derived from long-term observations and here we compare both original
and modified Kroupa IBPs with distributions extracted from the following studies\footnote{Our
definitions for the late-type binaries are as follows: all late-type, G and M-dwarfs have primary
masses in the ranges $[0.08,2.0]$, $[0.8,1.2]$, and $[0.08,0.6]$, respectively, all in units of M$_\odot$.}:
\begin{itemize}
\item all G-dwarf period: DM91 \citep[][]{DM_1991},
\item all M-dwarf period: FM92 \citep[][]{FM_1992},
\item all G-dwarf eccentricity: DM91,
\item short-period G-dwarf mass ratio: M1992 \citep[][]{M_1992},
\item long-period G-dwarf mass ratio: DM91,
\item all M-dwarf mass ratio: B2010 \citep[][]{B_2010},
\item all late-type-dwarf mass ratio: RG97 \citep[][]{RG_1997}.
\end{itemize}
Note that we compare the mass ratio distribution of M-dwarf binaries with the data from
\citet[][]{B_2010}, that provide a similar distribution to that from \citet[][]{FM_1992}.
For G-dwarfs, we opted for not including the data from
\citet[][]{R_2010} for the following reason. As already pointed out by \citet{Marks_2011b},
\citet[][]{R_2010} found a mass-ratio distribution for binaries with a solar-type primary
that is different from \citet[][]{DM_1991}. Since the range of primary masses
in both works are similar, their mass-ratio distributions should be also similar if
the periods are comparable. The reason for that is likely due to different approaches
in both works. The survey performed by \citet[][]{DM_1991} lasted 13 years and they
found accurate orbital solutions for their binaries. On the other hand, \citet[][]{R_2010}
compiled data from different sources and techniques. Given that, we preferred
to compare our models with data from \citet[][]{DM_1991}\footnote{See also
the comments in \citet{Kroupa_2009}.}.
Additional observational works could be also included in comparison.
For instance, the results by \citet{Halbwachs_2003} who investigated early-type F- and K-dwarf binaries
and those by \citet{Fisher_2005} who presented an incomplete survey of spectroscopic binaries.
In order to use those results, one could apply the Kroupa IBP to this data set using stimulated evolution,
but only by also modelling the complicated biases and selection effects, which is, however, out of scope of
the present investigation. Finally, \citet{Rastegaev_2010} presented a survey of population II (sub-dwarf)
stars. The authors note that the population II period distribution is narrower and
biased towards shorter periods than the population I
distribution (their Fig. 10). Even though analysing population II stars is not an objective here,
it will be interesting to assume the universality hypothesis and then to investigate which population
of star clusters may, if at all, account for the observed population II binary distributions.
The mass function of metal-poor (population II) star clusters may have been different to that observed
today for population I. In such an investigation, it will be necessary to model all selection effects and
biases inherent to the analysis by \citet{Rastegaev_2010}. A particular aspect of such a modelling will be
the inclusion of population I stars and binaries which have acquired large proper motions,
e.g. by being ejected out of their population I birth clusters. Such a contribution is likely to be found
in a proper-motion selected sample in addition to population II, and may skew or alter the deduced binary
star distribution functions. Again, modeling population II via the Kroupa IBP and stimulated evolution
is not in the scope of this investigation, but is doable in future works.
In Fig. \ref{FIG05}, we compare the observational distributions used in this work with our averaged models,
after stimulated evolution, concerning the original Kroupa IBP and the modified Kroupa IBP.
Vertical lines in the figure correspond to the entire range of all 10 realizations outcomes
(for the IBPs), and to Poisson errors (for observational data).
Note that both IBPs provide good agreement with observations. In addition, both IBPs give
similar distributions with the exception of the mass ratio distributions.
In panel (c), we show the mass ratio distribution of short-period binaries whose primaries
are G-dwarfs. Note that the fraction of binaries with high mass ratio $q \gtrsim 0.9$ is reduced,
this is expected given our initial motivation for changing the pre-main-sequence eigenevolution prescription.
In panel (e) and (f) we show mass ratio distributions of all binaries whose primaries are
M-dwarfs and late-type-dwarfs, respectively. As for G-dwarfs, we also see a reduction in the
fraction of high mass ratio binaries. This is again expected due to the revised pre-main-sequence eigenevolution
derived here.
Finally, Fig. \ref{FIG05} clearly shows that none of the previous results achieved with the
original Kroupa IBP are damaged when our modified Kroupa IBP is adopted. In other words,
our modified Kroupa IBP is not only consistent with observational data of
late-type binaries in the Galactic field, but also provides
qualitatively similarly good description when compared with the original Kroupa IBP.
\subsection{Globular cluster colour-magnitude diagrams}
\label{gccmd}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\linewidth]{./CMD1N.eps}
\includegraphics[width=0.48\linewidth]{./CMD2N.eps}
\includegraphics[width=0.48\linewidth]{./CMD3N.eps}
\includegraphics[width=0.48\linewidth]{./CMD4N.eps}
\includegraphics[width=0.48\linewidth]{./CMD5N.eps}
\includegraphics[width=0.48\linewidth]{./CMD6N.eps}
\end{center}
\caption{Present-day CMDs from simulated observations of the six GC models
(Table 2) evolved with the modified Kroupa IBP (Section \ref{modified}).
The gray areas indicate the regions used to generate the colour distributions in Fig. \ref{FIG06.2}.
Note that there is no clear and pronounced binary sequence due
to short-period low-mass binaries with mass ratios of unity in such CMDs,
specially close to the turn-off.
For more detail see Section \ref{gccmd}.
}
\label{FIG06.1}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\linewidth]{./C1mod.eps}
\includegraphics[width=0.49\linewidth]{./C2mod.eps}
\includegraphics[width=0.49\linewidth]{./C3mod.eps}
\includegraphics[width=0.49\linewidth]{./C4mod.eps}
\includegraphics[width=0.49\linewidth]{./C5mod.eps}
\includegraphics[width=0.49\linewidth]{./C6mod.eps}
\includegraphics[width=0.95\linewidth]{./cummod.eps}
\end{center}
\caption{Colour distributions derived from the CMDs shown in Fig. \ref{FIG06.1},
following the procedure described in Section \ref{problem_model}. Notations as
in \ref{FIG03.2}.
Only stars with $(B-V)'>0.05$ are used to compute the above distributions.
Note that in all colour distributions in the six models following the modified
Kroupa IBP are are consistent uniform distributions. This is supported by the
statistical test, which does not allows us to reject the hypothesis that they are
uniform with more than 99 \% of confidence (see $p$-values).
For more detail see Sections \ref{gccmd}.
}
\label{FIG06.2}
\end{figure}
After showing that the modified Kroupa IBP provides good agreement
with observational data of Galactic field binaries (i.e. our modifications
do not worsen previously established constraints), we can
turn to the solution of our initial problem.
In Section \ref{problem}, we presented the motivation for our investigation
showing the present-day CMDs and colour distributions derived from them of
a set of six GC models with different initial and present-day properties.
All six models there were set with the original Kroupa IBP (Section \ref{original}).
While comparing with observation of the nearest GCs (M4 and NGC6397), we showed
that models set with the original Kroupa IBP are not able to predict realistic binary
distributions in GCs. This is because it provides not only a clearly visible binary
main-sequence (due to short-period low-mass binaries with mass ratios equal to
unity) which is not seen in the CMDs of real GCs, but it also predicts CMD colour
distributions below the turn-off peaked at the right edge of the distributions (associated
again with high-q binaries).
We can at this point perform a similar analysis as done in Section \ref{problem},
but taking into account our modified Kroupa IBP (Section \ref{modified}).
We thus simulate models with the same six initial cluster conditions and perform
photometry of present-day models. The six model, now set with the modified
Kroupa IBP, properties are shown in Table 2 and Fig. \ref{FIG06.1} exhibits
their present-day CMDs.
Note that present-day GC model properties are similar for both set
of models (compare with Table 1), which indicates that our modifications
do not play a key role in the cluster dynamical evolution. However,
while comparing Figs. \ref{FIG06.1} and \ref{FIG03.1} we do notice huge
differences.
First, that pronounced and clearly visible binary sequence
close and above the turn-off is not present in models set with the modified
Kroupa IBP. Second, we do not clearly see binary sequences below the
turn-off associated with high-q binaries in all models.
This is a result of our revised pre-main-sequence eigenevolution prescription which
gives a mass-dependent strength of the changes to binary properties (Section
\ref{method}). This leads to a reduced fraction of high mass ratio binaries
as well as a uniform spread in the mass ratio in the range between 0.9 and 1.0,
for those binaries that would have equal mass components.
Proceeding further, we generated colour distribution is the same way
described in Section \ref{problem_model}, which are shown in Fig. \ref{FIG06.2}.
Notice that all histograms seem to be consistent with uniform distributions,
as in observations. Again, we applied one-sample Kolmogorov-Smirnov tests
for uniformity. The $p$-values of these tests are displayed in the bottom
panel of Fig. \ref{FIG06.2}. The tests do not permit us to rule out the
null hypothesis that they are uniform with more than 99 per cent of confidence,
even though we can reject the null hypothesis for model 5 with
more than 98 per cent of confidence.
We conclude then that our modifications to the Kroupa IBP bring present-day
GC models closer to real GCs, which corresponds to a step forward towards a better
prescription for initial cluster conditions.
We emphasize that the solution to the problem with respect to present-day GC CMDs
(without worsening previous results)
presented here open the possibility that a universal IBP might be at the origin of the binary
population in both the Galactic field and in GCs.
This is consistent with the conclusions reached by \citet{Leigh_2015} and is important for a
better understanding of the star formation process in proto-clusters.
\section{CONCLUSIONS AND PERSPECTIVE}
\label{discussion}
The original description of pre-main-sequence eigenevolution of a birth
binary population by \citet{Kroupa_1995b}, while being consistent with
Galactic-field and open cluster binary-star data, is found here to
lead to too many $q=1$ systems which implies that the CMDs of GCs
ought to have pronounced binary sequences above the turn-off and to
the right of the MS. This is not observed to be the case. In this
contribution we revise th Kroupa pre-main-sequence eigenevolution
model such that it is consistent with the Galactic-field, open
cluster and GC data.
To achieve this, we assumed a mass-dependent strength for
pre-main-sequence eigenevolution such that the lower the primary mass,
the weaker the changes in the birth population. This procedure results in a
smaller fraction of short-period binaries with equal mass components which
provides qualitatively similarly good description of
observations of Galactic field late-type binaries and GCs.
We also assumed distributions for massive binaries such that they
follow directly distributions derived from observations of O-dwarfs in open
stellar clusters. This is consistent with the fact that pre-main-sequence
timescales are extremely short for massive stars.
Finally, we emphasize that our modified Kroupa IBP should not change results
achieved with the original Kroupa IBP with respect to young star clusters
and GCs. In addition, this paper corresponds to the first step towards a better description
of the IBP that should seed population synthesis codes (after dynamical population
synthesis) and star cluster evolution codes, as well as a better understanding of
clustered star formation processes such as energy and angular momentum redistribution
within very young binary systems.
In following investigations of the modified Kroupa IBP,
we will verify the influence of parameters that control
binary stellar evolution such as the energy budget during
the common-envelope phase and the angular momentum loss formalisms
for interacting binaries, on populations of white dwarf-main sequence
post-common-envelope phase binaries and cataclysmic variables.
\section*{Acknowledgements}
We would like to kindly thank Michael Marks for useful discussions and
suggestions. We would also like
to thank an anonymous referee for numerous comments and suggestions.
DB was supported by the CAPES foundation, Brazilian Ministry of Education
through the grant BEX 13514/13-0 and by the National Science Centre, Poland,
through the grant UMO-2016/21/N/ST9/02938.
AA would like to acknowledge support from the National Science Centre, Poland,
through the grant UMO-2015/17/N/ST9/02573 and partial support from Nicolaus
Copernicus Astronomical Centre's grant for young researchers.
\bibliographystyle{mnras}
|
\section{Introduction} Problems of material optimization governed by Maxwell's equation \corr{have recently been} studied in the literature. In particular, for time-harmonic electromagnetic fields we refer to \cite{Diaz2010}, where an optimal distribution of two materials with distinct properties was computed based on the \corr{so-called} SIMP approach \cite{Bendsoe1994}\corr{. This approach} was originally developed for the topology optimization of elastic structures and is based on interpolation between the desired material properties and an appropriate penalization scheme rendering undesired intermediate material properties unattractive with respect to the particular cost function. A similar technique has been applied to the transient problem discussed, \corr{for example} in \cite{Hassan2014}. Again the goal here was to find an optimal distribution of two isotropic materials. Potential applications of structural optimization techniques in the context of electromagnetics range from inverse problems, where distribution of material is reconstructed by the information given by the scattered electromagnetic fields \cite{OhinKwon2002}, to optimal material layout to improve the properties of optical devices \cite{Byun2004} or nanoparticles \cite{Pendry}. \par In this paper we are interested in a more general class of material optimization problems, in the framework of which a complex-valued permittivity tensor for a given point in the design domain is specified by a function of a finite number of parameters. Particular \corr{realizations have led} to problems of free material optimization \cite{Zowe1997,Greifenstein2016,Bendsoe1994,Ringertz1993}, which \corr{have so far} been studied solely in the context of linear elasticity, to optimal material orientation problems, \corr{see} e.g. \cite{Pedersen1989}, and to so called discrete multi-material optimization as treated in literature by so called DMO methods, see, e.g. \cite{Stegmann2005,Hvejsel2011}. Rather than formulating the optimization problems \corr{directly in the design parameters} and using a derivative based optimization algorithm like SNOPT \cite{Gill2005a} or MMA \cite{Svanberg1987} in a ``black box'' way, in this article, a new algorithmic concept for the solution of the envisaged class of design problems is developed. The motivation for the development of \corr{this} new solution approach is \corr{the fact that} the material tensors typically depend on the design parameters in a non-linear way and thus the parametrization may \corr{result in numerous} poor local optima, see \cite{Pedersen1989}, in which algorithms applied in a black-box way may \corr{become} trapped. \par In order to prevent this, the following concept is suggested: the \corr{principal} idea is to formulate the design problem \corr{directly in terms of the material tensors}, while the associated parametrization is hidden in the definition of the admissible set. Then, in the course of a sequential approximation algorithm, FMO-type models \corr{(see \cite{Stingl2009a})} of the objective \corr{as a} function of the material tensors are derived and \corr{are used to generate a sequence of sub-problems.} Due to the potentially non-convex parametrization each sub-problem is a constrained non-linear optimization problem, which may exhibit an unknown number of local optima. We show that based on the properties of the particular approximations these sub-problems can be solved to global optimality with a reasonable effort, partially with \corr{the} analytical solution, for important classes of parametrizations. \par The manuscript is structured as follows: In \cref{sec:prereq} the Helmholtz-type state equation, based on the time-harmonic Maxwells equation, is given in its weak formulation and the dependency on the material tensor is \corr{highlighted}. Then, in \cref{sec:optimization} the class of optimization problems of interest is stated\corr{,} including a detailed description of the general structure of the objective function as well as \corr{the} structure of the set of admissible materials, which is based on a graph. The discussion is continued \corr{with} a short note on the discretization of the state and the optimization problem as well as regularization issues. \corr{\Cref{sec:algorithm}} \corr{constitutes} the heart of this article. Based on convex first-order hyperbolic approximation as well as a \corr{so-called} sequential global programming technique, an optimization algorithm is \corr{stated for} which a global convergence result can be established. Subsequently, parametrization-dependent solutions \corr{to the} sub-problems taking the graph-structure of the admissible set into account are derived. To show the capabilities of the algorithm\corr{,} two examples are discussed in \cref{sec:example}. These \corr{include both} the design of a cloaking for a scatterer made from an increasing number of anisotropic materials \corr{and the} tomographic reconstruction of an unknown material distribution consisting of a background material, a dielectric and an absorbing material. \par Throughout this paper, we indicate by $\SC$ the space of symmetric, two\hyp{}dimensional and complex-valued tensors. The term \corr{$\langle A,B \rangle := \Re(\tr(A^H B))$ denotes the standard scalar product in $\SC$ and $\| A \|^2_F = \langle A,A\rangle$ denotes the induced Frobenius norm}. \par For a real-valued continuously differentiable function $v:\SC \to \R$ we define the derivative of $v$ in a direction $Y\in \SC$ with respect to $B$ as \begin{align*} \del{v(B)}{B}[Y] := \lim_{\nu \to 0} \frac{v(B+\tau Y) - v(B)}{\nu}. \end{align*} We note that for the tuple $\B\in \SCK$ with $\noDesign\in \N$ and a real-valued continuously differentiable function $v:\SCK \to \R$\corr{, the} directional derivative of $v$ in direction $\Y \in \SCK$ with respect to $\B$ is given as \begin{align*} \del{v(\B)}{\B}[\Y] = \sum_{i=1}^\noDesign \del{v(\B)}{(\B)_i}[(\Y)_i]. \end{align*} Finally, we define the extended norm $\|\B\|^2_{F^\noDesign} := \sum_{i=1}^\noDesign \|(\B)_i\|^2_F$ for $\B\in \SCK$.\section{Prerequisites}\label{sec:prereq} The propagation of electromagnetic waves is described by Maxwell's equation \cite{Jackson1963}. In this paper we restrict ourselves to the time-harmonic propagation of so-called transverse magnetic waves (TM) for a given wavenumber $\omega$, where we assume that the electromagnetic field is given by a scalar function depending only on two spatial dimensions. With these assumptions, Maxwell's equation simplifies to the Helmholtz equation for the magnetic field. \par The relative permittivity $\eps$, which in this article is the material property of interest, is a complex- and tensor-valued function of space. For modeling purposes an additional tensor valued function $B$ is introduced\corr{,} whose values are given by the inverse of the permittivity \corr{at} each point. In general, we assume that the material tensor at a point is symmetric, \ie $B: \R^2 \to \SC$. For scattering applications an incident magnetic field $u_I:\R^2\to \C$ is given, which solves Maxwell's equation for the given background material $B_b: \R^2 \to \SC$. \corr{The} Helmholtz equation is actually defined \corr{on the whole of} $\R^2$, thus we introduce a perfectly matched layer (PML) \cite{Berenger1994} surrounding the domain of interest\corr{,} including the scattering object. \par \begin{figure}[t]\centering \includegraphics{computational_domain.pdf} \caption{General geometrical setting: PML, background and design domain} \label{fig:domains} \end{figure} \par The computational domain $\Omega = \Omega_D \cup \Omega_C\subset\R^2$ is subdivided into a design domain $\Omega_D$ and a non-design domain $\Omega_C$. \corr{The} non-design domain $\Omega_C = \Omega_\PML\cup \Omega_b \cup \Omega_P$ \corr{in turn} consists of three subdomains. The perfectly matched layer $\Omega_\PML$ \corr{completely} encloses the background domain $\Omega_b$ and both are equipped with background material tensor-valued function $B_b$. Moreover\corr{,} the scatterer domain $\Omega_P$ and the design domain $\Omega_D$ are both embedded in the background domain $\Omega_b$ (see \cref{fig:domains}). The tensor valued function $B_P:\Omega_P\to\SC$ in the scatterer domain is assumed to be independent of the design, whereas the tensor function $B:\Omega_D\to \SC$ associated with the design domain $\Omega_D$ will be subject to optimization. For the sake of notation, we combine both functions to the piecewise tensor-valued function $B_C:\Omega_C\to \SC$ with \begin{align*} B_C & = \begin{cases} B_b & \text{ in } \Omega_b \cup \Omega_\PML, \\ B_P & \text{ in } \Omega_P. \end{cases} \end{align*} Using this\corr{,} we state the Helmholtz equation in weak form: \begin{align}\label{eqn:weakbilinear} {\text{Find } u \in H^1_0(\Omega,\C) \ \st } & a(B;u,\phi) + a_C(u,\phi) = l(B;\phi)+l_C(\phi) \quad \text{ for all } \phi\in H^1_0(\Omega)\corr{.} \nonumber \end{align} Here, we explicitly point out the dependency on $B$ and note the subdivision of the bilinear and linear forms into design domain contributions \begin{align*} a(B;u,\phi) & = \int_{\Omega_D} B \nabla u \cdot \nabla \phi - \omega^2 u\phi \dx, & l(B;\phi) & = -\int_{\Omega_D} B\nabla u_I\cdot \nabla \phi \dx \end{align*} and non-design domain contributions \begin{align*} a_C(u,\phi) & = \int_{\Omega_C} B_C A^\omega_\eps \nabla u \cdot \nabla \phi - \omega^2 A^\omega_\mu u\phi \dx, \\ l_C(\phi) & = \int_{\Omega} B_b\nabla u_I\cdot \nabla \phi \dx-\int_{\Omega_C} B_C \nabla u_I\cdot \nabla \phi \dx. \end{align*} By the definition of $B_C$ we observe that the right hand side of \eqref{eqn:weakbilinear} \corr{vanishes} in $\Omega_b\cup \Omega_\PML$. The functions $A^\omega_\eps$ and $A^\omega_\mu$ describe the wavelength dependent PML \cite{Monk2003} based on a squared layer with distance $d$ from the origin and are defined as follows: \begin{align*} A^\omega_\epsilon(x,y) &= \begin{pmatrix} \frac{s(y)}{s(x)} & 0 & \\ 0 & \frac{s(x)}{s(y)} \end{pmatrix}, & A^\omega_\mu(x,y) & = s(x) s(y), & s(t) &= 1 - \frac{ \sigma_0 \max(0,|t|-d)}{\imath\omega}. \end{align*} Particular choices of the positive scalar $\sigma_0$ and $d$ depend on the particular application, see \cref{sec:example}. The definition of $s$ implies that $A^\omega_\epsilon\equiv\mathds1$ and $A^\omega_\mu \equiv 1$ in $\Omega \setminus \Omega_\PML$. \section{A general material optimization problem}\label{sec:optimization} \par \begin{figure}\centering \scalebox{0.7}{ \includegraphics[]{graphstructure.pdf} } \caption{Example graph of admissible set $\GraphSet$}\label{fig:graph} \end{figure} \par We start with the description of the set of admissible material tensors $\GraphSet$, which is structured by a graph $(V,E)$ with vertices $V$ and edges $E = \{e_1,\dotsc,e_\noEdges\} \subset V\times V$. We assume that every vertex $v\in V$ is part of at least one edge and is associated with a predefined material tensor $B^{(v)}\in \SC$, cf. \cref{fig:graph}. We further define $\NodeImage:=\{B^{(v)}\mid v\in V\}$, $\indexEdges := \{1,\dotsc, \noEdges\}$ and introduce the following parametrization: \begin{definition}[parametrization of $\GraphSet$]\label{ass:properties} We call the mapping $\Param: \indexEdges \times [0,1] \to \SC$ parametrization of $\GraphSet$ iff the following holds: \begin{itemize} \item $\Param$ is twice continuously differentiable with respect to the second variable. \item interpolation property: \begin{equation*} \Param(l,0) = B^{e_l^{(1)}} \text{ and } \Param(l,1) = B^{e_l^{(2)}} \quad \forall l\in\indexEdges, \end{equation*} where $e_l^{(1)}$ and $e_l^{(2)}$ denote the first and second node of the $l$-th edge, respectively. \item $\Param$ is injective on $\indexEdges\times(0,1)$, \ie \begin{equation*} \forall (k,s),(l,t)\in \indexEdges\times(0,1): \Param(k,s) = \Param(l,t) \Rightarrow (k,s)=(l,t). \end{equation*} \end{itemize} \end{definition} We denote \corr{the image generated by the parametrization on the $l$-th edge by $\EdgeImage_l$}, \ie \begin{align*} \EdgeImage_l = \{\Param(l,\delta)\mid \delta \in [0,1]\}, \quad \forall l\in\indexEdges. \end{align*} Thus the set of admissible material tensors can be written as \begin{align*} \GraphSet = \{ \Param(l,\delta) \mid (l,\delta) \in \indexEdges\times [0,1] \} = \EdgeImage_1 \cup \dotsb \cup \EdgeImage_\noEdges \end{align*} and the set $\rmG$ of admissible tensor-valued material distribution reads as \begin{equation*} \rmG = \{ B :\Omega_D \to \SC \mid B(x) \in \GraphSet \text{ for } x \in \Omega_D\}. \end{equation*} Particular choices of $\rmG$ are given in \cref{sec:example}. In the following, we use the notation $\Param_l(\delta):=\Param(l,\delta)$ for all $l\in \indexEdges, \delta\in [0,1]$. Using this, we state an optimal design problem of tensor-valued coefficients over a Helm\-holtz-type equation in the domain $\Omega$ as follows: \begin{equation} \label{eqn:opt:prob}\left\{\quad \begin{aligned} \min_{B\in \rmG} ~ & \Jphys(B, u) + \eta \Jreg(B) + \gamma \Jgray(B) \\ \st~& u\in H^1_0(\Omega,\C) \text{ is the solution of } & \\ & \begin{multlined} a(B;u,\phi) + a_C(u,\phi) = l(B;\phi)+l_C(\phi) \quad \text{ for all } \phi\in H^1_0(\Omega) \end{multlined} \end{aligned}\right. \end{equation} The real-valued function $\Jphys$ is called \corr{the} objective functional and \corr{is} assumed to be Gateaux-differentiable on $\rmG \times H^1_0(\Omega,\C)$ with respect to \corr{both the} complex- and tensor-valued material distribution $B$ and the state variable $u$. \par Moreover\corr{,} $\Jreg$ and $\Jgray$ are Gateaux-diffentiable functionals \corr{that} penalize irregular and undesired material distributions $B$, respectively, and $\eta$ and $\gamma$ are non-negative scalars. We refer to \cref{sec:reg,sec:subprobsolve} for precise definitions. \subsection{Discretization}\label{sec:discrete} Let $\Triang$ be a regular triangulation of $\Omega$ with $\noElements$ triangular elements $\TriI{i}$, $1\le i\le \noElements$\corr{, where} the first $\noDesign$ triangles are located in the design domain $\Omega_D$. We assume that all subdomains in $\Omega$ are exactly approximated by the triangulation. \par The tensor-valued function $B$ is assumed to be constant on each triangle $\TriI{i}$, $i\in \indexDesign:=\{1,\dotsc,\noDesign\}$ of the design domain, \ie $B(x)| _{\Omega_D} = B_h(x):= \sum_{i=1}^\noDesign (\B)_i \chi_{\TriI{i}}(x)$ with a tuple $\B\in \SCK $ and the characteristic function $\chi_{\TriI{i}}$ of triangle $\TriI{i}$. Furthermore\corr{,} the state $u:\Omega\to \C$ is approximated by $u_h(x):=\sum_{i=1}^\noDoF (\u)_i \phi_i(x)$ with $\noDoF$ degrees of freedom entering the coefficient vector $\u \in \C^\noDoF$. In summary\corr{,} the optimization problem \eqref{eqn:opt:prob} is approximated by \begin{align}\label{eqn:opt:probTMdis} \left . \begin{aligned} \min_{\B \in \AdmSet} ~ & \Jphysdisc(\B, \u ) + \eta \Jregdisc(\B) + \gamma \Jgraydisc(\B) \\ \st ~ & (\SYS(\B) + \SYS_C ) \u = (\RHS(\B) + \RHS_C) \end{aligned}\qquad \right. \end{align} with the admissible set $\AdmSet := (\GraphSet)^\noDesign$ and discretized versions of the objective functional $\Jphysdisc(\B,\u) := \Jphys(B_h,u_h)$, regularization functional $\Jregdisc(\B):=\Jreg(B_h)$ and penalization functional $\Jgraydisc(\B):=\Jgray(B_h)$. The PDE constraint is approximated by a system of equations with the symmetric matrices $\SYS(\B),\SYS_C\in \C^{\noDoF\times\noDoF}$, which are defined entry-wise by \begin{align*} (\SYS(\B))_{ij} & = \sum_{k=1}^\noDesign\int_{\TriI k} B_k \nabla \phi_i \cdot \nabla \phi_j -\omega^2 \phi_i\phi_j \rmd x , & \text{for all } 1\le i,j \le \noDoF \\ (\SYS_C)_{ij} & = \int_{\Omega_C} B_C A^\omega_\epsilon \nabla \phi_i \cdot \nabla \phi_j - \omega^2 A^\omega_\mu \phi_i\phi_j \rmd x & \text{for all }1\le i,j \le \noDoF. \end{align*} Finally, the right hand side vectors $\RHS(\B),\RHS_C \in \C^\noDoF$ are defined by \begin{align*} (\RHS(\B))_{i} & = \sum_{k=1}^\noDesign\int_{\TriI k} B_k \nabla u_I\cdot \nabla \phi_i \rmd x & \text{for all } 1\le i \le \noDoF \\ (\RHS_C)_{i} & = \int_{\Omega_C} B_C\nabla u_I\cdot \nabla \phi_i \rmd x - \int_{\Omega} B_b \nabla u_I\cdot \nabla \phi_i \rmd x & \text{for all } 1\le i,j \le \noDoF. \end{align*} We note that the solution $\u$ of the discretized state problem \begin{align}\label{eqn:state:algebraic} (\SYS(\B) + \SYS_C ) \u = \RHS(\B) + \RHS_C \end{align} is uniquely defined by the material tuple $\B \in \AdmSet$ and thus we can rewrite the discretized optimization problem \eqref{eqn:opt:probTMdis} as \begin{align}\label{eqn:opt:probTMdisred} \min_{\B\in\AdmSet}~\Jtotal(\B) \tag{\text{$P^{\gamma}_h$}} \end{align} with the objective functional $\Jtotal(\B) := \Jphysdisc(\B, \u(\B)) + \eta \Jregdisc(\B) + \gamma \Jgraydisc(\B)$, where $\u(\B)$ is the unique solution of \cref{{eqn:state:algebraic}}. For later use\corr{,} we define $\nonsep(\B):= \Jphysdisc(\B, \u(\B)) + \eta \Jregdisc(\B)$ and \corr{briefly} note that the derivative of $\Jphysdisc(\B,\b u( \B))$ with respect to $\B$ can be computed by adjoint calculus or \corr{using the} implicit function theorem.\subsection{Regularizations}\label{sec:reg} In this section we present a possible regularization on the material distribution $B:\Omega_D\to \SC$. \corr{To do this,} we define the filtered material distribution \begin{align*} \tilde B(x) := \frac{\int_{\Omega_D} \kappa(x-y)B(y) \rmd y}{\int_{\Omega_D} \kappa(x-y)\rmd y} \end{align*} with a filter kernel $\kappa:\R^2\to\R$. For instance $\kappa$ can be defined as a "circular filter" $x\mapsto \max(0,r_0-\|x\| )$ with filter radius $r_0$. \corr{In this way,} we define a tracking-type filter regularization term on the design domain $\Omega_D$ \begin{align*} \Jreg(B) = \int_{\Omega_D} \big \| B(x) - \tilde B (x) \big \|_F^2 \dx. \end{align*} After discretization, the regularization term can be \corr{expressed in} quadratic form \begin{equation*} \Jregdisc(\B) = \sum_{i,j=1}^2 \B_{ij}^H M \B_{ij} \end{equation*} where $\B_{ij} := (((\B)_1)_{ij},\dotsc,((\B)_\noDesign)_{ij})$. The particular form of the symmetric matrix $M\in \R^{\noDesign\times \noDesign}$ depends on $\kappa$ and on the chosen quadrature rule. Again for later use, the directional derivative with $\Y\in \SCK$ and $\B\in\SCK$ of the discretized regularization term reads for all $k\in\indexDesign$ \begin{equation*} \del{\Jregdisc(\B)}{(\B)_k}[(\Y)_k] = 2 \Re \left (\sum_{i,j=1}^2 ((\Y)_k)_{ij} \sum_{l=1}^\noDesign M_{kl} ((\B)_l)_{ij} \right). \end{equation*} It is well known that for the existence of a solution of \eqref{eqn:opt:prob} in infinite dimensions, \corr{an $L^2$-regularization is generally not,} sufficient. Nevertheless\corr{,} we see the desired regularization effect in the discretized setting, compare also \corr{with} \cref{sec:example}. We finally note that we do not apply a standard filter scheme as in \cite{Haber1996a,Sigmund1997a}, because the filtered material distribution $\tilde B$ is typically not contained in the admissible set $\GraphSet$. \section{Optimization Algorithm}\label{sec:algorithm} Throughout this section we would like to derive an optimization algorithm which takes the specific structure of problem \cref{eqn:opt:probTMdis} into \corr{consideration}. We proceed as follows: we first define suitable separable approximations of the non-separable functions $\nonsep(\B)$ defined in the previous section. On the basis of these we define a series of sub-problems and the so-called sequential global programming algorithm. \corr{We then} derive a global convergence result for the latter and show how we can efficiently solve non-convex separable sub-problems for two specific material parametrization schemes. \par We \corr{begin with some additional} notation: let $v:\SCK \to \R$ be continuously differentiable on a subset $\AdmSet\subset \SCK$. For all $i \in \indexDesign$ we define real and imaginary differential operators entry-wise by \begin{align*} (\nabla ^{i,R} v(\B))_{k,l} & := \del{v(\B)}{(\B)_i} [\boldsymbol e_k \boldsymbol e_l^T], & (\nabla ^{i,I} v(\B))_{k,l} & := \del{v(\B)}{(\B)_i} [\imath (\boldsymbol e_k \boldsymbol e_l^T)] \quad \forall 1 \le k,l\le 2, \end{align*} where $\boldsymbol e_l$ denotes the $l$-th standard basis vector. \par We use the notation $Y^R = \Re(Y)$ and $Y^I = \Im (Y)$ for any complex-valued tensor $Y$. \corr{This together with the differentiability of $\nonsep$ yields} the directional derivative in direction $\Y\in\SCK$ at $\B\in\SCK$ of the reduced functional \corr{$\nonsep$} \begin{align*} \del{\nonsep(\B)}{\B}[\Y] = \sum_{i=1}^K \langle \nabla ^{i,R} \nonsep(\B),(\Y)_i^R\rangle + \langle \nabla ^{i,I} \nonsep(\B),(\Y)_i^I\rangle. \end{align*} \subsection{Generalized Convex Approximation} We briefly recapitulate a \corr{number of} results from \cite{Stingl2009a}\corr{,} starting with the definition of a convex first-order approximation. \begin{definition}[convex first-order approximation]\label{defn:firstorder} We call an approximation $g:\SCK \to \R$ of a continuously differentiable function $\nonsep:\SCK \to \R$ a convex first-order approximation at $\bar \B =(\bar B_1,...,\bar B_\noDesign) \in \SCK$, if the following assumptions are satisfied \begin{enumerate}[label=\alph*),itemindent=0.5cm] \item $g(\bar \B) = \nonsep(\bar \B)$ \item $\del{g(\bar \B)}{B_i} = \del{\nonsep(\bar \B)}{B_i} $ for all $i\in \indexDesign$ \item $g$ is convex \end{enumerate} \end{definition} \begin{definition}[hyperbolic approximation]\label{defn:hyper} Let $\nonsep:\SCK\to \R$ be continuously differentiable on $\AdmSet\subset \SCK$ and $\bar {\B} = (\bar B_1,\ldots \bar B_\noDesign) \in \AdmSet$. Moreover\corr{,} let asymptotes $L=l\mathds1\in \SC$ and $U=u \mathds1\in \SC$ be given such that \begin{equation*} \Re(L)\prec \Re(Y) \prec \Re(U)\quad \text{and} \quad \Im(L) \prec \Im(Y) \prec \Im(U) \end{equation*} for all $Y \in \GraphSet$. Let $\tau$ \corr{be} a non-negative real parameter. Then we define the hyperbolic approximation $\hyperApprox$ of $\nonsep$ at $\bar {\B}$ as \begin{align}\label{eqn:hyper} \hyperApprox(\B) := & \nonsep(\bar {\B}) + \sum_{i=1}^\noDesign \partHyperApproxIR((\B)_i^R) -\constHyperApproxIR + \partHyperApproxII((\B)_i^I) -\constHyperApproxII \end{align} where the contributions of the real and imaginary part of $B$ are defined for $s\in\{R,I\}$ as \begin{align*} \partHyperApproxIs((\B)_i^s) & := \Big\langle (U^s - \bar B^s_i)\nabla ^{i,s}_+\nonsep(\bar {\B})(U^s - \bar B^s_i) + \tau (B^s_i-\bar B^s_i)^2, (U^s - B^s_i)^{-1}\Big\rangle \nonumber\\ & \qquad + \Big\langle (L^s - \bar B^s_i ) \nabla ^{i,s}_-\nonsep(\bar {\B})(L^s - \bar B^s_i) - \tau ((\B)^s_i-\bar B^s_i)^2, (L^s - (\B)^s_i)^{-1}\Big\rangle\nonumber \\ \constHyperApproxIs & := \Big\langle \nabla ^{i,s}_+\nonsep(\bar {\B}) , (U^s - \bar B^s_i)\Big\rangle + \Big\langle \nabla ^{i,s}_-\nonsep(\bar {\B}) , (L^s-\bar B^s_i)\Big\rangle\nonumber\corr{.} \end{align*} Here $\nabla^{i,s}_+\nonsep(\bar \B)$ and $\nabla ^{i,s}_-\nonsep(\bar \B)$ are the projections of $\nabla^{i,s}\nonsep(\bar \B)$ onto $\S_+$ and $\S_-$, \ie the space of symmetric positive definite and symmetric negative definite tensors, respectively. \end{definition} \begin{definition}[separable function on $\SCK$]\label{defn:seperable} A function $v:\SCK \to \R$ is called separable on $\SCK$ iff there exist $\bar \B\in \SCK$ and $\bar v_i : \SC \to \R$ for all $i \in\indexDesign$ such that \begin{equation*} v(\B) = v(\bar {\B}) + \sum_{i=1}^\noDesign \bar v_i((\B)_i) \quad \forall \B\in \SCK. \end{equation*} \end{definition} \begin{theorem} The hyperbolic approximation $\hyperApprox$ of $\nonsep$ given in \cref{defn:hyper} is a convex first-order approximation according to \cref{defn:firstorder} and separable on $\SCK$ according to \cref{defn:seperable}. \end{theorem} \begin{proof} The theorem is a straightforward extension of \cite[Theorem 3.4]{Stingl2009a} to the case of complex-valued material tensors. \end{proof} \begin{remark}[Proximal point terms]\label{rem:proxpoint} The terms $\tau \langle (B^s_i-\bar B^s_i)^2, (U^s - B^s_i)^{-1}\rangle$ and $\tau \langle ((\B)^s_i-\bar B^s_i)^2, (L^s - (\B)^s_i)^{-1}\rangle$ in \cref{defn:hyper} act as proximal point terms; accordingly the parameter $\tau$ takes the role of a proximal point parameter. As shown for the real-valued case in \cite{Stingl2009a}\corr{,} for $\tau > 0$ the hyperbolic approximations introduced above become uniformly convex with a modulus of the type $c_0+\tau c_1$, where $c_0\in \R_{\geq 0}$ and $c_1\in \R_{>0}$ are appropriate constants. \end{remark}\par To establish a solution scheme for \eqref{eqn:opt:probTMdisred}, we define the model problem \begin{equation}\label{eqn:Pjp} \begin{aligned} \min_{\B\in\AdmSet} ~ &\totalHyperApprox(\B) \end{aligned} \tag{\ensuremath{P^{\tau,\gamma}_{h,\bar\B}}} \end{equation} with the objective functional $\totalHyperApprox(\B) := \hyperApprox(\B)+\gamma \Jgraydisc(\B)$, i.e. we have applied the hyperbolic approximation \eqref{defn:hyper} to the non-separable functional $\nonsep$. \par We are now in the position to state the so-called sequential global programming algorithm, see \cref{alg:SGP}. We note that in each major iteration\corr{,} a finite number of sub-problems of type \cref{eqn:Pjp} \corr{are} solved. The inner loop realizes a globalization strategy: whenever the solution of the sub-problem does not provide sufficient descent for the objective of the original problem, the proximal point parameter $\tau$ is increased. Of course, in practice the stopping criterion of the outer loop is relaxed by a small positive constant. \begin{algorithm}\caption{Sequential Global Programming}\label{alg:SGP} \begin{algorithmic} \State Choose $\B^1\in \AdmSet$, $j \gets 1$, $\theta >1$, $\delta>0$, $\gamma\ge 0$ \Loop \State Choose $\tau$ \State $\bar \B \gets \B^j$ \Loop \State Solve \eqref{eqn:Pjp} $\rightarrow \B^+$ to global optimality \If{$\Jtotal(\B^+) < \totalHyperApprox(\B^j) - \delta \| \B^+ -\B^j\|_{F^\noDesign}^2$} \State $\B^{j+1} \gets \B^+$, {\bf break inner loop} \EndIf \State $\tau \gets \theta \tau$ \EndLoop \If {$\|\B^{j+1} -\B^{j}\|_{F^\noDesign}^2 = 0$} \State $\B^* \gets \B^{j+1}$, {\bf break outer loop} \EndIf \State $j \gets j+1$ \EndLoop \end{algorithmic} \end{algorithm} \subsection{Convergence theory} In order to be able to prove a global convergence result for \cref{alg:SGP}\corr{,} we need a few technical definitions \corr{and} assumptions. \corr{This is because standard optimality conditions do not apply due to the potential non-smoothness of the grayness term and the structure of the feasible set in \cref{eqn:opt:probTMdis}}. \begin{definition}[tangential cone]\label{def:conv:tangentialcone} Let $l\in\indexEdges$ and $Y\in \GraphSet$, $\B\in \AdmSet$, then we define the following tangential cones: \begin{align*} \TangCone^{l}_Y & := \Big\{ D\in \SC \mid \exists (E^n)_{n\in\N} \subset \EdgeImage_l, (t^n)_{n\in\N} \subset \R_{>0},E^n \rightarrow Y, t^n\searrow 0 :\frac{E^{n}-Y}{t^n} \rightarrow D \Big\},\\ \TangCone^{l}_\B & := \Big\{ \D\in \SCK \mid (\D)_k \in \TangCone^l_{(\B)_k}, \forall k\in \indexDesign\Big\}. \end{align*} \end{definition} \begin{ass}[Assumptions on $\Jtotal$ and $\totalHyperApprox$]~\label{ass:conv} The functions $\Jtotal : \AdmSet \to \R$ and $\totalHyperApprox : \AdmSet \to \R$ are twice continuously \corr{directionally} differentiable at any $\B \in \AdmSet$ in any direction of the tangential cone $\TangCone^l_{\B}$ and for all $\bar \B \in \AdmSet$. \end{ass} We note that it is a simple exercise to show that the assumption on $\totalHyperApprox$ is satisfied for the grayness term and parametrizations discussed later in this section. Moreover\corr{,} the smoothness assumption on $\Jtotal$ is satisfied if the physical objective functional $\Jphysdisc$ is twice continuously differentiable. \begin{remark}[Inequalities for the objective functional and its approximation] Due to \cref{ass:conv} and the compactness of the set of admissible tensors $\GraphSet$ as well as the properties of the proximal point term discussed in \cref{rem:proxpoint}\corr{,} there exist $c_0\in \R_{\geq 0}$ and $c_1,c_2,c_3,c_4 \in \R_{>0}$ s.t. the following holds: \begin{align} \label{eqn:modulus} \secdel{\totalHyperApprox}{(\B)_k}[Y,Z] &\geq (c_0 + c_1 \tau)\|Y\|_F\|Z\|_F & &\forall \B \in \AdmSet, \forall Y,Z \in \TangCone_{(\B)_k}^l\\ \Big| \del{\Jtotal(\B)}{(\B)_k}[Y] \Big| & \leq c_2\|Y\|_F & &\forall \B \in \AdmSet, \forall Y \in \TangCone_{(\B)_k}^l \label{eqn:d_obj_func_lipschitz}\\ \label{eqn:dd_obj_func_lipschitz} \Big| \secdel{\Jtotal(\B)}{(\B)_k}[Y,Z] \Big| & \leq c_3\|Y\|_F\|Z\|_F & &\forall \B \in \AdmSet, \forall Y,Z \in \TangCone_{(\B)_k}^l \\ |\totalHyperApprox( \B) - \Jtotal( \B)| & \leq (c_4 + c_1\tau)\|\bar \B - \B\|_{F^\noDesign} ^2 & & \forall \B \in \AdmSet \label{eqn:approx_ineq}\\ \label{eqn:FOA_prop} \del{\totalHyperApprox(\B)}{(\B)_k}[Y] & = \del{\Jtotal(\B)}{(\B)_k}[Y] & & \forall Y \in \TangCone_{(\B)_k}^l \end{align} for all $(l,k,\tau) \in \indexEdges\times\indexDesign\times\R_{\geq 0}$ and $ \bar \B \in \AdmSet$\corr{.} \end{remark} Thus we can prove the following: \begin{lemma}[finite number of inner iterations] \label{lem:fnii} The inner loop terminates after a finite number of iterations. \end{lemma} \begin{proof} There exists $\tau_{\max}\in \R$ s.t. \begin{align*} \totalHyperApprox(\B) \geq \Jtotal(\B) \qquad \forall \B, \bar \B \in \AdmSet,\tau > \tau_{\max} \end{align*} and thus for $\hat \tau_{\max} := \tau_{\max} + \delta$ the stopping criterion of the inner loop of \cref{alg:SGP} is fulfilled for all $\tau > \hat \tau_{\max}$. This directly results from properties a) and b) in \cref{defn:firstorder} which hold for $\totalHyperApprox(\B)$, boundedness of the second derivative of $\Jtotal$ on the compact set $\AdmSet$ and the lower bound on the second derivative of $\totalHyperApprox$ depending on $\tau$ obtained from \cref{eqn:modulus}, \ie $\tau_{\max} = (c_3-c_0)(c_1)^{-1}$. \end{proof} \begin{theorem}[convergence results] Let $(\tau^m,\B^m)_{m\in\N}$ be a sequence generated by \cref{alg:SGP} and let \cref{ass:conv} be satisfied. Then \corr{there} exists $\Jtotal^* \in \R$ and $\B^* \in \AdmSet$ such that the following holds: \begin{enumerate} \item[a)] convergence of function values: $\quad \Jtotal(\B^m) \rightarrow \Jtotal^*$ \item[b)] convergence of material tensors: \begin{itemize} \item For $\delta > 0$: $\quad \B^m \rightarrow \B^* $ \item For $\delta = 0$: $\quad \B^{m_n} \rightarrow \B^* \quad$ for subsequence $(m_n)_{n\in\N}$\corr{.} \end{itemize} \end{enumerate} \end{theorem} \begin{proof}~ \begin{enumerate} \item[a)] Due to the monotonicity of $\Jtotal(\B^m)$, the continuity of $\Jtotal$ and the boundedness of the admissible set $\AdmSet$, we \corr{clearly have} convergence of the function values. \item[b)] The boundedness of $\AdmSet$ leads directly to the existence of a convergent subsequence. If $\delta > 0$\corr{,} then the stopping criterion of the inner loop of \cref{alg:SGP} leads to $\Jtotal(\B^{m+1}) \leq \Jtotal(\B^{m}) - \delta\|\B^m - \B^{m+1}\|^2_{F^\noDesign}$ and implies $\big|\Jtotal(\B^{m})-\Jtotal(\B^{m+1})\big| \geq \delta\|\B^m - \B^{m+1}\|^2_{F^\noDesign}$. Together with the convergence of the objective functional values, \ie $|\Jtotal(\B^m) - \Jtotal(\B^{m+1})| \rightarrow 0$\corr{,} we get \begin{align*} \|\B^m - \B^{m+1}\|^2_{F^\noDesign} \rightarrow 0. \end{align*} Together with the convergence along the subsequence to $\B^*$ we \corr{obtain} convergence of the whole sequence $(\B^m)_{m\in\N}$. \end{enumerate} \end{proof} Now, first order optimality conditions based on the tangential cone defined in \cref{def:conv:tangentialcone} can be written as: \begin{definition}[first order optimality]\label{def:firstorder} A material distribution $\B\in\AdmSet$ is called first order optimal to $\Jtotal$ iff \begin{align*} \del{\Jtotal(\B)}{(\B)_k}[(\Y)_k] \geq 0 \qquad \forall (l,k)\in \indexEdges \times \indexDesign, \Y \in \TangCone^l_\B\corr{.} \end{align*} \end{definition} \begin{theorem}[first order optimality of $\B^*$] Any accumulation point $\B^*$ of the sequence $(\B^k)_{k\in\mathds N}$ generated by \cref{alg:SGP} is first order optimal to $\Jtotal$. \end{theorem} \begin{proof} We will argue by contradiction \corr{and assume} $\B^*$ is not first order optimal. Then there exists an element index $k\in \indexDesign$, edge index $l\in \indexEdges$ and $(\Y)_k \in \TangCone^l_{(\B^*)_k}$ such that \begin{align} \del{\Jtotal(\B^*)}{(\B^*)_k}[(\Y)_k] = -\FOOptConst < 0. \label{eqn:LocDer} \end{align} Thus there exist sequences $(E^n)_{n\in\mathds N}$ and $(t^n)_{n\in\mathds N}$ as in \cref{def:conv:tangentialcone} of the tangential cone $\TangCone^l_{(\B^*)_k}$ with $E^n \rightarrow (\B^*)_k$ and $N_1\in \mathds N$ large enough that the following holds: \begin{align*} \del{\Jtotal(\B^*)}{(\B^*)_k}[E^n - (\B^*)_k] - \del{\Jtotal(\B^*)}{(\B^*)_k}[(\Y)_k] t^n & < \tfrac12 t^n \FOOptConst & \forall n > N_1\corr{.} \end{align*} Together with \eqref{eqn:LocDer}\corr{, we obtain}: \begin{align} \del{\Jtotal(\B^*)}{(\B^*)_k}[E^n - (\B^*)_k] &< -\tfrac12 t^n \FOOptConst & \forall n> N_1 \label{eq:approx_prop_6}. \end{align} Note that the existence of $N_1$ fulfilling this inequality is given by the properties of the tangential cone. To continue the proof, we derive an estimate for the directional derivative in \corr{the} direction of $E^n$ for a small change in the expansion point $\B^m$, \ie \begin{align*} &\Big|\del{\Jtotal(\B^m)}{(\B^m)_k} [E^n - (\B^m)_k] - \del{\Jtotal(\B^*)}{(\B^*)_k} [E^n - (\B^*)_k]\Big|\\ & \quad = \Big|\del{\Jtotal(\B^m)}{(\B^m)_k} [E^n - (\B^m)_k + (\B^*)_k - (\B^*)_k] \\ & \qquad - \del{ \Jtotal(\B^*)}{(\B^*)_k} [E^n - (\B^m)_k + (\B^*)_k - (\B^*)_k] - \del{\Jtotal(\B^*)}{(\B^*)_k} [(\B^m)_k-(\B^*)_k]\Big|\\ & \quad \leq c_2\|\B^* - \B^m\|_{F^\noDesign} + c_3\|\B^*-\B^m\|_{F^\noDesign} \Big(\|\E^n - \B^*\|_{F^\noDesign} + \|\B^*-\B^m\|_{F^\noDesign} \Big) \end{align*} where we have used \cref{eqn:d_obj_func_lipschitz} and \cref{eqn:dd_obj_func_lipschitz}. As $\B^m \rightarrow \B^*$ and $n\in \N$ is fixed, \corr{based on the latter estimate we can choose} $M_1(t^n)\in\N$ sufficiently large s.t. \begin{align} \Big|\del{\Jtotal(\B^m)}{(\B^m)_k} [E^n - (\B^m)_k] - \del{\Jtotal(\B^*)}{(\B^*)_k} [E^n - (\B^*)_k]\Big| < \tfrac14t^n\FOOptConst\quad \forall m> M_1(t^n) \label{eq:approx_prop_10}. \end{align} If $\B^m$ converges only along a sub-sequence, we redefine $(\B^m)_{m\in\N}$ by a sub-sequence thereof, \ie $\B^m := \B^{m_n}$ and continue with the same arguments. Thus by \eqref{eq:approx_prop_6} and \eqref{eq:approx_prop_10}, for a slight change of the expansion point the directional derivative \corr{remains} negative, \ie \begin{align} \del{\Jtotal(\B^m)}{(\B^m)_k}[E^n - (\B^m)_k] &< -\tfrac14t^n\FOOptConst & \forall m > M_1(t^n), \forall n>N_1\corr{.} \label{eqn:LocDer1} \end{align} Moreover, by \eqref{eqn:FOA_prop} and \eqref{eqn:approx_ineq} the following holds for the first order accurate model $\totalHyperApproxArg{\tau}{\B^m}$ with $\E^n:= (B^m_1,\dotsc,B^m_{k-1},E^n,B^m_{k+1},\dotsc,B^m_K)$ for all $(m,n,\tau) \in \N\times \indexDesign \times \R_{>0}$: \begin{align} \totalHyperApproxArg{\tau}{\B^m}(\E^n) & \leq \Jtotal(\B^m) + \del{\Jtotal(\B^m)}{(\B^m)_k}[E^n - (\B^m)_k] + (c_4 + c_1\tau)\|E^n - (\B^m)_k\|_F^2. \label{eqn:mod_prop} \end{align} As $E^n$ \corr{converges} to $(\B^*)_k$, there exists $N_2 > N_1$ such that \begin{align*} \|E^n - (\B^*)_k\|^2_F &\leq \frac{t^n \FOOptConst}{8(c_4+c_1\overline{\tau})} & \forall n> \corr{N_2} \end{align*} with $\overline{\tau}$ being the maximal possible $\tau$ of \cref{alg:SGP}. This maximum exists as shown in \cref{lem:fnii} and is bounded by the maximum of $\theta\tau_{\max{}}$ and the initial value for $\tau$, where $\theta > 1$ is the scaling parameter within the inner loop of \cref{alg:SGP}. \par By inequality \eqref{eqn:LocDer1} and the latter estimate we \corr{have} from \cref{eqn:mod_prop} \begin{align} \label{eqn:DoNotKnow} \totalHyperApproxArg{\tau}{\B^m}(\E^n) & < \Jtotal(\B^m) - \tfrac18t^n\FOOptConst & \forall m > M_1(t^n), \forall n>N_2. \end{align} As $\B^m \rightarrow \B^*$ for $m\rightarrow \infty$ we can choose $M_2(t^n) > M_1(t^n)$ \corr{sufficiently} large s.t. \begin{align*} |\Jtotal(\B^m) - \Jtotal(\B^*)| &< \tfrac{1}{16} t^n \FOOptConst & \forall m > M_2(t^n) , \forall n \in \N. \end{align*} Combining the last two inequalities and noting that \cref{eqn:DoNotKnow} holds for all $\tau$ we can choose $n>N_2$ and $m>M_2(t^n)$ for which the following inequality is satisfied: \begin{align*} \totalHyperApproxArg{\tau_{m}}{\B^m}(\E^n) < \Jtotal(\B^*) - \tfrac{1}{16} t^n \FOOptConst. \end{align*} Here, \corr{$\tau_m$ denotes} the actual proximal point parameter used in the $m$-th outer iteration of \cref{alg:SGP}. Finally noting that $\B^{m+1}$ is the global minimizer of the sub-problem $\min_{\B\in\AdmSet} \totalHyperApproxArg{\tau_m}{\B^m}(\B)$ and taking the inner stopping criterion in \cref{alg:SGP} into \corr{consideration}, we arrive at \[\Jtotal(\B^{m+1}) \leq \totalHyperApproxArg{\tau_{m}}{\B^m}(\B^{m+1}) \leq \totalHyperApproxArg{\tau_{m}}{\B^m}(\E^n) < \Jtotal(\B^*) - \tfrac{1}{16}t^n\FOOptConst.\] This is in contradiction to the monotonicity properties of the sequence of objective function values, i.e. $(\Jtotal(\B^m))_{m\in\N}$. Thus, $\B^*$ is first order optimal in the sense of \cref{def:firstorder}. \end{proof} \subsection{Solution of the subproblem}\label{sec:subprobsolve} In this section \corr{we describe how} the sub-problems in \cref{alg:SGP} can be efficiently solved. In order to do this, we fix the definition of the grayness functional and the parametrization of the feasible set. In this paper we restrict ourselves to \corr{a} grayness function of the following type: \begin{definition}[grayness penalization on the graph $\AdmSet$]\label{defn:grayfun} On $\AdmSet$ we define for all $\B = (B_1,\dotsc,B_\noDesign)\in \AdmSet$ a grayness penalization $\Jgraydisc$ by \begin{align*} \Jgraydisc(\B) & = \sum_{i=1}^K \tilde \Jgraydisc(B_i) & \text{with}&& \tilde \Jgraydisc(B_i) & := \sum_{l\in \indexEdges }\begin{cases} \Param_l^\mo (B_i) (1-\Param_l^\mo (B_i) ) & \text{if } B_i \in \EdgeImage_l\setminus\NodeImage \\ 0 & \text{otherwise}. \end{cases} \end{align*} \end{definition} We note that for parametrizations $(\psi_l)_{l\in \noEdges}$ satisfying the assumptions in \cref{ass:properties}, the directional differentiability on each edge of the parametrization required in \cref{ass:conv} is satisfied for the grayness functional stated in \cref{defn:grayfun}. Next, we reformulate \eqref{eqn:Pjp} in terms of the parametrization $\Param$: \begin{equation}\label{eqn:QC}\tag{\text{$Q^\gamma$}} \begin{aligned} \min_{\b l \in(\indexEdges)^\noDesign} \min_{\b \alpha\in[0,1]^\noDesign} & \ \hyperApprox (\Param_{l_1}(\alpha_1),\dotsc,\Param_{l_\noDesign}(\alpha_\noDesign)) + \gamma \sum_{i=1}^\noDesign \alpha_i (1- \alpha_i)\corr{.} \end{aligned} \end{equation} \begin{lemma} If $( \b \alpha^*,\b l^*)$ is a global optimal solution of \eqref{eqn:QC}, then $(\B)_i = \Param_{l_i^*}(\alpha_i^*)$ for all $i \in \indexDesign$ is a global optimal solution of \eqref{eqn:Pjp}. \end{lemma} \begin{proof} Since $\tilde \Jgraydisc(\Param_{l_i}(\alpha_i)) = \alpha_i(1-\alpha_i)$ for all $i\in \indexDesign$ by \cref{defn:grayfun}, \eqref{eqn:QC} is a reparametrization of \eqref{eqn:Pjp}. That is why the global optimal solutions coincide. \end{proof} Due to the separability of $\hyperApprox$ and $\Jgraydisc$, we find the global optimum for \eqref{eqn:QC} if we find the global optimal solution of \begin{equation}\label{eqn:Qil}\tag{\text{$Q^i_l$}} \begin{aligned} \min_{\alpha_i \in[0,1]} & \ \partHyperApproxIR (\Re(\Param_{l}(\alpha_i))) + \partHyperApproxII (\Im(\Param_{l}(\alpha_i)))+ \gamma \alpha_i (1-\alpha_i) \\ \end{aligned} \end{equation} for each element $i\in \indexDesign$ and edge index $l\in \indexEdges$ (see \cref{alg:submultigamm}). \corr{Note that constant terms with respect to $\b\alpha$ are neglected here}. \begin{algorithm}[ht]\caption{solution of \eqref{eqn:Pjp}}\label{alg:submultigamm} \begin{algorithmic}[1] \For {$i\in \indexDesign$} \For {$l \in \indexEdges$} \State $\beta\gets$ solve \eqref{eqn:Qil} globally \State $(\b j^*)_l \gets \partHyperApproxIR (\Re(\Param_{l}(\beta))) + \partHyperApproxII (\Im(\Param_{l}(\beta)))+ \gamma \beta (1-\beta) $ \State $(\b \alpha^*)_l \gets \beta$ \EndFor \State Find index $l^*$ s.t. $(\b j^*)_{l^*} \le (\b j^*)_{l}$ for all $l \in \indexEdges $ \State $B_i \gets \Param_{l^*}((\b \alpha^*)_{l^*})$ \EndFor \end{algorithmic} \end{algorithm} The following result is useful when the desired solution is of \corr{a} discrete nature, \ie $(\B^*)_i \in \NodeImage, \forall i\in \indexDesign$. \begin{remark}\label{rem:discreteSol} For every $\tau>0$ there exists a $\gamma_{\text{max}}>0$ sufficiently large such that for all $\gamma> \gamma_{\max}$ the global optimal solution $(\b l^*_\gamma,\b\alpha^*_\gamma)$ satisfies $\Param((l_\gamma^*)_i,(\alpha_\gamma^*)_i) \in \NodeImage$ for all $i \in \indexDesign$. \end{remark} \begin{proof} Let $v_i((\B)_i) := \partHyperApproxIR( (\B)_i^R)+ \partHyperApproxII((\B)_i^I)$. Since $v_i\circ \Param_{l} \in C^2([0,1])$ for all $l\in \indexEdges$ and $i\in\indexDesign$, its second derivative is bounded from above by $\sigma \in \R$. By choosing $\gamma_\text{max} = \tfrac\sigma2$ the second derivative of $\delta \mapsto v_i(\Param_{l} (\delta)) + \gamma \delta (1-\delta) $ is strictly negative for all $\gamma > \gamma_\text{max}$ for all $l\in \indexEdges$ and $i\in\indexDesign$. Thus the second order optimality conditions are never fulfilled and the global minimum of the latter function restricted to $[0,1]$ is located on the boundary. \end{proof} In the next section, we \corr{provide} strategies to obtain the global optimal solution of \eqref{eqn:Qil} for two particular choices of parametrizations. Moreover, to shorten the notation we define for $s\in\{R,I\}$ \begin{align}\label{eqn:derabbrev} \constParamApproxIsU & := (U^s - \bar B^s_i) \nabla ^{i,s}_+\nonsep(\bar {\B}) (U^s - \bar B^s_i), & \constParamApproxIsL & := (L^s - \bar B^s_i ) \nabla ^{i,s}_-\nonsep(\bar {\B})(L^s - \bar B^s_i), \end{align} since these terms are independent of the design parameters in both cases. \subsubsection{Rotational parametrization}\label{sec:rotparam} \begin{definition}[rotational parametrization]\label{defn:rotparam} Let the so-called reference material tensor $B^{(r)}\in \SC$ be a diagonal tensor. We call a parametrization $\Param$ of material tensors \corr{a} rotational parametrization based on $B^{(r)}$, iff \begin{align*} \Param(\delta) = R(\pi \delta) B^{(r)} R(\pi \delta)^T \qquad \delta\in [0,1] \end{align*} with rotation matrix $R\colon\R \to \operatorname{SO}(2)$: \begin{align*} R(\pi\delta) = \begin{pmatrix} \cos(\pi\delta) & -\sin(\pi\delta)\\\sin(\pi\delta)& \cos(\pi\delta) \end{pmatrix}. \end{align*} \end{definition} \begin{theorem}[global solution, rotational parametrization] For a rotational parametrization $\Param_l$ based on a real-valued diagonal reference tensor $B^{(r)} \in \SR$ \corr{with} asymptotes satisfying the assumptions of \cref{defn:hyper} and $\gamma=0$, the parametrized subproblem \eqref{eqn:Qil} has the global optimal solution \begin{equation*} \alpha^*_i = \frac1{2\pi}\begin{cases} \arctan(-\frac{b}{a}) + \pi & a < 0\\ \arctan(-\frac{b}{a})\mod 2\pi & a > 0\\ \frac\pi2\sign(b)+\pi & a = 0 \end{cases} \end{equation*} with the two parameters \begin{align*} a & = \tau \big[(\bar B_i)_{22}-(\bar B_i)_{11}\big] c_0 + c_{1}, & b &= \tau \big[(\bar B_i)_{12}+(\bar B_i)_{21}\big] c_0 + c_{2}\corr{.} \end{align*} \corr{Here,} \begin{align*} c_0 & = 2(B^L_{11} - B^U_{11}) B^{(r)}_{11} + 2(B^U_{22} - B^L_{22}) B^{(r)}_{22} & \\ & \quad + \big[ (B^U_{11} - B^U_{22})+(B^L_{22}-B^L_{11})\big]\big[(\bar B_i)_{11}+(\bar B_i)_{22}\big]\corr{,}\\ c_{1} & = (B^U_{11}-B^U_{22}) \big[(C_U^{i,R})_{22}-(C_U^{i,R})_{11}\big] +(B^L_{11}-B^L_{22}) \big[(C_L^{i,R})_{22}-(C_L^{i,R})_{11}\big]\corr{,}\\ c_{2} & = (B^U_{11}-B^U_{22}) \big[(C_U^{i,R})_{12}+(C_U^{i,R})_{21}\big] + (B^L_{11}-B^L_{22}) \big[(C_L^{i,R})_{12}+(C_L^{i,R})_{21}\big]\corr{,} \end{align*} and we use the abbreviations $ B^L = (L-B^{(r)})^\mo$, $B^U = (U-B^{(r)})^\mo$. Furthermore, $\sign$ denotes the sign function and $\mod$ the modulo operation, respectively. \end{theorem} \begin{proof} First, we compute all stationary points of the hyperbolic approximation \eqref{eqn:hyper}, \corr{thus we} can neglect terms which are constant with respect to material tensors $\B = (B_1,\dotsc,B_\noDesign)$. For a fixed element $i\in \indexDesign$ we obtain with \eqref{eqn:derabbrev} \begin{align*} \partHyperApproxIR (B_i) = & \left\langle \constParamApproxIRU + \tau (B_i-\bar B_i)^2 ,(U - B_i)^\mo\right\rangle + \left\langle \constParamApproxIRL - \tau (B_i-\bar B_i)^2 , (L-B_i)^\mo\right\rangle. \end{align*} Note that $\constParamApproxIRU$ and $\constParamApproxIRL$ are independent on $B_i$. Taking the rotational parametrization into account, \ie $B_i = R(\pi \alpha_i) B^{(r)} R(\pi \alpha_i)^T$, using the choice of asymptotes $L = l \mathds1$ and $U = u \mathds1$ and by properties of the scalar product as well as rotation matrices this can be rewritten as \begin{align*} \mathrlap{ \partHyperApproxIR (R(\pi \alpha_i) B^{(r)} R(\pi \alpha_i)^T) =}\quad & \\ & \left\langle R(\pi \alpha_i)^T \constParamApproxIRU R(\pi \alpha_i) + \tau (B^{(r)} -R^T(\pi \alpha_i)\bar B_iR(\pi \alpha_i))^2 ,(U - B^{(r)} )^\mo\right\rangle \nonumber \\ & + \left\langle R(\pi \alpha_i)^T \constParamApproxIRL R(\pi \alpha_i) - \tau (B^{(r)} -R^T(\pi \alpha_i)\bar B_iR(\pi \alpha_i))^2 ,(L - B^{(r)} )^\mo\right\rangle.\nonumber \end{align*} After straightforward calculus using angle sum identities, the derivative of $\partHyperApproxIR(B_i)$ with respect to the design parameter $\alpha_i$ has the form \begin{align}\label{eqn:statioary} \tot{ \partHyperApproxIR (R(\pi \alpha_i) B^{(r)} R(\pi \alpha_i)^T) }{\alpha_i} & = a \pi\sin(2\pi \alpha_i) + b \pi \cos(2\pi \alpha_i) \end{align} with the coefficients $a$, $b$ as given in the theorem. The stationary points of $\partHyperApproxIR(B_i)$ are given by the two roots of \eqref{eqn:statioary} in the interval $[0,1)$. Now we choose the root for which the second derivative of $\partHyperApproxIR(B_i)$ with respect to $\alpha_i$ is positive. This root is given by the formula for $\alpha_i^*$ \corr{stated} in the theorem. \end{proof} \subsubsection{Polynomial parametrization}\label{sec:polyparam} In this paragraph, we provide a solution scheme to solve the subproblem \eqref{eqn:QC} if the material tensor is parametrized by a polynomial on an edge of $\GraphSet$. We show that in this case the hyperbolic approximation $\hyperApprox$ is a rational polynomial and solve \eqref{eqn:Qil} by finding the roots of its derivative. \begin{definition}[polynomial parametrization] We call a parametrization $\Param$ of material tensors polynomial parametrization of order $k$, \ie $\Param \in \mathcal P ^k_{\SC}$, if \begin{align*} \Param(\delta) = B^{(1)}(1-\delta) + B^{(2)} \delta + \delta(1-\delta) \sum_{i=0}^{k-2} A_i \delta^i, \qquad \delta \in [0,1] \end{align*} with $B^{(1)}\in \SC$, $B^{(2)}\in \SC$ and interpolation coefficients $A_i\in \SC$, $0\le i \le k-2$. \end{definition} \begin{lemma} \label{lem:numden} Let the parametrization $\Param_l$ be given by a polynomial of order $k$ over ${\SC}$, \ie $\Param_l \in\mathcal P^k_{\SC} $. Then the contribution of the $i$-th element to the hyperbolic approximation \eqref{eqn:hyper}, for $Y:=\Param_l(\delta)$, is \begin{equation*} \partHyperApproxIR (Y^R)+ \partHyperApproxII (Y^I) = \frac{p^i_l(\delta)}{q^i_l(\delta)} \end{equation*} where \begin{equation*} \begin{aligned} p^i_l(\delta) & = \sum_{s\in\{R,I\}} \frac { N_L^{i,s}(\delta) q_l^i (\delta)}{ \det(L^s-Y^s)} + \frac{ N_U^{i,s}(\delta) q_l^i (\delta)}{\det(U^s - Y^s)} \in \mathcal P^{9k}_\R,\\ q^i_l(\delta) & = \prod_{s\in\{R,I\}}\det(L^s-Y^s) \det(U^s - Y^s) \in \mathcal P^{8k}_\R \end{aligned} \end{equation*} and \begin{align*} N^{i,s}_L(\delta) & = \langle \constParamApproxIsL - \tau (Y^s-\bar B_i^s)^2 , \adj(L^s- Y^s)\rangle \\ N^{i,s}_U(\delta) & = \langle \constParamApproxIsU + \tau (Y^s-\bar B_i^s)^2 , \adj(U^s - Y^s)\rangle. \end{align*} \end{lemma} \begin{proof} Together with the definitions for $\constParamApproxIsU$ and $\constParamApproxIsL$ from \eqref{eqn:derabbrev}, the contribution of the real and imaginary part of $Y:= \Param_l(\delta)$ to \eqref{eqn:hyper} is \begin{align*} \partHyperApproxIs(Y) = \langle \constParamApproxIsL - \tau (Y-\bar B_i^s)^2 , (L^s-Y)^\mo \rangle + \langle \constParamApproxIsU + \tau (Y-\bar B_i^s)^2 , ( U^s-Y)^\mo \rangle. \end{align*} for $s\in\{R,I\}$, respectively. Using the formula $A^{-1}\det(A) = \adj(A)$ for a matrix $A$, we \corr{obtain} \begin{align*} \partHyperApproxIs(Y) = & \det(L^s-Y)^\mo \langle \constParamApproxIsL - \tau (Y-\bar B_i^s)^2 , \adj(L^s-Y) \rangle \\ & + \det(U^s-Y)^\mo \langle \constParamApproxIsU + \tau (Y-\bar B_i^s)^2 , \adj(U^s-Y) \rangle. \end{align*} \corr{Using} the common denominator $q^i_l$, we \corr{have} the expression \begin{equation*} \partHyperApproxIR (Y^R)+ \partHyperApproxII (Y^I) = \frac{p^i_l(\delta)}{q^i_l(\delta)} \end{equation*} with the polynomials $p^i_l(\delta)$ and $q^i_l(\delta)$ \corr{as stated} in the lemma. \end{proof} \begin{algorithm}[ht]\caption{solution of \eqref{eqn:Qil} with polynomial parametrization}\label{alg:polysolve} \begin{algorithmic}[1] \State Find the $n$ real roots $\alpha^{(1)},\dotsc,\alpha^{(n)}$ in the interval $(0,1)$ of the polynomial \begin{equation*} (q^i_l)'(\delta) p^i_l(\delta) - q^i_l(\delta) (p^i_l)'(\delta) + \gamma (1-2\delta) q^i_l(\delta)^2 \end{equation*}\vspace*{-\baselineskip} \State $\b \beta \gets (0,\alpha^{(1)},\dotsc,\alpha^{(n)},1) $ \For{$1\le k \le n+2$} \State $j_k \gets\partHyperApproxIR (\Re(\Param_l(\beta_{k}))) + \partHyperApproxII (\Im(\Param_l(\beta_{k}))) + \gamma \beta_k(1-\beta_k)$ \EndFor \State Find index $n^*$ s.t. $j_{n^*} \le j_k$ for all $1\le k \le n+2 $ \State $\alpha^*_{l} \gets \beta_{n^*}$ \end{algorithmic} \end{algorithm} \begin{theorem} With the assumptions of \cref{defn:hyper} and \cref{lem:numden}, \cref{alg:polysolve} yields a global optimal solution $\alpha^*_l$ of \eqref{eqn:Qil}. \end{theorem} \begin{proof} By \cref{lem:numden} the objective functional of \eqref{eqn:Qil} can be expressed as \begin{align*} j^i_l(\delta) := \frac{p^i_l(\delta)}{q^i_l(\delta)} + \gamma \delta (1-\delta). \end{align*} Necessarily, the global minimum of \eqref{eqn:Qil} is located either at $0$, $1$ or a root of the derivative of $j^i_l$. By \cref{defn:hyper}, the denominator of $j^i_l$ has no roots in the interval $(0,1)$, thus the roots of the derivative of $j^i_l$ coincide with the roots of its numerator. The global minimum of \eqref{eqn:Qil} is then selected by comparing the objective functional values for all candidates\corr{,} including the boundary points $0$ and $1$. \end{proof} We finally note that under specific assumptions on the admissible material tensors, the degree of the polynomial of Step 1 of \cref{alg:polysolve} can be significantly reduced. For instance, in the case of isotropic, real-valued materials and linear interpolation the roots of a cubic polynomial \corr{must} be computed. Moreover\corr{,} in the general case\corr{,} to find the roots of a normalized polynomial, one possible approach is to compute eigenvalues of its companion matrix \cite{Edelman1995}. \section{Examples}\label{sec:example}~ In the following we discuss two different examples \corr{to illustrate this approach}. The purpose of the first example is twofold: First we \corr{wish} to investigate \corr{the performance of \cref{alg:SGP}} for optimization problems involving arbitrarily oriented anisotropic materials. In a second step, we want to examine \corr{the performance of the algorithm} when only a finite subset of orientations is admissible. In particular, we \corr{wish} to study \corr{the extent to which} the quality of the locally optimal solutions \corr{depends} on the finite number of orientations. This is important because as a consequence of \cref{rem:discreteSol} and \cref{lem:fnii} it is clear that for sufficiently large $\gamma$ every element of the set $\NodeImage^\noDesign$ is a local minimum of problem \cref{eqn:opt:probTMdisred}. \par The second example \corr{demonstrates} the capabilities of \cref{alg:SGP} \corr{when} the set of admissible material is parametrized by a complete graph with given complex-valued and isotropic material tensors at the nodes. \corr{To achieve this}, a material distribution is reconstructed by the information covered in the scattered magnetic field. Furthermore the effect of the regularization is investigated. \subsection{Cloaking of a scatterer} \begin{figure}\centering \scalebox{0.7}{ \includegraphics[]{GeomSetting.pdf} } \caption{Cloaking of a scatterer: Optimization setting}\label{fig:CSopt} \end{figure} The first example aims \corr{to minimize the} visibility of an absorbing core particle $\Omega_P$ \corr{using} optimal local orientation of an anisotropic material in the design domain $\Omega_D$ surrounding the particle (see \cref{fig:CSopt}). We assume that the background material tensor $B_b$ is real valued. \corr{The} visibility of a scattering object can be quantified by the amount of absorbed energy and scattered energy \cite{Mishchenko2014}. The absorbed energy is related to the absorption cross section which is defined as \begin{equation*} W^{\text{abs}} := -\int_{\d U} \frac12 \Re \left( E_T \times H_T^*\right)\cdot n \rmd \omega, \end{equation*} \ie the total energy flow, depending on the total electric $E_T$ and total magnetic field $H_T$, through the boundary of a neighborhood $U$ of the whole scatterer $\Omega_D \cup \Omega_P$, which can be chosen, for instance, as a \corr{ball}. \corr{The} scattered energy is proportional to the scattering cross section which reads \begin{equation*} W^{\text{sca}} := \int_{\d U} \frac12 \Re \left( E \times H^*\right)\cdot n \rmd \omega \end{equation*} with scattered field quantities $E$ and $H$. Note that both values, $W^{\text{ext}}$ and $W^{\text{sca}}$ are independent \corr{of} the choice of $U$ and non-negative. Absorption and scattering can be combined \corr{into} a common value\corr{,} which is called \corr{the} extinction cross section \begin{equation*} W^{\text{ext}} := W^{\text{abs}} + W^{\text{sca}} = \int_{\d U} \frac12 \Re \left( E_S \times H_I^* + E_I \times H_S^*\right)\cdot n \rmd \omega\corr{.} \end{equation*} \corr{This will serve} as the physical objective functional $\Jphys$ in this example after a \corr{number of} further adaptions. First we introduce the tensor-valued function $B_\Omega:\Omega\to \SC$ with \begin{equation*}B_\Omega(x) = \begin{cases} B_C(x) & x \in \Omega_C\\ B(x) & x\in \Omega_D. \end{cases} \end{equation*} Using transverse magnetic assumptions and switching to the two-dimensional setting, the extinction cross section transforms to \begin{align*} \Jphys(B,u) = -\Re \left( \frac\imath{2\omega} \int_{\d U} ( u^*_I (B_\Omega\nabla u + (B_\Omega-B_b)\nabla u_I) + u^* B_b\nabla u_I )\cdot n \rmd \omega \right). \end{align*} Moreover with partial integration and \eqref{eqn:weakbilinear}, we can proceed to a volume integral instead of the boundary integral and arrive at the objective functional in \corr{its} final form: \begin{align}\label{eqn:objTM} \Jphys(B,u) & = - \Re \left (\frac\imath{2\omega} \int_{U} \nabla u_I^H (B_\Omega - B_b) (\nabla u + \nabla u_I) \rmd x\right). \end{align} Note that the integral in the objective functional \eqref{eqn:objTM} can be restricted to $\Omega_D \cup \Omega_C$. \subsubsection{Optimization problem} We collect the results from the previous sections and give the full optimization problem for this example: \begin{align*} \min_{\B} ~ & \Jphysdisc(\B,\u) + \eta \Jregdisc(\B) + \gamma \Jgraydisc(\B) \\ \st ~ & (\SYS(\B) + \SYS_C) \u = \RHS(\B) + \RHS_C \\ & \corr{\text{where }}\B \text{ is parametrized by rotation angles} \end{align*} with the objective function $\Jphysdisc(\B ,\u) = \Re ( \u^H (V(\B)+V_P) + W(\B) + W_P)$ based on \eqref{eqn:objTM}. The vectors $V(\B),V_P\in \C^\noDoF$ are \corr{defined} element-wise as for all $1\le i \le \noDoF$ \begin{align*} (V(\B))_i & = \frac\imath{2\omega} \sum_{k=1}^{\noDesign} \int_{\TriI k} \nabla u_I^T (B_k^* - B_b^*) \nabla \phi_i \rmd x \\ (V_P)_i & = \frac\imath{2\omega} \int_{\Omega_P} \nabla u_I^T (B_P^* - B_b) \nabla \phi_i \rmd x \\ \intertext{and the scalars $W(\B),W_P\in \C$ are } W(\B) & = \frac\imath{2\omega} \sum_{k=1}^{\noDesign} \int_{\TriI k} \nabla u_I^T (B_k^* - B_b^*) \nabla u_I^* \rmd x \\ W_P & = \frac\imath{2\omega} \int_{\Omega_P} \nabla u_I^T (B_P^* - B^*_b) \nabla u_I^* \rmd x\corr{.} \end{align*} Moreover\corr{,} $\SYS,\SYS_C,\RHS$ and $\RHS_C$ are given in \cref{sec:discrete}. Together with the adjoint variable $\b p$ which solves the adjoint equation \begin{align*} (\SYS(\B) + \SYS_C) \b p = - (V(\B) + V_C)^* \end{align*} we \corr{obtain} the derivative of the physical objective functional in direction $\Y\in \SCK$ for all $k\in \indexDesign$ \begin{align*} \tot{\Jphysdisc(\B, \u(\B))}{(\B)_k}[(\Y)_k] & = \Re\left ( \int_{\TriI k} \Big ( \nabla_{\!\!h} \b p - \frac\imath{2\omega} \nabla u_I^* \Big)^T (\Y)_k \Big(\nabla_{\!\!h} \u + \nabla u_I\Big) \rmd x \right ). \end{align*} Here, $\nabla_{\!\!h} \b u = \sum_{i=1}^\noDoF u_i\nabla \phi_i$ and $\nabla_{\!\!h} \b p = \sum_{i=1}^\noDoF p_i\nabla \phi_i$. Based on this, the sub-problems in \cref{alg:SGP} can be established and solved \corr{using the} strategies discussed in \cref{sec:rotparam}. \subsubsection{Numerical results} The particle domain $\Omega_P$ with material tensor $B_P = (0.1 + 2\imath)^{-2}\mathds1$ is a \corr{ball} with radius 0.2, see \cref{fig:CSopt}. Inside the design domain $\Omega_D$, which is a \corr{ball} with radius 0.4, the material tensor is optimized with rotational parametrization based on the diagonal reference material tensor $B^{(r)} = \diag(1,2)^{-2}$. A box with a side length of 2 defines $\Omega_b$ and a layer with thickness 1 represents the PML both with material properties $B_b=\mathds1$. \par Furthermore, we choose a plane incident wave $u_I(x,y)= \sqrt2\exp(\imath \omega x)$ with wavelength $\lambda = 0.55,\omega = \frac{2\pi}\lambda$ and set the PML function to $s(t) = 1 - \frac{100}{\imath\omega}\max(0,|t|-1)$. \par The state and adjoint equation are solved \corr{using} the Finite Element Method (FEM) on triangluar cells and implemented in MATLAB \cite{Matlab2014}. We use linear Lagrange basis functions \cite{Zienkiewicz2013} to approximate the scalar fields. The triangulation of the computational domain is generated with the Delaunay triangulation tool Triangle \cite{Shewchuk1996} and \corr{the} triangulation process provides approx. $\num[scientific-notation = fixed]{1e5}$ triangular elements in total and approx. $\num[scientific-notation = fixed]{3e4}$ triangles in the design region $\Omega_D$. \par Since in this example we consider a rotational parametrization (\cref{defn:rotparam}), the admissible set is given as $\AdmSet=\GraphSet^\noDesign$ with $\GraphSet = \{ R(\pi\delta) B^{(r)} R(\pi\delta)^T\mid \delta\in[0,1]\}$. Thus the underlying graph has only one closed edge. The asymptotes for \cref{alg:SGP} are defined by $l=0$ and $u=100$ and thus satisfy the assumptions in \cref{defn:hyper}. We choose the regularization parametrer $\eta=100$, the circular filter radius $r_0 = 0.01$ and the grayness penalty factor $\gamma=0$. \begin{figure}\centering \parbox[b]{0.5cm}{ \includegraphics[height=0.155\textwidth]{programresults/cbar_rainbow_-2x2.pdf}\\\includegraphics[height=0.155\textwidth]{programresults/cbar_gray_0xpi2.pdf}} \subfigure[inital\label{subfig:clk:inital}]{{\includegraphics[width=0.31\textwidth]{{programresults/initial_black_new}.png}}} \subfigure[optimized\label{subfig:clk:final}]{{\includegraphics[width=0.31\textwidth]{{programresults/final_black_new}.png}}} \subfigure[close-up \label{subfig:clk:closeup}]{{\includegraphics[width=0.31\textwidth]{{programresults/closeup_black_new}.png}}} \caption{Optimization result for the cloaking of scatterer with anisotropic material} \end{figure} In the upper part of \cref{subfig:clk:inital} and \cref{subfig:clk:final} the total magnetic field is illustrated for the initial and the optimal design configuration, respectively. The backscattering and absorption of the particle including the coating layer is clearly visible for the initial configuration. In the lower \corr{section of both} figures the local orientation of the anisotropic tensor inside the coating layer is visualized. The gray scale colors represent the absolute value of the orientation angle between $0$ and $\frac\pi2$ and the dashes in the closeup (see \cref{subfig:clk:closeup}) support the illustration of the orientation angle. The extinction cross section associated with the optimized anisotropic coating layer is decreased by over $88\percent$ relative to the initial design. \par \corr{As previously mentioned,} we are also interested in \corr{the performance of \cref{alg:SGP}}, if we restrict the design to $L\in\N$ uniformly distributed admissible orientations, \ie $\tfrac{\pi l}{L}$ for all $0\le l\le L-1$. In order to investigate this, the underlying graph of $\GraphSet$ is divided into $L$ edges, \ie $\noEdges=L$, and we obtain the modified admissible set $\AdmSet$ with \begin{align*} \GraphSet = \bigcup_{l=0}^{L-1} \left\{ R\left(\tfrac\pi{L}(\delta+l)\right) B^{(r)} R\left(\tfrac\pi{L}(\delta+l)\right)^T\mid \delta\in[0,1]\right\} \end{align*} and $\NodeImage = \{ R\left(\tfrac{\pi l}{L}\right) B^{(r)} R\left(\tfrac{\pi l}{L}\right)^T\mid 0\le l\le L-1\}$. \corr{To ensure that only points located at nodes are considered}, in theory we could choose a penalty parameter $\gamma>\gamma_{\text{max}}$. \corr{However, as} we know from \cref{rem:discreteSol} that the global minimizer of each sub-problem is located in the set $\NodeImage^\noDesign$ in this case, rather than solving the sub-problem as \corr{described} in \cref{sec:polyparam}, we \corr{can evaluate} the model objective in all nodes and choose the one with the lowest function value for each element. \begin{table}\centering \caption{Progression of the relative cloaking with respect to number of admissible angles}\label{tbl:progression} \begin{tabular}{cc} \toprule number of angles & rel. cloaking \\\midrule 4 & 85.69\percent\\ 12 & 41.35\percent\\ 18 & 36.71\percent\\ 60 & 27.05\percent\\ 180 & 19.80\percent\\ 360 & 12.73\percent\\ continuous & 11.86\percent\\ \bottomrule \end{tabular} \end{table} The relative cloaking after optimization with different numbers of admissible angles is listed in \cref{tbl:progression}. It can be observed that \corr{with an increasing number of orientations, the optimal value of the cost function approaches the optimal value of the continuous problem.} This reveals that despite the \corr{apparent 'brute force'} approach to the solution of the sub-problem, the algorithm is obviously not trapped in local optima introduced by the highly non-convex grayness terms. \begin{figure}\centering \subfigure[18 angles]{ {\includegraphics[,width=0.25\textwidth]{{programresults/closeup_black_no_glyph_18}.png}}} \subfigure[180 angles]{ {\includegraphics[,width=0.25\textwidth]{{programresults/closeup_black_no_glyph_180}.png}}} \subfigure[continuous]{ {\includegraphics[,width=0.25\textwidth]{{programresults/closeup_black_no_glyph}.png}}} \caption{Comparison of optimization results for different number of admissible orientations}\label{fig:contvsdisc} \end{figure} In \cref{fig:contvsdisc} the close up of the optimization results for \corr{18 admissible angle, 180 admissible angles} and the continuous setting are compared. \subsection{Tomographic reconstruction} In this example we \corr{attempt to} reconstruct a known material distribution by using the electromagnetic field response on the artificial observation boundary $\d U$. \corr{To do this,} we define the physical objective functional following \cite{Cheney1999,Vauhkonen1998} as \begin{equation*} \Jphys(u;u_D) = \int_{\d U} |u(x) - u_D(x)|^2 \dx. \end{equation*} The objective functional measures the distance of a magnetic field $u:\Omega\to\C$ associated with a material configuration $B$ to the magnetic field $u_D:\Omega\to\C$ associated with a reference configuration $B_D$ on the observation boundary $\d U$. After discretization, we obtain \begin{align*} \Jphysdisc(\u; u_D) = \Re\left(\u^H Q \u + 2 \u^H V(u_D) + W(u_D)\right) \end{align*} where $Q\in \R^{\noDoF\times\noDoF},V(u_d)\in \C^{\noDoF}$ and $W(u_D)\in\R$ are defined element-wise for all $1\le i,j \le \noDoF$ by \begin{align*} (Q)_{ij} & := \int_{\d U} \phi_i \phi_j \rmd x, & (V(u_D))_{i} & := \int_{\d U} \phi_i u_D \rmd x, & W(u_D) & := \int_{\d U} |u_D|^2 \rmd x. \end{align*} This functional must be minimized for multiple incident plane waves $u_I(x;\omega,d) = \exp(\imath \omega d \cdot x)$ depending on wave number $\omega$ and direction $d$. Hence, let $\noWaveL$ wave numbers $\omega_1,\dotsc,\omega_\noWaveL$ and $\noDir$ directions $ d_1,\dotsc, d_\noDir$ be given, then the physical objective functional for multiple incident waves $\boldsymbol{J}^p_h$ is \begin{equation*} \boldsymbol{J}^p_h(\u^{11},\dotsc,\u^{\noWaveL\noDir}) = \sum_{l=1}^\noWaveL \sum_{m=1}^\noDir \Jphysdisc(\u^{lm};u^{lm}_D) \end{equation*} where $\u^{lm}$ is the solution of the state equation for wave number $\omega_l$, incident direction $d_m$ with incident wave $u_I(x) = u_I(x; \omega_l,d_m)$ and $u^{lm}_D$ is the corresponding reference solution, \ie $u_D^{lm}(x) = u_D(x; \omega_l,d_m)$. \subsubsection{Optimization problem} We \corr{again gather} the results of the previous sections, and extend the optimization problem \eqref{eqn:opt:probTMdis} to multiple incident waves: \begin{align*} \left \{ \qquad \begin{aligned} \min_{\B} ~ & \boldsymbol{J}^p_h(\u^{11},\dotsc,\u^{\noWaveL\noDir}) + \eta \Jregdisc(\B) + \gamma \Jgraydisc(\B) & \\ \st ~ &(\SYS(\B) + \SYS_C) \u^{lm} = (\RHS(\B) + \RHS_C) \\ & \text{with } \omega = \omega_l \text{ and } u_I(x) = u_I(x;w_l,d_m)\\ & \corr{\text{where }} \B \text{ is parametrized using polynomial interpolations} \end{aligned}\right. \end{align*} With $\b p^{lm}\in \C^\noDoF$ solving the adjoint equation \begin{align*} (\SYS(\B) + \SYS_C ) \b p ^{lm} & = - 2(Q \u ^{lm} + L(u^{lm}_D))^* \end{align*} for $\omega= \omega_l$, we \corr{are once more able} to give a formula for the derivative of the physical objective in direction $\Y\in\SCK$ for all $k\in\indexDesign$ as follows: \begin{align*} \tot{\boldsymbol{J}^p_h(\u^{11},\dotsc,\u^{\noWaveL\noDir})}{(\B)_k} [(\Y)_k] & = \sum_{l=1}^\noWaveL \sum_{m=1}^\noDir \Re\left ( \int_{\TriI{k}}(\nabla_{\!\!h} \u ^{lm} + \nabla u_I^{lm} )^T (\Y)_k\nabla_{\!\!h} \b p^{lm} \rmd x \right ). \end{align*} As in the previous example, we have used $\nabla_{\!\!h} \b u = \sum_{i=1}^\noDoF u_i\nabla \phi_i$, $\nabla_{\!\!h} \b p = \sum_{i=1}^\noDoF p_i\nabla \phi_i$ and additionally $u_I^{lm}(x) = u_I(x;\omega_l,d_m)$. \corr{With this as a basis}, again sub-problems can be formulated and solved using the techniques described in \cref{sec:polyparam}. \subsection{Numerical Results} The design domain $\Omega_D$ is a \corr{ball} with radius $0.4$ and is contained in a square with side length $2$. Furthermore a perfectly matched layer of thickness $1$ is used and the PML function $s(t)$ is defined as in the previous example. The scattered electromagnetic field $u_D$ of the reference configuration $B_D$ is calculated on the same mesh, but perturbed \corr{at} every spatial point with a weighted normal probability density function with mean value $0$ and variance $1$. The reference magnetic field $u_D$ is used for the evaluation of the objective function on the observation boundary defined by a sphere with radius $0.8$. Furthermore\corr{,} the scatterer is illuminated from $8$ directions in \SI{45}{\degree} steps and $8$ uniformly distributed wavelengths in the interval $[0.4,0.7]$. \par The set of admissible material tensors is the cyclic graph with three edges ($\indexEdges =\{1,\dotsc,3\}$) and linear interpolation of the isotropic materials $B^{(1)} = \mathds 1$, $B^{(2)} = (2)^{-2}\mathds1$ and $B^{(3)} = (1 +2\imath)^{-2}\mathds1$. \par \begin{figure}\centering \hfill \subfigure[material distribution $B_D$\label{subfig:mattarget}]{ {{\centering \parbox[b][0.25\textwidth][c]{0.5cm}{\includegraphics[height=0.16\textwidth]{programresults/cbar_mat_name.pdf}} \includegraphics[height=0.25\textwidth,angle=0]{programresults/target_shape2.png}}}} \hfill \subfigure[noisy scattered target field $u_D$\label{subfig:magnettarget}]{ \hspace{1em}{{\centering \parbox[b][0.25\textwidth][c]{0.5cm}{\includegraphics[height=0.16\textwidth]{programresults/cbar_rainbow_-2x2.pdf}} \includegraphics[height=0.25\textwidth,angle=0]{programresults/target_field.png}}}\hspace{1em}} \hfill \caption[asd]{Reference material configuration and reference magnetic field including noise.} \end{figure} \par \cref{subfig:mattarget} illustrates the material distribution $B_D$ inside the circular design domain, where white corresponds to $B^{(1)}$, green corresponds to $B^{(2)}$ and red corresponds to $B^{(3)}$, respectively. Moreover the outline of the design domain and material distribution is marked by black lines. The noisy scattered magnetic field for a fixed wavelength and illumination direction is depicted in \cref{subfig:magnettarget}. Here, the \corr{outermost} circle is the observation boundary $\d U$ where the objective functional is evaluated. \par \begin{figure}\centering \hfill \parbox[b][0.25\textwidth][c]{0.5cm}{\includegraphics[height=0.16\textwidth]{programresults/cbar_mat_name.pdf}} \hfill \subfigure[$\eta =0.1$]{\includegraphics[height=0.25\textwidth,angle=0]{{programresults/0.000001_0.000010_0.010000_0.100000}.png}} \hfill \subfigure[$\eta =1$]{\includegraphics[height=0.25\textwidth,angle=0]{{programresults/0.000001_0.000010_0.010000_1.000000}.png}} \hfill \subfigure[$\eta =10$]{\includegraphics[height=0.25\textwidth,angle=0]{{programresults/0.000001_0.000010_0.010000_10.000000}.png}} \hfill \caption[asd]{Effect of filter penalization factor $\eta$ on optimization result after 100 iterations, $\gamma= \num{e-6}$, $r_0 =0.01$, colored by material index\label{fig:regcomp}} \end{figure} \par First we wish to analyze the influence of the regularization. We set the circular filter radius to $r_0=0.01$, choose a grayness parameter of $\gamma = \num{e-6}$ and perform the optimization with various choices of regularization constants $\eta \in \{\num{0.1},\num{1},\num{10}\}$. In \cref{fig:regcomp}, the optimization result after 100 iterations is depicted for all choices; the effect of the regularization becomes apparent at the blurred interfaces. \corr{While the influence of the filter radius was also studied, this is not described here.} \par \begin{figure}\centering \hfill \subfigure[objective functional\label{subfig:objective}]{{\includegraphics[height=0.25\textwidth,angle=0]{programresults/tracking_obj.pdf}}} \hfill \subfigure[optimization result\label{subfig:optresult}]{ \includegraphics[height=0.25\textwidth,angle=0]{{programresults/rmin_0.010_pen_peri_001.0_pen_3e-07p}.png}} \hfill \subfigure[magnetic field on $\d U$ \label{subfig:observ}]{{\includegraphics[height=0.25\textwidth,angle=0]{programresults/tracking_boundary.pdf}}} \hfill \caption{Evolution of the objective functional with continuation procedure and magnetic field quantities on the observation boundary $\d U$ for single wavelength and illumination direction} \end{figure} \par During the final optimization process\corr{,} a continuation procedure like \corr{that} in \cite{Sigmund1998} was executed, \ie the algorithm \corr{was} repeatedly restarted with an increased grayness penalty factor $\gamma$ and the previous optimization result as starting point. Based on the previous investigations we \corr{fixed} $\eta=1$ and $r_0=0.01$, and \corr{initialized} the material tensor in the design domain with background material. In \cref{subfig:objective} the evolution of the objective functional value, normalized to the inital state, is depicted. The objective functional ({\tt total}) is composed of the extinctional cross section $\b J^p_h$ ({\tt physical}), the regularization term $\Jregdisc$ ({\tt filter}) and the grayness term $\Jgraydisc$ ({\tt grayness}); all contributions are plotted seperately. Due to the increment of the grayness penalty factor $\gamma$ ({\tt penalty}), the total objective functional increases but the physical and grayness contributions \corr{decrease successively}. Additionally\corr{,} the vertical lines in \cref{subfig:objective} mark the continuation procedure where either a predefined maximum number of iterations is exceeded or a stopping criterion is reached. The continuation process is ended when a so-called black/white solution ($\Jgraydisc=0$) is achieved. Thus\corr{,} we finally obtain a material distribution with tensor values corresponding to the nodes of the underlying graph (\cref{subfig:optresult}). The difference between \corr{the} target and optimized magnetic response, \ie the physical objective functional, drops to $0.2\percent$. This can be also qualitively observed in \cref{subfig:observ}, where for a single wavelength and illumination direction the magnetic response $u$ and the noisy target field $u_D$ are compared along the circular observation boundary $\d U$. \par Finally, we study the effect of the filter penalization on the optimization result in combination with the continuation strategy. \corr{ \cref{fig:filterpenopt} depicts the results of the optimization with continuation process, where the filter penalty factor $\eta\in \{0.1,10,100\}$ is varied.} \corr{It is evident} that the overall shape of the scatterer \corr{has been} more or less reconstructed. For the smallest factor $\eta = 0.1$, we see a ragged boundary, whereas for the largest factor $\eta=100$ the filtering was too strong and \corr{dominated} the tracking objective. \par In all cases the sharp features, like the edges of the star, are poorly reconstructed, which may be caused by the low number of wavelengths and illumination directions. \par \begin{figure}\centering \hfill \parbox[b][0.25\textwidth][c]{0.5cm}{\includegraphics[height=0.16\textwidth]{programresults/cbar_mat_name.pdf}} \hfill {\subfigure[$\eta=0.1$]{\includegraphics[height=0.25\textwidth,angle=0]{{programresults/rmin_0.010_pen_peri_000.1_pen_3e-07p}.png}}} \hfill {\subfigure[$\eta=10$]{\includegraphics[height=0.25\textwidth,angle=0]{{programresults/rmin_0.010_pen_peri_010.0_pen_3e-07p}.png}}} \hfill {\subfigure[$\eta=100$]{\includegraphics[height=0.25\textwidth,angle=0]{{programresults/rmin_0.010_pen_peri_100.0_pen_3e-07p}.png}}} \hfill \caption{Optimization result with variation of the filter penalty $\eta$ \label{fig:filterpenopt}} \end{figure} \section{Concluding Remarks} \corr{We have proposed a new algorithm for the solution of material optimization problems in electromagnetics.} The algorithm is flexible in the sense that it can be applied to material problems of \corr{a} discrete and continuous nature. Theoretical properties of the algorithm \corr{such as} global convergence have been discussed \corr{and} it was further shown in the course of numerical experiments that poor local optima introduced by seriously non-linear parametrizations can be avoided. \corr{An open question remains: how can the improvement of the new algorithm over the application of general purpose optimization algorithms be quantified? This will be investigated in the near future, based on extended numerical experiments and a thorough mathematical investigation of the properties of limit points.} \corr{In terms of} the application side it \corr{would be a natural next step} to apply the concept to three dimensional problems and to \corr{consider} efficient parallelization. The algorithmic concept discussed in the article seems particularly suited for the latter due to the block separable structure of the sub-problem, which \corr{must} be solved in each major iteration. \let\c\oldc\hbadness 20000\relax |
\section*{Acknowledgements}
\input{Acknowledgements}
\printbibliography
\end{document}
\section*{Acknowledgements}
\input{acknowledgements/Acknowledgements}
\clearpage
\printbibliography
\newpage \input{atlas_authlist}
\end{document}
\section{Introduction}
\label{sec:intro}
New high-mass states decaying into two photons are predicted in extensions of the Standard
Model (SM). The diphoton final state provides a clean experimental signature with excellent invariant-mass resolution and moderate backgrounds.
This Letter presents an update of the searches for new high-mass states decaying into two photons,
using both 2015 and 2016 proton--proton ($pp$) collision datasets recorded at a centre-of-mass energy $\sqrt{s}=13$~\TeV\ by the ATLAS detector at the CERN Large Hadron Collider (LHC),
corresponding to a total integrated luminosity of \intLumi. The analysis closely follows
that described in Ref.~\cite{Aaboud:2016tru}, and includes small improvements in the photon reconstruction, selection and energy calibration. From the many extensions of the SM that predict new high-mass resonances decaying into two photons,
the two benchmark signal models studied in Ref.~\cite{Aaboud:2016tru} are considered:
a spin-0 resonance ($X$) as predicted in theories with an extended Higgs sector~\cite{Hill:1987ea,Veltman:1989vw,Binoth:1996au,Schabinger:2005ei,Patt:2006fw,Pruna:2013bma,Lee:1973iz}, and the lightest Kaluza--Klein (KK)~\cite{ModernKK} spin-2
graviton excitation ($G^*$) of a Randall--Sundrum~~\cite{Randall:1999ee} model with one warped extra dimension, later referred to as RS1.
The ATLAS and CMS collaborations reported a modest excess in the diphoton invariant-mass spectra with respect to the SM continuum background near a mass value of 750~\GeV~\cite{Aaboud:2016tru,Khachatryan:2016hje}, using 3.2--3.3~\ifb\ of $pp$ collision data recorded in 2015 at the LHC.
The ATLAS result corresponds to a global significance of 2.1~standard deviations ($\sigma$).
The CMS result corresponds to a global significance of 1.6\,$\sigma$.
No significant excess is observed by CMS in 12.9~\ifb\ of data collected in 2016~\cite{Khachatryan:2016yec}.
In addition to searching for a resonant signal, data are interpreted using the model proposed by Arkani-Hamed, Dimopoulos and Dvali (ADD)~\cite{ArkaniHamed:1998rs}.
Motivated by the weakness of gravity, the ADD model predicts the existence of $n$ extra dimensions of space-time where only gravity can propagate.
If the ultraviolet cutoff scale (\ms) of the Kaluza--Klein spectrum is lower than the Planck scale in the $(4+n)$-dimensional space-time,
the extra dimensions may be detected via virtual KK graviton exchange before being observed via direct KK graviton emission.
The strength of gravity in the presence of extra dimensions is typically parameterized by $ \eta_G = F/\ms^4$,
where $F$ is a dimensionless parameter of order unity reflecting the dependence of the virtual KK graviton exchange on the number of extra dimensions.
Several theoretical formalisms exist in the literature~\cite{Giudice:1998ck,Han:1998sg,Hewett:1998sn}.
While the definition of $\eta_G$ is consistent, each formalism uses a different convention for $F$,
which consequently leads to a different definition of \ms.
The KK graviton exchange creates a set of finely spaced resonances,
which manifests itself as a non-resonant deviation from the expected SM background in the
diphoton mass distribution due to limited experimental resolution.
The effective diphoton cross section is the result of the SM and ADD amplitudes, as well as their interference. The interference term in the effective cross section is linear in $\eta_G$ and the pure KK graviton exchange term is quadratic in $\eta_G$.
The interference effect is assumed to be constructive in formalisms considered in this Letter.
Previous searches for an ADD graviton signal
in the diphoton decay channel were carried out by the ATLAS~\cite{Aad:2012cy} and CMS~\cite{Chatrchyan:2011fq} experiments based on LHC Run 1 data.
The substantial increase in the centre-of-mass energy in LHC Run 2 greatly enhances the sensitivity to the ADD scenario at higher mass scales~\cite{CMS-EXO-15-009}.
\section{Results}
\label{sec:results}
Since the results of the 2015 data analysis were published in Ref.~\cite{Aaboud:2016tru}, photon reconstruction and energy calibration have been improved. The results for both the spin-0 and spin-2 resonance searches are updated as summarized below. In the light of the modest excess observed in the 2015 data alone, results are reported first considering the 2015 and 2016 data samples individually, before reporting the combined results.
\begin{itemize}
\item \textbf{Spin-0}:
the largest local deviation from the background-only hypothesis in the 2015 dataset
is \localSigSpinZeroFifteen\,$\sigma$ at a mass of \bestMassSpinZeroFifteen~\GeV\ with a relative width of \bestRelativeWidthSpinZeroFifteen.
The results obtained on the same dataset with the old reconstruction and calibration, published in Ref.~\cite{Aaboud:2016tru}, presented a deviation of 3.9\,$\sigma$ at 750~\GeV\ with a relative width of 6\%.
\item \textbf{Spin-2}:
the largest local deviation from the background-only hypothesis in the mass range 700--800~\GeV\ in the 2015 dataset is \localSigSpinTwoFifteen\,$\sigma$ at a mass of \bestMassSpinTwoFifteen~\GeV\ with a \coupling\ value of \bestCouplingSpinTwoFifteen.
The results obtained on the same dataset with the old reconstruction and calibration, published in Ref.~\cite{Aaboud:2016tru}, presented a deviation of 3.8\,$\sigma$ at 750~\GeV\ with a \coupling\ value of 0.23.
\end{itemize}
In the 2016 dataset, the observations are summarized as follows.
\begin{itemize}
\item \textbf{Spin-0}:
the largest local deviation from the background-only hypothesis corresponds to a 2.0\,$\sigma$ narrow-width excess at 304~\GeV.
Within the mass interval 700--800~\GeV, the largest local deviation from the background-only hypothesis corresponds to a \localSigSpinZeroSixteen\,$\sigma$ \bestWidthSpinZeroSixteen\ excess at \bestMassSpinZeroSixteen~\GeV.
\item \textbf{Spin-2}: the largest local deviation from the background-only hypothesis
corresponds to a \localSigSpinTwoSixteen\,$\sigma$ excess at \bestMassSpinTwoSixteen~\GeV\ with a best-fit \coupling\ value of \bestCouplingSpinTwoSixteen.
\end{itemize}
The results combining the 2015 and 2016 datasets are summarized below. A complete set of tables with the full limit results, including those from additional width and \coupling\ scenarios not covered in this Letter, are available at the Durham HepData repository.
\begin{figure}[t]
\begin{center}
\subfloat[\label{fig:mgg_distributions_spin0}]{\includegraphics[width=.49\textwidth]{figures/mass_spin0_bkg.pdf}}
\subfloat[\label{fig:mgg_distributions_spin2}]{\includegraphics[width=.49\textwidth]{figures/mass_spin2_bkg.pdf}}
\end{center}
\caption{Distributions of the diphoton invariant mass for events passing (a) the spin-0 selection or (b) the spin-2 selection,
with the background-only fits superimposed.
The data points are plotted at the centre of each bin.
The error bars indicate statistical uncertainties.
The differences between the data and the fits are shown in the bottom panels.
The arrows in the lower panels indicate values outside the range by more than one standard deviation.
There is no data event with $m_{\gamma\gamma}>2700$~\GeV.
}
\label{fig:mgg_distributions}
\end{figure}
\begin{itemize}
\item \textbf{Spin-0}:
the diphoton invariant-mass distribution of the events passing the spin-0 selection is shown in Figure~\ref{fig:mgg_distributions_spin0}.
The compatibility of the data with the background-only hypothesis as a function of both the assumed mass and width values is
shown in Figure~\ref{fig:p0_2D_spin0}.
The largest deviation from the background-only hypothesis is observed at a mass of \bestMassSpinZero~\GeV\ for a narrow width,
with a local $p_0$ of \localSigSpinZero\,$\sigma$.
The corresponding global significance is null,
as the local deviation is less than the median largest deviation in background-only pseudo-experiments in the search region defined by 200--2700~\GeV\ in resonance mass and 0--10\% in relative width.
Figure~\ref{fig:limit_spin0} shows the upper limits on the signal fiducial cross section times branching ratio to two photons for a narrow-width (4~\MeV) spin-0 resonance as a function of its mass.
The limit on the fiducial cross section times branching ratio ranges from \maximumLimitSpinZeroNWA\
at 200~\GeV\ to about \minimumLimitSpinZeroNWA\ at 2700~\GeV.
The impact of the systematic uncertainties on the expected limit decreases with the resonance mass from 29\% at 200~\GeV\ to 5\% at 700~\GeV. Above 700~\GeV, the impact is typically 2--3\%.
\item \textbf{Spin-2}:
the invariant-mass distribution of the events passing the spin-2 selection is shown in Figure~\ref{fig:mgg_distributions_spin2}.
The local compatibility of the data with the background-only hypothesis
as a function of both the resonance mass and \coupling\ values
is shown in Figure~\ref{fig:p0_2D_spin2}.
The largest deviation from the background-only hypothesis in the combined dataset is observed for a mass of \bestMassSpinTwo~\GeV, and $\coupling$ of \bestCouplingSpinTwo,
corresponding to \localSigSpinTwo\,$\sigma$ local $p_0$.
The global significance, estimated from pseudo-experiments in the search region of 500--2700~\GeV\ in mass and 0.01--0.3 in \coupling,
is \globalSigSpinTwo\,$\sigma$.
Figure~\ref{fig:limit_spin2} shows the limits on the KK graviton cross section times branching ratio to two photons as a function of the resonance mass with $\coupling=0.1$.
The cross sections predicted by the benchmark model are computed at LO in QCD using \textsc{Pythia8}, and are shown on the same figure.
The uncertainty band on the predictions represents the PDF uncertainty estimated from the variations of the NNPDF23LO PDF set.
The observed limits on the cross section times branching ratio range from \maximumLimitSpinTwoSmallCouplings\ to about \minimumLimitSpinTwoSmallCouplings\ for a KK graviton mass between 500~\GeV\ and 5000~\GeV.
The RS1 model with $\coupling=0.1$ is excluded for $m_{G^{*}}$ below \limitRS, based on the observed limit determined with pseudo-experiments.
The impact of systematic uncertainties on the expected limit is below 5\% over the entire mass range, and typically 2–-3\% below 2~\TeV.
\item[] The spin-2 spectrum shown in Figure~\ref{fig:mgg_distributions_spin2} is also interpreted in the context of the ADD model.
A counting experiment is performed in the signal region $\mgg>2240~\GeV$.
In this region, four events are observed in data for $4.3\pm1.0$ expected.
The expected 95\% CL upper limit on the number of excess events is 5.4, and the observed limit has the same value.
The limit on the number of events can be translated into a lower limit on the ultraviolet cutoff scale \ms\ in the ADD model for different theoretical formalisms as summarized in Table~\ref{tab:limit_para}.
Besides the GRW formalism used for the simulation, the Han--Lykken--Zhang (HLZ)~\cite{Han:1998sg} formalism and the Hewett~\cite{Hewett:1998sn} formalism with positive interference are also considered.
A K-factor of about 1.4 was computed using ADD samples generated at LO and NLO using \textsc{MadGraph5\_aMC@NLO},
and is included in the results to indicate the potential impact from the higher-order calculation.
The uncertainty in the signal theory prediction typically varies the limit results by 8\%.
\end{itemize}
\begin{figure}[t]
\begin{center}
\subfloat[\label{fig:p0_2D_spin0}]{\includegraphics[width=.49\textwidth]{figures/2dp0_spin0.pdf}}
\subfloat[\label{fig:p0_2D_spin2}]{\includegraphics[width=.49\textwidth]{figures/2dp0_spin2.pdf}}
\end{center}
\caption{Compatibility, in terms of local $p_0$ quantified in standard deviations $\sigma$,
with the background-only hypothesis (a) as a function of the assumed signal mass $m_X$ and relative width $\Gamma_X/m_X$ for the spin-0 resonance search and (b) as a function of the assumed signal mass $m_{G^*}$ and \coupling\ for the spin-2 resonance search. Only positive excesses are considered.
}
\label{fig:p0_2D}
\end{figure}
\begin{figure}[!ht]
\begin{center}
\subfloat[\label{fig:limit_spin0}]{\includegraphics[width=.49\textwidth]{figures/limit_spin0_gamma0.pdf}}
\subfloat[\label{fig:limit_spin2}]{\includegraphics[width=.49\textwidth]{figures/limit_spin2_0d10.pdf}}
\end{center}
\caption{(a) Upper limits on the fiducial cross section times branching ratio to two photons at $\sqrt{s} = 13$~\TeV\ of a narrow-width ($\Gamma_X$ = 4~\MeV) spin-0 resonance as a function of its mass $m_X$. (b) Upper limits on the production cross section times branching ratio to two photons at $\sqrt{s} = 13$~\TeV\ of the lightest KK graviton as a function of its mass $m_{G^*}$ for $\coupling=0.1$.
For $m_{G^*}>2500~\GeV$, the observed and expected limits are determined with pseudo-experiments shown by the blue solid and dashed lines, respectively.
Predictions are shown for the RS1 model, where the grey shaded band represents the PDF uncertainty.
}
\label{fig:limit}
\end{figure}
\begin{table}[htbp]
\begin{center}
{\small
\begin{tabular}{|c|l||c|c|cccc|} \hline
& ADD formalism & GRW & Hewett & \multicolumn{4}{c|}{HLZ} \\
& Parameter & & positive & $n = 3$ &$n = 4$ &$n = 5$ &$n = 6$ \\ \hline \hline
Without K-factor & $\ms$ observed limit [\TeV] & 6.8 & 6.1 & 8.1 & 6.8 & 6.1 & 5.7 \\\hline
With K-factor & $\ms$ observed limit [\TeV] & 7.2 & 6.5 & 8.6 & 7.2 & 6.5 & 6.1 \\\hline
\end{tabular}
}
\caption{Observed 95\% CL lower limits on the ADD model parameter \ms\ for the GRW, the Hewett (with positive interference) and the HLZ formalisms.
For the HLZ formalism, the number of extra dimensions ($n$) is varied between 3 and 6.
The second row shows results based on MC samples generated at LO by \textsc{Sherpa 2.2.1}.
A K-factor was computed using ADD samples generated at LO and NLO using \textsc{MadGraph5\_aMC@NLO},
and is included in the results shown in the third row to indicate the potential impact from the higher-order calculation.
}
\label{tab:limit_para}
\end{center}
\end{table}
\section{Data and simulated event samples}
\label{sec:sample}
Data were collected in 2015 and 2016 using $pp$ collisions at a centre-of-mass energy of $\sqrt{s}=13$~\TeV\
with a bunch spacing of 25~ns.
The average number of $pp$ interactions per bunch crossing is 13 in 2015 and 25 in 2016,
with a peak instantaneous luminosity up to \peakLumi.
Events from $pp$ collisions are recorded using a diphoton trigger with transverse energy ($\et$)
thresholds of 35~\GeV\ and 25~\GeV\
for the \et-ordered leading and subleading photon candidates, respectively.
In the high-level trigger, the shapes of the energy depositions in the EM calorimeter are required to match those expected for electromagnetic showers initiated by photons.
The trigger has a signal efficiency close to 100\% for events fulfilling the final event selection, with an uncertainty below $0.4\%$.
After applying data-quality requirements,
the data sample corresponds to an integrated luminosity of \intLumiFifteen\ for the 2015 data and
\intLumiSixteen\ for the 2016 data.
The measurement of the integrated luminosity has an uncertainty of 2.1\% for the 2015 data and 3.4\% for the 2016 data.
The uncertainties in the 2015 and 2016 integrated luminosities are derived
following a methodology similar to that detailed in Ref.~\cite{Aaboud:2016hhf},
from a calibration of the luminosity scale using a $x$--$y$ beam-separation scan performed in August 2015, and a preliminary calibration using a scan performed in May 2016, respectively.
The correlation between the two years' luminosity uncertainties is taken into account.
Simulated Monte Carlo (MC) events are used for optimizing the search strategy~\cite{Aaboud:2016hhf},
and for the signal and background modelling studies detailed in Sections~\ref{sec:signal_modelling} and \ref{sec:background}, respectively.
Interference effects between the resonant signal and the background processes are neglected.
The spin-0 signal MC samples were generated using the effective-field-theory approach implemented in
\textsc{MadGraph5\_aMC@NLO}~\cite{mg5anlo} version 2.3.3 at next-to-leading order (NLO) in quantum chromodynamics (QCD).
From the Higgs characterization framework~\cite{HCframework},
CP-even dimension-five operators coupling the new resonance to gluons and photons were included.
Samples were generated with the NNPDF3.0 NLO parton distribution functions (PDFs)~\cite{nnpdf30nlo}, using the A14
set of tuned parameters (tune) of \textsc{Pythia 8.186}~\cite{Pythia8, ATL-PHYS-PUB-2014-021} for the parton-shower and hadronization simulation.
Simulated samples were produced for fixed values of the mass and width of the assumed resonance, spanning the range 200--2400~\GeV\
for the mass, and the range from 4~\MeV\ to 10\% of the mass for the decay width.
Choosing an improved signal model with an event generator different from the one used in Ref.~\cite{Aaboud:2016tru} provides a description of the signal which is less sensitive to modelling effects from the off-shell region.
The impact of this change is only visible in scenarios with a large signal decay width, with mass values at the \TeV\ scale.
Spin-2 signal samples for the RS1 model were generated using \textsc{Pythia 8.186},
with the NNPDF23LO PDF set~\cite{nnpdf23lo} and the A14 tune.
Only the lightest KK graviton excitation was generated.
Its mass $m_{G^{*}}$ was varied in the range between 500~\GeV\ and 5000~\GeV.
The dimensionless coupling \coupling, where $\mplRed = \mpl/\sqrt{8\pi}$ is the reduced Planck scale
and $k$ the curvature scale of the extra dimension, is assumed to be in the range 0.01 to 0.3.
For $\coupling < 0.3$, the KK graviton is expected to be a fairly narrow resonance~\cite{Davoudiasl:1999jd},
with a width given by $1.44 (\coupling)^2 m_{G^*}$.
The non-resonant KK graviton signal was simulated using the ADD model with the representation proposed by Giudice, Rattazzi and Wells (GRW)~\cite{Giudice:1998ck}.
The \textsc{Sherpa} event generator (version 2.1.1) was used to simulate both the SM and ADD processes at the same time, including their interference.
The ultraviolet cutoff scale \ms\ was varied between 3500~\GeV\ and 6000~\GeV\ in the simulation.
Events containing two prompt photons,
representing an irreducible background to the search,
were simulated using the \textsc{Sherpa}~\cite{Gleisberg:2008ta} event generator, version 2.1.1.
Matrix elements were calculated with up to two additional partons at leading order (LO) in QCD
and merged with the \textsc{Sherpa} parton-shower simulation~\cite{Schumann:2007mg}
using the ME+PS@LO prescription~\cite{Hoeche:2009rj}.
The gluon-induced box process was also included.
The \textsc{CT10} PDF set~\cite{Lai:2010vv} was used in conjunction with a dedicated parton-shower tune of \textsc{Sherpa}.
Samples of the photon+jet reducible background component were also generated using \textsc{Sherpa}, version 2.1.1,
with matrix elements calculated at LO with up to four additional partons.
The same PDF set, parton-shower tune and merging prescription as for the diphoton sample were used.
To study the dependence of data composition results (Section~\ref{sec:event_selection}) on the event generator,
\textsc{Pythia8} was also used to generate SM diphoton events,
based on the LO quark--antiquark $t$-channel
diagram and the gluon-induced box process, and photon+jet events.
The same PDF set and parton-shower tune as for the RS1 model signal samples were used.
The generated events were passed through a full detector simulation~\cite{Aad:2010ah} based on \textsc{Geant4}~\cite{Geant4}.
Pile-up from additional $pp$ collisions in the same or neighbouring bunch crossings was simulated by overlaying each MC
event with a variable number of simulated inelastic $pp$ collisions,
generated using \textsc{Pythia8} with the AZNLO tune~\cite{AZNLO:2014}. The MC events were weighted to reproduce
the distribution of the average number of interactions per bunch crossing observed in the data.
\section{Event selection}
\label{sec:event_selection}
The event selection criteria are the same as those described in Ref.~\cite{Aaboud:2016tru}.
Photon candidates are reconstructed from clusters of energy deposited in the EM calorimeter,
and may have tracks and conversion vertices reconstructed in the ID.
They are required to be in a fiducial region of the EM calorimeter defined by $|\eta| < 2.37$, not including the transition region $1.37<|\eta|<1.52$ between the barrel and end-cap calorimeters.
Compared to the processing used in Ref.~\cite{Aaboud:2016tru}, the reconstruction of converted photons from stand-alone TRT tracks was re-optimized to improve the efficiency and to cope with higher pile-up in the 2016 data-taking period.
For an average number of interactions per bunch crossing close to 40, the re-optimized algorithm maintains the same reconstruction efficiency for genuine photon conversions, while the rate of unconverted photons reconstructed as converted is reduced from around 20\% to around 5\%.
In addition, a small bias in the track parameters was corrected.
These changes lead to small event-by-event differences between the two reconstructions in the classification of the photons as either converted or unconverted.
The two photon candidates with the highest transverse energies in each event, satisfying $\et > 40$~\GeV\ and 30~\GeV, are retained. The energy measurement is based on a multivariate regression algorithm~\cite{Aad:2014nim} used to determine corrections to the energy of the clusters, developed and optimized on simulated events. The calibration of the energies deposited in each layer of the calorimeter, the overall energy scale and energy resolution are determined in situ. The regression algorithm was retrained to account for the small changes in the conversion reconstruction. For photons near the transition region between the barrel and end-cap calorimeters,
the information from the scintillators located in front of the end-cap cryostat usually improves the energy resolution by a few percent compared to Ref.~\cite{Aaboud:2016tru}, although in a few rare cases a significant shift of the measured energy may occur.
At \et\ values larger than 100--200~\GeV\ the energy resolution is dominated by the constant term
of the calorimeter energy resolution, which varies from 0.9\% to 2.0\% in different $\eta$ regions.
The uncertainty in the photon energy scale for $\et > 100$~\GeV\ is typically 0.5--1.5\% depending on $\eta$.
The uncertainty in the photon energy resolution is driven by the uncertainty in the constant term in the \et\ range relevant to this analysis.
At $\et = 300~\GeV$, the relative uncertainty is 30--40\% depending on $\eta$.
Photons are required to fulfil \emph{tight} identification criteria~\cite{Aaboud:2016yuq} based on variables that measure the shape of the electromagnetic showers in the calorimeter (``shower shapes''), in particular in the finely segmented first layer. The efficiency of the photon identification increases with \et\ from 90\% at 50~\GeV\ to 95\% at 200~\GeV. The associated $\eta$-dependent uncertainties were measured in the whole dataset from 2015 and 2016 using the same methods as in Ref.~\cite{Aaboud:2016yuq}. They vary between 0.2\% and 4\% below 200~\GeV, and between 1\% and 4\% above.
To further reject the background from jets misidentified as photons, the candidates are required to be isolated using both calorimeter and tracking detector information. The scalar sum of the \et\ of energy clusters within a cone of size $\Delta R = \sqrt{(\Delta\eta)^2 + (\Delta\phi)^2} = 0.4$ around the photon candidate, excluding the photon energy deposits and correcting for pile-up and underlying-event contributions~\cite{STDM-2010-08,Cacciari:2008gn,Cacciari:2009dp}, is required to be below $0.022 \et + 2.45$~\GeV,
where $\et$ is the transverse energy of the photon candidate. The sum of the tranverse momentum (\pt) of tracks within $\Delta R = 0.2$ of the photon candidate, not including tracks associated with a photon conversion, is required to be below $0.05 \et$.
In order to minimize pile-up effects and to improve the separation between signal and background processes, tracks with $\pt < 1$~\GeV\ are excluded from the sum. Tracks are also excluded if they have a large impact parameter with respect to the primary vertex identified by combining the photon direction measurements from the calorimeter with the tracking information from the ID as described in Ref.~\cite{Aaboud:2016tru}.
The efficiency of the combined isolation requirement for photons fulfilling the tight identification selection in signal MC
samples is 88\% to 97\% in the \et\ range from 100~\GeV\ to 500~\GeV, with an uncertainty around 1\%. Compared to Ref.~\cite{Aaboud:2016tru}, improvements in the selection of tracks associated with photon conversions lead to a few percent increase in the efficiency of the isolation requirement for genuinely converted photons.
Different kinematic selections~\cite{Aaboud:2016tru} are applied in the searches for spin-0 and spin-2 signals to exploit the kinematic properties of the decay photons. In the selection used to search for a spin-0 resonance, the transverse energy is required to be $\et > 0.4 m_{\gamma\gamma}$ for the leading photon
and $\et > 0.3 m_{\gamma\gamma}$ for the subleading photon, for a given value of the diphoton invariant mass $m_{\gamma\gamma}$.
With these requirements, \yieldSpinZero\ (\yieldSpinZeroHighMass)~events are selected in the data with $m_{\gamma\gamma} >150$~\GeV\ ($>200$~\GeV). In the selection used to search for spin-2 resonant and non-resonant signals, the transverse energy of each photon is required to be \et~$>$~55~\GeV.
With these requirements, \yieldSpinTwo\ diphoton events with $m_{\gamma\gamma} >$~200~\GeV\ are selected in the data.
The search for a non-resonant signal uses only events with $\mgg>2240$~\GeV,
which is optimized for the expected limit on the ultraviolet cutoff scale \ms\ in the ADD model using the signal and background samples described in Section~\ref{sec:sample}.
The composition of the selected data sample is studied using three methods detailed in Ref.~\cite{Aaboud:2016tru}.
The 2$\times$2D sideband method and the matrix method are based on sidebands constructed by inverting photon identification and isolation requirements.
The results from these two methods are in good agreement as shown in Figure~\ref{fig:bkg_composition}.
The selected samples consist mainly of events from diphoton production, with a purity estimated by the 2$\times$2D sideband method to be $(91^{+3}_{-7})$\% for the spin-0 selection and $(91^{+3}_{-8})$\% for the spin-2 selection,
increasing by a few percent with \mgg.
Uncertainties in these purity estimates originate from the statistical uncertainty in the data sample,
the definition of the control region failing the tight identification requirement,
the event generator dependence (difference between \textsc{Sherpa 2.2.1} and \textsc{Pythia8}),
the modelling of the isolation and shower-shape distributions,
and possible correlations between the isolation variables and the inverted identification criteria.
The remaining background is mostly composed of photon+jet and dijet production, with one or two jets misidentified as photons. Backgrounds from other sources are negligible.
The composition derived from the 2$\times$2D sideband method is used to select the function to model the background in the spin-0 resonance search.
In the background estimate used for the spin-2 resonance and non-resonant signal searches, the composition of the data
sample is determined by a third method, which exploits the isolation profile of the two photons in the calorimeter, with consistent results obtained for the spin-2 selection as mentioned above.
\begin{figure}[!]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/purity_spin0.pdf}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/purity_spin2.pdf}}
\end{center}
\caption{The diphoton invariant-mass distributions of the data are shown in the upper panels for (a) the spin-0
and (b) the spin-2 selections and their decomposition into contributions
from genuine diphoton ($\gamma\gamma$), photon+jet ($\gamma j$ and $j \gamma$) and dijet ($jj$) events as determined using the 2x2D sideband method.
The bottom panels show the purity of diphoton events as determined by the matrix method and the 2$\times$2D sideband method.
Each point in the distributions is plotted in the centre of the corresponding bin.
The total uncertainties, including statistical and systematic components added in quadrature, are shown as error bars.
}
\label{fig:bkg_composition}
\end{figure}
\part*{Auxiliary material}
\addcontentsline{toc}{part}{Auxiliary material}
\begin{figure}[!ht]
\centering
\includegraphics[width=1.0\columnwidth]{figures/aux/JiveXML_309516_3908489606-2p6_1GeVIndetCut.pdf}
\caption{Event display of the highest-invariant-mass ($m_{\gamma\gamma}=2.57$~\TeV) diphoton candidate event selected in data.
The event passes both the spin-0 selection and the spin-2 selection.
Only inner-detector tracks with $\pt>5$~\GeV, calorimeter cells with transverse energy deposits larger than 1~\GeV\ and
cells of the tile hadronic calorimeter with transverse energy above 3~\GeV\ are shown.
Both photon candidates are reconstructed as unconverted.
The leading photon candidate has $\ET=1345$~\GeV, $\eta=-0.09$, $\phi=-2.42$, $\ETiso = -0.15$~\GeV.
The subleading photon candidate has $\ET=1227$~\GeV, $\eta=-0.06$, $\phi=0.69$, $\ETiso = 2.29$~\GeV.}
\label{fig:evt_display_a}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=1.0\columnwidth]{figures/aux/JiveXML_303291_327585180-2p45_1GeVIndetCut.pdf}
\caption{Event display of the second-highest-invariant-mass ($m_{\gamma\gamma}=2.45$~\TeV) diphoton candidate event selected in data.
The event passes the spin-2 selection but not the spin-0 selection.
Only inner-detector tracks with $\pt>5$~\GeV, calorimeter cells with transverse energy deposits larger than 1~\GeV\ and
cells of the tile hadronic calorimeter with transverse energy above 3~\GeV\ are shown.
The leading photon candidate is reconstructed as converted and has $\ET=443$~\GeV, $\eta=2.24$, $\phi=-2.14$, $\ETiso = -1.32$~\GeV.
The subleading photon candidate is reconstructed as unconverted and has $\ET=412$~\GeV, $\eta=-1.18$, $\phi=-0.96$, $\ETiso = -0.32$~\GeV.}
\label{fig:evt_display_b}
\end{figure}
\FloatBarrier
\begin{figure}[!h]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/resolution_spin0_800GeV_nwa.pdf}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/resolution_spin0_800GeV_Gamma06.pdf}}\\
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/resolution_spin2_1000_kappa01.pdf}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/resolution_spin2_1000_kappa2.pdf}}
\end{center}
\caption{The normalized $m_{\gamma\gamma}$ distributions for different signal hypotheses
for a scalar resonance with a mass of 800~\GeV,
and (a) with a narrow decay width and (b) with relative width $\Gamma_{X}/m_{X}$ = 0.06,
and for a KK graviton predicted by the RS1 model with a mass of 1000~\GeV,
and (c) with \coupling\ = 0.01 and (d) with \coupling\ = 0.2.
In all cases, a parameterized signal model is superimposed using the convolution of the signal mass
line-shape with the overall detector resolution.}
\label{fig:dscb_fit}
\end{figure}
\begin{figure}[!h]
\begin{center}
\subfloat[]{\includegraphics[width=.49\textwidth]{figures/aux/mass_spin0_2015_bkg}}
\subfloat[]{\includegraphics[width=.49\textwidth]{figures/aux/mass_spin2_2015_bkg}}
\end{center}
\caption{Distributions of the diphoton invariant-mass for events from 2015 data passing (a) the spin-0 or (b) the spin-2 selection,
with the background-only fits superimposed. The differences between the data and the fits are shown in the bottom panels.
The arrows in the lower panels indicate values outside the range by more than one standard deviation.
There is no data event with $m_{\gamma\gamma}>2000$~\GeV.
}
\label{fig:mgg_distributions2015}
\end{figure}
\begin{figure}[!h]
\begin{center}
\subfloat[]{\includegraphics[width=.49\textwidth]{figures/aux/mass_spin0_2016_bkg}}
\subfloat[]{\includegraphics[width=.49\textwidth]{figures/aux/mass_spin2_2016_bkg}}
\end{center}
\caption{Distributions of the diphoton invariant-mass for events from 2016 data passing (a) the spin-0 or (b) the spin-2 selection,
with the background-only fits superimposed. The differences between the data and the fits are shown in the bottom panels.
The arrows in the lower panels indicate values outside the range by more than one standard deviation.
There is no data event with $m_{\gamma\gamma}>2700$~\GeV.
}
\label{fig:mgg_distributions2016}
\end{figure}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin0_0}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin0_0_02}} \\
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin0_0_06}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin0_0_10}}
\end{center}
\caption{Compatibility with the background-only hypothesis as a function of the assumed signal mass $m_X$
for different values of the relative width $\Gamma_X/m_X$ from the spin-0 resonance search.}
\label{fig:1dp0_spin0}
\end{figure}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin2_0_01}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin2_0_10}} \\
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin2_0_20}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/1dp0_spin2_0_30}}
\end{center}
\caption{Compatibility with the background-only hypothesis as a function of the assumed mass of the KK graviton predicted by the RS1 model,
for different values of \coupling\ from the spin-2 resonance search.}
\label{fig:1dp0_spin2}
\end{figure}
\begin{figure}[!h]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/acceptance_efficiency_spin0.pdf}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/acceptance_efficiency_spin2.pdf}}\\
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/acceptance_efficiency_add.pdf}}
\end{center}
\caption{Fiducial acceptance (full points) and product of the fiducial acceptance times the combined reconstruction and identification efficiency (``Detector selection'', open points)
as a function of the mass
for (a) a spin-0 resonant signal with narrow width under the spin-0 selection,
(b) a \mbox{spin-2} resonant signal with $\coupling=0.1$ under the spin-2 selection, and
(c) an ADD excess (GRW formalism), calculated as the difference between the sum of ADD and SM contributions (including their interference) and the SM contribution, under the spin-2 selection in the signal region $\mgg>2240$~\GeV.
For larger values of \ms, the acceptance decreases to about 58\% at 8~\TeV\ due to a larger contribution from the interference term, which has smaller acceptance.
The definition of fiducial region used for (a) is provided in the Letter text.
The definition used for (b) is provided in JHEP 09 (2016) 001,
and $\mgg>2240$~\GeV is required in addition for (c).
Only MC statistical uncertainties are included.
}
\label{fig:acceptance_efficiency}
\end{figure}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/limit_spin0_0_02}} \\
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/limit_spin0_0_06}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/limit_spin0_0_10}}
\end{center}
\caption{Upper limits, calculated using the asymptotic approximation, on the fiducial cross section of a spin-0 particle at $\sqrt{s} = 13$~\TeV\ times branching ratio to two photons as a function of the assumed mass $m_X$,
for different
values of the decay width divided by the mass.
}
\label{fig:spin0_limit}
\end{figure}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/limit_spin2_0_01}}\\
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/limit_spin2_0_20}}
\subfloat[]{\includegraphics[width=.5\textwidth]{figures/aux/limit_spin2_0_30}}
\end{center}
\caption{Upper limits, calculated using the asymptotic approximation, on the production cross section of a KK graviton at $\sqrt{s} = 13$~\TeV\ times branching ratio to two photons as a function of the assumed mass, for different
values of \coupling.
The predicted cross sections times branching ratio to two photons in the RS1 model, computed at LO, are also shown. The uncertainty
in the cross-section values represents the PDF uncertainty.
}
\label{fig:spin2_limit}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=.8\textwidth]{figures/aux/data_bkg_ADD_B_M}
\caption{
Observed invariant-mass distribution of diphoton events from the spin-2 selection,
with the predicted SM background and expected signals for one KK graviton predicted by the RS1 model with $\mG=4~\TeV$,
as well as two values of the ADD model (GRW formalism) parameter \ms\ (3.5~\TeV, 6~\TeV) superimposed.
The vertical dashed line and the arrow indicate the signal region of $\mgg>2240$~\GeV where the counting experiment is performed for the spin-2 non-resonant signal search.
}
\label{fig:bkgADD}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=.8\textwidth]{figures/aux/ADD_Limits_with_Kfactor}
\caption{95\% CL upper limits on the number of excess events as a function of the ADD model (GRW formalism) parameter \ms. The black horizontal line corresponds to the observed limit, the dashed line corresponds to the expected limit and the green (yellow) band to the $\pm 1\sigma$ ($\pm 2\sigma$) uncertainty in the expected limit. The red curve corresponds to MC predictions from samples with different \ms\ values
with the K-factor applied.
The band around it illustrates the PDF and scale uncertainties in the signal. When the prediction is greater than the limit line, the corresponding value of \ms\ is excluded.
}
\label{fig:limitADD}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=.8\textwidth]{figures/aux/bkg_systematics_summary_spin2.pdf}
\caption{Relative pre-fit uncertainties in the shape of the \mgg\ distribution of the predicted background for the spin-2 resonance and non-resonant searches. The uncertainties are shown in the mass range 200~\GeV\ to 5000~\GeV, which is corresponding to the range where the signal-plus-background fit is performed in the resonance search. The reducible background uncertainty corresponds to the uncertainty in the shape of the reducible background component. The uncertainty in the shape of the irreducible background results from uncertainties affecting the NLO diphoton computation (parton distribution functions and factorization and renormalization scales). The uncertainty in the purity corresponds to the impact of the relative normalization of the reducible background compared to the irreducible background. The uncertainty in isolation results from the uncertainty due to the choice of parton-level isolation cut in the \textsc{Diphox} NLO computation.}
\label{fig:spin_2_syst_summary}
\end{figure}
\begin{figure}[hbpt]
\centering
\includegraphics[width=0.8\columnwidth]{figures/aux/DiphoxSherpa}
\caption{Effective $k$-factor correction as a function of diphoton mass obtained from the ratio \textsc{Diphox}/\textsc{Sherpa} at truth level for the irreducible background estimation used in the spin-2 searches.
To focus on the shape comparison, the histogram has been normalized in such a way that that the integral of the ratio in the region 200--300~\GeV\ is equal to unity.
The solid line represents the result of a fit to a smooth function.
}
\label{fig:BkgkFactor}
\end{figure}
\begin{figure}[hbpt]
\centering
\subfloat[]{\includegraphics[width=0.48\columnwidth]{figures/aux/GravitonAnal_LeadFinal.pdf}}
\subfloat[]{\includegraphics[width=0.48\columnwidth]{figures/aux/GravitonAnal_SubLeadFinal.pdf}}
\caption{Modified isolation energies $\ETisomod = \ETiso - 0.022\et$ for (a) the leading and (b) the subleading photon candidate for events selected in the low-mass control region ($200\, \GeV < \mgg < 510\, \GeV$) of the spin-2 analyses.
The result of the two-dimensional template fit, used for the determination of the background composition in the spin-2 searches, is shown in black, together with the contributions of the different jet (green lines) and photon (red lines) components.
The bin-by-bin significance of the difference between data and the fit is shown in the bottom panels.
}
\label{fig:Fit2D}
\end{figure}
\begin{table}
\centering
\begin{tabular}{lc}
\hline
SM diphoton process [fb] & 1.555 $\pm$ 0.002 \\
\hline
\multirow{2}{*}{ADD scenario} & Total cross section times \\
& branching ratio to two photons [fb] \\
\hline
$\ms=3500$~\GeV & 16.47 $\pm$ 0.07 \\
$\ms=4000$~\GeV & 8.60 $\pm$ 0.04 \\
$\ms=4500$~\GeV & 5.09 $\pm$ 0.02 \\
$\ms=5000$~\GeV & 3.44 $\pm$ 0.01 \\
$\ms=5500$~\GeV & 2.58 $\pm$ 0.01 \\
$\ms=6000$~\GeV & 2.159 $\pm$ 0.002 \\
$\ms=6500$~\GeV & 1.927 $\pm$ 0.002 \\
$\ms=7000$~\GeV & 1.797 $\pm$ 0.002 \\
$\ms=7500$~\GeV & 1.716 $\pm$ 0.002 \\
$\ms=8000$~\GeV & 1.667 $\pm$ 0.002 \\
\hline
\end{tabular}
\caption{
Cross section of SM processes producing two prompt photons,
and cross section times branching ratio to two photons in the presence of ADD model
calculated at leading order using \textsc{Sherpa} for different values of the ADD model parameter \ms\ in the GRW formalism. The uncertainties correspond to the statistical uncertainties.
In all the calculations, the following requirements are included in \textsc{Sherpa}: both photons with $\et > 5$~\GeV, $\Delta R$ between the two photons larger than 0.2, $\mgg > 1800$~\GeV. In addition, the Frixione isolation (Phys. Lett. B 429 (1998) 369) for both photons satisfies $\delta_0 = 0.3$, exponent $n = 2$, $\epsilon = 0.025$.
}
\label{tab:add_xsection}
\end{table}
\begin{table}
\centering
\begin{tabular}{lcc}
\multirow{2}{*}{Selection} & Efficiency & Efficiency \\
& ($\ms=3500$~\GeV) & ($\ms=6000$~\GeV) \\
\hline
Skimming (including trigger) & 93\% & 95\% \\[3mm]
Two photons passing basic shower shape & \multirow{2}{*}{81\%} & \multirow{2}{*}{83\%} \\
requirements and fiducial selection & & \\[3mm]
Trigger matching & 81\% & 82\% \\[3mm]
Photon identification & 70\% & 70\% \\[3mm]
Photon isolation & 67\% & 66\% \\[3mm]
$\mgg>2240$~\GeV & 46\% & 51\% \\[3mm]
Photon $\et > 55$~\GeV & 46\% & 51\% \\[3mm]
\hline
\end{tabular}
\caption{Selection efficiency (in percentage) for simulated ADD events after each event selection requirement in the spin-2 non-resonant search.
The SM expectation has been subtracted from all the calculations.
The simulated \textsc{Sherpa} samples include the following event-generator-level requirements: both photon \et\ larger than 5~\GeV, $\Delta R$ between the two photons larger than 0.2, $\mgg > 1800$~\GeV. In addition, the Frixione isolation (Phys. Lett. B 429 (1998) 369) for both photons satisfies $\delta_0 = 0.3$, exponent $n = 2$, $\epsilon = 0.025$.
}
\label{tab:cutflow}
\end{table}
\section{Signal modelling}
\label{sec:signal_modelling}
For both spin-0 and spin-2 resonance searches, parametric models of the diphoton invariant-mass distributions are used in order to test for different resonance masses and widths or \coupling. The distribution for a signal of given mass and width is obtained by convolving the detector resolution with the predicted mass line-shape distribution at the particle level.
The detector resolution is modelled by a double-sided Crystal Ball (DSCB) function, composed of a Gaussian core with power-law tails~\cite{Aaboud:2016tru}.
The line-shape at the particle level for each signal model is taken to be the product of the analytic differential cross-section expression for particle production and decay to photons, and a parametrised form of the parton luminosity.
The signal model is in good agreement with invariant-mass distributions from simulated signal samples introduced in Section~\ref{sec:sample} with corresponding values of resonance mass and width (\coupling) for the spin-0 (spin-2) case.
Potential differences in the tails between the signal model and the simulation are found to have negligible impact on the extracted signal yield.
For relative widths comparable to or below the detector resolution of 1\% ($\coupling \lesssim 0.08$), the searches have limited sensitivity to the decay width.
In the spin-0 resonance search, a fiducial region at particle level that closely follows the selection criteria applied to the reconstructed data is defined for setting limits:
$|\eta| < 2.37$ and $\et > 0.4\mgg$ ($0.3\mgg$) for the leading (subleading) photon. In addition, an isolation requirement of $\ETiso < 0.05\et + 6~\GeV$ is applied to reproduce the selection applied at the reconstruction level.
The particle-level isolation is calculated using all particles with lifetime greater than 10~ps at the event-generator level in
a cone of $\Delta R = 0.4$ around the photon direction.
Compared to the fiducial region definition used in Ref.~\cite{Aaboud:2016tru},
the mass range requirement introduced to reduce the model-dependence from the off-shell region is removed since its impact is found to be negligible using the new \textsc{MadGraph5\_aMC@NLO} signal model.
The combined reconstruction and identification efficiency, defined as the ratio of the number of events fulfilling all the selections placed on reconstructed quantities to the number of events in the fiducial acceptance, varies from 64\% at 200~\GeV\ to 75\% at 2700~\GeV.
It is evaluated with signal MC samples corresponding to the narrow-width approximation (NWA, $\Gamma_X = 4~\MeV$), with a 2.8\% uncertainty assigned to cover variations in the decay width (0--10\%).
In the spin-2 resonance search, the results are evaluated assuming the acceptance as well as the reconstruction and identification efficiencies obtained by the MC simulation of KK graviton decays.
The product of these two terms is evaluated as a function of the KK graviton mass \mG\ using MC samples with $\coupling=0.2$.
It increases from 45\% at 500~\GeV\ to 65\% at 5000~\GeV, with a 2.9\% uncertainty that results from varying \coupling.
In the non-resonant signal search,
the acceptance and reconstruction and identification efficiencies are evaluated for the excess yield in the presence of an ADD signal, which is defined as the difference between the sum of ADD and SM contributions (including their interference) and the SM contribution.
The acceptance for the excess yields in the signal region increases from 58\% at $\ms=3.5~\TeV$ to 65\% at $\ms=5~\TeV$. For larger values of \ms, the acceptance decreases to about 58\% at 8~\TeV\ due to a larger contribution from the interference term, which has smaller acceptance. The combined reconstruction and identification efficiency for the excess yields at different $\ms$ values is approximately constant at 77\% within MC statistical uncertainties.
Uncertainties in the signal parameterization and in the acceptance and detector efficiency correction factors for the signal considered in each search are summarized in Table~\ref{tab:uncertainties_signal}.
\begin{table}
\centering
{\small
\begin{tabular}{lccc}
\hline
\multirow{2}{*}{Uncertainty source} & Spin-0 resonance & Spin-2 resonance & Spin-2 non-resonant \\
& [\%] & [\%] & [\%] \\
\hline
Signal mass resolution & 17--38 & 28--36 & -- \\
\hline
Signal photon identification efficiency & 1.3--3.0 & 2.6--3.1 & 3.2 \\
Signal photon isolation efficiency & 1.1--1.3 & 1.2--1.4 & 1.4 \\
Signal width dependence & 2.8 & 2.9 & -- \\
Trigger efficiency & 0.4 & 0.4 & 0.4 \\
Luminosity & \multicolumn{3}{c}{2.1 (2015), 3.4 (2016)} \\
\hline
Total uncertainty in signal yield & 4.6--5.4 & 5.3--5.5 & 4.8 \\
\hline
\end{tabular}
}
\caption{Summary of the relative systematic uncertainties (in percent). For mass-dependent uncertainties, the quoted values cover the range from 200~\GeV\ (500~\GeV) to 2700~\GeV\ (5000~\GeV)
for the spin-0 (spin-2) resonance search. The uncertainty in the total signal yield corresponds to the sum in quadrature of the individual components, not including the uncertainty in the mass resolution.
}
\label{tab:uncertainties_signal}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
Searches for new phenomena in high-mass diphoton final states
with the ATLAS experiment at the LHC are presented.
The proton--proton collision data corresponding to an integrated luminosity of \intLumi\ were
recorded in 2015 and 2016 at a centre-of-mass energy of $\sqrt{s}$ = 13~\TeV. Analyses optimized for the
search for spin-0 resonances with masses above 200~\GeV, for spin-2 resonances predicted by the Randall--Sundrum model with masses above 500~\GeV,
and for non-resonant Kaluza--Klein graviton signals in the Arkani-Hamed--Dimopoulos--Dvali scenario are performed.
The data are consistent with the Standard Model background expectation. At a mass around 750~\GeV, where the largest deviation from the background hypothesis was previously observed, no excess is seen in the 2016 data. In the combined dataset, the largest local deviation from the background-only hypothesis for the spin-0 (spin-2) resonance search is \localSigSpinZero\,$\sigma$ (\localSigSpinTwo\,$\sigma$) for a mass near \bestMassSpinZero~\GeV\ and narrow width (mass near \bestMassSpinTwo~\GeV\ and $\coupling=\bestCouplingSpinTwo$).
The global significance of this excess is null (\globalSigSpinTwo\,$\sigma$) for the spin-0 (spin-2) resonance search.
In the spin-0 resonance search, the observed 95\% CL upper limits on the fiducial cross section times branching ratio for a narrow-width signal range from \maximumLimitSpinZeroNWA\ at
200~\GeV\ to about \minimumLimitSpinZeroNWA\ at 2700~\GeV.
In the spin-2 resonance search, the observed limits on the cross section times branching ratio for $\coupling=0.1$ range from \maximumLimitSpinTwoSmallCouplings\ to about \minimumLimitSpinTwoSmallCouplings\ for a KK graviton mass between 500~\GeV\ and 5000~\GeV.
The RS1 model with $\coupling=0.1$ is excluded below $m_{G^{*}}=\limitRS$.
These results supersede those previously reported by ATLAS based on 2015 data.
In the ADD scenario, lower limits between \minimumLimitADD\ and \maximumLimitADD\ are set on the ultraviolet cutoff scale \ms,
depending on the number of extra dimensions and the theoretical formalism used.
\section{Statistical procedure}
\label{sec:stat}
In the spin-0 and spin-2 resonance searches, the numbers of signal and background events are estimated from maximum-likelihood fits of the signal-plus-background models to the corresponding
$m_{\gamma\gamma}$ distribution of the selected events.
In the search for an ADD graviton signal,
a counting experiment is performed in the region $\mgg>2240$~\GeV\
with the excess and the SM yields extracted using the methods discussed in Sections~\ref{sec:signal_modelling} and \ref{sec:background}, respectively.
Systematic uncertainties summarized in Tables~\ref{tab:uncertainties_signal} and \ref{tab:uncertainties_background}
are included in the fits via nuisance parameters constrained by Gaussian or log-normal penalty terms.
The $p$-value is determined from a profile-likelihood-ratio-test statistic~\cite{Cowan:2010js} as detailed in Ref.~\cite{Aaboud:2016tru}.
For the resonance searches, the local $p$-value for compatibility with the background-only hypothesis when testing a given
signal hypothesis ($p_0$) is evaluated based on the asymptotic approximation~\cite{Cowan:2010js}, and expressed in standard deviations in the following.
Global significance values are computed from background-only pseudo-experiments to account for the trial factors due to scanning both the signal mass and the width hypotheses.
The expected and observed 95\% confidence level (CL) exclusion limits
on the cross section times branching ratio to two photons are computed using
a modified frequentist approach CL$_{\text{s}}$~\cite{cls} with
the asymptotic approximation to the test-statistic distribution~\cite{Cowan:2010js} for all three searches.
For resonance searches, cross-checks with sampling distributions generated using pseudo-experiments are performed for a few signal mass points
across the searched mass range. Below 2500~\GeV, the difference between the observed and expected limits is at the 0.01~fb level, which is well covered by the $\pm 1\sigma$ limit band.
In the mass range above 2500~\GeV\ covered by the spin-2 resonance search, the asymptotic approximation is no longer valid due to the small number of data events.
The observed and expected limits in this region are hence determined with pseudo-experiments.
\section{ATLAS detector}
\label{sec:detector}
The ATLAS detector~\cite{PERF-2007-01, ATLAS-TDR-19} is a multi-purpose detector with a forward--backward symmetric
cylindrical geometry\footnote{The ATLAS experiment uses a right-handed coordinate system with its origin at the nominal interaction point (IP) in the centre of the detector and the $z$-axis along the beam pipe. The $x$-axis points from the IP to the centre of the LHC ring, and the $y$-axis points upward. Cylindrical coordinates $(r,\phi)$ are used in the transverse plane, $\phi$ being the azimuthal angle around the $z$-axis. The pseudorapidity is defined in terms of the polar angle $\theta$ as $\eta=-\ln\tan(\theta/2)$. The transverse energy
is defined as $\et = E \sin(\theta)$.}. The systems most relevant for the presented searches are the inner detector (ID), immersed in a 2~T
magnetic field produced by a thin superconducting solenoid, and the calorimeters.
The ID consists of fine-granularity silicon pixel and microstrip detectors covering
the pseudorapidity range $|\eta|<2.5$ complemented by a gas-filled straw-tube transition radiation tracker (TRT) at larger radii which covers the region $|\eta| < 2.0$ and provides electron identification capabilities. The electromagnetic (EM) calorimeter is a lead/liquid-argon sampling
calorimeter with accordion geometry. It is divided into a barrel section covering
$|\eta|<1.475$ and two end-cap sections covering $1.375<|\eta|<3.2$. For $|\eta|<2.5$, it is divided
into three layers in depth, which are finely segmented in $\eta$ and $\phi$. A thin presampler
layer, covering $|\eta|<1.8$, is used to correct for fluctuations in upstream energy losses.
Hadronic calorimetry in the region $|\eta|<1.7$ uses steel absorbers and scintillator tiles as the
active medium. Liquid-argon calorimetry with copper absorbers is used in the hadronic end-cap
calorimeters, which cover the region $1.5<|\eta|<3.2$. A forward calorimeter using copper and
tungsten absorbers with liquid argon completes the calorimeter coverage up to $|\eta|=4.9$.
The muon spectrometer, located beyond the calorimeters,
consists of three large air-core superconducting
toroid systems with precision tracking chambers providing
accurate muon tracking for $|\eta| < 2.7$ and fast detectors for triggering for $|\eta| < 2.4$.
Events are selected using a first-level trigger implemented in custom electronics,
which reduces the event rate to a design value of at most 100~kHz using a subset of detector information~\cite{Aaboud:2016leb}.
Software algorithms with access to the full detector information are then used in the high-level trigger to yield an average
recorded event rate of about 1~kHz.
\section{Background estimates}
\label{sec:background}
Two different methods~\cite{Aaboud:2016tru} are used to estimate the SM background contributions to the $m_{\gamma\gamma}$ distribution. The approach adopted in the spin-0 search, appropriate for a mass range with enough data close to the investigated resonance mass,
is based on a fit using a smooth functional form,
with parameter values determined simultaneously with the signal and background yields by the fit. The mass distribution is fitted in the range above 180~\GeV\ (or 150~\GeV\ when fitting 2015 data alone),
and the search range for the signal is 200--2700~\GeV.
The procedure detailed in Ref.~\cite{HIGG-2013-08} is used to check that the functional form is flexible enough to accommodate different physics-motivated underlying background distributions from MC simulations, within the uncertainties in the measured background composition and PDF set. The potential bias due to the choice of the functional form is estimated by the fitted signal yield (``spurious signal'') in these background distributions,
and is considered as a systematic uncertainty.
The spurious signal is required to be less than 30\% of the statistical uncertainty in the fitted signal yield (from the background distributions) over most of the investigated mass range.
Below 400~\GeV, where the statistical uncertainty in the MC sample used to determine the spurious signal uncertainty is comparable to the maximum spurious signal allowed, the criterion is relaxed to 50\%.
In the spin-2 resonance search and also in the non-resonant signal search,
both of which target KK graviton signals at the \TeV\ scale,
the small number of data events at high $m_{\gamma\gamma}$ values does not effectively constrain the invariant-mass distribution of the background.
The shape of the distribution of the irreducible diphoton background is thus predicted using the \textsc{Diphox}~\cite{Binoth:1999qq} computation at NLO in QCD,
which is found to be in good agreement with the predictions from \textsc{Sherpa} version 2.2.2~\cite{Siegert:2016bre} using the same order in QCD.
The background from photon+jet and dijet events is added using control samples from the data with the same method as discussed in Ref.~\cite{Aaboud:2016tru}.
An alternative method, based on the rate of jets misidentified as photons extracted from inclusive photon data and applied to similar control samples,
gives compatible results.
The different background components are combined according to the decomposition studies in data reported in Section~\ref{sec:event_selection} for the $m_{\gamma\gamma}$ distribution of the total background.
The normalization of the background is a free parameter in the maximum-likelihood fit to the data spectrum.
Uncertainty in the total background's shape results from uncertainties in both the shape and the relative normalization of each component~\cite{Aaboud:2016tru},
including the shape of the reducible background,
the relative normalization of the reducible and irreducible backgrounds,
the impact of the parton-level isolation requirement in \textsc{Diphox},
and the effect of the uncertainties in the scales and PDF set used in the \textsc{Diphox} computation.
These uncertainties are taken to be correlated across the entire mass range. Statistical uncertainties in the simulated samples are also considered in each bin.
The uncertainty from the parton-level isolation requirement in \textsc{Diphox} is constrained by the low-mass sideband of the data \mgg\ spectrum in the maximum-likelihood fit, and thus reduces to about 1\% across most of the mass range.
In the non-resonant signal search, the SM yield and the associated uncertainties are computed directly as integrals of the above background estimates in the signal region $\mgg>2240$~\GeV, before the likelihood maximization.
The systematic uncertainties in the description of the background shapes are summarized in Table~\ref{tab:uncertainties_background}.
\begin{table}
\begin{center}
{\small
\begin{tabular}{lcc}
\hline
Uncertainty source & \multicolumn{2}{c}{Spin-0 resonance}\\
\hline
\multirow{2}{*}{Spurious signal [events]} & Narrow width & 74--0.006\\
& 10\% relative width & 195--0.04\\
\hline
\hline
Uncertainty source & Spin-2 resonance [\%] & Spin-2 non-resonant [\%] \\
\hline
Scales and PDFs in \textsc{Diphox} computation & 1--19 & 20 \tspp \\[3mm]
Shape of the reducible background & 1--10 & 11 \tspp \\[3mm]
Relative normalization of & \multirow{2}{*}{1--2\, } & \multirow{2}{*}{2} \\
reducible and irreducible backgrounds & & \\[3mm]
Parton-level isolation requirement & \multirow{2}{*}{10--12\, } & \multirow{2}{*}{9} \\
in \textsc{Diphox} & & \\[3mm]
MC statistical uncertainty & \multicolumn{2}{c}{$<1$} \\
\hline
Total & 10--25 & 25 \\
\hline
\end{tabular}
}
\end{center}
\caption{Summary of pre-fit relative systematic uncertainties in the background estimation used in the searches.
For mass-dependent uncertainties, the quoted mass ranges cover 200~\GeV\ (500~\GeV) until 2700~\GeV\
for the spin-0 (spin-2) resonance search.
In the spin-2 searches the PDF uncertainty in the irreducible background component dominates the total uncertainty beyond 2700~\GeV, rising to 130\% at 5000~\GeV.}
\label{tab:uncertainties_background}
\end{table}
|
\section{Introduction}
Fix a field $\mathbb{K}$ of characteristic zero and let $S=\mathbb{K} [x,y,z]$ be the polynomial ring on three variables which we associate to the symmetric algebra of a three dimensional vector space $V$. The $X_3$ arrangement is the collection of the hyperplanes $H_1=\{ x=0\}$, $H_2=\{ y=0\}$, $H_3=\{z=0\}$, $H_4=\{x+y=0\}$, $H_5=\{x+z=0\}$, and $H_6=\{y+z=0\}$ in $V$. The associated matroid is called the rank 3 whirl (see \cite{Ox} Appendix and pictured in Figure \ref{X3pic}) and is the unique relaxation of the braid matroid whose free multiplicities were studied in \cite{DFMS16}. A multiplicity on $X_3$ is a function $\mathbf{m} :X_3 \to \mathbb{Z}_{>0}$ and let $\mathbf{m}=[m_1,\ldots ,m_6]$ denote the multiplicity vector on ${\cal A}$ where $\mathbf{m}(H_i)=m_i$ for all $1\leq i\leq 6$ respectively. In Figure \ref{X3pic} we show a real projective picture of $X_3$ with the corresponding multiplicities.
\begin{figure}
\begin{tikzpicture}[scale=.6]
\draw[thick] (0,0) circle (4cm);
\draw[thick] (-3.8,0)--(3.8,0);
\draw[thick] (0,-3.8)--(0,3.8);
\draw[thick] (-2.7,2.7)--(2.7,-2.7);
\draw[thick] (-1,3.3)--(-1,-3.3);
\draw[thick] (3.3,-1)--(-3.3,-1);
\node at (.5,3.5) {$m_1$};
\node at (3.5,.5) {$m_2$};
\node at (3.1,3.1) {$m_3$};
\node at (2.3,-1.8) {$m_4$};
\node at (-1.5,-3) {$m_5$};
\node at (.-3,-1.5) {$m_6$};
\draw[thick,fill] (6,-3) circle (5pt);
\draw[thick,fill] (9,-3) circle (5pt);
\draw[thick,fill] (12,-3) circle (5pt);
\draw[thick,fill] (9,3) circle (5pt);
\draw[thick,fill] (7.5,0) circle (5pt);
\draw[thick,fill] (10.5,0) circle (5pt);
\draw[thick] (6,-3)--(12,-3)--(9,3)--(6,-3);
\node[below, left] at (5.8,-3.2) {2};
\node[below] at (9,-3.2) {6};
\node[below,right] at (12.2,-3.2) {3};
\node[above] at (9,3.2) {1};
\node[left] at (7.3,0) {4};
\node[right] at (10.8,0) {5};
\end{tikzpicture}
\caption{A projective picture of the $X_3$ arrangement labeled with multiplicities and its associated matroid diagram}\label{X3pic}
\end{figure}
The problem of classifying free multiplicities on a free arrangement is usually very difficult (in Section \ref{defs} we cover the background material for freeness). At the moment there are only four main results where arrangements (or classes of arrangements) free multiplicities are classified.
\begin{enumerate}
\item In \cite{ATY09} Abe, Yoshinaga, and Terao show that the only arrangements with the property that all multiplicities are free are products of one or two dimensional arrangements.
\item In \cite{Y10} Yoshinaga shows that generic arrangements have no free multiplicities.
\item In \cite{A07} Abe classifies the free multiplicities on the deleted $A_3$ braid arrangement.
\item In \cite{DFMS16} the first author, Francisco, Mermin, and Schweig finish the classification of free multiplicities on the full $A_3$ braid arrangement started in \cite{ATW08} by Abe, Terao, and the second author and progressed in \cite{ANN09} by Abe, Nuida, and Numata.
\end{enumerate}
The idea of this note is to be a positive addition to this list. It turns out that the $X_3$ arrangement has a one dimensional moduli space (the realization space of the rank 3 whirl matroid, see Section \ref{defs} for details). The main result, Theorem~\ref{thm:X3Classification}, is not only a classification of the free multiplicities on $X_3$ but also a classification of the free multiplicities on all of its moduli. Since the characteristic polynomial of $X_3$ does not factor, no arrangement in the moduli of $X_3$ is free (by Terao's factorization theorem~\cite{Te81}). This is the first non-generic arrangement to have it's free multiplicities classified where (1) it's not free and (2) it has non-trivial moduli. One consequence of this classification is many multiplicities where freeness of the associated multi-arrangement is not combinatorially determined, contrary to Terao's conjecture for simple arrangements (see Example~\ref{ex:FreeNotComb}).
Our proof of the classification follows along the lines of two recent papers~\cite{D17,DFMS16} and also draws on work of Brandt and Terao~\cite{BT94} and Schenck and Stillman~\cite{SS96}. We prove that freeness of $D(X_3,\mathbf{m})$ is characterized by the vanishing of a certain homology module of a chain complex. The relevant chain complex is a modification (to the setting of multi-arrangements) of a chain complex appearing in work of Brandt and Terao~\cite{BT94}, where it is used to study formality. On the other hand, our analysis of this chain complex closely mirrors techniques in multivariate spline theory~\cite{SS96}. The techniques are not isolated to the $X_3$ arrangement and can be generalized to arbitrary multi-arrangements; this is the intended subject of a forthcoming paper.
Our paper is arranged as follows. In Section \ref{defs} we review basic definitions and enumerate the moduli of $X_3$. In section \ref{proofs} we introduce the homological machinery necessary for our classification and in section~\ref{sec:freemult} we classify all the free multiplicities on $X_3$. In section \ref{exams} we discuss free extensions of $X_3$; in particular we show that extensions of $X_3$ and its moduli satisfy Terao's conjecture.
{\bf Acknowledgements:} The authors want to thank Graham Denham for pointing out some free multiplicities on $X_3$ which was the main motivation for this project. We also want to thank Takuro Abe and Masahiko Yoshinaga for pointing out the use of the multiarrangement addition-deletion theorem with the deleted $A_3$ arrangement. The second author is partially supported by the Simons Foundation and the Office of Naval Research.
\section{Set up}\label{defs}
In this section we first review the basic definitions of free multiplicities. Then we discuss the moduli on $X_3$.
\subsection{Derivations on multiarrangements}
The module of derivations on $S$ is $\mathrm{Der}(S)=\{ \theta \in \mathrm{Hom}(S,S)\ |\ \forall f,g\in S,\ \theta (fg)=\theta(f)g+f\theta (g)\}$ and the module of derivations on the multiarrangement $({\cal A},\mathbf{m})$ where ${\cal A}=\{ H_i\}$, with defining linear forms $\alpha_i$, is $$D({\cal A},\mathbf{m})=\{\theta\in \mathrm{Der}(S)\ |\ \theta (\alpha_i)\in \alpha_i^{\mathbf{m}(H_i)} S\}. $$
\begin{defn} A multiplicity $\mathbf{m}$ on an arrangement ${\cal A}$ is \emph{free} for ${\cal A}$ if $D({\cal A},\mathbf{m})$ is a free module over the polynomial ring $S$. If $\mathbf{m}$ is a free multiplicity we say $({\cal A},\mathbf{m})$ is free.
\end{defn}
Write $Q({\cal A},\mathbf{m})=\prod_{H\in{\cal A}} \alpha_H^{\mathbf{m}(H)}$. We may refer to a multi-arrangement by its polynomial $Q({\cal A},\mathbf{m})$. The following is Saito's critierion for multi-arrangements.
\begin{prop}\cite{Z89}
Suppose $({\cal A},\mathbf{m})$ is a multi-arrangement and $\theta_1,\ldots,\theta_\ell\in D({\cal A},\mathbf{m})$. Write $\theta_i=\sum_{j} \theta_{ij}\partial_j$ in the basis elements $\partial_j=\partial/\partial x_j$ and let $M=M(\theta_1,\ldots,\theta_\ell)$ be the $\ell\times\ell$ matrix with entries $\theta_{ij}$. The following are equivalent.
\begin{enumerate}
\item $D({\cal A},\mathbf{m})$ is a free $S$-module with basis $\theta_1,\ldots,\theta_\ell$.
\item $\det(M)=kQ({\cal A},\mathbf{m})$ for some $k\neq 0\in\mathbb{K}$.
\end{enumerate}
\end{prop}
If ${\cal A}\subset V\cong\mathbb{K}^\ell$ is an arrangement and $H\in{\cal A}$ a hyperplane, the restriction ${\cal A}^H$ is the hyperplane arrangement in $H\cong \mathbb{K}^{\ell-1}$ with hyperplanes $\{H\cap H': H'\in{\cal A}\}$. The \textit{Zeigler multiplicity} on ${\cal A}^H$ is $\mathbf{m}(H')=\#\{H''\in{\cal A}: H''\cap H=H'\cap H\}$ and the Ziegler multi-restriction of ${\cal A}$ to $H\in{\cal A}$ is the multi-arrangement $({\cal A}^H,\mathbf{m})$ where $\mathbf{m}$ is the Ziegler multiplicity. If ${\cal A}$ is a central arrangement, denote by $L$ its lattice of flats; this is the poset of all intersections of hyperplanes of ${\cal A}$ ordered with respect to reverse inclusion. If $X\in L$ then ${\cal A}_X$ is the arrangement consisting of all hyperplanes containing $X$. The following criterion is due to Yoshinaga~\cite{Y04}.
\begin{thm}\label{thm:Yosh}
Suppose ${\cal A}$ is a central arrangement with maximal flat $c$ and $H\in{\cal A}$. Then ${\cal A}$ is free if and only if
\begin{enumerate}
\item The Ziegler multi-restriction $({\cal A}^H,\mathbf{m})$ is free and
\item ${\cal A}_X$ is free for every $X\neq c\in L$ with $X>H$.
\end{enumerate}
\end{thm}
\subsection{Moduli of $X_3$}
Consider the coefficient matrix $M=(a_{ij})$ where $H_i=\{a_{i1}x+a_{i2}y+a_{i3}z=0\}$ for $1\leq i\leq 6$. Since hyperplanes $H_1$, $H_2$, $H_5$ and $H_6$ form a generic subarrangement of $X_3$ we can fix coordinates so that the coefficient matrix is of the form $$M=\left[ \begin{array}{ccc}
1&0 &0\\
0&1&0\\
a_{31}&a_{32}&a_{33}\\
a_{41}&a_{42}&a_{43}\\
0&0&1\\
1&1&1\\
\end{array}\right]$$ Then we get the following consequences of the three triple points:
\begin{enumerate}
\item The triple point $\{H_1,H_3,H_5\}$ gives that $a_{32}=0$.
\item The triple point $\{H_2,H_3,H_6\}$ gives that $a_{31}=a_{33}$.
\item The triple point $\{H_1,H_2,H_4\}$ gives that $a_{43}=0$.
\end{enumerate} Since $a_{31}\neq 0$ and $a_{41}\neq 0$ (otherwise we would have more dependences than are combinatorially allowed for $X_3$) we can scale the rows so that the coefficient matrix is $$M=\left[ \begin{array}{ccc}
1&0 &0\\
0&1&0\\
1&0&1\\
1&a_{42}&0\\
0&0&1\\
1&1&1\\
\end{array}\right] .$$ Since we do not have any more multiple intersection points this is as far as we can reduce $M$. Hence the dimension of the moduli space of $X_3$ is 1. Moreover, by examining all other possible determinants of $M$ we see that as long as $a_{42}\neq 0, 1$ the corresponding matroid is isomorphic to that of $X_3$. If $a_{42}=0$ then $H_1=H_4$. If $a_{42}=1$ then we get a new dependency which makes the arrangement lattice equivalent to the braid arrangement.
Again changing coordinates and using $\alpha\neq 0,1$ for the one dimensional moduli we will use the following coefficient matrix for the remainder of the paper $$M=\left[ \begin{array}{ccc}
1&0 &0\\
0&1&0\\
0&0&1\\
1&-\alpha&0\\
1&0&1\\
0&1&1\\
\end{array}\right] $$ which can be viewed as tilting the ``diagonal line'' in Figure \ref{pic2}.
\begin{figure}
\begin{tikzpicture}[scale=.6]
\draw[thick] (0,0) circle (4cm);
\draw[thick] (-3.8,0)--(3.8,0);
\draw[thick] (0,-3.8)--(0,3.8);
\draw[thick, red] (-3.5,1.5)--(3.5,-1.5);
\draw[thick] (-1,3.3)--(-1,-3.3);
\draw[thick] (3.3,-1)--(-3.3,-1);
\node at (.5,3.5) {$m_1$};
\node at (3.5,.5) {$m_2$};
\node at (3.1,3.1) {$m_3$};
\node at (2.3,-1.8) {$m_4$};
\node at (-1.5,-3) {$m_5$};
\node at (.-3,-1.5) {$m_6$};
\draw[->,red,thick] (-1.2,2.4) to [out=180, in=70] (-2.7,1.3);
\node[red] at (-2,1.8) {$\alpha$};
\end{tikzpicture}
\caption{The $X_3$ arrangement with the moduli emphasized}\label{pic2}
\end{figure}
\begin{exm}[Freeness of multi-arrangements is not combinatorial]\label{ex:FreeNotComb}
Consider the multiplicity $\mathbf{m}=[3,3,3,1,1,1]$ on the moduli of the $X_3$ arrangement over the complex numbers, with corresponding multi-arrangement defined by $x^3y^3z^3(x-\alpha y)(x+z)(y+z)$. The multi-arrangement defined thus is free if and only if $\alpha\neq -1$, which shows that freeness of multi-arrangements is not a combinatorial property.
We can prove this claim using addition-deletion techniques for multi-arrangements from~\cite{ATW08} (if the reader is not familiar with these techniques, feel free to skip the following argument). First, consider the deletion $({\cal A},\mathbf{m}')$ with defining polynomial $x^3y^3z^3(x+z)(y+z)$. Then ${\cal A}$ is supersolvable with filtration ${\cal A}={\cal A}_3\supset{\cal A}_2\supset{\cal A}_1$, where ${\cal A}_1,{\cal A}_2,{\cal A}_3$ have defining polynomials $x$,$xz(x+z)$, and $xyz(x+z)(y+z)$, respectively. We can check that $({\cal A},\mathbf{m}')$ is free with exponents $(3,4,4)$ by~\cite[Theorem~5.10]{ATW08}. The Euler restriction (see~\cite{ATW08}) of $(X_3,\mathbf{m})$ to $V(x)=H_1$ is the multi-arrangement $x^3z^3(x+z)(x+\alpha z)$ (some care must be taken to see that $3$ is the proper exponent on $x$, but this follows from~\cite[Proposition~4.1]{ATW08}). Ziegler considered the multi-arrangement $x^3z^3(x+z)(x+\alpha z)$ to show that exponents of multi-arrangements are not combinatorial~\cite[Proposition~10]{Z89}; he showed that the exponents are $(3,5)$ if $\alpha=-1$ and $(4,4)$ if $\alpha\neq -1$. It then follows from the addition theorem~\cite[Theorem~0.7]{ATW08} that the multi-arrangement $x^3y^3z^3(x-\alpha y)(x+z)(y+z)$ is free if $\alpha\neq -1$ and from the restriction theorem~\cite[Theorem~0.6]{ATW08} that it is not free when $\alpha=-1$.
\end{exm}
\begin{remark}\label{rem:P1}
A significant obstruction to using addition-deletion techniques from~\cite{ATW08} to classify all free multiplicities on the moduli of $X_3$ is that exponents for multi-arrangements of four points in $\mathbb{P}^1$ are largely unknown. If the multiplicity vector $\mathbf{m}$ of an arrangement ${\cal A}$ of points in $\mathbb{P}^1$ satisfies certain inequalities and the points in ${\cal A}$ are `generic' then the second author and Yuzvinsky show that the exponents of $({\cal A},\mathbf{m})$ are $(\lfloor |\mathbf{m}|/2\rfloor,\lceil |\mathbf{m}|/2 \rceil)$~\cite{WY07}. However it is not easy to determine what makes the arrangement of points in $\mathbb{P}^1$ `generic'; it depends in particular on the multiplicity vector $\mathbf{m}$.
\end{remark}
\section{Homological characterization of freeness}\label{proofs}
In this section we prove that freeness of $(D(X_3),\mathbf{m})$ can be characterized by the vanishing of a certain homology of a chain complex. We devote this section to a careful description of the chain complex and a proof that the vanishing of this particular homology is both necessary and sufficient for freeness of $D(X_3,\mathbf{m})$. The techniques are close in spirit to~\cite{SS96}, where it is shown that vanishing homologies of a chain complex determine freeness of the module of splines.
We will denote by $L=L_{X_3}$ the intersection lattice of $X_3$, and we denote by $L_c$ intersections of codimension (or rank) $c$. The $X_3$ arrangement has three triple points $Y_1,Y_2,$ and $Y_3$ given as intersections of $\{H_1,H_2,H_4\},\{H_1,H_3,H_5\},$ and $\{H_2,H_3,H_6\}$, respectively. Let $L_2^{\text{\tiny{trip}}}=\{Y_1,Y_2,Y_3\}$ be the set of triple points. For every hyperplane $H_i$ in $X_3$ set $J(H_i)=\alpha_i^{\mathbf{m}(H_i)}S$. Furthermore, for a flat $Y\in L$ of rank two let $J(Y)=\sum_{Y\subset H} J(H)$.
\subsection{Hilbert-Burch resolutions}\label{ss:HB} For later use, we will need to have a good understanding of the syzygies of the ideals $J(Y_i)$, which we write explicitly below:
\[
\begin{array}{c}
J(Y_1)=\langle x^{m_1},y^{m_2},(x-\alpha y)^{m_4}\rangle\\
J(Y_2)=\langle x^{m_1},z^{m_3},(x+z)^{m_5}\rangle\\
J(Y_3)=\langle y^{m_2},z^{m_3},(y+z)^{m_6}\rangle.
\end{array}
\]
Each of these ideals is visibly codimension two (being supported at the relevant triple point in $\mathbb{P}^2$) and is Cohen-Macaulay. As such, each ideal $J(Y_i)$ has a \textit{Hilbert-Burch resolution} (see~\cite[Theorem~3.2]{E05}) of the form
\[
0\rightarrow S^2\xrightarrow{\phi_i} S^3 \rightarrow J(Y_i)\rightarrow 0,
\]
where $\phi_i$ is a $2\times 3$ matrix of forms whose $2\times 2$ minors are (up to multiplication by a non-zero constant) the given generators of $J(Y_i)$. The columns of $\phi_i$ are \textit{syzygies} on the generators of $J(Y_i)$. We will write
\[
\phi_1=
\begin{bmatrix}
A_1 & B_1\\
A_2 & B_2\\
A_3 & B_3
\end{bmatrix}
\qquad
\phi_2=
\begin{bmatrix}
C_1 & D_1\\
C_2 & D_2\\
C_3 & D_3
\end{bmatrix}
\qquad
\phi_3=
\begin{bmatrix}
E_1 & F_1\\
E_2 & F_2\\
E_3 & F_3
\end{bmatrix}.
\]
To make the meaning of these matrices clearer, consider the ideal $J(Y_1)=\langle x^{m_1},y^{m_2},(x-\alpha y)^{m_4}\rangle$. Then the columns of $\phi_1$ express the relations $A_ix^{m_1}+B_iy^{m_2}+C_i(x-\alpha y)^{m_4}=0$ for $i=1,2$ (notice the relative order of the generators is important). Since these are matrices for Hilbert-Burch resolutions, we also have (up to multiplication by a non-zero constant)
\[
A_1B_2-B_1A_2=(x-\alpha y)^{m_4}, A_1B_3-B_1A_3=y^{m_2}, A_2B_3-B_2A_3=x^{m_1},\mbox{etc.}
\]
It is standard practice in commutative algebra to omit redundant generators from the generating set. However, it is important for our analysis that we fix the given generators for the ideals $J(Y_i)$. We illustrate with two simple examples.
\begin{exm}\label{ex:HB}
Suppose $m_1=m_2=2$ and $m_4=1$, so $J(Y_1)=\langle x^2,y^2,(x-\alpha y)\rangle$. Clearly either $x^2$ or $y^2$ is an extraneous generator. There are several possible choices for what $\phi_1$ will look like, depending on what syzygies are chosen on $J(Y_1)$. We make the following choice:
\[
\phi_1=
\begin{bmatrix}
1 & 0\\
-\alpha^2 & -(x-\alpha y)\\
-(x+\alpha y) & y^2
\end{bmatrix},
\]
which expresses the fact that syzygies on $J(Y_1)$ can be generated by the Koszul syzygy between $y^2$ and $(x+y)$, followed by the expression of $x^2$ in terms of $y^2$ and $x-\alpha y$. Notice that the $2\times 2$ minors are $-(x-\alpha y),y^2,$ and $-x^2$. More generally, suppose $J(Y_1)$ is minimally generated by two of the three forms $x^{m_1},y^{m_2},(x-\alpha y)^{m_4}$; without loss suppose $y^{m_2}$ and $(x-\alpha y)^{m_4}$ generate $J(Y_1)$. Then $\phi_1$ has the form
\[
\phi_1=
\begin{bmatrix}
1 & 0\\
-F & -(x-\alpha y)^{m_4}\\
-G & y^{m_2}
\end{bmatrix},
\]
where $F$ and $G$ are polynomials so that $Fy^{m_2}+G(x-\alpha y)^{m_4}=x^{m_1}$.
Now suppose instead that $m_1=m_2=m_4=2$, so $J(Y_1)=\langle x^2,y^2,(x-\alpha y)^2\rangle$. This time none of the generators are redundant. Then one choice for $\phi_1$ is:
\[
\phi_1=
\begin{bmatrix}
x-2\alpha y & y\\
\alpha^2x & -2\alpha x+\alpha^2y\\
-x & -y
\end{bmatrix}.
\]
Notice that the $2\times 2$ minors are $-2\alpha(x+y)^2,2\alpha y^2,$ and $-2\alpha x^2$.
\end{exm}
\subsection{The chain complex}
We will now assemble the promised chain complex whose homologies will tell us about freeness of the multi-arrangement $(X_3,\mathbf{m})$. First consider the chain complex $\mathcal{S}=\mathcal{S}_0\xrightarrow{\delta^0}\mathcal{S}_1\xrightarrow{\delta^1}\mathcal{S}_2$ with modules
\[
\mathcal{S}_0\cong S^3 \qquad \mathcal{S}_1\cong S^6 \qquad \mathcal{S}_2\cong S^3.
\]
We will consider taking homologies beginning at the left-hand side, so will denote homologies as \textit{co}homologies. The free module $\mathcal{S}_0$ has a basis in correspondence with the variables $x,y,z$, $\mathcal{S}_1$ has basis in correspondence with hyperplanes (codimension one flats), and $\mathcal{S}_2$ has basis in correspondence with the triple points of $X_3$. The maps $\delta^i:\mathcal{S}_{i}\rightarrow \mathcal{S}_{i+1}$ are given by the matrices
\[
\delta^0=\bordermatrix{ & x & y & z\cr
H_1 & 1 & 0 & 0 \cr
H_2 & 0 & 1 & 0 \cr
H_3 & 0 & 0 & 1 \cr
H_4 & 1 & -\alpha & 0 \cr
H_5 & 1 & 0 & 1 \cr
H_6 & 0 & 1 & 1}
\qquad
\delta^1=\bordermatrix{ & H_1 & H_2 & H_3 & H_4 & H_5 & H_6 \cr
Y_1 & 1 & -\alpha & 0 & -1 & 0 & 0 \cr
Y_2 & 1 & 0 & 1 & 0 & -1 & 0 \cr
Y_3 & 0 & 1 & 1 & 0 & 0 & -1
}.
\]
Note that $\delta^0$ is the coefficient matrix of forms defining $X_3$ and $\delta^1$ is the matrix of relations around triple points. It is readily checked that the chain complex $\mathcal{S}$ is exact. In fact, it can be checked that the homologies of $\mathcal{S}$ govern when $X_3$ is $2$-formal in the sense of Brandt and Terao~\cite{BT94} (see also~\cite{T07}); as a consequence we see that $X_3$ is $2$-formal.
Now define the sub-complex $\mathcal{J}=\mathcal{J}_1\xrightarrow{\delta^1}\mathcal{J}_2$ of $\mathcal{S}$ with modules
\[
\mathcal{J}_1=\bigoplus\limits_{H\in L_1} J(H)\subset \mathcal{S}_1\qquad \mathcal{J}_2=\bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}}J(Y)\subset \mathcal{S}_2.
\]
The quotient complex $\mathcal{S}/\mathcal{J}$ has the form
\[
S^3\xrightarrow{\bar{\delta}^0} \bigoplus\limits_{H\in L_1} \dfrac{S}{J(H)} \xrightarrow{\bar{\delta}^1} \bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}}\dfrac{S}{J(Y)},
\]
where $\bar{\delta}^0,\bar{\delta}^1$ are the quotient maps. Notice that $D(X_3,\mathbf{m})$ is the kernel of the map $\bar{\delta}^0$, in other words $D(X_3,\mathbf{m})\cong H^0(\mathcal{S}/\mathcal{J})$.
\begin{remark}
The modules of the chain complex $\mathcal{S}/\mathcal{J}$ are a natural extension to multi-arrangements of the modules $\bigoplus_{Y\in L_c} D_c({\cal A}_Y)$ which appear in the paper of Brandt and Terao~\cite{BT94}.
\end{remark}
The main point of the chain complex $\mathcal{S}/\mathcal{J}$ is that its homologies control freeness of $D(X_3,\mathbf{m})$. First, we notice that we may consider instead homologies of the complex $\mathcal{J}$.
\begin{prop}\label{prop:HomologySwitch}
With the chain complexes $\mathcal{J},\mathcal{S},\mathcal{S}/\mathcal{J}$ as above, $D(X_3,\mathbf{m})\cong H^0(\mathcal{S}/\mathcal{J})\cong H^1(\mathcal{J})$, $H^1(\mathcal{S}/\mathcal{J})\cong H^2(\mathcal{J})$, and $H^2(\mathcal{S}/\mathcal{J})=0$.
\end{prop}
\begin{proof}
This follows directly from the long exact sequence in homology derived from the short exact sequence of complexes $0\rightarrow \mathcal{J}\rightarrow \mathcal{S}\rightarrow \mathcal{S}/\mathcal{J} \rightarrow 0$, the fact that the homologies of $\mathcal{S}$ vanish, and the fact that $D(X_3,\mathbf{m})\cong H^0(\mathcal{S}/\mathcal{J})$.
\end{proof}
The main result of this section is the following proposition.
\begin{prop}\label{prop:VanishingHomologies}
$D(X_3,\mathbf{m})$ is a free $S$-module if and only if $H^2(\mathcal{J})=0$.
\end{prop}
\begin{proof}
Let $\mathfrak{m}=\langle x,y,z\rangle$ be the maximal ideal of $S$. First we claim that if $H^2(\mathcal{J})$ is nonzero, then it is only supported at $\mathfrak{m}$ (hence has finite length). It is easiest to show this using $H^1(\mathcal{S}/\mathcal{J})$; by Proposition~\ref{prop:HomologySwitch}, $H^1(\mathcal{S}/\mathcal{J})\cong H^2(\mathcal{J})$. We will show that the localization $H^1(\mathcal{S}/\mathcal{J})_P$ at any homogeneous non-maximal prime is zero, using the fact that localization commutes with taking homology. To this end, localize the complex $\mathcal{S}/\mathcal{J}$:
\[
0\rightarrow S^3_P\xrightarrow{(\bar{\delta}^0)_P} \bigoplus\limits_{H\in L_1} \dfrac{S}{J(H)}_P \xrightarrow{(\bar{\delta}^1)_P}\bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}}\dfrac{S}{J(Y)}_P \rightarrow 0.
\]
If $J(H)\not\subset P$ or $J(Y)\not\subset P$, then the corresponding summand in the above chain complex vanishes. Suppose $P$ is codimension one. Then $P$ is principal, hence $P$ can contain $J(H)$ for at most one $H\in L_1$. If $P$ contains none of the ideals $J(H)$, for $H\in L_1$, then clearly $H^1(\mathcal{S}/\mathcal{J})_P=0$. So suppose there is some $H\in L_1$ so that $J(H)\subset P$ (since $P$ is codimension one, this must be the only such $J(H)$). Clearly in this case $(\bar{\delta}_0)_P:S^3_P\rightarrow S/J(H)_P$ is surjective, so again $H^1(\mathcal{S}/\mathcal{J})_P=0$. Now suppose $P$ is codimension two (so $P$ is the ideal of a point in projective two-space). If $P$ contains only one $H\in L_1$, then the same argument as above shows that $H^1(\mathcal{S}/\mathcal{J})_P=0$. If $P$ contains two or more hyperplanes of ${\cal A}$, then $P$ must be the ideal of the point of intersection of whatever hyperplanes $H$ satisfy $J(H)\subset P$. If the point is a simple point of ${\cal A}$, then there are only two hyperplanes $H,H'\in{\cal A}$ so that $J(H)\subset P$ and $J(H')\subset P$. So $(\mathcal{S}/\mathcal{J})_P$ looks like
\[
0\rightarrow S^3_P\xrightarrow{(\bar{\delta}^0)_P} \dfrac{S}{J(H)}_P\oplus\dfrac{S}{J(H')}_P,
\]
which is clearly exact on the right, so $H^1(\mathcal{S}/\mathcal{J})_P=0$. Finally, suppose $P$ is the ideal of a triple point $Y\in L_2^{\text{\tiny{trip}}}$ (intersection of $H,H',$ and $H''\in{\cal A}$). Then $(\mathcal{S}/\mathcal{J})_P$ looks like
\[
0\rightarrow S^3_P\xrightarrow{(\bar{\delta}^0)_P} \dfrac{S}{J(H)}_P\oplus\dfrac{S}{J(H')}_P\oplus\dfrac{S}{J(H'')}_P\xrightarrow{(\bar{\delta}^1)_P} \dfrac{S}{J(Y)}_P \rightarrow 0.
\]
A check yields that $\mbox{coker}((\bar{\delta}_0)_P)=(S/J(Y))_P$, so again $H^1(\mathcal{S}/\mathcal{J})_P=0$. It follows that $H^1(\mathcal{S}/\mathcal{J})$ is only supported at $\mathfrak{m}$, the homogeneous maximal ideal.
Now to show that freeness of $D(X_3,\mathbf{m})$ is equivalent to vanishing of $H^2(\mathcal{J})$, consider the following four-term exact sequence:
\begin{equation}\label{D-exact}
0\rightarrow D(X_3,\mathbf{m})\cong H^1(\mathcal{J})\rightarrow \bigoplus\limits_{H\in L_1} J(H)\xrightarrow{\delta^1} \bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}}J(Y) \rightarrow H^2(\mathcal{J}) \rightarrow 0.
\end{equation}
Write $\mathcal{J}_1$ for $\bigoplus\limits_{H\in L_1} J(H)$ and $\mathcal{J}_2$ for $J(Y_1)\oplus J(Y_2) \oplus J(Y_3)$. Then break this four term sequence into the two short exact sequences
\[
0\rightarrow K \rightarrow \mathcal{J}^2 \rightarrow H^2(\mathcal{J}) \rightarrow 0\qquad\mbox{and}\qquad 0\rightarrow D(X_3,\mathbf{m}) \rightarrow \mathcal{J}^1 \rightarrow K \rightarrow 0,
\]
where $K=\mbox{im}(\delta^1)$. By the long exact sequence in $\mbox{Ext}$ applied to the first short exact sequence, coupled with the fact that $\mbox{Ext}^j_S(J(Y_i),S)$ vanishes when $j>1$ (hence $\mbox{Ext}^j_S(\mathcal{J}_2,S)$ vanishes when $j>1$), we obtain $\mbox{Ext}^3_S(H^2(\mathcal{J}),S)\cong\mbox{Ext}^2_S(K,S)$. Similarly, applying the long exact sequence in $\mbox{Ext}$ to the second short exact sequence (and using that $\mbox{Ext}^j(\mathcal{J}_1,S)=0$ for $j>0$ since $\mathcal{J}_1$ is a sum of principal ideals) yields that $\mbox{Ext}^2_S(K,S)\cong \mbox{Ext}^1(D(X_3,\mathbf{m}),S)$. All in all, we see that $\mbox{Ext}^3_S(H^2(\mathcal{J}),S)\cong \mbox{Ext}^1_S(D(X_3,\mathbf{m}),S)$. Since $D(X_3,\mathbf{m})$ is a second syzygy, freeness of $D(X_3,\mathbf{m})$ is equivalent to vanishing of $\mbox{Ext}^1_S(D(X_3,\mathbf{m}),S)$, which is equivalent to vanishing of $\mbox{Ext}^3_S(H^2(\mathcal{J}),S)$. Since $H^2(\mathcal{J})\cong H^1(\mathcal{S}/\mathcal{J})$ is only supported at $\mathfrak{m}$, $\mbox{Ext}^3_S(H^2(\mathcal{J}),S)$ vanishes if and only if $H^2(\mathcal{J})=0$, and we are done.
\end{proof}
\section{Free multiplicities}\label{sec:freemult}
In this section we derive the full characterization of free multiplicities on $(X_3,\mathbf{m})$. The characterization hinges on the use of Proposition~\ref{prop:VanishingHomologies}, together with a precise description of the homology module $H^2(\mathcal{J})$. We begin by giving this description in terms of syzygies.
Denote by $e_H$ a generator for $J(H)$ (of degree $\mathbf{m}(H)$) and given a triple point $Y$ and a hyperplane $H$ passing through $Y$ we will denote by $e_{H,Y}$ (of degree $\mathbf{m}(H)$) a generator for $J(H)$ viewed as a sub-ideal of $J(Y)$. We will use the following commutative diagram to get a presentation for $H^2(\mathcal{J})$ using syzygies of the ideal $J(Y)$.
\begin{center}
\begin{tikzcd}
0 & 0 \\
\bigoplus\limits_{H\in L_1} J(H) \ar{r}{\delta^1}\ar{u} &\bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}} J(Y) \ar{u} \\
\bigoplus\limits_{H\in L_1} S[e_{H}] \ar{r}{\hat{\delta}^1}\ar{u} & \bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}}\left[\bigoplus\limits_{\substack{H\in L_1\\ Y\subset H}} S[e_{H,Y}] \right] \ar{u}\\
0\ar{u}\ar{r} &\bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}} \syz(J(Y)) \ar{u}\\
& 0\ar{u} \\
\end{tikzcd}
\end{center}
The snake lemma yields an exact sequence
\begin{equation}\label{eq:SNAKE}
\bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}} \syz(J(Y))\rightarrow \mbox{coker}(\hat{\delta}^1)\rightarrow \mbox{coker}(\delta^1)=H^2(\mathcal{J})\rightarrow 0.
\end{equation}
We now explicitly identify the image of the syzygy modules inside of $\mbox{coker}(\hat{\delta}^1)$, which we will show is isomorphic to $S^3$. As noted in \S~\ref{ss:HB}, each of the ideals $J(Y_i)$ is Cohen-Macaulay of codimension two, with Hilbert-Burch resolution of the form
\[
0\rightarrow S^2\xrightarrow{\phi_i} S^3 \rightarrow J(Y_i)\rightarrow 0,
\]
where, as in \S~\ref{ss:HB}, we will write
\[
\phi_1=
\begin{bmatrix}
A_1 & B_1\\
A_2 & B_2\\
A_3 & B_3
\end{bmatrix}
\qquad
\phi_2=
\begin{bmatrix}
C_1 & D_1\\
C_2 & D_2\\
C_3 & D_3
\end{bmatrix}
\qquad
\phi_3=
\begin{bmatrix}
E_1 & F_1\\
E_2 & F_2\\
E_3 & F_3
\end{bmatrix}.
\]
We see that the image of $\bigoplus_{Y\in L_2} \syz(J(Y))$ inside of $\bigoplus\limits_{Y\in L_2^{\text{\tiny{trip}}}}\left[\bigoplus\limits_{\substack{H\in L_1\\ Y\subset H}} S[e_{H,Y}] \right]$ is generated by the columns of the following matrix (we label the rows by pairs $(H_i,Y_i)$ corresponding to $e_{H_i,Y_i}$:
\[
\bordermatrix{ & & & & & \cr
[H_1,Y_1] & A_1 & B_1 & 0 & 0 & 0 & 0 \cr
[H_2,Y_1] & A_2 & B_2 & 0 & 0 & 0 & 0 \cr
[H_4,Y_1] & A_3 & B_3 & 0 & 0 & 0 & 0 \cr
[H_1,Y_2] & 0 & 0 &C_1&D_1& 0 & 0 \cr
[H_3,Y_2] & 0 & 0 &C_2&D_2& 0 & 0 \cr
[H_5,Y_2] & 0 & 0 &C_3&D_3& 0 & 0 \cr
[H_2,Y_2] & 0 & 0 & 0 & 0 &E_1&F_1\cr
[H_3,Y_3] & 0 & 0 & 0 & 0 &E_2&F_2\cr
[H_6,Y_6] & 0 & 0 & 0 & 0 &E_3&F_3\cr
}_\text{\large{.}}
\]
The map $\hat{\delta}^1$ is lifted from $\delta^1$ and has the following matrix:
\[
\bordermatrix{& H_1 & H_2 &H_3&H_4&H_5&H_6\cr
[H_1,Y_1] & 1 & 0 & 0 & 0 & 0 & 0 \cr
[H_2,Y_1] & 0 & -\alpha & 0 & 0 & 0 & 0 \cr
[H_4,Y_1] & 0 & 0 & 0 &-1 & 0 & 0 \cr
[H_1,Y_2] & 1 & 0 &0 & 0 & 0 & 0 \cr
[H_3,Y_2] & 0 & 0 & 1 & 0 & 0 & 0 \cr
[H_5,Y_2] & 0 & 0 & 0 & 0 &-1 & 0 \cr
[H_2,Y_3] & 0 & 1 & 0 & 0 & 0 & 0 \cr
[H_3,Y_3] & 0 & 0 & 1 & 0 & 0 & 0 \cr
[H_6,Y_3] & 0 & 0 & 0 & 0 & 0 & -1\cr
}_\text{\large{.}}
\]
This matrix clearly has full rank, so $\mbox{coker}(\hat{\delta}^1)\cong S^3$, generated for instance by the images of $e_{H_1,Y_1},e_{H_2,Y_1},$ and $e_{H_3,Y_2}$; call these $[e_{H_1,Y_1}],[e_{H_2,Y_1}],$ and $[e_{H_3,Y_2}]$. With respect to this basis we may represent $\mbox{coker}(\hat{\delta}^1)$ as the image of $S^9$ under the projection with matrix
\[
\begin{bmatrix}
1 & 0 & 0 & -1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 & \alpha & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & -1 & 0
\end{bmatrix}_\text{\large{.}}
\]
With this choice of projection, the image of the syzygies inside of $S^3$ is generated by the columns of the following matrix (rows are labeled by $[H_i,Y_1]$ corresponding to $[e_{H_i,Y_1}]$ for $i=1,2,4$):
\[
M=\bordermatrix{ & & & & & \cr
[H_1,Y_1] &A_1&B_1&-C_1&-D_1& 0 & 0 \cr
[H_2,Y_1] &A_2&B_2& 0 & 0 &\alpha E_1&\alpha F_1 \cr
[H_3,Y_2] & 0 & 0 &C_2 &D_2 &-E_2 &-F_2
}_\text{\large{.}}
\]
Together with Proposition~\ref{prop:VanishingHomologies} and the presentation~\eqref{eq:SNAKE}, we have proved the following result.
\begin{prop}\label{prop:HomologyPresentation}
With all notation as above, $D(X_3,\mathbf{m})$ is free if and only if the columns of $M$ generate the free module $S^3=\mbox{coker}(\hat{\delta}^1)$.
\end{prop}
Now we classify free multiplicities on $X_3$ (and all of its moduli).
\begin{thm}\label{thm:X3Classification}
For any $\alpha\neq 0,1$, freeness of $(X_3,\mathbf{m})$ implies $m_4=m_5=m_6=1$ and $m_1=m_2=m_3=n$ for some positive integer $n$. So free multiplicities are of the form $[n,n,n,1,1,1]$. Furthermore,
\begin{itemize}
\item if $\alpha$ is not a root of unity then $[n,n,n,1,1,1]$ is a free multiplicity for any $n>1$, and
\item if $\alpha$ is a root of unity with order $m$, then $[n,n,n,1,1,1]$ is a free multiplicity if and only if $n\not\equiv 1\mbox{ mod } m$. (In particular, if $\alpha=-1$, then $[n,n,n,1,1,1]$ is a free multiplicity if and only if $n$ is even.)
\end{itemize}
Moreover, when $(X_3,\mathbf{m})$ is free the exponents are $(n+1,n+1,n+1)$.
\end{thm}
\begin{remark}
In Theorem~\ref{thm:X3Classification}, if we fix a multiplicity $\mathbf{m}=[n,n,n,1,1,1]$ then $(X_3,\mathbf{m})$ is free as long as $\alpha$ avoids the roots of $t^{n-1}-1$. Moreover $(X_3,\mathbf{m})$ is not free for any $\alpha$ if $\mathbf{m}$ does not have the form $[n,n,n,1,1,1]$. So among the moduli of arrangements with the $X_3$ lattice and a fixed multiplicity $\mathbf{m}$, the multi-arrangements which are free form a (possibly empty) Zariski open set. We do not know if this is true in general. Yuzvinsky~\cite{Y93} has shown that this is the case for simple arrangements.
\end{remark}
\begin{proof}
We use Proposition~\ref{prop:HomologyPresentation}. The columns of $M$ generate $S^3$ if and only if there is a $3\times 3$ minor $M'$ of $M$ so that $\det(M')$ is a non-zero constant. Let us see what constraints this assumption places on the entries of $M'$.
Suppose first that $M'$ contains two columns corresponding to the same ideal; without loss of generality assume that
\[
M'=
\left(
\begin{array}{ccc}
A_1 & B_1 & F \\
A_2 & B_2 & G \\
0 & 0 & H
\end{array}
\right),
\]
where the third column is some other column of $M$. Then $\det(M')=H(A_1B_2-B_1A_2)$. However, by the structure of the Hilbert-Burch matrix (see~\S~\ref{ss:HB}) $A_1B_2-B_1A_2$ is a (non-zero) constant multiple of a generator of the ideal $J(Y_1)$, hence this implies $\det(M')$ is either zero or has positive degree, so cannot be non-vanishing constant. It follows that if $\det(M')$ is a non-zero constant, it must have a column corresponding to a single syzygy from each of the three ideals $J(Y_1),J(Y_2),$ and $J(Y_3)$. Without loss of generality, we may assume that $M'$ has the form
\[
M'=
\left(
\begin{array}{ccc}
A_1 &-C_1 & 0\\
A_2 & 0 & \alpha E_1 \\
0 &C_2 & -E_2
\end{array}
\right),
\]
with $\det(M')=-\alpha A_1E_1C_2-A_2C_1E_2$. Suppose that one of the entries off of the anti-diagonal of $M'$ vanishes; without loss suppose $A_1=0$. Then it follows that $A_2y^{m_2}+A_3(x+y)^{m_4}$ is a Koszul syzygy on the forms $y^{m_2}$ and $(x+y)^{m_4}$ (see Example~\ref{ex:HB}). However, this implies that the second term $A_2C_1E_2$ either has positive degree or vanishes. Both cases contradict that $\det(M')$ is a non-zero constant. It follows that all entries other than the anti-diagonal entries of $M'$ are non-zero. Therefore, since both terms of $\det(M')$ are homogeneous, the only way that $\det(M')$ is a non-zero constant is if all entries of $M'$ other than the anti-diagonal entries are non-zero constants.
Now recall the analysis in Example~\ref{ex:HB}. The presence of \textit{any} non-zero constant in the matrix $\phi_i$ expresses the fact that $J(Y_i)$ is minimally generated by two of the three given generators. For simplicity let us consider $J(Y_1)$. The fact that $A_1$ and $A_2$ are \textit{both} non-zero constants means that $J(Y_1)$ may be generated by $(x-\alpha y)^{m_4}$ along with \textit{either} $x^{m_1}$ or $y^{m_2}$. It is not difficult to check that $J(Y_1)$ can be generated by $y^{m_2}$ and $(x-\alpha y)^{m_4}$ if and only if $m_2+m_4\le m_1+1$. Likewise, $J(Y_1)$ can be generated by $x^{m_1}$ and $(x-\alpha y)^{m_4}$ if and only if $m_1+m_4\le m_2+1$. To satisfy both of these inequalities, we must have $m_4=1$ and $m_1=m_2$. A similar analysis with $J(Y_2)$, $J(Y_3)$ yields $m_5=1$ and $m_1=m_3$, $m_6=1$ and $m_2=m_3$, respectively. Hence the ideals of the triple points have the form
\[
J(Y_1)=\langle x^n, y^n, x-\alpha y\rangle \qquad J(Y_2)=\langle x^n,z^n,x+z\rangle \qquad J(Y_3)=\langle y^n,z^n,y+z\rangle,
\]
where $m_1=m_2=m_3=n$. This proves the first statement of the theorem. Now, if we assume $m_1=m_2=m_3=n$, we have
\[
\phi_1=
\left(
\begin{array}{cc}
1 & 0 \\
-\alpha^{n} & -(x-\alpha y) \\
-(x^{n}-\alpha^{n}y^{n})/(x-\alpha y) & y^{n}
\end{array}
\right),
\]
\[
\phi_2=
\left(
\begin{array}{cc}
1 & 0 \\
(-1)^{n+1} & -(x+z) \\
-(x^{n}-(-1)^{n}z^{n})/(x+ z) & z^{n}
\end{array}
\right),
\]
\[
\mbox{and}\qquad\phi_3=
\left(
\begin{array}{cc}
1 & 0 \\
(-1)^{n+1} & -(y+z) \\
-(y^{n}-(-1)^{n}z^{n})/(y+z) & z^{n}
\end{array}
\right).
\]
See Example~\ref{ex:HB} for details on the computation of $\phi_1,\phi_2,$ and $\phi_3$. This yields $A_1=C_1=E_1=1,A_2=-\alpha^n$, and $C_2=E_2=(-1)^{n+1}$, so $\det M'=-\alpha A_1E_1C_2-A_2C_1E_2=-\alpha (-1)^{n+1}(\alpha^{n-1}-1)$. This is zero (hence $[n,n,n,1,1,1]$ is not free) if and only if $\alpha^{n-1}=1$; equivalently $n\equiv 1 \mod m$, where $m$ is the order of $\alpha$.
Finally, the exponents for the free multiplicities can be calculated by two different methods. First, one can look at the short exact sequence in \eqref{D-exact} and compute the Hilbert functions to get the exponents. Second, one can use the multiarrangement addition-deletion theorem \cite{ATW08} with the deleted hyperplane being $x-\alpha y$. In this case the deletion is the deleted $A_3$ arrangement whose free multiplicities are classified by Takuro Abe in \cite{A07}.
\end{proof}
\begin{remark}\label{rem:Characteristic}
Theorem~\ref{thm:X3Classification} appears to hold over finite fields (of characteristic greater than two, since we assume $\alpha\neq 0,1$).
\end{remark}
\begin{remark}
It is possible to recover portions of Theorem~\ref{thm:X3Classification} using other techniques for multi-arrangements. We list a few here.
\begin{enumerate}
\item If $\alpha=-1$ and we assume $m_4=m_5=m_6=1$, then local and global mixed products~\cite{ATW07} can be used to show that $(X_3,\mathbf{m})$ is free if and only if $m_1=m_2=m_3=2k$ for some natural number $k$.
\item If $(X_3,\mathbf{m})$ satisfies that all sub-$A_2$ multi-arrangements are `balanced,' then local and global mixed products show that $\mathbf{m}$ is not a free multiplicity if one of $m_4,m_5,$ or $m_6$ is greater than one.
\item If $m_i=1$ for some $i=1,\ldots,6$, then the deletion of $(X_3,\mathbf{m})$ with respect to $H_i$ has the deleted $A_3$ arrangement as the base arrangement. The classification of all free multiplicities on the deleted $A_3$ arrangement by Abe~\cite{A07}, along with addition/deletion/restriction theorems for multi-arrangements~\cite{ATW08} yields some constraints on the $m_i$ (however, this method runs into the problem of Remark~\ref{rem:P1}).
\item If $(X_3,\mathbf{m})$ has a `heavy flag,' then $\mathbf{m}$ is not a free multiplicity~\cite[Corollary~5.2]{AK16}.
\end{enumerate}
We do not know if it is possible to recover Theorem~\ref{thm:X3Classification} entirely using other (non-homological) techniques for multi-arrangements.
\end{remark}
Next we examine a basis for the case where $\alpha=-1$ and $D(X_3,\mathbf{m})$ is free.
\begin{prop}\label{lem:Basis}
The derivations \[
\begin{array}{rl}
\theta_1 & =x^{2k+1}\partial_1+y^{2k+1}\partial_2+z^{2k+1}\partial_3\\
\theta_2 & =(y+z)(x^{2k}\partial_1-y^{2k}\partial_2-z^{2k}\partial_3)\\
\theta_3 & =x^{2k}z\partial_1-(x+y+z)y^{2k}\partial_2+xz^{2k}\partial_3
\end{array}
\] form a basis for the $S$-module $D(X_3,\mathbf{m})$ for $\alpha=-1$ where $\mathbf{m}=[2k,2k,2k,1,1,1]$.
\end{prop}
\begin{proof} This is a routine check with Saito's criterion \cite{S80}. \end{proof}
\section{Free extensions}\label{exams}
Given an arrangement ${\cal A}'$ of rank $r$, an \textit{extension} of ${\cal A}'$ is an arrangement ${\cal A}$ of rank $r+1$ so that there exists a hyperplane $H_0\in{\cal A}$ with ${\cal A}^{H_0}={\cal A}'$. In our case, where ${\cal A}'$ is defined by linear forms in $S=\mathbb{K}[x,y,z]$, we will assume ${\cal A}$ is defined by forms in $\mathbb{K}[x,y,z,w]$ and $H_0$ is defined by $w=0$. A \textit{free extension} of a multi-arrangement $({\cal A}',\mathbf{m})$ is a free arrangement ${\cal A}$ with a hyperplane $H_0\in{\cal A}$ so that the Ziegler multi-restriction satisfies ${\cal A}^{H_0}=({\cal A}',\mathbf{m})$. The free multiplicities of the $X_3$ arrangement do admit free extensions; we write these down in Proposition \ref{prop:Extensions}, but first we need a preliminary lemma.
\begin{lem}\label{lem:LocalAlongW}
Let $n>0$ be an integer and let ${\cal A}$ be the central arrangement given by the vanishing of forms
\[
\begin{array}{c}
x-a_1z,\ldots,x-a_nz,\\
y-b_1z,\ldots,y-b_nz,\\
Ax+By+Cz,z,
\end{array}
\]
where $a_1,\ldots,a_n,b_1,\ldots, b_n,A,B,C$ are all constants with $A,B\neq 0$. Then ${\cal A}$ is free if and only if there is a re-ordering of the $a_i$ so that $(a_1,b_1,1),\ldots,(a_n,b_n,1)$ all lie on the line $Ax+By+Cz=0$ (in $\mathbb{P}^2(\mathbb{K})$).
\end{lem}
\begin{proof}
We use Yoshinaga's criterion for freeness of $3$-arrangements~\cite{Y05}. Namely, a three-arrangement is free if and only if
\begin{enumerate}
\item $\chi({\cal A},t)=(t-1)(t-d_1)(t-d_2)$ and
\item The Ziegler restriction $({\cal A}^H,\mathbf{m})$ is free with exponents $(d_2,d_2)$.
\end{enumerate}
Restricting to the hyperplane $H=V(z)$, we get the multi-arrangement ${\cal A}^H=V(x)\cup V(y)\cup V(Ax+By)$ with multiplicities $\mathbf{m}=[n,n,1]$. By a characterization due to Wakamiko~\cite{W07}, $({\cal A}^H,\mathbf{m})$ is free with exponents $(n,n+1)$. Suppose that the line $Ax+By+Cz=0$ passes through $q$ of the $n^2$ intersection points of the grid defined by $x_1=a_1z,\ldots,x_n=a_nz,y_1=b_1z,\ldots,y_n=b_nz$. Then we compute
\[
\chi({\cal A},t)=(t-1)(t^2-(2n+1)t+n^2+2n-q),
\]
which factors as $(t-1)(t-n)(t-(n+1))$ if and only if $q=n$. Since $A,B\neq 0$, the re-ordering in the statement of the lemma is possible.
\end{proof}
\begin{prop}\label{prop:Extensions}
Let $n>1$ be an positive integer, $\mathbf{m}_n=[n,n,n,1,1,1]$, and $\alpha\neq 0,1\in\mathbb{K}$. If $(X_3,\mathbf{m}_n)$ has a free extension then $\alpha$ is a root of unity and $n=mt$ where $m$ is the order of $\alpha$ and $t$ is a positive integer. Suppose $A_1,\ldots,A_t\in\mathbb{K}^*$ are nonzero constants. Then the arrangement defined by the vanishing of the forms
\[
\begin{array}{ccc}
x-A_1w,x-\alpha A_1w,\ldots,x-\alpha^{m-1}A_1w,\\
\vdots\\
x-A_tw,x-\alpha A_tw,\ldots,x-\alpha^{m-1}A_tw,\\
y-A_1w,y-\alpha A_1w,\ldots,y-\alpha^{m-1}A_1w,\\
\vdots\\
y-A_tw,y-\alpha A_tw,\ldots,y-\alpha^{m-1}A_tw,\\
z+A_1w,z+\alpha A_1w,\ldots,z+\alpha^{m-1}A_1w\\
\vdots\\
z+A_tw,z+\alpha A_tw,\ldots,z+\alpha^{m-1}A_tw\\
x-\alpha y,x+z,y+z,w
\end{array}
\]
is a free extension of $(X_3,\mathbf{m}_n)$. Moreover, up to translations, every free extension has this form.
\end{prop}
\begin{proof}
Freeness of the arrangement ${\cal A}$ given by the indicated forms can be verified by applying Yoshinaga's criterion~\cite{Y04}; since $(X_3,\mathbf{m}_n)$ is free by Theorem~\ref{thm:X3Classification}, we check that ${\cal A}$ is locally free along $w=0$. This reduces to checking freeness of the rank three closed sub-arrangements of ${\cal A}$ along $w=0$; all of these are free by Lemma~\ref{lem:LocalAlongW}.
Now we show that any free extension must have this form. By Theorem~\ref{thm:X3Classification} and Theorem~\ref{thm:Yosh}, a free extension ${\cal A}$ of $(X_3,\mathbf{m})$ exists only if $\mathbf{m}=[n,n,n,1,1,1]$ for some integer $n>1$. In this case ${\cal A}$ has the form
\[
\begin{array}{c}
x-a_1w,\ldots,x-a_nw\\
y-b_1w,\ldots,y-b_nw\\
z-c_1w,\ldots,z-c_nw\\
x-\alpha y+Aw,x+z+ Bw,y+z+Cw,w,
\end{array}
\]
for some constants $a_1,\ldots,a_n,b_1,\ldots,b_n,c_1,\ldots,c_n,A,B,$ and $C$. Translating in the $x,y,$ and $z$ directions we may assume that $A=B=C=0$. We now check local freeness along $w=0$. Let ${\cal A}_{xyw},{\cal A}_{xzw},{\cal A}_{yzw}$ be the closed rank three sub-arrangements consisting of hyperplanes which contain the line $x=y=w=0$, $x=z=w=0$, and $y=z=w=0$, respectively. Projectively, these are all arrangements of the form considered in Lemma~\ref{lem:LocalAlongW}. Assuming that ${\cal A}_{xzw}$ and ${\cal A}_{yzw}$ are both free, Lemma~\ref{lem:LocalAlongW} allows a re-indexing of the $a_i,b_i$ so that the points $(a_i,0,c_i,1)$ for $i=1,\ldots,n$ lie on the line determined by $x+z$ and the points $(0,b_i,c_i,1)$ for $i=1,\ldots,n$ lie along the line determined by $y+z$. It follows that $a_i=b_i=-c_i$ for $i=1,\ldots,n$.
Now, by Lemma~\ref{lem:LocalAlongW}, ${\cal A}_{xyw}$ is free if and only if there are $n$ points among the $n^2$ points $(a_i,b_j,0,1)=(a_i,a_j,0,1)$ which lie along the line $x-\alpha y$. Set $X=\{a_1,\ldots,a_n\}$ and consider the action of $\alpha$ on $X$ by multiplication. We see that this action must be a permutation action. Since $|X|=n>1$, $\alpha$ must be a root of unity; we denote the order of $\alpha$ by $m$. Suppose $0\in X$. Then the multiplication action of $\alpha$ on $X-\{0\}$ is faithful, so $|X-\{0\}|=tm$ (for some integer $t$) and $n=|X|=tm+1$. But then $\mathbf{m}_n=[n,n,n,1,1,1]$ is not a free multiplicity by Theorem~\ref{thm:X3Classification} since $n\equiv 1 \mod m$. It follows that $0\notin X$, so $\alpha$ acts faithfully on $X$ by multiplication, and $n=|X|=mt$ for some integer $t$. Choosing representatives $A_1,\ldots,A_t$ in each $\alpha$-orbit gives the form in the statement of the proposition.
\end{proof}
\begin{remark}
Every free arrangement ${\cal A}$ in Proposition~\ref{prop:Extensions} has the property that ${\cal A}^{w=0}=X_3$ is not free. For instance, taking $\alpha=-1$ and $t=1$ provides a one-parameter family of free arrangements of rank four with ten hyperplanes whose restriction is not free. The first such example was a free arrangement of rank five with 21 hyperplanes provided by Edelman and Reiner~\cite{ER93}.
\end{remark}
\begin{remark}
Yoshinaga's extendability criterion~\cite{Y10} does not apply; the $X_3$ arrangement is locally $A_2$ but only has a positive system when $\alpha=-1$. Furthermore, even if $\alpha=-1$, the $(*)$ criterion of~\cite[Theorem~2.5]{Y10} does not hold for the multiplicities $[n,n,n,1,1,1]$ when $n$ is even.
\end{remark}
\begin{cor}
If ${\cal A}$ is an extension of $xyz(x-\alpha y)(x+z)(y+z)$ (for $\alpha\neq 0,1$), then it satisfies Terao's conjecture (see~\cite[Conjecture~4.138]{OT92}). That is, freeness of ${\cal A}$ is determined from its intersection lattice.
\end{cor}
\begin{proof}
This is implicit in the proof of Proposition~\ref{prop:Extensions}. Explicitly, all of the following steps are combinatorial:
\begin{itemize}
\item determining that the Ziegler multi-restriction has the multiplicity $[n,n,n,1,1,1]$ for $n>1$ (by Theorem~\ref{thm:X3Classification})
\item determining local freeness of ${\cal A}$ along $w=0$ (by Lemma~\ref{lem:LocalAlongW}), which implies $\alpha$ is a root of unity
\item determining that $n=mt$ for some $t$, where $m$ is the order of $\alpha$ (by Theorem~\ref{thm:X3Classification})
\item Finally, determining that ${\cal A}$ is free by Yoshinaga's criterion~\cite{Y04} and Theorem~\ref{thm:X3Classification}. \qedhere
\end{itemize}
\end{proof}
\
\bibliographystyle{amsplain}
|
\section{Introduction}
Ultracold neutral atoms are among the leading experimental platforms for the development of quantum technologies. Specifically, they offer vast opportunities for the engineering of synthetic many-body quantum systems, resulting in both the possibility to realize exact implementations of fundamental theoretical models -- in a quantum simulation perspective -- and to achieve new "extreme" states of matter \cite{qsim}. Experimental tools like optical lattices provide a prime method to control the atomic motion, laying a direct connection with solid-state physics \cite{lewenstein}. Remarkably, the interactions between the particles, and even new types of "synthetic" couplings with artificial fields, can be controlled by exploiting atomic-physics and quantum-optics techniques involving the manipulation of the internal atomic states.
Two-electron atoms, like alkaline-earth atoms and closed-shell rare-earth or lanthanide elements (such as ytterbium), feature a considerably large, accessible internal Hilbert space, with new advanced possibilities of quantum control, adding new tools to the quantum simulation \cite{gorkshov} and quantum information \cite{daley} toolboxes. Specifically, these atoms are characterized by long-lived electronic states (with lifetimes exceeding tens of seconds), that can be accessed by exciting the atoms on ultranarrow optical transitions. In particular, the optical clock transition connecting the ground state $g=$ $^1$S$_0$ to the long-lived state $e=$ $^3$P$_0$ has been used in the last two decades to engineer the best atomic clocks up-to-date, now outperforming conventional atomic clocks by several orders of magnitude \cite{reviewclocks1,reviewclocks2}.
Besides the metrological application, coherent manipulation of the electronic state opens new routes for the production of artificial magnetic fields \cite{gerbier} and, indeed, recently allowed for the demonstration of both spin-orbit coupling and tunable synthetic flux ladders \cite{spinorbit1,spinorbit2}, successfully extending the concept of synthetic dimensions \cite{celi,synthdim1,synthdim2} towards a new implementation. Recent experiments with fermionic strontium and ytterbium showed the possibility to take advantage from the interplay between the electronic degree of freedom and the nuclear spin, resulting in novel effects such as two-orbital spin-exchange \cite{spinex1,spinex2} and interaction tuning with a new kind of orbital Feshbach resonances \cite{ofr1,ofr2,ofr3}, possibly leading the way to novel states of matter \cite{gorkshov,iemini}.
Most of these applications crucially rely on the scattering properties of atoms in different electronic states. While state-dependent scattering lengths for strontium \cite{traverso,zhang} and fermionic $^{173}$Yb \cite{spinex1,spinex2} have been thoroughly characterized, a similar study for the bosonic isotopes of ytterbium is still lacking. Knowing the state-dependent scattering properties of two-electron atoms is also highly relevant for the development of optical atomic clocks, especially for clocks based on bosonic isotopes, where low-temperature collisions cannot be prevented by the preparation of the atomic sample in the same initial internal state.
Here we are reporting on high-resolution clock spectroscopy of $^{174}$Yb bosons trapped in the lowest band of a 3D optical lattice in a Mott insulator state. These studies allow us to resolve the site occupancy in the Mott state and use this information to determine scattering lengths and loss rate coefficients for collisions involving atoms in the $e$ state, that were previously unknown.
\section{Experimental setup}
To address the clock transition we start preparing a Bose-Einstein condensate (BEC) of $ \sim 2\times10^5$ $^{174}$Yb atoms via evaporative cooling in a crossed optical dipole trap with trapping frequencies $\omega_{x,y,z}=2\pi\times(92.8,72.6,86.3)$ Hz. At the end of the evaporation stage the degenerate gas is loaded into a 3D cubic optical lattice operated at the magic wavelength $\lambda_L=759.35$ nm \cite{Barber}, in order to have the same polarizability for the $g$ and the $e$ states. The lattice depth, measured in units of the lattice recoil energy $E_r=\hbar^2 k_L^2/2m$ (where $\hbar$ is the reduced Planck constant, $m$ is the atomic mass and $k_L=2\pi/\lambda_L$), can be tuned up to $s=40$. Once the lattice is loaded, the crossed optical trap is adiabatically turned off and the atoms remain trapped only by the lattice beams, in a Mott insulator state. After this process we end up with a sample consisting of $\sim 1.2\times10^5$ atoms, a number that can be further controlled by changing the waiting time in the lattice after the optical trap has been turned off.
\begin{figure}[tbp]
\centering
\includegraphics[width=.7\linewidth]{rabi_very_long.pdf}
\caption{Rabi oscillation after selective excitation of the singly-occupied sites of an atomic $^{174}$Yb Mott insulator in a deep 3D optical lattice. The number of $g$ atoms remaining after the clock excitation pulse is reported as a function of the pulse duration. The incomplete depletion that we observe at the $\pi$ pulse (after approx. 20 ms) can be explained by the presence of multiply-occupied lattice sites that are not probed in this measurement.}
\label{fig:Rabi}
\end{figure}
Spectroscopy is then performed illuminating the atoms with a 578 nm, $\pi$-polarized, laser pulse, resonant with the $\ket{g} \rightarrow \ket{e}$ clock transition. Being this transition strictly forbidden for bosons, the coupling is artificially induced \cite{taichenachev} by a magnetic field which ranges from 55 to 175 G, depending of the specific experiment. The excitation in the 3D lattice is performed in the Lamb-Dicke regime, with lattice depths spanning from $s=15$ to $s=40$. The linewidth of our clock laser \cite{cappellinilaser} is estimated to be a few tens of Hz on the timescale of the atomic excitation, as it allows to resolve spectroscopic features with a frequency width of a few tens of Hz. Figure \ref{fig:Rabi} shows a Rabi oscillation of the atomic population, showing a coherence time in the atom-laser interaction clearly exceeding $100$ ms. As the overall spectrum is recorded on a timescale of about 20 minutes, from the same data we can obtain information on the long-term stability of the laser during this measurement, with an estimated drift $<2$ Hz/min.
\begin{figure}[htbp]
\centering
\includegraphics[width=1\linewidth]{spectroscopy_new.pdf}
\caption{Clock spectroscopy of an $^{174}$Yb atomic sample in a 3D optical lattice. All the graphs report the fraction of $g$ atoms remaining after the clock excitation pulse. \textbf{a}) In a typical spectrum we observe several evenly-spaced resonances, that we attribute to an interaction energy shift depending on the number of particles contained in each lattice site. The origin of the frequency axis has been chosen to match the position of the highest-energy resonance. \textbf{b}) Clock spectroscopy performed after the application of a photoassociation (PA) pulse, which removes the atoms in multiply-occupied sites, allowing the unambiguous identification of the resonance in singly-occupied sites, corresponding to the unshifted $g-e$ transition frequency. \textbf{c)} Clock spectroscopy performed for different numbers of atoms contained in the sample. As the atom number is decreased, the relative weight of the lower-energy resonances decreases, in agreement with the identification of the highest-energy resonance with the excitation of singly-occupied sites.}
\label{fig:spectroscopy}
\end{figure}
\section{Experimental spectra and determination of state-dependent scattering lengths}
In order to acquire the spectra, the atoms are probed with long interrogation pulses extending from 500 ms to 1.5 s, depending on the specific experiment. The number of atoms remaining in the ground state is then recorded as a function of the laser frequency, causing the spectroscopic signal to appear as a decrease in the atom number. Figure \ref{fig:spectroscopy}a reports a typical spectrum acquired for a sample of $\sim 5\times10^4$ atoms trapped in a 3D optical lattice with $s=30$. The spectrum is characterized by several resonances that we ascribe to processes where a single $g$ atom in a lattice site occupied by $n$ particles is excited to the $e$ state, a transition that we represent with the $\ket{(n)g}\rightarrow\ket{(n-1)g\,e}$ notation. We identify the highest-energy resonance in the spectra as the excitation of the $\ket{g}\rightarrow\ket{e}$ process in singly-occupied sites. This attribution is justified by the results of two different experiments in which we probe the transition by eliminating multiply-occupied sites in the lattice and by changing the total number of atoms in the sample, respectively.
In a first experiment (figure \ref{fig:spectroscopy}b) a photoassociation (PA) pulse is shone before the clock spectroscopy, leaving in this way a sample containing only singly-occupied sites. PA is performed using a 5-ms-long pulse having a frequency red-detuned by 20 MHz with respect to the MOT frequency \cite{photoassociation} and an intensity of $3\times 10^{-3}$ W mm$^{-2}$. No other resonances with the exception of the highest-energy one are observable in the resulting spectrum, which demonstrates the validity of our hypothesis.
In a second experiment (figure \ref{fig:spectroscopy}c) we perform spectroscopy on samples containing different atom numbers. As the atom number is lowered (and the mean number of particles in each lattice site is reduced) the relative weight of the lower-energy resonances decreases, which indicates that these processes can be attributed to $\ket{(n)g}\rightarrow\ket{(n-1)g\,e}$ transitions with $n>1$. On the other hand, the weight of the highest-energy resonance increases as the atom number is reduced, which signals an increased relative number of singly-occupied sites, further confirming the validity of our hypothesis. In additional Rabi oscillation measurements, we have validated our attribution of the particle numbers associated to the $n>1$ resonances (sketched in figure \ref{fig:spectroscopy}), by verifying that the Rabi frequency associated to the different resonances scales with the expected behavior $\Omega_n = \sqrt{n} \Omega_1$ \cite{dicke}.
These results indicate that it is possible to use the optical clock transition of Yb to perform an accurate spectroscopy of an atomic Mott insulator state, with a similar approach to those reported in Refs. \cite{Mottketterle,Mottjapan} by using different radiofrequency or optical transitions.
\subsection{Measurement of e-g scattering length}
\label{section_age}
The identification of the various resonances observed in figure \ref{fig:spectroscopy}a represents the starting point for the determination of the $s-$wave scattering length $a_{eg}$ in $^{174}$Yb, whose value has still not been reported, to our knowledge. We determine this quantity by measuring the interaction shift $\Delta U_{eg}=U_{eg}-U_{gg}$, that can be obtained from a fit of the experimental spectra as the energy difference between the centers of the $n=2$ and of the $n=1$ resonances, respectively labeled as (2) and (1) in figure \ref{fig:spectroscopy}a.
In the framework of the Bose-Hubbard model, $\Delta U_{eg}$ is related to the scattering length $a_{eg}$ by the following relation:
\begin{equation}
\label{delta_U2}
\Delta U_{eg}=\frac{4\pi\hbar^2 } {m} \,( a_{eg}-a_{gg} )\, \int w^4(\mathbf{r})\,d\mathbf{r} \, ,
\end{equation}
where $a_{gg}$ is the scattering length for two ground-state atoms and $w(\mathbf{r})$ are the Wannier functions. The values of $\Delta U_{eg}$ measured for several depths of the 3D lattice are reported in figure \ref{fig:interaction_energy} (blue points). The experimental data have been fitted with equation \ref{delta_U2}, leaving the differential scattering length $a_{eg}-a_{gg}$ as the only free parameter. The result of the fit (solid line) corresponds to a best-fitting value
\begin{equation}
a_{eg}-a_{gg}= -10.19\,(0.13)\,a_0 \,.
\end{equation}
Combining this determination with the known value of $a_{gg}=+104.9\,(1.5)\,a_0$ reported in Ref. \cite{kitagawa2008} results in a value for the $g-e$ scattering length
\begin{equation}
a_{eg}= +94.7\,(1.6)\,a_0 \,.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{interaction_energy_3.pdf}
\caption{Interaction shifts of the clock transition frequency for $n=2$ (blue) and $n=3$ (red) interacting $^{174}$Yb atoms, relative to the single-particle excitation frequency. The solid blue curve is a fit of the $n=2$ experimental data with equation \ref{delta_U2}. The solid red curve is the calculated shift for $n=3$ assuming the results of the previous fit to the $n=2$ data and taking into account only two-body interactions. The dashed curve is a simultaneous fit of the two experimental datasets taking into account, in an effective way, the energy correction for three-body elastic interactions reported in Ref. \cite{Johnson2009} (see text for more details).}
\label{fig:interaction_energy}
\end{figure}
In figure \ref{fig:interaction_energy} we also plot the interaction shift of the $n=3$ resonance (labeled as (3) in figure \ref{fig:spectroscopy}a) relative to the $n=1$ resonance (red points). If only two-body interactions are taken into account, the total interaction energy of three interacting bosons should be $3U_{gg}$ if the particles are all in the $g$ state and $U_{gg}+2U_{eg}$ if one atom is excited to the $e$ state. This means that we expect the interaction energy shift to be $2(U_{eg}-U_{gg})=2\Delta U_{eg}$, i.e. twice that measured for the $n=2$ case. However, the expected shift (solid red line in figure \ref{fig:interaction_energy}), evaluated on the basis of the previous determination of $a_{eg}$, clearly deviates from the experimental points. We ascribe this deviation to an additional energy correction arising from elastic three-body effective interaction. This correction was introduced in Ref. \cite{Johnson2009} for the case of $n$ indistinguishable interacting bosons in the lowest vibrational state of a 3D optical lattice, and was observed experimentally in Refs. \cite{Johnson2009,ThreebodyBloch}. Following the derivation of \cite{Johnson2009}, for three particles in the same quantum state, at the first perturbative order, this correction reads
\begin{equation}
\label{delta_U3}
\delta U_{3}(a,s)= \frac{\beta \, U_{2}(a,s)^2}{\hbar \, \omega(s) / (2\pi)},
\end{equation}
where $U_{2}(a,s)$ is the two-particle interaction energy, which depends on the scattering length $a$ and the lattice depth $s$, $\omega(s)$ is the harmonic frequency characterizing the confinement within one lattice site and $\beta=-1.34$ is a constant. This expression allows us to evaluate the correction to the interaction energy for the $\ket{ggg}$ state. For the $\ket{gge}$ state, however, the three interacting particles are not identical, and the theoretical approach of Ref. \cite{Johnson2009} could not be adequate. If we assume that the three-body elastic interaction in the $\ket{gge}$ state could still be described by equation \ref{delta_U3}, with an average scattering length given by the geometric mean $(a_{gg} a_{eg} a_{eg})^{1/3}$, also the measured $\ket{ggg}\rightarrow\ket{gge}$ interactions shifts can be used to estimate $a_{eg}$. In particular, a combined fit of the data relative to the two- and three-particles interaction shifts as a function of the lattice depths gives now a very good agreement with both datasets (dashed lines in figure \ref{fig:interaction_energy}). From the fit we extract the only fit parameter $\Delta a_{eg-gg} = a_{eg}-a_{gg}= -10.08\,(0.05)\,a_0$, which is consistent with the previous determination that does not rely on the knowledge of the elastic three-body contribution.
\subsection{Measurement of e-e scattering length}
In our spectra, an increased coupling on the clock transition reveals the presence of an additional resonance that lies at a higher energy than the single-particle one, as shown in figure \ref{ee_twophotonspectroscopy} (blue points), at a frequency difference $\Delta f_{ee} \simeq 160$ Hz for $s \simeq 30$. We identify this resonance with the two-photon process $\ket{gg}\rightarrow\ket{ee}$, which transfers two particles trapped in the same lattice site from the $\ket{gg}$ to the $\ket{ee}$ state, via an intermediate $\ket{ge}$ state.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{ee_spectroscopy.pdf}
\caption{Single-frequency (blue) and two-frequency (red) clock spectroscopy of an atomic $^{174}$Yb sample trapped in a 3D optical lattice with $s=30$. In the single-color spectrum, we identify the weak resonance at positive detuning as a signature of the two-particle/two-photon $\ket{gg}\rightarrow\ket{ee}$ process. This identification is confirmed by the presence of a resonance having the same frequency in the two-color spectrum taken at frequencies $f-\delta f$ and $f+\delta f$ (see text for more details). We note that the two-color spectrum has been recorded at a higher laser intensity in order to make the two-photon peak visible, causing the structure of the single-photon sidebands to be unresolved because of power broadening.}
\label{ee_twophotonspectroscopy}
\end{figure}
The two-photon nature of this excitation is validated by a two-frequency spectroscopy experiment in which the transition is excited by a clock laser with two frequency components $f_{+,-}=f \pm \delta f$ simultaneously probing the atomic sample. The result of this experiment, performed fixing $\delta f = 800$ Hz and scanning $f$, is reported in figure \ref{ee_twophotonspectroscopy} (red points). The broad excitation profiles at the edges of the spectrum can be identified with two replicas of the single-photon absorption spectra, frequency-shifted by $+\delta f$ and $-\delta f$ respectively, as expected for a single-photon excitation driven by each of the two frequency components (the interactions sidebands are unresolved in this spectrum because of a larger laser intensity resulting in an increased power broadening). In addition, we still observe a weaker resonance at the same detuning $\Delta f_{ee} \simeq 160$ Hz as in the single-color spectrum, which is a strong indication of its two-photon nature. The absence of frequency shift for this resonance can be explained if we assume that a pair of atoms in the same lattice site absorbs simultaneously two photons with frequencies $f + \delta f$ and $f - \delta f$ each, in such a way that the total energy transferred to the system is $2 h f$ (where $h$ is the Planck constant), as in the case of a two-photon transition occurring for a single-color excitation.
The identification of the $\ket{gg} \rightarrow \ket{ee}$ resonance allows the experimental determination of the interaction energy shift $\Delta U_{ee} = U_{ee}-U_{gg}$ from the energy-conservation relation $2 h \Delta f_{ee} = \Delta U_{ee}$. Adopting an argumentation similar to the one carried on in section \ref{section_age}, it is possible to link this interaction shift to the scattering length $a_{ee}$ via the relation
\begin{equation}
\label{delta_Uee}
\Delta U_{ee}=\frac{4\pi\hbar^2 } {m} \,( a_{ee}-a_{gg} )\, \int w^4(\mathbf{r})\,d\mathbf{r} \, .
\end{equation}
Using this equation and averaging over several spectra (both single-color and two-color) acquired at a mean lattice depth $s=29.3 (0.3)$, we determine the differential scattering length
\begin{equation}
a_{ee} - a_{gg} = +21.8 \, (1.8) \, a_0 \, .
\end{equation}
Combining this measurement with the known value of $a_{gg}=+104.9\,(1.5)\,a_0$ reported in Ref. \cite{kitagawa2008} results in a value for the $ee$ scattering length
\begin{equation}
a_{ee} = +126.7 \, (2.3) \, a_0 \, .
\end{equation}
\begin{table}
\caption{\label{tabone}Summary of the measured s-wave scattering lengths of $^{174}$Yb for different interaction channels involving the $g=$ $^1$S$_0$ and the $e=$ $^3$P$_0$ states (values in units of the Bohr radius $a_0$).}
\begin{indented}
\lineup
\item[]\begin{tabular}{@{}*{3}{l}}
\br
Scattering channel & Measured value & Reference \cr
\mr
$a_{gg}$& $+104.9\,(1.5)$ & \cite{kitagawa2008} \cr
$a_{eg}$& $+94.7\,(1.6)$ & This work + \cite{kitagawa2008}\cr
$a_{ee}$& $+126.7 \, (2.3)$ & This work + \cite{kitagawa2008} \cr
\mr
$a_{eg}-a_{gg}$& $-10.19\,(0.13)$
& This work \cr
$a_{ee} - a_{gg}$& $+21.8 \, (1.8) $
& This work \cr
\br
\end{tabular}
\end{indented}
\end{table}
\section{Detection of state-dependent inelastic collisions}
In this section we present the measurement of the loss rate coefficients relative to the $e-g$ and $e-e$ interaction channels.
In these experiments, we detect the atoms in the metastable $^3$P$_0$ state ($e$) by repumping them to the $^3$D$_1$ state with a laser at 1389 nm. Atoms in the $^3$D$_1$ state can decay to any of the $^3$P$_J$ ($J=0,1,2$) states: while the atoms decayed to the $^3$P$_0$ state undergo another repumping cycle, atoms in the $^3$P$_1$ state decay to the ground state $^1$S$_0$ (in less than 1 $\mu$s), where they can be then detected. A small branching ratio limits the number of atoms lost into the metastable $^3$P$_2$ dark state, allowing for a high repumping efficiency of $e$-state atoms without any additional repumping laser. In our setup, we let the atoms interact with the 1389 nm light for a few ms during the time-of-flight, which grants a repumping efficiency $\gtrsim 90\%$. This is estimated by exciting an atomic sample in purely singly-occupied lattice sites to the $e$ state (with a $\pi-$pulse), then shining the repumping pulse and comparing the measured number of atoms to that of a non-excited $g$ state sample. Before the repumping pulse, the atoms in the ground state are blasted with a 100-$\mu$s-long pulse of imaging light during time of flight, so that only the $e$ state atoms are detected.
\subsection{Inelastic $e-g$ collisions}
In a first experiment, we investigate the $e-g$ losses by loading the atomic sample in a 3D optical lattice with an average depth of s $\simeq$ 29.7. The waiting times in the trap (before the lattice loading) and in the lattice are adjusted in order to maximize the number of doubly-occupied sites and at the same time minimizing the number of multiply-occupied sites. The atoms in the $\ket{gg}$ state are then excited to the $\ket{eg}$ state by means of a 10-ms-long $\pi$-pulse of the clock laser selectively resonant with the $\ket{gg}\rightarrow\ket{eg}$ transition, and after a variable holding time in the lattice the number of atoms in the $e$ and in the $g$ states is measured. The collected data are reported in the two panels of figure \ref{fig:eg_decay} as blue points and red points for the $e$ and $g$ states respectively, while the green points represent the number of atoms in the $g$ state without performing the excitation pulse to the $e$ state. The solid lines are exponential fits to the data as guides to the eye.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{decay_eg.pdf}
\caption{Inelastic $e-g$ collisions have been investigated by exciting atoms in the $\ket{gg}$ state to the $\ket{eg}$ state by selectively addressing the $\ket{gg}\rightarrow\ket{eg}$ transition in a 3D optical lattice ($s=27.5$). \textbf{a}) The number of atoms remaining in $g$ after the clock pulse (red) is compared with the number of atoms in the absence of clock excitation (green) as a function of the holding time in the lattice (the slow decay of the green points shows the finite single-particle lifetime in the lattice). Lines are exponential fits of the experimental data that must be intended as guides to the eye. \textbf{b}) The number of atoms detected in the $e$ state (blue) is shown as a function of the holding time in the lattice after the excitation of the transition. This number is compared with the difference between the number of $g$ atoms without and with the clock excitation (gray), which displays an approximately time-independent behavior. The solid blue line is an exponential fit of the experimental data.}
\label{fig:eg_decay}
\end{figure}
The data clearly show that atoms in the $e$ state decay on a timescale of the order of the second. To better understand this behavior, in figure \ref{fig:eg_decay}b we compare the number of atoms in the $e$ state (in blue) to the difference between the number of atoms in the $g$ state without and with the clock laser pulse (in gray). At short times, this difference is equal to the number of atoms in the $e$ state, as can be reasonably expected. On longer timescales of the order of $1$ s, instead, while the atoms in the $e$ state are lost, the difference between the $g$ atoms without and with the excitation is approximately constant, implying that the losses in the $e$ state are not caused by inelastic collisions with $g$ atoms, otherwise this difference should have increased blue(the small decrease could be attributed to the finite single-particle lifetime of the atoms in the lattice). The decay rate $\gamma = 1/\tau$ of the atoms in the $e$ state can be estimated with an exponential fit to the data, obtaining $0.66 (0.06)$ s$^{-1}$. This timescale is comparable with the tunneling rate $\sim 1$ Hz at the lattice depth of the experiment, suggesting that the $e$ atoms, initially in lattice sites with $g$ atoms, could be lost after tunnelling to neighboring sites via $e-e$ inelastic collisions. Since the timescale of the observed losses is determined by the tunnelling time before the actual interaction events, it is difficult to extract a reliable $e-e$ loss rate coefficient from those data.
Nevertheless, the observed dynamics allows us to give an upper bound to the $e-g$ inelastic loss rate coefficient. Two-body $e-g$ losses would be described by the rate equation $\dot{n}_g = -\beta_{eg} n_e n_g= -\gamma_{eg} n_g$, where $\beta_{eg}$ is the density-dependent loss rate coefficient. Requesting $\gamma_{eg} = \beta_{eg} n_e \ll \gamma$ and determining the in-site density $n_e$ from the calculated Wannier functions in the 3D lattice, we obtain
\begin{equation}
\beta_{eg} \ll 10^{-14} \; \mathrm{cm}^3/\mathrm{s} \; .
\end{equation}
\subsection{Inelastic $e-e$ collisions}
A different strategy had to be implemented for the determination of the $e-e$ losses. As a matter of fact, we could not coherently excite a detectable number of atoms in the $ee$ state by means of $\pi$-pulses on the $\ket{gg}\rightarrow\ket{ee}$ transition, possibly due to an insufficient broadening of the two-photon transition. We then switched to a different geometry and loaded the atoms in a 1D vertical optical lattice at a depth $s=27.5$, obtaining an array of 2D pancakes with a radial trapping frequency $\omega = 2\pi \times 34.5$ Hz. A fraction of $g$ atoms is excited to the $e$ state with a 10-ms-long pulse of clock laser light directed along the pancakes plane, followed by a variable hold time in the lattice. Finally, the number of atoms in the $e$ and $g$ states is measured. As shown in figure \ref{fig:ee_decay}, the data exhibit fast losses in the $e$ state (blue points) on the ms timescale, while the $g$ population (red points) is constant (the red solid line represent the mean of the experimental data). On this timescale, as seen in the previous experiment, other loss channels are negligible, so, assuming only $e-e$ 2-body inelastic scattering (that is the dominant loss mechanism observed with different Yb and Sr isotopes, see e.g. Refs. \cite{traverso,spinex1}, and attributed to principal-number-changing collisions), the losses can be modeled with the rate equation $\dot{n}_e = -\beta_{ee} n_e^2$, where $n_e$ is the density of $e$ atoms and whose solution is given by:
\begin{equation}
n_e(t) = \left( \frac{1}{n_{e0}} - \beta_{ee} t\right)^{-1} \; ,
\label{eq:2bodymodel}
\end{equation}
where $n_{e0}$ is the initial $e$ atom density and $\beta_{ee}$ is the 2-body loss rate coefficient. In order to extract a value for $\beta_{ee}$ from our data we developed a simplified theoretical model to determine the density in the pancakes from the measured atom number.
\begin{figure}
\centering
\includegraphics[width=.7\linewidth]{decay_ee.pdf}
\caption{
Number of $e$ atoms (in blue) and $g$ atoms (in red) as a function of the holding time in the lattice after a clock excitation pulse in an 1D lattice. The red line represents the mean value of $g$ atoms, while the blue line is the result of a fit to the $e$ data with the model in equation (\ref{eq:2bodymodel}), assuming the scaling in equation (\ref{eq:scalingnN}).}
\label{fig:ee_decay}
\end{figure}
Starting from the initial atom number $N_{g0} \simeq 180\times 10^3$ and density distribution of the sample in the optical dipole trap, we first determine the number of atoms in every pancake, assuming no population redistribution during the loading of the lattice. We then consider the density profile in the pancakes as a sum of a condensed part ($\sim 20\%$) and a thermal component at temperature $T_P \simeq 45$ nK, as determined from the experimental time-of-flight images integrated along the lattice axis. We also take into account the atom number reduction due to the 3 s wait time after loading the pancakes. We calculate the mean values of the density distribution in each pancake, and a final global average density $n_{g0}$ of the sample (before the clock excitation pulse) is determined by a weighted average over the pancake distribution. We then assume a linear relation between the atom number and the density (justified by the short timescale of the excitation), so that the $e$-state density $n_e(t)$ after the clock laser excitation can be determined as
\begin{equation}
n_{e}(t) = \frac{n_{g0}}{N_{g0}}N_{e}(t)
\label{eq:scalingnN}
\end{equation}
where $N_e(t)$ is the measured number of atoms in the $e$ state. Using this relation, we convert $N_e(t)$ into a density $n_e(t)$ and fit it with equation (\ref{eq:2bodymodel}), from which the parameter $\beta_{ee}$ can be determined as
\begin{equation}
\beta_{ee} = 1.6 (0.8)\times 10^{-11} \; \mathrm{cm}^3/\mathrm{s},
\end{equation}
to which we attribute a conservative error due to the several assumptions in our theoretical model. The blue line in figure \ref{fig:ee_decay} is the result of the fit, converted back to atom number following the scaling of equation \ref{eq:scalingnN}.
\section{Conclusions}
In conclusion, we have performed high-resolution spectroscopy of a Mott insulator of ultracold $^{174}$Yb bosons in a 3D optical lattice by exciting them on the ultranarrow $^1$S$_0$~$\rightarrow$~$^3$P$_0$ clock transition. The metrological character of the transition and the narrow spectroscopic signals that we have demonstrated allow for the characterization of the Mott insulator state and for the determination of the lattice sites occupancies.
Our spectroscopic resolution allowed us to precisely determine the scattering lengths for $e-g$ and $e-e$ collisions in ultracold $^{174}$Yb atoms, that were previously unknown. These results are important in quantum information and quantum simulation applications, as well as for the development of optical lattice clocks based on bosonic isotopes of two-electron atoms, where the simpler internal structure (due to the absence of a nuclear spin) could provide advantages over the more commonly used fermionic isotopes.
We have also detected the effect of inelastic collisions involving the atoms in the $e$ state. While the observed lifetimes appear to be severely limited by inelastic $e-e$ losses, no inelastic collisions in the $e-g$ channel could be observed on the timescale and with the sensitivity of our experiment. This system offers rich possibilities for quantum simulation, for instance for the investigation of two-component Bose-Hubbard models with different mobility of the species, e.g. for the study of impurity physics (a state-dependent lattice can be used to freeze the motion of the $e$ atoms in such a way to inhibit inelastic losses), or for the realization of dissipative lattice models when $e-e$ losses are taken into account.
On a more general perspective, the spectroscopic approach that we have employed could be extended in future works to the use of the same ultranarrow clock transition to probe excitation spectra of more complex quantum many-body states of either bosonic or fermionic atoms with metrological accuracy.
{\it Final note:} During the completion of the work, we became aware of very similar measurements performed by the BEC group at LKB \cite{gerbierdraft}, confirming our experimental findings.
\ack
We acknowledge insightful discussions with Fabrice Gerbier and Carlo Sias. Special acknowledgments to the LENS
Quantum Gases group. This work has been supported by ERC Consolidator Grant TOPSIM, INFN project FISH, and MIUR PRIN 2015C5SEJJ. We thank the Galileo Galilei Institute for Theoretical Physics for the hospitality and INFN for partial support during the completion of this work.
\clearpage
\section*{References}
|
\section{Introduction}
\label{introintro}
Computational modeling is widely adopted in various branches of engineering~\cite{velten2009mathematical}.
The aim of the computational modeling is to replace costly field experiments
with evaluations of a less costly computational code.
While wide adoption of the mathematical modeling significantly reduces time and costs required to perform a design process in the industrial engineering,
in many cases it still requires several days and HPC resources to perform an experiment~\cite{forrester}.
Nowadays engineers often construct surrogate models to replace expensive computational models with cheap but sufficiently accurate approximations~\cite{forrester,gorissen2010surrogate,erofeev2016ann}.
An engineer generates a~sample of points, evaluate values of a high fidelity computational code (a high fidelity function) at these points, and then use the generated sample and machinery of regression analysis to construct a surrogate model.
Gaussian process regression is an attractive framework for construction of nonlinear regression models~\cite{forrester,rasmussen2006} with a number of guaranteed theoretical properties~\cite{burnaev2013bernstein,van2008rates}.
Constructed model can be used to speed-up evaluations~\cite{sterling2014approximation,alestra2014df,grihon2013structureOPT}, for surrogate-based optimization~\cite{forrester,martin2005use}, uncertainty quantification~\cite{kawai2013kriging}, sensitivity analysis \cite{panin1,panin2,panin3} and adaptive design of experiments~\cite{burnaev2015adaptive}.
Maturity of this approach is further confirmed not only by numerous applications but also by availability of software packages that are dedicated to surrogate modeling and include Gaussian process regression-based approaches~\cite{belyaev2016gtapprox,gorissen2010surrogate,ageev2016minimization,bergh2013ansys}.
Nice property of Gaussian process regression is an ability to treat variable fidelity data~\cite{forrester2007multi,kennedy2000predicting,qian2006building,han2013improving,doyen1988porosity,burnaev2015surrogate}: one can construct a surrogate model of a~high fidelity function using data both from high and low fidelity sources (e.g., high fidelity evaluations can be obtained using a computational code with a~fine mesh, and low fidelity evaluations can be obtained using the same computational code with a~coarser mesh).
Recent results provide theoretical analysis of obtained models~\cite{zaytsev16arxive,zhang2015doesn} and of parameters estimates~\cite{burnaev2013bernstein}.
Gaussian process based variable fidelity modeling shares many common ideas with
multi-output Gaussian process regression~\cite{alvarez2011computationally,boyle05dependent,higdon2008computer,chang2014fast}.
For data with no specific structure we need $O(\sS^2)$ memory to store a surrogate model and $O(\sS^3)$ operations to construct it.
Due to this computational complexity usually not more than a few thousands of points are used when training Gaussian Process regression.
In the variable fidelity scenario samples are often large, as one evaluation of a low fidelity function is usually significantly cheaper than one evaluation of a high fidelity function.
Currently there are several ways to reduce memory and computational requirements for Gaussian process regression.
Nystr\"om approximation~\cite{drineas2005nystrom} is a popular approach to perform large sample Gaussian process regression inference~\cite{quinonero2005unifying,foster2009stable,sun2015review}.
The idea is to select a subsample of a full sample for which we can still perform Gaussian process regression inference, and then approximate the full sample covariance matrix and its inverse by some combination of the covariance matrix for the selected subsample and the cross-covariance matrix between points from the selected subsample and from the full sample.
Bayesian approximate inference provides an alternative fast estimate of the full sample likelihood that is then optimized to estimate model parameters~\cite{titsias2009variational,hensman2013gaussian}.
Another popular approach with thoroughly investigated theoretical properties is a covariance tapering~\cite{furrer2006covariance,shaby2012tapered}:
we set a covariance between points equal to zero in case a distance between them is above some threshold, so in such a way we obtain sparse covariance matrices, and we can efficiently process them with appropriate routines.
Hierarchical models also can alleviate the computational burden, as they split the sample into separate subsamples. However, in this case exact inference is possible also only if we make some specific assumptions about their covariance structure~\cite{park2010hierarchical,shi2005hierarchical,banerjee2008gaussian}.
In addition, fast exact inference is possible if training data has some specific structure:
for example, in \cite{belyaev2014exact,burnaevKap} authors describe an exact inference scheme to construct Gaussian process regression in case of samples with a Cartesian product structure.
However, as far as we know there are no approaches to large scale variable fidelity Gaussian process regression in case of data without any specific structure.
Another issue with Gaussian process regression is its bad extrapolation properties:
the model prediction at a~new point is a weighted sum of function values at the available training points with weights defined by covariances between these points~\cite{rasmussen2006};
i.e., the prediction can be determined only locally near the training points, and we need to be careful with test points that are far away from the training sample.
We propose two approaches that mitigate the sample size limitation and improve extrapolation properties of variable fidelity Gaussian process regression.
The first approach adopts the Nystr\"om approximation and relies on the results obtained for single fidelity data in the Sparse Gaussian process regression framework~\cite{foster2009stable,kumar2012sampling}.
The second approach uses a low fidelity function blackbox that provides low fidelity function evaluations on the fly: we improve prediction of a surrogate model at a new point using the low fidelity function value at this point.
While for heuristic models it is a common practice to incorporate a low fidelity function blackbox in this way~\cite{alexandrov2000first,madsen2001multifidelity,sun2010two,zahir2012variable,xu1992integrating}, Gaussian process regression doesn't support its direct usage.
As we are able to evaluate a low fidelity function at any point of a design space, we avoid using a large sample to cover the whole design space.
Instead, we just need to obtain a low fidelity sample that is sufficient for accurate estimation of Gaussian process regression model parameters.
We investigate computational complexity and compare accuracies of the proposed approaches using real and artificial data.
The real problem at hand is an optimization of a~rotating disk in an aircraft engine.
The disk shape optimization problem remains challenging and often involves usage of surrogate modeling of maximal stress and radial displacement of the disk~\cite{huang2011optimal,farshi2004optimum}.
We compare four approaches to construct surrogate models: Gaussian process regression, Gaussian process regression for variable fidelity data, and approaches presented in this paper --- sparse Gaussian process regression for variable fidelity data and Gaussian process regression for variable fidelity data with an available low fidelity function blackbox.
The paper contains the following sections:
\begin{itemize}
\item Section~\ref{sec:gpRegression} describes the Gaussian process regression framework;
\item Section~\ref{sec:variableFidelityGp} outlines the variable fidelity Gaussian process regression framework;
\item Section~\ref{sec:sparse} proposes an approach to construct sparse Gaussian process regression for variable fidelity data;
\item Section~\ref{sec:blackbox} describes our approach to variable fidelity Gaussian process regression with a low fidelity function blackbox;
\item Section~\ref{sec:problem} provides results of computational experiments for both real and artificial data;
\item Conclusions and directions for future research are given in Section~\ref{sec:conclusions}.
\end{itemize}
In Appendix we provide proofs of some technical statements and details on low and high fidelity models for the rotating disk problem.
\section{Gaussian process regression for single fidelity data}
\label{sec:gpRegression}
We consider a training sample $D = (\matX, \vecY) = \left\lbrace\vecX_i, y_i = y(\vecX_i)\right\rbrace_{i = 1}^{\sS}$,
where a point $\vecX \in \bbX \subseteq \bbR^{\iD}$ and a function value $y(\vecX) \in \bbR$.
We assume that $y(\vecX) = f(\vecX) + \varepsilon$,
where $f(\vecX)$ is a realization of Gaussian process, and $\varepsilon$ is a Gaussian white noise with a variance $\sigma^2$.
The goal is to construct a surrogate model for the target function $f(\vecX)$.
The Gaussian process $f(\vecX)$ is defined by its mean and covariance function
\[
k(\vecX, \vecX') = \cov (f(\vecX), f(\vecX')) = \mathbb{E} \left(f(\vecX) - \mathbb{E} (f(\vecX)) \right) \left(f(\vecX') - \mathbb{E} (f(\vecX'))\right).
\]
Without loss of generality we assume the mean value to be zero.
We also assume that the covariance function belongs to some parametric family $\{k_{\vecT}(\vecX, \vecX'), \vecT \in \Theta\subseteq \bbR^{p}\} $;
i.e., $k(\vecX, \vecX') = k_{\vecT}(\vecX, \vecX')$ for some $\vecT \in \Theta$.
Thus $y(\vecX)$ is also a Gaussian process~\cite{rasmussen2006} with zero mean and the covariance function
$\cov (y(\vecX), y(\vecX')) = k_{\vecT}(\vecX, \vecX') + \sigma^2 \delta(\vecX - \vecX')$, where $ \delta(\vecX - \vecX')$ is the Kronecker delta.
The multivariate squared exponential covariance function~\cite{rasmussen2006} \[k_{\vecT}(\vecX, \vecX') = \theta_0^2\exp\left(-\sum_{k=1}^d\theta_k^2(x_k-x_k')^2\right)\] is widely used in applications.
The covariance function parameters $\vecT$ and the variance $\sigma^2$ fully specify the data model. We use the Maximum Likelihood Estimation (MLE) of $\vecT$ and $\sigma^2$~\cite{bishop2006pattern,rasmussen2006} to fit the model; i.e.,
we maximize the logarithm of the training sample likelihood
\begin{equation}
\label{eq:likelihood}
\log p (\vecY| \matX, \vecT, \sigma^2) = - \frac{1}{2} \left(\sS \log 2 \pi + \log |\matK| + \vecY^T \matK^{-1} \vecY \right) \rightarrow \max_{\vecT, \sigma^2},
\end{equation}
where $\matK = \{k_{\vecT}(\vecX_i, \vecX_j) + \sigma^2 \delta(\vecX_i - \vecX_j)\}_{i, j = 1}^{\sS}$ is the matrix of covariances between values $\vecY(\matX)$ from the training sample and $|\matK|$ is its determinant. Here
$\sigma^2$ plays the role of a regularization parameter for the kernel matrix $ \{k_{\vecT}(\vecX_i, \vecX_j)\}_{i, j = 1}^{\sS}$, being a matrix of covariances between the values $f(\matX)$.
The recent theoretical paper~\cite{burnaev2013bernstein} and the experimental papers~\cite{bachoc2013cross,zaitsev2013properties,zaytsev2014properties} state that under general assumptions MLE parameters estimates~$\hat{\vecT}$
are accurate even if the sample size is limited and the model is misspecified.
Using estimates of $\vecT$ and $\sigma^2$ we can calculate the posterior mean and the covariances of $y(\vecX)$ at new points playing, respectively, the role of
a~prediction and its uncertainty.
The posterior mean $\mathbb{E} (\vecY(\matX^*) | \vecY(\matX))$ at the new points $\matX^* = \{\vecX^*_i\}_{i = 1}^{\sS^*}$ has the form
\begin{equation}
\label{eq:postMean}
\hat{\vecY}(\matX^*) = \matK(\matX^*, \matX)\cdot \matK^{-1} \vecY,
\end{equation}
where $\matK(\matX^*, \matX) = \{k_{\vecT}(\vecX^*_i, \vecX_j)\}_{i = 1, \ldots, \sS^*, j = 1, \ldots, \sS}$ are the covariances between the values $\vecY(\matX^*)$ and $\vecY(\matX)$.
The posterior covariance matrix $\mathbb{V} \left(\matX^*\right) = \mathbb{E}\bigl[(\vecY(\matX^*) - \mathbb{E} \vecY(\matX^*))^T (\vecY(\matX^*) - \mathbb{E} \vecY(\matX^*))\left|\right. \vecY(\matX)\bigr]$ has the form
\begin{equation}
\label{eq:postCov}
\mathbb{V} \left(\matX^*\right) = \matK(\matX^*, \matX^*) - \matK(\matX^*, \matX)\cdot \matK^{-1}\cdot \matK(\matX, \matX^*),
\end{equation}
where $\matK(\matX^*, \matX^*) = \{k_{\vecT}(\vecX^*_i, \vecX^*_j) + \sigma^2 \delta(\vecX_i^* - \vecX_j^*)\}_{i, j = 1}^{\sS^*}$ is the matrix of covariances between the values $\vecY(\matX^*)$.
Maximum likelihood estimation of a Gaussian process regression model
sometimes provides degenerate results --- a phenomenon closely connected to overfitting~\cite{zaitsev2013properties,zaytsev2014properties,neal1997monte,pepelyshev2010role}.
To regularize the problem and avoid inversion of large ill-conditioned matrices, one can impose a prior distribution on a Gaussian process regression model and then use Bayesian MAP (Maximum A Posteriory) estimates~\cite{burnaev2016surrogate,burnaev2015surrogate,bishop2006pattern}.
In particular in this paper we adopted the approach described in~\cite{burnaev2016surrogate}:
we impose prior distributions on all parameters of the covariance function and additional hyperprior distributions on parameters of the prior distributions.
Experiments confirm that such approach allows to avoid ill-conditioned and degenerate cases, that can occur even more often when processing variable fidelity data.
\section{Variable fidelity Gaussian process regression}
\label{sec:variableFidelityGp}
Now we consider the case of variable fidelity data: we have a sample of low fidelity function evaluations $D_l = (\matX_l, \vecY_l) = \left\lbrace \vecX^l_i, y_l(\vecX^l_i)\right\rbrace_{i = 1}^{\sS_l}$ and a sample of high fidelity function evaluations
$D_h = (\matX_h, \vecY_h) = \left\lbrace \vecX^h_i, y_h(\vecX^h_i) \right\rbrace_{i = 1}^{\sS_h}$
with
$\vecX^l_i, \vecX^h_i \in \bbR^{\iD}$,
$y_l(\vecX), y_h(\vecX) \in \bbR$.
The low fidelity function $y_l(\vecX)$ and the high fidelity function $y_h(\vecX)$ model the same physical phenomenon but with different fidelities.
Using samples of low and high fidelity functions values our aim is to construct a~surrogate model $\hat{y}_h(\vecX) \approx y_h(\vecX)$ of the~high fidelity function; moreover, we also would like to provide a corresponding uncertainty estimate \cite{nazarov2016,vovk2014}.
In this paper we consider a well-known variable fidelity data model (co-kriging) \cite{forrester2007multi}:
\begin{equation*}
y_l(\vecX) = f_l(\vecX) + \varepsilon_l, \,\,y_h(\vecX) = \rho y_l(\vecX) + y_d(\vecX),
\end{equation*}
where $y_d(\vecX) = f_d(\vecX) + \varepsilon_d$.
Here $f_l(\vecX)$, $f_d(\vecX)$ are
realizations of independent Gaussian processes with zero means and covariance functions $k_l(\vecX, \vecX')$ and $k_d(\vecX, \vecX')$, respectively, and
$\varepsilon_l$, $\varepsilon_d$ are Gaussian white noise processes with variances $\sigma_l^2$ and $\sigma_d^2$, respectively.
We also set
$
\matX =
\begin{pmatrix}
\matX_l \\
\matX_h
\end{pmatrix},$
$\vecY =
\begin{pmatrix}
\vecY_l \\
\vecY_h
\end{pmatrix} \!.
$
Then the posterior mean of high-fidelity values at new points has the form
\begin{equation}
\hat{\vecY}_h(\matX^*) = \matK(\matX^*, \matX) \cdot\matK^{-1} \vecY,
\label{eq:VFGPmean}
\end{equation}
where
\begin{align*}
&\matK(\matX^*, \matX) =
\begin{pmatrix}
\rho \matK_l(\matX^*, \matX_l)\,\, &\,\, \rho^2 \matK_l(\matX^*, \matX_h) + \matK_d(\matX^*, \matX_h)
\end{pmatrix},\\
&\matK(\matX, \matX) =
\begin{pmatrix}
\matK_l(\matX_l, \matX_l) \,\,&\,\, \rho \matK_l(\matX_l, \matX_h)\\
\rho \matK_l(\matX_h, \matX_l) \,\,& \,\,\rho^2 \matK_l(\matX_h, \matX_h) + \matK_d(\matX_h, \matX_h)
\end{pmatrix},
\end{align*}
$\matK_l(\matX_a, \matX_b)$, $\matK_d(\matX_a, \matX_b)$ are matrices of pairwise covariances between $y_l(\vecX)$ and $y_d(\vecX)$ and points from some samples $\matX_a$ and $\matX_b$, respectively.
The posterior covariance matrix has the form
\begin{equation}
\mathbb{V} \left(\matX^* \right) = \rho^2 \matK_l(\matX^*, \matX^*) + \matK_d(\matX^*, \matX^*) - \matK(\matX^*, \matX)\cdot \matK^{-1}\cdot \left( \matK(\matX^*, \matX) \right)^T.
\label{eq:VFGPvariance}
\end{equation}
To estimate covariance function parameters and noise variances for the Gaussian processes $f_l(\vecX)$ and $f_d(\vecX)$ we use the following general approach~\cite{forrester2007multi}:
\begin{enumerate}
\item Estimate parameters of the covariance function $k_l(\vecX, \vecX')$ by MLE with a sample $D = D_l$, see Section~\ref{sec:gpRegression}.
\item Calculate posterior mean values $\hat{y}_l(\vecX)$ of the Gaussian process $y_l(\vecX)$ for $\vecX\in\matX_h$,
\item Estimate parameters of the covariance function $k_d(\vecX, \vecX')$, defining the Gaussian process $y_d(\vecX)$, and $\rho$ by maximizing likelihood \eqref{eq:likelihood} with
$D = D_{\text{\rm diff}} = (\matX_h, \vecY_d = \vecY_h - \rho \hat{\vecY}_l (\matX_h))$ and $k(\vecX, \vecX') = k_d(\vecX, \vecX')$.
\end{enumerate}
As we have a big enough sample of low fidelity data, we assume that we can get accurate estimates of parameters of the covariance function $k_l(\vecX, \vecX)$, so we don't need to refine these estimates using high fidelity data.
\section{Sparse Gaussian process regression}
\label{sec:sparse}
To perform inference for a variable fidelity Gaussian process regression we have to invert the sample covariance matrix of size $\sS \times \sS$, where $\sS = \sS_h + \sS_l$.
This operation has complexity $O(\sS^3)$, so for samples containing more than several thousands of points we cannot construct a~Gaussian process regression in a reasonable time.
In order to construct a Gaussian process regression in case of large sample sizes we propose to use an approximation to the exact inference.
The Nystr\"om approximation~\cite{burnaev11sparse,foster2009stable} of all involved positive definite matrices~$\matK(\matX^*, \matX)$, $\matK$ and $\matK(\matX^*, \matX^*)$ allows one to construct such approximation. As a basic building block of the approximation we use the Cholesky decomposition in the form similar to~\cite{foster2009stable}. This approach provides guarantees of improved numerical stability and requires reasonable amount of computations.
Let us select from the initial sample a subsample $
\matX^1 =
\begin{pmatrix}
\matX_l^1 \\
\matX_h^1
\end{pmatrix},
\vecY^1 =
\begin{pmatrix}
\vecY_l(\matX_l^1) \\
\vecY_h(\matX_h^1)
\end{pmatrix}$ of \textit{base} points with a small enough size $\sS_1 = \sS^1_h + \sS^1_l$ so we can perform the exact inference for it. The simplest, rather robust and efficient way is to perform selection without repetitions among points from the initial samples with a probability to select a point being proportional to the corresponding self-covariance value.
Hence, by definition,
\begin{align*}
&\matK_{11} =
\begin{pmatrix}
\matK_l(\matX_l^1, \matX_l^1) \,\,&\,\, \rho \matK_l(\matX_l^1, \matX_h^1) \\
\rho \matK_l(\matX_h^1, \matX_l^1) \,\,&\,\, \rho^2 \matK_l(\matX_h^1, \matX_h^1) + \matK_d(\matX_h^1, \matX_h^1)\\
\end{pmatrix}, \\
&\matK_{1} =
\begin{pmatrix}
\matK_l(\matX_l^1, \matX_l) \,\,& \,\,\rho \matK_l(\matX_l^1, \matX_h) \\
\rho \matK_l(\matX_h^1, \matX_l) \,\,&\,\, \rho^2 \matK_l(\matX_h^1, \matX_h) + \matK_d(\matX_h^1, \matX_h)
\end{pmatrix}, \\
\matK^*_{1} &=
\begin{pmatrix}
\rho \matK_l(\matX^*, \matX_l^1) \,\,& \,\,\rho^2 \matK_l(\matX^*, \matX_h^1) + \matK_d(\matX^*, \matX^1_h)
\end{pmatrix}
\end{align*}
for some new points $\matX^* = \{\vecX^*_i\}_{i = 1}^{\sS^*}$ and so we get the Nystr\"om approximation of the matrices
$\matK(\matX^*, \matX)$, $\matK$ and $\matK(\matX^*, \matX^*)$, respectively:
\begin{equation*}
\hat{\matK}(\matX^*, \matX) = \matK_1^* \matK_{11}^{-1} \matK_1, \,\,
\hat{\matK} = (\matK_1)^T \matK_{11}^{-1} \matK_1, \,\,
\hat{\matK}(\matX^*, \matX^*) = \matK_1^* \matK_{11}^{-1} (\matK_1^*)^T.
\end{equation*}
We set
\[
\matR = \begin{pmatrix}
\frac{1}{\sigma_l} \matI_{\sS_l} & 0 \\
0 & \frac{1}{\sqrt{\rho^2 \sigma_l^2 + \sigma_d^2}} \matI_{\sS_h} \\
\end{pmatrix},
\]
where $\matI_{k}$ is an identity matrix of size $k$, $\matC_1 = \matR \matK_1$ and $\matV = \matC_1 \matV_{11}^{-T}$, $\matV_{11}$ is the Cholesky decomposition of $\matK_{11}$.
\begin{theorem}
\label{th:stat1}
For the posterior mean and the posterior covariance matrix the following Nystrom approximations hold
\begin{align}
&\hat{\vecY}^{NA}_h(\matX^*) = {\matK}^*_1 \matV_{11}^{-1} (\matI_{\sS_1} + \matV^T \matV)^{-1} \matV^T \matR \vecY, \label{eq:posteriorSparseMean}\\
&\mathbb{V}^{NA} \left(\matX^* \right) = {\matK}_1^* \matV_{11}^{-1} (\matI_{\sS_1} + \matV^T \matV)^{-1} \matV_{11}^{-T} {\matK_1^*}^T + (\rho^2 \sigma_l^2 + \sigma_d^2) \matI_{n^*}.\label{eq:posteriorSparseVar}
\end{align}
\end{theorem}
Note that there are other ways to apply the Nystr\"om approximation to the posterior covariance matrix, but these alternatives either lead to inaccurate approximations or have low numerical stability~\cite{foster2009stable}.
\begin{theorem}
\label{th:stat2}
Computational complexities of the posterior mean and the posterior covariance matrix calculations at one point using~\eqref{eq:posteriorSparseMean} and~\eqref{eq:posteriorSparseVar} are equal to $O(\sS \sS_1^2)$.
\end{theorem}
Proofs of these theorems are provided in Appendix~\ref{section:proof}.
Note that as we use the Nystr\"om approach and select base points at random from the initial sample we can get the following estimate of the approximation accuracy by directly applying results from~\cite{kumar2012sampling}.
\begin{theorem}
\label{th:nystrom_bound}
With probability $1 - \delta$ it holds that:
\begin{align*}
\frac{\|\matK(\matX^*, \matX^*) - \hat{\matK}(\matX^*, \matX^*) \|_2}{\|\matK(\matX^*, \matX^*) \|_2} &\leq
\frac{\|\matK(\matX^*, \matX^*) - \hat{\matK}_{\sS_1}(\matX^*, \matX^*) \|_2}{\|\matK(\matX^*, \matX^*) \|_2} + \Delta, \\
\frac{\|\matK(\matX^*, \matX) - \hat{\matK}(\matX^*, \matX) \|_2 }{\|\matK(\matX^*, \matX) \|_2} &\leq
\frac{\|\matK(\matX^*, \matX) - \hat{\matK}_{\sS_1}(\matX^*, \matX) \|_2}{\|\matK(\matX^*, \matX) \|_2} + \Delta,
\end{align*}
where $\Delta$ is of order $O\left(\frac{1}{\sqrt{\sS}}\right) O\left(\sqrt{\log \frac{1}{\delta}}\right)$, $\|\cdot\|_2$ is $l_2$ matrix norm,
and $\hat{\matK}_{\sS_1}(\matX^*, \matX^*)$ is the best approximation with respect to $l_2$ matrix norm, having rank~$\sS_1$.
\end{theorem}
\section{Variable Fidelity Gaussian process regression with a low fidelity function blackbox}
\label{sec:blackbox}
Suppose that we have a blackbox for the low fidelity function $y_l(\vecX)$ that estimates the low fidelity function value at any point from the design space $\bbX \subseteq \bbR^{\iD}$ on the fly.
Let us assume that we have already constructed a Variable fidelity Gaussian process surrogate model and can calculate predictions using
\eqref{eq:VFGPmean} and \eqref{eq:VFGPvariance}.
We can not use a huge sample of low fidelity function values due to the high computational cost of the Gaussian process regression.
Instead, in order to improve the prediction accuracy we can update the posterior mean and the posterior variance of $y_h(\vecX)$ at the new point $\vecX$ with the low fidelity function value $y_l(\vecX)$
at this point, calculated by the blackbox.
Let us describe a computationally efficient procedure to calculate the update. We set
\[
\vecK_l(\vecX, \matX) =
\begin{pmatrix}
\matK_l(\vecX, \matX_l) \\
\rho \matK_l(\vecX, \matX_h) \\
\end{pmatrix},
\]
where $\vecX$ is the new point.
For a sample with the additional point $\vecX$ we get the expanded covariance matrix:
\[
\matK_{\text{\rm exp}} =
\begin{pmatrix}
\matK & \vecK_l \\
\vecK_l^T & k_l(\vecX, \vecX)
\end{pmatrix}.
\]
Suppose we know Cholesky factors $\matL$ and $\matL^{-1}$ of the initial training sample covariance matrix $\matK$ and its inverse $\matK^{-1}$, respectively.
To perform computations efficiently, we update these Cholesky factors and then update the posterior mean and the posterior variance values for the expanded sample.
For the matrix $\matK_{\sS}\in\mathbb{R}^{\sS \times \sS}$ and its Cholesky decomposition using a standard approach (see Appendix~\ref{section:blackbox_chol_update}) we can get the updated Cholesky decomposition of the expanded
matrix $\matK_{\sS + 1}\in\mathbb{R}^{(\sS + 1) \times (\sS + 1)}$ in $O(\sS^2)$ steps if the initial matrix $\matK_{\sS}$ is located in the upper left corner of the new matrix $\matK_{\sS + 1}$.
To update the inverse of the Cholesky decomposition we also need~$O(\sS^2)$ operations, as the expanded Cholesky factor is lower triangular and differs from the initial Cholesky factor only in the last row.
Therefore, we can calculate the matrix $\matK_{\text{\rm exp}}^{-1}$ in $O(\sS^2)$ operations.
An expanded vector of covariances between the new point $\vecX$ and the initial training sample has the form
\[
\vecK_{\text{\rm exp}} =
\begin{pmatrix}
\rho \matK_l(\vecX, \matX_l) \\
\rho^2 \matK_l(\vecX, \matX_h) + \matK_d(\vecX, \matX_h) \\
\rho k_l(\vecX, \vecX)
\end{pmatrix}.
\]
We set
$
\vecY_{\text{\rm exp}} =
\left( \vecY^T, y_l(\vecX) \right)^{T},
$
where $y_l(\vecX)$ is calculated by the blackbox.
Then updated expressions for the posterior mean and the posterior variance are:
\begin{align}
&\hat{y}^{\text{\rm exp}}_h (\vecX) = \vecK_{\text{\rm exp}} \matK_{\text{\rm exp}}^{-1} \vecY_{\text{\rm exp}}, \label{eq:posteriorMeanBb}\\
&\mathbb{V}_{\text{\rm exp}} \left(\vecX \right) = \rho^2 \matK_l(\vecX, \vecX) + \matK_d(\vecX, \vecX) - \vecK_{\text{\rm exp}}^T \matK_{\text{\rm exp}}^{-1} \vecK_{\text{\rm exp}}.\label{eq:posteriorVarBb}
\end{align}
As the Cholesky factor for the updated model differs only in the last row we calculate \eqref{eq:posteriorMeanBb} and \eqref{eq:posteriorVarBb} in~$O(\sS^2)$ operations.
The total computational complexity is the sum of computational complexities of the Cholesky decomposition update and the posterior mean and variance recalculation, so for a Variable fidelity Gaussian process regression with a blackbox, representing the low fidelity function, the following theorem holds true.
\begin{theorem}
Suppose we know Cholesky factors $\matL$ and $\matL^{-1}$ of the initial training sample covariance matrix $\matK$ and its inverse $\matK^{-1}$, respectively.
Then we can calculate the posterior mean $\hat{y}^{\text{\rm exp}}_h (\vecX)$ via \eqref{eq:posteriorMeanBb} and the variance $\mathbb{V}_{\text{\rm exp}} \left(\vecX \right)$ via \eqref{eq:posteriorVarBb} in $O(\sS^2)$ operations, where $\sS = \sS_l + \sS_h$.
\end{theorem}
As we add only one point to the initial training sample, we expect that estimates of parameters of the Gaussian process regression model do not change significantly.
While in some cases it can be reasonable to add many points, this issue raises a complex question on how and when we should re-estimate Gaussian process parameters as we add more points.
Using blackbox for the low fidelity function we can get significantly more accurate approximation with a small additional computational cost.
\section{Numerical experiments}
\label{sec:problem}
We compare four approaches for a surrogate model construction, listed below:
\begin{itemize}
\item GP --- Gaussian Process Regression using only high fidelity data;
\item VFGP --- Variable Fidelity Gaussian Process Regression using both high and low fidelity data;
\item SVFGP --- Sparse VFGP, which is a version of VFGP for the case of large training samples, introduced in Section~\ref{sec:sparse};
\item BB VFGP --- VFGP with a low fidelity function realized by a black box, introduced in Section~\ref{sec:blackbox}.
In experiments we use the same design of experiments as in the case of VFGP, while for a surrogate model update for each new point we use a low fidelity function value at this point.
\end{itemize}
To estimate parameters of SVFGP we use a randomly selected subsample, while we use the full sample to perform approximate inference. In experiments we always use the multivariate squared exponential covariance function, see~\cite{rasmussen2006}.
We measure accuracy of surrogate models by RRMS (relative root mean square) error estimated by either the cross-validation procedure~\cite{hastie2005elements} or using a separate test sample not involved in model training process.
RRMS error typically lies between $0$ and $1$.
RRMS error for an accurate model is close to $0$, while RRMS error for an inaccurate model is close to or greater than~$1$.
In case of a one-dimensional output and a test sample $D_{\text{\rm test}} = \{\vecX^{\text{\rm test}}_i, y_{i}^{\text{\rm test}} = f_h(\vecX^{\text{\rm test}}_i)\}_{i = 1}^{\sS_t}$
the RRMS error of a~surrogate model $\hat{y}(\vecX)$ is equal to
\[
RRMS(D_{\text{\rm test}}, \hat{y}) = \sqrt{\frac{\sum_{i = 1}^{\sS_t} (\hat{y}_h(\vecX^{\text{\rm test}}_i) - y^{\text{\rm test}}_i)^2}{\sum_{i = 1}^{\sS_t} (\overline{y} - y^{\text{\rm test}}_i)^2}},
\]
where $\overline{y} = \frac{1}{\sS_t} \sum_{i = 1}^{\sS_t} y^{\text{\rm test}}_i$.
In this section we assess presented approaches using several surrogate modeling problems: two artificial problems and a real problem of surrogate modeling and optimization of a rotating disk from an aircraft engine. We want to validate whether our approaches fit into requirements described in Introduction section \ref{introintro}:
we examine model construction times and accuracy of surrogate models intended to solve problems of extrapolation and interpolation.
\subsection{Toy problem}
\label{sec:artificialProblemBbVfgp}
Here we consider the well-known test problem~\cite{forrester2007multi} to construct a variable fidelity surrogate model. Data is generated by the following high fidelity $y_h(x)$ and low fidelity $y_l(x)$ functions:
\begin{align*}
y_h(x) &= (6 x - 2)^2 \sin(12 x - 4), \\
y_l(x) &= 0.5 y_h(x) + 10 (x - 1).
\end{align*}
As the problem is simple and so large samples are not required we do not perform comparison with SVFGP in this subsection.
In order to evaluate accuracy of various algorithms we use the following procedure:
\begin{itemize}
\item Generate a high fidelity sample of size $n_h \leq 100$ with points uniformly distributed in $[0, 1]$. We consider $n_h = 6, 15$ and $30$,
\item Generate a low fidelity sample with points from the high fidelity sample and additional $(100 - n_{h})$ points uniformly distributed in $[0, 1]$,
\item Construct surrogate models using GP, VFGP and BB VFGP methods and estimate their accuracies on the test sample consisting of $1000$ high fidelity function values.
\end{itemize}
RRMS errors, provided in Table~\ref{table:rrmsArtificial}, are averaged over $50$ runs for each considered value of $n_h$.
We see that using the low fidelity function blackbox we can significantly improve accuracy for all considered values of $n_h$.
\begin{table}
\centering
\caption{Toy problem. RRMS errors for various high fidelity training sample sizes $n_h$}
\label{table:rrmsArtificial}
\begin{tabular}{cccc}
\hline
$n_h$ & $6$ & $15$ & $30$ \\
\hline
GP & 0.7102 & 0.0159 & $3.83e-04$ \\
VFGP & 0.3036 & $7.42e-04$ & $1.38e-04$ \\
BB VFGP & 0.1610 & $6.90e-07$ & $1.67e-07$ \\
\hline
\end{tabular}
\end{table}
\subsection{Artificial problem in case of large samples}
\label{sec:artificialProblemSparse}
To benchmark proposed approaches we use artificial test functions with multiple local peculiarities and input dimension $d = 6$, so we really need a rather big sample to get an accurate surrogate model. As the high fidelity function $y_h(\vecX)$ and the low fidelity function $y_l(\vecX)$ we consider
\begin{align*}
y_h(\vecX) &= 20 + \sum_{i = 1}^{d} (x_i^2 - 10 \cos (2 \pi x_i)) + \varepsilon_h,\,\vecX\in[0,1]^d, \\
y_l(\vecX) &= y_h(\vecX) + 0.2 \sum_{i = 1}^{d} (x_i + 1)^2 + \varepsilon_l,\,\vecX\in[0,1]^d.
\end{align*}
The high fidelity function is corrupted by the Gaussian white noise $\varepsilon_h$ with variance $0.001$, and
the low fidelity function is corrupted by the Gaussian white noise $\varepsilon_l$ with variance $0.002$.
We generate points in $[0,1]^d$ using Latin Hypercube Sampling~\cite{park1994optimal}.
To test extrapolation properties we limit the training sample points to the region with the range $[0, 0.5]$ instead of $[0, 1]$ for one of $6$ input variables.
The high fidelity sample size is $n_h = 100$ and the size of the subsample for SVFGP is $\sS_l^1=1000$ in all experiments.
Results are averaged over $5$ runs for each considered value of $n_l$. In Table~\ref{table:errorArtSparse} for VFGP, SVFGP, and BB VFGP approaches we provide RRMS values in the extrapolation and interpolation regimes, as well as training times.
One can see that RRMS errors of SVFGP are comparable with RRMS errors of VFGP for the same sample size, while the training time of SVFGP is tremendously smaller when the sample size is equal to $5000$, and for SVFGP the training time increases only slightly when the sample size increases.
For BB VFGP the training time in this experiment coincides with that of VFGP, while in case of $1000$ training points we get better results with BB VFGP than in case of $5000$ training points and VFGP. Also we can see that in the extrapolation regime we get significantly better results with BB VFGP.
\begin{table}
\centering
\caption{Surrogate modeling for large samples of artificial data}
\label{table:errorArtSparse}
\begin{tabular}{clll}
\multicolumn{4}{c}{RRMS errors in case of the interpolation regime} \\
\hline
$n_l$ & 1000 & 3000 & 5000 \\
\hline
VFGP & $0.0502 $ & $0.0170 $ & $0.0058 $ \\
SVFGP & $0.0502 $ & $0.0305 $ & $0.0260 $ \\
BB VFGP & $0.0010 $ & $0.00029 $ & $0.00017 $ \\
\hline
\\
\multicolumn{4}{c}{RRMS errors in case of the extrapolation regime} \\
\hline
$n_l$ & 1000 & 3000 & 5000 \\
\hline
VFGP & $0.3636 $ & $0.1351$ & $0.1028 $ \\
SVFGP & $0.3636 $ & $0.3281$ & $0.3586 $ \\
BB VFGP & $0.000998$ & $0.00113 $ & $0.00034 $ \\
\hline
\\
\multicolumn{4}{c}{The training time in seconds, } \\
\multicolumn{4}{c}{Ubuntu PC, Intel-Core $i7$, $16$ Gb RAM} \\
\hline
$n_l$ & 1000 & 3000 & 5000 \\
\hline
VFGP & $30.46$ & $852.70$ & $7283.27$ \\
SVFGP & $30.46$ & $33.42$ & $37.50$ \\
BB VFGP & $30.38$ & $842.97$ & $7672.60$ \\
\hline
\end{tabular}
\end{table}
\subsection{Rotating disk problem}
\label{sec:realProblem}
\subsubsection{Rotating disk model description}
A high speed rotating risk is an important part of an aircraft engine (see Figure~\ref{fig:engine}).
Three parameters define performance characteristics of the disk: the mass of the disk, the maximal radial displacement $u_{\text{\rm max}}$, the maximal stress $s_{\text{\rm max}}$~\cite{armand1995structural,mohan2013structural,belyaev2014exact}.
It is easy to calculate mass of the disk, as we know all geometrical parameters of the disk, while surrogate modeling of the maximal radial displacement and the maximal stress is needed since these characteristics are computationally expensive~\cite{mohan2013structural,huang2011optimal}.
So the focus here is on modeling the maximal radial displacement and the maximal stress.
Used parametrization of the rotating disk geometry consists of $8$ parameters: the radii $r_i$, $i = 1, \ldots, 6$, which control where the thickness of the rotating disk changes, and
the values $t_1$, $t_3$, $t_5$, which control the corresponding changes in thickness.
In the considered surrogate modeling problem we fix the radii $r_4$, $r_5$ and the thickness $t_3$ of the rotating disk, so the input dimension for the surrogate model is~$6$.
The geometry and the parametrization of the~rotating disk are shown in Figure~\ref{fig:rotating_disk}.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{engine_disk}
\caption{Aircraft engine. Rotating disk is shown by the bold rectangle at the right side of the figure}
\label{fig:engine}
\end{subfigure}%
~
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{rotating_disk}
\caption{Rotating disk geometry}
\label{fig:rotating_disk}
\end{subfigure}
\caption{Rotating disk problem}
\end{figure}
There are two available solvers for $u_{\text{\rm max}}$ and $s_{\text{\rm max}}$ calculation.
The low fidelity function is realized by the Ordinary Differential Equations (ODE) solver based on a simple Runge--Kutta's method.
The high fidelity function is realized by the Finite Element Model (FEM) solver.
A~single evaluation of the low fidelity function takes $\sim 0.01$ seconds, and a~single evaluation of the high fidelity function takes $\sim 300$ seconds.
More detailed comparison of the solvers is provided in Appendix~\ref{section:solvers}.
\subsubsection{Surrogate model accuracy}
\label{sec:surrogateModel}
\begin{table}
\centering
\caption{RRMS errors (with standard deviations) for the developed approaches}
\label{table:rrmsBBsampleSizeSmax}
\begin{tabular}{ccccc}
\multicolumn{5}{c}{Output $u_{\text{\rm max}}$} \\
\hline
$n_h$ & $20$ & $40$ & $60$ & $80$ \\
\hline
GP & $0.287 \pm 0.039$ & $0.143 \pm 0.031$ & $0.082 \pm 0.020$ & $0.095 \pm 0.023$ \\
VFGP & $0.212 \pm 0.075$ & $0.088 \pm 0.009$ & $0.064 \pm 0.007$ & $0.068 \pm 0.006$ \\
SVFGP & $0.125 \pm 0.029$ & $0.074 \pm 0.016$ & $0.041 \pm 0.007$ & $0.047 \pm 0.011$ \\
BB VFGP & $0.123 \pm 0.019$ & $0.053 \pm 0.008$ & $0.030 \pm 0.007$ & $0.034 \pm 0.006$ \\
\hline
\\
\multicolumn{5}{c}{Output $s_{\text{\rm max}}$} \\
\hline
$n_h$ & $20$ & $40$ & $60$ & $80$ \\
\hline
GP & $0.505 \pm 0.10$ & $0.367 \pm 0.15$ & $0.251 \pm 0.049$ & $0.196 \pm 0.014$ \\
VFGP & $0.363 \pm 0.07$ & $0.261 \pm 0.06$ & $0.193 \pm 0.011$ & $0.123 \pm 0.043$ \\
SFGP & $0.190 \pm 0.06$ & $0.122 \pm 0.06$ & $0.119 \pm 0.015$ & $0.088 \pm 0.027$ \\
BB VFGP & $0.158 \pm 0.03$ & $0.162 \pm 0.03$ & $0.137 \pm 0.024$ & $0.078 \pm 0.020$ \\
\hline
\end{tabular}
\end{table}
In this section we compare our approaches SVFGP (Sparse variable fidelity Gaussian processes) and BB VFGP (Blackbox variable fidelity Gaussian processes) with GP based only on high fidelity data and VFGP baseline methods.
We use the Latin Hypercube approach to sample points.
The low fidelity training sample size is equal to $1000$,
the high fidelity training sample size $n_h$ is $20$, $40$, $60$, and $80$ in different experiments.
To estimate accuracy of the high fidelity function prediction we run the cross-validation procedure, applied to $140$ high fidelity data points (these points contain $n_h$ points used for training of surrogate models).
For each fixed sample size $n_h$ we generate $5$ splits of the data into training and test samples to estimate means and standard deviations.
For SVFGP, we use $n_l=5000$ low fidelity points in total, and randomly select $n_l^1=1000$ points from them to use as base points.
The results are given in Table~\ref{table:rrmsBBsampleSizeSmax} for $u_{\text{\rm max}}$ and $s_{\text{\rm max}}$ outputs:
VFGP outperforms GP, and both SVFGP and BB VFGP outperform VFGP in terms of RRMS error.
Therefore, we should decide which one to use, SVFGP or BB VFGP, by taking into account whether the blackbox for the low fidelity function is available, or whether one uses the surrogate model in extrapolation regime, etc.
\subsection{Optimization of the rotating disk shape}
The problem is to optimize the shape of the rotating disk:
\begin{align}
\label{eq:optProblem}
m, u_{\text{\rm max}} &\rightarrow \min_{r_1, \ldots, r_6, t_1, t_3, t_5}, \\
u_{\text{\rm max}} &\leq 0.3, s_{max} \leq 600, \nonumber \\
10 &\leq r_1 \leq 110, 120 \leq r_2 \leq 140, \nonumber \\
150 &\leq r_3 \leq 168, 170 \leq r_4 \leq 200, \nonumber \\
4 &\leq t_1 \leq 50, 4 \leq t_3 \leq 50, \nonumber \\
r_5 &= 210, r_6 = 230, t_5 = 32. \nonumber
\end{align}
This problem has multiple objectives, and we are looking for a Pareto frontier, not a single solution.
Single optimization run can be described as follows:
\begin{itemize}
\item Generate an initial high fidelity sample $D_h$ with $30$ points using the Latin Hypercube sampling;
\item Construct surrogate models using GP, VFGP, SVFGP and BB VFGP approaches using the generated high fidelity sample $D_h$ and a low fidelity sample $D_l$ of size $1000$ for GP, VFGP and BB VFGP and of size $5000$ for SVFGP;
\item Solve multiobjective optimization problem \eqref{eq:optProblem} using the constructed surrogate models as the target functions and the constraints;
\item Using the high fidelity solver calculate true values at Pareto frontiers, constructed on the previous step, to estimate quality of the models.
\end{itemize}
Due to properties of the used multiobjective optimization algorithm, sizes of the Pareto frontiers can slightly differ for different optimization runs, with an average size of a Pareto frontier equal approximately to $30$ points~\cite{druot2013multi}.
So we need about $50-60$ runs of the high fidelity function to solve this optimization problem ($30$ high fidelity function evaluations to generate the initial sample and $20-30$ high fidelity function evaluations to calculate the true values at the constructed Pareto frontier).
In order to recover a reference Pareto frontier we constructed an accurate surrogate model using $5000$ high fidelity evaluations on a uniform design over the whole design space and additional sampling in a region where points of the Pareto frontier are located.
So instead of using the high fidelity solver during optimization runs we used this surrogate model.
Examples of the obtained Pareto frontiers for a single optimization run is provided in Figure~\ref{fig:pareto}.
In these runs SVFGP and BB VFGP work better than GP and VFGP.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{30_15.pdf}
\end{subfigure}%
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{30_20.pdf}
\end{subfigure}
\caption{Pareto frontiers obtained by optimizing surrogate models constructed with GP, VFGP, SVFGP and BB VFGP approaches along with the reference Pareto frontier}
\label{fig:pareto}
\end{figure}
Results of optimization are given in Table~\ref{table:optimizationResults}.
We compare minimum values of different weighted sums of the two target variables $m$ and $u_{\text{\rm max}}$ averaged over $10$ optimization runs for different initial samples.
We obtain the best value of the mass $m$ output using SVFGP algorithm and the best value of $u_{\text{\rm max}}$ using BB VFGP algorithm while optimizations based on GP and VFGP work worse.
Also, with BB VFGP we produce significantly larger amount of feasible points compared to GP, VFGP and SVFGP, which typically leads to better Pareto frontier coverage with a similar number of high fidelity solver runs.
\begin{table}
\centering
\caption{Optimization results for different surrogate models along with minimal values for different optimization objectives. Also we indicate proportion of feasible points in a final Pareto frontier. The best values are indicated in bold font}
\label{table:optimizationResults}
\begin{tabular}{ccccc}
\hline
Objective & GP & VFGP & SVFGP & BB VFGP \\
\hline
$m$ & 16.62 & 15.69 & \bf 15.09 & 15.63 \\
$0.8 m + 0.2 u_{\text{\rm max}}$ & 73.65 & 70.74 & 70.71 & \bf 68.10 \\
$0.6 m + 0.4 u_{\text{\rm max}}$ & 125.10 & 117.37 & 116.21 & \bf 112.55 \\
$0.4 m + 0.6 u_{\text{\rm max}}$ & 176.55 & 163.89 & 161.18 & \bf 156.99 \\
$0.2 m + 0.8 u_{\text{\rm max}}$ & 228.00 & 210.33 & 206.12 & \bf 201.44 \\
$u_{\text{\rm max}}$ & 279.44 & 256.77 & 251.05 & \bf 245.89 \\
\hline
Proportion of the feasible points & 0.54 & 0.57 & 0.55 & \bf 0.75 \\
\hline
\end{tabular}
\end{table}
\section{Conclusions}
\label{sec:conclusions}
We presented two new approaches to variable fidelity surrogate modeling, which allow one to perform large sample inference for Variable Fidelity Gaussian process regression:
the first approach approximates a full sample covariance matrix and its inverse;
the second approach uses a low fidelity black box to update a surrogate model with a low fidelity function value at a point where one wants to estimate a high fidelity function, thus making usage of large low fidelity samples unnecessary.
Using developed approaches we can perform large sample inference for variable fidelity Gaussian process regression and construct more accurate surrogate models.
Our assessment of the proposed approaches by comparing them with state-of-the art methods demonstrate that we improve both accuracy of surrogate models and their training time.
Future directions of our research on surrogate modeling for variable fidelity data include
theoretical investigation and assessing of numerical stability of the proposed approaches, their
adaptation for the case of arbitrary number of fidelities in data~\cite{zaytsev2016reliable}, surrogate based optimization and adaptive design of experiments.
\vspace{15px}
|
\section{Introduction}\label{S:Introduction}
Wormholes are hypothetical handles or tunnels
in spacetime that connect different regions
of our Universe or completely different
universes altogether. That wormholes could
be actual physical structures was first
proposed by Morris and Thorne \cite{MT88}.
These could be described by the static and
spherically symmetric line element
\begin{equation}\label{E:line1}
ds^{2}=-e^{2\Phi(r)}dt^{2}+\frac{dr^2}{1-b(r)/r}
+r^{2}(d\theta^{2}+\text{sin}^{2}\theta\,
d\phi^{2}),
\end{equation}
using units in which $c=G=1$. Here
$\Phi=\Phi(r)$ is referred to as the
\emph{redshift function}, which must be
everywhere finite to avoid an event horizon.
The function $b=b(r)$ is called the
\emph{shape function} since it helps to
determine the spatial shape of the
wormhole when viewed, for example, in an
embedding diagram \cite{MT88}. The
spherical surface $r=r_0$ is the radius of
the \emph{throat} of the wormhole. Here
$b=b(r)$ must satisfy the following conditions:
$b(r_0)=r_0$, $b(r)<r$ for $r>r_0$, and
$b'(r_0)<1$, usually called the
\emph{flare-out condition}. This condition
can only be satisfied by violating the null
energy condition (NEC), defined as follows:
for the stress-energy tensor $T_{\alpha\beta}$,
we must have
\begin{equation*}
T_{\alpha\beta}k^{\alpha}k^{\beta}\ge 0
\end{equation*}
for all null vectors $k^{\alpha}$. For
Morris-Thorne wormholes, matter that
violates the NEC is called ``exotic."
The equation of state (EoS) for a standard
perfect fluid, $p=\omega\rho$, $0<\omega <1$,
was studied a long time ago by Chandrasekhar
\cite{sC72}. The realization that the
Universe is undergoing an accelerated
expansion \cite{aG01, sP99} has led to the
value of $\omega <-1/3$ due to the
Friedmann equation $\frac{\overset{..}{a}}{a}
=-\frac{4\pi}{3}(\rho+3p)$. This case is
known as \emph{quintessence dark energy}.
The value of $\omega=-1$ corresponds to the
existence of Einstein's cosmological
constant \cite{mC01}. The case that has
attracted the most attention in wormhole
physics is $\omega <-1$, referred to as
\emph{phantom dark energy} since this
case leads to a violation of the NEC:
given the null vector $(1,1,0,0)$,
$p+\rho =-\omega\rho +\rho <0$. The EoS
$p=\omega\rho$, $0<\omega <1$, refers to
ordinary (baryonic) matter, as well as to
dark matter.
In this paper we also make use of conformal
symmetry, the existence of a conformal Killing
vector $\xi$ defined by the action of
$\mathcal{L_{\xi}}$ on the metric tensor
\begin{equation}
\mathcal{L_{\xi}}g_{\mu\nu}=\psi(r)\,g_{\mu\nu};
\end{equation}
here $\mathcal{L_{\xi}}$ is the Lie derivative
operator and $\psi(r)$ is the conformal factor.
It is shown in this paper that the assumption of
conformal symmetry implies that a Morris-Thorne
wormhole is necessarily anisotropic. It is then
assumed that a perfect-fluid distribution with
EoS $p=\omega\rho$ can be adapted to an
inhomogeneous spacetime such as a wormhole by
letting $p$ be the lateral pressure $p_t$
instead of $p_r$, as assumed in previous
studies \cite{fL05, sS05}. It is subsequently
shown that a wormhole solution can exist only if
$\omega <-1$ or $0<\omega <1$.
\section{Conformal Killing vectors and wormhole
\\ construction}
This section consists of a brief discussion of the
assumption that our spacetime admits a one-parameter
group of conformal motions. First we need to recall
that these are motions along which the metric tensor
of the spacetime remains invariant up to a scale
factor, which is equivalent to stating that there
exists a set of conformal Killing vectors such that
\begin{equation}\label{E:Lie}
\mathcal{L_{\xi}}g_{\mu\nu}=g_{\eta\nu}\,\xi^{\eta}
_{\phantom{A};\mu}+g_{\mu\eta}\,\xi^{\eta}_{\phantom{A};
\nu}=\psi(r)\,g_{\mu\nu},
\end{equation}
where the left-hand side is the Lie derivative of the
metric tensor and $\psi(r)$ is the conformal factor.
Eq. (\ref{E:Lie}) shows that the vector $\xi$
characterizes the conformal symmetry since the metric
tensor $g_{\mu\nu}$ is conformally mapped into itself
along $\xi$. It must be emphasized that the
assumption of conformal symmetry has proved to be
fruitful in numerous ways, not only leading to new
solutions but to new geometric and kinematical
insights \cite{HPa, HPb, MS93, sR08, fR10, fR12}.
Another fairly recent discovery is that the Kerr
black hole is conformally symmetric \cite{CMS10}.
Exact solutions of traversable wormholes admitting
conformal motions are discussed in Ref.
\cite{RRKKI} by assuming a noncommutative-geometry
background. Two earlier studies assumed a
non-static conformal symmetry \cite{BHL07, BHL08}.
It is shown in Ref. \cite{pK15a} that the line element
\begin{equation}\label{E:line2}
ds^2=- e^{\nu(r)} dt^2+e^{\lambda(r)} dr^2
+r^2( d\theta^2+\text{sin}^2\theta\, d\phi^2)
\end{equation}
is particularly convenient for discussing the
consequences of the conformal-symmetry assumption.
In particular,
\begin{equation}\label{E:gtt}
e^{\nu}=Cr^2
\end{equation}
and
\begin{equation}\label{E:grr}
e^{\lambda}=\psi^{-2}.
\end{equation}
Moreover, the Einstein field equations are
\begin{equation}\label{E:E1}
\frac{1}{r^2}(1 - \psi^2)
-\frac{(\psi^2)^\prime}{r}= 8\pi \rho,
\end{equation}
\begin{equation}\label{E:E2}
\frac{1}{r^2}(3\psi^2-1)= 8\pi p_r,
\end{equation}
and
\begin{equation}\label{E:E3}
\frac{\psi^2}{r^2}+\frac{(\psi^2)'}{r}
=8\pi p_t.
\end{equation}
It is clear from Eq. (\ref{E:gtt}) that the
wormhole spacetime cannot be asymptotically
flat. So the wormhole material must be cut
off at some $r=a$ and joined to an exterior
Schwarzschild solution
\begin{equation}
ds^{2}=-\left(1-\frac{2M}{r}\right)dt^{2}
+\frac{dr^2}{1-2M/r}
+r^{2}(d\theta^{2}+\text{sin}^{2}\theta\,
d\phi^{2}).
\end{equation}
We see from line element (\ref{E:line1}) that
$M=\frac{1}{2}b(a)$. So for $e^{\nu}=Ca^2$,
we have $Ca^2=1-2M/a$ and the constant of
integration becomes
\begin{equation}
C=\frac{1}{a^2}\left(1-\frac{b(a)}{a}
\right).
\end{equation}
\section{Wormhole solution}
In this section we take a closer look at the
EoS $p=\omega\rho$ as it relates to wormholes.
In a cosmological setting, we are dealing with a
homogeneously distributed fluid. On the other
hand, for a given wormhole, the pressure may or
may not be isotropic, but a wormhole admitting
conformal motion is definitely not: suppose,
on the contrary, that $p_r=p_t$. Then from
Eqs. (\ref{E:E2}) and (\ref{E:E3}),
\begin{equation}
\frac{1}{r^2}(3\psi^2-1)=
\frac{\psi^2}{r^2}+\frac{(\psi^2)'}{r}.
\end{equation}
After simplifying, we obtain the differential
equation
\begin{equation*}
(\psi^2)'-\frac{2}{r}\psi^2=-\frac{1}{r},
\end{equation*}
which is linear in $\psi^2$ and readily solved
to obtain
\begin{equation*}
\psi^2=\frac{1}{2}+cr^2.
\end{equation*}
We will see a bit later that to obtain a
wormhole solution, the equation must satisfy
the initial condition $\psi^2(r_0)=0$, where
$r=r_0$ is the throat of the wormhole. The
result is
\begin{equation}
\psi^2(r)=\frac{1}{2}-\frac{1}{2r_0^2}r^2.
\end{equation}
Now observe that for $r>r_0$, $\psi^2(r_0)<0$,
which is impossible since $\psi(r)$ is a
real-valued function.
Returning now to the EoS $p=\omega\rho$
describing a homogeneous distribution, it
is emphasized in Refs. \cite{fL05, sS05}
that an extension to an inhomogeneous
spherically symmetric spacetime is possible
by making $p$ the radial pressure $p_r$,
so that the transverse pressure $p_t$ can
then be determined from the Einstein field
equations.
In this paper we follow the same strategy,
but instead of assuming that $p_r=\omega\rho$,
it is more convenient to use the EoS
\begin{equation}\label{E:EoS}
p_t=\omega\rho;
\end{equation}
($p_r$ is then determined from the Einstein field
equations.) Substituting in this equation,
\begin{equation}
\frac{1}{8\pi}\left(\frac{\psi^2}{r^2}+
\frac{(\psi^2)'}{r}\right)=\frac{\omega}{8\pi}
\left[\frac{1}{r^2}(1-\psi^2)-
\frac{(\psi^2)'}{r}\right].
\end{equation}
After simplifying, we obtain the differential
equation
\begin{equation}
(\psi^2)'+\frac{1}{r}\psi^2=
\frac{\omega}{1+\omega}\frac{1}{r},
\end{equation}
again linear in $\psi^2$. The solution is
\begin{equation}
\psi^2(r)=\frac{\omega}{1+\omega}+\frac{c}{r}.
\end{equation}
Comparing Eqs. (\ref{E:line1}) and
(\ref{E:line2}), observe that
\begin{equation}
b(r)=r(1-e^{-\lambda(r)})=r[1-\psi^2(r)]
\end{equation}
by Eq. (\ref{E:grr}). The condition
$b(r_0)=r_0$ at the throat implies that
$\psi^2(r_0)=0$, as noted earlier. It follows
that the shape function is
\begin{equation}
b(r)=r\left(1-\frac{\omega}{1+\omega}
+\frac{1}{r}\frac{r_0\omega}{1+\omega}\right).
\end{equation}
To check the flare-out condition, we need to
find $b'(r_0)$:
\begin{equation}
\left. b'(r_0)=1-\frac{\omega}{1+\omega}+
\frac{1}{r}\frac{r_0\omega}{1+\omega}+
r\left(-\frac{r_0\omega}{r^2(1+\omega)}
\right)\right|_{r=r_0}=
1-\frac{\omega}{1+\omega}.
\end{equation}
Now observe that $b'(r_0)<1$ only if
\begin{equation*}
-\frac{\omega}{1+\omega}<0,\quad
\omega\neq -1.
\end{equation*}
We conclude that $\omega <-1$ or $\omega >0$,
or, more completely,
\begin{equation}\label{E:omega}
\omega <-1\quad \text{or}\quad 0<\omega <1.
\end{equation}
According to Ref. \cite{MT88}, the flare-out
condition is equivalent to $p_r+\rho <0$, i.e.,
the NEC is violated for the null vector
$(1,1,0,0)$. Recalling the EoS $p_t=\omega\rho$,
Eq. (\ref{E:omega}) with $\omega <-1$ implies
that the NEC is violated for any null vector
of the form
\begin{equation*}
(1,0,a,1-a),\quad 0\le a\le 1.
\end{equation*}
For $0<\omega <1$, the NEC is actually met.
It is readily checked, however, that
\begin{equation*}
\left. p_r+\rho\right|_{r=r_0}=
-\frac{1}{8\pi}\frac{1}{r_0^2}
\frac{\omega}{1+\omega}<0
\end{equation*}
whenever
\[
\omega <-1\quad\text{or}\quad 0<\omega <1,
\]
as before.
It is interesting to note that
\[
p_r(r_0)=-\frac{1}{8\pi r_0^2},
\]
which is independent of $\omega$ and coincides
with $p_r(r_0)$ in Ref. \cite{MT88}.
\section{The shadow universe}
Even though our starting point was the lateral
pressure in the EoS $p_t=\omega\rho$, it soon
became apparent that $p_r+\rho <0$ whenever
$\omega <-1$ or $0<\omega <1$. In a cosmological
setting, we normally associate $\omega <-1$
with phantom dark energy and $0<\omega <1$ with
dark matter or normal (baryonic) matter. It is
well known that phantom energy can support
traversable wormholes since the NEC is violated
\cite{fL05, sS05}. So in the present situation,
the case $0<\omega <1$ is by far the more
interesting.
For the EoS $p_t=\omega\rho$, the conditions
$\omega <-1$ and $0<\omega <1$ cannot be
met simultaneously. Well outside the
galactic halo, however, the case $\omega <-1$
would apply since on large scales, the Universe
is undergoing an accelerated expansion. Inside
the galactic halo, on the other hand, dark
matter dominates since the galaxies themselves
do not participate in the expansion, being bound
gravitationally by predominantly dark matter.
So in the halo region, conformal-symmetry
wormholes can be supported by dark matter.
These wormholes would be part of what is often
called the shadow universe, normally invisible
to us. Detection of wormholes may nevertheless
be possible by means of gravitational lensing
\cite{VE00, pK14, pK15b}.
For completeness let us note that it is shown
in Ref. \cite{RKRI} that the Navarro-Frenk-White
model can be used to show that dark matter can
support traversable wormholes provided that we
confine ourselves to the outer regions of the
galactic halo. This restriction does not apply
to the present study.
\section{Conclusion}
This paper discusses a new wormhole solution
by making the common and presumably
reasonable assumptions that a homogeneously
distributed cosmic fluid can be adapted to
an anisotropic spacetime such as a wormhole
and that this spacetime admits a one-parameter
group of conformal motions. The pressure $p$
in the perfect-fluid EoS $p=\omega\rho$
becomes the lateral pressure $p_t$ instead
of the radial pressure $p_r$, as assumed in
previous studies. Since $p_t=\omega\rho$ in
this paper, $p_r$ is determined from the
Einstein field equations.
After showing that the assumption of
conformal symmetry implies that any
Morris-Thorne wormhole must be anisotropic,
the EoS $p_t=\omega\rho$ is used to
determine the conformal factor $\psi(r)$
and hence the shape function $b=b(r)$.
Neither could exist unless $\omega <-1$
or $0<\omega <1$. In both cases,
$p_r+\rho <0$, so that the NEC is
violated.
Well outside the galactic halo region
the EoS $p=\omega\rho$, $\omega <-1$,
is normally interpreted as phantom dark
energy, which is known to support
traversable wormholes. So the more
interesting case is $0<\omega <1$,
representing primarily dark matter in
the galactic halo region. This shows
that given the above conditions, dark
matter can support traversable wormholes.
|
\section{Introduction}
A number of wavefunction- and Green's function-based numerical techniques have been developed to address the problem of strongly interacting electrons in lattice models for condensed matter systems.~\cite{FehskeBook2008} Among the most widely and straightforwardly applied methods is small cluster exact diagonalization, which is particularly well suited to problems with strong interactions where the important physics remains local or relatively short-ranged. One explicitly constructs a many-body Hamiltonian from the full Hilbert space that consists of all allowed multi-particle configurations of single-particle states for the full lattice problem.~\cite{DagottoRMP} The usefulness of this technique is limited by the exponential growth of the Hilbert space dimension $D$ with increasing size of the clusters.
Full diagonalization of the Hamiltonian -- evaluating all eigenvalues and eigenvectors -- remains impractical for all but the smallest problems with a typical computational complexity of $O(D^{3})$. However, accurate information about the ground (lowest energy) state and several low lying excited states can be sufficient for zero, or low temperature properties of the model. To that end, iterative Krylov subspace methods, such as Lanczos or Arnoldi~\cite{golub2012matrix}, can be employed to reduce the computational burden.
Dynamical properties can be evaluated from the ground (and excited) state eigenfunctions using secondary numerical methods~\cite{jia2015using, wang2014real}, such as the continued fraction expansion~\cite{CFE} or bi-conjugate gradient stabilized techniques~\cite{BiCGS}.
A reduction of the computational complexity for these methods also comes from the fact that the Hamiltonian matrix is typically sparse, with a polynomial number of non-zero elements (usually only several hundred non-zero elements) per row or column, typically orders of magnitude smaller than the Hilbert space dimension $D$.
However, exponential growth of the Hilbert space means that the Hamiltonian matrix or even the wavefunctions might be too large to store in the memory available on a single compute node. This is certainly the case for the largest and most challenging problems for exact diagonalization. The nature of modern parallel computing environments with a large number of lightweight cores per node and limited memory per core (or per node) makes the issue of efficient memory utilization in exact diagonalization one of the most significant bottlenecks to performance and scalability.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\columnwidth]{Checkboard.pdf}
\caption{Checkerboard decomposition of the Hamiltonian matrix. The colors represent a range of indices in the many-body basis, for both the sparse matrix and wavefunctions, distributed between different compute nodes.
\label{fig:checkerboard}
\end{figure}
One solution to this problem has been a ``checkerboard'' decomposition of the Hamiltonian matrix as illustrated in Fig.~\ref{fig:checkerboard}. This decomposition distributes a different partition of the matrix and wavefunctions (blocks with the same color) to different processors, or compute nodes, to balance the memory distribution. When combined with graph partitioning and matrix reordering algorithms, such as those in the PARMETIS library~\cite{METIS}, this scheme maximizes data locality, reducing parallel communications.
The remaining performance and
memory bottleneck comes from the construction of the matrix from the many-body basis (Hilbert space) and the one-to-one mapping between each state and the appropriate index. A traditional binary search algorithm may be employed when the states are properly ordered; however, this method usually requires a hash table, which encodes the index-state map, resident in memory for each processor, or compute node, to minimize communications overhead. An ideal hashing scheme would allow for the evaluation of the index-state map, in both the forward and reverse directions, without any memory overhead.
In this paper, we effectively devise such a perfect hashing algorithm, \emph{Paradeisos}, to replace conventional search with a direct mapping. This eliminates the need to store a table for the mapping between the vector-matrix index and the many-body basis. The method becomes particularly advantageous for large problem sizes and in parallel computing environments where the ``checkerboard'' decomposition must be used for distributed matrix-vector storage.
We describe \emph{Paradeisos} in the framework of a single-band, or single-orbital, model for two species of interacting fermions (``spin-up'' $\uparrow$ and ``spin-down'' $\downarrow$) on a lattice. Extension of this algorithm to multi-orbital models, or models for restricted bosonic systems, can be accomplished with straightforward modifications.
The paper is organized as follows: first, we introduce the algorithm and the forward and backward mapping in the general case; second, we discuss the implementation of the algorithm in symmetry projected subspaces of the full Hilbert space; and finally, we explore the performance of the algorithm and its scaling with the number of parallel processes.
\section{Hilbert space and model Hamiltonian}
\label{sec:basis_construction}
\subsection{Many-body basis}
Start with the single-site fermionic problem.
There are four configurations due to the Pauli exclusion principle: $\lvert 0 \rangle$, $\lvert\uparrow\rangle$, $\lvert\downarrow\rangle$, $\lvert\uparrow \downarrow \rangle$ (or in occupation and spin direct product basis $\lvert 0 \rangle_\uparrow \otimes \lvert 0 \rangle_\downarrow $, $\lvert 1 \rangle_\uparrow \otimes \lvert 0 \rangle_\downarrow $, $\lvert 0 \rangle_\uparrow \otimes \lvert 1 \rangle_\downarrow $, $\lvert 1 \rangle_\uparrow \otimes \lvert 1 \rangle_\downarrow$). For a multi-site problem,
one typically works in the canonical ensemble with fixed total electron number. In general, for an $N$-site cluster with $n_\uparrow$ spin-up and $n_\downarrow$ spin-down electrons, the Hilbert space dimension
$D = C_N^{n_\uparrow} \cdot C_N^{n_\downarrow}$, where $C_m^n$ is the binomial coefficient.
Any basis element of the Hilbert space, represented in terms of the occupation of single-particle states, can be constructed from the vacuum $\vert 0 \rangle$ by repeated application of creation operators $c^\dagger_{i\sigma}$, where $\sigma \in \{\uparrow, \downarrow\}$ and $i \in \{0, 1, \ldots, N-1 \}$ indexes lattice positions, such that
\[\lvert \varphi \rangle = \underbrace{\ldots c^\dagger_{i\uparrow} \ldots}_{n_\uparrow} \underbrace{\ldots c^\dagger_{j\downarrow} \ldots}_{n_\downarrow} \lvert 0 \rangle. \]
Note the ``normal ordering'' convention with spin-down operators on the right, and we take the increasing site index right-to-left. The convention ensures proper
antisymmetrization for fermion exchange. As an illustration to which we will return throughout the discussion of the algorithm, consider a simple 4-site lattice model with two spin-up ($n_\uparrow = 2$) and three spin-down ($n_\downarrow = 3$) electrons. In this case, $D=24$ and one particular element of the Hilbert space basis
\[c^\dagger_{2\uparrow} c^\dagger_{0\uparrow} c^\dagger_{3\downarrow} c^\dagger_{1\downarrow} c^\dagger_{0\downarrow} \lvert 0\rangle = (0 \uparrow 0 \uparrow) \otimes (\downarrow 0 \downarrow \downarrow).\]
For practical computations, all basis elements can be enumerated by noting the correspondence of the occupation and spin direct product representation with bit sequences for unique integer values.
Turn again to the example of one particular basis element for the 4-site lattice model.
\[(0 \uparrow 0 \uparrow) \otimes (\downarrow 0 \downarrow \downarrow) \to 0101_2 \otimes 1011_2,\]
where each occurrence of $\uparrow$ or $\downarrow$ has been replace by a $1$ (occupied) with $0$ otherwise. The subscript $2$ on each sequence denotes binary notation for clarity.
A lexicographical binary representation can be constructed separately for every spin-up and spin-down element comprising the Hilbert space basis. For the 4-site example, results are presented in the following table.
\begin{center}
\begin{tabular}{c | l || c | l}
index & $\uparrow$ element & index & $\downarrow$ element \\
\hline
0 & $0011_2 = 3$ & 0 & $0111_2 = 7$ \\
1 & $0101_2 = 5$ & 1 & $1011_2 = 11$ \\
2 & $0110_2 = 6$ & 2 & $1101_2 = 13$ \\
3 & $1001_2 = 9$ & 3 & $1110_2 = 14$ \\
4 & $1010_2 = 10$ & & \\
5 & $1100_2 = 12$ & & \\
\end{tabular}
\end{center}
In general, the lexicographical next state,
advancing the index $p \rightarrow p+1$,
can be determined using just a few bit operations ~\cite{FermiFab2011, AndersonWebpage}. A tensor product of the spin-up and spin-down contributions defines the complete basis. This tensor product structure typically breaksdown within restricted symmetry subspaces, which we will address in a subsequent section.
\subsection{Many-body Hamiltonian}
Although the algorithm does not depend on the details of the many-body Hamiltonian, for simplicity, we consider the single-band Hubbard model. Note that the symmetry of the Hamiltonian affects the selection of the subspace.
The single-band Hubbard Hamiltonian $H$ includes both kinetic and on-site Coulomb interaction terms, written conventionally as
\begin{eqnarray}
\label{eq:hubbard}
H &=& H_K + H_I \nonumber\\
&=& -\sum_{ij, \sigma}\big( t_{ij} c^\dagger_{i\sigma} c_{j\sigma} + h.c. \big) + U \sum_{i} n_{i\uparrow} n_{i\downarrow},
\end{eqnarray}
where $n_{i\sigma} = c^\dagger_{i\sigma} c_{i\sigma}$ is the standard number operator. The kinetic terms include ``hopping'' of electrons on the lattice with an energy $t_{ij}$, typically restricted to nearest or next-nearest neighboring sites, and $U$ parameterizes the on-site interaction strength.
$H$ can be viewed as a one-to-many map of the Hilbert space into itself. Consider first the interaction terms. In a representation based on real-space occupation on the lattice, the interaction depends on the ``double occupancy'' -- how many sites have occupation by both $\uparrow$ and $\downarrow$ fermions. Comprised solely of number operators, this term represents a map of each many-body Hilbert space basis element to itself with a prefactor. In bit operations, the ``double occupancy'' (and by extension the Hamiltonian matrix elements) can be computed from the bitwise {\bf \texttt{AND}} between the $\uparrow$ and $\downarrow$ parts of the direct product.
The kinetic terms represent a more complicated map. The movement of fermions on the lattice, or ``hopping'', mixes basis elements through a change in the fermion occupation. One must determine not only the weights for the mapping (Hamiltonian matrix elements), but also identify the index (or indices) within the basis to which $H$ maps each element. Return to our example of the 4-site lattice; and consider a kinetic term with only nearest-neighbor hopping $t$ and sites linked together cyclically, as in a one-dimensional loop ($0 \leftrightarrow 1 \leftrightarrow 2 \leftrightarrow 3 \leftrightarrow 0$). For simplicity, we address only the spin-up kinetic term acting on our test element $0101_2 \otimes 1011_2$, where the first bit sequence of the product corresponds to spin-up. In this case,
\begin{align}
H_K^{\uparrow} & (0101_2 \otimes 1011_2) = \nonumber\\
-t &\big( 1001_2 + 0011_2 + 0110_2 - 1100_2 \big) \otimes 1011_2,\nonumber
\end{align}
where the minus sign in front of $1100_2$ results from the normal ordering convention and the antisymmetry of fermion exchange. One can immediately read-off the index for each state from the table created for this example; however, ideally one would like a more automated and less brute force method for determining the new indices under this mapping for a state-of-art size of quantum many-body problem.
\subsection{Index determination}
One may employ conventional binary search algorithm~\cite{BinarySearch}, to determine the indices to which a function such as $H$ will map each basis element. The binary search algorithm has time complexity $\mathcal{O}(\ln D)$. Since $D=C_N^{n_\uparrow} \cdot C_N^{n_\downarrow} < 4^N$, an upper bound for the time complexity of binary search is $\mathcal{O}(\ln 4^N)$ or $\mathcal{O}(N)$.
However, one must typically store the enumerated bit sequences for all basis elements in memory to reduce the communications overhead.
In the following sections we describe \emph{Paradeisos}, a direct ``forward mapping'' from a bit sequence to its basis index, mitigating the need for search, and by extension storage. Crucially, this algorithm has the same time complexity as traditionally binary search methods.
When combined with advanced matrix decomposition techniques for parallel computing environments, \emph{Paradeisos} has the potential to significantly reduce aggregate memory consumption and improve performance (see the section on ``Numerical Benchmarks").
\section{Mapping functions}
\label{sec:mapping_functions}
The key feature of \emph{Paradeisos} is a mapping from a bit sequence to an index. As opposed to binary search, which identifies an index from an ordered list through an iterative series of bisections, \emph{Paradeisos} determines the (absolute) index of an arbitrary bit sequence of the list by mathematically evaluating the total number of different configurations between it and a known pattern.
First, we define the distance between two elements $\varphi_1$ and $\varphi_2$ by the difference of their indices $\mathrm{dist}\{\varphi_1,\varphi_2\} = \mathrm{idx}(\varphi_2)-\mathrm{idx}(\varphi_1)$.
However, directly evaluating this distance is non-trivial. It can be straightforward in the special case of \emph{simple neighbors
, where two bit sequences differ by exchange of a single 1-bit across consecutive 0-bits. As an example consider two elements, each with 4 non-zero \emph{least significant bits} (LSBs) up to the position of this exchange. The first
\begin{equation*}
\begin{split}
|\varphi_1\rangle := \,&\ldots \overset{j_2}{0}\,0\ 0\underbrace{\overset{j_1}{{\bf {\bf 1}}}\,0\ 0\ 1\ 0\ 1\, \overset{i=0}{1_2}}_{n = 4} \\
\end{split}
\end{equation*}
has an index $\mathrm{idx}(\varphi_1)=p$, and the second
\begin{equation*}
\begin{split}
|\varphi_2\rangle :=\,&\dots \underbrace{\overset{j_2}{{\bf 1}}\,0\ 0\,\overset{j_1}{0}\ 0\ 0\ 1\ 0\ 1\, \overset{i=0}{1_2}}_{n = 4}
\end{split}
\end{equation*}
has an index $\mathrm{idx}(\varphi_2)=q$. The distance between these two bit sequences depends on the number of configurations of the $n$ bits between them, which is given simply by
\begin{equation}
\label{eq:index_free_hopping}
\mathrm{dist}\{\varphi_1,\varphi_2\} = q - p = C_{j_2}^n - C_{j_1}^n,
\end{equation}
with a convention that the bit position starts from $i=0$ and that $C_m^n = 0$ for $m < n$.
Note that Eq.~\eqref{eq:index_free_hopping} is independent of the particular bit configuration to the right of $j_1$, nor does it depend upon the configuration to the left of $j_2$ if the
total number of 1-bits were greater than $n$.
One can construct the mapping between a bit sequence and its corresponding index, via successive exchanges from a given sequence with a known index, by taking advantage of the distance between simple neighbors. By construction, index $0$ for the first element in a list of $N$-bit numbers with $n$ non-zero bits ($n < N$) belongs to the bit sequence with the $n$ LSBs set to $1$:
\begin{equation*}
\begin{split}
\lvert \varphi_0\rangle :=& \underbrace{\overset{N-1}{0} \ldots\ \overset{n}{0}\ \underbrace{\overset{n-1}{1}\ \ldots \overset{i = 0}{1_2}}_{n}}_{N}, \\
\mathrm{idx}(\varphi_0) &= 0
\end{split}
\end{equation*}
For any other $N$-bit sequence $\lvert \varphi \rangle$ in the ordered list, its $n$ 1-bits will occupy positions $j_{n-1} > j_{n-2} > \ldots > j_{1} > j_{0}$. Starting from $\lvert \varphi_0 \rangle$, one can construct a sequence of $n$ simple neighbors $\{\lvert \varphi_0^{(n)} \rangle, \dots, \lvert \varphi_0^{(1)} \rangle = \lvert \varphi \rangle\}$ by sequential exchange of the $m^{\mathrm{th}}$ LSB, $\{m=n, \ldots, 1\}$. Thus,
\begin{equation*}
\begin{split}
\lvert \varphi_0\rangle :=&\ \overset{N-1}{0} \ldots\ \overset{n}{0}\ \overset{n-1}{1}\ \ldots \overset{i = 0}{1_2} \\
\lvert \varphi_0^{(n)}\rangle :=&\ \ldots \overset{j_{n-1}}{1} \ldots\ \overset{n-1}{0}\ \overset{n-2}{1} \ldots \overset{i = 0}{1_2} \\
\lvert \varphi_0^{(n-1)}\rangle :=&\ \ldots \overset{j_{n-1}}{1} \ldots\ \overset{j_{n-2}}{1} \ldots \overset{n-2}{0}\ \overset{n-3}{1} \ldots \overset{i = 0}{1_2} \\
\vdots& \\
\lvert \varphi_0^{(2)}\rangle :=&\ \ldots \overset{j_{n-1}}{1} \ldots\ \overset{j_{n-2}}{1} \ldots \overset{j_{1}}{1} \ldots \overset{1}{0}\ \overset{i = 0}{1_2} \\
\lvert \varphi \rangle = \lvert \varphi_0^{(1)}\rangle :=&\ \ldots \overset{j_{n-1}}{1} \ldots\ \overset{j_{n-2}}{1} \ldots \overset{j_{1}}{1} \ldots \overset{j_{0}}{1} \dots_2,
\end{split}
\end{equation*}
and one can now read-off the index simply as the accumulated distance between simple neighbors
\begin{equation}\label{eq:idx}
\mathrm{idx}(\varphi) = C_{j_0}^1 + C_{j_1}^2 + \dots + C_{j_{n-2}}^{n-1} + C_{j_{n-1}}^n.
\end{equation}
\noindent
\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} forward map}\\
Goal: For $n$ electrons and $N$ sites,
map $\varphi \mapsto \textrm{idx}(\varphi)$\\
\hline
\textbf{function} {\it forward\_map}\,($\varphi$)\\
\indent \textbf{declare} $\mathrm{idx}$ = 0, $m$ = 0.\\
\indent \textbf{for} $i = 0$ to $N-1$:\\
\indent \indent \textbf{if} $i$-th bit in $\varphi$ is \texttt{TRUE}:\\
\indent \indent \indent $m$\ $\leftarrow$\ $m + 1$\\
\indent \indent \indent $\mathrm{idx}$\ $\leftarrow$\ $\mathrm{idx} + C_{i}^m$\\
\indent \indent \textbf{end}\\
\indent \textbf{end}\\
\indent \textbf{return} $\mathrm{idx}$ \\
\hline
\end{tabular}
\newline
\noindent For efficiency, the binomial coefficients $C_i^m$ for $i, m \in \{0, 1, \dots, N-2, N-1\}$ can be precomputed recursively $C_i^m = C_{i-1}^m + C_{i-1}^{m-1}$, and stored in a lookup-table. The time complexity for {\it Paradeisos} is O($N$), comparable to traditional binary search.
As an example, return to the problem of 2 spin-up and 3 spin-down electrons on a 4 site lattice and, for simplicity, consider only element $1100_2$ in the spin-up part of the Hilbert space basis.
After exchange of the highlighted (boldface/underline) bits,
\begin{equation*}
\begin{split}
\lvert \varphi_0 \rangle :=&\ 00\underline{\bf 1}1_2, \qquad \mathrm{idx(\varphi_{0})} = 0 \\
\downarrow& \\
\lvert \varphi_0^{(2)} \rangle :=&\ \underline{\bf 1}001_2, \qquad \mathrm{idx(\varphi_{0}^{(2)})} = C_3^2 = 3, \\
\end{split}
\end{equation*}
and
\begin{equation*}
\begin{split}
\lvert \varphi_0^{(2)} \rangle :=&\ 100\underline{\bf 1}_2, \qquad \mathrm{idx(\varphi_{0}^{(2)})} = C_3^2 = 3 \\
\downarrow& \\
\lvert \varphi \rangle = \lvert \varphi_{0}^{(1)} \rangle :=&\ 1\underline{\bf 1}00_2, \qquad \mathrm{idx(\varphi)} = C_3^2 + C_2^1 = 5.
\end{split}
\end{equation*}
One can refer back to the table to verify the result.
The index in the full Hilbert space, composed from tensor products between basis elements in the two spin sectors, can be computed from the known dimension of the spin-resolved Hilbert spaces. Given our convention with spin-down configurations occupying the first $N$ LSBs, $\mathrm{idx}(\varphi) = C_N^{n_{\downarrow}}\cdot\mathrm{idx}_{\uparrow}(\varphi_{\uparrow}) + \mathrm{idx}_{\downarrow}(\varphi_{\downarrow})$, where $\uparrow/\downarrow$ subscripts denote the appropriate spin-restricted subspace.
To be complete, \emph{Paradeisos} requires a ``backward mapping'' for determining a bit sequence from a known index. This map follows from the accumulated distance between simple neighbors to determine the position of non-zero bits. Pseudocode summarizes the procedure.
\newline\\
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} backward map}\\
Goal: For $n$ electrons and $N$ sites, map $\mathrm{idx}(\varphi) \mapsto \varphi$\\
\hline
\textbf{function} {\it backward\_map}\,($\mathrm{idx}$; $n$, $N$)\\
\indent \textbf{declare} $\varphi$ = $00 \ldots 00_2$, $m$ = $n$, $p=\mathrm{idx}$\\
\indent \textbf{for} $i = N-1$ to $0$:\\
\indent \indent \textbf{if} $p \geq C_i^m$:\\
\indent \indent\indent set $\varphi$'s $i$-th bit to \texttt{TRUE}\\
\indent \indent\indent $p$\ $\leftarrow$\ $p - C_i^m$\\
\indent \indent\indent $m$\ $\leftarrow$\ $m - 1$\\
\indent \indent \textbf{end}\\
\indent \textbf{end}\\
\indent \textbf{return} $\varphi$\\
\hline
\end{tabular}\\
\newline
\noindent One can verify this procedure in the previous example by inspection.
\section{Mapping within symmetry subspaces}\label{sec:symmetries}
One typically uses symmetries to reduce the dimension of the effective Hilbert space. These may include $\mathrm{SU(2)}$ spin symmetry~\cite{FehskeBook2008}, point group symmetries such as translational, rotational~\cite{FanoSym1} and inversion symmetry, or time reversal symmetry that partition the original Hilbert space and block diagonalize the Hamiltonian.
In the following sections, we illustrate the implementation of \emph{Paradeisos} within both translation and inversion symmetry restricted subspaces.
\subsection{Translational symmetry}
For regular lattice models defined on small clusters with periodic boundary conditions, translation symmetry can lead to a reduction in the Hilbert space dimension by a factor $\sim N$, the total number of lattice sites or unit cells. Translational invariance typically prompts one to choose a momentum space representation via a discrete fourier transform of the operators $c^{\dagger}_{\bf k}=\frac{1}{\sqrt{N}}\sum_{j} e^{i\textbf{k}\cdot{\bf r}_{j}c^{\dagger}_{j}}$ (similarly for the hermitian conjugate, annihilation operator $c_{\bf k}$). Transforming the single-band Hubbard model of Eq.~\ref{eq:hubbard} to this momentum space basis leads to
\begin{equation}
\label{eq:kspacehubbard}
H = \sum_{ \textbf{k}, \sigma}\varepsilon_{\bf k} c^\dagger_{\bf k\sigma} c_{\bf k\sigma} + \frac{U}{N}\sum_{\bf k,k',q}c^{\dagger}_{\bf k+q\uparrow}
c^{\dagger}_{\bf k'-q\downarrow}c_{\bf k'\downarrow}c_{\bf k\uparrow}
\end{equation}
where $\varepsilon_{\bf k}$ is the kinetic energy in momentum space, typically referred to as the bare band structure. For the single-band Hubbard model on a square lattice with only nearest, $t$, and next-nearest neighbor, $t'$, hopping terms, $\varepsilon_{\bf k} = -2t(\cos k_x +\cos k_y) -4t^\prime \cos k_x\cos k_y$.
The Pauli exclusion principle also applies in momentum space; so given a normal ordering of the discrete single-particle momenta $\{{\bf k}_0,\ldots,{\bf k}_{N-1}\}$, bit sequences relate to fermion occupation of single-particle momentum states. Note that the full Hilbert space dimension $D$ in the momentum space representation remains the same as that in the real space representation; however, inspection of the two terms in the Hamiltonian reveals an effective dimensional reduction. The kinetic term is now diagonal in momentum space, whereas the interaction term scatters fermions of momenta ${\bf k}$ and ${\bf k'}$ to new momenta ${\bf k+q}$ and ${\bf k'-q}$, while leaving the total momentum of the many-body state ${\bf K}$ unchanged. Thus, ${\bf K}$ can be used to partition the basis elements of the Hilbert space with $D_{\bf K} < D$
Restriction to a translational subspace is equivalent to fixing the total momentum $\mathbf{K}$, modulo the first BZ. To apply the \emph{Paradeisos} mapping functions in the restricted subspace, one must replace the binomial coefficients by an extended combinatorial $C_{i}^m(\textbf{k})$ following the recursion
\begin{equation}
C_i^m(\textbf{k}) = C_{i-1}^{m-1}(\textbf{k}-\textbf{k}_i) + C_{i-1}^m(\textbf{k}) ,
\end{equation}
where $C_{i}^m(\textbf{k})$ counts the number of ways to arrange $m$ electrons in a state with $i$-bits and a fixed total momentum $\textbf{k}$. The initial conditions for the recursion are
\begin{displaymath}
C_i^0(\textbf{k}) = \left\{ \begin{array}{l}
1 \textrm{, if $\textbf{k}=\textbf{0}\ (\mathbf{\Gamma}-\textrm{point})$}\\
0 \textrm{, otherwise }
\end{array} \right.
\end{displaymath}
and
\begin{displaymath}
C_i^1(\textbf{k}) = \left\{ \begin{array}{l}
1 \textrm{, if $\textbf{k}=\textbf{k}_i$}\\
0 \textrm{, otherwise}
\end{array} \right.
\end{displaymath}
Following these rules and Eq.~\ref{eq:idx}, an $n$ fermion state with occupation $j_0 < j_1 < \cdots < j_{n-2} < j_{n-1}$ and corresponding total momentum $\mathbf{K} = \textbf{k}_{j_0} + \cdots + \textbf{k}_{j_{n-1}}$ has an index (within the momentum subspace $D_{\bf K}$)
\begin{eqnarray}
\label{eq:basis_index_subspace}
\mathrm{idx}(\varphi) &=& C_{j_0}^1(\textbf{k}_{j_0}) + C_{j_1}^2(\textbf{k}_{j_0}+\textbf{k}_{j_1}) + \dots \nonumber\\
&&\ldots + C_{j_{n-2}}^{n-1}(\mathbf{K}-\mathbf{k}_{j_{n-1}}) + C_{j_{n-1}}^n(\mathbf{K}).
\end{eqnarray}
One can precompute the coefficients $C_{i}^m(\textbf{k})$ using dynamical programming and store them in a lookup-table. A pseudocode description of this forward mapping in the restricted subspace would be
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} k-space forward map}\\
Goal: For $n$ electrons and $N$ momenta, map \\
$\varphi \mapsto$ $\textrm{idx}(\varphi)$ \\
\hline
\textbf{function} {\it forward\_map\_k}\,($\varphi$)\\
\indent \textbf{declare} $\textrm{idx}$ = 0, $m$ = 0, \textbf{k}=0.\\
\indent \textbf{for} $i = 0$ to $N-1$:\\
\indent \indent \textbf{if} $i$-th bit in $\varphi$ is \texttt{TRUE}:\\
\indent \indent \indent $m$\ $\leftarrow$\ $m + 1$\\
\indent \indent \indent $\textbf{k}$\ $\leftarrow$\ mod($\textbf{k} + \textbf{k}_i$,BZ)\\
\indent \indent \indent $\textrm{idx}$\ $\leftarrow$\ $\textrm{idx} + C_{i}^m(\textbf{k})$\\
\indent \indent \textbf{end}\\
\indent \textbf{end}\\
\indent \textbf{return} $\textrm{idx}$, $\textbf{k}$\\
\hline
\end{tabular}
\newline\\
Following a similar procedure as in the full Hilbert space, the momentum restricted subspace backward mapping has a pseudocode
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} k-space backward map}\\
Goal: For $n$ electrons, $N$ momenta, \\
and total momentum $\textbf{K}$, map $\textrm{idx}(\varphi) \mapsto \varphi$\\
\hline
\textbf{function} {\it backward\_map\_k}\,($\textrm{idx}$; $n$, $N$, $\textbf{K}$)\\
\indent \textbf{declare} $\varphi$ = $00 \ldots 00_2$, $\textbf{k}$ = $\textbf{K}$, $m$ = $n$, $p=\textrm{idx}$\\
\indent \textbf{for} $i = N-1$ to $0$:\\
\indent \indent \textbf{if} $p \geq C_i^m(\textbf{k})$:\\
\indent \indent\indent set $\varphi$'s $i$-th bit to \texttt{TRUE}\\
\indent \indent\indent $p$\ $\leftarrow$\ $p - C_i^m(\textbf{k})$\\
\indent \indent\indent $m$\ $\leftarrow$\ $m - 1$\\
\indent \indent \indent $\textbf{k}$\ $\leftarrow$\ mod($\textbf{k}-\textbf{k}_i$,BZ)\\
\indent \indent \textbf{end}\\
\indent \textbf{end}\\
\indent \textbf{return} $\varphi$\\
\hline
\end{tabular}\\
The spin-full restricted momentum subspace can no longer be regarded as a direct product space between $\uparrow$ and $\downarrow$ momentum subspaces. Instead, the index of a basis element depends explicitly on both the $\uparrow$ and $\downarrow$ configurations. Noting the normal ordering convention with the $\downarrow$ bits followed by those for $\uparrow$ right-to-left, the full forward mapping procedure reduces to first performing a forward mapping for the $\downarrow$ portion of the basis element. The pseudocode for treating the remainder of the forward mapping including the $\uparrow$ portion of a basis element can be written as
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} k-space forward map (spin-full)}\\
Goal: For $n_\uparrow + n_\downarrow$ electrons and $N$ momenta, map \\%sites with total \\momentum $\textbf{K}$,
$\varphi = \varphi_\uparrow\otimes\varphi_\downarrow \mapsto \textrm{idx}(\varphi)$\\
\hline
\textbf{function} {\it forward\_map\_k\_spin}\,($\varphi_\uparrow$, $\varphi_\downarrow$) \\%; $\textbf{K}$)\\
\indent \textbf{declare} $\textrm{idx}$ = 0, $m$ = 0, \textbf{k}=0.\\
\indent \textbf{call} {\it forward\_map\_k}\,($\varphi_{\downarrow}$): $\textrm{idx}_{\downarrow}$, $\textbf{K}_{\downarrow}$; $\textbf{k}$\ $\leftarrow$\ $\textbf{K}_{\downarrow}$.\\
\indent \textbf{for} $i = 0$ to $N-1$:\\
\indent \indent \textbf{if} $i$-th bit in $\varphi_\uparrow$ is \texttt{TRUE}:\\
\indent \indent \indent $m$\ $\leftarrow$\ $m + 1$\\
\indent \indent \indent $\textbf{k}$\ $\leftarrow$\ mod($\textbf{k} + \textbf{k}_i$,BZ)\\
\indent \indent \indent $\textrm{idx}$\ $\leftarrow$\ $\textrm{idx} + \widetilde{C}_{i}^m(\textbf{k}; n_\downarrow)$\\
\indent \indent \textbf{end if}\\
\indent \textbf{end}\\
\indent \textbf{return} $\textrm{idx} + \textrm{idx}_{\downarrow}$ \\
\hline
\end{tabular}\\
\noindent Here one must work with an additional modification of the combinatorial function, which accounts for the spin-down combinations. This modified version takes the form
\begin{equation}
\widetilde{C}_i^m(\textbf{k}; n_\downarrow) = \sum_{\bf k_\downarrow}C_i^m(\textbf{k} - \textbf{k}_\downarrow)C_N^{n_\downarrow}(\textbf{k}_\downarrow).
\end{equation}
Again, these quantities may be precalculated and stored in a look-up table. A similar modification applies for the spin-full backward mapping whose pseudocode becomes
\newline\\
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} k-space backward map (spin-full)}\\
Goal: For $n_\uparrow + n_\downarrow$ electrons, $N$ momenta, \\
and total momentum $\textbf{K}$, map $\textrm{idx}(\varphi) \mapsto \varphi_\uparrow\otimes\varphi_\downarrow$\\
\hline
\textbf{function} {\it backward\_map\_k\_spin}\,($\textrm{idx}$; $n_\uparrow$, $n_\downarrow$, $N$, $\textbf{K}$)\\
\indent \textbf{declare} $\varphi_\uparrow$ = $00 \ldots 00_2$, $\textbf{k}$ = $\textbf{K}$, $m$ = $n_\uparrow$, $p=\textrm{idx}$\\
\indent \textbf{for} $i = N-1$ to $0$:\\
\indent \indent \textbf{if} $p \geq \widetilde{C}_{i}^m(\textbf{k}; n_\downarrow)$:\\
\indent \indent\indent set $\varphi_\uparrow$'s $i$-th bit to \texttt{TRUE}\\
\indent \indent\indent $p$\ $\leftarrow$\ $p - \widetilde{C}_{i}^m(\textbf{k}; n_\downarrow)$\\
\indent \indent\indent $m$\ $\leftarrow$\ $m - 1$\\
\indent \indent \indent $\textbf{k}$\ $\leftarrow$\ mod($\textbf{k}-\textbf{k}_i$,BZ)\\
\indent \indent \textbf{end if}\\
\indent \textbf{end}\\
\indent \textbf{call} {\it backward\_map\_k}\,($p$; $n_\downarrow$, $\textbf{k}$): $\varphi_\downarrow$ \\
\indent \textbf{return} $\varphi = \varphi_\uparrow\otimes\varphi_\downarrow$\\
\hline
\end{tabular}\\
\subsection{Inversion symmetry}
Inversion symmetry ($\mathcal{F}$) implies that a transformation ${\bf r} \to -{\bf r}$ (or ${\bf k} \to -{\bf k}$) leaves the Hamiltonian unchanged.
To simplify the presentation and without loss of generality, we confine our discussion to one-dimension in real space for a lattice (chain) with $N$ sites indexed sequentially $0, \dots ,N-1$. Inversion acting on the creation and annihilation operators appearing in the Hamiltonian returns $\mathcal{F}\left(c^{\dagger}_i\right) \to c^{\dagger}_{N-1-i}$ and $\mathcal{F}\left(c_i\right) \to c_{N-1-i}$.
Since $\mathcal{F}^2 = I$, $\mathcal{F}$ has eigenvalues $\pm 1$, which decomposes the Hilbert space into two orthogonal subspaces, corresponding to symmetric and antisymmetric combinations of the basis elements.
Note that certain basis elements may already be symmetric in $\mathcal{F}$.
Based on our conventions, the goal will be to compute the index of a given bit sequence representing the \emph{symmetric} subspace (a similar procedure can be employed for the antisymmetric subspace).
For $n$ electrons on $N$ sites, partition the representatives into three classes according to their bits in positions $N-1$ and $0$:
\begin{itemize}
\item[A:] \quad $ \lvert \varphi_{A} \rangle := \overset{N-1}{0} \dots \overset{i=0}{1_2}$
\item[B:] \quad $ \lvert \varphi_{B} \rangle := \overset{N-1}{0} \dots \overset{i=0}{0_2}$
\item[C:] \quad $ \lvert \varphi_{C} \rangle := \overset{N-1}{1} \dots \overset{i=0}{1_2}$
\end{itemize}
The symmetric states in class A are of the form
\begin{equation*}
\lvert \varphi_{A}^{\mathrm{s}} \rangle := \big( 0\ a_{N-2} \dots a_1\ 1_2 + 1\ a_1\ \dots a_{N-2}\ 0_2\big) /\sqrt{2},
\end{equation*}
which obviates the need for bit sequences of the form $\overset{N-1}{1} \dots \overset{i=0}{0_2}$ that are included by construction. Within class A, the index can be determined by standard \emph{Paradeisos} (forward mapping without symmetry). The index for elements in classes B and C can be determined by recursive partitioning for bits in positions $N-d$ and $d-1$, where $d$ is the recursion depth. For each recursion step the same classification scheme can be employed, which results in a tree with class A, B, or C nodes. Figure~\ref{fig:inversion_classification_tree} shows such a tree to illustrate this process.
\tikzstyle{level 1}=[level distance=2.0cm, sibling distance=4.5cm]
\tikzstyle{level 2}=[level distance=2.0cm, sibling distance=1.5cm]
\tikzstyle{level 3}=[level distance=2.25cm, sibling distance=0.65cm]
\tikzstyle{bag} = [text width=1em, text centered]
\tikzstyle{end} = [circle, minimum width=1pt, fill, inner sep=0pt]
\begin{figure}[ht]
\begin{tikzpicture}[grow=right]
\node[draw, circle, very thick] {$\lvert \varphi \rangle$ := ......$_2$}
child {
node[draw, circle] {1....1$_2$}
child {
node[draw, circle] {11..11$_2$}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.15em] {C}
}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.1em] {B}
}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.2em] {A}
}
edge from parent [very thin]
node[above] {C}
}
child {
node[draw, circle] {10..01$_2$}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.15em] {C}
}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.1em] {B}
}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.2em] {A}
}
edge from parent [very thin]
node[above, inner sep=0.2em] {B}
}
child {
node[draw, circle] {10..11$_2$}
edge from parent [very thin]
node[above] {A}
}
edge from parent
node[above, inner sep=1em] {C}
}
child {
node[draw, circle, very thick] {0....0$_2$}
child {
node[draw, circle, very thick] {01..10$_2$}
child {
node[end, label=right:
{}] {}
edge from parent [very thin]
node[above, inner sep=0.15em] {C}
}
child {
node[end, label=right:
{}] {}
edge from parent [very thin]
node[above, inner sep=0.1em] {B}
}
child {
node[end, label=right:
{}] {}
edge from parent [very thick]
node[above, inner sep=0.2em] {$\mathbf{A}$}
}
edge from parent [very thick]
node[above] {\bf{C}}
}
child {
node[draw, circle, very thin] {00..00$_2$}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.15em] {C}
}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.1em] {B}
}
child {
node[end, label=right:] {}
edge from parent [very thin]
node[above, inner sep=0.2em] {A}
}
edge from parent [very thin]
node[above, inner sep=0.2em] {B}
}
child {
node[draw, circle, very thin] {00..10$_2$}
edge from parent [very thin]
node[above] {A}
}
edge from parent [very thick]
node[above, inner sep=0.2em] {\bf{B}}
}
child {
node[draw, circle] {0....1$_2$}
edge from parent
node[above, inner sep=1em] {A}
};
\end{tikzpicture}
\caption{A tree up to depth $d=4$ (bits not shown for brevity) in the recursive classification of a bit sequence representing a basis element in the symmetric subspace. Branches typically terminate at an A class bit sequence from which the standard \emph{Paradeisos} forward mapping can be used to obtain the index of the state at that recursion depth. The highlight shows the path to a state such as $010110_2$, for example.}
\label{fig:inversion_classification_tree}
\end{figure}
There are a few things to note about this construction. First, each class A node is a leaf node (which contains a set of states) -- the recursion ends and \emph{Paradeisos} forward mapping can be used to determine the index at that depth. Leaf nodes can exist for class B and C, but only if the maximum recursion depth has been reached based on the number of bits $N$ in each sequence. Those leaves that end with B or C have a palindrome pattern, containing only one state. The height of the tree is $N/2 + 1$ for $N$ even or $(N+1)/2$ for $N$ odd, given that the number of bits yet to be classified at each recursion step decreases by $2$.
By convention, the indices for class A are the smallest, followed by those for class B, and lastly those for class C, at the same depth on the classification tree.
In this way, the standard order has been destroyed, and replaced by a set of continuous indices for all the representative elements in the symmetric subspace. One observes that class A contains $C_{i-2}^{m-1}$ representatives (the same number for either the symmetric or antisymmetric subspaces by construction) for $i$ ``active'' bits and $m$ non-zero bits at the current depth, as the inner bit sequence has no bearing on the A, B, or C classifications. Similarly, the number of representatives at a particular level in class B or C can be determined recursively by $S_{i}^{m}$ = $C_{i-2}^{m-1}$ + $S_{i-2}^{m}$ + $S_{i-2}^{m-2}$, where $S_{i}^{m}$ is the total number of representative elements in the symmetric subspace for $i$ total bits and $m$ non-zero bits. The three terms represent the number of elements for class A, B and C, respectively, at the next lower level. $S_{i}^{m}$ can be calculated in a bottom-up manner using dynamical programming.
In this way, the number of preceding basis elements can be calculated by a depth-first search of the tree, counting $C_{i-2}^{m-1}$ elements for a class A leaf and a single basis element for class B or C leaves.
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Decision Tree} for inversion symmetry} \\
Goal: construct root and node \emph{structure} \\
\hline
@ depth $d$: \\
\indent node.$\varphi_L$ ($d$ leftmost bits in $\varphi$) \\
\indent node.$\varphi_R$ ($d$ rightmost bits in $\varphi$) \\
\indent node.count (number of representatives) \\
\indent node.A (pointer to depth $d+1$, class A) \\
\indent node.B (pointer to depth $d+1$, class B) \\
\indent node.C (pointer to depth $d+1$, class C) \\
\indent node.parent (pointer to depth $d-1$) \\
\hline
\textbf{function} {\it decision\_tree}\,($n$, $N$)\\
\indent \textbf{for} $d = N/2$ to $0$: ($N$ even) \\
\indent \indent \textbf{construct} \emph{root} and \emph{nodes} @ $d$ \\
\indent \indent \indent \emph{assign} node.$\varphi_L(\varphi_R)$ \\
\indent \indent \indent \emph{assign} node.A(B,C) \\
\indent \indent \indent \emph{assign} node.parent \\
\indent \indent \textbf{calculate} node.count \\
\indent \indent \indent $C_i^m$ (class A) \\
\indent \indent \indent $S_{i}^{m}$ (class B or C) \\
\indent \textbf{end} \\
\indent \textbf{return} {\it root} \\
\hline
\end{tabular}
\newline
Unfortunately, one cannot simultaneously obtain the benefits of using both translation and inversion symmetry in \emph{Paradeisos}.
Invoking translation symmetry typically reduces the effective Hilbert space dimension much more than the reduction obtained from the inversion symmetric subspace. However, the inversion symmetric algorithm may be employed for some problems that lack translation invariance, such those with open boundaries.
To be complete, we present pseudocode for the forward and backward mapping \emph{Paradeisos} algorithms with inversion symmetry.
\noindent\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} inversion forward map}\\
Goal: For $n$ electrons and $N$ sites,
map $\varphi \mapsto \mathrm{idx}(\varphi) $\\
\hline
\textbf{function} {\it forward\_map\_inversion}\,($\varphi$)\\
\indent node $\leftarrow$ {\it decision\_tree}\,($n$, $N$) \\
\indent \textbf{for} $d = 0$ to $N/2 - 1$: ($N$ even)\\
\indent \indent \textbf{switch}($(N-1-d)$-th bit in $\varphi$, $d$-th bit in $\varphi$): \\
\indent \indent \textbf{case: } \texttt{FALSE}, \texttt{TRUE}:\\
\indent \indent \indent $\phi$ $\leftarrow$ ($d$+1)-th through $(N-2-d)$-th bits in $\varphi$ \\
\indent \indent \indent $\textrm{idx}$\ $\leftarrow$\ $\textrm{idx}$ + {\it forward\_map}($\phi$) \\
\indent \indent \indent \textbf{return} $\textrm{idx}$ \\
\indent \indent \textbf{case: } \texttt{FALSE}, \texttt{FALSE}:\\
\indent \indent \indent $\textrm{idx}$\ $\leftarrow$\ $\textrm{idx}$ + node.A.count \\
\indent \indent \indent node\ $\leftarrow$\ node.B \\
\indent \indent \textbf{case:} \texttt{TRUE}, \texttt{TRUE}:\\
\indent \indent \indent $\textrm{idx}$\ $\leftarrow$\ $\textrm{idx}$ + node.A.count + node.B.count \\
\indent \indent \indent node $\leftarrow$ node.C \\
\indent \textbf{end}\\
\indent \textbf{return} $\textrm{idx}$ \\
\hline
\end{tabular}\\
\newline
\begin{figure*}[ht]
\centering
\includegraphics[width=1.8\columnwidth]{Realspace_mem.pdf}
\caption{(Color online) Comparison of aggregate memory usage as a function of the number of processors for sparse matrix construction and diagonalization (obtaining the groundstate wavefunction), for a one-dimensional spin-1/2 Hubbard model in real space without invoking any Hilbert space reductions due to symmetry. The blue curves correspond to traditional binary search, green curves correspond to \emph{Paradeisos}, and red curves relate to \emph{Paradeisos} combined with a checkboard decomposition. The dashed lines indicate the memory required to store a single vector within the Hilbert space serially for each process. Note significant differences in scale with Hilbert space dimension (lattice size).}
\label{fig:RealspaceMem}
\end{figure*}
\noindent
\begin{tabular}{|l|}
\hline
\textbf{\emph{Paradeisos} inversion backward map}\\
Goal: For $n$ electrons and $N$ sites,
map $\mathrm{idx}(\varphi) \mapsto \varphi$\\
\hline
\textbf{function} {\it backward\_map\_inversion}\,(idx; $n$, $N$)\\
\indent node $\leftarrow$ {\it decision\_tree}\,($n$, $N$), $p = $ idx \\
\indent \textbf{for} $d = 0$ to $N/2 - 1$: ($N$ even)\\
\indent \indent \textbf{if} $p$ $\geq$ node.A.count + node.B.count: \\
\indent \indent\indent $p$ $\leftarrow$ $p$ $-$ node.A.count $-$ node.B.count\\
\indent \indent\indent node $\leftarrow$ node.C\\
\indent \indent \textbf{elseif} $p$ $\geq$ node.A.count: \\
\indent \indent\indent $p$ $\leftarrow$ $p$ $-$ node.A.count\\
\indent \indent\indent node $\leftarrow$ node.B\\
\indent \indent \textbf{else}: \\
\indent \indent \indent node $\leftarrow$ node.A \\
\indent \indent \indent $m$ $\leftarrow$ $n$ $-$ bit\_count(node.$\varphi_L$ $\wedge$ node.$\varphi_R$, 1) \\
\indent \indent \indent $\phi$ $\leftarrow$ {\it backward\_map}($p$; $m$, $N - 2d$) \\
\indent \indent\indent \textbf{return} node.$\varphi_L$ $\wedge$ $\phi$ $\wedge$ node.$\varphi_R$ \\
\indent \indent \textbf{end if}\\
\indent \textbf{end}\\
\indent \textbf{return} node.$\varphi_L$ $\wedge$ node.$\varphi_R$ \\
\indent ($\wedge$ represents concatenation of binary sequences) \\
\hline
\end{tabular}
\section{Numerical Benchmarks}
The goal in devising the \emph{Paradeisos} algorithm was to address one of the remaining bottlenecks for sparse matrix eigensolvers -- memory requirements for basis element hashing in traditional binary search that lead to a large communications overhead or significant aggregate memory consumption in large-scale parallel implementations. In this section we benchmark the performance of \emph{Paradeisos} against binary search, and investigate additional benefits when coupling the algorithm with the checkerboard decomposition for data parallelism. We concentrate specifically on how memory usage scales with the number of processors for single-band Hubbard model calculations on clusters of varying size, as the scaling slope usually determines the quality of a parallelization scheme.
In a parallel matrix-vector algorithm, the storage of a single vector of size $D$ on each processor characterizes the typical memory cost and can usually to be used as a reference to determine the necessity of a sophisticated parallelization scheme like the ``checkerboard" decomposition. Thus, we denote such a memory cost as grey lines in Figs.~\ref{fig:RealspaceMem} and ~\ref{fig:KspaceMem}. As shown in Fig.~\ref{fig:RealspaceMem}, the traditional binary search method (blue) has a relatively large scaling slope with system size, making the parallelization cost extremely expensive for large cluster calculations. Once the overhead cost exceeds the benefits of parallelization, adding more processors would not help to solve the memory problem. \emph{Paradeisos} (green) succeeds in reducing the slope.
However, the steady increase of aggregate memory with the number of processors still limits parallelization. When one applies \emph{Paradeisos} with the checkerboard decomposition (red), the situation changes dramatically. The scaling slope becomes relatively small and shows little change with growth in the dimension of the Hilbert space (dictated by the lattice size). The memory overhead remains effectively constant, an important consideration for parallel computations that then may be cheaply divided over many nodes or processors. For the largest problems, \emph{Paradeisos} improves the scaling beyond that of even storing a single vector in the Hilbert space for each process.
As shown in Fig.~\ref{fig:KspaceMem}, when applying symmetry reduction for small Hilbert space dimensions (the smallest lattice problems), \emph{Paradeisos} alone performs marginally better than binary search and even when combined with the checkerboard decomposition, primarily due to residual memory overhead from other parts of the sparse matrix diagonalization code. With increasing Hilbert space dimension (lattice size), the results remain qualitatively similar to the those from the real space implementation of the algorithm without symmetry reduction. Fig.~\ref{fig:slope} compares the memory scaling slope for various lattice sizes, showing the significant improvements in memory consumption that can be realized by combining \emph{Paradeisos} with the checkerboard decomposition.
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\columnwidth]{Kspace_mem.pdf}
\caption{(Color online) Comparison between traditional binary search (blue), \emph{Paradeisos} (green), and \emph{Paradeisos} combined with a checkerboard decomposition (red). Each curve shows overall memory consumption for a one-dimensional spin-1/2 Hubbard model on chains (lattices) of length (a) 12, (b) 14, and (c) 16 sites. Each one represents the memory required for solving the problem in the $K = 0$ momentum subspace, which contains the groundstate wavefunction.}
\label{fig:KspaceMem}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{mem_scaling_slope.pdf}
\caption{Summary of the memory scaling slopes as a function of system size from Figs.~\ref{fig:RealspaceMem} and \ref{fig:KspaceMem}. The solid dots represent results from the real space implementation, while the open symbols correspond to those obtained from invoking the translation symmetry reduction of the effective Hilbert space.}
\label{fig:slope}
\end{figure}
\section{Summary and Discussion}
We have proposed a perfect hashing algorithm -- a direct mapping between Hilbert space basis elements and their corresponding index -- for use in sparse matrix eigenvalue problems. Compared to the traditional binary search, the present algorithm provides a considerable savings in aggregate memory usage without incurring additional penalties to the time complexity. Moreover, in concert with a checkerboard decomposition scheme, the memory overhead can be negligible (effectively independent of problem size), implying efficient parallelization for large size compute environment. The Paradeisos algorithm is also compatible with additional point group symmetries -- which can highly reduce the Hilbert space dimension, and thus can be efficiently applied to most quantum many-body systems or models. The algorithm can be extended to apply on many-body bosonic basis states, at which a particle number truncation is necessary. The algorithm eschews storage, and as such, may be utilized in \emph{matrix-free} implementations of eigenvalue solvers.
\section{Acknowledgements}
\begin{acknowledgments}
This work was supported at SLAC and Stanford University by the U.S.~Department of Energy, Office of Basic Energy Sciences, Division of Materials Sciences and Engineering, under Contract No.~DE-AC02-76SF00515. Y.W.\ was supported by the Stanford Graduate Fellows in Science and Engineering, and C.M.\ acknowledges support from the Alexander von Humboldt foundation via a Feodor Lynen fellowship. The computational work was performed using the resources of the National Energy Research Scientific Computing Center supported by the U.S. Department of Energy, Office of Science, under Contract No.~DE-AC02-05CH11231.
\end{acknowledgments}
|
\section{Introduction}
A \emph{complex manifold} is a differentiable manifold $M$ whose transition functions are holomorphic. If $x+iy$ denotes local coordinates around $p\in M$, the multiplication by $i$ induces an endomorphism $J_p$ of the tangent space $T_p(M)$ given by
\begin{equation}
\label{eq:cpx_mfd}
\left\{
\begin{aligned}
\frac{\partial}{\partial_x}|_p&\longmapsto\frac{\partial}{\partial_y}|_p\\
\frac{\partial}{\partial_y}|_p&\longmapsto-\frac{\partial}{\partial_x}|_p
\end{aligned}
\right.
\end{equation}
Writing the Cauchy-Riemann equations on the intersection of two overlapping charts, one obtains that $J_p$ is globally defined. The endomorphism $J:T(M)\rightarrow T(M)$, locally defined by \eqref{eq:cpx_mfd}, is an example of almost complex structure. More generally, an \emph{almost complex structure} on a real differentiable manifold $M$ is a linear bundle morphism $J$ of the tangent vector bundle $T(M)$ satisfying $J^2=-\mathop{\mathrm{Id}}$. The pair $(M,J)$ is called an \emph{almost complex manifold}, and whenever $J$ arises from \eqref{eq:cpx_mfd} for certain holomorphic coordinates $x+iy$, one says that the almost complex structure $J$ is \emph{integrable}.
One often refers to the above discussion by the sentence: ``on a complex manifold the multiplication by $i$ gives a compatible integrable almost complex structure''.
The problem of deciding if a given almost complex structure $J$ is integrable is nowadays an easy task, due to the celebrated Newlander-Nirenberg Theorem \cite{NeNCAC}. Much more challenging is the problem of deciding whether a differentiable manifold admits \emph{any} integrable almost complex structure: in real dimension $>4$, the only known general obstruction is the trivial one, that is, the existence of an almost complex structure.
The existence of an almost complex structure $J$ on a manifold $M$ is possible only if the dimension $m$ of $M$ is even, because of $(\det J)^2=(-1)^m$. Moreover, $J$ is equivalent to a $\GLC{m/2}$-structure on $T(M)$. In fact, any endomorphism $J_p$ on the tangent space $T_p(M)$ such that $J_p^2=-\mathop{\mathrm{Id}}$ admits an \emph{adapted} basis $\{v,J(v)\}$, and with respect to this basis
$J_p=
\begin{pmatrix}
0,-1\\
+1,0
\end{pmatrix}$. Thus, all the $J_p$ stick together giving an almost complex structure $J$ if and only if the structure group $\GLR{2n}$ can be reduced to $\GLC{n}$. See \cite[Remark 2 at page 8]{BrySRG} for a nice discussion regarding the terminology ``almost'' and ``integrable'' for $G$-structures.
Very clear references for basic facts on almost complex structures are the classical Kobayashi-Nomizu \cite[IX.1 and IX.2]{KoNFD2} and the Besse \cite[2.A]{BesEiM}.
This paper originates from the talk ``Almost Complex Structures on Spheres'' given by the second author at the \href{http://www.mathematik.uni-marburg.de/~agricola/Hopf2017/}{MAM1 workshop ``(Non)-existence of complex structures on $S^6$''}, held in Marburg from March 27th to March 30th, 2017. The plan of this beautiful workshop was to discuss and summarize various results of the past 30 years about the existence or non-existence of integrable almost complex structures on spheres, and this talk was about the classical result of Borel and Serre ruling out all spheres besides $S^2$ and $S^6$, namely, the result from where the story began:
\begin{thm}[Borel and Serre, 1953 \cite{BoSGLP}]\label{thm:main}
The sphere $S^n$ admits an almost complex structure if and only if $n=2$ or $n=6$.
\end{thm}
The material in this paper comes from several sources, clearly stated whenever necessary. As such, no result is intended to be original. However, we tried to produce a clear, motivated and as much as possible self-contained exposition.
\section{Almost complex structures on $S^2$ and $S^6$}
An almost complex structure on $S^2$ can be constructed in several ways. One can observe that since $S^2$ is orientable, its structure group $\GLR{2}$ can be reduced to $\SO{2}=\U{1}\subset\GLC{1}$. Or, one can observe that $S^2$ is diffeomorphic to $\CP{1}$.
An explicit construction of $J$ is possible by means of a vector product in $R^3$. The feasibility of this approach, that we are now going to describe, is that the same construction leads to an almost complex structure on $S^6$.
Denote by $\mathbb{H}$ and $\mathbb{O}$ the quaternions and the octonions, respectively. We will denote the product of $u,v\in\mathbb{H}$ or $u,v\in\mathbb{O}$ by juxtaposition $uv$, and the conjugation by $u^*$. Note then that $(uv)^*=v^*u^*$, and $uu^*=|u|^2$, where $|\cdot|$ is the standard Euclidean norm. By polarization, we get $2\langle u,v\rangle=uv^*+v^*u$, which implies the useful identity $uv^*=-v^*u$ whenever $u\perp v$. On the octonions, we have a non-trivial \emph{associator}, that is, $[u,v,w]\;\shortstack{{\tiny\upshape def}\\=}\;((uv)w)-(u(vw))$ is not identically zero if $u,v,w\in\mathbb{O}$. A key property of this associator is that it is \emph{alternating}, that is, $[u,v,w]=0$ whenever two amongst $u,v,w$ are equal.
Look at the spheres $S^2$ and $S^6$ as unit spheres in $\Im\mathbb{H}$, $\Im\mathbb{O}$, respectively:
\[
S^2=\{x\in\Im\mathbb{H}\text{ such that } |x|^2=1\},\qquad S^6=\{x\in\Im\mathbb{O}\text{ such that } |x|^2=1\}
\]
Let $V$ denote either $\Im\mathbb{H}$ or $\Im\mathbb{O}$. A cross product $u\times v$ can be defined on $V$ by:
\[
u\times v\;\shortstack{{\tiny\upshape def}\\=}\;\Im(uv)=\frac{1}{2}(uv-vu)\qquad u,v\in V
\]
\begin{prp}
Let $p\in S^n\subset V$, $n=2$ or $n=6$. Define $J\in\End{T(S^n)}$ by $J_p(v)\;\shortstack{{\tiny\upshape def}\\=}\; b\times v$, where $v\in T_p(S^n)\subset V$. Then $J$ is an almost complex structure on $S^n$, that is, $J_p^2=-\mathop{\mathrm{Id}}$ for any $p\in S^n$.
\end{prp}
\begin{proof}
First, we prove that if $u\perp v$, then $u\times v=uv$:
\[
(uv)^*=v^*u^*\stackrel{v\in\Im}{=}-vu^*\stackrel{u\perp v}{=}uv^*=-uv\implies\Re(uv)=0\implies u\times v=\Im(uv)=uv
\]
Now, since $p\perp v$, we have $J_p^2(v)=p(pv)\stackrel{[p,p,v]=0}{=}(pp)v=-(pp^*)v=-(|p|^2)v=-v$.
\end{proof}
\begin{rem}
The algebras $\mathbb{H}$ and $\mathbb{O}$ belong to an infinite family of algebras $A_n$ of dimension $2^n$ arising from the \emph{Cayley-Dickson process}. One could wonder whether the above construction works for any other $A_n$, apart from $\mathbb{H}=A_2$ and $\mathbb{O}=A_3$. The answer is no, essentially because in order to define a cross product, you need $A_n$ be alternative, and this happens only for $n$ up to 3. To deepen this part of the story, see \cite[Appendix IV.A, page 140 and forward]{HaLCaG}, \cite[Hurwitz 1898 paper, in German]{HurNGW}, \cite[Section 2.2]{BaeOct}, \cite{BrGVCP}, and the recent \cite{MclWDC}.
\end{rem}
\begin{rem}
Denote by $N_J$ the Nijenhuis tensor of $J$. Then, the almost complex structure on $S^2$ is integrable, because it is the complex structure induced by $\CP{1}$ (this can be seen geometrically, because $p\times v$ is a $\pi/2$ counterclockwise rotation of $v$ on the plane $p^\perp$, or observing that in complex dimension one $N_J$ is 0, thus $(S^2,J)$ is a compact complex curve of genus 0, the only one being $\CP{1}$), whereas the almost complex structure on $S^6$ is not integrable \cite{EcFISP}. The following statement is then true \emph{a fortiori}:
\[
N_J=0\iff [u,v,w]=0\qquad\forall u,v,w\in V
\]
This suggests a direct relation between $N_J(u,v)$ and some expression involving associators. A natural one would be $\langle N_J(u,v),w\rangle=[u,v,w]$. However, we have not been able so far to find such a relation.
\end{rem}
\section{Characteristic classes}
According to Dieudonné \cite[IV.1]{DieHAD}, characteristic classes arose in 1935 in Stiefel's dissertation \cite{StiRFN}, as a tool to investigate whether a manifold $M^n$ admits $m\le n$ linearly independent nowhere vanishing vector fields. Looking at the $(r+1)$-skeleton of $M$, after orthonormalizing, such $m$ vector fields define maps $\alpha$ from disks $D^{r+1}$ into the Stiefel manifold $\StR{m}{n}$ of orthonormal $m$-frames in $R^n$. Since this argument reduces the problem to whether these $m$ vector fields can be extended from the sphere $S^r$ to the disk $D^{r+1}$, the problem of existence can be read in terms of homotopy groups $\pi_r(\StR{m}{n})$, for $r\ge n$. Since $\pi_{r}(\StR{m}{n})=0$ for $r<n-m$, such $m$ vector fields always exist on the $(m-n)$-skeleton of $M$, and the first obstruction appears as $\alpha\in\pi_{n-m}(\StR{m}{n})$. This $\alpha$ is what was called the \emph{characteristic} of the $m$-frame by Stiefel, giving birth to the name \emph{characteristic classes}.
\subsection{Classifying spaces}
The above argument is the obstruction-theoretic definition of \emph{Stiefel-Whitney} characteristic classes of the tangent bundle of a manifold $M$. More generally, characteristic classes are associated to a (real or complex) vector bundle $E\rightarrow B$: they are defined as cohomology classes of $B$ induced by any classifying map $\phi:B\rightarrow \BG{\GL{n}}$, where $\BG{\GL{n}}=\BG{\GLR{n}}$ or $\BG{\GLC{n}}$ is the classifing space for real or complex vector bundles.
In the following, we sketch the notions of classifying maps, classifying spaces and universal bundles in the general setting of principal $G$-bundles. For a historical description going from 1935 Whitney paper \cite{WhiSpS} to 1956 Milnor beautiful construction \cite{MilCUB,MilCII}, see \cite[III.2E-2G]{DieHAD}. The main reference for this part is \cite{HatVBK}.
Let $G$ be a topological group, acting freely on a \emph{weakly contractible} topological space $\EG{G}$ (that is, with trivial homotopy groups). The corresponding principal $G$-bundle $\pi_G:\EG{G}\rightarrow\BG{G}$ is called a \emph{universal $G$-bundle}, and the base space $\BG{G}=\EG{G}/G$ a \emph{classifying space} of $G$. The reason for the name ``universal bundle'' is the following theorem.
\begin{thm*}
Any principal $G$-bundle $\pi:E\rightarrow B$ over a CW-complex $B$ is isomorphic to a pull back of a universal $G$-bundle under some continuous map $f:B\rightarrow\BG{G}$. Moreover, two pull backs are isomorphic if and only if the corresponding maps are homotopic.
\end{thm*}
The map $f$ in the theorem is called a \emph{classifying map} for the principal bundle $E\rightarrow B$. For $B$ a CW-complex, the reason why the universal bundle exists is the \emph{Brown's representability theorem}, a result in Category Theory \cite{BroCoT}. However, hypothesis can be relaxed in several directions: for instance, $B$ can be a paracompact topological space instead of a CW-complex.
The need for a universal bundle arise from the classification of fiber bundles with a fixed $G$ as structure group. After reducing to the associated principal $G$-bundle $E\rightarrow B$ one can, heuristically, look for a principal $G$-bundle $\EG{G}\rightarrow\BG{G}$ \emph{depending only on $G$} such that \emph{any principal $G$-bundle} is induced by $\EG{G}\rightarrow\BG{G}$ by means of a bundle map $\phi$:
\begin{center}
\begin{equation}\label{dgr:univ}
\begin{tikzpicture}[scale=1]
\node (A) at (0,1) {$E=\phi^*(\EG{G})$};
\node (B) at (2,1) {$\EG{G}$};
\node (C) at (0,0) {$B$};
\node (D) at (2,0) {$\BG{G}$};
\path[->]
(A) edge (B)
(C) edge node[above]{$\phi$} (D)
(A) edge (C)
(B) edge (D);
\end{tikzpicture}
\end{equation}
\end{center}
Using the left and the right action of $G$ respectively on $E$ and $\EG{G}$, and the canonical projections, one obtains the following diagram:
\begin{center}
\begin{tikzpicture}[scale=1]
\node (A) at (0,1) {$E=\phi^*(\EG{G})$};
\node (B) at (5,1) {$\EG{G}$};
\node (C) at (0,0) {$B$};
\node (D) at (5,0) {$\BG{G}$};
\node (E) at (3,1) {$E\times\EG{G}$};
\node (F) at (3,0) {$E\times_G\BG{G}$};
\path[->]
(E) edge (A)
(E) edge (B)
(E) edge (F)
(F) edge node[above]{$\pi$} (C)
(F) edge node[above]{$p$} (D)
(A) edge (C)
(B) edge (D);
\end{tikzpicture}
\end{center}
Thus, if one can find a section $s$ of $\pi$, the classifying map is $\phi=p\circ s$: but such a section $s$ exists whenever all the homotopy groups of $\EG{G}$ vanish - here is where the need for $\EG{G}$ to be weakly contractible comes from. As said before, under mild hypothesis on the base space $B$ and the group $G$, such a weakly contractible space $\EG{G}$ always exists.
Moreover, since everything behaves well with respect to homotopies, one obtains a bijective correspondence between principal bundles and homotopy classes of continuous maps:
\[
\{\text{isomorphism classes of principal $G$-bundles over $B$}\}\stackrel{1-1}{\longleftrightarrow} [B:\BG{G}]
\]
Concrete examples of classifying spaces can be given using the notion of Grassmannian and Stiefel manifold. For any $n\le k$, denote by $\GrR{n}{k}$ the set of $n$-dimensional vector subspaces of $\mathbb{R}^k$, and consider the set
\[
\TauR{n}{k}\;\shortstack{{\tiny\upshape def}\\=}\;\{(v,l)\in\mathbb{R}^k\times\GrR{n}{k}\text{ such that } v\in l\}
\]
The map sending $(v,l)\in\TauR{n}{k}$ into $l\in\GrR{n}{k}$ defines a \emph{tautological} vector bundle $\TauR{n}{k}\rightarrow\GrR{n}{k}$. Because of Gram–Schmidt, its associated $\GLR{n}$-principal bundle and $\Or{n}$-principal bundle, having as fiber over $l\in\GrR{n}{k}$ respectively the isomorphisms and the isometries from $\mathbb{R}^n$ to $l$, have the same homotopy type. The latter is called the \emph{Stiefel manifold} $\StR{n}{k}$ of ordered orthonormal $n$-frames in $\mathbb{R}^k$. Using the fibration $\StR{n-1}{k-1}\hookrightarrow\StR{n}{k}\twoheadrightarrow S^{k-1}$, one sees that $\StR{n}{k}$ is compact, and that the first non-trivial homotopy group of the Stiefel manifold is $\pi_{k-n}(\StR{n}{k})$. Considering the sequence of inclusions $\mathbb{R}\subset\mathbb{R}^2\subset\mathbb{R}^3\subset\dots$, one can define the direct limit versions of the above: the tautological bundle $\EGLR{n}\rightarrow\BGLR{n}$ over the infinite Grassmannian and the associated infinite Stiefel manifold $\EO{n}$. Note that $\EO{n}$ is no more compact, but is weakly contractible.
Consider the tangent bundle $T(S^n)\rightarrow S^n$ of the sphere $S^n\subset\mathbb{R}^{n+1}$. For any $x\in S^n$, we have $T_x(S^n)\subset\mathbb{R}^{n+1}$, and this gives a map $\phi:S^n\rightarrow\GrR{n}{n+1}$. The fiber on $x\in S^n$ are the vectors that are tangent to the sphere in $x$. Read in $\mathbb{R}^{n+1}$, these vectors are the elements of $\phi(x)$, that is, the fiber over $x\in S^n$ is ``tautologically'' the fiber of $\TauR{n}{n+1}\rightarrow\GrR{n}{n+1}$. In this way, diagram \eqref{dgr:univ} becomes
\begin{center}
\begin{tikzpicture}[scale=1]
\node (A) at (0,1) {$T(S^n)$};
\node (B) at (3,1) {$\TauR{n}{n+1}$};
\node (C) at (0,0) {$S^n$};
\node (D) at (3,0) {$\GrR{n}{n+1}$};
\path[->]
(A) edge (B)
(C) edge (D)
(A) edge (C)
(B) edge (D);
\end{tikzpicture}
\end{center}
The above argument works whenever $B$ is a smooth manifold, because of the Whitney embedding theorem. More generally, one can show that every rank $n$ vector bundle $E\rightarrow B$ over a paracompact topological space $B$ is induced by $\EGLR{n}\rightarrow\BGLR{n}$. Since the total space of the associated $\GLR{n}$-principal bundle has the same homotopy type as $\EO{n}$, and $\EO{n}$ is weakly contractible, we have that $\BGLR{n}$ is at the same time the classifying space $\BG{\GLR{n}}$ for real vector bundles and the classifying space $\BG{\Or{n}}$ for real Euclidean vector bundles. A similar argument shows that $\BG{\GLC{n}}=\BG{\U{n}}=\BGLC{n}$ is the classifying space for complex vector bundles, and complex Hermitian bundles as well. In summary:
\begin{align}
\VectR{n}{B}&=[B:\BGLR{n}]\label{eq:vectreal}\\
\VectC{n}{B}&=[B:\BGLC{n}]\label{eq:vectcpx}
\end{align}
\subsection{Stiefel-Whitney and Chern classes}
In theory, formulas \eqref{eq:vectreal} and \eqref{eq:vectcpx} solve the classification problem for vector bundles. In real life, the set of homotopy classes of maps is very difficult to compute. However, the notion of universal bundle $\EG{G}\rightarrow\BG{G}$ can be used to define $G$-bundle invariants.
\begin{dfn}\label{dfn:classes}
Let $E\rightarrow B$ be a $G$-principal bundle, and $f:B\rightarrow\BG{G}$ a classifying map. Cohomology classes $c$ of $B$ induced by cohomology classes $c'$ of $\BG{G}$, namely $c=f^*(c')\in H^*(B;R)$, are called \emph{characteristic classes} of $E\rightarrow B$. Here $R$ is any ring of coefficients.
\end{dfn}
This definition works because different classifying maps are homotopic, and two homotopic maps induce the same homomorphism on cohomology. Characteristic classes of vector bundles are then the pull-back by any classifying map of cohomology classes of the Grassmannian. In the following theorem, we describe the cohomology ring structure of $\BGLR{n}$ and $\BGLC{n}$ with coefficients in $\mathbb{Z}$ and $\mathbb{Z}_2$.
\begin{thm}\label{thm:cohomology}
\begin{align}
H^*(\BGLR{n};\mathbb{Z}_2)&=\mathbb{Z}_2[w_1,\dots,w_n],\qquad w_i\in H^i(\BGLR{n};\mathbb{Z}_2)\label{frm:st-wh}\\
H^*(\BGLC{n};\mathbb{Z})&=\mathbb{Z}[c_1,\dots,c_n],\qquad c_i\in H^{2i}(\BGLC{n};\mathbb{Z})\label{frm:chern}
\end{align}
\end{thm}
Theorem \ref{thm:cohomology} can be proved in several ways. One can use the sphere bundles $S(\EGLR{n})\rightarrow\BGLR{n}$, $S(\EGLC{n})\rightarrow\BGLC{n}$ and their associated Gysin sequences. Or one can use the \emph{splitting principle} \cite[from page 273]{BoTDFA}, as for instance in \cite[Theorem 7.1]{MiSChC}. A third possibility is using cellular cohomology, and a cellular decomposition of the Grassmannian by means of \emph{Schubert cells}, see for instance the very clear description in \cite[pages 31--34]{HatVBK}.
Combining definition \ref{dfn:classes} and theorem \ref{thm:cohomology}, we define the Stiefel-Whitney and the Chern classes. For Stiefel-Whitney classes, this approach was used for the first time by Pontrjagin in \cite{PonCCM,PonCCD}, see \cite[IV.3.C]{DieHAD}.
\begin{dfn}\label{dfn:st-wh}
Let $E\rightarrow B$ be a rank $n$ real vector bundle, $f:B\rightarrow\BGLR{n}$ a classifying map and $w_i$ the generator of $H^i(\BGLR{n};\mathbb{Z}_2)$ as in formula \eqref{frm:st-wh}. The cohomology classes $w_i(E)\;\shortstack{{\tiny\upshape def}\\=}\; f^*(w_i)\in H^i(B;\mathbb{Z}_2)$, for $i=1,\dots,n$, are called the \emph{Stiefel-Whitney classes} of $E\rightarrow B$. The \emph{total Stiefel-Whitney class} is then $w(E)\;\shortstack{{\tiny\upshape def}\\=}\; 1+w_1(E)+\dots+w_n(E)$.
\end{dfn}
\begin{dfn}\label{dfn:chern}
Let $E\rightarrow B$ be a rank $n$ complex vector bundle, $f:B\rightarrow\BGLR{n}$ a classifying map and $c_i$ the generator of $H^{2i}(\BGLC{n};\mathbb{Z})$ as in formula \eqref{frm:chern}. The cohomology classes $c_i(E)\;\shortstack{{\tiny\upshape def}\\=}\; f^*(c_i)\in H^{2i}(B;\mathbb{Z})$, for $i=1,\dots,n$, are called the \emph{Chern classes} of $E\rightarrow B$. The \emph{total Chern class} is then $c(E)\;\shortstack{{\tiny\upshape def}\\=}\; 1+c_1(E)+\dots+c_n(E)$.
\end{dfn}
\begin{rem}\label{rem:axiom}
Denote by $\cdot$ the product in the cohomology ring (for instance, the cup product for singular cohomology, or the wedge product for de Rham cohomology). Stiefel-Whitney and Chern classes satisfies the following fundamental properties.
\begin{enumerate}
\item This is called \emph{naturality}, and follows from homotopic invariance of the classifying map: if $g:B'\rightarrow B$ is a continuous map, then
\[
w(g^*E)=g^*w(E)\quad|\quad c(g^*E)=g^*c(E)
\]
\item The \emph{Whitney product formula} follows from the splitting principle \cite[page 66]{HatVBK}:
\begin{equation}\label{eq:product}
w(E\oplus E')=w(E)\cdot w(E')\quad|\quad c(E\oplus E')=c(E)\cdot c(E')
\end{equation}
\item The \emph{normalization property} follows from the very definition: $w_1(\BGLR{1})$ is the generator of $H^1(\BGLR{1};\mathbb{Z}_2)$ | $c_1(\BGLC{1})$ is the generator of $H^2(\BGLC{1};\mathbb{Z})$.
\end{enumerate}
\end{rem}
\begin{rem}
As usually happens, definition is not as important as properties, and for characteristic classes this is particularly true: in fact, the above properties can be used to define characteristic classes axiomatically. This approach was first taken by Hirzebruch for Chern classes in \cite[pages 60--63]{HirNTM}, see \cite[End of IV.1.E]{DieHAD}.
\end{rem}
\begin{rem}\label{rem:conj}
The \emph{conjugate bundle} $\bar{E}\rightarrow B$ has the same underlying real vector bundle $E_\mathbb{R}$, and the conjugate complex multiplication on fibers: $z\cdot_{\bar{E}}v\;\shortstack{{\tiny\upshape def}\\=}\;\bar{z}\cdot_E v$. In terms of associated almost complex structure $J$, we have $E=(E_\mathbb{R},J)$ and $\bar{E}=(E_\mathbb{R},-J)$ \cite[page 123]{KoNFD2}. For the conjugate bundle, Chern classes are given by
\begin{equation}\label{frm:conj}
c_k(\bar{E})=(-1)^k c_k(E)
\end{equation}
This formula comes from the fact that the identity map $\mathop{\mathrm{Id}}:E\rightarrow\bar{E}$ is conjugate linear: $\mathop{\mathrm{Id}}(z\cdot_E v)=\bar{z}\cdot_{\bar{E}}\mathop{\mathrm{Id}}(v)$ \cite[page 167]{MiSChC}.
\end{rem}
\begin{rem}\label{rem:sta-tri}
Denote by $\trivial{n}$ the trivial vector bundle $\mathbb{R}^n\times B\rightarrow B$ or $\mathbb{C}^n\times B\rightarrow B$ (which one, will always be clear from the context).
A rank $n$ vector bundle $E$ is \emph{stably trivial} if for a certain $k$ one has $E\oplus\trivial{k}=\trivial{n+k}$. Since the constant map induces the trivial bundle, naturality and product formula in remark \ref{rem:axiom} gives that Stiefel-Whitney and Chern classes are trivial for any stably trivial vector bundle.
This is in particular true for tangent bundles $T(S^n)$ of spheres, because $T(S^n)\oplus\trivial{1}=\trivial{n+1}$, where $\trivial{1}$ is the normal bundle of the embedding $S^n\subset\mathbb{R}^{n+1}$.
\end{rem}
\subsection{Pontryagin classes}
If you are wondering why $H^*(\BGLR{n};\mathbb{Z})$ doesn't appear in theorem \ref{thm:cohomology}, the reason is that its ring structure is a bit more complicated. In short, $H^*(\BGLR{n};\mathbb{Z})$ contains torsion classes of order 2 which are ``essentially'' Stiefel-Whitney classes - namely, images under the cohomology map $H^i(\BGLR{n};\mathbb{Z}_2)\rightarrow H^{i+1}(\BGLR{n};\mathbb{Z}_2)$ induced by the short exact sequence
\[
0\rightarrow\mathbb{Z}_2\stackrel{2\cdot}{\rightarrow}\mathbb{Z}_4\rightarrow\mathbb{Z}_2\rightarrow 0
\]
Factoring out these torsion classes, one is left with a free algebra as in theorem \ref{thm:cohomology}:
\begin{equation}\label{frm:pontr}
\frac{H^*(\BGLR{n};\mathbb{Z})}{\text{torsion}}=\mathbb{Z}[p_1,\dots,p_k],\qquad p_i\in H^{4i}(\BGLR{n};\mathbb{Z}),\qquad n=2k \text{ or } n=2k+1
\end{equation}
For a thorough description of the ring structure of $H^*(\BGLR{n};\mathbb{Z})$, see \cite{BroCBB}.
\begin{dfn}\label{dfn:pontr}
Let $E\rightarrow B$ be a rank $n=2k$ or $n=2k+1$ real vector bundle, $f:B\rightarrow\BGLR{n}$ a classifying map and $p_i$ the generator of $H^{4i}(\BGLR{n};\mathbb{Z})/\text{torsion}$ as in formula \eqref{frm:pontr}. The cohomology classes $p_i(E)\;\shortstack{{\tiny\upshape def}\\=}\; f^*(p_i)\in H^{4i}(B;\mathbb{Z})$, for $i=1,\dots,k$, are called the \emph{Pontryagin classes} of $E\rightarrow B$. The \emph{total Pontryagin class} is then $p(E)\;\shortstack{{\tiny\upshape def}\\=}\; 1+p_1(E)+\dots+p_k(E)$.
\end{dfn}
Pontryagin classes can be also defined by means of Chern classes. Consider the complexification $E_\mathbb{C}\;\shortstack{{\tiny\upshape def}\\=}\; E\otimes\mathbb{C}$, and notice that $E_\mathbb{C}=\overline{E_\mathbb{C}}$. Then, since conjugation swaps sign for odd-indexed Chern classes by formula \eqref{frm:conj}, the only non-torsion Chern classes are
\begin{equation}\label{eq:pontr}
p_i(E)\;\shortstack{{\tiny\upshape def}\\=}\; (-1)^i c_{2i}(E_\mathbb{C})\in H^{4i}(B;\mathbb{Z}),\qquad i=1,\dots,[n/2]=k
\end{equation}
\begin{rem}\label{rem:sta-tri-pontr}
Since Pontryagin classes are defined by means of Chern classes, remark \ref{rem:sta-tri} still holds: all Pontryagin classes of a stably trivial bundle vanish. In particular, for any $n$ we have $p(T(S^n))=1$, or, in other words, all Pontryagin classes of spheres vanish.
\end{rem}
To close this section, a trivia on \emph{rational} Pontryagin classes of tangent bundles: they are topological invariants, as shown in \cite{NovMFA}, so that one would expect a topological definition. But such a definition is still missing. However, a combinatorial definition was given in \cite{ThoCCP}.
\subsection{Euler class}
Whenever we have a distinguished orientation on a real vector bundle $E\rightarrow B$, another characteristic class appears: the \emph{Euler class} $e(E)$. In terms of classifying space, the Euler class belongs to the cohomology of $\BG{\SO{n}}=\BGSO{n}$, where $\BGSO{n}$ is the Grassmannian of \emph{oriented} $n$-dimensional vector subspaces of $\mathbb{R}^{\infty}$. Whenever $n$ is odd, $E$ admits an orientation reversing isomorphism, and this translates to $e(E)=e(-E)=-e(E)$, meaning that $e(E)$ is a torsion class. For this reason, the Euler class is often considered only for even $n=2k$.
The following is the cohomology structure up to torsion of $\BGSO{2k}$:
\begin{equation}\label{frm:euler}
\frac{H^*(\BGSO{2k};\mathbb{Z})}{\text{torsion}}=\mathbb{Z}[p_1,\dots,p_{k-1},e],\quad p_i\in H^{4i}(\BGSO{2k};\mathbb{Z}),\quad e\in H^{2k}(\BGSO{2k};\mathbb{Z})
\end{equation}
If $B$ is an oriented smooth manifold of dimension $d$, with Euler characteristic
\[
\chi(B)\;\shortstack{{\tiny\upshape def}\\=}\;\sum_{i=0}^d(-1)^i\rank{H_i(B;\mathbb{Z})}
\]
and $[B]\in H_d(B;\mathbb{Z})$ the fundamental homology class of $B$, we have
\[
\chi(B)=\langle e(T(B),[B]\rangle
\]
where the brackets denote the usual Kronecker pairing. This is the reason of the name for the Euler characteristic class.
When $E\rightarrow B$ is complex of rank $n$, the underlying real vector bundle $E_\mathbb{R}$ has rank $2n$, and $e(E_\mathbb{R})\in H^{2n}(B;\mathbb{Z})$ can be identified with the top Chern class of $E$:
\begin{equation}\label{eq:topchern}
e(E_\mathbb{R})=c_{n}(E)
\end{equation}
\section{Ruling out $S^{4k}$}
The same argument showing that $S^4$ does not admit an almost complex structure holds for every $S^{4k}$, and this latter is somehow clearer to state. Therefore, in this section we consider $S^{4k}$, and prove that it doesn't admit any almost complex structure.
The core of the proof is the fact that an almost complex structure on $S^{4k}$ would imply the last Pontryagin class to be twice the Euler class.
\begin{lem}\label{lem:eulpon}
If there is an almost complex structure on $S^{4k}$, then
\[
(-1)^k p_k(T(S^{4k}))=2e(T(S^{4k}))
\]
\end{lem}
\begin{proof}
Assume that an almost complex structure $J$ on $S^{4k}$ exists. Then each tangent space is a complex vector space \cite[page 114]{KoNFD2}, and $(T(S^{4k}),J)$ is a complex vector bundle that here we denote briefly by $T$. Since $S^{4k}$ has non-trivial cohomology only in dimension $4k$, all Pontryagin and Chern classes vanish except $p_k(T(S^{4k}))$ and $c_{2k}(T)$. Moreover, extending $J$ to a complex linear endomorphism of $T(S^4)\otimes\mathbb{C}$, we obtain the $\pm i$-eigenspaces splitting $T(S^4)\otimes\mathbb{C}=T\oplus\bar{T}$ \cite[IX, proposition 1.5]{KoNFD2}. Thus
\[
\begin{split}
1+(-1)^k p_k(T(S^{4k}))\stackrel{\eqref{eq:pontr}}{=}c(T(S^{4k})\otimes\mathbb{C})&=c(T\oplus \bar{T})\stackrel{\eqref{eq:product}}{=}c(T)\cdot c(\bar{T)})\\
&=(1+c_{2k}(T))\cdot(1+c_{2k}(T))=1+2c_{2k}(T)\stackrel{\eqref{eq:topchern}}{=}1+2e(T(S^{4k}))
\end{split}
\]
and this proves the lemma.
\end{proof}
\begin{thm}\label{thm:hirzebruchs4k}
The spheres $S^{4k}$ for $k\geq 1$ do not admit any almost complex structure.
\end{thm}
\begin{proof}
If $S^{4k}$ admits an almost complex structure, we get $(-1)^k p_k(T(S^{4k}))=2e(T(S^{4k}))$ from lemma \ref{lem:eulpon}. But $e(T(S^{4k}))$ is non-trivial on even-dimensional spheres, because $\chi=2\neq 0$, and this contradicts remark \ref{rem:sta-tri-pontr} stating that all Pontryagin classes of spheres are trivial. More explicitly, denoting by $[S^{4k}]$ the fundamental class of $S^{4k}$ we have
\[
\langle p_k(T(S^{4k})),[S^{4k}]\rangle=(-1)^k\langle 2e(T(S^{4k}),[S^{4k}]\rangle=(-1)^k 2\chi(S^{4k})=(-1)^k 4\neq 0
\]
\end{proof}
\subsection{Ruling out $S^{4k}$ via Hirzebruch Signature Theorem}
Another nice way to prove theorem \ref{thm:hirzebruchs4k} is by using the Hirzebruch Signature Theorem. In short, lemma \ref{lem:eulpon} contradicts the fact that the signature of $S^{4k}$ must be zero. For completeness, we give here a brief sketch of this argument.
Let $M^{4k}$ be a connected, compact, smooth manifold of dimension $4k$. The singular cohomology cup product on $H^{2k}(M;\mathbb{R})$ and the Poincaré duality induce a symmetric bilinear form:
\begin{align}
\beta:H^{2k}(M;\mathbb{R})\times H^{2k}(M;\mathbb{R})&\rightarrow H^{4k}(M;\mathbb{R})\cong H^0(M;\mathbb{R})=\mathbb{R}\\
(x,y) &\mapsto x\cup y
\end{align}
By Poincaré duality, $\beta$ is non-degenerate. The \emph{signature} $\sigma(M)$ of $M$ is the signature of $\beta$.
In the following, we briefly describe a relation between the signature of $M$ and the Pontryagin classes of $TM$ which is known as the \emph{Hirzebruch Signature Theorem}. For a more systematic discussion see \cite[Chapter I]{HirTMA}.
Consider the power series of
\[
Q(z) = \frac{\sqrt{z}}{\tanh\sqrt{z}} = 1+\sum_{k=1}^\infty (-1)^{k-1} \frac{2^{2k}}{(2k)!} B_k z^k
\]
where $B_k$ are the \emph{Bernoulli numbers}. In particular, $B_k>0$ and $\neq \frac{1}{2}$ for all $k$. Let $p_1,\dots,p_m$ be variables for $m \in \mathbb{N}$. By the fundamental theorem of symmetric polynomials there are variables $\beta_1,\dots,\beta_m$ such that
\[
1+p_1z+p_2z^2+\ldots+p_mz^m = \prod_{i=1}^m (1+\beta_i z),
\]
i.e.\ we may consider $p_i$ as elementary symmetric polynomials in $\beta_1,\dots,\beta_m$. Now consider the coefficient of $z^k$ in the product $\prod_{i=1}^mQ(\beta_i z)$. Clearly it is a symmetric polynomial in $\beta_1,\ldots,\beta_m$, homogeneous of weight $k$. Therefore it can be expressed as a polynomial $L_k^m(p_1,\dots,p_k)$ in a unique way and moreover it does not depend on $m$ for $k\leq m$. Consequently we write $L_k\;\shortstack{{\tiny\upshape def}\\=}\; L^m_k$ for $k\leq m$. The coefficient $s_k$ of $p_k$ in $L_k$ is determined by
\[
\sum_{k=0}^\infty s_k z^k = \frac{1}{2} + \frac{1}{2}\frac{2\sqrt{z}}{\sinh 2\sqrt{z}}
\]
hence (see \cite[page 12]{HirTMA})
\[
s_0 =1,\quad s_k = \frac{2^{2k}(2^{2k-1}-1)}{(2k)!}B_k,\; k\geq 1
\]
The first polynomials $L_k$ are given by
\[
L_1 = \frac{1}{3}p_1, \quad L_2=\frac{1}{45}\left( 7p_2 -p_1^2\right), \quad L_3 = \frac{1}{3^2\cdot 5\cdot 7}\left(62p_3 -13p_2p_1 +2p_1^3\right),\quad \dots
\]
\begin{dfn}
The \emph{$L$-class} of $M^{4k}$ is the cohomology class
\[
L_k(p_1,\dots,p_k) \in H^{4k}(M^{4k};\mathbb{Q}),
\]
where $p_i = p_i(TM^{4k})$.
\end{dfn}
\begin{thm}\label{thm:signa}\cite[Hirzebruch Signature Theorem]{HirTMA}
Let $M^{4k}$ be a compact and oriented manifold and let $[M] \in H_{4k}(M)$ be the fundamental class of $M$. Then we have
\[
\sigma(M) = \langle L_k(p_1,\dots,p_k), [M]\rangle
\]
In particular $L_k(p_1,\dots,p_k)$ is integral.
\end{thm}
\emph{Alternative proof of theorem \ref{thm:hirzebruchs4k}.}
Since the cohomology of $S^{4k}$ is trivial except in dimensions $0$ and $4k$, all Pontryagin classes of $TS^{4k}$ are zero, except possibly the top class $p_k$, hence $L_k(S^{4k})=s_kp_k$.
Moreover the signature $\sigma(S^{4k})$ vanish, since $H^{2k}(S^{4k};\mathbb{Z})=0$.
Combining lemma \ref{lem:eulpon} with the signature formula in theorem \ref{thm:signa}, we obtain a contradiction as follows
\[
0 =\sigma(S^{4k})=\langle s_kp_k,[S^{4k}]\rangle=(-1)^k 2\langle s_k e(T(S^{4k})),[S^{4k}]\rangle=(-1)^k 2s_k\chi(S^{4k})=(-1)^k 4s_k \neq 0\hfill\qed
\]
\section{K-Theory, Bott Periodicity Theorem and the Chern character}
We start with the definition of topological $K$-theory. Let $B$ be a compact CW-complex and let $F(B)$ denote the free abelian group of isomorphism classes of complex vector bundles over $B$ (in the following the expression \emph{vector bundle} shall always mean complex vector bundle). We define $K(B)$ as the quotient qroup of $F(B)$ by the subgroup generated by elements of the form $[E\oplus F] - [E] -[F]$, where $E,F$ are vector bundles over $B$ and $[E]$ is the isomorphism class of $E$. It is easy to see that vector bundles $E\to B$ and $F \to B$ represent the same class in $K(B)$ if and only if $E$ and $F$ are \emph{stably isomorphic}, i.e.\ there is an $n \in \mathbb{N}$ and a trivial bundle $\trivial{n} \to B$ of rank $n$ such that $E\oplus \trivial{n}$ is isomorphic to $F \oplus \trivial{n}$. Finally we will denote with $x_E$ the stable class of $E$ in $K(B)$.
In the following paragraphs we would like to present some properties of $K(B)$ which will be needed in the proof of Theorem \ref{thm:main}.
\subsection{Ring structure of $\mathbf{K(B)}$}
By construction $K(B)$ is an abelian group with respect to direct sums of vector bundles. The zero element is the trivial vector bundle of rank $0$. In addition $K(B)$ has also a multiplication coming from tensor product of vector bundles. More precisely, for elements in $K(B)$ represented by vector bundles $E$ and $F$, their product in $K(B)$ will be represented by $E\otimes F$. We extend this definition distributively to all elements of $K(B)$ and we obtain a well-defined multiplication on $K(B)$. This makes $K(B)$ into a commutative ring with unit (where the unit is represented by the trivial line bundle $\trivial{1} \to B$).
If $B'$ is another compact CW-complex and $f \colon B \to B'$ a continuous map, then it induces a ring homomorphism $K(f):K(B') \to K(B)$ by pulling back vector bundles. Hence $K$ can be regarded as a contravariant functor from the category of compact CW-complexes into the category of commutative rings.
\subsection{Bott periodicity in $K$-theory}
Let $B$ and $B'$ be compact CW-complexes. There is an \emph{external product} $\mu:K(B)\otimes K(B') \to K(B\times B')$, where $K(B)\otimes K(B')$ is the tensor product of rings, which is a ring again. If $a \otimes b \in K(B) \otimes K(B')$ then
\[
\mu(a\otimes b) = p_B^*(a)p_{B'}^*(b)
\]
where $p_B$ and $p_{B'}$ are the projections of $B\times B'$ onto $B$ and $B'$ respectively. We extend $\mu$ linearly to all elements of $K(B) \otimes K(B')$. Taking $B'=S^2$ we obtain
\begin{thm}[Bott periodicity, cf. Corollary 2.2.3 in \cite{AtiKTh}]\label{thm:BottPeriodicity}
The homomorphism
\[
\mu \colon K(B) \otimes K(S^2) \to K(B \times S^2)
\]
is an isomorphism of rings.
\end{thm}
\subsection{The Chern character}
Let $L \to B$ and $L' \to B$ be two complex line bundles. The first Chern class is an isomorphism from the group of isomorphism classes of line bundles over $B$ \footnote{where the multiplication is defined by the tensor product of line bundles} to $H^2(B;\mathbb{Z})$, in particular $c_1(L\otimes L') = c_1(L) + c_1(L')$. In the same way, we would like to have a natural ring homomorphism from $K$-theory to (ordinary) cohomology. The product formula for line bundles suggest that we take the exponential of the first Chern class to preserve the multiplication of $K(B)$ into cohomology. Hence we define the \emph{Chern character of a line bundle} to be
\[
\mathop{ch}(L)=e^{c_1(L)} = 1 + c_1(L) +\frac{c_1(L)^2}{2!} + \frac{c_1(L)^3}{3!}+ \ldots \in H^*(B;\mathbb{Q})
\]
and since $B$ was assumed to be compact, the formal sum will be finite. Clearly we obtain
\[
\mathop{ch}(L\otimes L')=e^{c_1(L\otimes L')} = e^{c_1(L) + c_1(L')} = e^{c_1(L)}e^{c_1(L)} = \mathop{ch}(L)\cup \mathop{ch}(L').
\]
We extend this definition linearly to direct sums of line bundles, i.e. if $E=L_1\oplus\ldots\oplus L_n$, then
\[
\mathop{ch}(E) = \sum_i \mathop{ch}(L_i) = n + (t_1+\ldots +t_n) + \ldots + (t_1^k+\ldots+t_n^k)/k! + \ldots
\]
where $t_i = c_1(L_i)$. In this notation, the total Chern class is given by $c(E)=(1+t_1)(1+t_2)\cdot\ldots\cdots(1+t_n) = 1 + \sigma_1 + \ldots + \sigma_n$, where $\sigma_j = c_j(E)$ is the $j$-th elementary symmetric polynomial in the $t_i$'s. Moreover, if $\nu_k$ denotes the $k$-th Newton polynomial (cf.\ \cite{MiSChC}) it satisfies the relation
\[
t_1^k+\ldots+t_n^k = \nu_k(\sigma_1,\ldots,\sigma_k) =\nu_k(c_1(E),\ldots,c_k(E)).
\]
Hence the \emph{Chern character of $E$} becomes
\begin{equation}
\mathop{ch}(E) = \dim E + \sum_{k> 0}\frac{1}{k!}\nu_k(c_1(E),\ldots,c_k(E)).
\end{equation}
Since the right hand side depends only on the vector bundle $E$, we take this as a definition for the Chern character for an arbitrary vector bundle over $B$. Using the \emph{splitting principle} it is easy to show
\begin{prp}[cf. Proposition 4.2 in \cite{HatVBK}]\label{prop:chSumsAndProducts}
For complex vector bundles $E$ and $E'$ over $B$ we have
\begin{enumerate}
\item $\mathop{ch}(E\oplus E') = \mathop{ch}(E) + \mathop{ch}(E')$,
\item $\mathop{ch}(E\otimes E') = \mathop{ch}(E) \cup \mathop{ch}(E')$.
\end{enumerate}
\end{prp}
Note that for Chern classes we have the identity $c(E\oplus\trivial n) = c(E)$, hence $c_k(E \oplus\trivial n) = c_k(E)$ for all $k$. It follows that the Chern character is well-defined on the stable class $x_E \in K(B)$ of $E$ and therefore it descends to a map $\mathop{\mathrm{ch}}\colon K(B) \to H^*(B;\mathbb{Q})$. From Proposition \ref{prop:chSumsAndProducts} we obtain
\begin{prp}\label{prop:chRingHom}
The Chern character $\mathop{\mathrm{ch}}\colon K(B) \to H^*(B;\mathbb{Q})$ is a ring homomorhism.
\end{prp}
\begin{rem}\label{rem:naturality}
Firstly, let $\ch{B}$ denote the image of $\mathop{\mathrm{ch}}\colon K(B) \to H^*(B;\mathbb{Q})$. We call $\ch{B}$ \emph{integral} if $\ch{B} \subset H^*(B;\mathbb{Z})$. Secondly, let $B'$ be another compact CW-complex and $f:B \to B'$ a continuous map, then by the naturality of Chern classes we obtain the naturality of the Chern character, i.e.\ for $y \in K(B')$ we have $f^*(\ch{y}) = \ch{K(f)(y)}$.
\end{rem}
\section{Proof of Theorem \ref{thm:main} }
We start with an observation that $\ch{S^2}$ is integral. For if $E \to S^2$ is a complex vector bundle then $\ch{x_E}= n + c_1(E) \in H^*(S^2;\mathbb{Z})$. Moreover we have
\begin{prp}\label{prop:ch(S2n)isIntegral}
$\ch{S^{2n}}$ is integral.
\end{prp}
\begin{proof}
Let $f \colon S^2 \times\ldots\times S^{2} \to S^{2n}$ be a map of degree $1$, where the cartesian product is taken $n$ times. Then the induced map on cohomology $f^* \colon H^*(S^{2n}
) \to H^*(S^2\times\ldots\times S^2)$ is an monomorphism (if the coefficient group in $H^*$ is not specified, it is assumed to be the rationals) . Using the Künneth-formula we also have an isomorphism of rings $H^*(S^2\times\ldots\times S^2) \cong H^*(S^2)\otimes\ldots\otimes H^*(S^2)$, where the tensor product is taken $n$ times.
Moreover by the Bott periodicity theorem, cf. Theorem \ref{thm:BottPeriodicity}, the map $\mu$ induces inductively an isomorphism
\[
K(S^2\times\ldots\times S^n) \to K(S^2)\otimes\ldots\otimes K(S^2).
\]
and the following diagramm commutes
\begin{center}
\begin{tikzpicture}[scale=2]
\node (A) at (0,1) {$K(S^2\times\ldots\times S^2)$};
\node (B) at (4,1) {$K(S^2)\otimes\ldots\otimes K(S^2)$};
\node (C) at (0,0) {$H^*(S^2\times\ldots\times S^2)$};
\node (D) at (4,0) {$H^*(S^2)\otimes\ldots\otimes H^*(S^2)$.};
\path[->,font=\scriptsize,>=angle 90]
(A) edge node[above]{Bott periodicity} (B)
(A) edge node[right]{$\mathop{\mathrm{ch}}$} (C)
(B) edge node[right]{$\mathop{\mathrm{ch}}\otimes\ldots\otimes \mathop{\mathrm{ch}}$} (D)
(C) edge node[above]{Künneth isomorphism} (D);
\end{tikzpicture}
\end{center}
Since $\ch{S^2}$ is integral, it follows that the image of $\mathop{\mathrm{ch}}\otimes\ldots\otimes \mathop{\mathrm{ch}}$ lies in $H^*(S^2;\mathbb{Z})\otimes\ldots\otimes H^*(S^2;\mathbb{Z})$. Hence $\ch{S^2\times\ldots\times S^2}$ has to be integral.
The naturality of Chern character (cf. Remark \ref{rem:naturality}) implies that the diagramm
\begin{center}
\begin{tikzpicture}[scale=2]
\node (A) at (0,1) {$K(S^{2n})$};
\node (B) at (2,1) {$K(S^2\times\ldots\times S^2)$};
\node (C) at (0,0) {$H^*(S^{2n})$};
\node (D) at (2,0) {$H^*(S^2\times\ldots\times S^2)$.};
\path[->,font=\scriptsize,>=angle 90]
(A) edge node[above]{$K(f)$} (B)
(A) edge node[right]{$\mathop{\mathrm{ch}}$} (C)
(B) edge node[right]{$\mathop{\mathrm{ch}}$} (D)
(C) edge node[above]{$f^*$} (D);
\end{tikzpicture}
\end{center}
also commutes. From above we have that $\ch{S^2\times\ldots\times S^2}$ is integral and since $f^*$ is an isomorphism we conclude that $\ch{S^{2n}}$ has to be integral too.
\end{proof}
Suppose now, there is a complex vector bundle $E$ of rank $n$ such that $E_\mathbb{R} = TS^{2n}$. Then, due to the cohomology of $S^{2n}$ all Chern classes have to be zero except possibly $c_n(E)$. From the recursive formula for Newton polynoms, cf.\cite[p.195]{MiSChC}, it follows that only $\nu_n$ is not zero and is given by $\nu_n = \pm n c_n(E)$. Hence $\ch{x_E}=n \pm \frac{c_n(E)}{((n-1)!)}$ and therefore
\[
\frac{c_n(E)}{(n-1)!} \in H^{2n}(S^{2n};\mathbb{Z})\cong\mathbb{Z}.
\]
Since $\langle c_n(E),[S^{2n}]\rangle = \langle e(E_\mathbb{R})),[S^{2n}]\rangle = \chi(S^{2n})= 2$, we conclude that $(n-1)!$ has to divide $2$, hence $n\leq 3$.
This completes the proof of Therem \ref{thm:main}.
\begin{rem}
The proof of Theorem \ref{thm:main} uses only the homotopy type of $S^{2n}$. By the generalized Poincare conjecture every manifold homotopy equivalent to $S^{2n}$ is homeomorphic to $S^{2n}$. But since Milnor \cite{MilMHS} it is known that a manifold can be homeomorphic but not diffeomorphic to the standard sphere $S^{2n}\subset\mathbb{R}^{2n+1}$. These manifolds are called \emph{exotic spheres}. Thus Theorem \ref{thm:main} is also true if $S^{2n}$ is replaced by an exotic sphere.
\end{rem}
|
\section{Introduction}
The pair contact process with diffusion (PCPD) is an interacting particle system
with diffusion, pair annihilation ($2A \rightarrow 0$), and creation by pairs ($2A \rightarrow 3A$),
which exhibits an absorbing phase transition.
The PCPD was first studied by Grassberger in 1982~\cite{G1982},
but it had not attracted much interest until Howard and T\"auber~\cite{HT1997}
introduced a so-called `bosonic' PCPD.
Since $2A \rightarrow 3A$ dynamics in the bosonic PCPD drives particle density in the active phase to
infinity in finite time, a more controllable model was desired to study the phase transition.
Carlon et al.~\cite{CHS2001} introduced a present form of the PCPD on a lattice with hard core exclusion
to keep the density from diverging.
The conclusion in Ref.~\cite{CHS2001} was already
controversial, which has triggered a lot of numerical and analytical studies of the PCPD~\cite{H2001,O2000,PHK2001,H2001_291,PK2002,DdM2002,KC2003,O2003,BC2003,JvWDT2004,PHP2005a,PHP2005b,H2006,PP2006,KK2007,SB2008,PP2009,SB2012,GCDD2014,SCP2014Cde}.
It turned out that the PCPD, especially in one dimension, is influenced by strong corrections to scaling.
This, in turn, makes it difficult for numerical studies to lead to a consensus about the universality class
of the one dimensional PCPD. Accordingly, many scenarios had been suggested
in the early stage of research (for a review of early discussions, see~\cite{HH2004,PP2008EPJB}).
The controversy remains unabated, but by now only two competing theories have survived.
In one theory, the PCPD is claimed to form a different universality class from the directed percolation (DP)
universality class~\cite{KC2003,PHP2005a,PHP2005b,PP2006,PP2009,SCP2014Cde}.
This theory is supported by the following facts:
First, the upper critical dimension of the PCPD is 2~\cite{OMS2002}, while that of the DP is 4.
Second, diffusion bias changes the universality class of the one-dimensional PCPD, whereas
such a bias can be asymptotically removed by a Galilean transformation in the DP class~\cite{PHP2005a}.
Third, there are nontrivial crossover behaviors between the PCPD and the DP models~\cite{PP2006,PP2009}.
In the other theory, the PCPD is claimed to belong to
the DP class~\cite{BC2003,H2006,SB2008,SB2012}.
An argument to support this theory was put forward by Hinrichsen~\cite{H2006},
which will be called {\em the DP argument} for later reference.
Since the DP argument has motivated the present work, we repeat it here for completeness.
The DP argument is based on a numerical observation that the dynamic exponent $z$ of the PCPD
in one dimension is smaller than 2. Since critical clusters spread as $t^{1/z}$ and isolated particles
spread diffusively as $t^{1/2}$, a comparison of these two scales
suggests that the critical spreading will eventually dominate the critical behavior over
the diffusive spreading
and, accordingly, diffusion is irrelevant in the sense of the renormalization group (RG).
Although it is in principle possible for a pair to be formed purely by diffusion of isolated particles,
the irrelevance of diffusion implies that such events will hardly occur in the long time limit.
Thus, it does not matter whether a particle can diffuse or not, as far as critical phenomena are concerned.
If particles are not allowed to diffuse, the model is the pair contact process
(PCP)~\cite{J1993}, which has infinitely many absorbing states (IMAS) and belongs to the DP class.
In this sense, diffusion can at best affect corrections to scaling
and the one-dimensional PCPD should exhibit the same universal behavior as the PCP or,
in the context of universality, as the DP.
The aim of this paper is to figure out if diffusing particles in the PCPD are really indistinguishable in the RG sense from the frozen isolated particles of the PCP, as claimed in the DP argument.
If isolated particles can be regarded as frozen particles of the PCP, introducing
dynamics which make diffusing particles frozen should not yield any other singular behavior than
the typical DP critical behavior. Motivated by this idea, we introduce and study numerically
a variation of the PCPD in which
diffusing particles can mutate irreversibly to immobile species with rate $w$.
The case with $w=0$ will correspond to the PCPD and
the case with nonzero $w$, which includes the PCP ($w=1$), has IMAS. The exact definition of
the model will be given in Sec.~\ref{Sec:model}.
We will present simulation results of the model in Sec.~\ref{Sec:results}, focusing on the existence
of a crossover around $w=0$.
Section~\ref{Sec:sum} summarizes and concludes this work.
\section{\label{Sec:model}Model}
This section introduces a one-dimensional two-species lattice model
with periodic boundary conditions.
The system size will be denoted by $L$, and each species will
be denoted by $A$ and $B$, respectively.
Each site can be one of the three possible states; $A$-occupied, $B$-occupied, and vacant states.
No multiple occupancy is allowed.
The system evolves stochastically under the following rules:
\begin{align}
\label{Eq:rules}
&\varnothing A \stackrel{w/2}{\longrightarrow} \varnothing B,\quad
A \varnothing \stackrel{w/2}{\longrightarrow} B\varnothing,\quad
\varnothing A \stackrel{D}{\longleftrightarrow}A \varnothing,\\
&XY \stackrel{p}{\longrightarrow} \varnothing \varnothing,\quad
XY\varnothing \stackrel{\sigma}{\longrightarrow} XYA,\quad
\varnothing XY \stackrel{\sigma}{\longrightarrow} AXY,
\nonumber
\end{align}
where $\varnothing$ stands for a vacant site; $A$ ($B$) represents a site occupied
by species $A$ ($B$); $X$ and $Y$ can be any of $A$ and $B$; and the variables
above the arrows mean the transition rates of the corresponding events.
In all simulations in this paper, we set $\sigma = (1-p)/2$ and $D=(1-w)/2$
with two parameters $w$ and $p$ ($0\le w \le 1$ and $0 \le p \le 1$).
We always use the configuration with all sites occupied by species $A$ as an
initial condition and all discussions are based on this initial condition.
The extension to higher dimensions is straightforward, but most discussions in this paper
are limited to the one dimensional model.
For the case with $w=0$, species $B$ cannot be generated by the rules, so this
case is identical to the PCPD studied in Ref.~\cite{PHP2005a}.
When $w$ is strictly positive, $B$ particles can appear by the dynamics and
any configuration only with isolated $B$ particles are absorbing.
Hence, the model with nonzero $w$ has infinitely many absorbing states, which is characterized
by the nonzero density of $B$ species in all phases.
Following the convention, we will call the density of $B$ species, denoted by $\rho_b$, a nonorder field.
In particular, if $w=1$, isolated $A$ particles can at best mutate to $B$ particles, so
the two species play the exactly same role as particles in the PCP~\cite{J1993}.
Since our main concern is the regime where $0\le w \ll 1$,
we will call this model the PCPD with a nonorder field (PNF).
To simulate the model, we made a list of active pairs. By an active pair we mean
two consecutive sites that take one of 6 forms, $\varnothing A$, $A\varnothing $,
$AA$, $AB$, $BA$, and $BB$. The list contains the information as to where active pairs are
located. The size of the list at time $t$ is denoted by $N_t$.
At $t=0$ when all sites are occupied by $A$, the size of the list is the same as the system
size $L$. At time $t$, we choose one of the active pairs at random with equal probability.
If only one site is occupied (by $A$ by definition) in the chosen active pair, this particle mutates
to $B$ with probability $w/2$ or hops to the empty site inside the chosen active pair with probability
$D = (1-w)/2$. With probability $\frac{1}{2}$, however, nothing happens. If
both sites of the chosen active pair are occupied,
with probability $p$ these two sites become vacant ($XY \rightarrow
\varnothing \varnothing$) or with probability $1-p$ there will be an attempt to branch an $A$ particle to one of its
nearest neighbors, to be called a target site, which is chosen at random with equal probability.
If the target site is empty,
$A$ particle will be placed there. Otherwise, nothing happens. After the above step,
time increases by $1/N_t$. The above procedure will continue until
either the system falls into one of the absorbing states or time exceeds
the preassigned observation time.
In all simulations whose results will be presented in the next section, however,
the system size is so large that no simulation ends up with an absorbing state at the
end of the preassigned observation time.
We are mainly interested in the behavior of pair density defined as
\begin{equation}
\rho_p(t) \equiv \frac{1}{L} \sum_{i=1}^L \left \langle
\left (1-\delta_{X_i(t),\varnothing}\right )
\left (1-\delta_{X_{i+1}(t),\varnothing}\right )\right \rangle,
\end{equation}
where $X_i(t)$ means the state (one of $A$, $B$, and $\varnothing$)
at site $i$ at time $t$,
$\delta$ is the Kronecker delta symbol, and $\langle \ldots \rangle$ stands for
the average over all realizations.
The order parameter is then defined as
\begin{align}
\rho_s = \lim_{t\rightarrow \infty} \lim_{L\rightarrow \infty} \rho_p(t),
\label{Eq:order_param}
\end{align}
which is zero (nonzero) in the absorbing (active) phase.
Note that $\rho_s$ defined in Eq.~\eqref{Eq:order_param} is the order parameter
used in both the PCP and the PCPD.
Before presenting simulation data, we would like to argue that the PNF with nonzero
$w$ should belong to the DP class, mimicking the DP argument with a slight modification.
As long as $w$ is nonzero, an isolated $A$ particle can mutate to $B$ within finite time of $O(1/w)$
for any $p$. Since the correlation time diverges at the critical point,
diffusion of $A$ particles which should be terminated in finite time is irrelevant
in the RG sense.
Thus, the critical behavior of the PNF with nonzero $w$ cannot be distinguishable from the PCP which
belongs to the DP class.
Note that, in the original DP argument, the dynamic exponent has to be assumed to be smaller than 2, which
makes the extension to higher dimensional systems problematic.
In our argument, however, we do not have to assume the value of the dynamic
exponent, so it can also be applicable to higher dimensions.
\section{\label{Sec:results}Results}
This section presents simulation results of the PNF defined by the rules in Eq.~\eqref{Eq:rules}.
We use $p$ as a tuning parameter with $w$ fixed and
the critical point will be denoted by $p_c(w)$.
The system size is $L=2^{20}$ and
no realization has ended up with an absorbing state within the observation time,
which minimally guarantees that the finite size effect is not important.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig1.eps}
\caption{\label{Fig:DP1e4} (Color online) Semilogarithmic plots of $\rho_p(t) t^\delta$ vs. $t$ around the
critical point for $w=10^{-4}$, where $\delta = 0.1595$ is the critical decay exponent of the DP.
For $p=0.133~05$ (middle curve), the curve becomes flat from $t=10^5$ for about
three log decades and the other curves veer up ($p=0.133~045$) or down ($p=0.133~055$).
Thus, we conclude that the critical point is $p_c=0.133~050(5)$ and
the model belongs to the DP class.
}
\end{figure}
\begin{table}[b]
\caption{\label{Table:pc} Critical points of the PNF for various $w$'s. The numbers
in parentheses indicate the errors of the last digits.}
\begin{ruledtabular}
\begin{tabular}{llll}
$w$&$p_c(w)$&$w$&$p_c(w)$\\
\hline
0&$0.133~519(3)$\footnotemark[1]&0.1&0.103~635(5) \\
0.0001&0.133~050(5)&0.2&0.093~545(5) \\
0.0003&0.132~595(5)&0.3&0.087~872(2) \\
0.0008&0.131~83(1) &0.4&0.084~265(5) \\
0.001&0.131~580(5)&0.5&0.081~815(5) \\
0.002&0.130~54(1) &0.6&0.080~090(5) \\
0.004&0.128~92(1) &0.7&0.078~860(5) \\
0.008&0.126~45(1) &0.8&0.077~990(5) \\
0.01 &0.125~410(5)&0.9&0.077~410(5) \\
0.05 &0.112~56(1) &1&$0.077~0905(5)$\footnotemark[2]
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{From Ref.~\cite{PHP2005a}.}
\footnotetext[2]{From Ref.~\cite{PP2007}.}
\end{table}
We first show the behavior of $\rho_p(t)$ around the critical point for $w=10^{-4}$.
In Fig.~\ref{Fig:DP1e4}, we plot $\rho_p(t) t^{\delta}$ against $t$ around the critical point,
where $\delta = 0.1595$ is the critical decay exponent of the DP class.
In what follows, the symbol $\delta$ is reserved for the DP critical decay exponent.
The numbers of independent runs are 16, 24, and 24 for $p=0.133~045$, $0.133~05$, and $0.133~055$, respectively.
At $p=0.133~05$, the curve is almost flat for more than three log decades, while two
other curves veer up or down in the long time limit.
Hence, we conclude that this case indeed belongs to the DP class with $p_c = 0.133~050(5)$,
where the number in parentheses is the uncertainty of the last digit.
We also studied the critical behavior for various values of $w$ in a similar manner.
The resulting critical points are summarized in Table~\ref{Table:pc}.
Since both the PCP and the PNF with $0 < w < 1$ belong to the DP universality class,
there should not be a significant change due to the variation of $w$ when $(1-w) \ll 1$.
To make this point strong, we present the behaviors of $\rho_p(t)$ at the corresponding
critical points to various values of $w$ in Fig.~\ref{Fig:rp}.
Indeed, the critical decays of $\rho_p(t)$ for $w \ge 0.9$ are barely discernible.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig2.eps}
\caption{\label{Fig:rp} (Color online) Double-logarithmic plots of $\rho_p$ vs. $t$ at the critical points
for various $w$'s from $w=0$ to $w=1$ (bottom to top).
Three curves corresponding
to $w=0.5, 0.9$ and 1 are barely discernible, whereas difference among curves
corresponding to $w<0.1$ are conspicuous.
}
\end{figure}
A symptom of crossover is already observed in Fig.~\ref{Fig:rp} for $w \ll 1$.
When $w < 0.1$, $\rho_p(t)$'s at $p=p_c(w)$ transiently follow the PCPD behavior and then eventually deviate from the PCPD curve.
In particular, the time when the critical PNF starts to deviate from the PCPD curve increases as $w$ gets smaller and
there is no symptom of saturation up to $w = 10^{-4}$.
This behavior seems consistent with the crossover scaling ansatz~\cite{PP2006}
\begin{align}
\rho_p(t;\Delta,w) = t^{-\alpha} F_c\left ( \Delta w^{-1/\phi}, t w^{\nu_\|/\phi}\right ),
\end{align}
where $\Delta \equiv p_0 - p$ with $p_0$ to be the critical point of the model with $w=0$ (that is, the critical point of the PCPD), $\alpha$
is the critical decay exponent of the PCPD, $\nu_\|$ is the correlation time exponent of the PCPD,
$\phi$ is the crossover exponent, and $F_c$ is a scaling function.
Since the scaling ansatz suggests a power-law behavior of the phase boundary for small $w$ in such a manner that
\begin{equation}
\label{Eq:cross}
p_0 - p_c(w) \sim w^{1/\phi},
\end{equation}
we can estimate $\phi$ by analyzing the behavior of the phase boundary for
small $w$ without resorting to the values of $\alpha$ and $\nu_\|$.
Although $\alpha$ was estimated as $\approx 0.19$ in Ref.~\cite{PHP2005a},
which is larger than the more accurate estimate in Ref.~\cite{SCP2014Cde},
one should note that the accuracy of $p_0$ in Ref.~\cite{PHP2005a} was
attained without resorting much to the accuracy of $\alpha$; the critical
point was actually estimated by finding two $p$'s belonging to the absorbing and active phases, respectively.
Thus, we can safely use the value $p_0$ from Ref.~\cite{PHP2005a} to find the crossover exponent $\phi$.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig3.eps}
\caption{\label{Fig:phi} Phase boundary of the PNF in the $wp$-plane.
Symbols represent the critical points obtained by simulations (see Table~\ref{Table:pc}) and lines which connect two nearest symbols
are for guides to the eyes.
The phase boundary approaches the ordinate continuously with infinite slope, while it approaches the
PCP critical point (corresponding to $w=1$) with
finite slope. Inset: Plot of $p_0 - p_c$ vs. $w$ on a double logarithmic scale. For
comparison, Eq.~\eqref{Eq:cross} with $1/\phi \approx 0.62$ is also drawn.
}
\end{figure}
In Fig.~\ref{Fig:phi}, we depict the phase boundary of the PNF in the $wp$-plane.
The phase boundary approaches the ordinate with infinite slope,
which suggests that $1/\phi$ in Eq.~\eqref{Eq:cross}
is smaller than 1. The fitting of the phase boundary gives $1/\phi = 0.62(2)$; see the inset
of Fig.~\ref{Fig:phi}. Hence, we conclude that there is indeed a crossover behavior
for $w\ll 1$ and the time for the critical PNF to deviate from the critical decay of the PCPD
diverges as $w$ gets smaller.
Although the number of absorbing states in itself does not depend on $w$ for any $L$,
the so-called natural density~\cite{JD1993} of the nonorder field $\rho_b$
is expected to vary with $w$. Since the discontinuity of the natural density
of a nonorder field can give a crossover even within the same universality
class~\cite{PP2007}, we need to check if the natural density changes abruptly at $w=0$.
At the critical point, the density $\rho_b(t)$ of $B$ species is expected to have
asymptotic behavior as~\cite{OMSM1998}
\begin{align}
\rho_b(t) -\rho_B(w) \sim t^{-\delta},
\label{Eq:rhob}
\end{align}
where $\rho_B(w)$ is the natural density at the critical point for given $w$.
If we plot $\rho_b(t)$ as a function of $t^{-\delta}$,
a straight line is expected and the extrapolation of the straight line to $t\rightarrow \infty$
can give $\rho_B(w)$. In Fig~\ref{Fig:rhoB}, we show the result
of this extrapolation for $w= 10^{-4}$. Applying the same method,
we obtained $\rho_B$ for other values of $w$. In the inset of Fig.~\ref{Fig:rhoB},
we show how $\rho_B$ behaves with $w$. As can be easily guessed,
$\rho_B$ is an increasing function of $w$. Furthermore, it turns out that $\rho_B$
exhibits a power-law behavior as $\sim w^{\chi}$ with an estimate $\chi \approx 0.51$, which
suggests that $\rho_B$ is continuous at $w=0$.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig4.eps}
\caption{\label{Fig:rhoB} Plot of $\rho_b$ vs. $t^{-\delta}$ at the
critical point $p_c(w)$ for $w=10^{-4}$. Close to the ordinate, the curve becomes almost straight.
By a linear extrapolation using Eq.~\eqref{Eq:rhob}, we found the natural density $\rho_B$.
The fitting result is drawn as a straight line.
Inset: Plot of $\rho_B$ vs. $w$ on a double logarithmic scale. $\rho_B$ for small $w$ is fitted
by a power function $a w^\chi$ with a result $\chi \approx 0.51$.
}
\end{figure}
We now argue that $\chi$ is exactly 0.5.
First, we take the existence of nonzero $\rho_B(w)$ for granted.
Now, consider a situation where an isolated $A$ particle appears (by the given dynamics) in the sea of $B$ particles.
This isolated $A$ particle will either mutate to $B$ in time of $O(1/w)$ or meet another $B$
particle in time of $O(1/\rho_B^2)$ ($\sqrt{Dt} \sim 1/\rho_B$ with $1/\rho_B$ to be the order of mean distance between $A$ and the nearest $B$ particle in this situation). If $\rho_B \ll \sqrt{w}$,
the chance of meeting a $B$ particle before mutation occurs is very low.
Thus, we can regard isolated $A$ particles as frozen $B$ particles.
In this case, $\rho_B$ is purely determined
by the dynamics of pairs and should not depend on $w$, which contradicts to the assumption
$\rho_B \ll \sqrt{w}$.
If $\rho_B \gg \sqrt{w}$, diffusing $A$ particle will find a frozen $B$ particle before it mutates and
the newly formed pair begins so-called defect dynamics.
Since the system is at the critical point, the defect dynamics are eventually terminated and
the number of $B$ particles is likely to decrease by 1 at the end of the defect dynamics
when it is compared with the number of $B$ particles at the time a pair is formed by diffusion.
As the number of $B$ particles can only decrease as long as $\rho_B$ is finite,
$\rho_B$ should decrease to zero, which again is contradictory to the assumption $\rho_B \gg \sqrt{w}$.
Thus, a consistent conclusion is attainable only for $\rho_B \sim \sqrt{w}$.
Finally, we would like to check if mere presence of a nonorder field
is enough to result in a crossover even within the same universality class.
To this end, we slightly modify the PCP (that is, the PNF with $w=1$) in such a way that
whenever isolated particles appear by a pair annihilation,
each isolated particle is removed with probability $1-\varepsilon$ and
it remains there with probability $\varepsilon$.
Note that up to two isolated particles can appear
by a single pair-annihilation event (if two $A$ particles
in the middle of a local configuration $\varnothing AAAA \varnothing$ are removed by pair annihilation,
two $A$'s at the end of this cluster become isolated).
Once an isolated particle which appears right after a pair-annihilation event is decided not to disappear,
it remains frozen until another particle come to one of its nearest neighbors.
We use the same transition rates as in Eq.~\eqref{Eq:rules} for creation and annihilation events.
For convenience, we will call this model the modified PCP (MPCP).
The MPCP with $\varepsilon=1$ is just the PCP. When $\varepsilon=0$, there is no isolated particles and this
becomes a model with a single absorbing state (we take a configuration with every site occupied by a particle
as an initial state). Just like the PNF, the number of absorbing states does not depend
on $\varepsilon$, though the natural density of a nonorder field
should depend on $\varepsilon$. For any value of $\varepsilon$, MPCP is expected to belong to the DP
universality class.
Using the DP exponent $\delta$, we estimated critical points by studying the behavior
of $\rho_p t^\delta$ as in Fig.~\ref{Fig:DP1e4}.
The critical points are summarized in Table~\ref{Table:pc_D}.
To see whether there is a nontrivial crossover for small $\varepsilon$, we plotted
critical decay behavior for various $\varepsilon$'s in Fig.~\ref{Fig:pc_dp}.
Unlike the PNF around $w=0$, there is no symptom of a diverging time scale for $\varepsilon \ll 1$.
Furthermore, $p_c$ of the MPCP has finite slope at $\varepsilon = 0$ unlike
the PNF at $w=0$.
Thus, we conclude that mere introduction of a nonorder field is not sufficient to trigger a crossover.
\begin{table}[t]
\caption{\label{Table:pc_D} Critical points of the MPCP for various $\varepsilon$'s. The numbers
in parentheses indicate the errors of the last digits.}
\begin{ruledtabular}
\begin{tabular}{llll}
$\varepsilon$&$p_c$&$\varepsilon$&$p_c$\\
\hline
0 &0.064~162(1) &0.8 &0.074~046(1) \\
0.05&0.064~6925(5) &0.9 &0.075~5325(15) \\
0.1 &0.065~234(1) &0.95&0.076~303(1) \\
0.2 &0.066~3465(5) &1 &0.077~0905(5)$\footnotemark[1]$ \\
0.5 &0.069~957(1) &&
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{From Ref.~\cite{PP2007}.}
\end{table}
\section{\label{Sec:sum}Summary and discussion}
We studied a variation of the pair contact process with diffusion (PCPD)
by introducing a nonorder field assocated with infinitely many absorbing states; see Eq.~\eqref{Eq:rules}.
We called this model the PCPD with a nonorder field (PNF).
We analyzed absorbing phase transitions for various values of $w$.
When $w$ is nonzero, the PNF has infinitely many absorbing states (IMAS) and was found to belong
to the directed percolation (DP) universality class.
When $w$ is close to 1, critical decay of the pair density $\rho_p(t)$ does not show any significant
change from the pair contact process (PCP).
This is also manifest by the finite slope of the phase boundary at $w=1$; see Fig.~\ref{Fig:phi} around $w=1$.
On the other hand, a singular behavior representing a crossover appears when nonorder field is introduced
to the PCPD (the PNF with $w=0$). This crossover is described by the crossover exponent $\phi$,
whose numerical value was found as $1/\phi \approx 0.62$ from the analysis of the phase boundary at $w=0$
(see Fig.~\ref{Fig:phi} around $w=0$ and its inset).
We argued that the natural density $\rho_B$ at the critical point changes continuously as $\sqrt{w}$
when $w \ll 1$, which is consistent with simulation results.
To support that this crossover is originated from the difference of the universality classes,
we also presented simulation results of another model which belongs to the DP class irrespective of
whether the number of absorbing states is infinite or finite.
Unlike the PNF around $w=0$, no crossover was observed in this case.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig5.eps}
\caption{\label{Fig:pc_dp} (Color online) Double-logarithmic plots of $\rho_p$ vs. $t$ at the
critical point of the MPCP for various $\varepsilon$'s from $\varepsilon=0$ to $\varepsilon=1$ (top to bottom). As a guide to the eyes, a line segment with slope $-0.1595$ is
also drawn. Unlike the PNF, the critical behavior for small $\varepsilon$ is hardly
discernible from that for $\varepsilon=0$; see also Fig.~\ref{Fig:rp}.
Inset: Plot of $p_c$ as a function of $\varepsilon$ for the MPCP. Unlike the PNF, the slope at
$\varepsilon = 0$ is finite.
}
\end{figure}
In Refs.~\cite{PP2006,PP2009}, the existence of the nontrivial crossover from the PCPD to the DP
was invoked to be the evidence of the existence of the fixed point of the PCPD distinct from
the DP in one dimension. The same conclusion is also arrived at in this paper.
Furthermore, our work clarifies the difference between diffusing particles in the PCPD
and frozen particles in the PCP, unlike the anticipation from the DP argument.
It is worth while to discuss the role of a nonorder field which sometimes triggers
a crossover within the same universality class~\cite{PP2007} and sometimes not.
This difference can be understood as follow:
The nonorder field in this paper cannot play a role of an order parameter for all cases, whereas
the nonorder field in Ref.~\cite{PP2007} becomes an order parameter as soon as
crossover dynamics are introduced. That is, the crossover dynamics in Ref.~\cite{PP2007} make
the irrelevant nonorder field a relevant order parameter, which is the origin of
the crossover within the same universality class.
observed in Ref.~\cite{PP2007} reflects that a nonorder field becomes relevant in the RG sense
by crossover dynamics while absence of role change of nonorder field in this paper does not
give a nontrivial crossover behavior within the same universality class.
As a final remark, we would like to compare the three different crossovers in this work and in Ref.~\cite{PP2006,PP2009}
to the also three different crossovers from the directed Ising (DI) class to the DP class~\cite{BB1996, KHP1999, OM2008,PP2008PRE}. Note that three different mechanisms from the DI to the DP are qualitatively
identical to those from the PCPD to the DP. In this regard, the crossover in this paper seems different
from that in Ref.~\cite{PP2006}, though the values of crossover exponents are quite close; 0.62
in this paper and 0.57 in Ref.~\cite{PP2006}.
\begin{acknowledgments}
This work was supported by the Basic Science Research Program through the
National Research Foundation of Korea~(NRF) funded by the Ministry of
Science, ICT and Future Planning~(Grant No. 2014R1A1A2058694)
and by the Catholic University of Korea, research fund 2017.
The author furthermore thanks the Regional Computing Center of
the University of Cologne (RRZK) for providing computing time on the
DFG-funded High Performance Computing (HPC) system CHEOPS as well as support.
The author would also like to thank Korea Institute for Advanced Study (KIAS)
for its support and hospitality during his stay there on sabbatical leave (2016-2017).
\end{acknowledgments}
|
\section{Introduction}
\vspace{-3mm}
Quantum measurement theory, by which we mean the projection postulate and the formalism that follows from it, has an odd history; for decades physicists tried to avoid it, motivated by the curious philosophical issues that it raises~\cite{Schrodinger35, Schrodinger35t, Einstein35, Einstein36, Landau65, Gillespie86}. As experiments advanced, however, the action of measurements on individual systems in real-time became impossible to ignore, and quantum measurement theory was fully embraced in the quantum optics, quantum information, and related communities~\cite{mikeandike, WM10, Jacobs14}. In taking measurement theory on board, it appears that practitioners generally adopted the point of view that quantum measurements introduce into quantum mechanics some element that is otherwise absent.
Recent conversations with a number of physicists lead us to believe that there was, and still is, widespread confusion as to what this element is. Indeed it is not so surprising that such a confusion might exist, given the ways that measurements tend to be referred to in the literature~\footnote{The following passages taken from papers in the areas of fault-tolerant quantum computation and the quantum-to-classical transition illustrate the way that measurements are often referred to: ``... we make an effort to show that fault tolerance can be achieved without using classical operations or measurements. We do this because of two reasons. One is purely theoretical: One would like to know that measurements and classical operations are not essential and that the quantum model is complete in the sense that it can be made fault-tolerant within itself.''; ``... we need to add a new component, and that component will be measurement.''; ``... we find that continuous measurement can effectively obtain classical mechanics from quantum mechanics.''.}.
In fact, as far as dynamical processes are concerned --- that is, anything that can be achieved by any physical process or protocol --- measurements introduce nothing new, and are thus not ``outside'' of quantum mechanics. Any protocol that employs explicit measurements can be rewritten entirely in terms of unitary processes, and is thus contained within the theory of quantum mechanics. Similarly any experiment that uses measurements to generate some physical process can be reproduced by an experiment that uses only unitary processes (with the sole exception of a single measurement performed at the end of the process, a measurement that does not affect the dynamical description). If a physical measurement-based process is designed to produce a single outcome, then even the measurement at the end is unnecessary. These statements follow from the fact that the so-called ``measurement problem'' of quantum theory has no impact on the physical predictions of the theory. This does not mean that the measurement-based and unitary versions of an experiment will be equally easy to implement --- that depends on the available technology.
Nevertheless, practical considerations motivate problems that involve constraints. These constraints may distinguish in some way between measurements and unitary processes so as to render them inequivalent as far as the problem is concerned. But the only way to distinguish between the two, from a physical point of view, is to expand the definition of measurement so that a ``measurement'' does more than merely select out one of a set of mutually orthogonal subspaces. A prime example of a problem that involves practically-motivated constraints is fault-tolerant quantum computation (FTQC), in which quantum gates, the source of unitary operations, come with an error probability. In this case the definition of measurement is expanded to include \textit{amplification}. Amplification is the act of taking a set of states that have only small differences in energy and transforming them into a set of states in which the energy differences are much larger. The importance of amplification is that it allows the information encoded in the set of states to be stored and processed on a classical computer. The larger energy scale is what allows the classical computer to store and process information error-free. Since systems with a larger energy scale are harder to protect from decoherence, amplification is associated with a loss of coherence in a given basis, and this is of course entirely compatible with making measurements. (The fact that classical computers are always decohered in a given basis is the reason that one equates the conversion of information from a mesoscopic quantum device to a classical computer with a measurement process.) For quantum computation, since mesoscopic quantum gates must maintain coherence, it is natural to restrict these gates so that they cannot include amplification. Since measurement is now the sole provider of amplification, and amplification is a physical process (projection is not), measurement now brings a unique physical element to the problem and thus has a special role. This special role is to allow classical information (information stored in a given basis) to be processed error-free.
In fact, as we discuss below, amplification is not required for error-free classical processing. This can be achieved with mesoscopic (and thus error-prone) gates by using them to construct fault-tolerant circuits. In view of this, unitary circuits can provide all the properties of measurement-based protocols that we have considered so far: measurement, processing of the classical measurement results, and feedback consisting of operations applied to the computational qubits. Given that there may be technological advantages in implementing FTQC without measurements~\cite{DiVincenzo07, Paz10}, it is an interesting question as to precisely what effective measurement errors and feedback errors unitary circuits can achieve when used to replace measurements, for a given gate error. The above fact also begs the question as to why all high-threshold FTQC schemes developed to-date involve measurements. We devote some space in what follows to these two questions.
There is one further special property with which measurements (actually amplification) are commonly imbued in analyses of FTQC schemes. This special property is the assumption that macroscopic classical gates are much faster than quantum gates. That is, a mesoscopic quantum gate that performs a given classical logic operation (AND, OR, NOT, \textit{etc}) is assumed to take significantly longer than the equivalent macroscopic gate. This confers on measurements the ability to provide much faster classical processing than possible with unitary circuits. Although this assumption may be true of present-day devices, it is not clear that it is justified in the longer term. While the energy scale of macroscopic systems is (by definition) larger than that of mesoscopic quantum systems, and this difference might lead to a fundamental gap between the speed of mesoscopic and macroscopic gates, it is not clear that it does so. Certainly the fundamental timescale of mesoscopic superconducting qubits --- the separation of their energy levels --- is in the GHz range, and thus similar to the clock speed of classical electronics~\cite{Vijay12, Lecocq15, Kelly15, Ofek16, Huang14}.
In the following we explain in detail, and thus elucidate and justify, the various statements and claims made in the above discussion. We explain how measurement-based processes are rewritten as unitary processes, especially those involving post-selection for which this translation is less obvious. Two areas in which the relationship between measurement-based (those that employ amplification) and purely unitary (purely mesoscopic) protocols are most relevant are those of FTQC and quantum feedback control. For the former we discuss in some detail how purely mesoscopic schemes can be designed to reproduce the behavior of amplification-based schemes. For the latter we consider a range of open questions that are somewhat different from those posed by FTQC. We also elucidate the difference between the ``measurement problem'' and the question of the physical properties of measurements.
\vspace{-2mm}
\section{Physical processes}
\vspace{-2mm}
A specification of the way in which the state of a physical system changes with time is called a \textit{dynamical process}. We will refer to a dynamical process as a \textit{physical process}: the state of a system is a complete description of its physical properties, and so a physical process is something that describes how the physical properties of a system change with time. Quantum Mechanics is a probabilistic theory, so our physical processes must be able to describe random changes in the state of a system. We therefore define a physical process as follows: it is something that, given an initial state at time $t$, specifies all the possible states at a time $t' > t$, as well as the probability with which each will occur. Central to our discussion here is the fact that a \textit{physical process is completely defined by the final states and the probabilities of these states.} The question of \textit{which} of the final outcomes actually occurs is irrelevant for the purposes of specifying the process. Specifying which of the outcomes occurs is equivalent to simulating a single ``run'' of the process, but since one does not know which realization of the process will occur before the run, and all the possible realizations are already specified by the process, the mere act of simulating a run does not provide any further information about the physical behavior.
In quantum mechanics a physical process, which is generated by the fundamental forces (the laws of physics), is specified by a unitary transformation, along with a partitioning of Hilbert space into a set of mutually orthogonal subspaces. The subspaces define the mutually exclusive outcomes, and the probabilities of these outcomes are given by the Born rule: they are the square moduli of the projections of the final state of the system onto each of the subspaces. The unitary transformation and the set of subspaces completely determines the physical process. In fact, to be more precise, the process is already fully defined by the unitary transformation alone (and thus by the laws of physics): this unitary transformation tells us, for every way in which we may choose to break up Hilbert space into a set of mutually exclusive outcomes, the probabilities of these outcomes.
To look ahead a little, quantum measurement theory is the machinery --- consisting essentially of sets of projectors --- that actually performs the projections onto the mutually exclusive subspaces, and thus picks realizations of the physical process. The ``use of measurements'' in any protocol or process means merely that one is explicitly introducing projection operators so as to describe the process in terms of an individual realization (the state projected onto one of the subspaces). The operation of the protocol is then described as an explicit stochastic process with classical probabilities. One can alternatively remove the projection operators from the analysis, so that the evolution is described in the full Hilbert space that contains all the possibilities simultaneously. In that case no explicit classical probabilities need be involved, except to record the probabilities of the final outcomes of the protocol (if needed). But in both cases the physical process is the same: the only role of measurement is to select out a subspace and thus pick a realization. For this reason, from a fundamental point of view, all protocols that involve measurements are automatically also protocols that are purely unitary. The two kinds of protocols are inseparable.
\vspace{-2mm}
\subsection*{Where is the ``measurement problem''?}
\vspace{-2mm}
The ``measurement problem'' arises from the fact that the unitary evolution of quantum mechanics can, in general, place systems in superpositions of any chosen basis of orthogonal states. Since mutually orthogonal states can be distinguished by an observer, upon doing so an observer must find the universe to be in one of these states, and not a superposition. However, the evolution of quantum mechanics does not provide any dynamical process (any evolution via a fundamental force) that ``collapses'' a superposition of a set of states to one of these states when a measurement is made. As a result measurement is described merely as a process in which the possible states of an observer become correlated with those of the system being measured. The Born rule, which connects quantum states to our observed reality, says that the observer will find the universe to be in one of a set of possible orthogonal states and determines the probability for each one. Because the observer's possible states are correlated with a set of states of the system, if we select one of these states as reality we necessarily select one state for the system, and the observer will find the system to be in that state.
Thus quantum theory determines the possible states that will be observed, and the respective probabilities for these states, but provides no mechanism where by one of the states actually becomes reality, since the joint system of the observer, the system, and the rest of the universe continues to exist in a superposition of all the possibilities. This description of the universe is at variance with our everyday experience: quantum mechanics asserts that the universe is \textit{actually} in a superposition of a certain set of possible orthogonal states, while we \textit{actually} observe the universe to be in just one of these states. As pointed out above, however, this conundrum (usually referred to as the ``quantum measurement problem'') is philosophical, not physical, since it is separate from the question of the predictions of the theory as to the dynamics of physical processes: a physical process is defined only by the set of possible states and their respective probabilities. We are therefore not concerned in this article with the quantum measurement problem.
\vspace{-2mm}
\section{Unitary implementations of measurement-based feedback}
\vspace{-2mm}
\label{uimp}
To gain insight into how measurement-based protocols map into unitary protocols it is useful to consider a simple measurement-based feedback procedure. A measurement on a quantum system $\mathcal{S}$, whose state is given by the density matrix $\rho$, is described by a set of ``measurement operators'' $\{A_n\}$, in which the final state of the system for outcome $n$ is given by
\begin{equation}
\tilde{\rho}_n = \frac{A_n \rho A_n^\dagger}{P_n} ,
\end{equation}
where
\begin{equation}
P_n = \mbox{Tr}[A_n^\dagger A_n \rho ]
\end{equation}
is the probability that this outcome occurs. Once an observer has obtained the measurement result (the value of $n$), then he or she can apply a unitary transformation to the system that depends on this result. Denoting each of these unitary operations by $U_n$, the complete dynamics resulting from the measurement and the ``feedback'' operation is
\begin{equation}
\tilde{\rho}_n' = \frac{U_n A_n \rho A_n^\dagger U_n}{P_n} .
\end{equation}
We can implement the above measurement and feedback operation unitarily using a second ``auxiliary'' quantum system, $\mathcal{A}$, with dimension $n$. Denoting a basis for $\mathcal{A}$ by $\{ |n\rangle_{\ms{a}} \}$, we first prepare $\mathcal{A}$ in the state $|0\rangle_{\ms{a}}$, and then apply a joint unitary $V$ to $\mathcal{S}$ and $\mathcal{A}$. By choosing $V$ appropriately we can arrange that the final joint state of the two systems is
\begin{equation}
\sigma = V (\rho \otimes |0\rangle_{\ms{a}} \langle 0 |_{\ms{a}} )= \sum_{n,m} A_n \rho A_m^\dagger \otimes |n\rangle_{\ms{a}} \langle m |_{\ms{a}} .
\end{equation}
The relationship between the set of operators $\{ A_n\}$ and the required unitary $V$ is quite simple (see, e.g.~\cite{Jacobs14}). The state $\sigma$ is one for which the system $\mathcal{S}$ is in the state $\tilde{\rho}_n = A_n \rho A_n^\dagger/P_n$ when $\mathcal{A}$ is in the state $|n\rangle$ . To see this, first we separate out the diagonal terms in the sum:
\begin{equation}
\sigma = \sum_{n} A_n \rho A_n^\dagger \otimes |n\rangle_{\ms{a}} \langle n |_{\ms{a}} + \sum_{k\not= m} A_k \rho A_m^\dagger \otimes |k\rangle_{\ms{a}} \langle m |_{\ms{a}}.
\end{equation}
If we want to examine the state of $\mathcal{S}$ when $\mathcal{A}$ is in state $|n\rangle_{\ms{a}}$, then we need to examine the joint density matrix restricted to the subspace in which $\mathcal{A}$ is in state $|n\rangle_{\ms{a}}$. This is given by the term in the expression for $\sigma$ above that contains only the outer product $|n\rangle_{\ms{a}} \langle n|_{\ms{a}}$. The state of the system in this subspace is the coefficient of $|n\rangle_{\ms{a}} \langle n|_{\ms{a}}$, and is thus $A_n \rho A_n^\dagger$. This state is not normalized, however. To normalize it we divide it by its trace with is $\mbox{Tr}[A_n^\dagger A_n \rho] = P_n$. The result is the state $\tilde{\rho}_n$ defined above.
To determine the probability that $\mathcal{A}$ is in state $|n\rangle_{\ms{a}}$ we use the Born rule, and thus calculate the trace of the state obtained by projecting $\sigma$ into the subspace defined by $|n\rangle_{\ms{a}}$. Since this projected state is just $A_n \rho A_n^\dagger$, this trace is $\mbox{Tr}[A_n^\dagger A_n \rho] = P_n$.
To summarize the above analysis, if we define the outcomes of our physical process by partitioning the Hilbert space into the subspaces given by the states $\{ |n\rangle_{\ms{a}}\}$ of $\mathcal{A}$, then the state $\sigma$ precisely encodes the physical state corresponding to having the outcomes $\tilde{\rho}_n$ occurring with probability $P_n$, and thus to the physical process given by the measurement with operators $\{ A_n \}$.
To apply to $\mathcal{S}$ the feedback unitaries $U_n$ all we have to do is to apply a joint unitary to the two systems that applies the unitary $U_n$ to the system when $\mathcal{A}$ is in state $|n\rangle_{\ms{a}}$. This unitary is
\begin{equation}
W = \sum_{n} U_n \otimes |n\rangle_{\ms{a}} \langle n |_{\ms{a}} .
\end{equation}
Applying $W$ to $\sigma$ the joint state of the two systems becomes
\begin{equation}
\sigma' = \sum_{n} U_n A_n \rho A_m^\dagger U_m^\dagger \otimes |n\rangle_{\ms{a}} \langle m |_{\ms{a}} ,
\end{equation}
which, as we have seen above, describes the physical situation in which the system is in one of the states $ \tilde{\rho}_n'$ with the corresponding probabilities $P_n$. From a physical point of view the unitary process is thus equivalent to the measurement-based feedback process above, in that an observer who measures in the basis $|n\rangle_{\ms{a}}$ will see the same outcomes with the same probabilities.
\vspace{-2mm}
\subsection*{Physical processes with a single outcome}
\vspace{-2mm}
Often the purpose of a protocol is to produce a single final state rather than a number of possibilities. In such protocols, especially those that must deal with random external influences, there may be a large number of intermediate possibilities that must somehow be reduced to a single final outcome. If we wish to reduce all outcomes of all subsystems taking part in our process to a single outcome, then this is only possible if we reverse the evolution of the external systems that have generated the multiple possibilities. But so long as we have enough auxiliary systems taking part in our process then we can always place the system of interest in a single final state, no matter the number of intermediate possibilities. This can be done, for example, by using an auxiliary system to implement a feedback process similar to the one described above. If system $\mathcal{S}$ is in a mixture of some set of orthonormal states $|n\rangle_{\ms{s}}$, each having probability $p_n$, then $\mathcal{S}$ can be transformed into a single pure state $|\psi\rangle_{\ms{s}}$ in the following way. First we initialize the auxiliary system in the state $ |0\rangle_{\ms{a}}$. Then we define a unitary operator
\begin{align}
U = \sum_{n} |n\rangle_{\ms{s}} \langle n |_{\ms{s}} \otimes |n\rangle_{\ms{a}} \langle 0 |_{\ms{a}} + \tilde{U}
\end{align}
designed to correlate system $\mathcal{S}$ with the auxiliary. Here the operator $\tilde{U}$ is chosen merely to ensure that $U$ is unitary. For example we could set
\begin{align}
\tilde{U} = & \; |0\rangle_{\ms{s}} \langle 0 |_{\ms{s}} \otimes \sum_{n\not=0} |n\rangle_{\ms{a}} \langle n |_{\ms{a}} \nonumber \\
& \; + \sum_{n\not=0} |n\rangle_{\ms{s}} \langle n |_{\ms{s}} \otimes \biggl( |0\rangle_{\ms{a}} \langle n |_{\ms{a}} + \sum_{k\not=0,n} |k\rangle_{\ms{a}} \langle k |_{\ms{a}} \biggr) .
\end{align}
Now we apply $U$ to the two systems. Since the auxiliary is initially in the state $ |0\rangle_{\ms{a}}$ the operator $\tilde{U}$ has no effect and the result is the correlated state
\begin{align}
\sigma_U = \sum_{n} p_n |n\rangle_{\ms{s}} \langle n |_{\ms{s}} \otimes |n\rangle_{\ms{a}} \langle n |_{\ms{a}} .
\end{align}
We now define the unitary operator
\begin{align}
V = \sum_{n} |\psi\rangle_{\ms{s}} \langle n |_{\ms{s}} \otimes |n\rangle_{\ms{a}} \langle n |_{\ms{a}} + \tilde{V},
\end{align}
in which, once again, $\tilde{V}$ is chosen merely to ensure that $V$ is unitary. Applying $V$ to the two systems implements the ``feedback'' process and produces the final state
\begin{align}
\sigma_{\ms{fin}} = |\psi\rangle_{\ms{s}} \langle \psi |_{\ms{s}} \otimes \Bigl[ \sum_{n} p_n |n\rangle_{\ms{a}} \langle n |_{\ms{a}} \Bigr] .
\end{align}
Note that this feedback process actually transfers the initial state of the system into the auxiliary. Since the entire process is unitary (and thus reversible), it must also transfer the initial state of the auxiliary into the system (up to a local unitary transformation), which it does. We could, of course, combine the two unitary operations together to obtain the ``swap'' unitary $S = VU$. Here we implemented the ``swap'' in two parts to show how it is achieved by using a correlating process (essentially a measurement) followed by a ``feedback'' process. Since all basic physical processes are logically reversible, to reset a system to a predefined pure state one must always swap the entropy in the system to another system. This is the origin of Landauer's erasure principle, in which erasing the information in a bit (reseting a bit) requires raising the entropy of the environment~\cite{MyEprint, Jacobs14}.
\vspace{-2mm}
\section{Translating measurement-based protocols into unitary protocols}
\vspace{-2mm}
In the analysis above we showed how any physical measurement-based feedback process can be implemented with a purely unitary process. One can clearly construct a unitary protocol that reproduces \textit{any} protocol involving measurements merely by using the above equivalence to replace all occurrences of a measurement with a unitary process involving an auxiliary quantum system. Nevertheless, it appears to be a natural tendency to assume that the unitary processes that result from replacing measurements with auxiliary systems would be much more complex and cumbersome to implement than the equivalent measurement-based protocols because of the need to ``carry around'' all the Hilbert space and use all those auxiliary systems. However a simple line of reasoning shows that this is not the case.
Consider a unitary protocol $\mathcal{U}$ that performs the same physical process as a measurement-based protocol $\mathcal{M}$. The parts of $\mathcal{U}$ that implement each of the measurement and feedback subprocesses, in the manner described in Section~\ref{uimp} above, do not require coherence between the states of the basis used for the auxiliary. We can thus allow the all the auxiliary systems to decohere in the appropriate bases. This decoherence does not change the complexity of the implementation. It does however, render the unitary implementation identical to the measurement-based implementation. The auxiliary quantum systems can now be thought of as classical systems performing the feedback operations. In fact, the unitary protocol is now an exact description of the measurement-based protocol, and as such the unitary protocol is no more complex that the measurement-based one.
In a measurement-based protocol the macroscopic classical systems that perform the measurement and feedback operations are continually being reused. The information contained in the classical bits is thus repeatedly erased, which ultimately involves dumping this entropy to an environment. In exactly the same way, the qubits in a unitary protocol can be reused by swapping their states into a cold reservoir to reset them to zero. Since parts of a unitary protocol may be decohered and/or erased by interactions with a bath, doing so is likely to be the simplest way to perform the protocol. The resulting evolution is still unitary, but if one traces out the bath(s) involved the resulting description of the protocol is no-longer unitary. To avoid confusion that might arise from referring to such protocols as unitary, we will instead use the term ``measurement free''. This emphasizes the key property, which is that the protocol does not use measurements, without concerning ourselves with whether baths are used to allow reset operations or induce decoherence.
Even given that unitary and measurement-based protocols can have the same structure, it is not immediately clear from the above analysis how post-selection protocols are implemented unitarily. We discuss the implementation of these protocols in next two sections.
\vspace{-2mm}
\subsection{``Static'' vs. ``Dynamic'' unitary circuits}
\vspace{-2mm}
Post-selection is the term used for a process in which a collection of systems is prepared in some joint state, some of the systems are measured, and depending on the measurement result the rest of the systems are either used as part of an ongoing process, or discarded. If the rest of the systems are discarded, then the process is repeated until the measurement returns a result that allows the systems to be used. In this way an iterative process with a variable number of iterations is used to prepare quantum states that are required for some process, such as a fault-tolerant quantum computation.
Unitary implementations of quantum computing are assumed to use a circuit (meaning a specified sequence of quantum gates applied to a number of qubits) that is fixed at the start of the computation and then implemented. Post-selection, however, involves running a unitary circuit to prepare the state of a collection of qubits, and then deciding whether or not to run the circuit again based on the result of a measurement. These two kinds of implementation seem somewhat different, so let us refer to an implementation that involves a fixed gate sequence as a ``static'' circuit, and one in which the gate sequence is changed as the process proceeds based on (random) measurement outcomes as a ``dynamic'' circuit.
A first sight it may seem that the amplification of measurement provides a unique ability to modify the gates performed on qubits because these gates are implemented with macroscopic controls. These controls apply macroscopic forces to the qubits to implement the gates in the fastest way. We cannot merely replace the measurement and feedback processes that choose the gates with mesoscopic unitary processes because the mesoscopic elements cannot control the macroscopic elements that implement the gates. But let us suppose for a moment that we do wish to use a mesoscopic qubit $Q$ to control which gates are applied to and between other qubits. Since quantum gates are typically two-qubit gates, having the action of these gates controlled by another qubit merely means replacing the two-qubit gates with three-qubit gates. We then arrange things so that all the gates in the circuit to be controlled are controlled by the same qubit, namely Q. But we also know that two qubit gates are universal for unitary computation, so all the three qubit gates can be replaced by two-qubit gates. There is therefore no problem at all with having Q control the action of gates between other qubits. The complexity of the circuit has been increased somewhat (there are now more two-qubit gates than there were before), but the additional complexity is presumably no more than the complexity of the classical controller that would otherwise have to act based on the results of a measurement.
Now note that to implement post-selection by having one or more qubits, Q, control whether or not a given circuit, C, is applied to some other qubits, we do need the ability to run circuit C repeatedly. We can, of course, do this merely by continually repeating the fixed sequence of gates that constitute C. While this gate sequence is fixed, the action of the gates, and thus the circuit C, is different each time because this action depends on the value of the control qubit(s) Q. The circuit C is part of a larger circuit L that is also run each time C is run. The circuit L includes the gates that correlate the control qubit(s) Q with a subset of the other qubits (so that Q effectively stores the result of a measurement on these qubits), and thus updates the value of Q each time it is run. We now have a perfectly deterministic circuit, L, that is able to implement a repeat-until-success post-selection protocol. It does so by virtue of the fact that its action on each repetition depends on information stored on previous repetitions. In fact, by introducing a process of repeating a fixed circuit an indefinite number of times to perform a computation we are merely employing the procedure already used by classical computers. It is precisely this method of operation that allows classical computers to perform universal computation (and thus post-selection) with fixed circuits. We see that the distinction between ``static'' and ``dynamic'' unitary circuits is merely that the former consist of a fixed sequence of gates applied once, whereas the latter consist of a fixed sequence of gates applied repeatedly until the computation is complete.
\vspace{-3mm}
\subsection{Post-selection with a fixed-time circuit}
\vspace{-1mm}
In the previous discussion we saw that unitary circuits can be used to perform post-selection processes in the same ``dynamic'' repeat-until-success manner employed by measurement-based protocols. However, we feel it is worth noting that one can also perform post-selection processes with ``static circuits'', meaning a gate sequence that is performed only once and that does not involve the repetition of state preparations. This is achieved by having all the preparations done simultaneously at the start (in parallel), which requires effectively fixing the number of repetitions that the post-selection process is allowed to use. Fixing the number of repetitions does not make a significant difference to the output of the protocol for two reasons. The first is that the number of repetitions that a dynamic procedure can use to prepare a state for use in a computation is in any case limited by practical considerations. The second is that the probability that the procedure does not produce a correct state for the computation reduces exponentially with the number of repetitions, so in practice only a small number is required.
Let us consider an example in which the post-selection protocol involves i) preparing two identical states of $m$ qubits each, ii) applying gates that interact the combined set of $2m$ qubits, iii) performing a measurement on one of the $m$-qubit subsets, and iv) keeping or discarding the other subset depending on the result of the measurement. To perform this protocol with a fixed circuit we apply $N$ copies of the circuit that performs steps i) and ii) simultaneously on $N$ respective sets of $2m$ qubits. We then mimic step iii (the measurement) by applying (again simultaneously) to each of the $N$ sets a circuit that correlates one of their two $m$-qubit subsets a single control qubit (each of the $N$ sets has its own control qubit, which stores the result ``keep'' or ``discard''). We now construct a register of $m$ new qubits, R, that will contain the output of our protocol. We then go to each of the $N$ sets in turn, and perform a swap operation that swaps its unmeasured subset of $m$ qubits into R, conditional on the state of the control qubit for that set. In this way, each of the $m$ subsets is loaded into the output register in turn, but only if it contains a state that has passed the ``measurement test''. The larger the value of $N$ the less likely it is that the register does not contain the desired output state, with the likelihood dropping exponentially in $N$. The action of this circuit is the same as that of the measurement-based repeat-until-success protocol with the maximum number of repetitions fixed at $N$ (except for the time that the output states have to spend sitting in the output register, which is linear in $N$).
\vspace{-2mm}
\subsection{Unitary error-free classical processing}
\vspace{-1mm}
To fully replace measurements with unitary circuits we must have an explicit construction by which the unitary circuits can perform perfect (that is, fault-tolerant) classical computation. This could be achieved by using the concatenation methods of FTQC, but these have a component-wise error threshold of around $1\times 10^{-3}$~\cite{Antonio16x}. Further, the information that we wish to process, which is stored in the computational basis states of mesoscopic qubits, must be encoded first so that fault-tolerant processing can be performed on it. We must therefore have an explicit circuit for performing this encoding, and consider the probability that this circuit encodes the state incorrectly. This error probability for encoding is precisely equivalent to the measurement error in a measurement-based protocol. We recently presented an explicit scheme for encoding classical information and processing it reliably using unitary gates~\cite{Cruikshank16x}. This method builds upon von Neumann's ``multiplexing'' method~\cite{vonNeumann56}, appears to be quite feasible to implement (unlike previous multiplexing protocols), and achieves an error threshold close to von Neumann's ideal value of $\epsilon=1/6$ for a "majority organ" (see below) in an error-correcting network. We now briefly summarize the performance of this method.
Our encoding and processing circuits are based on the simplest classical coding scheme, namely repetition coding, in which the state of the qubit is simply stored in some number of qubits. The state can be faithfully recovered so long as more than half the qubits retain the correct state. The circuit shown in Fig.~\ref{fig1} achieves the encoding process by copying the state of a single qubit to $3^n$ qubits by using a ``cascade'' of ``AMP'' gates, in which each AMP copies the classical information in a single qubit to two others (using, e.g., two CNOT's). The probability that there is an error in the resulting encoding is $\approx 0.51 p$, in which $p$ is the probability that there is a (quantum) error in at least one of the outputs of the AMP gate~\cite{Cruikshank16x}. Thus the equivalent measurement error incurred by this encoding circuit is $\approx 0.51 p$, and thus less than that of an individual AMP gate.
To process the information in the repetition code fault-tolerantly we must have an error correction procedure (that is, a unitary circuit) which is applied periodically to the $3^n$ bits in the code. To achieve this error correction von Neumann used a ``majority organ'' that takes three input bits and, so long as it works correctly, outputs three bits with the value shared by the majority of the inputs~\cite{vonNeumann56}. We construct this irreversible three-bit gate with two three-bit unitary gates- a reversible majority counting gate, and an AMP gate to copy the correct output to two ancillary bits. Applying the majority organ to triples of the $3^n$ code bits certainly corrects errors, but it also produces correlations between the bits belonging to each triple, degrading the code. Von Neumann got around this problem by randomly regrouping the bits into new triples each time one desires a correction operation. In this way, so long as there are enough bits in the code, the bits input to each majority gate remain sufficiently independent to achieve a stable steady-state. Nevertheless the procedure demands highly complex wiring with the result that its scalability is unclear~\cite{Pippenger90, Han11} (see also~\cite{Nikolic02, Roy05, Boykin05, Beiu07, Bhaduri07}). It turns out that it is possible to use majority gates in a manner that is much simpler and has a performance which is quite similar. By i) arranging the $3^n$ code qubits in an $n$-dimensional hypercube, ii) defining a single correction operation as that of sequentially applying majority gates along each of the $n$ dimensions of the hypercube in turn, and iii) defining the code ``bits'' as the $3^{n-1}$ groups containing three mutually correlated bits, one obtains a high-threshold fault-tolerant correction circuit that is highly efficient in that it enables a compact (wiring) configuration~\cite{Cruikshank16x}. As an example, an 81-(qu)bit code (n=4) turns a three-bit unitary gate error of $ p = 0.4\%$ into a logical error probability of $1.5\times 10^{-11}$.
Finally we must have a way to process the information stored in the repetition code. This is not difficult, since a logical operation on a logical (coded) bit is obtained merely by applying that operation to each of the coding (qu)bits. In fact, the AMP and majority gates are universal for classical computation. Since all inputs of a three-input gate must be correct in order to produce a correct logical result, the threshold is slightly smaller for universal computation and was calculated to be $p = 5.5\%$~\cite{Cruikshank16x} or $\epsilon\approx 13\%$.
\subsection{Summary: measurement-based discrete feedback vs. a unitary implementation}
To reproduce measurement-based feedback operations using unitary circuits requires encoding the information, processing it, and then using the result to apply an operation to one or more qubits (the ``feedback''). We discussed the first two above. To apply the feedback operation to some ``target'' qubit we can simply use any one of the code qubits, since the error-correction circuit maintains each on of them as a copy of the encoded logical bit, with some (steady-state) probability of error. Using the error-correction circuits described above, this probability is $q \approx 0.51 p$. If the feedback operation is then performed by applying a two-qubit gate to the chosen code qubit and the target qubit, the total probability of error is that of the two-qubit gate (which we take to be $p$) plus that of the coding qubit, and is thus $\approx 1.51 p$.
We can now summarize the equivalent measurement error and feedback error achievable with unitary circuits using majority gates with error probability of $p$. The equivalent measurement error (as discussed above) is no more than $0.51 p$, the processing has an arbitrarily low error, and there is an error on the feedback operation of $0.51 p$ in addition to the basic gate error (which is that incurred by the equivalent classical feedback operation). To put this succinctly, unitary circuits can implement any measurement and feedback operation, and do so with an error probability very similar to the unitary gate error probability $p$.
\begin{figure}[t]
\leavevmode\includegraphics[width=1\hsize]{fig1.png}
\caption{A circuit that encode a single classical bit (which may be stored in a qubit) into a repetition code of $3^n$ qubits, in which the encoding error is less than $0.51$ of the error probability of the AMP gate. We also show the AMP gate in terms of controlled-not gates. Of course, the bits in the resulting code contain mutual correlations. These correlations can be effectively eliminated (for a subset of $3^{n-1}$ bits) by using majority counting gates to realize a stable error-correction circuit. The result is a repetition code with effectively $3^{n-1}$ bits.}
\label{fig1}
\end{figure}
\subsection{Fault-tolerant quantum computation}
In their analysis of FTQC, Aharanov and Ben-Or used an explicitly measurement-free implementation~\cite{Aharonov98, Aharonov08}. Their purpose was to demonstrate that scalable FTQC was possible, and not to provide a scheme with a high threshold, and their scheme had a threshold of $p \approx 10^{-12}$. Since then quite a few fault-tolerant schemes have been devised with the purpose of obtaining higher thresholds~\cite{Shor96, Steane99, Knill05, Aliferis06b, Raussendorf07a, Svore07, Fowler09, Fujii10, Paetznick13, Stephens14}, with the most recent claiming a threshold of $p \approx 0.01$. All of these latter FTQC schemes employ measurements, thus begging the question as to whether measurements are required for high threshold FTQC. As we have explained above, there can be no difference between measurement-based and unitary schemes that originates from quantum theory itself; any difference can only be due to technological factors.
Now consider the technological difference that arises between measurement-based and unitary FTQC schemes when measurements come with the ability to perform amplification. We have discussed above the fact that any amplification-based process that forms part of an FTQC protocol can be replaced by mesoscopic unitary circuits, and this can be done in such a way that the resulting error probabilities are very similar to those of the basic unitary quantum gates. Therefore, as far as error probabilities go, one expects the ultimate limits to the thresholds for unitary and measurement-based protocols to be similar. However, the error rate is not the only property of the feedback processes that is important for FTQC thresholds. The time required by these processes is also important. Any time that the computational qubits spend waiting for classical information to be processed incurs the accumulation of errors and thus affects the threshold. The relative thresholds of measurement-based and unitary FTQC protocols will thus depend, among other factors, on the relative time required by amplification as compared with encoding, something that again depends on technology.
\section{Measurements vs. unitary quantum control: open questions}
We have discussed the relationship between measurements and unitary processes from a fundamental point of view, and from a more practical (technological) point of view in the context of quantum computation. There is another area in which the practical relationship between measurements and unitary processes is especially salient, and that is \textit{feedback control}~\cite{Jacobs13, Jacobs14}. Quantum computation certainly uses feedback control, but when referring to feedback control as a subject in itself one usually has in mind applications of feedback for controlling the \textit{state} of a system (rather than to dynamical processes that enable the processing of information, as in FTQC). The difference is that the processing of information requires that the control protocol does not know the state of the system (it must operate correctly for any input state) whereas this is not the case if one wants merely to prepare a given state or realize a specific evolution. Because of this the problems typically studied under the term ``quantum feedback control'' tend to be much simpler as far as the protocols are concerned, but have other features that distinguish them from those in quantum computation.
Feedback control is usually concerned with continuous-time processes, rather than the discrete gates of computation, and the relevant constraints are restrictions on the sizes of the available forces and the rates at which these forces can be changed. The questions of interest tend to involve the fidelities that can be achieved under the relevant constraints in the presence of specified damping and/or noise sources. The problems are thus mainly of a dynamical nature rather than an information theoretic one. Nevertheless, since feedback protocols must usually transfer entropy from the system to the controller, environment, or other auxiliary systems, this area of research lies on the border between quantum dynamics, thermodynamics, and information theory~\cite{Strasberg13, Horowitz14, Brandner15, Horowitz15, Gong16}.
As in quantum computation it is natural to define measurements as including amplification. Measurements thus bring with them error-free processing of the information obtained from the system. Further, depending on the technological constraints relevant to a particular problem, there may be some differences between the fidelity or speed of the control operations that can be applied to a quantum system by a macroscopic, as opposed to a mesoscopic, device~\cite{Balouchi16x}.
The second difference between measurement-based controllers and mesoscopic quantum controllers is that measurement-based controllers are more restricted in the way in which they can interact with the system being controlled. As we have emphasized in this article, measurements do not provide any fundamentally new dynamics. In fact, the requirement that a controller interact with a system via making measurements on it actually restricts the dynamics that the controller can employ. For example, the controller cannot use joint dynamical processes that generate entanglement between the system and the controller.
We see that the question of the usefulness of measurements for control is somewhat different from that of their usefulness for FTQC. While measurement (more precisely amplification) provides stronger control forces and error-free processing of the information obtained about the system, a controller that is a classical system, and whose interactions with the system must effectively involve measurements, cannot access the full range of dynamical interactions available to a controller that is a quantum system. When comparing the performance of measurement-based and coherent control one must also decide whether the constraint on the strength of the interaction between the system and the quantum controller should be the same as that for the probe system that makes the measurements for the measurement-based controller. Certainly setting these to be the same makes sense for present technologies, and the results we describe below are obtained under this assumption.
To-date the question of the relative advantages of measurement-based and coherent feedback control is largely unexplored. Some results have been obtained for ideal controllers (that is, controllers that are not themselves subject to noise, and are able to make minimal-noise (quantum limited) measurements, etc.). It is known that ideal linear quantum controllers outperform classical (measurement-based) linear controllers for controlling linear quantum systems~\cite{Hamerly12, Hamerly13, Jacobs14b}. It is known that when there is a bound on the maximum eigenvalues of the interaction between the system and the controller, a quantum controller can outperform a measurement-based controller~\cite{Jacobs14c}. It is also known that there are some tasks, especially for information processing, that quantum controllers can perform that measurement-based controllers cannot~\cite{Wiseman94d, Naoki14}. There has been virtually no research on regimes in which measurement-based control is superior to that by mesoscopic quantum systems; this may well be the case when mesoscopic controllers are subject to various kinds of noise. In fact there are many open questions regarding the operation and performance of coherent controllers and their relationship to measurement-based control. For classical controllers it is clear that the controller must obtain information about the system in order to effect control, and this information is easily quantified and its dynamics elucidated using established techniques. For quantum controllers the relationship between control achieved and information obtained is not at all clear; whether information is irrelevant to quantum controllers, or whether there are measures of quantum information that are relevant to quantum control is unknown. Exactly what mechanism(s) allow quantum controllers to outperform classical controllers has been scarcely elucidated, and whether there is any relationship with entanglement that may be generated between the controller and the system is unknown. In classical controllers the data obtained from measurements of the system must often be processed with complex circuits to effect good control, while those coherent control methods explored to date use virtually no processing. The value of processing to quantum controllers, or equivalently the value of increasing the complexity of quantum controllers has not been explored. The ultimate limits to control imposed by finite resources, such as interaction strength, are also largely unknown~\cite{Wang13, Balouchi16x}.
\section{Conclusion}
We have elucidated in detail how measurements, defined as the action of projecting onto a subspace to obtain a measurement result, do not generate any physical effect that cannot be obtained via a unitary process. Because of this, no physical process or protocol can require the use of measurements for its implementation, unless by ``measurement'' one means more than the selection of an outcome along with any outcome-dependent actions. However, real measurements of mesoscopic systems in the laboratory always involve not only projection but also amplification. While projection does not effect the physics of a process, amplification does since it is a physical process itself.
We have also explained how amplification processes can be replaced by mesoscopic encoding processes, which shows that while amplification is physical it doesn't necessarily imbue measurement-based protocols with novel power.
One way to summarize the main concepts we have discussed is to consider a statement that is fairly common in the literature, and has been referred to as ``the principle of deferred measurement''. This statement is ``measurements that are applied part-way through any physical process may always be removed and replaced by measurements at the end of the process''. If the reader has fully digested our discussion, we would expect that he or she might comment on this statement in the following way: ``If by measurement one means the application of projection operators in the manner of the standard quantum measurement formalism, then measurements at the end of the process are no more necessary than those in the middle; they are equivalent merely to defining the basis with which an observer would wish to correlate his or her measuring apparatus. The phrase `make a measurement at the end of the process' is somewhat misleading when referring only to applying a projector, since it does not involve any physical process and thus any action. If one is instead referring to measurement as the act of amplifying a signal (one of a set of quantum states) so as to correlate these states with a macroscopic measuring device, then from a practical point of view it is not necessarily the case that measurements can be removed from a process and placed at the end. Since amplification is a physical process, in general it will depend on the available technology as to whether the use of amplification as part of implementing a process is advantageous or not.''
We have also considered the question as to whether unitary protocols that maintain coherence have any unique power over that of measurement-based protocols. In the most general context the answer is yes, since quantum computing can perform tasks that classical computation cannot. Here we considered this question in the context of the dynamical control of quantum systems by other systems, often referred to as (measurement-based or coherent) feedback control. In this case it does appear that unitary protocols have some advantages over measurement-based protocols, at least in certain situations, but relatively little is yet known. As with the question of the relative advantages of measurement-base versus unitary protocols for implementing fault-tolerant quantum computation, it seems likely that in practice the answer will be dominated by technological rather than fundamental considerations.
\vspace{2mm}
\section*{Acknowledgments}
\vspace{-1mm}
KJ would like to thank Tanmoy Bhattacharya and Emil Mottola, of Los Alamos National Laboratory, for valuable discussions circa 2000 on the place of measurement in quantum physics. BC was supported by an appointment to the Student Research Participation Program at the U.S Army Research Laboratory, administered by the Oak Ridge Institute for Science and Education through an interagency agreement between the U.S Department of Energy and USARL.
|
\section{\label{intro}Introduction}
Experiments to measure the rate of neutrinoless double-beta ($0\nu\beta\beta$\xspace) decay are
increasing in number and scale \cite{avi08}, in part because they offer the only
real hope of determining whether neutrinos are Majorana particles. The rate of
decay, however, depends on nuclear matrix elements that must be accurately
calculated to allow experimentalists to plan efficiently and interpret results.
At present, the predictions of various nuclear models for the matrix elements
differ by factors of up to three~\cite{engel16,vog12}, and it is not possible to
estimate the theoretical uncertainty in any of them. Improving the accuracy of
matrix-element calculations has become an important goal for the
nuclear-structure community.
The methods used to calculate the matrix elements include the shell
model~\cite{Iwata16, sen16, men08}, the interacting boson model
(IBM)~\cite{bar15}, the quasiparticle random phase approximation (QRPA)
~\cite{Mustonen13, sim13, Hyvarinen15}, and the generator coordinate method
(GCM)~\cite{rod10, vaq13, yao15, hin14}. Both the QRPA and the GCM have been
used in conjunction with energy density functional (EDF)
theory~\cite{Mustonen13, vaq13, yao15}. These methods, which allow large
single-particle spaces (and thus allow unrestricted collective deformation and
pairing) but do not contain all kinds of correlations, yield matrix elements
that are usually larger than those of the shell model, which employs only a few
single-particle levels but allows arbitrarily complex correlations within them.
Recent work to compare the shell-model with the GCM \cite{men14} suggests that
the extra valence-space correlations, particularly those due to isoscalar
pairing \cite{men16}, are responsible for much of the difference in the
predictions of the two methods.
In an attempt to include all relevant physics, Ref.~\cite{hin14} proposed
combining the virtues of the large-single-particle-space methods with those of
the shell model by using the neutron-proton pairing amplitudes as generator
coordinates in the GCM approach, within spaces of two or more major shells.
Though promising, the results in that paper were obtained with the restriction
(as in prior GCM work on $0\nu\beta\beta$\xspace decay \cite{rod10,vaq13,yao15}) that deformation
be axially symmetric and with a simple pairing-plus-quadrupole interaction.
Here we include a measure of triaxial deformation parameter as a generator
coordinate and work with genuine shell-model interactions. We use a one-shell
calculation to compare our results, obtained with carefully fit and widely used
interactions \cite{men08}, to those of exact diagonalization with the same
interactions. Having verified the accuracy of our approach, we move on to a
two-shell spaces, in which exact diagonalization is not currently possible.
Because no careful Hamiltonian tuning has yet been done for the $fp-sdg$ model
space appropriate for $A\approx 80$, we use the perturbation-theoretic methods
discussed in Ref.\ \cite{Tsunoda14} to derive an interaction. Our results are
the first for $^{76}$Ge and $^{82}$Se of a realistic shell-model-like
calculation in more than one shell, and the first exploration of the effects of
triaxiality on $\beta\beta$ decay. Recent experiments indicate that the
deformation of $^{76}$Ge is indeed triaxial \cite{Toh13}.
This paper is organized as follows: Section~\ref{mode} presents a brief overview
of the $0\nu\beta\beta$\xspace matrix elements and of the GCM with a Hamiltonian.
Section~\ref{calc1} contains results in a single shell, which we use to test our
approach by comparing with exact diagonalization. Section \ref{calc2}, which
presents $0\nu\beta\beta$\xspace matrix elements in two shells, is the heart of the paper.
Section~\ref{summary} contains a preview of future work and a summary.
\section{\label{mode}Matrix Elements and Methods}
In the closure approximation, the quantity we need is the matrix element of a
two-body operator between the ground states of the initial and final nuclei. If
the decay is produced by the exchange of a light-Majorana neutrino with the
usual left-handed currents,
we can write the nuclear matrix element as \cite{pan96,sim08}
\begin{widetext}
\begin{align}
\label{eqn:ME}
M^{0\nu}&=M^{0\nu}_{\text{GT}}-\frac{g^2_V}{g^2_A}M^{0\nu}_{\text{F}}
+ M^{0\nu}_T \\
&= \frac{2R}{\pi g^2_{A}}\int_0^{\infty}\!\! q \, dq \,
\bra{F} \sum_{a,b}\frac{j_0(qr_{ab})\left[h_{\text{GT}}(q)\vec{\sigma}_a\cdot
\vec{\sigma}_b + h_{\text{F}}(q)\right]+j_2(qr_{ab})h_T(q)\left[3\sigma_1 \cdot
\vec{r}_{ab} \sigma_2 \cdot \vec{r}_{ab} - \sigma_1 \cdot \sigma_2\right]}
{q+\overline{E}-(E_I+E_F)/2}\tau^+_a\tau^+_b\ket{I} \,, \nonumber
\end{align}
\end{widetext}
where GT, F, and T refer to the Gamow-Teller, Fermi, and tensor parts of the
matrix element. The vector and axial coupling constants are given by $g_V=1$ and
$g_A \approx 1.27$, $|I\rangle$ and $|F\rangle$ are the ground-states of the
initial and final nuclei, $r_{ab}$ is the distance between nucleons $a$ and $b$,
$j_0$ and $j_2$ are the usual spherical Bessel functions, $\bar{E}$ is an
average excitation energy (to which the matrix element is not sensitive), and
the nuclear radius $R=1.2A^{1/3}$ fm makes the matrix element dimensionless. The
functions $h_{\text{F}}(q)$, $h_{\text{GT}}(q)$, and $h_\text{T}(q)$ contain
nucleon form factors and forbidden corrections to the weak current. We modify
our wave functions at short distances with the ``Argonne'' correlation
function~\cite{sim09}. A detailed presentation of the form of the matrix
element can be found in Ref.\ \cite{sim08}.
The crucial ingredients in Eq.\ \eqref{eqn:ME} are the initial and final ground
states $\ket{I}$ and $\ket{F}$. To obtain them, we use a shell-model effective
Hamiltonian $H_{\text{eff}}$ in a valence space whose size we are free to
choose.
The first step in the GCM procedure is to generate a set of reference
quasiparticle vacuua $\ket{\varphi(q_1,q_2,\ldots)}$ that provide the minimum
energy such states can have while constrained to also have expectation values
$q_i = \braket{\mathcal{O}_i}$ for a set of collective operators
$\mathcal{O}_i$. Here we take the operators $\mathcal{O}_i$ to be:
\begin{equation}
\label{eq:params}
\begin{aligned}
\mathcal{O}_1 &= Q_{20}\,,&
\mathcal{O}_2 &= Q_{22} \,, \\
\mathcal{O}_3 &= \frac{1}{2}(P_0+P_0^\dag) \,, &
\mathcal{O}_4 &= \frac{1}{2}(S_0+S_0^\dag) \,,
\end{aligned}
\end{equation}
where
\begin{equation}
\label{eq:constr-ops}
\begin{aligned}
Q_{2M} &= \sum_a r^2_a Y^{2M}_a \,, \\
P_0^\dag &= \frac{1}{\sqrt{2}}\sum_l \sqrt{2l+1} \, [c^\dag_l
c^\dag_l]^{L=0,J=1,T=0}_{000} \,, \\
S_0^\dag &= \frac{1}{\sqrt{2}}\sum_l \sqrt{2l+1} \, [c^\dag_l
c^\dag_l]^{L=0,J=0,T=1}_{000} \,,
\end{aligned}
\end{equation}
with $M$ labeling the angular-momentum $z$-projection and $a$ labeling nucleons,
and the brackets signifying the coupling of orbital angular momentum, spin, and
isospin to various values, each of which has $z$-projection zero. The operator
$c^\dag_l$ creates a particle in the single-particle level with orbital angular
momentum $l$. The operator $P_0^\dag$ creates a correlated isoscalar pair, and
the operator $S_0^\dag$ a correlated isovector neutron-proton pair. We actually
only constrain one of the two pair amplitudes at a time: the isoscalar amplitude
when computing $M^{0\nu}_{GT}$ (and $M^{0\nu}_T$, which is small) and the
isovector amplitude when computing $M^{0\nu}_{F}$. The usual deformation
parameters $\beta$ and $\gamma$ are related to $q_1 \equiv \braket{Q_{20}}$ and
$q_2 \equiv \braket{Q_{22}}$ by $\beta = (\chi b^2 / \omega_0) \sqrt{q_1^2 + 2
q_2^2}$ (with $b$ the oscillator length, given by $2R/\sqrt{5} \,
(3A/2)^{-1/6}$, $\omega_0 =41.2 A^{-1/3}$, $\chi=0.4$) and $\gamma =
\tan^{-1}(\sqrt{2}q_2/q_1)$.
To efficiently include the effect of neutron-proton pairing, we start, as in
Ref.\ \cite{hin14}, from a Bogoliubov transformation that mixes neutrons and
protons, i.e.\ from quasiparticle operators of the (schematic) form
\begin{equation}
\label{eqn:operator}
\alpha^{\dagger}\sim u_pc^{\dagger}_p+v_pc_p+u_nc^{\dagger}_n+v_nc_n.
\end{equation}
In the full equations single-particle states are summed over, so that each of
the coefficients $u$ and $v$ are replaced by matrices, as described in
Ref.\ \cite{goo79}. We then solve constrained Hartree-Fock-Bogoliubov (HFB)
equations, minimizing expectation values of the form
\begin{equation}
\label{eqn:H}
\begin{split}
\braket{H^{\prime}}&=\braket{H_{\text{eff}}}-\lambda_Z \left(\braket{N_Z}-Z
\right) -\lambda_N \left(\braket{N_N} - N\right) \\
&-\sum_i \lambda_i \left(\braket{\mathcal{O}_i}-q_i\right) \,,
\end{split}
\end{equation}
where the $N_Z$ and $N_N$ are the proton and neutron number operators,
$\lambda_Z$ and $\lambda_N$ are corresponding Lagrange multipliers, the sum over
$i$ includes up to three of the 4 $\mathcal{O}_i$ in Eq.\ \eqref{eq:params}, and
the other $\lambda_i$ are Lagrange multipliers to constrain the expectation
values of those operators to $q_i$. We solve these equations many times,
constraining each time to a different point on a mesh in the space of $q_i$.
Having obtained a set of HFB vacua with various amounts of axial deformation,
triaxial deformation, and isoscalar/isovector pairing, we construct the GCM
state by superposing projected HFB vacua:
\begin{equation}{\label{eqn:GCMwf}}
\ket{\Psi^{J}_{NZ\sigma}}=\sum_{K,q}f^{JK}_{q\sigma}\ket{JMK;NZ;q} \,,
\end{equation}
where $|JMK;NZ;q\rangle\equiv
\hat{P}^J_{MK}\hat{P}^N\hat{P}^Z|\varphi(q)\rangle$ and $q$ is short for the set
of all $q_i$. Here, the $\hat{P}^{\prime}$s are projection operators onto
states with well-defined angular momentum $J$ and $z$-component $M$, neutron
number $N$, and proton number $Z$~\cite{ring80}. The weight functions
$f^{JK}_{q\sigma}$, where $\sigma$ enumerates states with the same quantum
numbers, follow from the Hill-Wheeler equations~\cite{ring80}
\begin{equation}
\label{eqn:HWeq}
\sum_{K^{\prime},q^{\prime}}\Bigl\{\mathcal{H}_{KK^{\prime}}^J(q;
q^{\prime})-E^J_{\sigma}\mathcal{N}^J_{KK^{\prime}}(q;
q^{\prime})\Bigr\}f^{JK^{\prime}}_{q^{\prime}\sigma}=0,
\end{equation}
where the Hamiltonian kernel $\mathcal{H}_{KK^{\prime}}^J(q; q^{\prime})$ and the norm kernel $\mathcal{N}_{KK^{\prime}}^J(q; q^{\prime})$ are given by
\begin{equation}
\quad
\begin{aligned}
\mathcal{H}_{KK^{\prime}}^J(q; q^{\prime}) &=
\bra{\varphi(q)}H_{\text{eff}}\hat{P}^{J}_{KK^{\prime}}\hat{P}^{N}\hat{P}^{Z}
\ket{\varphi(q^{\prime})} \,, \\
\mathcal{N}_{KK^{\prime}}^J(q; q^{\prime}) &=
\bra{\varphi(q)}\hat{P}^{J}_{KK^{\prime}}\hat{P}^{N}\hat{P}^{Z}
\ket{\varphi(q^{\prime})} \,.
\end{aligned}
\quad
\end{equation}
To solve Eq.\ \eqref{eqn:HWeq}, we first diagonalize the norm kernel
$\mathcal{N}$ and then use the nonzero eigenvalues and corresponding
eigenvectors to construct a set of ``natural states.'' Finally, we diagonalize
the Hamiltonian in the space of these natural states to obtain the GCM states
$|\Psi^{J}_{NZ\sigma}\rangle$ (for details, see Refs.~\cite{Rodriguez10a,
Yao10}). We carry out this entire procedure in both the initial and final
nucleus, using the lowest $J=0$ states in each as the ground states between
which we sandwich the $0\nu\beta\beta$\xspace operator to obtain the matrix element $M^{0\nu}$ from
Eq.\ \eqref{eqn:ME}.
\section{\label{calc1}Tests in a Single Shell}
Before undertaking a two-major-shell calculation, we need to test our GCM with a
realistic interaction in a model space small enough to allow exact
diagonalization. We begin by performing GCM calculations in the $pf$-shell,
comprising the $0f_{7/2}$, $0f_{5/2}$, $1p_{3/2}$, and $1p_{1/2}$ orbits. Using
use the KB3G interaction~\cite{Poves01}, which accounts successfully for the
spectroscopy, electromagnetic and Gamow-Teller transitions, and deformation of
$pf$-shell nuclei~\cite{cau05}, we compute the $0\nu\beta\beta$\xspace matrix elements of
$^{48}$Ca, $^{54}$Ti, and $^{54}$Cr. Although the last two nuclei are not
candidates for an experiment, they offer opportunities to test the GCM. Because
these nuclei show no evidence of triaxial deformation, we need only use the
axial quadrupole moment $q_1 \equiv \braket{Q_{20}}$ and isoscalar pairing
amplitude $\phi\equiv q_3=1/2\braket{P_0+P_0^\dag}$ as generator coordinates for
the computation of $M^{0\nu}_{GT}$.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig1_NME_Ca}
\caption{\label{fig:caticr}GCM results for the Gamow-Teller part of $0\nu\beta\beta$\xspace matrix
elements of $^{48}$Ca, $^{54}$Ti, and $^{54}$Cr, compared with the results of
exact diagonalization.}
\end{figure}
\begin{figure}[t]
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=3.0in]{Fig2_PES_76Ge_GCN}
\label{fig:side:a}
\end{minipage}%
\\
\vspace{0.2in}
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=3.0in]{Fig2_PES_76Se_GCN}
\label{fig:side:b}
\end{minipage}
\caption{\label{PES} Projected potential-energy surfaces produced by the GCN2850
interaction, with the isoscalar pairing amplitude $\phi=0$, in the
($\beta,\gamma$) plane for $^{76}$Ge and $^{76}$Se. }
\end{figure}
Figure \ref{fig:caticr} shows the GT matrix elements that result from this
procedure, alongside those coming from exact diagonalization. To highlight the
effects of isoscalar pairing in the GCM, we present the results of two separate
GCM calculations. In the first, as in Ref.\ \cite{men16}, we set all the
two-body matrix elements of the Hamiltonian with angular momentum $J=1$ and
isospin $T=0$ to zero, because those are the ones through which isoscalar
pairing acts. The resulting GT matrix elements overestimate the exact one
substantially. In the second calculation, we use the full KB3G interaction,
with the result that the matrix element decreases, coming quite close to the
exact one. The sensitivity to isoscalar pairing, pointed out long ago for the
QRPA in Refs.\ \cite{vog86} and \cite{eng88} and more recently for the GCM and
shell model in Refs.\ \cite{hin14} and ~\cite{men16}, shows that the
neutron-proton mixing in our HFB states is essential. The good agreement with
exact diagonalization suggests that once it is included, we are not omitting
anything of importance.
We turn now to one of the nuclei in which we are really interested: $^{76}$Ge,
used or to be used in many $\beta\beta$\xspace
experiments~\cite{kla01a,Aalseth02,GERDA13,MAJORANA14}. Shell model
calculations of the $0\nu\beta\beta$\xspace decay of this nucleus
\cite{men08,caurier08,sen16,sen14b} have usually been set in the so-called
$f5pg9$ space, comprising the $0f_{5/2}$, $1p_{3/2}$, $1p_{1/2}$, and $0g_{9/2}$
orbits, and have employed either the JUN45~\cite{Honma09} or
GCN2850~\cite{Gniady} Hamiltonian. The $f5pg9$ model space is not a complete
major shell; it includes levels from two different major shells and is missing,
in particular, the spin-orbit partners of the $0f_{5/2}$ and $0g_{9/2}$ orbits.
We discuss the effects of including these and other orbits later.
As we already mentioned, both theory \cite{Guo07,Shen11} and experiment
\cite{Chou93a,Toh13} indicate triaxial deformation in low-lying states of
even-even Ge and Se isotopes near $A=76$. Our calculations predict it as well.
Figure~\ref{PES} displays the $^{76}$Ge and $^{76}$Se quantum-number-projected
potential-energy surfaces (PES's), at isoscalar-pairing amplitude $\phi=0$,
produced by the GCM with the GCN2850 interaction. The minimum is at
$\beta_2=0.23, \gamma=24^{\circ}$ in $^{76}$Ge, a result that agrees well with
those of EDF-based GCM calculations~\cite{Guo07}, and at at $\beta_2=0.28,
\gamma=45^{\circ}$ in $^{76}$Se. In addition to this ``static'' triaxial
deformation, dynamical triaxial effects arise from the $\gamma$-soft PES's in
both isotopes. The GCM, which mixes states with a range of $\gamma$ values,
incorporates dynamical effects.
\begin{figure}[b]
\centering
\includegraphics[width=\columnwidth]{Spectra_GCN}
\caption{\label{fig:def} Low-lying excitation spectra of $^{76}$Ge and $^{76}$Se
produced in one shell by the GCM with the GCN2850 interaction, with and
without triaxial deformation (labeled by the parameter $\gamma$). The results
from the exact diagonalization of the shell-model Hamiltonian appear for
comparison \cite{men-exact}. }
\end{figure}
Our complete calculations include as generator coordinates both deformation
parameters $q_1$ and $q_2$ (or equivalently $\beta$ and $\gamma$) as well as one
of the proton-neutron-pairing parameters $q_3$ and $q_4$. We can assess the
effects of triaxial shape fluctuations by including or excluding triaxially
deformed configurations from the set of GCM basis states. Including them has
clear effects on spectroscopy. Fig \ref{fig:def} shows the spectra of low-lying
$0^+$ and $2^+$ states in the two important $A=76$ isotopes with GCN2850;
triaxial shapes, though they have a relatively small effect on the first excited
$0^+$ state, lower the second such state significantly in both nuclei, and in
$^{76}$Se by
over an MeV. The values for the strength $B(E2; 0^+\rightarrow 2^+)$ are
affected in a similar way. With triaxial deformation (and with the usual
effective charges $e^\text{eff}_p = 1.5 e$ and $e^\text{eff}_n = 0.5 e$) the
values in $e^2b^2$ are $0.172$ in $^{76}$Ge (vs.\ the exact-diagonalization
value of 0.158) and $0.275$ in $^{76}$Se (vs.\ the exact value of $0.209$).
Without triaxial deformation the numbers are smaller: $0.154$ in $^{76}$Ge, and
$0.268$ in $^{76}$Se.
Triaxial deformation has a non-negligible effect on the $0\nu\beta\beta$\xspace matrix element as
well. As Table \ref{t:one-shell} shows, our full GCM calculation gives values
for the matrix elements $M^{0\nu}$ that are about $15\%$ smaller than the
results obtained without triaxially deformed configurations. The full matrix
elements, though slightly suppressed, are in good agreement with those of exact
diagonalization (in this calculation only, we neglected the very small matrix
element $M^{0\nu}_T$). The GCM approach with neutron-proton pairing indeed
captures most of correlations around the Fermi surface that are important for
$0\nu\beta\beta$\xspace decay. The small discrepancy may be due to fluctuations in like-particle
pairing, which we do not treat here but which, according to the EDF-based work
of Ref.\ \cite{vaq13}, increase $0\nu\beta\beta$\xspace matrix elements slightly. We could include
those fluctuations, but at the cost of a considerable increase in computing
time.
\begin{table}[t]
\centering
\begin{tabular}{ldd}
\toprule &\multicolumn{1}{c}{GCN2850}&\multicolumn{1}{c}{JUN45} \\ \midrule
Axial GCM & 2.71 & 3.42 \\
Triaxial GCM & 2.33 & 2.94 \\
Exact & 2.81 \ \text{\cite{men08}} & 3.37 \ \text{\cite{sen14b}} \\
\bottomrule
\end{tabular}
\caption{\label{t:one-shell}The matrix elements $M^{0\nu}$ produced in the GCM
by GCN2850 and JUN45 for the decay of $^{76}$Ge, with and with out triaxial
deformation as a generator coordinate, and by those same interactions with exact
diagonalization.}
\end{table}
\section{Results in Two Shells}
\label{calc2}
The promise of the Hamiltonian-based GCM is an eventual \textit{ab initio}
calculation. Here we take a step in that direction by working in the full
$fp-sdg$ two shell space. The number of states for $A=76$ nuclei in this space
is still too large for exact diagonalization.
Before considering Ge and Se, we make one more test, for $^{48}$Ca, the one
experimental candidate in which an exact two-shell calculation is almost
possible at present. Ref.\ \cite{Iwata16} uses the SDPFMU-DB interaction, with
the omission of some cross-shell excitations, to compute the $0\nu\beta\beta$\xspace matrix element
nearly exactly. Our GCM result, 1.082, is close to 1.073, the result of Ref.\
\cite{Iwata16}, and suggests in addition that the cross-shell excitations
neglected in that paper really are unimportant. With some confidence in the
performance of the GCM in two shells, we turn to the decay of $^{76}$Ge.
The first issue we must grapple in this mid-shell nucleus is what to use for the
valence-space Hamiltonian. Ref.\ \cite{hin14} used a multi-separable collective
Hamiltonian that we wish to improve on here. The size of the two-shell space,
however, makes the usual procedure, in which shell-model Hamiltonians are tuned
to data, difficult to follow; furthermore, there are no well-tested Hamiltonians
for this space on the market. The first step in the usual approach is to produce
an initial valence-space Hamiltonian, traditionally in many-body perturbation
theory. Deficiencies in the many-body method are then remedied by tuning
single-particle energies and interaction matrix elements to experimental data.
Here we must settle for adjusting only single-particle energies. The tuning of
interaction matrix elements requires repeated calculations that are simply too
time consuming.
Although nonperturbative methods such as the in-medium similarity
renormalization group can produce shell-model Hamiltonians
\cite{tsukiyama12,bog14}, they have not been tested systematically for valence
spaces larger than one major harmonic-oscillator shell. We therefore use the
Extended Krenciglowa-Kuo (EKK) variant of many-body perturbation theory
\cite{Tsunoda14}, suitable for non-degenerate valence spaces, to construct an
effective Hamiltonian from a third-order $Q$-box in the $pf-sdg$ shell. We
begin from the 1.8/2.0 two- plus three-nucleon (3N) interaction of
Refs.~\cite{Hebeler11,Simonis16}; the interaction reproduces ground-state
energies across the light- and medium-mass regions of the nuclear chart
\cite{Simonis17}. With $\hbar\omega\!=\!10$ MeV, a space of 13 major shells for
intermediate-state sums is enough to ensure convergence.
The monopole components of our valence-space Hamiltonian are particularly
sensitive to the initial three-nucleon interaction \cite{ots10}, which one
generally reduces to effective zero-, one- and two-body parts via normal
ordering with respect to some independent-particle reference state
\cite{hagen07}. For shell model calculations, the usual reference state is the
inert core, containing all orbitals below the valence space. As discussed in
Refs.~\cite{Stroberg16,Stroberg17}, however, this choice completely omits
three-nucleon interactions among the valence particles themselves, and if the
target nucleus (here $^{76}$Ge or $^{76}$Se) is far from the core, the neglected
effects become sizable. Since we can choose to normal-order with respect to any
reference we want, it makes sense to include more orbitals, so that we better
capture the bulk effects of three-nucleon interactions among valence particles.
Thus, we take the reference state to be the (fictional) inert core corresponding
to $^{56}$Ni. While this nucleus is still some distance in proton and neutron
number from the $A=76$ nuclei, we expect it to make a better reference than than
the $^{40}$Ca core. We call the resulting valence-space interaction $pfsdg$ and
use it exclusively in the following.
\begin{figure}
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=\textwidth]{neutron_occupancy}
\end{minipage}%
\\
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=\textwidth]{proton_occupancy}
\end{minipage}
\caption{\label{fig:occ} The occupancies of valence neutron and proton orbits
produced by the interaction $pfsdg$ (see text) for $^{76}$Ge and $^{76}$Se,
following the adjustment of the single-particle energies for levels in the lower
shell. The measured occupancies are from Refs.~\cite{Schiffer08, Kay09}.}
\end{figure}
Perhaps because of the non-ideal reference state, the single-particle energies
that emerge from the perturbative procedure are poor. The proton sub-shell gap
at $Z=34$ is too large, causing the proton pairing mean field to disappear no
matter what the deformation and leading to occupation numbers that differ
significantly from measured values. To remedy the problem, we adjust the
single-particle energies of the orbits in the lower shell to reproduce those
values, while leaving the two-body part of the Hamiltonian untouched. Figure
\ref{fig:occ} shows the occupation numbers after adjustment, for both the
projected HFB state with the minimum energy (which we used to make the
adjustments) and in the final GCM states. Though the occupations change a
little when the HFB states are mixed in the GCM, they remain close to the
experimental values.
\begin{figure}[t]
\begin{minipage}{\columnwidth}
\includegraphics[width=\textwidth]{PES_pfsdg_76Ge}
\end{minipage}\\
\begin{minipage}{\columnwidth}
\includegraphics[width=\textwidth]{PES_pfsdg_76Se}
\end{minipage}
\caption{\label{fig:pes-2-shell} Potential-energy surfaces for $^{76}$Ge (top)
and $^{76}$Se (bottom) with the Hamiltonian $pfsdg$.}
\end{figure}
Figure \ref{fig:pes-2-shell} shows the resulting potential-energy surfaces. The
minimum occurs at smaller deformation than in the one-shell example so that, as
the calculated spectra in Fig.\ \ref{fig:spec-two-shell} show, the energy of the
first $2^+$ state rises noticeably. This worsens the agreement with experiment
slightly in $^{76}$Ge, but improves it in $^{76}$Se. The excited $0^+$ states
are also generally better reproduced in the two-shell calculation.
The $B(E2)$ values calculated in two shells are not better than those from one
shell, however. With the same effective charges as before, we find noticeably
smaller values: $0.126$ in $^{76}$Ge and $0.221$ in $^{76}$Se. The
corresponding experimental values are larger, $0.274$ and $0.432$. Evidently,
the $E2$ operator must be renormalized more in two shells than in one, a result
that is consistent with the smaller deformation in the two-shell calculation but
is nevertheless a little surprising.
We turn finally to the $0\nu\beta\beta$\xspace matrix elements, which appear in Table
\ref{t:two-shell}. The total matrix element, once triaxial deformation is
included, is only slightly larger than that from GCN2850 in a single shell.
Though our interaction is clearly not perfect, the result suggests that
enlarging the space further may not dramatically change the matrix element,
though gradual but continual changes with the addition of successive shells
cannot be ruled out. It also shows the importance of including triaxial shapes
in larger spaces.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{spectra-jason}
\caption{\label{fig:spec-two-shell} Calculated low-lying excitation spectra of
$^{76}$Ge and $^{76}$Se produced by the Hamiltonian $pfsdg$ alongside
experimental data~\cite{NNDC}.}
\end{figure}
\begin{table}[b]
\centering
\begin{tabular}{r@{\hskip 1cm}dd}
\toprule &\multicolumn{1}{c}{axial} & \multicolumn{1}{c}{full} \\ \midrule
$M^{0\nu}_{\rm GT}$ & 3.25 & 2.01 \\
$-\frac{g_V^2}{g_A^2}M^{0\nu}_{\rm F}$ & 0.43 & 0.35 \\
$M^{0\nu}_{\rm T}$ &-0.03 &-0.02 \\ \midrule
Total $M^{0\nu}$ & 3.65 & 2.34 \\ \bottomrule
\end{tabular}
\caption{\label{t:two-shell}GCM results for thte Gamow-Teller ($M^{0\nu}_{\rm
GT}$), Fermi ($M^{0\nu}_{\rm F}$), and tensor ($M^{0\nu}_{\rm T}$) $0\nu\beta\beta$\xspace matrix
elements for the decay of $^{76}$Ge in two shells, with and without triaxial
deformation.}
\end{table}
Figure \ref{fig:monu} summarizes our $\beta\beta$\xspace results. For the decay of $^{48}$Ca,
as noted, we reproduce the exact shell model results nearly perfectly in both
one and two shells. For the decay of $^{76}$Ge (and $^{82}$Se) in a single
shell, the GCM reproduces the exact result well enough, with two different
effective interactions. And in two shells, with a brand new effective
interaction, it obtains a result that is only slightly different from the
GCN2850 result in one shell.
An important caveat, in addition to those already mentioned: We really ought to
be using an effective $0\nu\beta\beta$\xspace operator to accompany our effective interaction, as
in Refs.\ \cite{hol13c} and \cite{engel09}. Those papers lead us to suspect a
change of 20\% or less from an effective decay operator in two shells. In any
event, because we made significant phenomenological adjustments to the
single-particle energies in the prototype calculation here, we cannot
systematically construct the decay operator that should accompany the effective
interaction.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{Fig7_NME-compilation}
\caption{\label{fig:monu} GCM matrix elements $M^{0\nu}$ compared with those of
the shell-model (SM), with either the JUN45~\cite{sen14b}, CN2850~\cite{men08}
KB3G~\cite{men16}, or SDPFMU-DB~\cite{Iwata16} interactions. The term $pfsdg$
denotes the two-shell interaction used here for $A\approx 80$ nuclei.}
\end{figure}
\section{\label{summary}Summary}
The perfect many-body method will include all possible correlations in an
infinitely-large space. One step on the way to that ideal is to enlarge the
single-particle space for the shell-model, a method that includes all
correlations within that space. Here we have approximately diagonalized a
shell-model Hamiltonian in two major shells. Tests in a single shell, and in
two shells for $^{48}$Ca, show that the our approximation method includes the
most important correlations. Our first-of-its-kind two-shell calculation of the
$0\nu\beta\beta$\xspace matrix element in $^{76}$Ge suggests a small effect from the extra
single-particle orbitals.
There are at least two ways forward from here. We should use a better effective
Hamiltonian, either by normal-ordering with respect to an ensemble reference
\cite{Stroberg17} that better includes bulk effects of three-nucleon forces far
from closed shells, or by careful tuning of the interaction. The second option,
besides being very difficult, would make it impossible to develop a consistent
effective operator, but the first should be pursued. One might also use our GCM
wave functions as a starting point for refinement by the ``multi-reference''
version of the In-Medium Similarity Renormalization Group. Work in that
direction is in progress.
\section{Acknowledgments}
We would like to thank J. Men\'endez for providing us the unpublished effective
interaction GCN2850, J.\ Simonis for providing matrix elements of normal-ordered
three-nucleon interactions, and N. Hinohara, M. Horoi, J. Men\'endez, T. R.
Rodr{\'i}guez, and J. M. Yao for helpful discussions. This work has been
supported by U.S.\ Department of Energy grants DE-FG0297ER41019, DE-SC0008641,
and DE-SC0004142, by by the National Research Council of Canada, and by NSERC.
We used allocations of computing resources at the U.S.\ National Energy Research
Scientific Computing Center (NERSC) and the J\"ulich Supercomputing Center
(JURECA) to carry out computations. Finally, we thank the Institute for Nuclear
Theory at the University of Washington for its hospitality and the Department of
Energy for partial support during the completion of this work.
|
\section{#1}\setcounter{equation}{0}}
\newcommand{\begin{eqnarray}\displaystyle}{\begin{eqnarray}\displaystyle}
\newcommand{\end{eqnarray}}{\end{eqnarray}}
\renewcommand{\theequation}{\arabic{equation}}
\def{\hbox{ 1\kern-.8mm l}}{{\hbox{ 1\kern-.8mm l}}}
\def{\hbox{ 0\kern-1.5mm 0}}{{\hbox{ 0\kern-1.5mm 0}}}
\def\,{\rm Tr}\, {\,{\rm Tr}\, }
\def\,{\rm Tr}_M^{q,y}\, {\,{\rm Tr}_M^{q,y}\, }
\def\tau{\tau}
\def\eta{\,{\rm e}}
\def\hat{n}{\hat{n}}
\def\tilde{g}{\tilde{g}}
\def{\mathcal A}{{\mathcal A}}
\def{\mathcal J}{{\mathcal J}}
\def{\mathcal K}{{\mathcal K}}
\def{\mathcal N}{{\mathcal N}}
\def{\mathcal O}{{\mathcal O}}
\def{\mathcal U}{{\mathcal U}}
\def\tilde{\mathcal O}{\tilde{\mathcal O}}
\def{\mathbb Z}{{\mathbb Z}}
\def{\rm Tr}{{\rm Tr}}
\def\slashed{\slashed}
\def\left{\left}
\def\right{\right}
\def\iota{\iota}
\def\phi{\phi}
\def\Delta{\Delta}
\def\sigma{\sigma}
\def\gamma{\gamma}
\def\Gamma{\Gamma}
\def\mu{\mu}
\def\bar{\mu}{\bar{\mu}}
\def\bar{w}{\bar{w}}
\def\eta{\eta}
\def\alpha{\alpha}
\def\dot{\alpha}{\dot{\alpha}}
\def\bar{a}{\bar{a}}
\def\tilde{\eta}{\tilde{\eta}}
\def\bar{Q}{\bar{Q}}
\def\o[#1]{{\rm O}\left({#1}\right)}
\def\dotl[#1,#2]{\left\langle #1, #2 \right\rangle}
\def\dotlb[#1,#2]{[ #1, #2 ]}
\def\dotp[#1,#2]{(#1) \cdot (#2)}
\def\>{\rangle}
\def\<{\langle}
\def\tilde{l}{\tilde{l}}
\def\tilde{\mu}{\tilde{\mu}}
\def\mathbf{X}{\mathbf{X}}
\def\text{S}^2{\text{S}^2}
\def\ss2/\mathbb{Z}_N{\text{S}^2/\mathbb{Z}_N}
\def\text{AdS}_2{\text{AdS}_2}
\def\text{AdS}_2/\mathbb{Z}_N{\text{AdS}_2/\mathbb{Z}_N}
\def\ss2\otimes\ss2{\text{S}^2\otimes\text{S}^2}
\def\adss\otimes\ss2{\text{AdS}_2\otimes\text{S}^2}
\def\l(\ss2\otimes\ss2\r)/\mathbb{Z}_N{\left(\text{S}^2\otimes\text{S}^2\right)/\mathbb{Z}_N}
\def\l(\adss\otimes\ss2\r)/\mathbb{Z}_N{\left(\text{AdS}_2\otimes\text{S}^2\right)/\mathbb{Z}_N}
\def\mathbb{Z}_N{\mathbb{Z}_N}
\def\mathcal{A}{\mathcal{A}}
\def\bar{s}{\bar{s}}
\def\kappa{\kappa}
\def\tilde{k}{\tilde{k}}
\def\tilde{m}{\tilde{m}}
\DeclareMathOperator{\sech}{sech}
\DeclareMathOperator{\cosech}{cosech}
\DeclareMathOperator{\expo}{exp}
\begin{document}
\baselineskip 24pt
\begin{center}
{\Large \bf Logarithmic Corrections to Entropy of Magnetically Charged AdS$_4$ Black Holes }
\end{center}
\vskip .6cm
\medskip
\vspace*{4.0ex}
\baselineskip=18pt
\centerline{\large \rm Imtak Jeon$^a$, Shailesh Lal$^b$}
\vspace*{4.0ex}
\centerline{\large \it $^a$Harish-Chandra Research Institute}
\centerline{\large \it Chhatnag Road, Jhusi,
Allahabad 211019, India}
\centerline{\large \it $^b$LPTHE -- UMR 7589, UPMC Paris 06,}
\centerline{\large \it Sorbonne Universit{\'e}s, Paris 75005, France}
\vspace*{1.0ex}
\centerline{\small E-mail: <EMAIL>, <EMAIL>}
\vspace*{5.0ex}
\centerline{\bf Abstract} \bigskip
We compute logarithmic corrections to the entropy of a magnetically charged extremal black hole in AdS$_4 \times S^7$ using the quantum entropy function and discuss the possibility of matching against recently derived microscopic expressions.
\vfill \eject
\baselineskip=18pt
\tableofcontents
\section{Introduction}
Providing a microscopic interpretation to the Bekenstein Hawking formula in the context of certain classes of supersymmetric extremal black holes in flat space has been a main success of string theory as a theory of quantum gravity \cite{Strominger:1996sh,Dijkgraaf:1996it,Pioline:2005vi,Shih:2005qf,David:2006yn,Sen:2008ta}. The expression for the microscopic entropy obtained by explicit enumeration and counting of black hole microstates in these cases contains the area law as the leading formula, but also contains higher-derivative and quantum corrections to it. We refer the reader to \cite{Sen:2007qy} for a review of these developments as well as more exhaustive references. Importantly, since extremal black holes expectedly possess an AdS$_2$ factor in the near horizon geometry, one may use the AdS$_2$/CFT$_1$ correspondence to provide an alternative, but equivalent, definition of the quantum entropy of extremal black holes in string theory. This proposal is known as the quantum entropy function, and
for extremal black holes carrying charges $\vec{q}\equiv q_i$ \cite{Sen:2008yk,Sen:2008vm},
\begin{equation}\label{qef}
d_{hor}\left(\vec{q}\right)\equiv\left\langle\exp\left[i\oint q_i d\theta\mathcal{A}_\theta^i\right] \right\rangle_{\text{AdS}_2}^{finite},
\end{equation}
where $d_{hor}$ is the full quantum degeneracy associated with the black hole horizon, and $\mathcal{A}_\theta^i$ is the component of the $i^{\text{th}}$ gauge field along the boundary of the AdS$_2$. In this picture the entropy associated to the horizon degrees of freedom of an extremal blackhole is essentially the free energy corresponding to the partition function \eqref{qef}. The superscript `finite' reminds us that the quantity on the right hand side of \eqref{qef} is naively divergent due to the infinite volue of AdS$_2$ but this divergence may be regulated in accordance with general principles of the AdS/CFT correspondence and a cutoff-insensitive finite part extracted, which is then identified to $d_{hor}$ \cite{Sen:2008yk,Sen:2008vm,Sen:2009vz}. The path integral is carried out over all fields that asymptote to the black hole near horizon geometry. In the context of supersymmetric extremal black holes in flat space, this has been evaluated using saddle point techniques \cite{Sen:2009vz,Banerjee:2010qc,Banerjee:2011jp,Sen:2011ba,bmpvlog,Gupta:2013sva,Gupta:2014hxa} as well as supersymmetric localization \cite{Dabholkar:2010uh,Dabholkar:2011ec,Gupta:2012cy,Dabholkar:2014ema,Murthy:2015yfa,Gupta:2015gga,Murthy:2015zzy} and the answer matched with microscopic results wherever available. Importantly, even in the cases where the full microscopic formula is unavailable, this quantity may be evaluated at least using saddle-point methods to gain some insight into the full microscopic formula \cite{Sen:2011ba,Gupta:2014hxa}.
In contrast, the situation for entropy computations extremal black holes in AdS space is relatively in its infancy. While the Wald entropy may be computed for such black holes using for example the entropy function formalism, the quantum corrections to it are still unknown. In this situation, it would clearly be of interest to evaluate \eqref{qef} for such black holes to obtain the set of quantum corrections to the Wald formula. Such computations are further motivated by the recent proposal for a \textit{microscopic} computation of a CFT$_3$ index argued to capture the quantum entropy of an AdS$_4$ extremal black hole \cite{Benini:2015eyy}.
In this note we will focus on the computation of \eqref{qef} in the semi-classical approximation where the black hole has a large length scale $a$ associated with it, which sets the scale for the area $A_H$ of the event horizon. In the case of extremal black holes in flat space, this large length scale arises when the charges of the black hole are taken to be large. In the present case of AdS black holes, this corresponds to taking the the rank $N$ of the gauge group in the dual CFT to be large, while the charges themselves are not scaled.
One saddle--point of the path integral \eqref{qef} is the black hole's near--horizon geometry itself. By evaluating the on--shell action on this field configuration, it is possible to show that \cite{Sen:2008yk,Sen:2008vm}
\begin{equation}
d_{hor}\simeq e^{S_{\text{Wald}}},
\end{equation}
and hence the quantum entropy function produces as the leading contribution, Wald's formula for black hole entropy \cite{Wald:1993nt}. Now, the next term in the large--charge expansion of the black hole entropy is the so called \textit{log term}, i.e.\footnote{Here we take the leading answer to be Bekenstein--Hawking.}
\begin{equation}
S_{\text{quantum}}= {A_{H}\over 4} + c\ln A_{H},
\end{equation}
where $c$ is a coefficient which depends on the details of the quantum gravity that the black hole is embedded in. For example, the same four--dimensional black hole which is a quarter--BPS black hole in $\mathcal{N}=4$ supergravity may be embedded as a one--eighth BPS black hole in $\mathcal{N}=8$ supergravity and while the leading Bekenstein--Hawking answer is the same for the black hole, the log terms computed in both theories are different \cite{Banerjee:2011jp}, and match with the microscopic computations carried out respectively in $\mathcal{N}=4$ and $\mathcal{N}=8$ string compactifications. This matching is an important test of the consistency of the quantum entropy function proposal.
The main reason why the log term is an important contribution to the microscopic formula is that it is a genuinely quantum correction to the Bekenstein Hawking formula, but is determined completely from one-loop fluctuations of massless fields of two-derivate supergravity, which essentially constitute the IR data of the black hole \footnote{The term `massless' has to be carefully defined on curved manifolds. The more precise statement is that the eigenvalues of the kinetic operator should scale as ${1\over a^2}$, which is how the eigenvalues of the kinetic operator over a massless scalar, $-\Delta$ would scale.}. To see this, let us recall some elements of a scaling argument presented in \cite{Sen:2012dw}. Consider the $\ell$-loop free energy for a theory defined on a $D$ dimensional background with a length scale $a$ associated with it. A typical Feynman graph contributing to this quantity would scale as (see \cite{Sen:2012dw} for details)
\begin{equation}\label{scaling}
\ell_P^{(D-2)(\ell-1)}a^{-(D-2)(\ell-1)}\int^{a/\sqrt{\epsilon}} d^{D\ell}\tilde{k} \,\tilde{k}^{2-2\ell} F\left(\tilde{k}\right),
\end{equation}
where $\tilde{k}=ka$ and $\lbrace\,k\,\rbrace$ are the loop momenta, and $F$ is a function which approaches 1 at large values of its arguments. By focusing on the regime where all loop momenta are of the same order, and working at large $\tilde{k}$, we see that a $\ln a$ term arises from the $\tilde{k}^{2\ell-2-D\ell}$ term in the $\tfrac{1}{\tilde{k}}$ expansion of $F$ at large $\tilde{k}$, and the full $a$ dependence of this term is
\begin{equation}
\left({1\over a}\right)^{(D-2)(\ell-1)}\ln a,
\end{equation}
which is highly suppressed in the large $a$ limit unless $\ell=1$. This verifies the above claim that only one-loop fluctuations give rise to the log term. Further, by considering different scaling regimes where various subsets of loop momenta scale to be much larger than the rest, one may verify the fact that only the two derivative sector of massless fields contributes to the log term. However, we do not do so here and instead refer the reader to \cite{Sen:2012dw} for those details.
Therefore, as argued earlier, the log term may be regarded as an IR probe of the microscopic theory, in the sense that any putative microscopic description of the black hole must correctly reproduce not only the leading Bekenstein Hawking area law, but also the log correction to it.
In this note we shall compute the log term for a class of magnetically charged extremal black holes which asymptote to AdS$_4\times S^7$ for which a complete expression for the microscopic entropy has recently been computed via the computation of a topologically twisted index in ABJM theory \cite{Benini:2015eyy}. We omit details of the microscopic formula, referring the reader to \cite{Benini:2015eyy} as well as the companion papers \cite{Benini:2015noa,Benini:2016rke}. Further, the only features of the near horizon geometry which shall be relevant to us are that it is AdS$_2\times S^2\times S^7$ where the $S^7$ is bundled over the $S^2$ and that the AdS$_2$, $S^2$ and $S^7$ have a common length scale $a$ associated to them. That is, the metric over the full 11 dimensional near horizon geometry can be brought to the form $g_{\mu\nu} = a^2 g^{(0)}_{\mu\nu}$ where the metric $g^{(0)}$ and the coordinates are $a$ independent.
Further the $S^2$ has $SO(3)$ isometry, while the $S^7$ has $U(1)^4$ isometry. These inputs suffice for the macroscopic computation of the log term as a prediction for the microscopic formula of \cite{Benini:2015eyy}. We shall finally discuss a few aspects of the proposed match. \footnote{While this draft was being readied for submission, we learned of \cite{lpz} where this comparison has been carried out in a numerical scheme to find a mismatch.} Details of the full black hole solution are reviewed in \cite{Benini:2015eyy}.
\section{The Log Term from the Quantum Entropy Function}
We will now describe how the log term may be extracted from the path integral \eqref{qef}. To do so, it is useful to phrase the problem more generally. In particular, we consider a theory in $D$ dimensions, with a dynamical field $\Phi$, admitting a saddle-point which is a background with length scale $a$.
\begin{equation}
\mathcal{Z}\left[\Phi\right]=\int\left[\mathcal{D}\Phi\right] e^{-{1\over\hbar}S\left[\Phi\right]}.
\end{equation}
Then, as argued from Equation \eqref{scaling}, to extract the term in the free energy proportional to $\ln a$, it is sufficient to concentrate on the one-loop partition function of the theory. The techniques for this analysis are well-known and we refer the reader to \cite{bmpvlog,Bhattacharyya:2012ye,Gupta:2013sva,Gupta:2014hxa} for accounts of how these computations are carried out.
Firstly, the one-loop partition function is then given by
\begin{equation}
\mathcal{Z}_{1-\ell}={{\det}'\mathcal{O}}^{-{1\over 2}}\cdot\left(\mathcal{Z}_{\text{zero}}\right)^{n_0},
\end{equation}
where $\det'\mathcal{O}$ is the determinant of $\mathcal{O}$ evaluated over its non--zero modes, $n_0$ is the number of zero modes of $\mathcal{O}$, and$\mathcal{Z}_{\text{zero}}$ is the residual zero-mode integral. Therefore
\begin{equation}
\ln\mathcal{Z}_{1-\ell}=-{1\over 2}\ln{\det}'\mathcal{O} + n_0\ln\mathcal{Z}_{\text{zero}},
\end{equation}
The final result is that the coefficient of the $\ln a$ term in the free energy computed about this saddle-point depends on $K\left(t;0\right)$, which is the $t^0$ coefficient of the heat kernel expansion of the kinetic operator $\mathcal{O}$ about this saddle-point, and $\beta_{\Phi}$, which determines how the zero mode contribution $\mathcal{Z}_{\text{zero}}$ to the path integral scales with $a$.
\begin{equation}
\mathcal{Z}_{\text{zero}} = a^{\beta_{\Phi}}\hat{\mathcal{Z}}_{\text{zero}},
\end{equation}
where $\hat{\mathcal{Z}}_{\text{zero}}$ does not scale with $a$.
We eventually obtain the formula
\begin{equation}\label{logterm}
\ln\mathcal{Z}_{1-\ell}= K\left(t;0\right)\ln a+\left(\beta_{\Phi}-1\right)n_0\ln a,
\end{equation}
It is well known that in odd--dimensional spacetimes, $K\left(0;t\right)=0$ and hence we have the formula
\begin{equation}
\ln\mathcal{Z}_{1-\ell}=\left(\beta_{\Phi}-1\right)n_0\ln a.
\end{equation}
The extension to the case of multiple fields $\lbrace\,\Phi\,\rbrace$ having zero modes is apparent.
\begin{equation}\label{logtermzeromodes}
\ln\mathcal{Z}_{1-\ell}=\sum_{\phi\in\lbrace\,\Phi\,\rbrace}\left(\beta_{\phi}-1\right)n^{\phi}_0\ln a,
\end{equation}
where $n^{\phi}_0$ is the number of zero modes of the kinetic operator over the field $\phi$. The specific values for $\beta_{\Phi}$ that will be relevant to us are for the vector, the graviton, the three-form and the gravitino. These are given by
\begin{equation}\label{betas}
\beta_{v}={D-2\over 2},\quad \beta_{m}={D\over 2},\quad \beta_{f}=D-1,\quad \beta_C = {D\over 2}-3.
\end{equation}
Here $v$ denotes the vector field, $m$ the metric or the graviton, $f$ the gravitino, and the corresponding $\beta$ values
have been listed in Equation (2.37) of \cite{bmpvlog}. $C$ denotes the three-form field and its $\beta$ value may be determined exactly in the same manner as the previous fields \cite{bmpvlog,Bhattacharyya:2012ye}. We start with the expression for the normalization of the field $C_{MNP}$ in $D$-dimensions.
\begin{equation}
\int\left[\mathcal{D}C_{MNP}\right]\,\expo\left[-\int d^dx\,\sqrt{g} g^{MU}g^{NV} g^{PW} C_{MNP}C_{UVW}\right]=1.
\end{equation}
Here the metric scales as $g_{MN}= a^2 g^{(0)}_{MN}$ where $g^{(0)}$ does not scale with $a$. Therefore we have
\begin{equation}
\int\left[\mathcal{D}C_{MNP}\right]\,\expo\left[-a^{D-6}\int d^dx\,\sqrt{g^{(0)}} g^{(0)MU}g^{(0)NV} g^{(0)PW} C_{MNP}C_{UVW}\right]=1.
\end{equation}
Hence the correctly normalized integration measure is
\begin{equation}
\prod_{x,\left(MNP\right)}d\left(a^{{D\over 2}-3} C_{MNP}\right).
\end{equation}
From this we can obtain that the zero mode of $C_{MNP}$ corresponds to
\begin{equation}
\beta_{C} = {D\over 2}-3.
\end{equation}
\section{Counting the Number of Zero Modes}
From the above discussion it is clear that if we are to extract the contribution to the (logarithm of the) quantum entropy function which scales as $\ln a$, where $a$ is the length scale associated with the radii of AdS$_2$, $S^2$ and $S^7$, it is enough to compute the zero modes of the fields appearing in the path integral \eqref{qef}. Zero modes can in principle appear in the massless spectrum of AdS$_2$ fields obtained from fields of the 11 dimensional supergravity reduced on to $S^2\times S^7$. In this section we will enumerate these zero modes and compute their contribution to the quantum entropy of magnetically charged AdS$_4$ black holes. In isolating these zero modes, a special role is played by the so-called \textit{discrete modes} of the Laplacian for spin-1 and spin-2 fields, and the Dirac operator for spin-$\tfrac32$ fields on AdS$_2$ explicitly enumerated in \cite{Camporesi:1994ga,Camporesi:1995fb} respectively, and counting the total number of zero modes is essentially equivalent to counting the total number of discrete modes of these fields. These modes have also been listed in \cite{Banerjee:2011jp,bmpvlog,Gupta:2014hxa}. An important observation for us is that it turns out that naively the number of zero modes for each of these fields turns out to be infinite. However, this divergence turns out to be essentially equivalent to the the volume divergence of the free energy and may be regulated in the same way. Two slightly different, but equivalent, procedures for doing this are available in \cite{Banerjee:2010qc,Banerjee:2011jp,bmpvlog} and \cite{Gupta:2013sva,Gupta:2014hxa} and we shall use those results for the regularized number of zero modes in the computations that follow.
\subsection*{Bosonic Zero Modes}
The bosonic fields are the graviton $h_{MN}$ and the 3-form $C_{MNP}$. Quantization of the graviton gives rise to a ghost vector field but this has no zero modes. Quantization of the 3-form $C$ gives rise to a ghost 2-form $B$ with Grassmann odd statistics, and a ghost-for-ghost 1-form $A$ with Grassmann even statistics. We use the following conventions: $M$ is an 11-d vector index, $\mu$ is an AdS$_2$ index, $a$ is an $S^2$ index and $i$ is an $S^7$ index. Finally $\alpha$ is either $a$ or $i$.
Consider the metric zero modes first. The graviton $h_{MN}$ decomposes into the AdS$_2$ graviton $h_{\mu\nu}$, 3 massless AdS$_2$ vectors $h_{\mu a}$, and 4 massless AdS$_2$ vectors $h_{\mu i}$, along with the AdS$_2$ scalars $g_{ia},\,g_{ij}$ and $g_{ab}$. In counting the number of massless vectors, we used the fact that these are given by
\begin{equation}
h_{\mu\alpha} = v_{\mu}k_{\alpha},
\end{equation}
where $k_{\alpha}$ is a Killing vector along an internal direction. Since the internal space has SU(2) $\otimes$ U(1)$^{\otimes 4}$ isometry, there are $3 + 4 =7$ Killing vectors. Each massless vector field on AdS$_2$ contributes
\begin{equation}
n_0^v = -1,
\end{equation}
hence there are $-7$ zero modes from the $h_{\mu\alpha}$. Also there are $-3$ zero modes from the AdS$_2$ metric $h_{\mu\nu}$. Therefore total number of metric zero modes is
\begin{equation}
n_0^m = -7 -3 =-10.
\end{equation}
Then the contribution to the log term from the 11d metric zero modes is
\begin{equation}\label{logm}
\delta Z|_{\text{metric}} = \left(\beta_m-1\right) n_0^m \ln a = \left({11\over 2}-1\right)\left(-10\right) = -45\ln a.
\end{equation}
We next consider the 3-form field $C_{MNP}$. Its quantization was carried out in \cite{Siegel:1980jj,ThierryMieg:1980it,Copeland:1984qk} and is reviewed in Section 3.1 of \cite{Bhattacharyya:2012ye}. We will just require the following result. The quantization of a $p$-form requries $p$ generalized ghost fields which are $(p-j)$-forms, where $j$ runs from $1$ to $p$. Further the logarithmic contribution of all these fields to the free energy may be packaged into the expression
\begin{equation}
\Delta F = \sum_{j=0}^{p}\left(-1\right)^j\left(\beta_{p-j}-j-1\right)n^0_{\mathcal{O}_{p-j}}\ln a,
\end{equation}
where $\mathcal{O}$ is the kinetic operator. For $j=1,\ldots,p$ which are the generalized ghost fields, this is just the Hodge Laplacian. For $j=0$, which is the physical field, this can have couplings to background fluxes as well. We now consider the reduction of $C_{MNP}$ onto AdS$_2$. This firstly leads to $C_{\mu\nu a}\,C_{\mu\nu i}$ which are two-forms on AdS$_2$ which are Hodge duals of scalars. These have no zero modes. Next, we consider a 3-form which is the wedge product of a 1-form along AdS$_2$ with a harmonic 2-form along $S^2\times S^7$.
\begin{equation}
C^{(3)} = C_{(1)}^{(AdS_2)}\wedge C_{(2)}^{(S^2\times S^7)}.
\end{equation}
The number of such harmonic forms is given by the second Betti number $b_2$ of the manifold, which is 1 in this case, as may be readily seen from multiplying the Poincar\'e polynomials of $S^2$ and $S^7$.
Hence there is a single massless vector fields along AdS$_2$ from the 11 dimensional 3-form field. Finally we have the scalars $C_{abi},\,C_{aij},\,C_{ijk}$ which don't have zero modes. Hence the 3-form contributes
\begin{equation}\label{logc}
\delta Z|_{\text{C}} = \left(\beta_C-1\right) n_0^C \ln a = \left({11\over 2}-3-1\right)\left(-1\right) = -{3\over 2}\ln a
\end{equation}
to the log term. We next consider the ghost $B_{MN}$ which arises from the quantization of $C$. This decomposes into the following massless fields on AdS$_2$. First we have $B_{\mu\nu}$ which contributes no zero modes. Next we may obtain a massless 1-form on AdS$_2$ from this field by decomposing $B_{MN}$ into a wedge product of an AdS$_2$ 1-form with a harmonic 1-form on $S^2\times S^7$. The number of such harmonic 1-forms is the first Betti number of $S^2\times S^7$ which is zero. Finally we have the scalars $B_{ab},\,B_{ai}$ and $B_{ij}$, which contribute no zero modes. Therefore the contribution to the log term is
\begin{equation}\label{logb}
\delta Z|_{\text{B}} = -\left(\beta_B-2\right) n_0^B \ln a = 0.
\end{equation}
The overall minus sign is on account of Grassmann odd statistics of this field. Finally we have the ghost-for-ghost field $A_M$ from the quantization of $B$. This leads to one massless vector field on AdS$_2$ and therefore
\begin{equation}\label{loga}
\delta Z|_{\text{A}} = \left(\beta_A-3\right) n_0^A \ln a = \left({9\over 2}-3\right)\left(-1\right) = -{3\over 2}\ln a.
\end{equation}
We therefore add \eqref{logm}, \eqref{logc}, \eqref{logb} and \eqref{loga} to obtain
\begin{equation}\label{logbos}
\delta Z = \left(-45-\frac{3}{2}-\frac{3}{2}\right)\ln a = -48\ln a.
\end{equation}
The residual scalar generalized ghost has no zero modes.
\subsection*{Fermionic Zero Modes}
To count fermion zero modes, we need to compute the regularized number of discrete modes $\xi^{(k)+}_\mu$ and $\hat{\xi}^{(k)+}_\mu$, $k=1,2,\ldots$, on AdS$_2$. The relevant computations are available in \cite{Banerjee:2011jp,bmpvlog,Gupta:2014hxa} and we only mention final results. Firstly, it may be shown that the regularized number of modes of both $\xi$ and $\hat{\xi}$ is given by $-1$. Further, these modes should be tensored with the spinors associated with directions transverse to AdS$_2$. This will give rise to additional multiplicity factors. To determine the multiplicity, we note first that the near horizon geometry of the black hole has a superconformal symmetry $su(1,1|1)$ with fermionic generators $G_{n}^\alpha$ where $\alpha = \pm$ and $n\in\mathbb{Z}+\tfrac12$. The gravitino zero modes we consider are associated with the generators $G_n^\alpha$ where $\vert n\vert \geq \tfrac32$. In particular, we identify $G^\alpha_{n}$ with $n\geq\tfrac32$ with $\xi^{k}$ where $n=k+\tfrac12$ and $G^\alpha_{n}$ with $n\leq -\tfrac32$ with $\xi^{k}$ where $n=-k-\tfrac12$. Hence, there is an overall multiplicity factor of 2 coming from $\alpha$ taking values $+$ and $-$. Therefore the number of fermion zero modes is
\begin{equation}
n_0^f = (-1-1)\times 2 = -4.
\end{equation}
Then the contribution to the log term from the fermionic zero modes is
\begin{equation}\label{logf}
\delta Z|_{\text{fermions}} = -\left(\beta_f-1\right) n_0^f \ln a = -\left(10-1\right)\left(-4\right) = +36\ln a.
\end{equation}
The overall minus is on account of Grassmann odd statistics.
Adding \eqref{logbos} and \eqref{logf} we see that the log term is given by
\begin{equation}
\Delta F = \left(-48 +36\right)\ln a = -12\ln a.
\end{equation}
\section{On the comparison with Microscopics}
We have so far computed the log term for magnetically charged AdS$_4$ extremal black holes using the quantum entropy function. In this section we shall very briefly discuss how in principle a match with the proposed microscopic answer of \cite{Benini:2015eyy} may be carried out. At first glance one may expect that the large $N$ expansion of the logarithm of the topologically twisted index in the CFT computed in \cite{Benini:2015eyy}
may be matched term by term with the large $a$ expansion of the logarithm of the quantum entropy function. We note however, that several steps should in principle be necessary to carry out before a match can be meaningfully proposed.
Firstly, the index computed by \cite{Benini:2015eyy} measures the black hole entropy in the grand canonical ensemble as it employs the AdS$_4$/CFT$_3$ correspondence. In contrast, the natural boundary conditions for the quantum entropy function \eqref{qef} pick out the \textit{microcanonical} ensemble \cite{Sen:2008vm}. While the choice of ensemble is irrelevant in the strict large $N$ limit, it is typically important when finite $N$ effects are taken into account. Indeed examples exist where the choice of ensemble is explicitly shown to affect the value of the log term \cite{Sen:2012dw}. Therefore, as a first step, we expect that one would need to go to the microcanonical ensemble when computing the CFT answer to match with the quantum entropy function.
Second, it is not obvious how the naive large $N$ scalings of the log of the index are to be reproduced by the quantum entropy function. In particular, it appears from the analysis of \cite{Benini:2015eyy} that $\ln Z$ scales as
\begin{equation}
\ln Z \sim N^{3/2}+ \mathcal{O}\left(N\ln N\right).
\end{equation}
The $N^{3/2}$ term precisely matches with the Bekenstein Hawking entropy of the black hole, but the possible subleading term of order $N\ln N$ is particularly surprising from the point of view of the general scalings of contributions to the quantum entropy as expected from the quantum entropy function. In particular, on going through the scaling analysis of Equation \eqref{scaling}, it is apparent that we do not expect a term of the form $N\ln N$, and that this term should drop from the CFT$_3$ answer if a match with the quantum entropy function is to be possible.
Remarkably, it seems that a numerical estimate of the large $N$ behaviour of the index produces the pattern one would naturally expect from the quantum entropy function scalings, including an absence of the $N\,\ln N$ term, and the presence of the $\ln N$ term, albeit with a mismatching coefficient \cite{lpz}. This remarkable feature should certainly be better understood by carrying out a systematic large $N$ expansion of the CFT index while accounting for the choice of ensemble as we have indicated above. It would be interesting to return to these questions in the future.
\section*{Acknowledgements}
We would like to thank Rajesh Gupta, Nick Halmagyi, Euihun Joung, Tomoki Nosaka, Chiung Hwang, Leopoldo Pando Zayas, Ashoke Sen, Piljin Yi and Alberto Zaffaroni for helpful discussions. SL thanks the Korea Institute of Advanced Study and Kyung Hee University for hospitality while this work was carried out. SL's work is supported by a Marie Sklodowska Curie 2014 Individual Fellowship.
|
\section{Introduction}\label{sec:intro}
Several methods of generation of \emph{two-dimensional languages} (also called \emph{array languages} or \emph{picture languages})
have been proposed in the literature, extending the techniques and results of formal string language theory.
A picture is considered as a rectangular array of terminal symbols in the two-dimensional plane.
Models based on grammars or automata as well as those based on theoretical properties of the string languages are well-known and have been extensively investigated. We refer the interested readers to books and surveys like the ones by Rosenfeld~\cite{ros:pic}, Wang~\cite{wan:arr}, Rosenfeld~and~Siromoney~\cite{ros:pic2}, Giammarresi~and~Restivo~\cite{gia:two}, or Morita~\cite{mor:two}.
For example, regular \emph{string} languages (also known as recognizable string languages) can be characterized in terms of local languages and projections.
Based on a similar idea,
the class REC of recognizable \emph{picture} languages (see Giammarresi~and~Restivo~\cite{gia:rec}) was proposed as a two dimensional counterpart of regular string languages.
In this work, we attempt to generalise a class of string languages to the two-dimensional case, which also provides several desirable features and has therefore attracted considerable interest over the last three decades
in the formal language theory community as well as in the learning theory community: Angluin's pattern languages (see \cite{ang:fin2}). \par
In this context, a \emph{pattern} is a string over an alphabet $\{x_1, x_2, x_3, \ldots\}$ of \emph{variables}, e.\,g., $\alpha := x_1\,x_1\,x_2\,x_2\,x_1$. For some finite alphabet $\Sigma$ of \emph{terminal symbols}, the \emph{pattern language} described by $\alpha$ (with respect to $\Sigma$) is the set of all words over $\Sigma$ that can be derived from $\alpha$ by uniformly substituting the variables in $\alpha$ by (non-empty) terminal words. For example, if $\Sigma := \{\ta, \tb, \tc\}$, then $u := \tb \tc \tb\tb \tc \tb\tc \tc \ta\tc \tc \ta\tb \tc \tb$ and $v := \ta \tb\ta \tb\ta \tb \ta\ta \tb \ta\ta \tb$ are words of the pattern language given by $\alpha$, since replacing $x_1$ by $\tb \tc \tb$ and $x_2$ by $\tc \tc \ta$ turns $\alpha$ into $u$ and replacing $x_1$ by $\ta \tb$ and $x_2$ by $\ta \tb \ta$ turns $\alpha$ into $v$. On the other hand, the word $\tc \ta \tb \ta \tc \ta \tb \ta \tb \ta \tb \ta$ is not a member of the pattern language of $\alpha$.\par
One of the most notable features of pattern languages is that they have natural and compact human readable descriptors (or generators), namely the patterns. In particular, this advantage becomes evident when patterns are compared to other language descriptors as, e.\,g., grammars or automata, which are usually quite involved even though the language they describe is rather simple. Nevertheless, patterns can compete with common automata models and grammars in terms of expressive power and their practical relevance is demonstrated by the widespread use of so-called extended regular expressions with backreferences, which implicitly use the concept of patterns and are capable of defining all pattern languages.\footnote{In fact, these extended regular expressions with backreferences are nowadays a standard element of most text editors and programming languages (cf. Friedl~\cite{fri:mas}). } \par
The main goal of this paper is to generalise the concept of patterns as language descriptors to the two-dimensional case, while preserving the desirable features of (one-dimensional) pattern languages, i.\,e., the simplicity and compactness of their descriptors. The work done so far on two-dimensional languages demonstrates that there are difficulties that seem to be symptomatic for the task of generalising a class of string languages to the two-dimensional case. Firstly, such a generalisation is usually accompanied with a substantial increase in complexity of the descriptors (e.\,g., when extending context-free or contextual grammars to the two-dimensional case (see Fernau~et~al.~\cite{FerFreHol98a}, Freund~et~al.~\cite{fre:con})) and, secondly, there are often many competing and seemingly different ways to generalise a specific class of string languages, which all can be considered natural (e.\,g., it is still on debate what the appropriate two-dimensional counterpart of the class of regular languages
might be (see Giammarresi~et~al.~\cite{Giaetal96}, Matz~\cite{mat:rec})). Our two-dimensional patterns, to be introduced in this work, are as simple and compact as their one-dimensional counterparts. Although there are several different possibilities of how these two-dimensional patterns can describe two-dimensional languages, one of these sticks out as the intuitively most natural one. Hence, the model of Angluin's pattern languages seems to be comparatively two-dimensional friendly.\par
Besides the conceptional contribution of this paper, we present a comparison between the expressive power of different classes of two-dimensional pattern languages and an investigation of their closure properties. We conclude the paper by outlining further research questions and possible extensions to the model of two-dimensional pattern languages.
\section{Preliminaries}\label{sec:def}
In this section, we briefly recall the standard definitions and notations regarding one- and two-dimensional words and languages.\par
Let $\mathbb{N} := \{1, 2, 3, \ldots \}$ and let $\mathbb{N}_{0} := \mathbb{N} \cup \{0\}$. For a finite alphabet $\Sigma$, a \emph{string} or \emph{word} (\emph{over $\Sigma$}) is a finite sequence of symbols from $\Sigma$, and $\varepsilon$ stands for the \emph{empty string}. The notation $\Sigma^+$ denotes the set of all nonempty strings over $\Sigma$, and $\Sigma^*:=\Sigma^+ \cup \{ \varepsilon \}$. For the \emph{concatenation} of two strings $w_1, w_2$ we write $w_1 \cdot w_2$ or simply $w_1 w_2$. We say that a string $v \in \Sigma^*$ is a \emph{factor} of a string $w \in \Sigma^*$ if there are $u_1, u_2 \in \Sigma^*$ such that $w = u_1 \cdot v \cdot u_2$. If $u_1$ or $u_2$ is the empty string, then $v$ is a \emph{prefix} (or a \emph{suffix}, respectively) of $w$. The notation $|w|$ stands for the length of a string $w$.\par
A \emph{two-dimensional word} (or \emph{array}) \emph{over $\Sigma$} is a tuple
\begin{equation*}
W := ((a_{1,1}, a_{1,2}, \ldots, a_{1,n}), (a_{2,1}, a_{2,2}, \ldots, a_{2,n}), \ldots, (a_{m,1}, a_{m,2}, \ldots, a_{m,n}))\,,
\end{equation*}
where $m, n \in \mathbb{N}$ and, for every $i$, $1 \leq i \leq m$, and $j$, $1 \leq j \leq n$, $a_{i, j} \in \Sigma$. We define the \emph{number of columns} (or \emph{width}) and \emph{number of rows} (or \emph{height}) of $W$ by $|W|_c := n$ and $|W|_r := m$, respectively. The \emph{empty array} is denoted by $\lambda$, i.\,e., $|\lambda|_c = |\lambda|_r = 0$. For the sake of convenience, we also denote $W$ by $[a_{i, j}]_{m, n}$ or by a matrix of one of the following forms:
\begin{equation*}
\begin{smallmatrix}
a_{1,1} & a_{1,2} & \ldots & a_{1,n} \\
a_{2,1} & a_{2,2} & \ldots & a_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
a_{m,1} & a_{m,2} & \ldots & a_{m,n}
\end{smallmatrix}
,\:\:\:\:\:\:
\begin{bsmallmatrix}
a_{1,1} & a_{1,2} & \ldots & a_{1,n} \\
a_{2,1} & a_{2,2} & \ldots & a_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
a_{m,1} & a_{m,2} & \ldots & a_{m,n}
\end{bsmallmatrix}\,.
\end{equation*}
If we want to refer to the $j^{\text{th}}$ symbol in row $i$ of the array $W$, then we use $W[i, j] = a_{i, j}$. By $\Sigma^{++}$, we denote the set of all nonempty arrays over $\Sigma$, and $\Sigma^{**}:=\Sigma^{++} \cup \{ \lambda \}$. Every subset $L \subseteq \Sigma^{**}$ is an \emph{array language}. \par
Let $W := [a_{i, j}]_{m, n}$ and $W' := [a'_{i, j}]_{m', n'}$ be two non-empty arrays over $\Sigma$.
The \emph{column concatenation} of $W$ and $W'$, denoted by $W \varobar W'$, is undefined if $m \neq m'$ and is the array
\begin{equation*}
\begin{smallmatrix}
a_{1,1} & a_{1,2} & \ldots & a_{1,n} & b_{1,1} & b_{1,2} & \ldots & b_{1,n'} \\
a_{2,1} & a_{2,2} & \ldots & a_{2,n} & b_{2,1} & b_{2,2} & \ldots & b_{2,n'} \\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots \\
a_{m,1} & a_{m,2} & \ldots & a_{m,n} & b_{m',1} & b_{m',2} & \ldots & b_{m',n'}
\end{smallmatrix}
\end{equation*}
otherwise. The \emph{row concatenation} of $W$ and $W'$, denoted by $W \varominus W'$, is undefined if $n \neq n'$ and is the array
\begin{equation*}
\begin{smallmatrix}
a_{1,1} & a_{1,2} & \ldots & a_{1,n} \\
a_{2,1} & a_{2,2} & \ldots & a_{2,n} \\
\vdots & \vdots & \ddots & \vdots \\
a_{m,1} & a_{m,2} & \ldots & a_{m,n} \\
b_{1,1} & b_{1,2} & \ldots & b_{1,n'} \\
b_{2,1} & b_{2,2} & \ldots & b_{2,n'} \\
\vdots & \vdots & \ddots & \vdots \\
b_{m',1} & b_{m',2} & \ldots & b_{m',n'}
\end{smallmatrix}
\end{equation*}
otherwise. Intuitively speaking, the vertical line and the horizontal line in the symbols $\varobar$ and $\varominus$, respectively, indicate the edge where the arrays are concatenated. In order to denote that, e.\,g., $U \varominus V$ is undefined, we also write $U \varominus V = \undef$.
Furthermore, for every array $U$, $U \varobar \lambda = \lambda \varobar U = U \varominus \lambda = \lambda \varominus U = U$ and
$U \varobar \undef = \undef \varobar U = U \varominus \undef = \undef \varominus U = \undef$. Algebraically speaking, if $\Sigma^{**}_{\undef}:=\Sigma^{**}\cup\{\undef\}$, then
$(\Sigma^{**}_{\undef},\varobar,\lambda)$ and $(\Sigma^{**}_{\undef},\varominus,\lambda)$ both form monoids with $\undef$ as an absorbing element.
\begin{example}\label{concatExample1}
Let
\begin{equation*}
W_1 :=
\begin{bsmallmatrix}
\ta & \tb & \ta \\
\tb & \tc & \ta \\
\ta & \tb & \tb \\
\end{bsmallmatrix}
, W_2 :=
\begin{bsmallmatrix}
\tb & \tc \\
\tb & \ta \\
\tc & \ta
\end{bsmallmatrix}
, W_3 :=
\begin{bsmallmatrix}
\ta & \tb & \tc \\
\tc & \tb & \tb
\end{bsmallmatrix}
\text{ and } W_4 :=
\begin{bsmallmatrix}
\ta & \ta \\
\ta & \tb
\end{bsmallmatrix}\,.
\end{equation*}
Then $W_1 \varominus W_2$ and $W_1 \varobar W_3 = \undef$, but
\begin{equation*}
W_1 \varobar W_2 =
\begin{bsmallmatrix}
\ta & \tb & \ta & \tb & \tc \\
\tb & \tc & \ta & \tb & \ta \\
\ta & \tb & \tb & \tc & \ta
\end{bsmallmatrix}
\text{ and } W_1 \varominus W_3 =
\begin{bsmallmatrix}
\ta & \tb & \ta \\
\tb & \tc & \ta \\
\ta & \tb & \tb \\
\ta & \tb & \tc \\
\tc & \tb & \tb
\end{bsmallmatrix}\,.
\end{equation*}
\end{example}
\begin{example}
Let $W_1, W_2, W_3$ and $W_4$ be defined as in Example~\ref{concatExample1}.
\begin{align*}
&(W_1 \varobar W_2) \varominus (W_3 \varobar W_4) = (W_1 \varominus W_3) \varobar (W_2 \varominus W_4) =
\begin{bsmallmatrix}
\ta & \tb & \ta & \tb & \tc \\
\tb & \tc & \ta & \tb & \ta \\
\ta & \tb & \tb & \tc & \ta \\
\ta & \tb & \tc & \ta & \ta \\
\tc & \tb & \tb & \ta & \tb
\end{bsmallmatrix}\,,\\
&(W_1 \varobar W_2) \varominus (W_4 \varobar W_3) =
\begin{bsmallmatrix}
\ta & \tb & \ta & \tb & \tc \\
\tb & \tc & \ta & \tb & \ta \\
\ta & \tb & \tb & \tc & \ta \\
\ta & \ta & \ta & \tb & \tc \\
\ta & \tb & \tc & \tb & \tb
\end{bsmallmatrix}\,,\:\:
(W_1 \varominus W_4) \varobar (W_2 \varominus W_3) = \undef\,.
\end{align*}
\end{example}
Next, we define some operations for array languages. The row and column concatenation for array languages $L_1$ and $L_2$ is defined by $L_1 \varominus L_2 := \{U \varominus V \mid U \in L_1, V \in L_2, U \varominus V \neq \undef\}$ and $L_1 \varobar L_2 := \{U \varobar V \mid U \in L_1, V \in L_2, U \varobar V \neq \undef\}$, respectively. For an array language $L$ and $k \in \mathbb{N}$, $L^{\varominus k}$ denotes the $k$-fold row concatenation of $L$, i.\,e., $L^{\varominus k} := L_1 \varominus L_2 \varominus \ldots \varominus L_k$, $L_i = L$, $1 \leq i \leq k$. The $k$-fold column concatenation, denoted by $L^{\varobar k}$, is defined analogously. The \emph{row} and \emph{column concatenation closure} of an array language $L$ is defined by $L^{\varominus *} := \bigcup_{k = 1}^{\infty} L^{\varominus k}$ and $L^{\varobar *} := \bigcup_{k = 1}^{\infty} L^{\varobar k}$, respectively. Obviously, the row and column concatenation closure of an array language correspond to the Kleene closure of a string language. \par
Now, we turn our attention to some geometric operations for arrays. The \emph{transposition} of an array $U$, denoted by $\transpose{U}$, is obtained by reflecting $U$ along the main diagonal.
The \emph{$\varominus$-reflection} and \emph{$\varobar$-reflection} of $U$, denoted by $U^{\varominus\texttt{R}}$ and $U^{\varobar\texttt{R}}$, respectively,
are obtained by reflecting $U$ along the horizontal and vertical axis, respectively.
The \emph{right turn} and \emph{left turn} of $U$, denoted by $U^{\curvearrowright}$ and $U^{\curvearrowleft}$, respectively, is obtained by turning $U$ through $90$ degrees to the right and to the left, respectively. For example, if $U :=
\begin{bsmallmatrix}
\ta & \tb & \tc & \td \\
\te & \tf & \tg & \tth
\end{bsmallmatrix}$,
then
\begin{align*}
\transpose{U} =
\begin{bsmallmatrix}
\ta & \te \\
\tb & \tf \\
\tc & \tg \\
\td & \tth \\
\end{bsmallmatrix}\,,
U^{\varominus\texttt{R}} =
\begin{bsmallmatrix}
\te & \tf & \tg & \tth \\
\ta & \tb & \tc & \td \\
\end{bsmallmatrix}\,,
U^{\varobar\texttt{R}} =
\begin{bsmallmatrix}
\td & \tc & \tb & \ta \\
\tth & \tg & \tf & \te \\
\end{bsmallmatrix}\,,
U^{\curvearrowright} =
\begin{bsmallmatrix}
\te & \ta \\
\tf & \tb \\
\tg & \tc \\
\tth & \td \\
\end{bsmallmatrix}\,,
U^{\curvearrowleft} =
\begin{bsmallmatrix}
\td & \tth \\
\tc & \tg \\
\tb & \tf \\
\ta & \te \\
\end{bsmallmatrix}\,,
(U^{\curvearrowright})^{\curvearrowright} =
\begin{bsmallmatrix}
\tth & \tg & \tf & \te \\
\td & \tc & \tb & \ta
\end{bsmallmatrix}\,.
\end{align*}
We address left and right turn also as \emph{quarter-turns} below.
Moreover, the twofold right turn (displayed right-most in the example above)
is also known as a \emph{half-turn}.
\par
A special operation considered in the context of arrays is the \emph{conjugation} of an array $U \in \Sigma^{**}$ with $|\Sigma| = 2$, denoted by $U^{\texttt{C}}$, which means that the two symbols of $\Sigma$ are exchanged in $U$, e.\,g.,
$\left(\begin{bsmallmatrix}
\ta & \tb & \ta & \ta \\
\tb & \ta & \ta & \tb \\
\end{bsmallmatrix}\right)^{\texttt{C}} =
\begin{bsmallmatrix}
\tb & \ta & \tb & \tb \\
\ta & \tb & \tb & \ta \\
\end{bsmallmatrix}$.
This can be also viewed as a quite restricted form of a two-dimensional morphism defined in the next section.\par
All these operations for arrays are extended to array languages in the obvious way. \par
Next, we briefly summarise the concept of (one-dimensional) pattern languages as introduced in \cite{ang:fin2} by Angluin. Technically, the version of pattern languages used here are called \emph{nonerasing terminal-free} pattern languages (for an overview of different versions of one-dimensional pattern languages, the reader is referred to \cite{mat:pat} by Mateescu and Salomaa).\par
A (one-dimensional) \emph{pattern} is a string over an alphabet $X := \{x_1, x_2, x_3, \ldots\}$ of \emph{variables}, e.\,g., $\alpha := x_1\,x_1\,x_2\,x_2\,x_1$. In Section~\ref{sec:intro}, we have seen an intuitive definition of the language described by a pattern $\alpha$. This intuition can be formalised in an elegant way by using the concept of (\emph{word}) \emph{morphisms}, i.\,e., mappings $h : \Sigma_1^+ \rightarrow \Sigma_2^+$, which satisfy $h(u \, v) = h(u)\,h(v)$, for all $u, v \in \Sigma_1^+$. In this regard, for some finite alphabet $\Sigma$, the (one-dimensional) \emph{pattern language} of $\alpha$ (with respect to $\Sigma$) is the set $L^{\oneD}_{\Sigma}(\alpha) := \{h(\alpha) \mid h : X^+ \rightarrow \Sigma^+ \text{ is a morphism}\}$. An alternative, yet equivalent, way to define pattern languages is by means of factorisations. To this end, let $\alpha := y_1\,y_2 \ldots y_n$, $y_i \in X$, $1 \leq i \leq n$. Then $L^{\oneD}_{\Sigma}(\alpha)$ is the set of all words $w \in \Sigma^+$ that have a \emph{characteristic} factorisation for $\alpha$, i.\,e., a factorisation $w = u_1\,u_2 \cdots u_n$, such that, for every $i$, $1 \leq i \leq j \leq n$, $y_i = y_j$ implies $u_i = u_j$. It can be easily seen, that these two definitions are equivalent. However, for the two-dimensional case, we shall see that a generalisation of these two approaches will lead to different versions of two-dimensional pattern languages. The class of all one-dimensional pattern languages over the alphabet $\Sigma$ is denoted by $\mathcal{L}^{\oneD}_{\Sigma}$.
We recall the example pattern $\alpha = x_1\,x_1\,x_2\,x_2\,x_1$ and the words $u := \tb \tc \tb\tb \tc \tb\tc \tc \ta\tc \tc \ta\tb \tc \tb$ and $v := \ta \tb\ta \tb\ta \tb \ta\ta \tb \ta\ta \tb$ of Section~\ref{sec:intro} Since $h(\alpha) = u$ and $g(\alpha) = v$, where $h$ and $g$ are the morphisms induced by $h(x_1) := \tb \tc \tb$, $h(x_2) := \tc \tc \ta$ and $g(x_1) := \ta \tb$, $g(x_2) := \ta \tb \ta$, we can conclude that $u, v \in L^{\oneD}_{\Sigma}(\alpha)$, where $\Sigma := \{\ta, \tb, \tc\}$.
\section{Two-Dimensional Pattern Languages}\label{sec:2DPatternLanguages}
As already mentioned, this work deals with the task of generalising pattern languages from the one-dimensional to the two-dimensional case. In order to motivate our approach to solve this task, we first spent some effort on illustrating the general difficulties and obstacles that arise.\par
Abstractly speaking, a pattern language for a given pattern $\alpha$ is the collection of all elements that satisfy $\alpha$. Thus, a sound definition of how elements satisfy patterns directly entails a sound definition of a class of pattern languages. In the one-dimensional case, the situation that a word satisfies a pattern is intuitively clear and it can be defined in several equivalent ways, i.\,e., a word $w$ satisfies the pattern $\alpha$ if and only if
\begin{itemize}
\item $w$ can be derived from $\alpha$ by uniformly substituting the variables in $\alpha$,
\item $w$ is a morphic image of $\alpha$,
\item $w$ has a characteristic factorisation for $\alpha$.
\end{itemize}
We shall now demonstrate that for a two-dimensional pattern, i.\,e., a two-dimensional word over the set of variables $X$, e.\,g., $\alpha :=
\begin{bsmallmatrix}
x_1&x_2\\
x_3&x_1
\end{bsmallmatrix}$,
these concepts do not work anymore or they describe fundamentally different situations. For instance, the basic operation of substituting a single symbol in a word by another word cannot that easily be extended to the two-dimensional case. For example, the replacements $x_1 \mapsto \begin{bsmallmatrix}\ta&\ta\end{bsmallmatrix}, x_2 \mapsto \begin{bsmallmatrix}\tc\\\tc\end{bsmallmatrix}$ and $x_3 \mapsto \begin{bsmallmatrix}\tb\end{bsmallmatrix}$ may turn $\alpha$ into one of the following objects,
\begin{equation*}
\begin{bsmallmatrix}
&&\tc&\\
\ta&\ta&\tc&\\
&\tb&\ta&\ta
\end{bsmallmatrix} ,
\begin{bsmallmatrix}
\ta&\ta&\tc\\
&&\tc\\
\tb&\ta&\ta
\end{bsmallmatrix},
\begin{bsmallmatrix}
&&\tc\\
\ta&\ta&\tc\\
\tb&\ta&\ta
\end{bsmallmatrix},
\begin{bsmallmatrix}
&&&\tc\\
\ta&\ta&&\tc\\
&\tb&\ta&\ta
\end{bsmallmatrix},
\end{equation*}
which are not two-dimensional words, since they all contain holes or are not of rectangular shape and, most importantly, are not uniquely defined. On the other hand, it is straightforward to generalise the concept of a morphism to the two-dimensional case:
\begin{definition}\label{homoDef}
A mapping $h: \Sigma_1^{++} \rightarrow \Sigma_2^{++}$ is a two-dimensional morphism if it satisfies $h(V \varobar W) = h(V) \varobar h(W)$ and $h(V \varominus W)=h(V) \varominus h(W)$ for all $V, W \in \Sigma_1^{++}$.
\end{definition}
Hence, we may say that a two-dimensional word $W$ satisfies a two-dimensional pattern $\alpha$ if and only if there exists a two-dimensional morphism which maps $\alpha$ to $W$. Unfortunately, this definition seems to be too strong as demonstrated by the following example. From an intuitive point of view, the two-dimensional word
$W := \begin{bsmallmatrix}
\ta& \ta& \tb& \ta& \ta& \tb\\
\ta& \ta& \tb& \ta& \ta& \tb\\
\tc& \tc& \tc&\tc&\tc&\tc
\end{bsmallmatrix}$
should satisfy the two-dimensional pattern
$\alpha := \begin{bsmallmatrix}
x_1 & x_1\\
x_2 & x_2
\end{bsmallmatrix}$,
but there is no two-dimensional morphism mapping $\alpha$ to $W$. This is due to the fact that, as pointed out by the following proposition (which has also been mentioned by Siromoney~et~al. in \cite{sir:pic}), a two-dimensional morphism is a mapping with a surprisingly strong condition.
\begin{proposition}\label{morphismProp}
Let $\Sigma_1 := \{a_1, a_2, \ldots, a_k\}$ and $\Sigma_2$ be alphabets. If a mapping $h : \Sigma_1^{++} \rightarrow \Sigma_2^{++}$ is a two-dimensional morphism, then
\begin{equation*}
|h(\begin{bsmallmatrix} a_1 \end{bsmallmatrix})|_c = |h(\begin{bsmallmatrix} a_2 \end{bsmallmatrix})|_c = \ldots = |h(\begin{bsmallmatrix} a_k \end{bsmallmatrix})|_c \text{ and } |h(\begin{bsmallmatrix} a_1 \end{bsmallmatrix})|_r = |h(\begin{bsmallmatrix} a_2 \end{bsmallmatrix})|_r = \ldots = |h(\begin{bsmallmatrix} a_k \end{bsmallmatrix})|_r\,.
\end{equation*}
\end{proposition}
\begin{proof}
We prove the statement of the proposition by contraposition. To this end, we assume that, for some $i, j$, $1 \leq i < j \leq k$, $|h(\begin{bsmallmatrix} a_i \end{bsmallmatrix})|_r \neq |h(\begin{bsmallmatrix} a_j \end{bsmallmatrix})|_r$, which implies $h(\begin{bsmallmatrix} a_i \end{bsmallmatrix}) \varobar h(\begin{bsmallmatrix} a_j \end{bsmallmatrix}) = \undef$. Hence, since $h(\begin{bsmallmatrix} a_i \end{bsmallmatrix} \varobar \begin{bsmallmatrix} a_j \end{bsmallmatrix}) = h(\begin{bsmallmatrix} a_i & a_j \end{bsmallmatrix}) \in \Sigma_2^{++}$, we can conclude that $h(\begin{bsmallmatrix} a_i \end{bsmallmatrix} \varobar \begin{bsmallmatrix} a_j \end{bsmallmatrix}) \neq h(\begin{bsmallmatrix} a_i \end{bsmallmatrix}) \varobar h(\begin{bsmallmatrix} a_j \end{bsmallmatrix})$, which contradicts the morphism property. Similarly, if $|h(\begin{bsmallmatrix} a_i \end{bsmallmatrix})|_c \neq |h(\begin{bsmallmatrix} a_j \end{bsmallmatrix})|_c$, then $h(\begin{bsmallmatrix} a_i \end{bsmallmatrix} \varominus
\begin{bsmallmatrix} a_j \end{bsmallmatrix}) \neq h(\begin{bsmallmatrix} a_i \end{bsmallmatrix}) \varominus h(\begin{bsmallmatrix} a_j \end{bsmallmatrix})$.
\end{proof}
Similarly as in the string case, homomorphisms $h: \Sigma_1^{++} \rightarrow \Sigma_2^{++}$ are uniquely defined by giving the images $h(\Sigma_1)$.
If in particular $h(\Sigma_1)\subseteq\Sigma_2$, we term the resulting
morphism a \emph{letter-to-letter morphism}, while in the even more restricted
case when the restriction $h_{\Sigma_1}$ of $h$ to $\Sigma_1$ yields
a surjective mapping $h_{\Sigma_1}:\Sigma_1\to\Sigma_2$, $h$ is referred to as
a \emph{projection}.
We can conclude that the existence of a two-dimensional morphism seems to be a reasonable sufficient criterion for the situation that a two-dimensional word satisfies a two-dimensional pattern, but not a necessary one. \par
In fact, it turns out that characteristic factorisations provide the most promising approach to formalise how a two-dimensional word satisfies a two-dimensional pattern. Recall the example pattern
$\alpha = \begin{bsmallmatrix}
x_1 & x_1\\
x_2 & x_2
\end{bsmallmatrix}$
from above. Since $\alpha = (\begin{bsmallmatrix} x_1 \end{bsmallmatrix} \varobar \begin{bsmallmatrix} x_1 \end{bsmallmatrix}) \varominus (\begin{bsmallmatrix} x_2 \end{bsmallmatrix} \varobar \begin{bsmallmatrix} x_2 \end{bsmallmatrix})$, a characteristic factorisation of a two-dimensional word $U$ for $\alpha$ is a factorisation of the form $U = (V_1 \varobar V_1) \varominus (V_2 \varobar V_2)$. On the other hand, since $\alpha = (\begin{bsmallmatrix} x_1 \end{bsmallmatrix} \varominus \begin{bsmallmatrix} x_2 \end{bsmallmatrix}) \varobar (\begin{bsmallmatrix} x_1 \end{bsmallmatrix} \varominus \begin{bsmallmatrix} x_2 \end{bsmallmatrix})$, we could as well regard a factorisation $U = (V_1 \varominus V_2) \varobar (V_1 \varominus V_2)$ as characteristic for $\alpha$. For the sake of convenience, we say that the former factorisation is of \emph{column-row type} and the latter one is of \emph{row-column type}. Obviously, the two-dimensional word $W$ from above has a characteristic factorisation of column-row
type and a characteristic factorisation of row-column type (with respect to $\alpha$):
\begin{align*}
&W = (V_1 \varobar V_1) \varominus (V_2 \varobar V_2) = (\begin{bsmallmatrix} \ta&\ta&\tb \\ \ta&\ta&\tb \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \ta&\ta&\tb \\ \ta&\ta&\tb \end{bsmallmatrix}) \varominus (\begin{bsmallmatrix} \tc&\tc&\tc \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \tc&\tc&\tc \end{bsmallmatrix}) =
\begin{bsmallmatrix}
\ta& \ta& \tb& \ta& \ta& \tb\\
\ta& \ta& \tb& \ta& \ta& \tb\\
\tc& \tc& \tc&\tc&\tc&\tc
\end{bsmallmatrix}\,,\\
&W = (V_1 \varominus V_2) \varobar (V_1 \varominus V_2) = (\begin{bsmallmatrix} \ta&\ta&\tb \\ \ta&\ta&\tb \end{bsmallmatrix} \varominus \begin{bsmallmatrix} \tc&\tc&\tc \end{bsmallmatrix}) \varobar (\begin{bsmallmatrix} \ta&\ta&\tb \\ \ta&\ta&\tb \end{bsmallmatrix} \varominus \begin{bsmallmatrix} \tc&\tc&\tc \end{bsmallmatrix}) =
\begin{bsmallmatrix}
\ta& \ta& \tb& \ta& \ta& \tb\\
\ta& \ta& \tb& \ta& \ta& \tb\\
\tc& \tc& \tc&\tc&\tc&\tc
\end{bsmallmatrix}\,.
\end{align*}
As a matter of fact, for every two-dimensional word $U$ there exists a characteristic factorisation for $\alpha = \begin{bsmallmatrix}
x_1 & x_1\\
x_2 & x_2
\end{bsmallmatrix}$ of column-row type if and only if there exists a characteristic factorisation for $\alpha$ of row-column type. However, this is a particularity of $\alpha$ and, e.\,g., for
$\alpha' = \begin{bsmallmatrix}
x_1 & x_2 & x_3\\
x_2 & x_3 & x_1
\end{bsmallmatrix}$ and
$W' := \begin{bsmallmatrix}
\ta& \ta& \ta& \tb& \tc\\
\tb& \tc&\ta& \ta& \ta\\
\end{bsmallmatrix}$,
there exists a characteristic factorisation of column-row type $W' = (V_1 \varobar V_2 \varobar V_3) \varominus (V_2 \varobar V_3 \varobar V_1) = (\begin{bsmallmatrix} \ta& \ta& \ta \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \tb \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \tc \end{bsmallmatrix}) \varominus (\begin{bsmallmatrix} \tb \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \tc \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \ta& \ta& \ta \end{bsmallmatrix})$, but no characteristic factorisation of row-column type. Furthermore, the column-row factorisation of $W'$ is somewhat at odds with our intuitive understanding of what it means that a two-dimensional word satisfies a two-dimensional pattern. This is due to the fact that factorising $W'$ into $(\begin{bsmallmatrix} \ta& \ta& \ta \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \tb \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \tc \end{bsmallmatrix}) \varominus (\begin{bsmallmatrix} \tb \end{bsmallmatrix} \varobar \begin{bsmallmatrix}
\tc \end{bsmallmatrix} \varobar \begin{bsmallmatrix} \ta& \ta& \ta \end{bsmallmatrix})$ means that we associate the two-dimensional factors $\begin{bsmallmatrix} \ta& \ta& \ta \end{bsmallmatrix}$, $\begin{bsmallmatrix} \tb \end{bsmallmatrix}$ and $\begin{bsmallmatrix} \tc \end{bsmallmatrix}$ with the variables $x_1$, $x_2$ and $x_3$, respectively, but in the pattern $\alpha'$ the vertical neighbourship relation between the occurrence of $x_2$ in the first row and the occurrence of $x_3$ in the second row is not preserved in $W'$ with respect to the corresponding two-dimensional factors $\begin{bsmallmatrix} \tb \end{bsmallmatrix}$ and $\begin{bsmallmatrix} \tc \end{bsmallmatrix}$. More precisely, while a column-row factorisation preserves the horizontal neighbourship relation of the variables, it may violate their vertical neighbourship relation, where for row-column factorisations it is the other way around. Consequently, if we want both the vertical as well as the horizontal neighbourship relation to be
preserved, we should require that the two-dimensional word $U$ can be disassembled into two-dimensional factors that induce both a column-row as well as a row-column factorisation. More precisely, we say that $U$ satisfies
$\alpha' = \begin{bsmallmatrix}
x_1 & x_2 & x_3\\
x_2 & x_3 & x_1
\end{bsmallmatrix}$
if and only if there exist two-dimensional words $V_1, V_2$ and $V_3$, such that $U = (V_1 \varobar V_2 \varobar V_3) \varominus (V_2 \varobar V_3 \varobar V_1) = (V_1 \varominus V_2) \varobar (V_2 \varominus V_3) \varobar (V_3 \varominus V_1)$, which we call a \emph{proper} characteristic factorisation of $U$. \par
We are now ready to formalise the ideas developed so far and we can finally give a sound definition of two-dimensional pattern languages. Although we consider the class of two-dimensional pattern languages that results from the proper characteristic factorisations as the natural two-dimensional counterpart of the class of one-dimensional pattern languages, we shall also define the other classes of two-dimensional pattern languages which were sketched above. \par
For the definition of two-dimensional patterns, we use the same set of variables $X$ that has already been used in the definition of one-dimensional pattern languages. An \emph{array pattern} is a non-empty two-dimensional word over $X$ and a \emph{terminal array} is a non-empty two-dimensional word over a \emph{terminal} alphabet $\Sigma$. If it is clear from the context that we are concerned with array patterns and terminal arrays, then we simply say pattern and array, respectively.
Any mapping $h : X \rightarrow \Sigma^{++}$ is called a \emph{substitution}. For any substitution $h$, by $h_{\varobar, \varominus}$, we denote the mapping $X^{++} \rightarrow \Sigma^{++}$ defined in the following way. For any $\alpha := [y_{i, j}]_{m, n} \in X^{++}$, we define
\begin{align*}
h_{\varobar, \varominus}(\alpha) :=\:&(h(y_{1,1}) \varobar h(y_{1,2}) \varobar \ldots \varobar h(y_{1,n})) \varominus\\
&(h(y_{2,1}) \varobar h(y_{2,2}) \varobar \ldots \varobar h(y_{2,n})) \varominus \ldots \varominus\\
&(h(y_{m,1}) \varobar h(y_{m,2}) \varobar \ldots \varobar h(y_{m,n}))\,.
\end{align*}
Similarly, $h_{\varominus, \varobar} : X^{++} \rightarrow \Sigma^{++}$ is defined by
\begin{align*}
h_{\varominus, \varobar}(\alpha) :=\:&(h(y_{1,1}) \varominus h(y_{2,1}) \varominus \ldots \varominus h(y_{m,1})) \varobar\\
&(h(y_{1,2}) \varominus h(y_{2,2}) \varominus \ldots \varominus h(y_{m,2})) \varobar \ldots \varobar\\
&(h(y_{1,n}) \varominus h(y_{2,n}) \varominus \ldots \varominus h(y_{m,n}))\,.
\end{align*}
Intuitively speaking, both mappings $h_{\varobar, \varominus}$ and $h_{\varobar, \varominus}$, when applied to an array pattern $\alpha$, first substitute every variable occurrence of $\alpha$ by a terminal array according to the substitution $h$ and then these $m \times n$ individual terminal arrays are assembled to one terminal array by either first column-concatenating all the $n$ terminal arrays in every individual row and then row-concatenating the resulting $m$ terminal arrays, or by first row-concatenating all the $m$ terminal arrays in every individual column and then column-concatenating the resulting $n$ terminal arrays.\par
Let $\alpha \in X^{++}$, $W \in \Sigma^{++}$ and let $h : X \rightarrow \Sigma^{++}$. The array $W$ is a (1) \emph{column-row image} of $\alpha$ (\emph{with respect to $h$}), (2) a \emph{row-column image} of $\alpha$ (\emph{with respect to $h$}) or (3) a \emph{proper image} of $\alpha$ (\emph{with respect to $h$}) if and only if (1) $h_{\varobar, \varominus}(\alpha) = W$, (2) $h_{\varominus, \varobar}(\alpha) = W$ or (3) $h_{\varobar, \varominus}(\alpha) = h_{\varominus, \varobar}(\alpha) = W$, respectively. The mapping $h$ is called a \emph{column-row substitution for $\alpha$ and $W$}, a \emph{row-column substitution for $\alpha$ and $W$} or a \emph{proper substitution for $\alpha$ and $W$}, respectively. We say that $W$ is a column-row, a row-column or a proper image of $\alpha$ if there exists a column-row, a row-column or a proper substitution, respectively, for $\alpha$ and $W$. \par
A nice and intuitive way to interpret the different kinds of images of array
patterns is to imagine a grid to be placed over the terminal array.
The vertical lines of the grid represent a column concatenation and the
horizontal lines of the grid represent a row concatenation of the corresponding factorisation.
This means that every rectangular area of the grid corresponds to an occurrence
of a variable $x$ in the array pattern or, more precisely,
to the array $h(x)$ substituted for $x$. The fact that an array satisfies a pattern
is then represented by the situation that each two rectangular areas of the
grid that correspond to occurrences of the same variable must have identical
content. In Figure~\ref{fig-illustration}, an example for each a morphic image,
a proper image, a column-row image and a row-column image of a $5 \times 4$
pattern is represented in this illustrative way.
\begin{figure}[tb]
\begin{center}
\begin{tikzpicture}[scale=.4]
\coordinate (coord0) at (0,0);
\coordinate (xshift) at (8.5,0);
\coordinate (width) at (7.5,0);
\coordinate (height) at (0,12.5);
\draw ($(0,0)$) -> ($(0,0) + (width)$);
\draw ($(0,0) + (width)$) -> ($(0,0) + (width) - (height)$);
\draw ($(0,0) + (width) - (height)$) -> ($(0,0) - (height)$);
\draw ($(0,0) - (height)$) -> ($(0,0)$);
\draw ($(0,0) + (1.875,0)$) -> ($(0,0) + (1.875,0) - (height)$);
\draw ($(0,0) + (3.75,0)$) -> ($(0,0) + (3.75,0) - (height)$);
\draw ($(0,0) + (5.625,0)$) -> ($(0,0) + (5.625,0) - (height)$);
\draw ($(0,0) - (0, 2.5)$) -> ($(0,0) - (0, 2.5) + (width)$);
\draw ($(0,0) - (0, 5)$) -> ($(0,0) - (0, 5) + (width)$);
\draw ($(0,0) - (0, 7.5)$) -> ($(0,0) - (0, 7.5) + (width)$);
\draw ($(0,0) - (0, 10)$) -> ($(0,0) - (0, 10) + (width)$);
\draw ($(0,0) + (xshift)$) -> ($(0,0) + (width) + (xshift)$);
\draw ($(0,0) + (width) + (xshift)$) -> ($(0,0) + (width) - (height) + (xshift)$);
\draw ($(0,0) + (width) - (height) + (xshift)$) -> ($(0,0) - (height) + (xshift)$);
\draw ($(0,0) - (height) + (xshift)$) -> ($(0,0) + (xshift)$);
\draw ($(0,0) + (2,0) + (xshift)$) -> ($(0,0) + (2,0) - (height) + (xshift)$);
\draw ($(0,0) + (3,0) + (xshift)$) -> ($(0,0) + (3,0) - (height) + (xshift)$);
\draw ($(0,0) + (4.5,0) + (xshift)$) -> ($(0,0) + (4.5,0) - (height) + (xshift)$);
\draw ($(0,0) - (0, 2)+ (xshift)$) -> ($(0,0) - (0, 2) + (width)+ (xshift)$);
\draw ($(0,0) - (0, 3.5)+ (xshift)$) -> ($(0,0) - (0, 3.5) + (width)+ (xshift)$);
\draw ($(0,0) - (0, 8)+ (xshift)$) -> ($(0,0) - (0, 8) + (width)+ (xshift)$);
\draw ($(0,0) - (0, 11)+ (xshift)$) -> ($(0,0) - (0, 11) + (width)+ (xshift)$);
\draw ($(0,0) + 2*(xshift)$) -> ($(0,0) + (width) + 2*(xshift)$);
\draw ($(0,0) + (width) + 2*(xshift)$) -> ($(0,0) + (width) - (height) + 2*(xshift)$);
\draw ($(0,0) + (width) - (height) + 2*(xshift)$) -> ($(0,0) - (height) + 2*(xshift)$);
\draw ($(0,0) - (height) + 2*(xshift)$) -> ($(0,0) + 2*(xshift)$);
\draw ($(0,0) + (2.5,0) + 2*(xshift)$) -> ($(0,0) + (2.5,0) - (0, 3) + 2*(xshift)$);
\draw ($(0,0) + (5,0) + 2*(xshift)$) -> ($(0,0) + (5,0) - (0, 3) + 2*(xshift)$);
\draw ($(0,0) + (5.5,0) + 2*(xshift)$) -> ($(0,0) + (5.5,0) - (0, 3) + 2*(xshift)$);
\draw ($(0,0) + (1,0) - (0, 3) + 2*(xshift)$) -> ($(0,0) + (1,0) - (0, 5) + 2*(xshift)$);
\draw ($(0,0) + (3.5,0) - (0, 3) + 2*(xshift)$) -> ($(0,0) + (3.5,0) - (0, 5) + 2*(xshift)$);
\draw ($(0,0) + (6,0) - (0, 3) + 2*(xshift)$) -> ($(0,0) + (6,0) - (0, 5) + 2*(xshift)$);
\draw ($(0,0) + (2,0) - (0, 5) + 2*(xshift)$) -> ($(0,0) + (2,0) - (0, 6.5) + 2*(xshift)$);
\draw ($(0,0) + (3.5,0) - (0, 5) + 2*(xshift)$) -> ($(0,0) + (3.5,0) - (0, 6.5) + 2*(xshift)$);
\draw ($(0,0) + (5,0) - (0, 5) + 2*(xshift)$) -> ($(0,0) + (5,0) - (0, 6.5) + 2*(xshift)$);
\draw ($(0,0) + (2.5,0) - (0, 6.5) + 2*(xshift)$) -> ($(0,0) + (2.5,0) - (0, 9.5) + 2*(xshift)$);
\draw ($(0,0) + (5,0) - (0, 6.5) + 2*(xshift)$) -> ($(0,0) + (5,0) - (0, 9.5) + 2*(xshift)$);
\draw ($(0,0) + (6.5,0) - (0, 6.5) + 2*(xshift)$) -> ($(0,0) + (6.5,0) - (0, 9.5) + 2*(xshift)$);
\draw ($(0,0) + (1,0) - (0, 9.5) + 2*(xshift)$) -> ($(0,0) + (1,0) - (height) + 2*(xshift)$);
\draw ($(0,0) + (4,0) - (0, 9.5) + 2*(xshift)$) -> ($(0,0) + (4,0) - (height) + 2*(xshift)$);
\draw ($(0,0) + (6,0) - (0, 9.5) + 2*(xshift)$) -> ($(0,0) + (6,0) - (height) + 2*(xshift)$);
\draw ($(0,0) - (0, 3)+ 2*(xshift)$) -> ($(0,0) - (0, 3) + (width)+ 2*(xshift)$);
\draw ($(0,0) - (0, 5)+ 2*(xshift)$) -> ($(0,0) - (0, 5) + (width)+ 2*(xshift)$);
\draw ($(0,0) - (0, 6.5)+ 2*(xshift)$) -> ($(0,0) - (0, 6.5) + (width)+ 2*(xshift)$);
\draw ($(0,0) - (0, 9.5)+ 2*(xshift)$) -> ($(0,0) - (0, 9.5) + (width)+ 2*(xshift)$);
\draw ($(0,0) + 3*(xshift)$) -> ($(0,0) + (width) + 3*(xshift)$);
\draw ($(0,0) + (width) + 3*(xshift)$) -> ($(0,0) + (width) - (height) + 3*(xshift)$);
\draw ($(0,0) + (width) - (height) + 3*(xshift)$) -> ($(0,0) - (height) + 3*(xshift)$);
\draw ($(0,0) - (height) + 3*(xshift)$) -> ($(0,0) + 3*(xshift)$);
\draw ($(0,0) + (2,0) + 3*(xshift)$) -> ($(0,0) + (2,0) - (height) + 3*(xshift)$);
\draw ($(0,0) + (3,0) + 3*(xshift)$) -> ($(0,0) + (3,0) - (height) + 3*(xshift)$);
\draw ($(0,0) + (6,0) + 3*(xshift)$) -> ($(0,0) + (6,0) - (height) + 3*(xshift)$);
\draw ($(0,0) - (0, 2)+ 3*(xshift)$) -> ($(0,0) - (0, 2) + (2,0) + 3*(xshift)$);
\draw ($(0,0) - (0, 1.5) + (2,0) + 3*(xshift)$) -> ($(0,0) - (0, 1.5) + (3,0) + 3*(xshift)$);
\draw ($(0,0) - (0, 2.5) + (3,0) + 3*(xshift)$) -> ($(0,0) - (0, 2.5) + (6,0) + 3*(xshift)$);
\draw ($(0,0) - (0, 4) + (6,0) + 3*(xshift)$) -> ($(0,0) - (0, 4) + (7.5,0) + 3*(xshift)$);
\draw ($(0,0) - (0, 4.5)+ 3*(xshift)$) -> ($(0,0) - (0, 4.5) + (2,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 6)+ (2,0) + 3*(xshift)$) -> ($(0,0) - (0, 6) + (3,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 7)+ (3,0) + 3*(xshift)$) -> ($(0,0) - (0, 7) + (6,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 5)+ (6,0) + 3*(xshift)$) -> ($(0,0) - (0, 5) + (7.5,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 8)+ 3*(xshift)$) -> ($(0,0) - (0, 8) + (2,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 8) + (2,0) + 3*(xshift)$) -> ($(0,0) - (0, 8) + (3,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 9) + (3,0) + 3*(xshift)$) -> ($(0,0) - (0, 9) + (6,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 8.5) + (6,0) + 3*(xshift)$) -> ($(0,0) - (0, 8.5) + (7.5,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 12)+ 3*(xshift)$) -> ($(0,0) - (0, 12) + (2,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 11)+ (2,0) + 3*(xshift)$) -> ($(0,0) - (0, 11) + (3,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 11.5)+ (3,0) + 3*(xshift)$) -> ($(0,0) - (0, 11.5) + (6,0)+ 3*(xshift)$);
\draw ($(0,0) - (0, 10)+ (6,0) + 3*(xshift)$) -> ($(0,0) - (0, 10) + (7.5,0)+ 3*(xshift)$);
\node[anchor=center] at ($(0, 0) + (3.75, 0) - (height) - (0,0.8)$) {morphic image};
\node[anchor=center] at ($(0, 0) + (3.75, 0) - (height) - (0,0.8) + (xshift)$) {proper image};
\node[anchor=center] at ($(0, 0) + (3.75, 0) - (height) - (0,0.8) + 2*(xshift)$) {column-row image};
\node[anchor=center] at ($(0, 0) + (3.75, 0) - (height) - (0,0.8) + 3*(xshift)$) {row-column image};
\end{tikzpicture}
\end{center}
\caption{Illustrating possible image partitions.}
\label{fig-illustration}
\end{figure}
Alternatively, we can interpret the property that a terminal array $W$ is a certain type of image of an array pattern as a tiling of $W$. More precisely, $W$ satisfies a given array pattern $\alpha$ with $n$ different variables if and only if $n$ tiles can be allocated to the $n$ variables of $\alpha$ such that combining the tiles as indicated by the structure of $\alpha$ yields $W$. The grids depicted in Figure~\ref{fig-illustration} then illustrate the structure of such a tiling.
The definitions of the corresponding classes of pattern languages are now straightforward:
\begin{definition}
Let $\alpha \in X^{++}$ be an array pattern. We define the following variants of two-dimensional pattern languages:
\begin{itemize}
\item $L_{\Sigma, h}(\alpha) := \{W \in \Sigma^{**} \mid W \text{ is a morphic image of } \alpha\}$,
\item $L_{\Sigma, p}(\alpha) := \{W \in \Sigma^{**} \mid W \text{ is a proper image of } \alpha\}$,
\item $L_{\Sigma, r}(\alpha) := \{W \in \Sigma^{**} \mid W \text{ is a column-row image of } \alpha\}$,
\item $L_{\Sigma, c}(\alpha) := \{W \in \Sigma^{**} \mid W \text{ is a row-column image of } \alpha\}$,
\item $L_{\Sigma, rc}(\alpha) := L_{\Sigma, r}(\alpha) \cup L_{\Sigma, c}(\alpha)$.
\end{itemize}
For a pattern $\alpha$, we also denote the above languages by \emph{$Z$ pattern language of $\alpha$}, where $Z \in \{h, p, r, c, rc\}$.
For every $x \in \{r, c, rc, p, h\}$, we define $\mathcal{L}_{\Sigma, x} := \{L_{\Sigma, x}(\alpha) \mid \alpha \in X^{++}\}$ and $\mathcal{L}_{x} := \{\mathcal{L}_{\Sigma, x} \mid \Sigma \text{ is some alphabet}\}$.
\end{definition}
Since, for a fixed array pattern $\alpha$, every morphic image is a proper image and every proper image is a row-column image as well as a column-row image, the following subset relations between the different types of pattern languages hold (in the following diagram, an arrow denotes a subset relation):\par
\begin{tikzpicture}
[every path/.style={>=latex},every node/.style={text centered}]
\node (0) at (0,0) {$\:$};
\node (1) at (2,0) {$L_{\Sigma, h}(\alpha)$};
\node (2) at (5,0) {$L_{\Sigma, p}(\alpha)$};
\node (3a) at (8,-0.3) {$L_{\Sigma, c}(\alpha)$};
\node (3b) at (8,0.3) {$L_{\Sigma, r}(\alpha)$};
\node (4) at (11,0) {$L_{\Sigma, rc}(\alpha)$};
\draw[->] (1) -- (2);
\draw[->] (2) -- (3a);
\draw[->] (2) -- (3b);
\draw[->] (3a) -- (4);
\draw[->] (3b) -- (4);
\end{tikzpicture}
\begin{remark}
As indicated in the introductory part of this section, we consider the class of $p$ pattern languages as the most natural class of two-dimensional pattern languages. Another observation that supports this claim is that the $p$ pattern languages are compatible, in a certain sense, to the one-dimensional pattern languages. More precisely, for a \emph{one-dimensional} (i.\,e., $1 \times n$) array pattern $\alpha$ the set $L_{\Sigma, p}(\alpha) \cap \{W \in \Sigma^{++} \mid |W|_r = 1\}$ coincides with the one-dimensional pattern language of $\alpha$. This does not hold for the $h$ pattern languages (since in the one-dimensional case the words variables are mapped to can differ in length), but holds for the $r$, $c$ and $rc$ pattern languages. However, as pointed out above, the $r$, $c$ and $rc$ pattern language of a given pattern $\alpha$ may contain arrays that, from an intuitive point of view, do not satisfy $\alpha$.
\end{remark}
\section{General Observations}
In this section, we state some general lemmas about two-dimensional morphisms and array pattern languages, which shall be important for proving the further results presented in this paper. First, we refine Proposition~\ref{morphismProp}, by giving a convenient characterisation for the morphism property for mappings on arrays. To this end, we define a substitution $h$ to be \emph{$(m, n)$-uniform} if, for every $x \in X$, $|h(x)|_{r} = m$ and $|h(x)|_{c} = n$ and a substitution is \emph{uniform} if it is $(m, n)$-uniform, for some $m, n \in \mathbb{N}$.
\begin{lemma}\label{lem-hom-charac}
A mapping $h:\Sigma^{**} \to \Gamma^{**}$ is a two-dimensional morphism if and only if $h = g_{\varominus, \varobar} = g_{\varobar, \varominus}$, where $g : \Sigma \to \Gamma^{**}$ is a uniform substitution.
\end{lemma}
\begin{proof}
We first observe that if $g$ is uniform, then $g_{\varominus, \varobar} = g_{\varobar, \varominus}$ obviously holds (so it is sufficient to prove the statement of the lemma only for one of these two mappings). Furthermore, if $g$ is uniform, then, for every $U, V \in \Sigma^{**}$, $g_{\varobar, \varominus}(U \varominus V) = g_{\varobar, \varominus}(U) \varominus g_{\varobar, \varominus}(V)$ and $g_{\varobar, \varominus}(U \varobar V) = g_{\varobar, \varominus}(U) \varobar g_{\varobar, \varominus}(V)$, which proves the \emph{if} direction. In order to prove the \emph{only if} direction, we assume that $h$ is a two-dimensional morphism and we define a substitution $g$ by $g(b) := h(b)$, $b \in \Sigma$. Furthermore, let $\widehat{g} \in \{g_{\varominus, \varobar}, g_{\varobar, \varominus}\}$. We now show that $h$ equals $\widehat{g}$ by induction. By definition, for every $b \in \Sigma$, $h(b) = \widehat{g}(b)$. Now let $U, V \in \Sigma^{**}$ with $h(U) = g'(U)$ and $h(V) = g'(V)$. Then $h(U \varominus V) = h(U) \varominus h(V) = g'(U) \varominus g'(V) = g'(U \varominus V)$ and, analogously, $h(U \varobar V) = h(U) \varobar h(V) = g'(U) \varobar g'(V) = g'(U \varobar V)$. By induction, it follows that $h = \widehat{g}$. Consequently, we can conclude that $g$ is uniform.
\end{proof}
The next lemma states that the composition of two two-dimensional morphisms is again a two-dimensional morphism.
\begin{lemma}\label{lem-hom-closure}
Let $h_1:\Gamma^{**}\to \Sigma^{**}$ and $h_2:\Sigma^{**}\to\Delta^{**}$ be two-dimensional morphisms.
Then, the composition $h_{1,2}:=h_1\circ h_2:\Gamma^{**}\to\Delta^{**}$ is also a two-dimensional morphism.
\end{lemma}
\begin{proof}
We first observe the following. If $g$ and $f$ are some uniform substitutions, then $h := g \circ f$ is a uniform substitution as well. Furthermore, it can be easily verified that $g_{\varominus, \varobar} \circ f_{\varominus, \varobar} = h_{\varominus, \varobar}$. With Lemma~\ref{lem-hom-charac}, this directly implies the statement of the lemma.
\end{proof}
It is intuitively clear that the structure of a pattern fully determines the corresponding pattern language and the actual names of the variables are irrelevant, e.\,g., the patterns $\begin{bsmallmatrix} x_1 & x_2 & x_1 \\ x_2 & x_3 & x_3 \end{bsmallmatrix}$ and $\begin{bsmallmatrix} x_7 & x_3 & x_7 \\ x_3 & x_5 & x_5 \end{bsmallmatrix}$ should be considered identical. In the following we formalise this intuition. Two array patterns $\alpha := [y_{i, j}]_{m,n}$ and $\beta := [z_{i, j}]_{m',n'}$ are \emph{equivalent up to a renaming}, denoted by $\alpha \sim \beta$, if and only if $m = m'$, $n = n'$ and, for every $i, j, i', j'$, $1 \leq i, i' \leq m$, $1 \leq j, j' \leq n$, $y_{i, j} = y_{i', j'}$ if and only if $z_{i, j} = z_{i', j'}$.
\begin{lemma}\label{equalityLemma}
Let $z, z' \in \{h, p, r, c, rc\}$, let $\Sigma$ be an alphabet with $|\Sigma| \geq 2$ and let $\alpha, \beta \in X^{++}$. If $L_{\Sigma, z}(\alpha) = L_{\Sigma, z'}(\beta)$, then $\alpha \sim \beta$.
\end{lemma}
\begin{proof}
We assume that $L_{\Sigma, z}(\alpha) = L_{\Sigma, z'}(\beta)$ and note that this implies that $|\alpha|_c = |\beta|_c = m$ and $|\alpha|_r = |\beta|_r = n$. This is due to the fact that if $|\alpha|_c < |\beta|_c$ or $|\alpha|_r < |\beta|_r$, then the array obtained from $\alpha$ by replacing every variable by a single symbol $\ta \in \Sigma$ is in $L_{\Sigma, z}(\alpha)$, but not in $L_{\Sigma, z'}(\beta)$. We further assume that $\alpha \nsim \beta$, which implies that there are $i, j, i', j'$ with $1 \leq i, i' \leq m$ and $1 \leq j, j' \leq n$, such that $\alpha[i, j] = \alpha[i', j']$ and $\beta[i, j] \neq \beta[i', j']$ (or $\alpha[i, j] \neq \alpha[i', j']$ and $\beta[i, j] = \beta[i', j']$, for which an analogous argument applies). We now define a substitution $h : X \rightarrow \Sigma$ in the following way. For every $x \in X$, if $\beta[i, j] = x$, then $h(x) := \ta$ and if $\beta[i, j] \neq x$, then $h(x) := \tb$. We observe that, since $h_{\varominus, \varobar}(\beta)$ is a morphic image, a proper image, a row-column image and a column-row image of $\beta$, $h_{\varominus, \varobar}(\beta) \in L_{\Sigma, z'}(\beta)$. Furthermore, $h_{\varominus, \varobar}(\beta)[i, j] \neq h_{\varominus, \varobar}(\beta)[i', j']$. On the other hand, for every $W \in L_{\Sigma, z}(\alpha)$, with $|W|_c = |h_{\varominus, \varobar}(\beta)|_c$ and $|W|_r = |h_{\varominus, \varobar}(\beta)|_r$, $W[i, j] = W[i', j']$ is
satisfied. Thus, $h_{\varominus, \varobar}(\beta) \notin L_{\Sigma, z}(\alpha)$, which implies that $L_{\Sigma, z}(\alpha) \neq L_{\Sigma, z'}(\beta)$.
\end{proof}
For every $z, z' \in \{h, p, r, c, rc\}$, $z \neq z'$, $\alpha \sim \beta$ does not necessarily imply $L_{\Sigma, z}(\alpha) = L_{\Sigma, z'}(\beta)$, as pointed out by, e.\,g., $L_{\Sigma, h}(\begin{bsmallmatrix} x_1 & x_2 \end{bsmallmatrix}) \neq L_{\Sigma, p}(\begin{bsmallmatrix} x_1 & x_2 \end{bsmallmatrix})$ or $L_{\Sigma, p}(\begin{bsmallmatrix} x_1 & x_2 \\ x_3 & x_1 \end{bsmallmatrix}) \neq L_{\Sigma, c}(\begin{bsmallmatrix} x_1 & x_2 \\ x_3 & x_1 \end{bsmallmatrix})$. On the other hand, since, for every $z, \in \{h, p, r, c, rc\}$, $\alpha \sim \beta$ obviously implies $L_{\Sigma, z}(\alpha) = L_{\Sigma, z}(\beta)$, two $z$ pattern languages are equivalent if and only if they are described by two patterns that are equivalent up to a renaming.\par
In the remainder of this work, we do not distinguish anymore between patterns that are equivalent up to a renaming, i.\,e., from now on we say that $\alpha$ and $\beta$ are equivalent, denoted by $\alpha = \beta$ for simplicity, if they are actually the same arrays or if they are equivalent up to a renaming.
\section{Comparison of Array Pattern Language Classes}
In this section, we provide a pairwise comparison of our different classes of array pattern languages and, furthermore, we compare them with the class of recognisable array languages, denoted by REC, which is one of the most prominent classes of array languages. For a detailed description of REC, the reader is referred to the survey \cite{gia:two} by Giammarresi~and~Restivo. Next, we show that, for every alphabet $\Sigma$ with $|\Sigma| \geq 2$, the language classes REC, $\mathcal{L}_{\Sigma, h}$, $\mathcal{L}_{\Sigma, p}$, $\mathcal{L}_{\Sigma, r}$, $\mathcal{L}_{\Sigma, c}$ and $\mathcal{L}_{\Sigma, rc}$ are pairwise incomparable. More precisely, for every $\mathcal{L}_1, \mathcal{L}_2 \in \{\text{REC}, \mathcal{L}_{\Sigma, h}, \mathcal{L}_{\Sigma, p}, \mathcal{L}_{\Sigma, r}, \mathcal{L}_{\Sigma, c}, \mathcal{L}_{\Sigma, rc}\}$ with $\mathcal{L}_1 \neq \mathcal{L}_2$, we show that $\mathcal{L}_1 \setminus \mathcal{L}_2 \neq \emptyset$, $\mathcal{L}_2 \setminus \mathcal{L}_1 \neq \emptyset$ and $\mathcal{L}_1 \cap \mathcal{L}_2 \neq \emptyset$. The non-emptiness of the pairwise intersections of these language classes can be easily seen:
\begin{proposition}
For every $z \in \{h,p,r,c,rc\}$, $L_{z, \Sigma}(\begin{bsmallmatrix} x_1 \end{bsmallmatrix}) = \Sigma^{++}$ and $\Sigma^{++} \in \text{REC}$.
\end{proposition}
It remains to find, for every $\mathcal{L}_1, \mathcal{L}_2 \in \{\text{REC}, \mathcal{L}_{\Sigma, h}, \mathcal{L}_{\Sigma, p}, \mathcal{L}_{\Sigma, r}, \mathcal{L}_{\Sigma, c}, \mathcal{L}_{\Sigma, rc}\}$, a separating language $L_1 \in \mathcal{L}_1 \setminus \mathcal{L}_2$ and a separating language $L_2 \in \mathcal{L}_2 \setminus \mathcal{L}_1$. We first present all these separating languages in a table and then we formally prove their separating property. In rows $2$ to $6$ of the following table, if a pattern $\alpha$ is the entry that corresponds to the row labeled by class $\mathcal{L}_{\Sigma, z}$ and the column labeled by class $\mathcal{L}_{\Sigma, z'}$, where $z, z' \in \{h, p, r, c, rc\}$, $z \neq z'$, then this means that $L \in \mathcal{L}_{\Sigma, z} \setminus \mathcal{L}_{\Sigma, z'}$. Row $1$, on the other hand, contains recognisable array languages that are not array pattern languages.
\begin{center}
\begin{tabular}{| c !{\vrule width 1pt} c | c | c | c | c | c |}
\hline
& REC & \textbf{$\mathcal{L}_{\Sigma, h}$} & $\mathcal{L}_{\Sigma, p}$ & $\mathcal{L}_{\Sigma, r}$ & $\mathcal{L}_{\Sigma, c}$ & $\mathcal{L}_{\Sigma, rc}$ \\ \noalign{\hrule height 1pt}
REC & -- & $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\}$ & $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\}$ & $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\}$ & $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\}$ & $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\}$ \\ \hline
$\mathcal{L}_{\Sigma, h}$ & $\begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$ & -- & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ \\ \hline
$\mathcal{L}_{\Sigma, p}$ & $\begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & -- & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ \\ \hline
$\mathcal{L}_{\Sigma, r}$ & $\begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & -- & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ \\ \hline
$\mathcal{L}_{\Sigma, c}$ & $\begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & -- & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ \\ \hline
$\mathcal{L}_{\Sigma, rc}$ & $\begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & $\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}$ & -- \\ \hline
\end{tabular}
\end{center}
\begin{lemma}\label{nonRECLang}
$L_{\Sigma, h}(\begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}) \notin \text{REC}$.
\end{lemma}
\begin{proof}
In this proof, we use the characterisation of REC by local array languages and projections (see Giammarresi~and~Restivo \cite{gia:two} and also the next section). Let $\alpha := \begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$ and $L := L_{\Sigma, h}(\alpha) = \{W \varominus W \mid W \in \Sigma^{++}\}$. Suppose $L \in REC$. Then there is a local array language $L'$ over an alphabet $\Gamma$ so that $L$ is a projection of $L'$. For the sake of convenience, we define $r := |\Gamma|$ and $s := |\Sigma|$. For every $m, n \in \mathbb{N}$, let
\begin{equation*}
L_{m,n} := \{W \varominus W \mid W \in \Sigma^{++} \text{ and } |W|_r = m, |W|_c = n\} \subseteq L\,.
\end{equation*}
Obviously, $|L_{m,n}| = s^{mn}$. Now let $L'_{m,n}$ be the set of pictures in $L^\prime,$ whose projections are in $L_{m,n}$. In the arrays of $L^\prime_{m,n},$ there are at most $r^{2n}$ possibilities of how the $m^{\text{th}}$ and $(m+1)^{\text{th}}$ row can look like. For sufficiently large $m,$ $s^{mn} > r^{2n}.$ Thus, there exist two arrays $W_1 \varominus W_1$ and $W_2 \varominus W_2,$ $W_1 \neq W_2$ in $L_{m,n}$ such that the corresponding arrays $W_1^\prime \varominus W_1^\prime$ and $W_2^\prime \varominus W_2^\prime$ in $L^\prime_{m,n}$ have the same $m^{\text{th}}$ row and the same $(m + 1)^{\text{th}}$ row. Hence, since $L'$ is a local array language, $W_1^\prime \varominus W_2^\prime, W_2^\prime \varominus W_1^\prime \in L^\prime_{m,n}$ and therefore $W_1 \varominus W_2, W_2 \varominus W_1 \in L_{m,n} \subseteq L$, which is a contradiction.
\end{proof}
It can be easily verified that, for every $z \in \{p,r,c,rc\}$, $L_{\Sigma, z}(\alpha) = L_{\Sigma, h}(\alpha)$, where $\alpha := \begin{bsmallmatrix} x_1 \\ x_1 \end{bsmallmatrix}$. Hence, for every $z \in \{h, p, r, c, rc\}$, $L_{\Sigma, z}(\alpha) \notin \text{REC}$, which implies the first column of the table. Furthermore, for every $z \in \{h,p,r,c,rc\}$, $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\} \notin L_{\Sigma, z}(\alpha)$, but $\{\begin{bsmallmatrix} \ta \end{bsmallmatrix}\} \in \text{REC}$, which implies the first row of the table. \par
We point out that, by Lemma~\ref{equalityLemma}, for every $z, z' \in \{h,p,r,c,rc\}$, $z \neq z'$, if there exists a pattern $\beta$ with $L_{\Sigma, z}(\beta) \neq L_{\Sigma, z'}(\beta)$, then $L_{\Sigma, z}(\beta) \in \mathcal{L}_{\Sigma, z} \setminus \mathcal{L}_{\Sigma, z'}$ and $L_{\Sigma, z'}(\beta) \in \mathcal{L}_{\Sigma, z'} \setminus \mathcal{L}_{\Sigma, z}$. Consequently, in order to prove the remaining entries of the table, it is sufficient to identify, for every $z, z' \in \{h,p,r,c,rc\}$, $z \neq z'$, a pattern $\beta$ with $L_{\Sigma, z}(\beta) \neq L_{\Sigma, z'}(\beta)$, which is done by the following two lemmas.
\begin{lemma}\label{hIncompLemma}
For every $z \in \{p,c,r,cr\}$, $L_{\Sigma, h}(\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix}) \neq L_{\Sigma, z}(\begin{bsmallmatrix} x_1&x_2 \\ x_3&x_4 \\ \end{bsmallmatrix})$.
\end{lemma}
\begin{proof}
Let $\beta :=
\begin{bsmallmatrix}
x_1&x_2 \\
x_3&x_4 \\
\end{bsmallmatrix}$ and let $W :=
\begin{bsmallmatrix}
\ta & \ta & \ta \\
\ta & \ta & \ta \\
\ta & \ta & \ta \\
\end{bsmallmatrix}
$. For every $z \in \{p,c,r,cr\}$, $W \in \mathcal{L}_{\Sigma, z}(\beta)$, since $g_{\varominus, \varobar}(\beta) = g_{\varobar, \varominus}(\beta) = W$, where $g$ is defined by
\begin{equation*}
g(x_1) :=
\begin{bsmallmatrix}
\ta & \ta \\
\ta & \ta \\
\end{bsmallmatrix},\:\:\\
g(x_2) :=
\begin{bsmallmatrix}
\ta \\
\ta \\
\end{bsmallmatrix},\:\:\\
g(x_3) := \begin{bsmallmatrix}
\ta & \ta \\
\end{bsmallmatrix},\:\:\\
g(x_4) := \begin{bsmallmatrix}
\ta \\
\end{bsmallmatrix}\,.
\end{equation*}
By Lemma~\ref{lem-hom-charac}, it is obvious that there does not exist any morphism $h$ with $h(\beta) = W$. Thus, $W \notin L_{\Sigma,h}(\beta)$ and, for every $z \in \{p, c, r, cr\}$, $L_{\Sigma,h}(\beta) \neq L_{\Sigma,z}(\beta)$.
\end{proof}
\begin{lemma}\label{prcIncompLemma}
For every $z, z' \in \{p,c,r,cr\}$, $z \neq z'$, $L_{\Sigma, z}(\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \end{bsmallmatrix}) \neq L_{\Sigma, z'}(\begin{bsmallmatrix} x_1&x_2 \\ x_2&x_1 \\ \end{bsmallmatrix})$.
\end{lemma}
\begin{proof}
Let $\gamma :=
\begin{bsmallmatrix}
x_1&x_2 \\
x_2&x_1
\end{bsmallmatrix}$ and let
$W_1 :=
\begin{bsmallmatrix}
\ta & \ta\\
\ta & \ta\\
\ta & \ta\\
\end{bsmallmatrix},
W_2 :=
\begin{bsmallmatrix}
\ta & \ta & \ta\\
\ta & \ta & \ta\\
\end{bsmallmatrix}$.
We observe that $g_{\varominus, \varobar}(\gamma) = W_1$, $g'_{\varobar, \varominus}(\gamma) = W_2,$ where $g, g'$ are defined by
\begin{equation*}
g(x_1) :=
\begin{bsmallmatrix}
\ta \\
\end{bsmallmatrix},\,
g(x_2) :=
\begin{bsmallmatrix}
\ta\\
\ta\\
\end{bsmallmatrix},\,
g'(x_1) := \begin{bsmallmatrix}
\ta\\
\end{bsmallmatrix},\,
g'(x_2) := \begin{bsmallmatrix}
\ta & \ta \\
\end{bsmallmatrix}\,.
\end{equation*}
Thus, $W_1 \in L_{\Sigma, c}(\gamma)$, $W_2 \in L_{\Sigma, r}(\gamma)$ and $W_1, W_2 \in L_{\Sigma, rc}(\gamma)$. On the other hand, $W_1, W_2 \notin L_{\Sigma, p}(\gamma)$, since every proper image of $\gamma$ must have an even number of columns and an even number of rows. Consequently, for every $z \in \{r,c, rc\}$, $L_{\Sigma, p}(\gamma) \neq L_{\Sigma, z}(\gamma)$. Similarly, $W_1 \notin L_{\Sigma, r}$ and $W_2 \notin L_{\Sigma, c}$, since every column-row image of $\gamma$ must have an even number of rows and every row-column image of $\gamma$ must have an even number of columns. This implies that, for every $z, z' \in \{c,r,cr\}$, $z \neq z'$, $L_{\Sigma, z}(\gamma) \neq L_{\Sigma, z'}(\gamma)$, which concludes the proof.
\end{proof}
\section{Closure Properties of Array Pattern Languages}
The research of closure properties of classes of formal languages is a classical topic in this area. However, the number of natural properties is richer in the case of arrays compared to the more conventional string case. Thus, in this section, we classify the operations that shall be investigated in this regard according to whether or not they correspond to string language operations. \par
First, in Section~\ref{sec:stringcaseops}, we investigate operations that correspond to string language operations. These are the Boolean operations of union, intersection and complementation, and also two special cases of (inverse) morphisms: letter-to-letter morphisms, and more special surjective letter-to-letter morphisms, known as
\emph{projections} in the terminology of array languages, and, more generally, the two-dimensional morphisms as defined in Section~\ref{sec:2DPatternLanguages}\par
Next, in Section~\ref{sec:similartostringcaseops}, we take a closer look at operations similar to string language operations. More precisely, we investigate closure under concatenation and concatenation closure (or Kleene star), which constitute classical operations for string languages, but, with respect to the array case, we encounter an important difference, namely, there are two different types of concatenations: row and column concatenation. In particular, the concatenation of two arrays could be undefined (just because the dimensions do not match), but the concatenation of the two corresponding languages need not be empty.\par
Finally, in Section~\ref{sec:arraycaseops}, we investigate \emph{operations special to arrays}, that are usually not considered or even defined for string languages. These are mainly geometric operations like quarter turn, half turn, reflection or transposition of an array.
\subsection{String Language Operations}\label{sec:stringcaseops}
We first point out that, due to Lemma~\ref{lem-transfer} below, whenever a non-closure result is known for terminal-free non-erasing string pattern languages, this would straightforwardly transfer to the array case. We will therefore focus on finding proofs for the string case for non-closure properties, and conversely, we will try to give proofs for the array case for closure properties.
Interestingly enough, (non-)closure properties have not been studied for the (classical) terminal-free non-erasing string pattern languages, all published
proofs that we are aware of for this topic use terminals or erasing. So, our study also contributes to the theory of string pattern languages.
Conversely, if we do not manage to find proofs or examples as required for the mentioned approach, this implicitly always raises an open classical string language question. For any mode $z \in \{h, p, r, c, rc\}$ and any pattern $\pi$, let $L_{\Sigma, z}^{\oneD}(\pi)$ denote those arrays from $L_{\Sigma, z}(\pi)$ that have just one row, i.\,e., $L_{\Sigma, z}^{\oneD}(\pi) := \{W \in L_{\Sigma, z}(\pi) \mid |W|_r = 1\}$. Clearly, such arrays can be interpreted as strings and vice versa. In this sense, $L_{\Sigma, z}^{\oneD}(\pi)$ and the string language $L^{\oneD}_{\Sigma}(\pi)$ generated by the pattern $\pi$ coincide, as long as $z\neq h$. For $z=h$, we encounter the special case that all inserted words have to be of the same length. Let us formulate this more formally:
\begin{lemma}\label{lem-transfer}
Let $\pi$ be an array pattern of height one.
Then, $\pi$ is, at the same time, a string pattern.
Moreover, for any $z \in \{p, r, c, rc\}$,
$L_{\Sigma, z}^{\oneD}(\pi)=L^{\oneD}_{\Sigma}(\pi)$, while
$L_{\Sigma, h}^{\oneD}(\pi)\subseteq L^{\oneD}_{\Sigma}(\pi)$.
\end{lemma}
We shall now prove non-closure properties for $L^{\oneD}_{\Sigma}(\pi)$, which directly carry over to the classes $L_{\Sigma, z}(\pi)$, $z \in \{h, p, r, c, rc\}$ (for some operations, however, the class $L_{\Sigma, h}(\pi)$ constitutes a special case, which is treated separately). To this end, we will mostly focus on two patterns: $\alpha=xyx$ and $\beta=xxy$. The next lemma states an immediate observation for these patterns.
\begin{lemma}
Over the terminal alphabet $\Sigma=\{\ta,\tb\}$,
let $L_s(\alpha)$ ($L_s(\beta)$) denote
the shortest words that can be described by $\alpha$ and $\beta$, respectively, disallowing erasing.
Then,
\begin{eqnarray*}
L_s(\alpha)&=&\{\ta \ta \ta, \ta \tb \ta, \tb \ta \tb, \tb \tb \tb\}\,,\\
L_s(\beta)&=&\{\ta \ta \ta, \ta \ta \tb, \tb \tb \ta, \tb \tb \tb\}\,.
\end{eqnarray*}
\end{lemma}
\begin{proposition}\label{prop-union}
For any non-unary alphabet $\Sigma$, $\mathcal{L}^{\oneD}_{\Sigma}$ is not closed under union.
\end{proposition}
\begin{proof}
Without loss of generality, $\{\ta, \tb\}\subseteq\Sigma$. In the following argument, we actually focus on
$\Sigma=\{\ta, \tb\}$, but this can be easily extended to the more general case.
Assume that there was a pattern $\gamma$ with $\mathcal{L}^{\oneD}_{\Sigma}(\gamma)=\mathcal{L}^{\oneD}_{\Sigma}(\alpha)\cup \mathcal{L}^{\oneD}_{\Sigma}(\beta)$.
Then, $$L_s(\gamma)=L_s(\alpha)\cup L_s(\beta)=\{\ta \ta \ta, \ta \ta \tb, \ta \tb \ta, \tb \ta \tb, \tb \tb \ta, \tb \tb \tb\}.$$
This means that $\gamma$ contains exactly three variable occurrences (with more, these words cannot be generated, with less, shorter
words could be generated), i.e., $\gamma=y_1\,y_2\,y_3$, $y_i \in X$, $1 \leq i \leq 3$. As $L_s(\gamma)\neq\{\ta, \tb\}^3$, some of these variables must coincide, which leads to a contradiction. More precisely, if $y_1 = y_2$ or $y_2 = y_3$, then $\ta \tb \ta$ cannot be generated and if $y_1 = y_3$, then $\ta \ta \tb$ cannot be generated.
Hence, $\gamma$ with $L_s(\gamma)=L_s(\alpha)\cup L_s(\beta)$ cannot exist.
\end{proof}
Now if there was a pattern $\gamma$ such that $L_{\Sigma, z}(\gamma)= L_{\Sigma, z}(\alpha)\cup L_{\Sigma, z}(\beta)$, $z \in \{p, r, c, rc\}$, then, by Lemma~\ref{lem-transfer}, this would imply $L_{\Sigma}^{\oneD}(\gamma)= L_{\Sigma}^{\oneD}(\alpha)\cup L_{\Sigma}^{\oneD}(\beta)$, contradicting Proposition~\ref{prop-union}. We point out that in the proof of Proposition~\ref{prop-union}, we do not use any replacements by words of different lengths to obtain our contradiction. Hence, this argument is also valid in the case when $z=h$.
\begin{corollary}\label{cor-clos-union}
None of the array pattern language classes under consideration (over some non-unary alphabet) is closed under union.
\end{corollary}
We proceed with the intersection operation.
\begin{proposition}\label{prop-intersection} For any non-unary alphabet $\Sigma$,
$\mathcal{L}^{\oneD}_{\Sigma}$ is not closed under intersection.
\end{proposition}
\begin{proof}
The argument resembles the previous proof.
Assume that $\gamma$ describes $L^{\oneD}_{\Sigma}(\alpha)\cap L^{\oneD}_{\Sigma}(\beta)$.
Notice that $L_s(\gamma)=\{\ta \ta \ta, \tb \tb \tb\}$, which clearly implies that $\gamma=xxx$.
However, $\ta \ta \tb \ta \ta\in (L^{\oneD}_{\Sigma}(\alpha)\cap L^{\oneD}_{\Sigma}(\beta))\setminus L^{\oneD}_{\Sigma}(\gamma)$.
\end{proof}
Notice that the replacement words we used for deriving a contradiction are of different lengths, meaning that $\ta \ta \tb \ta \ta \in L_{\Sigma}(\alpha)$ because of the replacement $x\mapsto \ta \ta$ and $y\mapsto \tb$, but
$\ta \ta \tb \ta \ta\in L_{\Sigma}(\beta)$ because of $x\mapsto \ta$ and $y\mapsto \tb \ta \ta$. Hence, we cannot conclude non-closure for the $h$-mode in the following corollary:
\begin{corollary}
None of the array pattern language classes under consideration (over some non-unary alphabet and apart from the $h$-case) is closed under intersection.
\end{corollary}
Indeed, the $h$-mode plays a special r\^ole, as can be seen by the following
result.
\begin{proposition}\label{prop-hom-intersection}Let $\Sigma$ be some alphabet.
Then, $\mathcal{L}_{\Sigma, h}$
is closed under intersection.
\end{proposition}
\begin{proof}
Assume that $a\in\Sigma$.
Let $\alpha,\beta$ be two array patterns. Let $m_\alpha$ be the height (number of rows) of $\alpha$ and $n_\alpha$ be the width (number of columns) of $\alpha$.
Likewise, $m_\beta$ and $n_\beta$ are understood.
Then, the width of the smallest arrays in $L_{\Sigma, h}(\alpha)\cap
L_{\Sigma, h}(\beta)$ equals $n=\operatorname{lcm}(n_\alpha,n_\beta)$,
and their height equals $m=\operatorname{lcm}(m_\alpha,m_\beta)$.
This can be easily seen by substituting each variable in $\alpha$
by the unique array of width $n'_\alpha=\frac{\operatorname{lcm}(n_\alpha,n_\beta)}{n_\alpha}$ and height $m'_\alpha=\frac{\operatorname{lcm}(m_\alpha,m_\beta)}{m_\alpha}$
over the alphabet $\{a\}$
into the pattern $\alpha$, as $n=n_\alpha \times n'_\alpha$,
$m=m_\alpha \times m'_\alpha$, and a similar substitution within $\beta$.
For each variable $x$ that occurs in $\alpha$, take $m'_\alpha \times n'_\alpha$ new variables
$x_{i,j}$ with $1\leq i\leq m'_\alpha$ and $1\leq j\leq n'_\alpha$.
Define a morphism $h_\alpha$ by replacing the variable $x$ by the following array of height $n'_\alpha$ and width $m'_\alpha$, consisting of the previously introduced variables:
$$\begin{smallmatrix}
x_{11} & x_{12} & \cdots & x_{1n'_\alpha}\\
x_{21} & x_{22} & \cdots & x_{2n'_\alpha}\\
\vdots & \vdots & \vdots & \vdots\\
x_{m'_\alpha 1} & x_{m'_\alpha 2} & \cdots & x_{m'_\alpha n'_\alpha}
\end{smallmatrix}
$$
Hence, $h_\alpha(\alpha)$ is some array of height $m$ and of width $n$.
Accordingly, one can define a morphism $h_\beta$ such that $h_\beta(\beta)$ is again some array of height $m$ and of width $n$.
Due to Lemma~\ref{lem-hom-closure}, $L_{\Sigma, h}(h_\alpha(\alpha))\subseteq L_{\Sigma, h}(\alpha)$ and $L_{\Sigma, h}(h_\beta(\beta))\subseteq L_{\Sigma, h}(\beta)$ (*).
Namely, if $U\in L_{\Sigma, h}(h_\alpha(\alpha))$, then there exists some two-dimensional morphism $g$ such that $U=g(h_\alpha(\alpha))$, i.e., there also
exists some two-dimensional morphism $f:=h_\alpha\circ g$ with $U=f(\alpha)$, so that $U\in L_{\Sigma, h}(\alpha)$.
Now, define an array pattern $\gamma$ of height $m$ and of width $n$, consisting exclusively of variable entries, as follows:
The variables occurring at positions $(i,j)$ and $(i',j')$ in $\gamma$ (where $1\leq i,i'\leq m$ and $1\leq j,j'\leq n$) are identical if and only if the corresponding variables in at least one of the arrays $h_\alpha(\alpha)$ or $h_\beta(\beta)$ are identical.
We claim that $L_{\Sigma, h}(\gamma)=L_{\Sigma, h}(\alpha)\cap
L_{\Sigma, h}(\beta)$.
As $\gamma$ is obtained from $h_\alpha(\alpha)$ by identifying certain variables, due to (*) we find that:
$$L_{\Sigma, h}(\gamma)\subseteq L_{\Sigma, h}(h_\alpha(\alpha))\subseteq L_{\Sigma, h}(\alpha),$$
and likewise for $\beta$, so that $L_{\Sigma, h}(\gamma)\subseteq L_{\Sigma, h}(\alpha)\cap
L_{\Sigma, h}(\beta)$.
Conversely, we have already argued that the smallest arrays in $L_{\Sigma, h}(\alpha)\cap
L_{\Sigma, h}(\beta)$ have height $m$ and width $n$.
More generally, any array $U\in L_{\Sigma, h}(\alpha)\cap
L_{\Sigma, h}(\beta)$ has height $k\cdot m$ and width $l\cdot n$.
As $U\in L_{\Sigma, h}(\alpha)$, there is some two-dimensional morphism $h_{U,\alpha}$ such that
$U=h_{U,\alpha}(\alpha)$. Moreover, for each variable $x$ in $\alpha$, $|h_{U,\alpha}(x)|_r=k_\alpha=\frac{k\cdot m}{m_\alpha}$
and $|h_{U,\alpha}(x)|_c=l_\alpha=\frac{l\cdot n}{n_\alpha}$. We can make an analogous reasoning with $\beta$, introducing
the constants $k_\beta=\frac{k\cdot m}{m_\beta}$ and $l_\beta=\frac{l\cdot n}{n_\beta}$ for the morphism $h_{U,\beta}$.
As $U=h_{U,\alpha}(\alpha)=h_{U,\beta}(\beta)$, entries in $U$ must coincide both according to $\alpha$ and to $\beta$.
This is exactly reflected in the construction of $\gamma$ provided above, so that $h_{U,\alpha}=h_{\alpha}\circ h'$
and $h_{U,\beta}=h_{\beta}\circ h'$ for some two-dimensional morphism $h'$ with $h'(\gamma)=U$.
Hence, $U\in L_{\Sigma, h}(\gamma)$.
\end{proof}
Arguments as in Propositions~\ref{prop-union} and~\ref{prop-intersection}
can be given for any non-trivial binary set operation, for instance, symmetric difference or set difference.
This also gives the according result for complementation, but there is also an easier argument in that case.
Notice that, as non-erasing pattern languages or array patterns cannot reasonably cope with the empty word or the empty array,
we disregard this in the complement operation.
\begin{proposition}\label{prop-complementation}
For any alphabet $\Sigma$, $\mathcal{L}^{\oneD}_{\Sigma}$ is not closed under complementation.
\end{proposition}
\begin{proof}
Let $\Sigma$ be some alphabet with $a\in\Sigma$.
Consider the pattern $\alpha=xy$.
The complement (disregarding the empty word) contains as shortest words a word of the form $a$.
This implies that a pattern $\beta$ describing the complement must have length one, i.e., $\beta=x$.
Hence, $L(\beta)=\Sigma^+$, which is not the complement of $L(\alpha)$.
\end{proof}
\begin{corollary}
None of the array pattern language classes under consideration (over any alphabet) is closed under complementation.
\end{corollary}
Notice that in the other cases (but complementation), we cannot cope with unary alphabets.
This might need some different arguments.\par
We shall now turn to operations that are described by different kinds of morphisms. For the array case, codings (or projections) is a common such operation.
\begin{theorem}\label{thm-projection}
Any of our array pattern language classes (over arbitrary alphabets) is closed under projections.
\end{theorem}
\begin{proof}
The following proof works for every $z \in \{h, p, r, c, rc\}$. Let $\alpha$ be some array pattern and let $\pi:\Sigma\to\Gamma$ be a surjective mapping that describes some projection. We shall now show that $\pi(L_{\Sigma, z}(\alpha))=L_{\Gamma, z}(\alpha)$.\par
Namely, consider some array $U\in L_{\Sigma, z}(\alpha)$.
$U$ is obtained from $\alpha$ by replacing any variable $x$ occurring in $\alpha$ by some
array $h(x)$. If we replace $x$ by $\pi(h(x))$ instead (which has the same dimensions as $h(x)$),
we can see that $\pi(U)\in L_{\Gamma, z}(\alpha)$.\par
Conversely, fix some letter from $\pi^{-1}(\ta)$ for each $\ta\in\Gamma$ for now and denote it by $\kappa(\ta)$.
This is possible, as $\pi$ is a surjective mapping.
As $\pi(\kappa(\ta))=\ta$ for each $\ta\in\Gamma$ by construction, for any $U\in L_{\Gamma, z}(\alpha)$ we can
describe some $V\in L_{\Sigma, x}(\alpha)$ such that $\pi(V)=U$, just by taking $V=\kappa(U)$.
\end{proof}
The result does not generalize to (string) morphisms where each image is of the same length.
\begin{proposition}\label{prop-length2-morphisms} For any non-unary alphabet $\Sigma$,
$\mathcal{L}^{\oneD}_{\Sigma}$ is not closed under
morphisms that map every letter to a word of length two.
\end{proposition}
\begin{proof}
Consider the pattern $x$ on the alphabet $\Sigma=\{\ta, \tb\}$, describing the universal language $\Sigma^+$, and the morphism $h$ with $h(\ta)=\ta \tb$ and $h(\tb)=\ta \tb$. Then, $h(\Sigma^+)=\{\ta \tb\}^+$, which is not a pattern language, as an easy analysis shows.
\end{proof}
\begin{corollary}
None of the array pattern language classes under consideration (over some non-unary alphabet) is closed under two-dimensional morphisms.
\end{corollary}
This is also true for the more general operation of substitution, with the same examples.
Let us also remark that Proposition~\ref{prop-length2-morphisms}
did not rely on the fact that we restricted our attention to one specific non-unary alphabet $\Sigma$.
However, if we have a specific alphabet, then we can even state:
\begin{proposition}\label{prop-bijective-coding}
Any of our array pattern language classes (over some fixed alphabet $\Sigma$) is closed under some projection $\pi:\Sigma\to\Sigma$
if and only if $\pi$ is a bijection.
\end{proposition}
\begin{proof}
If $\pi$ is some bijection, then the argument given in the proof of Theorem~\ref{thm-projection} applies. Recall now that for finite sets as $\Sigma$, $\pi$ is a bijection if and only if $\pi$ is a surjection. If $\pi$ is not a bijection, then $\pi$ is not a surjection; hence, there is some $b \in \Sigma\setminus\pi(\Sigma)$, Consider the pattern $\alpha=x$. Clearly, $\pi(L_{\Sigma, z}(x))=(\pi(\Sigma))^{**}$, $z \in \{h, p, r, c, rc\}$, but any array pattern language over $\Sigma$ will also contain patterns with the letter $b$.
\end{proof}
This immediately implies the following two results:
\begin{corollary}
Any of our array pattern language classes (over
binary alphabets) is closed under conjugation.
\end{corollary}
\begin{corollary}
None of our array pattern language classes (over some fixed alphabet $\Sigma$) is closed under all letter-to-letter morphisms.
\end{corollary}
Alternatively, we could also look at inverse morphisms. Here, we already get negative results for inverse codings.
\begin{proposition}\label{prop-one-dim-inverse-coding} For any alphabet $\Sigma$ with at least four letters,
$\mathcal{L}^{\oneD}_{\Sigma}$ is not closed under
inverse letter-to-letter morphisms.
\end{proposition}
\begin{proof}
Consider the coding
$h: \ta \mapsto 1, \tb\mapsto 1, \tc\mapsto 2, \td\mapsto 3$. The pattern $xx$ generates (over the alphabet $\{1,2,3\}$) the language $L=\{ww\mid w\in \{1,2,3\}^+\}$. However, $L'=h^{-1}(L)$ is not a pattern language. The shortest words in $L'$ are of length two, so that a hypothetical pattern for $L'$ must be of the form $yz$ or $yy$. In the first case, $\tc \td\notin L'$ is produced, while in the second case, $\ta \tb\in L'$ cannot be produced.
\end{proof}
\begin{corollary}
None of the array pattern language classes under consideration (over some sufficiently large alphabet) is closed under inverse (two-dimensional) morphisms.
\end{corollary}
\subsection{Operations Similar to String Language Operations}\label{sec:similartostringcaseops}
Let us first turn to the concatenation operation.
As a warm-up, we first consider the string case.
\begin{lemma}\label{lem-string-concatenation} For any alphabet $\Sigma$,
$\mathcal{L}^{\oneD}_{\Sigma}$ is closed under concatenation.
\end{lemma}
\begin{proof}
Consider two patterns $\alpha$ and $\beta$. After renaming, we can assume that $\alpha$ and $\beta$ do not contain any
identical variables.
Then, $L^{\oneD}_{\Sigma}(\alpha\beta)=L^{\oneD}_{\Sigma}(\alpha)L^{\oneD}_{\Sigma}(\beta)$.
\end{proof}
The first thing one should note is that the concatenation of two arrays could be undefined (i.\,e., if their dimensions do not match), even though the concatenation of the two according languages need not be empty. However, we can prove:
\begin{theorem}\label{thm-catenation} Fix some alphabet $\Sigma$.
\begin{itemize}
\item $\mathcal{L}_{\Sigma,r}$ is closed under row concatenation $\varominus$;
\item $\mathcal{L}_{\Sigma,c}$ is closed under column concatenation $\varobar$;
\item $\mathcal{L}_{\Sigma,p}$ and $\mathcal{L}_{\Sigma,h}$ are closed both under row and under column concatenation.
\end{itemize}
\end{theorem}
\begin{proof}
We only prove the first item. The others follow similarly. Observe that in the case of $\mathcal{L}_{\Sigma,h}$, we need Lemmas~\ref{lem-hom-charac} and~\ref{lem-hom-closure} to finish the argument. Let $\alpha$ and $\beta$ be two array patterns. We can assume that the variable alphabets of $\alpha$ and $\beta$ are disjoint.
We want to construct an array pattern $\gamma$ such that
$$L_ {\Sigma,r} (\gamma)=L_ {\Sigma,r}(\alpha)\varominus L_ {\Sigma,r}(\beta).\qquad (*)$$
Let $m_\alpha$ be the number of rows of $\alpha$ and $n_\alpha$ be the number of columns of $\alpha$.
Accordingly, $m_\beta$ is the number of rows of $\beta$ and $n_\beta$ is the number of columns of $\beta$.
If $n_\alpha=n_\beta$, then we can set $\gamma=\alpha\varominus\beta$ to satisfy $(*)$ as in the string case of Lemma~\ref{lem-string-concatenation}.
More generally, set $n=\operatorname{lcm}(n_\alpha,n_\beta)$.
$n$ is the width of the smallest arrays in $L_ {\Sigma,r}(\alpha)\varominus L_ {\Sigma,r}(\beta)$.
More generally speaking, any array in $L_ {\Sigma,r}(\alpha)\varominus L_ {\Sigma,r}(\beta)$ has some width
that is a multiple of $n$. We are going to exploit this property by constructing two array patterns $\alpha'$
and $\beta'$ of width $n$ such that
$$L_ {\Sigma,r}(\alpha)\varominus L_ {\Sigma,r}(\beta)
=L_ {\Sigma,r}(\alpha')\varominus L_ {\Sigma,r}(\beta'),$$
so that we can apply our previous reasoning, defining now $\gamma=\alpha'\varominus\beta'$.
Consider the two-dimensional morphism $h_\alpha$ that maps every variable $x$ of $\alpha$ on an array of height one and
width $n'_\alpha=\frac{\operatorname{lcm}(n_\alpha,n_\beta)}{n_\alpha}$,
more precisely, onto $x_1x_2\cdots x_{n'_\alpha}$.
Here, $x_1,\dots,x_{n'_\alpha}$ are ``new variables.''
Then, $h_\alpha(\alpha)$ is an array pattern of width $n=n_\alpha\cdot n'_\alpha$.
Similarly, we define a morphism $h_\beta$, yielding an array pattern $h_\beta(\beta)$ of width $n=n_\beta\cdot n'_\beta$.
Now, any array $U$ in $L_ {\Sigma,r}(\alpha')$ has a width that is a multiple of $n$ and also belongs to $L_ {\Sigma,r}(\alpha)$,
and conversely any $U\in L_ {\Sigma,r}(\alpha)$ that has a width that is a multiple of $n$ belongs to $L_ {\Sigma,r}(\alpha')$.
Together with similar statements for the pattern $\beta$, the claim follows.
\end{proof}
It is not a coincidence that for $\mathcal{L}_{\Sigma,r}$ and $\mathcal{L}_{\Sigma,c}$,
we had to focus on the ``correct'' concatenation operation in the preceding theorem.
More precisely, we can show:
\begin{theorem}
\label{thm-concatenation-nonclosure}
Fix some non-unary alphabet $\Sigma$.
\begin{itemize}
\item $\mathcal{L}_{\Sigma,r}$ is not closed under column concatenation $\varobar$;
\item $\mathcal{L}_{\Sigma,c}$ is not closed under row concatenation $\varominus$;
\item $\mathcal{L}_{\Sigma,rc}$ is neither closed under row nor under column concatenation.
\end{itemize}
\end{theorem}
\begin{proof}
Again, we only prove the first item; the others can be seen in a similar fashion.
Consider the array patterns $\alpha=\begin{bsmallmatrix}
x_1 & x_2\\
x_2 & x_3
\end{bsmallmatrix}$
and $\beta=\begin{bsmallmatrix}
y_1 & y_2\\
y_2 & y_3
\end{bsmallmatrix}$.
Notice that $L_{\Sigma,r}(\alpha)=L_{\Sigma,r}(\beta)$,
There are arrays of width four and height two in $L:=L_{\Sigma,r}(\alpha)\varobar L_{\Sigma,r}(\beta)$,
and these are the smallest arrays in $L$. Hence, any array pattern $\gamma$ with $L=L_{\Sigma,r}(\gamma)$ has width four and height two.
Let $\gamma=\begin{bsmallmatrix}
z_1 & z_2 & z_3 & z_4\\
z_5 & z_6 & z_7 & z_8
\end{bsmallmatrix}
$, with possibly some of the variables being the same.
As $\begin{bsmallmatrix}
\ta & \tb & \ta & \tb\\
\tb & \ta & \tb & \ta \\
\end{bsmallmatrix}
\in L
$, but $\begin{bsmallmatrix}
\ta & \tb & \ta & \ta\\
\tb & \ta & \tb & \ta \\
\end{bsmallmatrix}
\notin L
$ and also
$\begin{bsmallmatrix}
\ta & \ta & \ta & \tb\\
\tb & \ta & \tb & \ta \\
\end{bsmallmatrix}
\notin L
$, $z_2=z_5$ and $z_4=z_7$.
More generally, it can be verified that the array pattern
$\zeta=\begin{bsmallmatrix}
z_1 & z_2 & z_3 & z_4\\
z_2 & z_6 & z_4 & z_8
\end{bsmallmatrix}
$ describes those and only those arrays of width four and height two that belong to $L$.
However, $U=\begin{bsmallmatrix}
\ta & \ta & \tb & \ta & \tb & \tb\\
\tb & \ta & \tb & \tb& \ta & \tb
\end{bsmallmatrix}$
is an array from $L_{\Sigma,r}(\zeta)$ (namely, consider $z_1=\ta\ta$, $z_2=\tb\ta$, $z_3=z_4=z_6=\tb$ and $z_8=\ta\tb$) that does not belong to $L$, as this would mean that
$U=U_1\varobar U_2$ for some array $U_1$ or $U_2$ of width two and height two from $L_{\Sigma,r}(\alpha)=L_{\Sigma,r}(\beta)$.
However, neither $U_1=\begin{bsmallmatrix}
\ta & \ta\\
\tb & \ta
\end{bsmallmatrix}
$
nor $U_2=\begin{bsmallmatrix}
\tb & \tb\\
\ta & \tb
\end{bsmallmatrix}
$
belongs to $L_{\Sigma,r}(\alpha)=L_{\Sigma,r}(\beta)$.
\end{proof}
Notice that the proofs of negative closure properties necessitate a non-unary alphabet to work.
We now turn to the Kleene closure. Here, we can again first show a non-closure result for the string case that then readily transfers to the array cases.
\begin{lemma}
\label{lem-Kleene}
Let $\Sigma$ be a non-unary alphabet.
Consider $\alpha=xx$. Then, $(L^{\oneD}_{\Sigma}(\alpha))^+\notin \mathcal{L}^{\oneD}_{\Sigma}$.
\end{lemma}
\begin{proof}
The shortest words in $L:=(L^{\oneD}_{\Sigma}(\alpha))^+$ are of length two. Hence, there are only two different possibilities for any pattern $\beta$ with $L^{\oneD}_{\Sigma}(\beta)=L$: If $\beta=\alpha=xx$, then $\ta\ta\tb\tb\in L\setminus L^{\oneD}_{\Sigma}(\beta)$, while if $\beta=xy$, then $\ta\tb\in L^{\oneD}_{\Sigma}(\beta)\setminus L$.
\end{proof}
\begin{proposition}\label{prop-Kleene}
Let $\Sigma$ be a non-unary alphabet.
Then, none of the array language families
$\mathcal{L}_{\Sigma,x}$ with $x\in\{r,c,rc,p,h\}$ is closed under
column concatenation closure nor under row concatenation closure.
\end{proposition}
\begin{proof}
Consider $\mathcal{L}_{\Sigma,r}$. Due to Lemmas~\ref{lem-transfer} and~\ref{lem-Kleene}, this class is not closed under row concatenation closure.
For the case of column contentation closure, reconsider the proof of Theorem~\ref{thm-concatenation-nonclosure}.
There, we have presented a language $L\in\mathcal{L}_{\Sigma,r}$ such that $L\varobar L\notin \mathcal{L}_{\Sigma,r}$.
But that argument also shows that the column contentation closure of $L$ does not belong to $\mathcal{L}_{\Sigma,r}$.
The other cases are simillarly seen.
For the case of morphisms, observe that the contradiction in Lemma~\ref{lem-Kleene} was derived by
substituting the variables by words of the same length.
\end{proof}
\subsection{Operations Special to Arrays}\label{sec:arraycaseops}
Recall that the transposition operation is first defined for arrays (or patterns) and can then be lifted to languages and even to language classes.
Nearly by definition, we find:
\begin{lemma}\label{lem-transpose}
Let $\Sigma$ be some alphabet. Let $\alpha$ be a pattern. Then,
$\transpose{L_{\Sigma, r}(\alpha)}=L_{\Sigma, c}(\transpose{\alpha})$
and
$\transpose{L_{\Sigma, c}(\alpha)}=L_{\Sigma, r}(\transpose{\alpha})$.
\end{lemma}
\begin{corollary}\label{cor-transpose}
Let $\Sigma$ be some alphabet. Then,
$\transpose{\mathcal{L}_{\Sigma, r}}= \mathcal{L}_{\Sigma, c}$ and
$\transpose{\mathcal{L}_{\Sigma, c}}= \mathcal{L}_{\Sigma, r}$.
\end{corollary}
Since $\alpha :=
\begin{bsmallmatrix}
x_1 & x_2 \\
x_2 & x_1 \\
\end{bsmallmatrix}$ is identical to its transposition and, as shown in the proof of Lemma~\ref{prcIncompLemma}, describes an $r$ pattern language (a $c$ pattern language), which is not a $c$ pattern language (not an $r$ pattern language, respectively), we can conclude the following:
\begin{proposition}\label{prop-transposition}
Let $\Sigma$ be an alphabet.
Neither $\mathcal{L}_{\Sigma, c}$ nor $\mathcal{L}_{\Sigma, r}$
are closed under transposition.
\end{proposition}
\begin{proof}
Consider the pattern $\alpha :=
\begin{bsmallmatrix}
x_1 & x_2 \\
x_2 & x_1 \\
\end{bsmallmatrix}$.
As $\alpha=\transpose{\alpha}$, $$\transpose{L_{\Sigma, r}(\alpha)}=L_{\Sigma, c}(\transpose{\alpha})=L_{\Sigma, c}(\alpha)\notin \mathcal{L}_{\Sigma, r},$$
as we have shown in Lemma~\ref{prcIncompLemma}.
Symmetrically, the other claim follows.
\end{proof}
\begin{proposition}\label{prop-transposition-a}
For any alphabet $\Sigma$ and $x\in\{h,p,rc\}$, $\mathcal{L}_{\Sigma, x}$ is closed under transposition.
\end{proposition}
\begin{proof}
For $h$ and $p$, this claim is immediate from the fact that we have proper factorizations.
For the case $x=rc$, we use Lemma~\ref{lem-transpose}.
Let $\alpha$ be some pattern.
Then,
\begin{eqnarray*}
\transpose{L_{\Sigma, rc}(\alpha)}&=&\transpose{({L_{\Sigma, r}(\alpha)}\cup {L_{\Sigma, c}(\alpha)})}\\
&=&L_{\Sigma, c}(\transpose{\alpha})\cup L_{\Sigma, r}(\transpose{\alpha})\\
&=& L_{\Sigma, rc}(\transpose{\alpha})
\end{eqnarray*}
This immediately implies the claim.
\end{proof}
With respect to purely geometric operations as turns and reflections, we find the following:
\begin{proposition}\label{prop-turns}
Let $\Sigma$ be some alphabet.
\begin{itemize}
\item $\mathcal{L}_{\Sigma,rc}$, $\mathcal{L}_{\Sigma,p}$ and $\mathcal{L}_{\Sigma,h}$ are closed under quarter-turn.
\item For every $x\in\{r,c,rc,p,h\}$, $\mathcal{L}_{\Sigma,x}$ is closed under half-turn and reflections.
\item $\mathcal{L}_{\Sigma,r}$ and $\mathcal{L}_{\Sigma,c}$ are closed
neither under left nor under right turn.
\end{itemize}
\end{proposition}
\begin{proof}
For the positive closure results, simply observe that the language described by the quarter-turn, by the half-turn or by a reflection of the array pattern $\alpha$ is just the quarter-turn, the half-turn or the reflection of the language described by $\alpha$.\par
For the non-closure properties, by symmetry it suffices to show that there is a language in $\mathcal{L}_{\Sigma,r}$
whose quarter-turn is not in $\mathcal{L}_{\Sigma,r}$.
To this end, consider $L:=L_{\Sigma,r}\left(\begin{bsmallmatrix}x&y\\y&x\end{bsmallmatrix}\right)$.
Observe that the quarter-turn of $L$ is the same as $L_{\Sigma,c}\left(\begin{bsmallmatrix}x&y\\y&x\end{bsmallmatrix}\right)$,
which was proven not to be in $\mathcal{L}_{\Sigma,r}$ in Lemma~\ref{prcIncompLemma}.
\end{proof}
The positive closure properties can be easily observed by applying the geometric operation directly on the array pattern. In order to show non-closure of $\mathcal{L}_{\Sigma,r}$ and $\mathcal{L}_{\Sigma,c}$ with respect to left and right turn, it is again sufficient to observe that the pattern $\alpha$ from above is identical to its left or right turn and then apply a similar argument as in the proof of Lemma~\ref{prcIncompLemma}.\par
Due to symmetry, it does not matter if we consider horizontal or vertical reflections. Notice that both half-turns and reflections coincide in the string case in any meaningful, non-trivial interpretation; in that case,
the operation is also known as mirror image.
\section{Future Research Directions}
A thorough investigation of the typical decision problems for two-dimensional pattern languages like the membership, inclusion and equivalence problem is left for future research. It can be easily seen that the NP-completeness of the membership problem for string pattern languages carries over to $\mathcal{L}_{\Sigma,x}$, $x \in \{p, r, c, rc\}$. On the other hand, for a given array pattern $\alpha$ and a terminal array $W$, the question whether or not $W \in L_{\Sigma, h}(\alpha)$ can be decided in polynomial time by checking whether $W$ is a morphic image of $\alpha$ with respect to a $(\frac{|W|_r}{|\alpha|_r}, \frac{|W|_c}{|\alpha|_c})$-uniform substitution. As shown by Lemma~\ref{equalityLemma}, the equivalence problem for all the classes $\mathcal{L}_{\Sigma,x}$ with $x \in \{h, p, r, c, rc\}$ and $|\Sigma| \geq 2$ can be easily solved by simply comparing the patterns. However, for every $z, z' \in \{h, p, r, c, rc\}$, $z \neq z'$, the problem to decide for given patterns $\alpha$ and $\beta$ whether or not $L_{\Sigma, z}(\alpha) = L_{\Sigma, z'}(\beta)$ might be worth investigating. The inclusion problem for terminal-free nonerasing string pattern languages is still open. Hence, with respect to the inclusion problem, a positive decidability result for two-dimensional pattern languages implies a positive decidability result for terminal-free nonerasing string pattern languages.\par
For string pattern languages it is common to use terminal symbols in the patterns as well as to consider the \emph{erasing} case, i.\,e., variables can be replaced by the empty word. The $p$ pattern languages can be adapted to the erasing case by allowing variables to be substituted by the empty array. Furthermore, the situation of having a terminal symbol at position $(i, j)$ of an array pattern simply forces all the variables in the $i^{\text{th}}$ row to be substituted by arrays of height $1$ and all the variables in the $j^{\text{th}}$ column to be substituted by arrays of width $1$. As in the string case, it is likely that in the two-dimensional case the difference between erasing and nonerasing substitutions and patterns with and without terminal symbols lead to different language classes with different decidability properties, too. \par
Finally, we wish to point out that it is straightforward to generalise our different classes of two-dimensional pattern languages to the three-dimensional or even $n$-dimensional case.
\bibliographystyle{plain}
|
\section{Introduction}
\label{intro}
In the coming years, the next generation of radio telescopes are expected to survey large areas of sky to a sensitivity which could see the discovery of hundreds of transient and variable sources \citep{Fender2015, Metzger2015}.
Previous discoveries of radio transients proved that relying solely on radio data makes it extremely difficult to validate both the nature, as well as the reality of the source (\citealt{Bower2007}; \citealt{Stewart2016}) and that it is vital they are supported by simultaneous multiwavelength observations or rapid follow up.
On the other hand, with a rate of transient/variable candidates as high as expected, it will not be possible to follow up each event with other multiwavelength instruments. However, if the nature of the source is constrained as early as possible, decision-making strategies can be created to follow up the type of events relevant to specific science goals.
The importance of the automatic classification of light-curves has been discussed before in the context of deep surveys carried at optical wavelengths, with attempts to include automatic classification of transients into optical surveys (\citealt{saglia2012, djorgovski2012}). Extended investigations of different ways to approach this task include developing classification techniques based on machine learning methods \citep{richards}, Bayesian theory \citep{pichara}, density approach (eliminating direct measurements of features such as magnitude or color; \citealt{kugler}), fitting non-complete light-curves \citep{Lo2014} and other non-parametric techniques \citep{varughese}.
Those methods have been developed using archival optical light-curves as training datasets, as well as X-ray transients (\citealt{Lo}, with a supervised learning technique implemented). For radio data, \citet{rebbapragada} test different classification algorithms on simulated light-curves in preparation for the ASKAP VAST survey. Most of these methods, although they work effectively on sparse and noisy datasets, require the light-curve/flare to be (mostly) complete. In this work, however, we are focused on investigating whether an initial classification of the source based on the partial information about its light-curve is possible.
For any transient or variable candidate discovered in a blind radio survey, the most basic measurable property of the light-curve is its variability time-scale.
Here we present a potential technique of classifying these types of events by measuring the rise/decline rates of their flares, which could be used as the early step in a more complex classification pipeline. At the moment however, due to the relatively small data sample used to develop the method, it is not a final classification solution.
As we have previously shown (\citealt{paper1}; hereafter PFK15), there is a correlation between the luminosity ($L$) and rise/decline rates ($\tau$) that exists between different classes of synchrotron emitting sources, approximately $L \propto \tau^{5}$.
It covers a wide range of events, from nearby flare stars to extragalactic supermassive black holes, with the more massive and luminous objects displaying variability on progressively longer time-scales.
Although it has been expected previously that more massive and luminous objects vary on much longer time-scales compared to intrinsically faint sources \citep{vdl}, the analysis presented in PFK15 showed a clear distinction between different classes of objects occupying the luminosity -- time-scale parameter space.
Because at the time of discovery of a radio event its distance is unknown (unless a counterpart at other wavelengths is immediately identified), the information about the luminosity is not available, and the discussed relation (Figure~3 in PFK15) is reduced to a distribution of the variability time-scales.
This result offers an opportunity to develop a method which could be used to perform an initial classification of an unknown object based on its radio data only.
Approximate classification by time-scale and luminosity has its analogues in the field of optical transients (e.g. Fig 1 of \citealt{Rau2009}).
It should be noted, that the results presented in PFK15 were based on the analysis of a sample of single flaring events, manually selected from various radio light-curves. However, if we were to base an early classification on those results, the analysis should be done in a reproducible way, such that the bias introduced by identifying flares by eye is minimized. In order to do that, we have developed a basic piece of software which automatically identifies and selects flaring events from the radio light-curves in our sample.
Our routine is based on a simple thresholding approach, where any variability of the light-curve above a chosen flux density threshold is defined as a `flare'. However, this routine is designed to work specifically on the diverse dataset analysed in this study.
In future surveys, where data quality is more consistent, complex algorithms will be more appropriate to be used as part of the detection pipeline.
An important aspect of the presented analysis which needs to be carefully addressed is the bias associated with our sample of radio light-curves. The number of objects analysed within each class represents the frequency and quality of observations rather than the actual sky density of those sources.
Therefore, in order to accurately predict the probability of finding a given class of objects, this effect needs to be accounted for. Although the exact areal densities of analysed classes of objects are not well known, we attempt to estimate those values and convolve them with the obtained distribution of rise/decline rates across a range of considered classes. This final distribution provides an initial reference point against which we can compare the measurements of time-scales taken for any newly discovered transient/variable candidate.
In Section~\ref{data} we describe the sample of light-curves used in the analysis. Section~\ref{flarefinding} provides details of the simple flare finding routine we developed to select flares from our sample of light-curves, as well as an overview of alternative methods for the flare selection. We describe the analysis of selected flares in Section~\ref{analysis} and give details of the estimation of expected areal densities in Section~\ref{se}. Final results are presented in Section~\ref{results} and discussed in Section~\ref{discussion}. Conclusions are given in Section~\ref{conclusions}.
\section{Data}
\label{data}
The data included in the analysis span a broad range of radio variable sources,
observed at frequencies between 5 and 8 GHz, compiled from the literature and the Green Bank Interferometer (GBI) archive\footnote{\url{ftp://ftp.gb.nrao.edu/pub/fghigo/gbidata/gdata/}}.
Overall, the analysed light-curves originate from a range of instruments and different observing programmes, therefore, there is no consistency in their sampling intervals, length of time series or sensitivities.
The number of sources from different classes is as follows: AGN (28), tidal disruption events (2), supernovae (13), GRB afterglows (4), classical novae (7), dwarf nova (1), X-ray binaries (18), magnetar (1), RSCVn (3), algol (3), magnetic CVs (2) and flare stars (4), giving the total number of light-curves from synchrotron emitting sources of 92.
A detailed list of sources included in the sample, together with corresponding references, can be found in PFK15, with the exception of one tidal disruption event not included in the previous work - details of this source are summarized in Table~\ref{tdesample}.
Additionally, listed in Tables~\ref{sctsample} and \ref{esesample} are scintillating sources and extreme scattering events (ESE). Because the discussed luminosity--time-scale correlation applies to intrinsic variability only, these sources are not included in the main analysis. However, we use this sample to investigate the time-scales of extrinsic variability and discuss the extent of overlap with time-scales of synchrotron flares (Section~\ref{discussion}).
The errors of the flux density measurements have been provided for all the GBI data sets, and listed in
selected papers. In cases where the errors were not specifically stated, we have estimated
them based on the published figures, which, in majority of the cases meant taking ten per cent of the flux density measurement.
The following section describes the basic flare finding software, which automatically selects flaring events
from the described sample of radio light-curves, in order to analyse variability time-scales across the classes of objects.
\section{Automatic selection of flares}
\label{flarefinding}
\subsection{Overview of flare identification methods}
There are a number of possible approaches to measuring time-scales, and in particular flaring time-scales, in time-series data. Perhaps the simplest and most intuitive approach is to adopt a given threshold and designate any contiguous sections of the time-series above that threshold as a `flare'. A time-scale measurement can then be obtained by measuring the time spent above the threshold level, but of course this requires that we wait for the flare to subside. Alternatively we may attempt to fit a model of some kind to any data points already recorded. This approach has often been adopted in the past due to its relative simplicity, robustness, and ease of implementation. Depending on the model-fitting requirements the thresholding approach can also be extremely computationally efficient, and hence suitable for real-time applications. \\
However, there are several draw-backs to the basic thresholding approach. First and foremost, finding a suitable threshold level can be a somewhat arbitrary process. If sources of noise are well characterised then a threshold can be chosen on the basis of a calculated false-positive rate, but this may not be possible in all cases. Even if instrument characteristics are well known, an astronomical source may display low-level intrinsic variability which is qualitatively different to the rapid flux-rise-and-decay characteristic of a flaring event. As such it may be necessary to devise some kind of calibration technique (as we do in this work) or even to manually pick a suitable threshold level. The second problem is that of choosing a suitable model for flare-fitting, which again is often somewhat arbitrary, although hopefully motivated by knowledge of the physical processes at work. Additionally, simple thresholding provides no means of separating multiple-superimposed flares, which degrades the accuracy of any model fits. Finally, a point-wise threshold will always lose some sensitivity compared to more sophisticated approaches, if enough is known about the flare-morphology to apply an effective matched-filter. In the more general context of time-scale measurement, model-free metrics such as autocorrelation time and other metrics for aperiodic variability \citep{findeisen2015} can be applied, but typically require a large amount of variable time-series data to work effectively. Since we are interested in time-scales as an early-time classification tool, we do not consider these approaches further here.\\
More advanced change-or-flare-detection techniques typically take a probabilistic approach. As such they give a more rigorous and informative measure of the time-series data, at the cost of additional computational time and complexity. One possibility might be application of the Bayesian Blocks algorithm \citep{scargle1998} for a model-free probabilistic method of change-detection. This side-steps the problem of threshold-determination, but again does not provide an early-time estimate of characteristic time-scale. If a reasonable model can be chosen, then Monte-Carlo methods can be applied to estimate the likelihood of a flare presence with excellent sensitivity \citep{pitkin2014}, or even to dissect the time-series into multiple superimposed flares \citep{Huppenkothen2015}, though this requires significant computational time. A thorough investigation of the relative accuracy and efficacy of such methods would be interesting, but is outside the scope of this work. \\
In our case, advanced flare finding techniques are not suitable to work on the diverse sample of light-curves we analyse. For this study we apply a thresholding approach with a simple exponential model. This provides a reasonably robust time-scale estimation for modest computational effort, albeit with some limitations to applicability, as detailed below.
\begin{figure*}
\includegraphics[width=0.49\textwidth]{Fig1a.eps}
\includegraphics[width=0.49\textwidth]{Fig1b.eps}
\includegraphics[width=\textwidth]{Fig1c.pdf}
\caption{Example figure showing the detection of flaring events in the radio light-curve of the binary SS~433~(GBI). {\em Upper panel, left}: light-curve of the source and flaring events selected by the Automatic Flare Finder (AFF). Marked with the dashed green line is the estimated quiescent flux level (labelled `median' in the key). The blue and red lines indicate the variation above the quiescent level by 3$\sigma$ and 5$\sigma$ respectively. The black symbols on each flare mark: start/end of the flare (triangles), data point triggering the flare detection (pentagons) and the peak of the flare (stars).
{\em Upper panel, right}: histogram of flux density values for the light-curve after sigma clipping has been applied. Flare-threshold values of 3$\sigma$ and 5$\sigma$ are marked by the blue and red vertical lines respectively. Curve overplotted in blue represents a Gaussian distribution with median and median absolute deviation matched to the sigma-clipped data. Curve overplotted in green represents a Gaussian distribution of the same median, but with quiescent variability matched to the formal error bars on the flux-density measurements, clearly showing that the formal error bars underestimate the low-level variability in this light-curve.
{\em Lower panel}: close up view of two flares selected from the light-curve, showing clearly the trigger, beginning, peak and end of the flare markers.}
\label{AFF}
\end{figure*}
\subsection{Automatic Flare Finder (AFF) for PFK15 dataset}
\label{aff}
The flare-identification routine we tested is simple, but nonetheless provided a reproducible means of characterising flare time-scales for this work\footnote{The flare finding software along with the subsequent scripts used to model the probability distributions is available at {\url{https://github.com/4pisky/variability-timescale-analysis-paper}}}.
Using more advanced methods discussed in the previous section would be impractical, since our datasets are very diverse (Section \ref{data}) and do not meet the requirements of such complex techniques.
It should be stressed that, while this simple method is sufficient to select flares from our sample of light-curves, it is not intended to be used as a generic tool.
It uses a simple thresholding approach, as follows:
given a pre-determined quiescent or background flux level estimate, $b$,
and an estimate of the signal variation due to noise, $\sigma$,
the routine first steps through the data looking for data points $x$ where the flux $f_x$ is more than $5\sigma$ above the quiescent flux:
\begin{equation}
x: \quad f_x > b + 5\sigma\, \textrm{;}
\label{eq:threshold}
\end{equation}
these are referred to as the `trigger' points.
Once a trigger is found, the algorithm searches for the nearest data points before and after the trigger which have flux less than
$1\sigma$ above the quiescent flux-level, these are designated as the flare start and end.
The tricky part of this process is determining a suitable quiescent / background flux level, and estimating the levels of signal variation present when the source is not undergoing an outburst.
For a source which has a well-sampled long-term light-curve encompassing extended periods of quiescence, the first problem -- `background' flux-level estimation -- is straightforward; simply taking the median flux level provides a reasonable estimate.
However, when analysing a light-curve which displays flares for a significant fraction of data-period, the median may overestimate the background level.
The second problem, estimating the quiescent low-level variance, is harder.
Again considering the ideal case of a well-sampled long-baseline light-curve, at some level we could simply use the formal errors on the data points and designate anything greater than $5\sigma$ above the median a flare.
However, many of the light-curves analysed for this paper display low-level variation which does not subjectively qualify as a flare, but which is nonetheless larger than would be expected from the formal errors, and appears persistent over multiple data points (e.g. SS 433, as shown in Figure~\ref{AFF}) -- we designate this as intrinsic quiescent variation (though varying telescope systematic noise-levels cannot be ruled out without access to the original raw data).
To avoid the need to manually change the $\sigma$-threshold for each dataset according to the level of quiescent variation, we required an alternative method of estimating the low-level variation.
For this purpose, we employed the sigma-clipping routine from the {\em astropy} library \citep{astropy}, using the default clipping-threshold of $3\sigma$, and iterating until convergence. This means estimating the median value of the whole light-curve, clipping all the points which are above $3\sigma$ and repeating the procedure until there are no points to reject. We have used median absolute deviation as a measure of the data variability $\sigma$.
For some of the light-curves we analysed this does a good job of masking the high-flux outlier data points representing flares, however, alternative methods of finding the low-level variation for light-curves which require more specific approach are discussed in Section~\ref{prep}.
We then estimate the median and median absolute deviation of the remaining unmasked `quiescent' data, and use these quantities for the values of $b$ and $\sigma$ in Equation~\ref{eq:threshold}. Figure~\ref{AFF} shows an example of the flares identified in a GBI light-curve for the X-ray binary SS~433.
\section{Analysis}
\label{analysis}
The quality of the data described in Section \ref{data} varies significantly across the sample.
Data provided by the GBI consists of noisy light-curves for both slowly (AGN) and rapidly
varying sources (XRBs, RSCVn), observed on time-scale of years. Data compiled
from the literature, although much less noisy, in many cases shows pre-selected flares,
sometimes with the background emission subtracted (GRBs, SNe).
The diversity of available light-curves presented a substantial challenge in designing
software in a way that would allow us to detect flaring events with no manual intervention
during the process. Even the simple thresholding method described in Section \ref{aff} sets a number of constraints on the light-curve which make it difficult to implement across the variety of the datasets.
Firstly, it requires the light-curve to be long enough in relation to the duration of the flare,
such that the background flux level (excluding flares) can be accurately estimated.
Secondly, it needs to be sufficiently noise-free in order to avoid false detections. In this section we describe the general
adjustments that have been made to the flare identification routine and/or data sample
in order to make the selection of flares as optimal as possible.
\subsection{Data preparation}
\label{prep}
Testing the AFF on all the available light-curves identified a set of issues which allowed us to divide the whole data sample into four groups. Those groups, together with list of adjustments made and example plots are discussed below.
\subsubsection{Short duration flares selected from GBI (15\% of all light-curves; 84\% of analysed flares)}
Light-curves in this group consist of multiple, short duration
($\approx$~days) flares, observed on a time-scale of years.
This set of data as the only one in our sample presented no issues for the AFF.
Figure~\ref{gbim} shows an example of such light-curve, with a number of
identified flares.
\begin{figure}
\includegraphics[scale=0.415]{Fig2.eps}
\caption{An example of a light-curve showing multiple short flares for a periodic X-ray binary LS I +63$^{\circ}$303.
Flaring events selected from this 8.3~GHz GBI light-curve by the AFF are marked with the black stars. The background level has been estimated using sigma clipping method described in Section~\ref{aff}.}
\label{gbim}
\end{figure}
\begin{figure}
\includegraphics[scale=0.415]{Fig3a.eps}
\includegraphics[scale=0.415]{Fig3b.eps}
\caption{{\em Upper}: An example of a light-curve showing multiple flares with the quiescent level insufficiently well characterised for the automatic software to identify flares. (3C120, \citealt{3C}). {\em Lower}: The same light-curve, where the quiescent flux level
has been estimated as the $15^{\mathrm{th}}$ percentile of the flux measurements, and, for calculating the flare-detection-trigger amplitude, we set $\sigma$ equal to the mean of the formal error bars on the flux density measurements.}
\label{litm}
\end{figure}
\begin{figure}
\includegraphics[scale=0.415]{Fig4a.eps}
\includegraphics[scale=0.415]{Fig4b.eps}
\caption{{\em Upper}: An example of a light-curve showing long time-scale flares (0224+671, GBI).
With the background and the quiescent low-level variability estimated as in lower panel of Figure~\ref{litm}, some odd noisy data points are falsely identified as flares.
{\em Lower}: The same light-curve smoothed, with one major flare recognized by the software.}
\label{gbis}
\end{figure}
\subsubsection{Flares compiled from the literature (30\% of all light-curves; 9\% of analysed flares)}
\label{datalits}
This group comprises the light-curves of repeating, as well as several cataclysmic events.
However, information about the background level in this sample is limited to the close vicinity of the flares,
and therefore is insufficient for the correct estimation of the background emission and the low-level variability.
In most cases, the automatically estimated values were too high when compared to the
peak flux of the outbursts, and as a result, the flares were not recognized
as varying above chosen (5$\sigma$) threshold.
In order to resolve that problem, we have estimated the quiescent flux level of the light-curve as the
15$^{\mathrm{th}}$ percentile of the flux density measurements.
For calculating the flare-detection-trigger amplitude, we set $\sigma$ equal to the mean of the formal error bars on the flux density measurements.
An example of such a light-curve, showing both original attempt to select flares by AFF,
as well as flares selected with the alternative approach is shown in Figure~\ref{litm}.
\begin{figure}
\includegraphics[scale=0.415]{Fig5.eps}
\caption{Example of a light-curve (RS Oph, \citealt{RSOph}) limited to the flaring event, with no information about the background flux density. }
\label{lits}
\end{figure}
\subsubsection{Long duration flares selected from GBI (14\% of all light-curves; 2\% of analysed flares)}
\label{datagbis}
This group of light-curves consists of long duration ($\approx$~months) flares,
collected from the GBI database. As in the previous group, the datasets did not
provide enough information for the estimation of the background and the quiescent variability, and required
both of these parameters to be calculated in the same way as described in Section~\ref{datalits}.
However, here the additional difficulty was introduced by scattered noise in the data.
Some of the noisy data points have been falsely recognised by the AFF as the beginning and/or
end of the flare. The upper panel of the Figure~\ref{gbis} shows an example of such false detections.
This problem has been solved by smoothing the data, with a window size of
$\approx$~20 days. This size of the window has been chosen based on the expected
time-scale of the AGN intrinsic variability. Variabilities in those types of objects,
observed on time-scales shorter
than 20 days are more likely to be associated with scintillation.
Lower panel of Figure~\ref{gbis} shows the result of smoothing, with flaring events
identified by the AFF.
\subsubsection{Pre-selected flares compiled from the literature (41\% of all light-curves; 5\% of analysed flares)}
This group of datasets
contains mainly single, pre-selected flaring events, with no information about the
light-curve before and after the outburst. Included here are flares from supernovae, GRBs, classical novae and tidal disruption events. In those cases, we decided
to measure rise and decline time-scales directly (Figure~\ref{lits}), using method described in Section~\ref{sec:mvt}.\\
For light-curves specified in Sections~\ref{datalits} and \ref{datagbis}, the sigma clipping routine as described in Section~\ref{aff} failed to detect 70~per~cent of the flares due to insufficient information about the background. We note that setting the background level in this alternative way is a subjective choice and can introduce some bias in the analysis.
Therefore, we have repeated the analysis for a range of parameters, setting the quiescent level to 10, 15, 20 and 30th percentile of the flux density. The results were then investigated 'by eye' to decide which approach selected the highest percentage of flares. While each of those estimations caused some of the flares to be missed, we found that the 15th percentile was the optimal choice for the discussed dataset. Figure~\ref{missed-flare-example} in Appendix~\ref{additional_figures} gives an overview of flares selected from an example light-curve depending on the background estimation. It shows that setting the threshold too low can merge poorly separated flares into one, while too high values results in missing fainter flares.
We have also checked how different choices of the background in those groups of light-curves affect the final results -- see Figure~\ref{background-est-var} in Appendix~\ref{additional_figures} and refer to the analysis in Sections~\ref{sec:mvt} and \ref{results}.
Since the discussed datasets form the entire sample of flare stars, as well as most of AGN light-curves, the effects of changing the flare detection threshold are most visible in those two populations of sources (left panel in Figure~\ref{background-est-var}).
The centre of the diagram, which is dominated by flares that do not suffer the background uncertainty (fast GBI light-curves, individual flares from GRBs, SNe and Novae) shows minimal difference. The right panel of Figure~\ref{background-est-var} demonstrates that the choice of background does not have a significant impact on final probability distributions, due to the affected classes of object populating the extreme ends of the time-scale parameter space.\\
Overall, the automatic selection of flares fails in about 15--20~per~cent of cases, missing flares that can be easily identified by eye. Those are mainly superimposed or close together flares which can not be separated with a simple thresholding method, such as the ones presented in Figure~\ref{background-est-var}.
\subsection{Measuring variability time-scales}
\label{sec:mvt}
We have run the AFF on the whole sample of light-curves adjusted as described in previous section, selecting 1290 single outbursts for further analysis, which, together with 38 pre-selected flares, gives the total number of 1328 flaring events from 86 distinct objects.
In order to measure rise and decline time-scales for each of the selected flares, we attempted fitting exponential functions to the data:
\begin{equation}
F = A e^{st}+B,
\end{equation}
\noindent where amplitude $A$, background $B$ and characteristic time-scale $s$ are the free parameters of the fit.
Exponential fits provide distance independent measurements of the rise/decline rates, and, are physically motivated \citep[][where the flux density of a source is an exponential function of the optical depth.]{vdl}
However, we found that in most cases the exponential fit would fail. One of the likely reasons
for that is the quality of the data, where time sampling is often insufficient for fitting the three parameters model.
\noindent Therefore, we decided to estimate the background emission and subtract it from the flare prior to the fit. One obvious choice of this background is the quiescent flux density level $b$ estimated earlier for each of the light-curves (Sections~\ref{aff} and \ref{prep}). However, for poorly sampled light-curves, the beginning/end of the flare (defined as in Section \ref{aff}) can sometimes fall below this background level (for example, the first flare detected in lower panel of Figure~\ref{litm}). Subtracting this background would result in selecting only part of the flare for further analysis.
In order to make sure that the entire flare is selected, we decided to calculate a separate background level for each flare $b_{\mathrm{flare}}$ either as the first percentile $p^{\mathrm{1st}}$ of the flux values of the light-curve ($F_{\mathrm{all}}$) or the minimum flux density measurement of a given flare ($F_{\mathrm{flare}}$):
\begin{equation}
b_{\mathrm{flare}} = \mathrm{min} \left(p^{\mathrm{1st}}\left(F_{\mathrm{all}}\right); \mathrm{min}\left(F_{\mathrm{flare}}\right)\right).
\end{equation}
\noindent The background emission calculated this way is then subtracted from the flux density measurements of every flare selected by the AFF.
\begin{figure}
\includegraphics[scale=0.42]{Fig6.pdf}
\caption{Distribution of the $\mathrm{log}\, \chi^{2}_{\mathrm{red}}$ values for rise and decline rates calculated as in Equation~\ref{fiteq}.}
\label{chi2}
\end{figure}
\noindent For each single flare we measure the rise and decline rates, $s$, by fitting the linear
function in the log-lin space (see Fig~1.~in PFK15):
\begin{equation}
\mathrm{ln} \left( F - b_{\mathrm{flare}} \right) = st+const.
\label{fiteq}
\end{equation}
\noindent No measurements have been taken for rise/decline phases consisting of three data points or less. Excluding flares for which neither rise or decline phase could be measured decreased the final sample to 804 flaring events, with 564 rise and 649 decline phases fitted successfully. Fits to the data have been done
using the {\em curve fit} routine from {\em scipy.optimize} package, and include error bars of the flux measurements.
Figure~\ref{chi2} shows the distribution of $\chi^{2}_{\mathrm{red}}$ values for
measurements of rise and decline phases of the flares. It should be noted that for $\approx$~30~per~cent of measurements the $\chi^{2}_{\mathrm{red}}$ is higher than~10. This might partly be caused by underestimating error bars of flux measurements. However, we expect that the main reason is that the simple two parameter model used in the analysis often fails to accurately fit the data, which may require more complex approach.
\noindent Having a set of measurements for each flare, consisting of
peak flux density, rise/decline time-scales and reduced $\chi^{2}_{\mathrm{red}}$,
we can map those results
with a corresponding class to which the source belongs, its distance and the
observed frequency. This enables us to calculate peak radio luminosity of each flare. Results
of those measurements for the rising phase are shown in Figure~\ref{LT}, where we plot
peak radio luminosity against rise time $\tau$ of the event (where $\tau=1/s$). The overall correlation of the form
$L \propto \tau^{6}$ is steeper than $L \propto \tau^{5}$ obtained by manual measurements.
In particular, the relation presented here shows more scatter for relatively lower-luminosity
classes such as XRBs or RSCVn, than it did in results reported in PFK15.
This effect is clearly demonstrated in Figure~\ref{TH}, where histograms of the rise rates measured automatically
are overlaid with those corresponding to manual measurements (in case of manual measurements, the level of background emission for each flare has been estimated by eye and subtracted from the flux density measurements). It should be noted, that the scatter observed in time-scales of the different types of object is partly due to a wide range of time-scales covered by individual sources. A variation of Figure \ref{LT} is shown in Figure \ref{single-source-scatter}, where
single-source measurements for several classes are highlighted.
We have tested whether rejecting measurements with high and/or low $\chi^{2}_{\mathrm{red}}$ values decreases that scatter, however, neither the final range of the time-scales populated by each class or the overall time-luminosity relation have been significantly changed by those constraints. Therefore, we have decided to include all the measurements.
The observed scatter in measured time-scales is discussed in more detail in Section~\ref{discussion} and
a summary of the manual and automatic results can be found in Table~\ref{manvsauto}.
\noindent In the analysis described in this section, we have automatically reproduced results of the variability
time-scales measurements reported in PFK15.
From this point we focus on converting those results into a method which could be used to perform an initial classification of radio transient and variable sources.
\noindent The diagram shown in Figure~\ref{chart} gives an overview of steps included in the analysis described in this section, as well as a list of remaining steps which are discussed in Sections~\ref{se} and \ref{results}.
\begin{figure*}
\includegraphics[width=\textwidth]{Fig7.pdf}
\caption{Peak radio luminosity plotted against exponential rise time-scale for a range of synchrotron emitting sources, approximately of the form $L \propto \tau^{6}$.}
\label{LT}
\end{figure*}
\begin{table}
\caption{Parameters of the fits to the rise and decline phases (with the following formula: $\log(L_{\mathrm{peak}}) = a \times \log\tau + b$), for both manual (a$_\mathrm{M}$, $b_\mathrm{M}$) and automatic (a$_\mathrm{A}$, $b_\mathrm{A}$) methods. }
\begin{threeparttable}[b]
\begin{tabular}{|l|@{}c|@{}c|c|@{}c}
\hline
& a$_\mathrm{M}$ $\pm \delta$ a$_\mathrm{M}$ & b$_\mathrm{M}$ $\pm \delta$ b$_\mathrm{M}$ & a$_\mathrm{A}$ $\pm \delta$ a$_\mathrm{A}$ & b$_\mathrm{A}$ $\pm \delta$ b$_\mathrm{A}$ \\
\hline
Rise & 5.21 $\pm$ 0.15 & 31.08 $\pm$ 1.84 & 6.01 $\pm$ 0.12 & 31.49 $\pm$ 1.42 \\
Decline & 5.09 $\pm$ 0.22 & 29.44 $\pm$ 2.88 & 6.70 $\pm$ 0.22 & 29.11 $\pm$ 2.11 \\
\hline
\end{tabular}
\end{threeparttable}
\label{manvsauto}
\end{table}
\section{Sky Densities}
\label{se}
We aim to use the results obtained in Section~\ref{sec:mvt} as a base for the probabilistic classification method by modelling the measured time-scale distributions as Gaussian functions, while assuming that for each class we have reasonably well measured the width of the distribution.
However, the compiled dataset consists of observations originating from various surveys,
performed with a wide range of instruments with different sensitivities, fields of view and other parameters which contribute to certain preferences in the available sample of light-curves. As such, the analysed dataset does not accurately represent the underlying populations of considered classes of object.
Although it is not possible to properly account for all these biases, in this section we attempt to estimate the areal densities of objects within each class. This correction, although basic, provides more realistic picture of time-scale distribution for different populations of radio transient and variable sources.
\\
\subsection{Method of estimating areal densities}
\label{sec:method}
{\em Extragalactic events.} Areal density estimates for most of the extragalactic classes of objects
are available in the literature.
Here we extrapolate these results to our chosen flux density limit
of 0.1~mJy, assuming that events of the extragalactic origin are distributed as:
\begin{equation}
N \propto S^{- \frac{3}{2}},
\end{equation}
\noindent where $N$ is the expected number of sources observed with the
flux density limit $S$.
\noindent{\em Galactic objects.} Space densities of Galactic objects have been
studied previously
and are available in the literature.
In order to convert spatial distribution into areal density, we need
to consider the volume of the Galaxy each class populates. First,
we calculate the maximum distance at which source belonging to
considered class can be observed, with assumed flux density limit.
We estimate a typical radio luminosity of the class and use the
following relation to evaluate the distance:
\begin{figure}
\includegraphics[scale=0.54]{Fig8.pdf}
\caption{Dashed histograms represent distribution of automatically measured rise rates for a range of classes of objects. For comparison, overplotted in grey are previously reported \citep{paper1} manual measurements.}
\label{TH}
\end{figure}
\begin{figure}
\includegraphics[scale=0.31]{Fig9.pdf}
\caption{Illustration of the process flow showing the main steps of the analysis presented in Section~\ref{analysis}, from the compiled sample of the data to measuring rise/decline rates. The remaining steps, listed in lower panel, are described in Sections~\ref{se}~and~\ref{results}. Rectangular frames correspond to parts of the analysis done manually; automatic steps are listed in rounded frames.}
\label{chart}
\end{figure}
\begin{equation}
d_{\mathrm{max}} = \sqrt{\frac{L}{4 \pi F_{\mathrm{lim}}}} .
\end{equation}
Unlike for extragalactic sources, the space density $\rho$ of galactic sources depends on their position in the Galaxy and drops as we move out of the Galactic plane and away from the Galactic center.
The density profile can be described as
\begin{equation}
\rho(z) = \rho_0 \mathrm{e}^{\left(-|z|/h\right)},
\end{equation}
\noindent where $z = d \mathrm{sin}b$ is the distance from the Galactic plane ($d$ is distance to the source, $b$ is the Galactic latitude), $h$ is the scale height (describing a distance over which the number of sources drops by factor of $e$) and $\rho_0 = \rho(0)$ is the local space density of objects (we assume that the radial dependence of $\rho$ is negligible). With this assumption we calculate the volume of space occupied by a given class of objects within distance $d$ (details of the calculation of generalised volume of space depending on the density profile can be found in \citealt{vmax}).
Finally, including a duty cycle estimate for each class of object gives the
actual number of sources visible in the sky in a single snapshot.
Table~\ref{rate} summarises areal densities evaluated for each class.
Details of the calculations are described in the following sections.
\subsection{Types of transients and variables}
\label{tt}
\subsubsection{Tidal Disruption Events}
\label{tde}
There have been several radio detections of tidal disruption events (TDEs) to date, however, only two of them are sampled well enough to be included in this analysis.
\citet{Frail2012} estimated the sky rate of such events to be
0.1~deg$^{-2}$ at the 0.3~mJy flux density limit. Using the method described in
Section~\ref{sec:method} we estimate that at the 0.1~mJy limit,
the TDE rate is $\approx$~0.52~deg$^{-2}$.
\subsubsection{Gamma Ray Bursts}
\label{grb}
The areal density of Orphan GRB Afterglows -- classical GRBs with the explosion
axis directed away from the line of sight, which should dominate, is derived by \citet{Frail2012}
at the 0.3~mJy flux limit for a 10$^{\circ}$ beam opening angle.
Their result of 0.01~deg$^{-2}$ scaled down to 0.1~mJy gives 0.052~deg$^{-2}$ in a single snapshot.
\subsubsection{Supernovae}
\label{sn}
\citet{Frail2012} give areal densities for SNII, SNIbc and SN~1998bg-like sources. Adding all those sky
densities and extrapolating the results to our flux density limit, gives a rate of 0.21~deg$^{-2}$ at a single epoch.
\subsubsection{Active Galactic Nuclei}
\label{agn}
In order to estimate the surface density of radio-variable AGN, we looked at the number of all variable objects found in the FIRST
survey data. Out of 1600 variable sources reported by \citet{FIRST}, 489 were identified to be AGN and
variable galaxies. Another 120 are believed to be highly variable quasars. The analysed observations
covered 8444~deg$^{2}$, at the sensitivity of 0.15~mJy. The source detection threshold
chosen for the analysis was 1~mJy. Extrapolating this result to 0.1~mJy with the spherical approximation
described in Section~\ref{sec:method}, gives the AGN sky density of 2.3~deg$^{-2}$ at single epoch.
\subsubsection{X-ray binaries}
\label{xrb}
\noindent There are about 1000 X-ray binaries within our Galaxy, which, when above the flux density limit of 0.1~mJy chosen in this work, are in outburst and can be detected in a radio survey \citep[e.g.][]{xrbrate}. With the estimated duty cycle of XRBs of $\approx$~1\% we get $\approx$~10 sources in the sky, at a single epoch. This gives a rate of about 2.4~$\times$10$^{-4}$~deg$^{-2}$ objects at 0.1~mJy flux density limit.
\subsubsection{Non-magnetic CVs}
\label{dn}
\noindent {\em Dwarf Novae}. Having only one radio light curve of a Dwarf Nova (SS Cyg),
we use the peak radio luminosity measured for that source in order to estimate
the maximum distance to which objects of this class can be detected at 0.1~mJy flux density limit. Following the steps described in Section~\ref{sec:method} we get d$_{\mathrm{max}}$~=~380~pc. \citet{PK2012} derive space density of non-magnetic
cataclysmic variables to be 4~$\times$~10$^{-6}$~pc$^{-3}$ with the scale height of 260~pc. The estimated
duty cycle of Dwarf Novae is 1-15~per~cent (Servillat et al. 2011) -- here we assume it is $\approx$~10~per~cent.
Applying the method discussed in Section~\ref{sec:method} we get the areal
density of 0.0013~deg$^{-2}$.\\
\noindent{\em Nova-likes}. Nova-like CVs were until recently not thought to be radio sources. However, \citet{kording2011} and \citet{coppejans2015} have now shown that several of these systems are radio sources at a level of less than about 0.2~mJy, with variability detected in two systems. Nova-likes will therefore show up as variable sources in sensitive radio surveys. We will not consider them here, however, because we do not expect to see many of them as transients. The reasons are their low space density (more than an order of magnitude less than DNe) and the fact that Nova-likes do not show frequent large amplitude variability.
\begin{table*}
\caption{Estimated areal densities of the studied classes of object at 0.1~mJy flux density
limit, together with the mean and standard deviation parameters calculated for each class and used to estimate their Gaussian distribution, for both rise and decline phases.}
\centering
\begin{tabular}{p{2cm}p{1.5cm}p{7cm}p{1cm}p{1cm}p{1cm}p{1cm}}
\hline
Class & Rate (deg$^{-2}$) & References & log$\,\mu_{\mathrm{R}}$ (day) & log$\,\sigma_{\mathrm{R}}$ (day)& log$\,\mu_{\mathrm{D}}$ (day) & log$\,\sigma_{\mathrm{D}}$ (day) \\
\hline
AGN & $2.3$ & \citet{FIRST} & 2.6 & 0.51 & 2.57 & 0.48 \\
TDE & $0.52$ & \citet{Frail2012}& 2.07 & 0.46 & 2.84 & 0.25\\
GRB afterglow & $0.052$ & \citet{Frail2012} & 1.11 & 0.17 & 1.94 & 0.46\\
SN & $0.21$ & \citet{Frail2012} & 1.47 & 0.54 & 2.43 & 0.68 \\
SGR & $10^{-8}$ & \citet{Olausen2014}, \citet{Ofek2007} & -- & -- & 0.68 & 0.43\\
XRB & $2.4 \times 10^{-4}$ & \citet{xrbrate} & 0.0044 & 0.73 & 0.07 & 0.69 \\
Dwarf Nova & $0.0013$ & \citet{PK2012}, \citet{Servillat2011} & -- & -- & 0.92 & 0.43\\
Classical nova & $0.0023$ & \citet{Roy2012} & 1.60 & 0.32 & 2.39 & 0.35\\
RSCVn & $0.011$ & \citet{Williams2013}, \citet{Favata1995}, \citet{Ottmann1992} & 0.45 & 0.47 & 0.46 & 0.45\\
Algol & $0.062$ & \citet{Duerbeck1984} & 0.031 & 0.51 & 0.15 & 0.50\\
Flare star & $0.0064$ & \citet{Reid2007}, \citet{Osten2008} & -1.85 & 0.67 & -1.86 & 0.43\\
Magnetic CVs: & & & & & & \\
Polars & 0.038 & \citet{polarsdc}, \citet{Pretorius2013} & -- & -- & -- & -- \\
IPs & 0.0008 & \citet{ipspacedens}, \citet{Pretorius2013} & -- & -- & -- & -- \\
Polars + IPs & 0.039 & -- & -1.81 & 0.23 & -1.93 & 0.069 \\
\hline
\end{tabular}
\label{rate}
\end{table*}
\subsubsection{RSCVn}
\label{rscvn}
The space density of RSCVns is approximately 6 $\times$ 10$^{-5}$ pc$^{-3}$ \citep{Favata1995}. The duty cycle of the class is estimated to be $\approx$ 10 per cent and the typical monochromatic radio luminosity of sources 10$^{16}$ erg s$^{-1}$ Hz$^{-1}$ \citep{Williams2013}. Converting space density into areal density as in Section~\ref{sec:method}, with the scale height of 325 pc for RSCVns \citep{Ottmann1992}, we get the sky density of 0.011 deg$^{-2}$.\\
\subsubsection{Algol binaries}
\label{algol}
We use a value of $\approx$~5$\times$10$^{-6}$~pc$^{-3}$ for the space density of algol binaries, as defined by \citet{Duerbeck1984}.
To calculate the maximum distance of detection at the 0.1~mJy flux density limit we use the average monochromatic luminosity
of objects in our sample (1.3$\times$10$^{17}$~erg~s$^{-1}$~Hz$^{-1}$). There are no good estimations of duty cycle for algol binaries --
based on the light-curves of sources in our sample we estimate it to be $\approx$~23~per~cent.
Correcting for the scale height of algols, which is $\approx$~400~pc \citep{Duerbeck1984}, we get the
final areal density of 0.062~deg$^{-2}$.
\subsubsection{Novae}
\label{cn}
The rate of Novae in the Milky Way is estimated to be $\approx$~35~year$^{-1}$ \citep{Roy2012}.
In order to calculate the maximum distance at which a nova could be detected, with a limiting flux density of 0.1~mJy, we assume that the typical radio luminosity it can reach in an outburst is $\approx$~1.2~$\times$~10$^{20}$erg~s$^{-1}$~Hz$^{-1}$, that is, the average peak radio luminosity of novae in the sample.
Because the derived distance is 30~kpc, we can estimate that approximately all 35 novae per year in our Galaxy could be
detected at the assumed flux limit. The final rate in a single snapshot is evaluated by multiplying a
rate per year by the duration of nova. Based on the properties of light curves in our sample,
we expect to detect a typical nova in the outburst for $\approx$~1000 days at 5-8~GHz frequency.
For the parameters listed above, we estimate that the areal density of novae is 0.0023~deg$^{-2}$.
\subsubsection{Soft Gamma-ray repeaters}
\label{sgr}
\noindent The sky density of SGRs is not well constrained. \citet{Ofek2007} gives the upper limit for the rate of giant flares (similar to SGR~1806-20 event) of 5~$\times$~10$^{-4}$~year$^{-1}$ per SGR.
With the number of SGR sources in the Galaxy reported by \citet{Olausen2014} of 15\footnote{\url{http://www.physics.mcgill.ca/~pulsar/magnetar/main.html}}, and the duration of the radio flare analysed in our sample ($\sim$~20~days) we get the areal density of approximately 10$^{-8}$~deg$^{-2}$.\\
\subsubsection{Magnetic CVs}
\noindent {\em Polars}. One of the two magnetic CVs in our sample -- V834 Cen -- belongs to the subclass of polar magnetic CVs.
With the luminosity measured in the analysis (2.4 $\times$ 10$^{17}$ erg s$^{-1}$ Hz$^{-1}$), we estimate the maximum distance of $d_{\mathrm{max}}$~=~1.4~kpc.
For the space density of 10$^{\mathrm{-6}}$~pc$^{-3}$ \citep{Pretorius2013}, duty cycle of 0.5~\citep{polarsdc} and the scale height equal to 260~pc \citep{Pretorius2013}, we estimate the sky density for these types of objects of 0.038~deg$^{-2}$.\\
\noindent {\em Intermediate polars}. Sky density of IPs is calculated based on the AE Aqr.
With the average measured luminosity of that source (5.8~$\times$~10$^{16}$~erg~s$^{-1}$~Hz$^{-1}$), the estimated maximum distance for the detection is
d$_{\mathrm{max}}$~=~700~pc.
With space density of 10$^{-7}$~pc$^{-3}$ \citep{ipspacedens}, scale height 120~pc \citep{Pretorius2013} and duty cycle of 1 (persistent sources) we get the estimated sky density of 0.0008~deg$^{-2}$.
\subsubsection{Flare Stars}
\label{fs}
Space density of flare stars given by \citep{Reid2007} is 0.08~pc$^{-3}$. The average luminosity of flare stars in our sample is 4.7$\times$10$^{13}$~erg~s$^{-1}$~Hz$^{-1}$.
With such low luminosities, at the 0.1~mJy flux limited survey they can be detected up to a distance of $d_{\mathrm{max}}~\approx$~20~pc. Within that distance, which is relatively small compared
to the size of the Galaxy, we can assume a uniform, spherical distribution of sources. The duty cycle of flare stars is not well constrained (\citealt{Osten2008}, \citealt{Hilton2010}) -- assuming the upper limit of 10~per~cent, we get the areal density of 0.0064~deg$^{-2}$.
\noindent Flare stars can also produce bright coherent bursts, which at the sub-milijansky detection threshold can be detected from a distance of several hundred parsecs \citep{Osten2008}.
Although these types of events are quite rare (duty cycle less than 1 per cent), because of their high brightness, the areal density of coherent flares can reach approximately 1.2~deg$^{-2}$ at the 0.1~mJy flux density limit. On the other hand, they are expected to evolve rapidly ($\sim$~60~s) and should only be common in surveys exploring very short time-scales.\\
\section{Probability distribution of variability time-scales}
\label{results}
For a transient/variable candidate discovered in a blind radio survey, it is very unlikely that the
information about its distance, and therefore the luminosity, will be known (at least immediately).
Thus, all of the earliest information will be contained in the variability of the
light-curve and spectral distribution of the source. Correspondingly, if we can not
take into account information about the luminosity of the source, the relation shown in
Figure~\ref{LT} is reduced to a one dimensional histogram of variability time-scales (Figure~\ref{TH}),
increasing the uncertainty in separation between different classes of objects. This result, however, can be
converted into the probability distribution and convolved with the expected areal densities
of radio sources.
In order to represent the variability time-scales as a probability distribution,
for each of the considered classes of object we have calculated the mean value~$\mu$
and the standard deviation~$\sigma$ (listed in Table~\ref{rate}) of the variability time-scales~$\tau$.
Using these values we have assigned the following
Gaussian probability distribution to each class:
\begin{equation}
P(\tau) = \frac{1}{\sqrt{2 \pi } \sigma} e^{-\frac{(\tau - \mu)^{2}}{2 \sigma^{2}}}.
\label{rawprob}
\end{equation}
\begin{figure*}
\includegraphics[scale=0.5]{Fig10a.pdf}
\vspace{0.6cm}\\
\includegraphics[scale=0.5]{Fig10b.pdf}
\caption{ {\em Upper:} Relation shown in Figure~\ref{LT} converted into the probability distribution of source belonging to given class of objects based on the rise time-scale, normalised to 100 per cent. {\em Lower:} The same relation, convolved with the estimated sky densities of objects (Table~\ref{rate}). These estimated distributions have been calculated with the flux density limit of 0.1~mJy.}
\label{RP}
\end{figure*}
\noindent The upper panel in Figure~\ref{RP} shows the probability distribution of all the classes of objects in our sample,
normalized to 100~per~cent. The dashed line represents the distribution of TDEs, for which only one flare with well sampled rise time was available. In this case, the mean value of the
distribution has been approximated by the variability time-scale measured for that source,
while the standard deviation has been calculated as the mean value of the
standard deviations measured for all the other classes of objects.
The probability distribution obtained in Equation~\ref{rawprob} depends strongly on biases in our sample.
In order to convert it into the actual probability distribution, which takes into account expected number of sources in the sky,
we multiply the distribution obtained in Equation~\ref{rawprob} by the areal density of objects $R$, estimated in Section~\ref{se}
\begin{equation}
P_{\mathrm{corr}}(\tau) = R \times P(\tau).
\end{equation}
Given a small number of light-curves representing each subclass of Magnetic CVs, we add
their sky densities and calculate their combined probability distribution instead of considering them separately.
With this approach, we avoid adding the uncertainties associated with assigning Gaussian distribution based on approximated parameters.
The lower panel in Figure~\ref{RP} shows the probability distribution of all the types of objects in our sample, corrected
for the estimated sky rates of different classes. It is worth noting, that sources like XRBs or Novae, even though they make up a significant fraction of the original sample, due to
their low expected sky densities, are very unlikely to be found in a blind radio
transient survey.
The lower panel in Figure \ref{RP} provides a template of the variability time-scales which could be used as an indicator of the newly discovered sources class.
At the moment, the number of available flares used in designing this classification
method is uneven across a range of objects included. However, it can be gradually
improved with new sources detected in future surveys, where additional data could help to limit the uncertainties of the initial classification.
This would be especially desirable in the case of classes which have few well-studied light-curves to date, such as tidal disruption events, for which at this point, the variability time-scales are not well constrained.
Tables~\ref{probR} and \ref{probD} give the probability information for each of the classes of objects
(for rise and decline phases respectively) on a time-scale $\approx$~10$^{-4}$--10$^{4}$~days, with a logarithmic time-step of~0.5.
The presented results depend on a range of parameters, which we can not accurately account for in our analysis. Although we have corrected for the expected areal densities of objects down to 0.1~mJy, for higher sensitivity surveys intrinsically faint sources will become progressively more common, with the number of Galactic sources increasing up to the point where the entire population within the Galaxy can be detected.
The form of the probability distributions which can be applied to a given survey will also depend on the observing frequency, which might favour detection of sources showing certain variability time-scales, and -- for Galactic populations -- pointing location. At the moment, these have to be considered separately within a given survey, but, in the future might be included as part of a more advanced pipeline.
\section{Discussion}
\label{discussion}
Figure~\ref{TH} shows that creating a simple automatic flare finder allowed us to include a larger number of flares
in the analysis, compared with the manual selection.
Although there is an overall agreement between both methods,
comparison of results obtained
automatically (Figure~\ref{LT}) and manually (Figure~3, PFK15) shows a higher
degree of scatter for several classes of objects in the automatic approach.
Light-curves corresponding to these sources (XRB, RSCVn, Algol) consist of tens of single outbursts,
which in the manual analysis we have limited to those with the best data quality.
The flare finding software, however, had identified all of the flaring
events in the same light-curves, increasing the range of obtained time-scales for
each source. Although most of the measurements correspond to real flares,
in marginal cases a false detection has also been included.
For example, an odd noisy data point within a flare can be mistaken for the end
of the respective flare and result in underestimating measured decline time. In other extreme cases,
several outbursts superimposed such that the flux level does not drop low enough
for the AFF to separate them, will be measured as one long flare rather than single events,
and make the rise/decline times appear longer than they really are.
Because several of those exceptions identified in our analysis could not be
removed other than by manually excluding them, we have decided to leave them in the final
results in order to keep the process purely automatic. In this way we also begin to learn what pitfalls might befall a future automated system.
\begin{figure}
\includegraphics[scale=0.2465]{Fig11.eps}
\caption{Variation of Figure \ref{LT}, where all events originating from synchrotron emission are marked with blue circles, and additional measurements corresponding to scintillating sources and extreme scattering events (ESEs) are plotted in purple and red respectively. As discussed in Section \ref{discussion}, scintillating sources together with extreme scattering events appear to populate most of the time-scale parameter space and cover the range of rise rates typical for all the classes, from low luminosity stars to AGN. Scintillation and ESEs have not been included in the fit.}
\label{mainsct}
\end{figure}
This work is exploring the parameter space typical for flares originating from incoherent emission processes.
Variability from sources of coherent emission such as pulsars is usually detected in high-time resolution observations, and, with duration of milliseconds--seconds, is outside of the time-scales regime considered in this analysis.
However, \citet{coppejans2015} showed that a flare from nova-like source TT Ari, evolving on time-scale of minutes, is most likely produced by a coherent emission mechanism. The characteristic exponential rise time of that flare ($\approx$~0.005~day) shows that in some cases, coherent flares can overlap the time-scales of incoherent events.
One of the main challenges we might come across while
classifying a newly discovered source by its variability rates can be scintillation.
Figure~\ref{mainsct} shows the relation between the time-scales and peak radio luminosities for our extended sample of objects, where a sample of scintillating sources and extreme scattering events are plotted along with the previously shown (Figure~\ref{LT}) synchrotron events. It shows that the parameter space of rise rates for these types of
sources overlaps strongly with most of the analysed astrophysical flares. It might be
possible to distinguish between those based on their spectral properties --
while the amplitude of synchrotron flares is lower at lower frequencies, \citet{sctexp} show that the observed variability due to interstellar scintillation
can be stronger at lower frequencies.
Similarly, if originating from intrinsic variability, the light-curve should evolve on longer time-scales at lower frequencies, which is not the case for ESEs, where the peak flux density is observed simultaneously at all wavelengths \citep{ESE1}.
Another way of separating discussed types of events lies in the shape of the flare -- those originating
from astrophysical explosions are usually characterised by longer decaying times
comparing to their rise times, while in the scintillation both phases are more symmetrical.
Although the entire process aims at automating the rise/decline rate measurements, several steps have been done `by eye' (listed in rectangular frames in Figure~\ref{chart}). Those include splitting the complete sample of light-curves into four distinct groups (each of which required different method of estimating the quiescent flux level) and applying smoothing for one of them. The remaining steps (listed in rounded frames in Figure~\ref{chart}), such as selecting flaring events from the data and measuring rise/decline rates (including background subtraction) have been done automatically.
It should be noted that all the manual adjustments were only required due to the nature of the compiled sample.
The flare finding routine applied to the dataset was necessarily simple, however, even this unsophisticated thresholding approach needed to be somewhat flexible in order to work on a variety of available light-curves.
In the future transients/variables surveys the observations carried out with an individual telescope will provide more consistent datasets, and the flare identification method can be replaced by more advanced techniques, such as those discussed in Section~\ref{flarefinding}.
Additionally, as the aim of presented classification method is to give an earliest indication of the sources nature, a potential flare should be identified as soon as the variability above the quiescent flux level arises. With the long term monitoring of the background prior to outburst, it is possible that an iteration of a presented flare finding routine could be used as part of a transients/variables pipeline, however, the exact form of the software will depend on the surveys design.
At the moment the analysis is based on a fairly limited sample of radio flares and an incomplete representation of certain types of sources.
Including more data could restrict the individual distributions and reduce the resulting overlap, providing clearer classification.
Alternatively, it can not be ruled out that extending the number of flares will lead to even greater scatter of measurements within respective classes, decreasing the separation of their probability distributions.
At this point, we can not definitively say which of these scenarios is more likely. Gradually updating presented result with new sources might help to provide a more conclusive distribution.
\section{Conclusions}
\label{conclusions}
This work builds upon the previously reported result, which showed a clear correlation
between the luminosity and time-scale of variable radio objects, ranging from nearby,
intrinsically faint to distant and luminous sources (PFK15).
Here we proposed a method of classifying radio transient and variable sources using the rise and/or decline rates of their flares.
It is based on the analysis of existing light-curves of synchrotron events, where the individual flares were selected automatically in order to reduce bias associated with choosing data by hand and make the result easy to reproduce.
We convolved the measurements of rise and decline time-scales with the expected areal densities of sources down to a 0.1~mJy flux density limit, providing probability distributions of variability time-scales for a wide range of classes.
We have shown that comparison of the variability time-scales measured for a transient/variable source with the presented distribution could point -- with a given probability -- to its nature.
The result shows that the variability-based classification can be used -- to some extent -- to differentiate between classes of objects.
We have also investigated time-scales of variability caused by propagation effects such as scintillation and extreme scattering events. We showed that these time-scales partly overlap with the parameter space of synchrotron events and might require further analysis to separate from astrophysical flares.
In the future, this method or a derivative could be incorporated into automated transient detection
pipeline \citep[e.g. ][]{trap} and used as part of a more complex detection/classification software by setting initial constraints
on the class of the newly discovered source.
The number of transient and variable sources found in future SKA surveys could reach up to 1000 per week \citep{Burlon2015, Fender2015, Perez-Torres2015}, hence, the initial, time-scale based classification might be valuable in the process of selecting interesting events.
This early indication could then be further verified and narrowed down
by including additional information such as an optical flux measurement (Stewart et al. {\em submitted}), location in the sky or comparison with the archival data.
\section*{Acknowledgements}
We thank the referee for helpful comments and suggestions which helped to improve this paper.
This project was funded by European Research Council Advanced Grant 267697
"4 PI SKY: Extreme Astrophysics with Revolutionary Radio Telescopes".
The Green Bank Interferometer is a facility of the National Science Foundation
operated by the NRAO in support of NASA High Energy Astrophysics programs.
This research has made use of data from the MOJAVE database that is maintained
by the MOJAVE team (Lister et al., 2009, AJ, 137, 3718). This research has made use
of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion
Laboratory, California Institute of Technology, under contract with the National Aeronautics
and Space Administration.
|
\section{Introduction}
Galactic dust obscures a large fraction of the sky, called the Zone of Avoidance \citep[ZOA,][]{1961gala.book.....S}. This region, apparently devoid of galaxies, used to be
avoided by observers because of the inherent difficulties, resulting from the strong dust extinction, in analyzing the galaxies there. However, to understand local flows and the Cosmic Microwave Background \citep[CMB,][]{1965ApJ...142..419P} dipole \citep[e.g.][]{1971Natur.231..516H,1976BAAS....8Q.351C,1977PhRvL..39..898S,2014A&A...571A..27P}, coverage of the whole sky is essential. Numerous studies in the past have indeed shown the relation between surrounding structures and the CMB dipole \citep{1991AIPC..222..421L,2006ApJ...645.1043K,2006MNRAS.368.1515E,2011ApJ...741...31B,2012MNRAS.427.1994G}. The lack of data in the ZOA (which covers about 10--15\% of the sky depending on the wavelength of observations) means that astronomers must rise to this challenge.
Many local large-scale structures are indeed bisected by the Galactic Plane. Two of them, Perseus-Pisces and the Great Attractor, actually lie across the ZOA on exactly opposite sides of the Local Group \citep[e.g.][]{1994AJ....108..921P,1994A&A...286...17S,1997AJ....113..905P}. Because they are partially obscured, determining the winner in this tug-of-war on the Local Group constitutes a daunting task \citep[see e.g.][]{1992AGAb....7..177S,1994ASPC...67...81S}. Similarly, identifying the connections between structures lying across the ZOA, if they exist,
is essential to measure the longest structures and verify their compatibility both in terms of size and abundance with predictions from the standard cosmological model or from the observations of fluctuations in the CMB \citep{2000A&ARv..10..211K,2011MNRAS.417.2938S}. At present, most CMB dipole estimates assume a homogeneously filled ZOA or an extrapolation of structures on either side of the ZOA, which prevents them from achieving high precision measurements. Thus there exists a real need to map the structures across the ZOA \citep[see][for a complete review on the subject]{2000A&ARv..10..211K}.
A great deal of effort has already been deployed to observe galaxies within the ZOA \citep[e.g.][for a non-extensive list with the last two references as complete reviews]{1994ASPC...67...99K, 2014MNRAS.443...41W,2016AJ....151...52S,2006MNRAS.369.1741D,2016MNRAS.460..923R, 2016MNRAS.457.2366S,2000A&ARv..10..211K,2005RvMA...18...48K} or at least to reduce its extent by using observations at less-affected wavelengths \citep[e.g.][]{2012AJ....144..133S, 2014MNRAS.444..527S,2014ApJ...792..129N}. A complementary approach is to use observations
surrounding the ZOA to reconstruct the missing information in it \citep{1994ASPC...67..185H,2006MNRAS.373...45E}. In this paper, we propose a method for
quantifying the probability of the presence of structures in different areas of the ZOA.
The study is conducted within the framework of
constrained simulations of the local Universe \citep[e.g.][]{2001ApJS..137....1B,2010arXiv1005.2687G}.
Unlike ordinary cosmological simulations, these simulations stem from initial conditions constrained by observational data to resemble the
local Universe. These observational data can be either radial peculiar velocities of galaxies \citep[]{2002ApJ...571..563K,2003ApJ...596...19K, 2014MNRAS.437.3586S} or redshift catalogs \citep[]{2010MNRAS.406.1007L,2013MNRAS.435.2065H}. The initial conditions are retrieved either backwards \citep{1987ApJ...323L.103B,1991ApJ...380L...5H,1992ApJ...384..448H,1993ApJ...415L...5G,1996MNRAS.281...84V,1998ApJ...492..439B,2008MNRAS.383.1292L} or forwards \citep{2013MNRAS.429L..84K,2013MNRAS.435.2065H,2013MNRAS.432..894J, 2013ApJ...772...63W}. In the latter case, the initial density field is sampled from a probability distribution function consisting of a Gaussian prior and a likelihood. A complete overview of the different methods is given in \cite{2014ApJ...794...94W} and an extension of the backwards method to non-Gaussian fields is proposed in \citet{1995MNRAS.277..933S}.
Recently, \citet{2016MNRAS.455.2078S} released simulations constrained with the second radial peculiar velocity catalog \citep{2013AJ....146...86T} of the Cosmic Flows project \citep[e.g.][]{2012ApJ...749...78T,2012ApJ...749..174C} using a backwards method. Although estimating peculiar velocities constitutes an observational challenge, there is a double advantage in using them: first they are highly linear and correlated on larger scales than the density field. Consequently, the Zone of Avoidance can still be recovered when it becomes larger than the correlation length of the density field with a sufficiently sampled velocity field; second, peculiar velocities are excellent tracers of the underlying gravitational field as they account for both the baryonic and the dark matter. The backwards method permits producing a large number of realizations in a short amount of (computing) time which is paramount for such statistical studies. Running the simulations from the large set of initial conditions produced from the method is the most demanding of computer time.
Actually, as shown in Section~2, running such constrained simulations is not an absolute necessity when the goal is to study the smoothed high density peaks of the large-scale structure at redshift zero. There is thus no reason either to use forwards methods with high computational demands to get the reconstruction of the local Universe with the non-linearities. Basically, studying the linear reconstruction of the local Universe obtained with a backwards method\citep[e.g.][]{1995ApJ...449..446Z,1999ApJ...520..413Z}, or more precisely the constrained realizations of the latter \citep{1987ApJ...323L.103B,1992ApJ...384..448H,1996MNRAS.281...84V}, to actually have access to structures in the ZOA up to large distances as well as to their probability, is sufficient. Indeed, the reconstruction technique by definition goes to the null field (the overdensity $\delta$ equals 0) in the absence of data or in presence of very noisy data while, by extension, the constrained realizations technique statistically restores the missing structures. Namely, while the Wiener Filter reconstruction presents smoother and smoother structures up to reaching the null field with increasing distances from the center of the dataset, the constrained realizations compensate this smoothing by combining the constraints with a random field thus allowing us to fully access structures even beyond the volume sampled by the dataset. Detailed equations are given in Appendix A. The goal of this paper is to show that this restoration is not statistically random and depends on the surrounding structures up to large distances. Hence, it is possible to predict where the structures are more likely to be in the ZOA, as shown in Section~3. In Section~4, theoretical predictions are compared with observations, and the conclusions of this work are presented in Section~5.
\section{Methodology}
\subsection{Preparing Constrained Realizations}
The scheme used by the CLUES project to build constrained initial conditions and run the simulations is described in detail in
\citet{2016MNRAS.455.2078S}. Since only some parts of the process are of interest here, the steps to produce realizations constrained by
observational radial peculiar velocity catalogs and a brief description of their purpose are the only aspects recapitulated here:
\begin{enumerate}
\item grouping of galaxies in the radial peculiar velocity catalog to remove non-linear virial motions
\citep[e.g.][]{2015AJ....149...54T,2015AJ....149..171T} that would cause problems when using the linear reconstruction method;
\item minimization of the biases \citep{2015MNRAS.450.2644S} in the grouped radial peculiar velocity catalog; and
\item production of linear density and velocity fields constrained by the grouped and bias-minimized observational peculiar velocity catalog combined with a random realization to statistically restore the missing structures using the Constrained Realization technique \citep[CR,][]{1987ApJ...323L.103B,1991ApJ...380L...5H,1992ApJ...384..448H} .
\end{enumerate}
The Planck cosmology framework \citep[$\Omega_m$=0.307, $\Omega_\Lambda$=0.693, H$_0$=67.77, $\sigma_8$~=~0.829,][]{2014A&A...571A..16P} is used as a prior.
\subsection{Constrained Realizations versus Constrained Simulations}
Running constrained simulations is useful to study the formation and evolution of the large-scale structures as well as the small non-linear
scales. However, when interested only in the large-scale structure present today, there is no need to simulate the full formation history
of these structures---the linear reconstruction is sufficient. This reconstruction can be obtained with the Wiener Filter technique
\citep[linear minimum variance estimator, abbreviated as WF;][]{1995ApJ...449..446Z,1999ApJ...520..413Z}. Where data are
available, the WF-reconstructed field reproduces the observed large scale structure \citep[e.g.][]{1999ApJ...520..413Z,2012ApJ...744...43C}. The WF method
is also able to reconstruct across the ZOA provided that the latter does not cover too large a volume \citep[i.e. nearby, at low
redshift;][]{1994ASPC...67..185H,2000ASPC..218..173Z}. The further from the observer, the worse the extrapolation is, and in fact structures are
not reconstructed in the ZOA at large distances---the WF simply goes to the null field. By contrast, the CR technique restores statistically structures in that zone by adding a random field.
The goal of this paper is to determine if there are some regions in the ZOA that have a higher probability of hosting a structure than
others. To pursue this work, statistical analyses are required. This constitutes another argument in favor of studying the structures
directly in the fields obtained with the CR technique, since running lots of constrained and random (for statistical comparison) simulations
consumes a lot of computer time. One might argue that CR fields are not suitable substitutes for full simulated fields at the level of interest
for this paper, and of course CR fields present only the linear structures and obviously are in no way exact substitutes for the fully
simulated fields. However, as long as CR fields and simulated fields present a very similar large-scale structure for high-density peaks in
the ZOA, one can be used for the other in statistical tests.
\begin{figure}
\centering
\hspace{-1.54cm}\includegraphics[width=0.56\textwidth]{dcosmicvarianceZOAALLSimuCRcolor.eps}
\vspace{-1cm}
\caption{Mean (points) and scatter (error bars) of the standard deviations ($\sigma$) obtained from cell-to-cell comparisons carried out on pairs of overdensity fields smoothed on a 5{ \textit{h}$^{-1}$~Mpc}\ scale at
$z$=0 for the whole 500{ \textit{h}$^{-1}$~Mpc}\ box (top), for the 320{ \textit{h}$^{-1}$~Mpc}\ sub-box (middle), and for the Zone of Avoidance (bottom). The mean standard deviations are given as a function of the overdensity threshold. Pairs consist of two constrained simulations (filled blue square), one constrained simulation and one constrained realization (filled black
circle) and one constrained simulation and one constrained realization sharing the same random part (filled red triangle). }
\label{fig:CRvsSimu}
\end{figure}
To demonstrate that CR fields can be taken as substitutes for simulated fields for our purpose, 15 simulations from \citet{2016MNRAS.455.2078S}
and the 15 corresponding CR fields (so that simulations and CRs share the same random part) are compared. We apply a cloud-in-cell scheme on a
256$^3$ grid to the particle distributions of the 15 simulations at $z$=0 with a subsequent Gaussian smoothing on a scale of 5{ \textit{h}$^{-1}$~Mpc}. The
resulting smoothed density field of any constrained simulation is then converted to the overdensity field. The CR technique indeed only
produces the overdensity field. Each simulated overdensity field is then compared to that of another constrained simulation, to that of a
constrained realization, and to that of the constrained realization sharing the same random field. For each pair of each category
(simulation/simulation; simulation/CR; simulation/CR sharing the same random field), we build an overdensity-overdensity plot (i.e.\ cells of
overdensity field \#1 versus cells of overdensity field \#2). If the two fields were identical, all points would lie on the 1:1 relation (cell
value in field \#1 is equal to cell value in field \#2). The difference between two fields is determined as the one-sigma, hereafter 1$\sigma$,
scatter (or standard deviation) around this 1:1 relation. We repeat this procedure for all the pairs of each category and then the mean and the
variance of the 1$\sigma$ scatters are derived.
In the top panel of Fig.~\ref{fig:CRvsSimu}, the result is three points (blue, black and red symbols for each of the
categories simulation/simulation, simulation/CR, simulation/CR sharing the same random field), with error bars, computed only for cells with positive overdensities.
The procedure is repeated in smaller sub-boxes of size 320~{ \textit{h}$^{-1}$~Mpc}, where most of the observational constraints are, and in the ZOA (i.e.\ cells are
compared only in this area: the ZOA is taken to be a 20$^\circ$ cone with the apex at the center of the box and oriented in the same direction in the
realizations as in the observations. The resulting points are shown in the middle and bottom panels of Fig.~\ref{fig:CRvsSimu}. We repeat the process,
changing the threshold to select higher and higher overdensity regions.
Figure~\ref{fig:CRvsSimu} shows that the 1$\sigma$ scatters increase with the threshold in accord with the expectation that, although the
large-scale structure is reproduced, high-density peaks might be shifted by a few megaparsecs from one realization to another, and from the
simulation to the CR. In addition, because simulations also probe the non-linear regime, and hence allow larger fluctuations, the 1$\sigma$
scatters are higher when comparing two simulations than when comparing a simulation and a CR. That is, the chance of getting a large contrast
between cells coming from two different simulations is higher than for cells coming from a simulation and a CR; and indeed in the latter,
linear-regime, case the fluctuations are reduced.
The main result from Fig.~\ref{fig:CRvsSimu} is that the differences between two simulations, or between a constrained realization and a
simulation built with different random fields, are larger on average than the difference between the simulation and the constrained
realization sharing the same seed (at least for the high density peaks of interest in this paper). Admittedly, the differences are more
pronounced when considering the full boxes rather than the smaller boxes or the ZOA. In the first case, it is expected because of the decrease in
cosmic variance when working with constrained simulations compared to working with random simulations \citep[e.g.][]{2016MNRAS.455.2078S}. The inner parts of the boxes are
more constrained (have more constraints) and thus they resemble more to the local Universe as well as to each other in the different
simulations (hence the smaller 1$\sigma$ scatters than for the full boxes). This is also expected in the second case because the ZOA is not as
well constrained as the other parts of the box given the scarcity of the data, especially at increasingly higher distances. The difficulty of reconstructing across the ZOA increases linearly with the distance from the observer. In addition, the error bars of the 1$\sigma$ scatter increase with the reduction of the number of cells compared between two realizations,
(i.e. with the reduction of the compared zones).
In that context, the scatters between the simulation and the CR sharing the same random field are even lower than between two
simulations, or between a CR and a simulation with different random fields. This suggests that for our purpose (a statistical study of high-overdensity peaks in the ZOA)
one can be taken as a substitute for the other. Taking CR fields as substitutes for the simulated
fields to derive statistics on the large-scale structure in the ZOA, we produce 200 CR fields on 256$^3$ grids as well as 200 random fields
(sharing the same random part as the corresponding CR field) on the same grid size at a small computational cost.
\section{Probability of Structures}
\subsection{Random vs Constrained Probabilities}
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{probhistmoreCRv2.eps}\\
\vspace{-.2cm}
\caption{Histograms of the number of cells with an overdensity above the threshold given in the top right corner of each panel (1.0, 1.5, or 2.0) as a function of the number of realizations. The dotted black line and solid black line represent, respectively, the probability distribution functions of cells in, respectively, random and
constrained realisations having an overdensity value above the threshold. The bottom panels are zooms of the positive tails of the
distributions. The red dashed line shows the distribution of the residual between the constrained and the random probabilities for a
given cell. The blue dot-dashed lines are the distributions if cells are only counted if they exhibit an overdensity value above the
threshold {\em only} in the CR and {\em not} in the RR sharing the same random seed. In this case a cell is counted only when the
overdensity is entirely due to the constraints: the random part alone would not have induced an overdensity above the given threshold. See
the text for a more detailed explanation.}
\label{fig:hist}
\end{figure*}
The first goal is to quantify the constraining power of the CR technique in the ZOA. That is, at a given location in the ZOA, do the CR fields
present a structure with a greater probability than the random fields, or are the structures simply randomly distributed in a manner consistent
with the cosmological model? With 200 constrained realizations and 200 random realizations, it is straightforward to study the probability that
a cell is overdense (i.e. hosting a structure) in both the random and constrained cases.
To determine these probabilities in the ZOA, cells belonging to the ZOA need first to be identified. Since we smooth the realizations over
5{ \textit{h}$^{-1}$~Mpc}, cells of 5{ \textit{h}$^{-1}$~Mpc}\ filling up the entire ZOA constitute the optimal choice. We split the ZOA into cells of 5{ \textit{h}$^{-1}$~Mpc}\ and compute their
overdensity value. We count the number of times a cell belonging to the ZOA has a value above a given threshold over 200 CRs. For clarity and
simplicity, the paraphrase `threshold-pass cell' is used to mean hereafter `a cell with a value above a given overdensity
threshold'. Finally, looping over all the ZOA cells, the number of cells that are threshold-pass in a certain number of CRs is obtained and the
probability distribution function of a cell to be threshold-pass is calculated. The same procedure is repeated with the random realizations
(RRs) to obtain the random probability distribution function of a cell to be threshold-pass.
If the constraints were not affecting the ZOA at all, i.e. if constraining the realizations was not giving some insight into the ZOA structures, then the constrained and random probability distribution
functions should be statistically identical. This would imply that the structures in the ZOA are completely randomly distributed and abide only by the prior cosmological model in both the random and the constrained cases. Figure~\ref{fig:hist} shows that this is not the case: the histograms of probabilities or the distributions of threshold-pass cells
(one threshold per panel in the top row) as a function of the number of realizations in which they are found differ between the constrained
(solid black line) and random (dotted black line) realizations. For instance, the top left panel of Fig.~\ref{fig:hist} shows that the
random probability for a cell to be threshold-pass at the 1.0 overdensity level follows a Gaussian distribution with a mean at 15\%
(30 out of 200 realizations, dotted black line). Thus there is on average about a 15$\pm$3\% chance of finding an overdensity greater than
1.0 in a cell. The constrained probability however does not follow a Gaussian distribution, the solid black line distribution is
skewed (skewness 1.5) and flatter (kurtosis 11), confirming that constraining the realization also affects the ZOA. The middle and right
panels of the top row on Fig. \ref{fig:hist} confirm that such observations are valid also for higher overdensity thresholds.
From these different probability distribution functions, it becomes clear that cells with probabilities in the high tail of the distribution
for constrained realizations, most likely, host a structure that is produced by the constraints surrounding the ZOA. The bottom row of
Fig.~\ref{fig:hist} zooms in on the high tail of the probability distribution functions calculated at a given threshold to highlight the existence of such cells.
To reinforce our claims, we pursue our quest for statistical differences between CRs and RRs in the ZOA by deriving two more probability distributions that are hereafter called the global probability and the individual probability:
\begin{itemize}
\item The global probability for a given cell is defined as the difference between its constrained and random probabilities. At a fixed overdensity threshold, the number of RRs in which a given cell is
threshold-pass is subtracted from the number of CRs in which this cell is also threshold-pass. The procedure is repeated over all the different cells in the ZOA to derive the global probability
distribution function.
\item The individual probability for a given cell is a bit more complex. Considering one CR, a given cell is counted as threshold-pass only in the case that in the RR sharing the same random seed as the CR, this same cell is {\em not} threshold-pass---i.e. the cell is threshold-pass in the CR only. In other words, we require that the constraints force the cell to become threshold-pass and the random component used in the CR technique to restore the missing structures would not have induced an overdensity greater than the threshold in that cell. The process is then repeated over the 200 realizations and then over all the cells filling the ZOA to derive the individual probability distribution function.
\end{itemize}
The global and individual probability distribution functions are plotted in Fig.~\ref{fig:hist} as the red dashed and blue dot-dashed lines
respectively. These curves confirm the statistical difference between CRs and RRs in the ZOA. Whatever the threshold level is, the global
probability distribution function is a Gaussian centered on zero with (for instance) a standard deviation of 8\% for a 1.0 overdensity
threshold. This distribution implies that some constrained cells have on average a greater (or lower) probability to be threshold-pass than their random
counterparts. This indicates that structures (both over- and under-densities) are more likely to exist in the ZOA because of the constraints than can be explained as solely due to the prior
cosmological model and its associated statistical fluctuations. It confirms that the random and constrained probability distribution functions differ and that the probability of each cell
to be threshold-pass is different in the random and constrained cases. The individual probability distribution function goes further since the random seed is also fixed. With a mean
at 13\% and a standard deviation of 8\% for a 1.0 overdensity threshold, this distribution is even more remarkable: some cells are clearly
threshold-pass solely because of the constraints: the random realizations have a very low probability of generating overdensities in
the cells that are in the high tail of the individual probability distribution function on a 1-to-1 random seed basis.
Since there is a clearly demonstrated difference between threshold-pass cells in CRs and RRs, we can proceed to pursue the analysis of
structures in the constrained field to locate the most probable structures in the ZOA.
\subsection{Locations of Constrained Structures}
\begin{figure*}
\vspace{-0.9cm}
\centering
\hspace{-1.5cm} \includegraphics[width=0.55\textwidth]{cf2gvpecc1pt5_CR_256_500_stackedZ_SGXYZud.eps}
\hspace{-0.58cm}
\includegraphics[width=0.55\textwidth]{random_256_500_stackedZ_SGXYZud.eps}\\
\vspace{-0.5cm}
\hspace{-0.58cm} \hspace{-1.5cm} \includegraphics[width=0.55\textwidth]{standdev200crZ_SGXYZ.eps}
\includegraphics[width=0.55\textwidth]{probmultiplot1_1pt0_00CR.eps}
\caption{Top panels: Averaged overdensity fields (black solid contours) of 200 constrained (left) and 200 random (right) realization fields in the 2{ \textit{h}$^{-1}$~Mpc}\ thick X-Y, Y-Z and X-Z supergalactic slices centered on the observer. The green contours show the transition between overdensities and underdensities ($\delta$ = 0), the dashed contours stand for the underdensity and the red lines delimit the Zone of Avoidance in the different slices,
corresponding to $(l,b)=(-10^\circ,+10^\circ)$ in Galactic coordinates. The name of a few local structures is given in blue on top of the averaged constrained field (GW stands for `Great Wall' and PP for `Perseus-Pisces'). Bottom panels: Left: Standard deviation of the 200 constrained overdensity fields (black solid contours). Labels give the value of the standard deviation. ; Right: Dots stand for 1.0-overdensity threshold-pass regions with a 3$\sigma$ signal; see the text for a
more detailed explanation. Dot sizes are proportional to the probability of cells to be threshold-pass. Their location in the third dimension (Z, X and Y directions from left to right, top to
bottom respectively) is indicated by the color scale.}
\label{fig:space}
\end{figure*}
In the rest of this section, we focus on cells that have a threshold-pass probability with a 3$\sigma$ level of significance. For
instance, for the 1.0-overdensity threshold, the probability that the cell is threshold-pass must be at least 24\% (=15\% + 3$\times$3\%,
i.e.\ the mean of the random probability distribution function plus three times its standard deviation) for the cell to be considered. The
probability distribution being quasi-Gaussian, a positive 3$\sigma$ level is an appropriate choice for selecting the statistically significant cells in the positive tail of the distribution and removing
the others. To locate these cells (or potential structures) in space, the top left panel of Fig.~\ref{fig:space} shows as black contours the
average overdensity field of the 200 CRs in three 2{ \textit{h}$^{-1}$~Mpc}\ thick supergalactic slices centered on the observer. Keeping the same contour
values, the same plot is shown for the average overdensity field of the 200 RRs on the right side. The red lines identify the ZOA in the different supergalactic
slices. While the average of the 200 RRs presents a
quasi-null field (green contours) everywhere, there are, as expected from our statistical study of simulations, clearly distinct structures
in the average of the CRs, including in the ZOA that is of interest here. Such an observation means that similar structures are repeatedly at the same
location in the CRs while the locations of similar structures differ substantially from one RR to another. Consequently, while they annihilate each other when averaging in the random case, they reinforce each other in the constrained case. This claim is further supported by the standard deviation of the 200 constrained realizations shown in the bottom left panel of Fig.~\ref{fig:space}: the constrained realization technique decreases the cosmic variance between the different constrained overdensity fields with an increasing effect from the outer to the inner part (where most of the constraints are). However these plots of the slice structures do not show either the entire ZOA nor the likelihood of a structure in the ZOA of the local
Universe. Although the standard deviation indeed highlights the part of the box that is the most constrained, it does not exhibit particular information on each one of the structures especially those in the ZOA. To get additional information, the $>$3$\sigma$ cells are plotted in the bottom panel of Fig.~\ref{fig:space} with a size proportional to the
probability that they are threshold-pass and a color indicating their distance in the supergalactic direction perpendicular to the plotted slice.
Structures with the highest probabilities (bigger dots) of containing structures are located at nearer distances along the line of sight,
simply because the technique is better able to reconstruct structures when the ZOA is less extended (i.e. nearby). However there are some more
distant structures that also exhibit probabilities with 3$\sigma$ significance. Such structures, further from the regions where
constraints are (since the ZOA covers a greater physical extent), need to be considered more fully, and compared with observational
expectations and predictions.Such comparisons are conducted in the next section.
\section{Observed and Predicted Structures}
Observers working on the ZOA prefer working in Galactic coordinates (longitude and latitude). To compare the theoretical results
with observations, we convert the grids in supergalactic coordinates to Galactic coordinates, and split the ZOA into small cells of Galactic
longitude, Galactic latitude, and distance, $(l,b,d)$. We compute the overdensities in these cells and gather the cells into partial spherical
shells of depth 10{ \textit{h}$^{-1}$~Mpc}. We then determine the overdensity contours on the projected distribution of cells. Several of these contour maps of
overdensity in partial spherical shells, centered on distances between 75 to 215{ \textit{h}$^{-1}$~Mpc}, are shown in Figs.~\ref{fig:gl_gb1}, \ref{fig:gl_gb2}, and \ref{fig:gl_gb3}.
Galaxies and clusters observed in the ZOA from various surveys are plotted on top of the contours. colors and symbols are as follows:
(1)~small red filled circles are part of our Nan\c cay observations of the Puppis 3 cluster;
(2)~small green filled circles are from the HIZOA-S survey \citep{2016AJ....151...52S};
(3)~small blue filled circles are from \citet{2006MNRAS.369.1131R};
(4)~small light blue filled circles are from the 2MASS redshift survey \citep{2012ApJS..199...26H} and
(5)~the big red filled circle is a cluster from the CIZA (Clusters In the Zone of Avoidance) project \citep{2002ApJ...580..774E}. Note that
small shifts in the Galactic longitude of galaxy points were applied in two panels (second panel of Fig. \ref{fig:gl_gb2} and second panel of
Fig. \ref{fig:gl_gb3}). These shifts were less than 5$^\circ$. The accuracy of the WF is about 2-3 { \textit{h}$^{-1}$~Mpc}. A small 5$^\circ$ shift at distances greater than 50 { \textit{h}$^{-1}$~Mpc}\ is thus completely justified and even expected.
In the rest of this section, the probability quoted for each detected structure is the percentage of CRs in which the structure is detected at the 1.0 overdensity threshold.
\subsection{The Puppis 3 cluster, other galaxy groups and a Cygnus concentration}
\begin{figure}
\vspace{-2.3cm} \hspace{-1cm}
\includegraphics[width=0.6\textwidth]{gl_gb75.eps}\\
\vspace{-4.2cm}
\hspace{-1cm}\includegraphics[width=0.6\textwidth]{gl_gb95.eps}\\
\vspace{-4.2cm} \hspace{-1cm}
\includegraphics[width=0.6\textwidth]{gl_gb95with2MRS.eps}\\
\vspace{-2.cm}
\caption{Galactic coordinate representation of the average overdensity field of 200 constrained realizations (black contours). The partial
spherical shells shown are 10{ \textit{h}$^{-1}$~Mpc}\ thick and are centered on structures in the Zone of Avoidance (ZOA) at distances of 75{ \textit{h}$^{-1}$~Mpc}\ and 95{ \textit{h}$^{-1}$~Mpc}. Red, green and light blue filled circles are galaxies observed in three different ZOA surveys. There is a remarkable agreement between the computed overdensity field and the observations of Puppis~3 in the top
panel. The two bottom panels probe distances of 95{ \textit{h}$^{-1}$~Mpc}\ where two galaxy groups have previously been identified (middle panel). The bottom panel suggests the presence of a
mass concentration in the Cygnus constellation. Galaxies observed on both sides of this concentration reinforce its existence. The observations and the average
reconstruction are in very good agreement.}
\label{fig:gl_gb1}
\end{figure}
The top panel of Fig.~\ref{fig:gl_gb1} shows a remarkable agreement between the observed galaxies (red filled circles) in the Puppis~3
cluster at about 70{ \textit{h}$^{-1}$~Mpc}, first hinted at by \citet{2004MNRAS.347..541C} and confirmed by \citet{2016AJ....151...52S}, and the average
reconstructed field (black solid contours); the Puppis~3 cluster has a 21.5\% probability in the reconstruction, namely it is significant at the 2$\sigma$ level. Interestingly enough, we note that \citet{2006MNRAS.373...45E} also theoretically predicted that cluster (C11 in their paper). The agreement is remarkable in the sense that they base their reconstruction on the 2MASS redshift survey while we base ours on the second radial peculiar velocity catalog of Cosmicflows, namely the two datasets are of different natures and do not share the same spatial distribution of datapoints.
The middle panel of the figure is remarkable in two ways: (1)~\emph{two} distinct clumps of galaxies at 95{ \textit{h}$^{-1}$~Mpc}\ are not only clearly observed but also
reconstructed with about 25\% probabilities (more than a 3$\sigma$ significance level); and (2)~they correspond to galaxy \emph{groups} with lower density than clusters and so would not
necessarily be expected to appear in the reconstruction. These groups were identified by \citet{2007ApJ...655..790C} in the Two Micron All Sky
Survey Extended Source Catalog \citep{2006AJ....131.1163S} using a percolation algorithm.
Finally, the bottom panel of the figure suggests the presence of a mass concentration in the Cygnus constellation close to Deneb and the North America Nebula (NGC7000). Galaxies from the 2MASS redshift survey (light blue filled circles) observed on both sides of this concentration strengthen its most probable existence. The structure is significant at more than the 2$\sigma$ level (23\% probability). Actually, \citet{2006MNRAS.373...45E} predicted a structure at about 100 { \textit{h}$^{-1}$~Mpc}, (l,b)=(100$^\circ$,-7$^\circ$) that they call C13, i.e. not far from the aforementioned mass concentration. Looking at this exact location, the probability to have a structure is in our case as high as 28\% (3$\sigma$ significance level).
\subsection{The Cygnus A cluster and two high density peaks}
\begin{figure}
\vspace{-2.3cm}\hspace{-1cm}
\includegraphics[width=0.6\textwidth]{gl_gb165.eps}\\
\vspace{-4.2cm}
\hspace{-1cm}\includegraphics[width=0.6\textwidth]{gl_gb170b.eps}
\vspace{-2.cm}
\caption{Galactic coordinate representation of the average overdensity field of 200 constrained realizations (black contours). The partial
spherical shells shown are 10{ \textit{h}$^{-1}$~Mpc}\ thick and are centered on structures in the Zone of Avoidance (ZOA) around
$(l,b)=(130^{\circ},0^{\circ})$ at a distance of 165{ \textit{h}$^{-1}$~Mpc}\ (top panel) and around $(l,b)=(80^{\circ},0^{\circ})$ at a distance of 170{ \textit{h}$^{-1}$~Mpc}\
(bottom panel). The red filled circle in the bottom panel shows the location of the Cygnus~A cluster observed in the CIZA project at a
distance about 170{ \textit{h}$^{-1}$~Mpc}. There is a remarkable agreement between the computed overdensity field and the location of Cygnus~A. The top panel
shows two reconstructed structures that need to be observationally confirmed.}
\label{fig:gl_gb2}
\end{figure}
The excellent agreement between the observations and the average reconstructed field at distances less than 100{ \textit{h}$^{-1}$~Mpc}\ encourages us to
probe greater distances. Figure~\ref{fig:gl_gb2} shows two slices centered at 165 and 170{ \textit{h}$^{-1}$~Mpc}. The top panel shows two structures at
about $(l,b)=(140^{\circ},-4^{\circ})$ and $(l,b)=115^{\circ},-1^{\circ})$ with probabilities of about 22\% and
24\% (respectively 2 and 3$\sigma$ significance level). Examining the various available online catalogs of galaxies and
clusters in the ZOA, we were unable to identify particular objects in this region due to a lack of available redshift measurements to confirm
the distances of the objects predicted at these locations. However, here again, there is a good agreement with the predictions from \citet{2006MNRAS.373...45E} although at the edge of the 2MASS redshift survey: they predicted two clusters (C31 and C33 in their paper) at very similar locations as those identified above. This new agreement inspires us to make some deeper comparisons between their predictions and ours although it is not the principal subject of this paper. The results are gathered in Appendix B and reveal astonishingly similar predictions. Combined with the agreement between the X-ray observation
of the Cygnus~A cluster and its probability of 29.5\% (significance of quasi 5$\sigma$) in the averaged reconstructed field visible in the bottom panel of the same figure, its stimulates further investigations at even larger distances, deeper in the ZOA.
\subsection{The Vela Supercluster}
Motivated by the recent announcement of the discovery of a supercluster in the Vela region of the ZOA \citep{2017MNRAS.466L..29K}, we plot in
Fig.~\ref{fig:gl_gb3} partial spherical shells at distances of 180 and 215{ \textit{h}$^{-1}$~Mpc}\ to probe this region. In the top panel of the figure, at the
claimed location of the Vela supercluster (denoted by the red dotted line ellipse centered at $(l,b)=(272.5\pm20^{\circ},0\pm10^{\circ})$,
there is a structure with a 27.5\% probability (i.e.\ 55 of the 200 CRs show the Vela supercluster structure or a 4$\sigma$ significance level). Note the very small
difference between the predicted position of the structure and that announced by \citet{2017MNRAS.466L..29K} based on
observations. \citet{2017MNRAS.466L..29K} describe the region as ``two merging wall-like structures, interspersed with clusters and
groups''. They observed the second wall to be at about 220{\rm km/s}. Remarkably, in the 10{ \textit{h}$^{-1}$~Mpc}\ thick slice centered on 215{ \textit{h}$^{-1}$~Mpc},
this second wall is present with a 23\% probability (almost a 3$\sigma$ significance level). A few galaxies
observed by \citet{2006MNRAS.369.1131R} indeed hint at the extension of a structure deep into the ZOA. Probing intermediate distances between
these two shells, we do not find any prominent structures, reinforcing the claim of two separate walls. All in all, this study reinforces the
claims of \citet{2017MNRAS.466L..29K}: the central core of the Vela supercluster seems to be hidden behind the thickest dust layers of the
Milky Way. If the Vela supercluster is as massive as the Shapley supercluster, it would reduce the difference in both direction and
amplitude between the observed and predicted values of the Local Group dipole motion.
\begin{figure}
\vspace{-2.3cm}\hspace{-1cm}
\includegraphics[width=0.6\textwidth]{gl_gb180.eps}\\
\vspace{-4.2cm}
\hspace{-1cm}\includegraphics[width=0.6\textwidth]{gl_gb215.eps}
\vspace{-2.cm}
\caption{Galactic coordinate representation of the average overdensity field of 200 constrained realizations (black contours). The partial
spherical shells shown are 10{ \textit{h}$^{-1}$~Mpc}\ thick and are centered on structures in the Zone of Avoidance (ZOA) around
$(l,b)=(260^{\circ},0^{\circ})$ at distances of 180{ \textit{h}$^{-1}$~Mpc}\ (top panel) and 215{ \textit{h}$^{-1}$~Mpc}\ (bottom panel). The red dotted line ellipse in the top
panel indicates the location of the recently discovered Vela supercluster. There is a remarkable agreement between the prediction
and the observation. The bottom panel shows the second wall beyond the center of the Vela supercluster.}
\label{fig:gl_gb3}
\end{figure}
\section{Conclusion}
The ZOA is very often discarded in studies of the local large-scale structure, either because observationally it is too complex and
difficult or because theoretically it is replaced by extrapolated or statistically random structures. However, in the upcoming era of high
precision cosmology, whole-sky coverage is increasingly essential to address issues related to the origin of the CMB dipole, to the largest
structures that possibly exist behind the ZOA, and to the local bulk flows. Consequently, more and more observational programs are taking on
the challenge of observing behind (or through) the Galactic dust. In this paper, we study the ZOA from a theoretical perspective, and
confront observations via the use of constrained realizations of the local Universe. These constrained realizations (CRs) stem from three
components: (1)~local observations of the large-scale structures within 200{ \textit{h}$^{-1}$~Mpc}, used as constraints; (2)~a random field to statistically
compensate for missing structures where the data are too sparse and too noisy; and (3)~a prior cosmological model. These constrained
realizations reproduce the linear part of the local overdensity field within 200{ \textit{h}$^{-1}$~Mpc}\ very well and, thus, constitute a sufficiently good simulacrum
of the local Universe for our statistical exploration of structures in the ZOA.
This paper is based on the statistical study of 200 CRs and their 200 random-realization counterparts (RRs; 200 random
realizations sharing the same prior cosmological model and random seed, calculated at a small computational cost). To show that in the random case the structures in the ZOA differ completely from one realization to another, while in the constrained case the structures are distributed in agreement with the surrounding environment, different recipes lead to different probability distribution functions. They are based on the overdensity values of cells of size 5{ \textit{h}$^{-1}$~Mpc}\ filling up the entire ZOA region. These values are compared to a given threshold and probability of each cell
being threshold-pass (i.e.\ having a value above the threshold) is
estimated from the 200 realizations. The probability distribution functions are defined as follows:
\begin{itemize}
\item constrained probability distribution: the procedure described above is looped over the 200 CRs and then reiterated over all the
cells. The result is the probability distribution of cells in the ZOA being threshold-pass.
\item random probability distribution: same as before but for the 200 RRs.
\item global probability distribution: for a given cell and threshold, once the procedure is looped over the 200 CRs on the one hand and over
the 200 RRs on the other hand, the two resulting counts are subtracted. The procedure is repeated for all the cells. It gives the
probability distribution of a cell in the ZOA being threshold-pass once the random probability has been subtracted.
\item individual probability distribution function: for a given cell,
threshold and random seed, the cell is counted as being threshold-pass if it meets the requirement of having an overdensity value above the
threshold only in the CR and not in the RR sharing the same random seed. The procedure is reiterated over the 200 CR/RR pairs and then
over all the cells in the ZOA. It results in the probability distribution of a cell in the ZOA being threshold-pass solely because
of the constraints; the random part alone would not have induced an overdensity sufficiently high for the cell to be threshold-pass.
\end{itemize}
If the constraints did not affect the reconstruction of the ZOA, the constrained and random probability distributions would be identical;
this is not the case. While the random probability distribution is a Gaussian, the constrained one is skewed and flatter,
revealing that cells in the high tail of the constrained distribution most likely host a structure---their probability is increased due to the constraints.
That the random and constrained probability distributions differ is confirmed by the shape of the global probability distribution
function. This function is a Gaussian; this implies that some cells have a probability to host a structure that is greater over 200 constrained realizations than these same
cells in 200 random realizations associated with the same prior cosmological model. The individual probability distribution is even more
remarkable: it confirms that some cells host structures by the sole action of the constraints; the random part alone would not cause the cells to be threshold-pass.
The probability derived for each cell allows us to locate the reconstructed structures in the ZOA that have the highest probability
(i.e.\ in which we can have the greatest confidence). These structures are confronted with observations. As expected, since the reconstruction
technique works well nearby, where the physical extent of the ZOA is least, there is good overall agreement between reconstructed and
observed structures close by. For instance, the Puppis~3 cluster at about 70{ \textit{h}$^{-1}$~Mpc}\ is successfully reconstructed. More notably, this
agreement extends farther away: the Cygnus~A cluster at about 170{ \textit{h}$^{-1}$~Mpc}\ is equally well reconstructed. The agreement for these two known
structures confirm the effectiveness of the techniques and encourages us to study even larger distances and deeper into the ZOA. Looking at
180{ \textit{h}$^{-1}$~Mpc}, we find a structure reconstructed with 27.5\% probability (i.e. at the 1.0 overdensity threshold there is a structure in 27.5\% of the CRs) that is almost centered on the location of the
recently discovered Vela supercluster \citep{2017MNRAS.466L..29K}. The signal for this structure is significant at the 4$\sigma$ level. This prediction is remarkable in the sense that it is inferred with observational data, 98\% of which are within 160{ \textit{h}$^{-1}$~Mpc}\ and 50\% of which are within 61{ \textit{h}$^{-1}$~Mpc}. Furthermore, probing slightly larger distances
still, we also see the second wall at about 220{ \textit{h}$^{-1}$~Mpc}\ behind the centre of the Vela supercluster also brought to light by \citep{2017MNRAS.466L..29K}; this second wall is reconstructed despite
being outside the region covered by the data used as constraints.
This study lends support to the existence of two merging wall-like structures forming the Vela supercluster, with a central core exactly behind the deepest part of the ZOA. This supercluster might resolve the misalignment of clustering dipoles and account for the residual bulk flows.
\section*{Acknowledgements}
The authors gratefully acknowledge the Gauss Centre for Supercomputing
e.V. (www.gauss-centre.eu) for providing computing time on the GCS
Supercomputer SuperMUC at LRZ Munich. The authors thank the referee for constructive comments that helped to improve the manuscript. This research has made use of the
SIMBAD database and the VizieR catalog access tool, operated at CDS,
Strasbourg, France as well as of the TOPCAT interactive graphical tool
developed by Mark Taylor. JS acknowledges support from the Astronomy
ESFRI and Research Infrastructure Cluster ASTERICS project, funded by
the European Commission under the Horizon 2020 Programme (GA 653477). RKK thanks the South African National Research Foundation for their support.
|
\section{Introduction}
In a one-class classification problem statement, it is assumed that we mainly use data of one, normal, class to build a decision function that describes characteristics of the data. On the other hand data of the other abnormal class are either not used at all or used to refine the obtained data description. The one-class classification is broadly applied to many real application domains namely image processing, network intrusion detection, user verification in computer systems, machine fault detection, etc. In most of the real-world applications for one-class classification, the number of normal data samples is much larger than that of abnormal data samples or abnormal samples are even not available at all \cite{burnaev}. The reason is that collecting normal data is inexpensive and easy in comparison to collecting abnormal data. For example in machine fault detection applications, a normal data can be collected directly under the normal condition of a machine while collecting an abnormal data requires observation of machines until they break. Since the prevalence of one class, in a one-class classification, the boundary decision primarily comes from the dominant class compared to a binary classification, where the data of both classes are used to construct the boundary decision.
One of the most popular approaches for one-class classification is based on support vector models. These models have a nice inherited property allowing to build simple linear decision boundaries in highly non-linear infinite-dimensional spaces thanks to the kernel trick. The most wide-spreadly used kernels are Gaussian kernels. Although these techniques became extremely popular in recent years, there still exists an open question of the kernel bandwidth selection, which, as in binary classification, is crucial for model performance. Despite similarity to the binary classification problems, the standard approaches for bandwidth selection, like cross-validation, can not be used for the one-class classification models due to specific nature of the data (absence of second, abnormal, class data).
In this paper we concentrate on Support Vectors Data Description (SVDD) \cite{tax2004support} which is a one-class
classification algorithm from the family of support vector models. The main difference from a standard one-class SVM is that in a linear case SVDD builds a sphere surface around the normal data instead of a linear hyperplane, separating the normal data from the origin (in a feature space).
Several approaches for kernel selection were proposed in the literature. Some of them are intended for binary classification but can be generalized to the one-class problem, some are specially designed for one-class classification problem.
For example, in \cite{steinwart2005classification} authors proposed an empirical risk for one-class SVM and proved that under certain conditions it converges to the real risk. Another approach \cite{lukashevich2009using} is based on an idea that a kernel width should be selected in such a way that a fraction of outliers and support vectors in the final model should be close to their corresponding theoretical estimates, depending on parameters of the SVDD algorithm.
In \cite{xiao2014two} authors try to optimize the decision region directly in order to get neither too sparse nor too dense area. All the mentioned approaches imply building and comparing several models w.r.t. some criterion, while in \cite{evangelista2007some} a direct optimization method for kernel parameters is proposed without explicit model building.
In this paper we propose an original method for SVDD model selection based on empirical risk approximation using oversampling with SMOTE \cite{chawla2002smote} and develop a methodology for SVDD model selection. Finally, we generalize several kernel selection methods from binary-class case to the case of one-class classification
and perform extensive comparison of these approaches using both synthetic and real-world data.
\section{Anomaly Detection using SVDD}
We consider Support Vector Data Description (SVDD) proposed in \cite{tax2004support} for the one-class classification problem. Let us have a dataset $D = \{X_1,\ldots, X_l\},$ $X_i\in \mathbb{R}^n$. Let $\phi_{\gamma}(\cdot)$ be some mapping to a high-dimensional space depending on scalar parameter $\gamma\in \mathbb{R}_{+}$. In this space we solve the following optimization problem:
\begin{equation}
\begin{cases}
R^2 + \frac{1}{\nu l}\sum_{i=1}^l \xi_i \to \min_{R,a}\\
\|\phi_{\gamma}(X_i) - a \|^2 \leqslant R^2 + \xi_i &, i = 1,\ldots, l,\\
\xi_i \geqslant 0, & i = 1, \ldots, l.
\end{cases}\label{eq:svdd}
\end{equation}
Physical meaning of such an optimization problem statement is that we are looking for a minimum volume ball in high-dimensional space defined by $\phi_{\gamma}(\cdot)$, containing the dataset.
Herewith we allow some points to be outside the boundary of the ball and controll their number by $\nu$. The penalty in (\ref{eq:svdd}) is equivalent to $l_1$ penalty that zeros out some of the parameters. In our case it means that several points will belong exactly to the surface of the ball \cite{tax2004support}. We can re-write \eqref{eq:svdd} in dual form and find that a decision function, predicting whether a point is an anomaly or not, is defined through the kernel $K_{\gamma}(X,X') = \langle\phi_{\gamma}(X),\phi_{\gamma}(X')\rangle$ and has the form $f_{\gamma}(X) = \mathrm{sign}\left(\sum_{i=1}^l\alpha_iK_{\gamma}(X,X_i)-\rho\right)$ for some $\alpha_i\geq0$ and $\rho$. A typical example of the kernel we are using is $K_{\gamma}(X,X') = \exp\left(-\|X-X'\|^2/\gamma\right)$, $\gamma>0$.
\section{Model Selection Techniques}
In this section we give brief descriptions of several model selection approaches for SVDD, which are based on optimization of some risk functions.
\subsection{Support Vectors and Anomalies Fraction Optimization}
Complexity of the model is described by the number of support vectors $X_i\in D$ corresponding to $\alpha_i>0$. In the original paper \cite{tax2004support} authors proved that parameter $\nu$ in (\ref{eq:svdd}) is an upper bound for the fraction of elements in the train data marked as anomaly (outliers) and a lower bound for the fraction of elements used as support vectors.
Let us define the risk function as follows
\begin{equation}
R^{\textrm{SV}}_{\gamma} = (\nu - f_{\textrm{SV},\gamma})^2,
\end{equation}
where $f_{\textrm{SV},\gamma}$ is a fraction of support vectors in the dataset, defining the current decision function $f_{\gamma}(X)$.
Minimizing $R^{\textrm{SV}}_{\gamma}$ we can balance complexity and classification accuracy on the train set.
\subsection{Empirical Risk}
Another approach would be to reduce the one-class classification to binary classification \cite{steinwart2005classification}. It can be proved that the following empirical risk converges to a real risk of this binary classification problem
\begin{equation}\label{empiric}
R_{\gamma}^{\textrm{empirical}} = \frac{1}{(1-\nu) \cdot l}\sum_{i=1}^n [f_{\gamma}(X_i) = -1] +
\frac{1}{\nu}\mathbb{E}_{\mu}[f_{\gamma}(X_i) = 1],
\end{equation}
where $\mathbb{E}_{\mu}[f(X_i) = 1]$ denotes average error of classification on anomaly elements assuming that they have distribution $\mu$.
This value is estimated using Monte Carlo simulation.
Generally, for modeling $\mu$ it is reasonable to use a least favorable uniform distribution.
\subsection{Risk based on Oversampling}
In empirical risk (\ref{empiric}) the same dataset is used for a model construction and its risk estimation leading to a biased estimates.
That is why for the risk assessment we propose to use sampling based on SMOTE \cite{chawla2002smote}, an oversampling technique, designed for the imbalanced binary classification. We define the risk as
\begin{equation}\label{smote}
R_{\gamma}^{\textrm{SMOTE}} = \frac{1}{(1-\nu) \cdot m}\sum_{i=1}^m \left[f_{\gamma}\left(\tilde{X}_i\right) = -1\right] +
\frac{1}{\nu}\mathbb{E}_{\mu}[f_{\gamma}(X_i) = 1]
\end{equation}
New examples $\tilde{X}_i$ of a normal data are generated synthetically based on the training set and the SMOTE algorithm. The first summand
in (\ref{smote}) is proportional to the number of synthetic normal elements marked by the classifier as anomalies. The second summand is the same as in (\ref{empiric}).
\subsection{Kernel Matrix Optimization}
Another popular idea is to optimize a kernel matrix directly without building an anomaly detection model. For example, in \cite{evangelista2007some} a simple statistic was proposed:
\begin{equation}\label{kernel_variance}
R_{\gamma}^{\mathrm{kernel}} = \dfrac{\bar{k}_{\gamma}}{s^2_{\gamma}},
\end{equation}
where $\bar{k}_{\gamma} = \frac{\sum_{i=2}^n\sum_{j=i}^n K_{\gamma}(X_i, X_j)}{(n-1)(n-2)}$ and $s^2_{\gamma} = \frac{\sum_{i=2}^n\sum_{j=i}^n (\bar{k}_{\gamma} - K_{\gamma}(X_i, X_j))^2}{(n-1)(n-2)}
$ are a normalized sum of the kernel matrix elements and a normalized variance of the kernel matrix elements correspondingly.
\subsection{Kernel Polarization}
Another approach for tuning the kernel matrix is a so-called polarization optimization. It was originally proposed for binary classification problems but can be naturally generalized for one class classification problems \cite{steinwart2005classification}.
In order to calculate the polarization we generate an artificial anomaly data $\{X_{l+1},\ldots,X_{2l}\}$ from a uniform distribution.
All elements of the initial data set are marked as normal data and finally we have a set of elements $(X_i,y_i)$, $i = 1,\ldots,2l$, where $y_i = 1$ for $i = 1,\ldots,l$ and $y_i = -1$ for $i = l+1,\ldots,2l$. The polarization risk is calculated as
\cite{baram2005learning}
\begin{equation}
R_{\gamma}^{\mathrm{polar}} = -\sum_{i, j=1}^{2l} y_i \cdot K_{\gamma}(X_i, X_j) \cdot y_j.
\end{equation}
\section{Numerical Experiments}
In this section we describe our settings for the numerical experiments and provide results of kernel selection techniques comparison.
\subsection{Validation Error}
The real risk of anomaly detection consists of two parts: misclassification of points from a normal distribution as anomalies and misclassification of abnormal data as a normal.
For measuring the quality of a decision function we will use sum of these errors on independent validation data set normalized by the number of elements.
Let $\left\{X^{\mathrm{normal}}_1, \cdots, X^{\mathrm{normal}}_s\right\}$ be the set of normal elements in the validation set and $\left\{X_1^{\mathrm{anomaly}}, \cdots, X_m^{\mathrm{anomaly}}\right\}$ be the set of anomaly elements in the validation set. Then the real risk estimate is calculated as follows:
\begin{equation}
R^{\mathrm{val}}_{\gamma} = \frac{1}{s}\sum_{i=1}^s\left[f_{\gamma}\left(X_i^{\mathrm{normal}}\right) = -1\right]
+ \frac{1}{m}\sum_{j=1}^m\left[f_{\gamma}\left(X_j^{\mathrm{anomaly}}\right) = 1\right].
\end{equation}
The validation risk behavior depends on a dimensionality $n$ of a data. A typical behavior is presented in figure~\ref{fig:validation-risk}. Generally there is a plateau of almost the same validation error values for some range of $\gamma>0$.
\subsection{Behavior of the Risk Functions}
In order to understand a typical behavior of the introduced risk functions we estimate $R^{\mathrm{val}}_{\gamma}$ of the constructed model for different values of $\gamma$ and compare it to the estimates of the risks.
In these experiments we fix the value of $\nu$ to $0.1$, meaning that we have $10\%$ of anomalies.
We generate $100$ points of normal data from a mixture of two $5$-dimensional normal distributions with means at $[1, 1, 1, 1, 1]^{T}$ and $[-1, -1, -1, -1, -1]^{T}$, and identity covariance matrices. Also we generate $100$ anomalies from the uniform distribution on $[-5,5]^{5}$. For the validation a huge dataset from the same distributions is drawn.
Empirical risk (see figure~\ref{fig:empirical-risk-artificail}) looks very similar to the validation error, ranges of plateaus mostly coincide. Risk based on SMOTE oversampling (see figure~\ref{fig:smote-risk-artificial}) looks even more similar to the real validation error. But in either cases choosing the minimizer of the risk as an optimal value for $\gamma$ we obtain similar results. In case of support vectors based risk (see figure~\ref{fig:support-risk-artificial}) the plateau of the risk is much wider than those of the validation error. The kernel risk (see figure~\ref{fig:kernel-risk-artificial}) behaves very smoothly with a single minimum as it was shown in \cite{evangelista2007some} but this mimimum is biased with respect to the optimal value of $\gamma$. Finally, the polarization risk (see figure~\ref{fig:polarization-risk-artificial}) also gives a single minimum and the optimum value belongs to the plateau of the validation error.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{test_risk_dim.png}
\caption{Validation error behavior}
\label{fig:validation-risk}
\end{subfigure} ~ \begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{empiric_risk.png}
\caption{Empirical risk}
\label{fig:empirical-risk-artificail}
\end{subfigure}
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{smote_risk.png}
\caption{SMOTE risk}
\label{fig:smote-risk-artificial}
\end{subfigure} ~ \begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{support_risk.png}
\caption{SV risk}
\label{fig:support-risk-artificial}
\end{subfigure}
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{kernel_risk.png}
\caption{Kernel risk}
\label{fig:kernel-risk-artificial}
\end{subfigure} ~ \begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{polarization.png}
\caption{Polarization risk}
\label{fig:polarization-risk-artificial}
\end{subfigure}
\caption{Dependency of a risk function and a validation error on $\gamma$}
\end{figure}
\subsection{Distribution of Anomalies}
We use a uniform distribution of anomalies
when calculating the proposed risk functions.
It is natural to test how these risk functions depend on the distribution of anomalies. We use a normal distribution with variance, comparable to the range of the normal data, to generate anomalies in the learning data sample. At the same time when estimating risks we use uniformly distributed synthetic anomalies. From figure \ref{normal-anomaly} we see that it is still possible to use uniformly generated synthetic anomalies even if true anomalies are generated from another distribution; and we can select a reasonable $\gamma$ as the largest value from the plateau of the corresponding risk function.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{empiric_normal.png}
\caption{Empiric risk}
\label{fig:gull}
\end{subfigure} ~ \begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{smote_normal.png}
\caption{SMOTE risk}
\label{fig:tiger}
\end{subfigure}
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{polarization_normal.png}
\caption{Polarization risk}
\label{fig:gull}
\end{subfigure} ~ \begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{uniform_empiric.png}
\caption{Empiric risk}
\label{fig:gull}
\end{subfigure}
\begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{uniform_smote.png}
\caption{SMOTE risk}
\label{fig:tiger}
\end{subfigure} ~ \begin{subfigure}[b]{0.40\textwidth}
\includegraphics[width=\textwidth]{uniform_polarization.png}
\caption{Polarization risk}
\label{fig:gull}
\end{subfigure}
\caption{Dependency of a risk function and a validation error on $\gamma$ for a non-uniform anomaly distribution}
\label{normal-anomaly}
\end{figure}
\subsection{Real World Data}
Real world data were taken from
\url{http://homepage.tudelft.nl/n9d04/occ/index.html}.
These data sets are generated from multi class classification problems like ``Iris Dataset'', ``Sonar'', etc. For these problems we considered each class as normal and added anomalies
from a uniform distribution to get a one-class classification problem. Borders of the uniform
distribution were taken as double borders of the normal class. Number of anomalies were equal to $5\%$, $10\%$ and $15\%$ of a normal class size. Finally we constructed $96$ datasets with smallest value of $n$ equal to $5$ and biggest value of $n$ equal to $1909$. For some of the datasets $l\ll n$ (datasets, obtained from ``Leukemia'' and ``Colon'' classification problems). For majority of datasets $l\leq 1000$, but also there are several big datasets like those, obtained from ``Spam Base'' ($l = 4600$)
and ``Concordia'' ($l=4000$) classification problems. Also there are data sets with a very small
$l$ like that obtained from ``Leukemia'' ($l = 72$) classification problem.
To compare model selection methods on real data we use Dolan-More curves \cite{dolan-more} which are built in the following way. Let $\{R_1, \ldots, R_K\}$ be the set of considered model selection methods, $\{D_1, \ldots, D_M\}$ be the set of tasks (datasets), $q_{ti}$ be the quality of the method $i$ on the dataset $t$. For each method $i$ we introduce $p_i(\beta)$, a fraction of datasets, on which the method $i$ is worse than the best one not more than $\beta$ times:
$$
p_i(\beta) = \frac{1}{T} \left| \left\{t: q_{ti} \geq \frac{1}{\beta} \max\limits_{i} q_{ti} \right\}\right|, \ \beta \geq 1.
$$
For example, $p_i(1)$ is a fraction of datasets where the method~$i$ is the best. A graph of $p_i(\beta)$ is called Dolan-More curve for the method $i$. This definition implies that the higher the curve, the better the method.
Note that Dolan-More curve for a particular method depends on other methods considered in comparison.
\begin{figure}[t!]
\begin{center}
\includegraphics[resolution=900, scale=0.6]{Dolan_More.png}
\end{center}
\caption{Dolan-More curves for comparison of model selection methods on real data}
\label{dolan2002benchmarking}
\end{figure}
From figure \ref{dolan2002benchmarking} we can see that perfomances of the SMOTE risk and of the Empirical risk are quite similar, but the curve for the SMOTE risk is slightly higher than that for the Empirical risk. The kernel risk and the Support Vectors risk provides worse perfomance.
\section{Conclusions}
We considered Support Vectors Data Description as a one-class
classification algorithm; generalized several model selection methods from binary classification for SVDD model selection
and performed extensive comparison of these approaches using synthetically generated data and real-world data sets.
\textbf{Acknowledgement}: The research was conducted in IITP RAS and supported solely by the Russian Science Foundation grant (project 14-50-00150).
\renewcommand\refname{\uppercase{References}}
|
\section{Introduction}
Cases of scientific misconduct are identified every year.
Scientific papers are retracted because of errors, or for suspected fraud, ranging from plagiarism and minor manipulations to faking the data and disguising the results.
It has been shown that, however, among the retracted articles indexed in PubMed, only $21.3$\% are retracted due to error, while $67.4$\% were removed due to misconduct, among which suspected fraud amounts to $43.4$\%, the others being due to duplicate publications or plagiarism \cite{fang:misconduct:12}.
In a recent paper, \newcite{markowitz_linguistic_2015} proposed the first analysis of writing style in fraudulent papers across authors and disciplines.
They approached the question of whether these authors have a specific writing style, from a psychological perspective.
They found that these papers exhibit a higher rate of jargon, make a higher use of references, and have a lower readability rate, suggesting that the authors try to obfuscate their writing, making them harder to read and analyze.
They report classification results using a leave-one-out strategy over the dataset, with a classification accuracy of $57.2$\%.
As suggested in the paper, we propose to improve this performance by evaluating different classification models.
In this paper, we first show that much better results can be obtained using a simple bag-of-words representation and Logistic Regression.
Our best model is a syntax-enhanced trigram-model. We also show that the leave-one-out strategy used by the authors leads to an over-estimation of model precision, and we report new results based on a more robust strategy, taking into account the low number of instance available; namely a {\em nested} cross-validation \cite{varma_bias_2006,scheffer_error_1999}. We also considered semantic and discourse features, but we did not observe improvements with such features.
Of course, that a bag-of-words model outperforms a model based on psychologically motivated features, may simply be the result of overfitting. We present an extensive feature analysis to validate our models, as well as to test psychologically motivated hypotheses from the literature.
\paragraph{Contributions}
(i) We present a simple model with high accuracy, and show that it implicitly captures the previously-proposed psychologically-motivated features.
(ii) We show that adding semantics and discourse features does not lead to improvements.
(iii) On the other hand, our feature analysis suggests that the models {\em do} learn to focus on concepts that are intuitively related to scientific misconduct, e.g., that scientific fraud contains less comparison.
\section{Related work}
\newcite{markowitz_linguistic_2015} were the first to study writing style in fraudulent papers.
They gathered a corpus of $253$ articles indexed in PubMed that have been retracted for fraudulent data, as well as $253$ unretracted papers (see Section~\ref{sec:data}).
They define five indicators of obfuscation, and show that fraudulent papers tend to demonstrate a higher rate of linguistics obfuscation, corresponding to a lower readability, an higher use of jargon and a higher degree of abstraction.
Linked to studies on deception identification, they also report a lower rate of positive emotion terms and a higher rate of causal terms (e.g. ``depend", ``induce", ``manipulated") in fraudulent papers.
The readability score was computed using Coh-Metrix \cite{mcnamara2013coh}, while the other scores were based on the Linguistic Inquiry and Word Count (LIWC; \cite{pennebaker2007linguistic}), a dictionary associating a word to various scores such as abstraction (a word is considered as jargon if it is not found in the dictionary).
Finally, they report $57.2$\% in accuracy using these five indicators as features, a score that we show is probably a little too optimistic, since it is based on a leave-one-out procedure (see Section~\ref{sec:expe}).
We extend their work by first showing that a simple unigram model outperforms their model by a large margin, but also by considering more indicators, including discourse and syntax, and by showing, as mentioned, that their scores were probably over-estimated due to their validation strategy.
Our work is also inspired by another related field of research concerned with deception detection.
\newcite{mihalcea2009lie}
built three datasets consisting of $100$ true and $100$ deceptive short statements on three different topics (abortion, death penalty, best friend).
Using only unigrams, they report 70.8\% accuracy in a 10-fold cross validation.
They found that specific word classes, as defined in the LIWC, were predictive of deceptive texts, especially classes indicating detachment from self or related to certainty.
\newcite{feng2012syntactic} investigate syntactic features, using lexicalized and unlexicalized production rules in addition to shallow features (words unigram and bigram, and POS unigram).
They experiment on truthful and deceptive reviews from TripAdvisor, either gold \cite{ott2011finding} or retrieved using a fake review detector \cite{feng2012distributional}, reviews automatically extracted from Yelp,
and the corpus introduced in \cite{mihalcea2009lie}.
They report scores between $64.3$ and $91.2$\% accuracy, depending on the dataset.
They found that, for all datasets, syntax helps, and
that deceptive reviews more frequently use VP, SBAR and WHADVP.
We also consider $n$-gram features, syntactic features, as well as discourse features.
Our task is however a bit different, since authors of fraudulent papers are not directly lying, rather trying to conceal their fraud.
Moreover, our documents are longer and are of a different genre, i.e. scientific articles.
\section{Data}
\label{sec:data}
We use the dataset proposed in \cite{markowitz_linguistic_2015} containing $253$ publications retracted for data fraud and $253$ unretracted publications.
These publications were taken from the PubMed archives from $1973$ through $2013$.
The unretracted papers are extracted by considering one retracted paper and taking a control paper published the same year, in the same journal, and with some common keywords when possible.
When no such paper exists (around $19$\% of the papers), a paper from an adjacent year, or using the same words in the abstract, was selected.
The data used is the pre-processed version presented in \cite{markowitz_linguistic_2015}: Words were converted from British English to American English forms. Brackets, parentheses, and percent signs were removed. Periods were removed from certains words, such as `Dr.' or Ìnc.'.
The documents only contain the main body text (no section titles, figures, or tables).
\section{Methodology}
We investigate different types of features, from $n$-grams to discourse.
In large vocabulary feature spaces, we perform feature reduction, to reduce sparsity.
We then provide an analysis of the features to identify the most informative indicators.
\paragraph{Word features}
We use word $n$-grams as features,
with $n \in \{1,2,3\}$. In order to test the hypotheses presented in previous studies, we also use lexicons to extract information about the tokens.
We use the General Inquirer \cite{stone:general:1966} to extract words expressing a \textit{polarity} -- the features built represent the polarity between positive, negative, both and neutral --, and words corresponding to a \textit{causal} term.
We also use this lexicon to map the words to a more general semantic category (\textit{Inquirer}).
We identify all the personal \textit{pronouns}
using manually defined lists.
Finally, we also include as features \textit{hedge} and modal words, also using a pre-defined list.\footnote{\url{https://github.com/wooorm/hedges/blob/master/index.json}}
\paragraph{Syntactic features}
In order to obtain syntactic information, we parse the data using UDPipe\footnote{\url{http://ufal.mff.cuni.cz/udpipe}} \cite{udpipe:2016}, and a prebuilt model available online for English.\footnote{UD 1.2, \url{https://lindat.mff.cuni.cz/repository/xmlui/handle/11234/1-1659}}
We follow \cite{johannsen:cross:2015} in extracting all subtrees of up to three tokens (\textit{treelets}).
\paragraph{Discourse features}
Finally, we automatically annotate all the data with discourse connectives and explicit discourse relations using simple models trained on the Penn Discourse Treebank (PDTB) \cite{prasad:penn:2008}, a corpus of news articles from the Wall Street Journal.
Discourse coherence is an indicator of the quality of a text \cite{lin:automatically:2011}, of its reasoning that could reveal an attempt to deceive. Some specific semantic relations could also be good indicators (e.g. Cause).
We used models to identify the discourse connectives (\textit{Connectives}) and to identify the explicit discourse relation\footnote{We ignore the non explicit relations for which the in-domain scores are very low -- around 40-57\% in accuracy \cite{rutherford:improving:2015,lin:pdtb:2014}.} (\textit{Explicit relations}) they trigger, either among the $4$ coarse-grained classes (\textit{lvl1}) at the top of the hierarchy of sense
or using the $11$ more fine-grained relations at the second level (\textit{lvl2}).
Our models use Logistic Regression and the connective and the surrounding words and their POS as features \cite{lin:recognizing:2009}.
They are trained on the sections 2-21 of the PDTB.
Our results on the section 23 are close to the state-of-the-art \cite{pitler:using:2009,pitler:easily:2008,lin:pdtb:2014}:
$92.9$\% in accuracy for identifying the connectives, $95.1$\% for the level-1 relations, and $86.2$\% for the level-2 relations.
\paragraph{Feature analysis}
In addition to presenting accuracies obtained with these feature sets, we also perform a feature analysis. For this purpose we use a combination of correlation coefficients, logistic regression coefficients, and stability selection
\cite{meinshausen2010stability} -- a method that consists in repeatedly fitting the model across different random subsamples, and counting how many times features are selected in $\ell_1$-regularized logistic regression models.
For stability selection, we use the implementation available in scikit-learn \cite{scikit-learn} with its default parameters, run it on the whole dataset and keep features selected more than $50$\% of the time.
We indicate the size of the original vocabulary and the number of selected features for each category in Table \ref{table:vocab}.
\begin{table}[t!]
\resizebox{\textwidth}{!}{
\begin{tabular}{lrr}
\toprule
Category & \# Orig. feat. & \# Selec. feat. \\
\midrule
Unigrams & $65,798$ & $118$ \\
2-3-grams & $1,745,188$ & $154$ \\
Polarity & $4$ & $-$ \\
Causal & $68$ & $-$ \\
Inquirer & $180$ & $-$ \\
Pronouns & $7$ & $-$ \\
Hedges & $121$ & $-$ \\
\midrule
Treelets & $50,522$ & $136$ \\
\midrule
Connectives & $70$ & $-$ \\
Explicit relations lvl1 & $4$ & $-$ \\
Explicit relations lvl2 & $10$ & $-$ \\
\bottomrule
\end{tabular}}%
\caption{Size of the original vocabulary and number of selected features for $n$-grams and treelets.}
\label{table:vocab}
\end{table}
\section{Classification}
\label{sec:expe}
\paragraph{Representation}
We test separately count vectorizations with each set of features -- unigrams, 2-3-grams, polarity, causality, Inquirer categories, pronouns (grouping per person, or considering each lemma), treelets, connectives, hedge words, level-1 relations and level-2 relations, and combinations of these features.
\paragraph{Model}
We use a binary logistic regression classifier, optimizing the norm ($\ell_1$ or $\ell_2$) and strength ($c \in \{0.001, 0.005, 0.01, 0.1, 0.5, 1, 5, 10, 100\}$ of the regularization term on held-out data.
\begin{figure*}[!ht]
\begin{subfigure}{.3\textwidth}
\centering
\includegraphics[width=\linewidth]{plot_fraud_loo-nloounigrams}
\caption{Unigrams}
\label{fig:sfig1}
\end{subfigure}%
\begin{subfigure}{.3\textwidth}
\centering
\includegraphics[width=\linewidth]{plot_fraud_loo-nloo2-3-grams}
\caption{2-3-grams}
\label{fig:sfig2}
\end{subfigure}
\begin{subfigure}{.3\textwidth}
\centering
\includegraphics[width=\linewidth]{plot_fraud_loo-nlootreelet}
\caption{Treelets}
\label{fig:sfig3}
\end{subfigure}%
\caption{Accuracy difference between LOO and Nested LOO for each trial for different features.}
\label{fig:diffloo}
\end{figure*}
\paragraph{Validation schemes}
\newcite{markowitz_linguistic_2015} report results with a leave-one-out strategy (LOO).
However, LOO often under-estimates the error rate.
We compare with a {\em nested}~cross-validation procedure that can provide an almost unbiased estimate of the true error \cite{varma_bias_2006,scheffer_error_1999}.
Specifically, we use two cross-validation loops: the inner loop is used for tuning the hyper-parameters, and the outer loop estimates the generalization error.
The data are first split into $N$ folds,
the fold $k$ ($1 \leq k \leq N$ ) is the current evaluation set, and the $N-1$ other folds are used as training data and split into $M$ folds used for model fitting.
The best model is then evaluated on fold $k$.
Final scores are averages over the $N$ folds.
For comparison with \newcite{markowitz_linguistic_2015}, we report performance with LOO and with nested cross-validation using
LOO as outer loop, the inner loop being a random $5$-fold cross-validation.
We repeat each evaluation $10$ times, and report a mean over these trials.
\begin{table}[t!]
\resizebox{\textwidth}{!}{
\begin{tabular}{lll}
\toprule
System & LOO & N-LOO \\
\midrule
\cite{markowitz_linguistic_2015}& $57.2$ & - \\
\midrule
Unigrams & $72.1$ & $71.7$ \\
2-3-grams & $70.8$ & $69.6$ \\
Polarity & $50.0$ & $45.3$ \\
Causal & $59.9$ & $58.4$ \\
Inquirer & $58.7$ & $54.3$ \\
Pronouns & $54.5$ & $52.2$ \\
Hedges & $56.7$ & $54.1$ \\
\midrule
Treelets & $72.9$ & $71.7$ \\
\midrule
Connectives & $60.1$ & $58.3$ \\
Explicit Relations lvl1 & $54.3$ & $53.2$ \\
Explicit Relations lvl2 & $54.5$ & $54.3$ \\
\midrule
\midrule
1-2-3-grams+treelets & $\mathbf{76.3}$ & $\mathbf{76.0}$ \\
All & $70.3$ & $69.8$ \\
\bottomrule
\end{tabular}}%
\caption{Results (accuracy, in \%).}%
\label{tab:results}
\end{table}
\paragraph{Results}
Our results are summarized in Table \ref{tab:results}.
Our results are generally higher than the $57.2$\% reported in \cite{markowitz_linguistic_2015}, with at best $71.7$\% with a nested LOO and a single group of features (\textit{unigrams} or \textit{treelets}) and $76.0$\% when $n$-grams and treelets are combined.
Using all the $n$-grams already leads to a better accuracy score ($+1.3$\%) compared to using only \textit{unigrams} ($73.0$\% in accuracy for \textit{1+2-3-grams} with N-LOO).
On the other hand, combining discourse features to the $n$-grams does not allow improvements over using only the $n$-grams ($72.8$\% with N-LOO for \textit{1+2-3-grams+Connectives+Explicit Relations lvl1}).
The scores obtained with LOO are over-estimate performance, compared to nested cross-validation, see for example Figure \ref{fig:diffloo}: Even if the differences are low, they are consistent across the trials and the feature sets.
\section{Feature analysis}
We use Pearson's $\rho$ (w.~Bonferroni correction) to establish what features are predictive of fraud and non-fraud.
We report the values for the features cited in Table~\ref{tab:indicators}.
\paragraph{Hedging} There is an interesting contrast between adverbial hedges ({\em conceivably, presumably, surely, effectively}) and verbal hedges ({\em suggest}) indicative of fraud, and adverbial hedges ({\em practically, occasionally}) and verbal hedges {\em assume, speculate}) indicative of non-fraud: It seems adverbs and verbs used in fraud are for interpreting the data on behalf of the reader, whereas the adverbs and verbs indicative of fraud are more observer-aware (e.g.,~{\em we speculate}). This suggest that a fraud strategy is to hide observer’s bias, rather than being explicit about it.
\paragraph{Comparison} Both the discourse relation and the Inquirer class for comparison are predictive of non-fraud. Scientific fraud thus seems less likely to compare.
On the other hand, neither the causal relations or the presence of causal terms were significantly linked to fraudulent papers.
\paragraph{Therefore vs. since} A peculiar, but statistically significant difference between fraud and non-fraud articles, is that fraud articles prefer {\em therefore} over {\em since}, and vice versa. We speculate that it may be a fraud strategy to make the reasoning more verbose by separating out premises (because the authors are, consciously or not, afraid the readers will not accept them). This is in slight contrast with or qualifies the main hypothesis in \newcite{markowitz_linguistic_2015}, that fraudulent writers try to obfuscate their writing.
\paragraph{Other markers of fraud} Many technical concepts were highly correlated with fraud, but we suspect these are cases of overfitting. More interestingly, the bigram {\em described previously} was among the top-5 most highly correlated features, indicating fraud. From our syntactic treelets, proper nouns and interjections were both slightly indicative of fraud ($p<0.01$).
\paragraph{Other markers of non-fraud} From our syntactic treelets, conjunctions of numbers were indicative of non-fraud, suggesting maybe a higher level of technical detail. Non-fraud articles are also more likely to use the pronoun {\em they}, as compared to {\em we}, compared to fraud papers.
\begin{table}[t!]
\resizebox{\textwidth}{!}{
\begin{tabular}{llll}
\toprule
\multicolumn{3}{c}{Hedges} \\
assume & -0.121 & p=0.006 \\
practically & -0.118 & p=0.008 \\
occasionally & -0.112 & p=0.012 \\
conceivably & 0.089 & p=0.045 \\
assumed & -0.086 & p=0.052 \\
surely & 0.077 & p=0.083 \\
effectively & 0.075 & p=0.090 \\
presumably & 0.058 & p=0.195 \\
\midrule
\multicolumn{3}{c}{Inquirer} \\
compare & -0.158 & p=0.0003 \\
\midrule
\multicolumn{3}{c}{Explicit Relations lvl1} \\
comparison & -0.096 & p=0.031 \\
cause & 0.008 & p=0.863 \\
\midrule
\multicolumn{3}{c}{Connectives} \\
since & -0.102 & p=0.022 \\
therefore & 0.064 & p=0.147 \\
\midrule
\multicolumn{3}{c}{2-3-grams} \\
described previously & 0.115 & p=0.009 \\
\midrule
\multicolumn{3}{c}{Treelets} \\
intj & 0.126 & p=0.004 \\
propn & 0.110 & p=0.013 \\
\midrule
\multicolumn{3}{c}{Pronouns} \\
we & 0.071 & p=0.112 \\
they & -0.059 & p=0.182 \\
\end{tabular}}%
\caption{Pearson $\rho$ and original $p$-value (before Bonferroni correction) for some features.}
\label{tab:indicators}
\end{table}
\section{Conclusion}
We show that a simple unigram model outperforms previous work on scientific fraud detection. Overall, more high-level linguistic features, beyond syntactic treelets, do not lead to improvements, but we also presented a feature analysis showing, for example, that comparison and explanation (at the semantic and discourse level) are indicators of non-fraud, and that fraudulent writing uses slightly different hedging strategies.
|
\section{Introduction}
We consider a nonlocal model on a bounded domain involving the Riesz
fractional derivative (i.e., the fractional Laplacian).
For theory and numerical analysis of
general nonlocal models, we refer to the review paper \cite{DGLZ12} and references therein.
Particularly, several applications are modeled by partial differential equations involving the fractional
Laplacian: obstacle problems from symmetric $\alpha$-stable L\'evy processes \cite{Tankov03,Levendorskii04,Pham97};
image denoisings \cite{GH15}; fractional kinetics and anomalous transport \cite{Zaslavsky02};
fractal conservation laws \cite{Droniou10,BKW01}; and geophysical fluid dynamics \cite{CC03,Constantin02,CCW01,HPGS95}.
In this paper, we consider a class of fractional boundary problems on
bounded domains where
the fractional derivative comes from the fractional Laplacian defined on
all of $\mathbb R^d$. The motivation for these problems is illustrated by
an evolution equation considered by Meuller \cite{mueller} of the form:
\begin{align}
u_t &= -\widetilde\Lambda_su +f(u),\quad\text{in } {\mathbb R}^+\times
D,\label{eq1}\\
u&=0,\quad \hbox{ in } D^c. \label{eq2}
\end{align}
Here $D$ is a convex polygonal domain in
${\mathbb R}^d$, $D^c$ denotes its complement and
$$\widetilde \Lambda_s u := \left( (-\Delta)^s\widetilde u\right)|_D$$
with $\widetilde u $ denoting the extension of $u$ by
zero to ${\mathbb R}^d$. This fractional Laplacian on
${\mathbb R}^d$ is defined using the Fourier transform ${\mathcal F}$:
\begin{equation}
{\mathcal F}( (-\Delta)^s f)(\zeta) = |\zeta|^{2s} {\mathcal F}(f)(\zeta).
\label{Deldef}\end{equation}
The formula \eqref{Deldef} defines an unbounded operator $(-\Delta)^s$
on $L^2({\RR^d})$ with domain of definition
$$D((-\Delta)^s):=\{f\in L^2({\RR^d})\ : \
|\zeta|^{2s} {\mathcal F} ( f) \in L^2({\RR^d})\}.$$
It is clear that the Sobolev space
$$
H^{2s}({\RR^d}) := \left\lbrace f \in L^2({\RR^d}) \ : \ (1+|\zeta|^2)^{s} {\mathcal F}(f) \in L^2({\RR^d})\right\rbrace
$$
is a subset of $D((-\Delta)^s)$ for any $s\ge 0$. Note that $(-\Delta)^sv$ for
$s=1$ and $v\in H^2({\RR^d})$ coincides with the negative Laplacian applied
to $v$.
The term $-\widetilde\Lambda_s$ along with the ``boundary condition"
\eqref{eq2} represents the generator of a symmetric $s$-stable
L\'evy process which is killed when it exits $D$ (cf. \cite{mueller}). The
$f(u)$ term in \eqref{eq1} involves white noise and will be
ignored in this paper.
The goal of this paper is to study the numerical approximation of solutions of
partial differential equations on bounded domains involving the
fractional operator $\widetilde \Lambda_s$ supplemented with the boundary conditions \eqref{eq2}. As finite element approximations to parabolic
problems are based on approximations to the elliptic part, we shall
restrict our attention to the elliptic case, namely,
\begin{equation} \begin{aligned}
\widetilde \Lambda_s u &=f, \qquad \text{in } D,\\
u&=0, \qquad \text{in } D^c.\\
\end{aligned}
\label{mainP}
\end{equation}
The above system is sometimes referred to as the \emph{``integral"} fractional Laplacian problem.
We note that the variational formulation of \eqref{mainP} can be defined
in terms of the classical spaces
$\widetilde H^s(D)$ consisting of the functions defined in $D$ whose
extension by zero are in $H^s({\RR^d})$. This is to find $u\in
\widetilde H^s(D)$ satisfying
\begin{equation}
a(u,\phi) = \int_D f \phi\, dx, \qquad \hbox{for all } \phi \in \widetilde H^s(D),
\label{mainWP}
\end{equation}
where
\begin{equation}\label{e:variational1}
a(u,\phi) = \int_{{\RR^d}} [(-\Delta)^{s/2} \widetilde u ][(-\Delta)^{s/2}\widetilde\phi]\, dx
\end{equation}
with $\widetilde u$ and $\widetilde \phi$ denoting the extensions by 0.
We refer to Section~\ref{s:analytical} for the description of model problems.
The bilinear form $a(\cdot,\cdot)$ is obviously bounded on ${\tH^s}(D)\times {\tH^s}(D)$ and, as discussed in Section~\ref{s:notation}, it
is coercive on ${\tH^s}(D)$.
Thus, the Lax-Milgram Theory guarantees existence
and uniqueness.
We consider finite element approximations of \eqref{mainWP}. The use of
standard finite element approximation spaces of continuous functions
vanishing on $\partial D$ is the natural choice. The convergence
analysis is classical once the regularity properties of solutions to
problem \eqref{mainWP} are understood (regularity results for
\eqref{mainWP} have been studied in \cite{AB15} and
\cite{ros-oton}). However, the implementation of the resulting
discretization suffers from the fact that, for $d>1$, the entries of the
stiffness matrix, namely, $a(\phi_i,\phi_j)$, with $\{\phi_k\}$ denoting
the finite element basis, cannot be computed exactly.
When $d=1$, $s\in (0,1/2)\cup (1/2,1)$ and, for example, $D=(-1,1)$,
the bilinear form can be written
in terms of Riemann-Liouville fractional derivatives (cf. \cite{KST06}), namely,
\begin{equation} a(\phi_i,\phi_j) = \frac{(\partial_L^{s}\phi_i,\partial_R^{s}\phi_j)_D
+(\partial_L^{s}\phi_j,\partial_R^{s}\phi_i)_D}{ 2 \cos(s\pi)} . \label{onedstiff}
\end{equation}
Here $(\cdot,\cdot)_D$ denotes the inner product on $L^2(D)$ and for $t\in (0,1)$ and $v\in H^1_0(D)$,
the left-sided and right-sided Riemann-Liouville fractional derivatives of order $t$ are defined by
\begin{equation}
\partial^{t}_Lv(x) = \frac 1 {\Gamma(1-t)}\frac{d}{dt} \int_{-1}^x \frac{v(y)}{(x-y)^{t} }\, dy
\label{onedL}
\end{equation}
and
\begin{equation}
\partial^{t}_Rv(x) = \frac 1 {\Gamma(1-t)}\frac{d}{dt} \int_x^1 \frac{v(y)}{(x-y)^{t} }\, dy .
\label{onedR}
\end{equation}
Note that the integrals in \eqref{onedL} and \eqref{onedR} can be easily computed when $v$ is
a piecewise polynomial, i.e, when $v$ is a finite element basis
function. The computation of the stiffness matrix in this case reduces
to a coding exercise.
A representation of the fractional Laplacian for $d\ge 1$ is given by
\cite{ST10}:
\begin{equation}\label{e:PV}
((-\Delta)^{s} \eta)(x) = c_{d,{s}} PV \int_{\RR^d} \frac {\eta(x)-\eta(y)}
{|x-y|^{d+2s}}\,
dy, \qquad \hbox{for all } \eta\in \mathcal S ,
\end{equation}
where $\mathcal S$ denotes the Schwartz space of rapidly decreasing
functions on ${\RR^d}$,
$PV$ denotes the principle value and $c_{d,{s}}$ is a normalization
constant. It follows that for $\eta,\theta\in \mathcal S$,
\begin{equation}
a(\eta,\theta)= ((-\Delta)^{s} \eta,\theta)=
\frac {c_{d,{s}}}{2} \int_{{\RR^d}}\int_{{\RR^d}}
\frac{(\eta(x)-\eta(y))(\theta(x)-\theta(y))}{|x-y|^{d+2{s}}}
\, dy\, dx.
\label{schwpair}
\end{equation}
A density argument implies that the stiffness entries are given by
\begin{equation}\label{stiff_int}
a(\phi_i,\phi_j)=
\frac {c_{d,{s}}}{2} \int_{{\RR^d}}\int_{{\RR^d}}
\frac{(\widetilde \phi_i(x)-\widetilde\phi_i(y))(\widetilde\phi_j(x)-\widetilde\phi_j(y))}{|x-y|^{d+2{s}}}
\, dy\, dx,
\end{equation}
where again $\widetilde \phi$ denotes the extension of $\phi$ by zero outside $D$.
It is possible to apply the techniques developed for the approximation
of boundary integral stiffness matrices \cite{BEM} to deal with some of the issues
associated with the approximation of the double integral above,
namely, the
application of special techniques for handling the singularity and
quadratures. However, \eqref{stiff_int} requires additional
truncation techniques as the non-locality of the kernel implies a
non-vanishing integrand over ${\RR^d}$. These techniques are used to
approximate \eqref{stiff_int} in \cite{DG13,AB15}. In particular,
\cite{AB15} use their regularity theory to do {\it a priori} mesh
refinement near the boundary to develop higher order convergence
under the assumption of exact evaluation of the stiffness matrix.
The method to be developed in this paper is based on a
representation of the underlying bilinear form given in
Section~\ref{s:alternative}, namely, for $s\in (0,1)$, $0\le r \le s$,
$\eta\in H^r({\RR^d})$ and $\theta\in H^{s-r}({\RR^d})$,
\begin{equation}
\int_{\RR^d} [(-\Delta)^{r/2}\eta][(-\Delta)^{(s-r)/2} \theta] \, dx
= c_s \int_0^\infty t^{2-2s} ((-\Delta)(I-t^2\Delta)^{-1}
\eta,\theta)\, \frac {dt}t \label{intid}
\end{equation}
where $(\cdot,\cdot)$ denotes the inner product on $L^2({\RR^d})$ (see,
also \cite{bacuta-thesis}). We note
that for $t>0$, $( I-t^2\Delta)^{-1}$ is a bounded map of $L^2({\RR^d})$
into $H^2({\RR^d})$ so that the integrand above is well defined for
$\eta,\theta\in L^2({\RR^d})$. In Theorem~\ref{t:alternate1}, we show
that for $\eta\in H^r({\RR^d})$ and
$\theta\in H^{s-r}({\RR^d})$, the formula \eqref{intid} holds and the right hand side integral converges
absolutely.
It follows that the bilinear form $a(\cdot,\cdot)$ is given
by
\begin{equation}
a(\eta,\theta)
= c_s \int_0^\infty t^{2-2s} ((-\Delta)(I-t^2\Delta)^{-1}
\widetilde \eta,\theta)_D\, \frac {dt} t,\qquad \hbox{for all } \eta,\theta\in {\tH^s}(D) .
\label{intidD}
\end{equation}
There are three main issues needed to be addressed in developing numerical
methods for \eqref{mainWP} based on \eqref{intidD}:
\begin{enumerate}[(a)]
\item The infinite integral with respect to $t$ must be truncated and
approximated by numerical quadrature;
\item At each quadrature node $t_j$, the inner product term in the integrand involves an elliptic problem
on ${\RR^d}$.
This must be replaced by a problem with vanishing boundary condition on a bounded truncated
domain $\Omega^M(t_j)$ (defined below);
\item Using a fixed subdivision of $D$, we construct subdivisions of the larger domain $\Omega^M(t_j)$ which
coincide with that on $D$. We then replace the problems on
$\Omega^M(t_j)$ of (b) above by their finite element approximations.
\end{enumerate}
We address (a) above by first making the change of variable $t^{-2}=e^y$
which results in an integral over ${\mathbb R}$. We then apply a sinc
quadrature obtaining the approximate bilinear form
\begin{equation}
a^k(\eta,\theta):=\frac {c_s k} 2 \sum_{j=-N^-}^{N^+} e^{sy_j}
((-\Delta) (e^{y_j}I-\Delta)^{-1} \widetilde \eta,\theta)_D,\qquad \hbox{for all }
\theta,\eta\in L^2(D),
\label{quad}
\end{equation}
where $k$ is the quadrature spacing, $y_j=kj$, and $N^-$ and $N^+$ are
positive integers. Theorem~\ref{t:quad_error} shows that for
$\theta\in{\tH^s}(D)$ and $\eta\in \widetilde{H}^\delta(D)$ with
$\delta\in(s,2-s]$, we have
$$
\begin{aligned}
&|a(\eta,\theta)-a^k(\eta,\theta)| \\
&\quad\le C(\delta,s, \mathsf{d})\big[e^{-2\pi\mathsf{d}/k}
+e^{\mathbin{(s-\delta)}N^+k/2}+e^{-s kN^-}\big]
\|\eta\|_{\widetilde{H}^\delta(D)}
\|\theta\|_{\widetilde{H}^s(D)},
\end{aligned}
$$
where $0<\mathsf{d}<\pi$ is a fixed constant.
Balancing of the exponentials gives rise to an
$O(e^{-2\pi\mathsf{d}/k})$ convergence rate with the relation ${N^+}+{N^-}=O(1/k^2)$.
The size of the truncated domain $\Omega^M(t_j)$ in (b) is determined by the decay of $(e^{y_j}I-\Delta)^{-1} f$ for functions $f$ supported in $D$. For technical
reasons, we first extend $D$ to a bounded convex (star-shaped with
respect to the origin) domain
$\Omega$ and set (with $t_j=e^{-y_j/2}$)
$$
\Omega^M(t_j):= \left\lbrace
\begin{array}{ll}
\left\lbrace (1+t_j(1+M))x \ : \ x \in \Omega \right\rbrace, &\qquad t_j\geq 1\\
\left\lbrace (2+M)x \ : \ x \in \Omega \right\rbrace, &\qquad t_j < 1.
\end{array}\right.
$$
Let $\Delta_t$ denote the
unbounded operator on $L^2(\Omega^M(t))$ corresponding to the Laplacian
on $\Omega^M(t)$ supplemented with vanishing boundary condition. We define
the bilinear form $a^{k,M}(\eta,\theta)$ by replacing
$(-\Delta)(e^{y_j}I-\Delta)^{-1}$ in \eqref{quad} by $(-\Delta_{t_j}) (e^{y_j}I
-\Delta_{t_j})^{-1}$. Theorem~\ref{t:truncate} guarantees that for
sufficiently large $M$, we have
$$|a^k(\eta,\theta)-a^{k,M}(\eta,\theta)|\le
Ce^{-cM}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)},\quad\qquad \hbox{for all } \eta,\theta\in
L^2(D).$$
Here $c$ and $C$ are positive constants independent of $M$ and $k$. This addresses (b).
Step (c) consists in approximating $(-\Delta_{t_j}) (e^{y_j}I
-\Delta_{t_j})^{-1}$ using finite elements.
In this aim, we associate to a subdivision of $\Omega^M(t_j)$ the finite element space $\mathbb V_h^M(t_j)$ and the restriction $a^{k,M}_h(\cdot,\cdot)$ of $a^{k,M}(\cdot,\cdot)$ to $\mathbb V_h^M(t_j) \times \mathbb V_h^M(t_j)$.
As already mentioned, the subdivisions of $\Omega^M(t_j)$ are constructed to coincide on $D$.
Denoting by $\mathbb V_h(D)$ the set of finite element functions restricted to $D$ and vanishing on
$\partial D$, our approximation to the solution of \eqref{mainWP} is the function $u_h\in \mathbb V_h(D)$ satisfying
\begin{equation}
a_h^{k,M}(u_h,\theta) = \int_D f \theta\, dx,\qquad \hbox{for all } \theta\in \mathbb V_h(D).
\label{fem}
\end{equation}
Lemma~\ref{l:ellipticity} guarantees the $\mathbb V_h(D)$-coercivity of the bilinear form $a^{k,M}_h(\cdot,\cdot)$. Consequently, $u_h$ is well defined again from the Lax-Milgram theory.
Moreover, given, for every $t_j$, a sequence of quasi-uniform subdivisions of $\Omega^M(t_j)$, we show (Theorem~\ref{t:discrete_consistency}) that for $v$ in $\widetilde H^\beta(D)$ with $\beta\in (s,3/2)$ and for $\theta_h\in \mathbb V_h(D)$,
$$
|a^{k,M}(v_h,\theta_h)-a^{k,M}_h(v_h,\theta_h)|\le C(1+\ln (h^{-1})) h^{\beta-s}
\|v\|_{\widetilde H^ \beta(D)}\|\theta_h\|_{\widetilde H^s(D)}.
$$
Here $C$ is a constant independent of $M,k$ and $h$, and $v_h\in \mathbb V_h(D)$ denotes the Scott-Zhang interpolation or the $L^2$ projection of $v$ depending on whether $\beta \in (1,3/2)$ or $\beta \in (s,1]$.
Strang's Lemma implies that the
error between $u$ and $u_h$ in the ${\tH^s}(D)$-norm is bounded by the error of the best approximation in ${\tH^s}(D)$ and
the sum of the consistency errors from the above three steps (see Theorem~\ref{t:total_error}).
The online of the paper is as follows. Section~\ref{s:notation} introduces notations of Sobolev spaces followed by Section~\ref{s:dotted_space} introducing the dotted spaces associated with elliptic operators.
The alternative
integral representation of the bilinear form is given in Section~\ref{s:alternative}. Based on this integral representation,
we discuss the discretization of the bilinear form and the associated consistency error in three steps (Sections~\ref{s:quad}, \ref{s:truncated} and \ref{s:FEM}). The energy error estimate for the discrete problem is given in Section~\ref{s:FEM}.
A discussion on the implementation aspects of the method together with results of numerical experiments
illustrating the convergence of the method are provided in Section~\ref{s:numerical}.
We left to Appendix
the proof of technical result regarding the stability and approximability of the Scott-Zhang interpolant in nonstandard norms.
\section{Notations and Preliminaries}\label{s:notation}
\subsubsection*{Notation}
We use the notation $D\subset {\RR^d}$ to denote the polygonal domain with Lipschitz boundary in problem \eqref{mainWP} and
$\omega \subset {\RR^d}$ to denote a generic bounded Lipschitz domain.
For a function $\eta: \omega \rightarrow {\mathbb R}$, we denote by $\tilde \eta$ its extension by zero outside $\omega$.
We do not specify the domain $\omega$ in the notation $\tilde \eta$ as it will be always clear from the context.
\subsubsection*{Scalar Products}
We denote by $(\cdot,\cdot)_{\omega}$ the $L^2(\omega)$-scalar product and
by $\|\cdot\|_{L^2(\omega)}:= (\cdot,\cdot)_{\omega}^{1/2}$ the associated norm.
The $L^2({\RR^d})$-scalar product is denoted $(\cdot,\cdot)_{\RR^d}$.
To simplify the notation, we write in short $(\cdot,\cdot):=(\cdot,\cdot)_{\RR^d}$ and $\|\cdot\|:=\|\cdot\|_{L^2({\RR^d})}$.
\subsubsection*{Sobolev Spaces.}
For $r>0$, the
Sobolev space of order $r$ on ${\RR^d}$, $H^r({\RR^d})$, is defined to
be the set of functions $\theta\in L^2({\RR^d})$ such that
\begin{equation}
\|\theta\|_{H^r({\RR^d})} := \bigg(\int_{\RR^d} (1+|\zeta|^2)^{r/2}
|\mathcal F(\theta)(\zeta)|^2\, d\zeta\bigg)^{1/2} < \infty.
\label{halpharrd}
\end{equation}
In the case of bounded Lipschitz domains, $H^r(\omega)$ with $r \in (0,1)$, stands for the Sobolev space of
order $r$ on $\omega$. It is equipped with the Sobolev--Slobodeckij norm, i.e.
\begin{equation}
\|\theta\|_{H^r(\omega)}:= \big(\|\theta\|_{L^2(\omega)}^2 +
|\theta|_{H^r(\omega)}^2\big)^{1/2} ,
\label{hbetanorm}
\end{equation}
where
$$ |\theta|_{H^r(\omega)}^2:=\int_\omega \int_\omega \frac {(\theta(x)-\theta(y))^2}
{|x-y|^{d+2r}} \, dx\,dy.$$
When $r\in (1,2)$ instead, the norm in
$H^r(\omega)$ is given by
$$\|\theta\|_{H^r(\omega)}^2 := \|\theta\|_{H^1(\omega)}^2 +
\int_{\omega} \int_{\omega} \frac
{|\nabla \theta(x)-\nabla \theta(y)|^2} {|x-y|^{d+2(r-1)}} \, dx\, dy ,$$
where $\|w\|_{H^1(\omega)}:=(\|w\|_{L^2(\omega)}^2+\||\nabla w|\|_{L^2(\omega)}^2)^{1/2}$.
In addition, $H^1_0(\omega)$ denotes the set of functions in $H^1(\omega)$ vanishing at $\partial \omega$, the boundary
of $\omega$.
Its dual space is denoted $H^{-1}(\omega)$.
We note that when we replace $\omega$ with ${\RR^d}$ and $r\in[0,2)$, the norms using the double integral
above are equivalent with those in \eqref{halpharrd} (see e.g. \cite{lions,mclean}).
\subsubsection*{The spaces ${\tH^r}(D)$.}
For $r\in (0,2)$, the set of functions in $D$ whose extension by zero are in $H^s({\RR^d})$ is denoted $\widetilde H^r(D)$.
The norm of ${\tH^r}(D)$ is given by $\|\tilde{\cdot}\|_{H^r({\RR^d})}$.
Note that for $r\in(0,1)$, \eqref{e:PV} implies that for $\phi$ in the Schwartz space ${\mathcal S}$,
\begin{equation}
((-\Delta)^r \phi,\phi)=
{|c_{d,r}|} \int_{{\RR^d}}\int_{\RR^d} \frac {(\phi(x)-\phi(y))^2}
{|x-y|^{d+2r}}\,
dx\, dy.
\label{tartar}
\end{equation}
Thus, we prefer to use
\begin{equation}\label{d:Hsnorm}
\| \phi \|_{\widetilde H^r(D)}:= \bigg( {|c_{d,r}|} \int_{{\RR^d}}\int_{\RR^d} \frac
{(\widetilde \phi(x)-\widetilde \phi(y))^2}
{|x-y|^{d+2r}}\,
dx\, dy \bigg)^{1/2}
\end{equation}
as equivalent norm on $\widetilde H^r(D)$ for $r\in(0,1)$.
This is justified upon invoking a variant of the
Peetre-Tartar compactness
argument on ${\tH^r}(D) \subset H^r(D)$.
\subsubsection*{Coercivity.}
Since $C_0^\infty(D) $ is dense in ${\tH^s}(D)$ for $s\in (0,1)$ \cite{grisvard}, \eqref{schwpair} and a
density argument imply that for
$\eta,\theta\in {\tH^s}(D)$, we have
$$a(\eta,\theta) =\frac {c_{d,s}}{2} \int_{{\RR^d}}\int_{\RR^d}
\frac{(\widetilde \eta(x)-\widetilde \eta(y))(\widetilde \theta(x)-\widetilde\theta(y))}{|x-y|^{d+2s}}
\, dy\, dx.
$$
In turn, from the definition \eqref{d:Hsnorm} of the $\widetilde H^s(D)$ norm, we directly deduce the coercivity of $a(\cdot,\cdot)$ on
${\tH^s}(D)$
\begin{equation}\label{e:coercivity_a}
a(\eta,\eta) = \| \eta \|^2_{\widetilde H^s(D)}, \qquad \forall \eta \in \widetilde H^s(D).
\end{equation}
\subsubsection*{Dirichlet Forms}
We define the Dirichlet form on $H^1(\omega)\times H^1(\omega)$ to be
$$
d_{\omega}(\eta,\phi) := \int_{\omega} \nabla \eta \cdot \nabla \phi\, dx.
$$
On $H^1({\RR^d})\times H^1({\RR^d})$ we write
$$
d(\eta,\phi):=d_{{\RR^d}}(\eta,\phi) := \int_{{\RR^d}} \nabla \eta \cdot \nabla \phi\, dx.
$$
\section{Scales of interpolation spaces}\label{s:dotted_space}
We now introduce another set of functions instrumental in the analysis of the finite element method described in Section~\ref{s:truncated} and \ref{s:FEM}.
In this section $\omega$ stands for a bounded domain of ${\RR^d}$.
Given
$f\in L^2(\omega)$, we define $\theta \in H^1_0(\omega)$ to be the unique
solution to
\begin{equation}
(\theta,\phi)_{\omega} + d_{\omega}(\theta,\phi) =
(f,\phi)_{{\omega}},
\qquad \hbox{for all } \phi \in H^1_0({\omega})
\label{varscale}
\end{equation}
and define $T_{\omega} : L^2({\omega})\rightarrow H^1_0({\omega}) $ by
\begin{equation}\label{e:Tomega}
T_\omega f=\theta.
\end{equation}
As discussed in \cite{kato1961}, this defines a densely
defined unbounded
operator on $L^2({\omega})$, namely $L_{\omega} f:=T_{\omega}^{-1} f$ for $f$
in
$$D(L_{\omega}) := \{ T_{\omega} \phi\ : \ \phi\in L^2({\omega})\}.$$
The
operator $L_{\omega}$ is self-adjoint and positive so its fractional powers
define a Hilbert
scale of interpolation spaces, namely, for $r\ge 0$,
$$\dH r({\omega}) := D(L_{\omega}^{r/2})$$
with $D(L_{\omega}^r)$ denoting the domain of $L_{\omega}^r$. These are Hilbert spaces
with norms
$$\|w\|_{\dH r({\omega})} := \|L_{\omega}^{r/2} w\|_{L^2({\omega})}.$$
The space $\dH 1({\omega})$ coincides with $H^1_0({\omega})$ while $\dH 0(\omega)$ with $L^2(\omega)$, in both cases with equal norms.
Hence for $r\in[0,1]$, we have
$$\dH r({\omega})= (L^2({\omega}),H^1_0({\omega}))_{r,2},$$
where $(L^2({\omega}),H^1_0({\omega}))_{r,2}$ denotes the interpolation
spaces defined using the real method.
Another characterization of these spaces stems from Corollary 4.10 in \cite{chandler}, which states that for $r\in [0,1]$,
the spaces ${\tH^r}(\omega)$ are interpolation spaces.
Since $\widetilde H^1(\omega)=H^1_0(\omega)$ and
$\widetilde H^0(\omega)=L^2(\omega)$, ${\tH^s}(\omega)$ coincides with
$\dH r(\omega)$. In particular, we have
\begin{equation}\label{e:dot-tilde}
C^{-1} \|\theta\|_{\dH {r}(\omega)}\le \|\theta\|_{\widetilde H^{r}(\omega)} \le C \|\theta\|_{\dH {r}(\omega)},
\end{equation}
for a constant $C$ only depending on $\omega$.
The intermediate spaces can also be characterized by expansions in
the $L^2({\omega})$ orthonormal system of eigenvectors $\{\psi_i\}$ for
$T_{\omega}$, i.e.,
$$\dH {r}({\omega})= \bigg \{ \phi
\in L^2({\omega})\ : \sum_{i=1}^\infty \lambda_i
^r |(\phi,\psi_i)_{\omega} |^2 <\infty\bigg\}.$$
Here $\lambda_i=\mu_i^{-1}$ where $\mu_i$ is the eigenvalue of
$T_{\omega}$ associated with $\psi_i$.
In this case, we find that
$$
\|\phi\|^2_{\dH r({\omega})} =\|L_{\omega}^{r/2}\phi\|^2_{L^2({\omega})} =\sum_{i=1}^\infty \lambda_i
^r |(\phi,\psi_i)_{\omega} |^2$$
and for $r\in (0,1)$, (see, e.g., \cite{00BZ})
$$\|\phi\|^2_{\dH r({\omega})}=\frac{2\sin\pi r}\pi \int_0^\infty t^{-2r} K_\omega(\phi,t) \frac
{dt}t.
$$
Here
$$K_\omega(\phi,t) :=\inf _{w\in H^1_0({\omega}) } (\|\phi-w\|_{L^2({\omega})}^2+t^2
\|w\|_{H^1({\omega})}^2).
$$
Note that if ${\omega}' \subset {\omega}$ then since the extension of a
function $\phi$ in $H^1_0({\omega}')$ by zero is in $H^1_0({\omega})$, the K-functional
identity implies that for all $r\in [0,1]$,
\begin{equation}
\|\widetilde \phi\|_{\dH r({\omega})}\le \|\phi\|_{\dH r({\omega}')},
\label{extenhs}
\end{equation}
where $\widetilde \phi$ denotes the extension by zero of $\phi$ outside $\omega'$.
The operator $T_{\omega}$ extends naturally to $F\in H^{-1}({\omega})$ by setting
$T_{\omega} F=u$ where $u\in H^1_0({\omega})$ is the solution of
\eqref{varscale} with $(f,\phi)_{\omega}$ replaced by $\langle F,\phi
\rangle $. Here $\langle \cdot,\cdot\rangle $ denotes the
functional-function pairing.
Identifying $f\in L^2({\omega})$ with the
functional $\langle F,\phi \rangle := (f,\phi)_{\omega}$, we define the
intermediate spaces for $r\in (-1,0) $ by
$$\dH {r}({\omega}) := (H^{-1}({\omega}),L^2({\omega}))_{1+r,2}$$
and set $\dH {-1}:= H^{-1}(\omega)$.
Since $T_{\omega}$ maps $H^{-1}({\omega})$ isomorphically onto $\dH 1({\omega})$ and
$L^2({\omega})$ isomorphically onto $\dH 2({\omega})$,
$T_{\omega}$ maps $\dH {-r}({\omega})$ isometrically onto $\dH {2-r}(\omega)$ for
$r\in [0,1]$.
Functionals in $H^{-1}({\omega})$ can also be characterized in terms of
the eigenfunctions of $T_{\omega}$, indeed,
$H^{-1}({\omega})$ is the set of linear functionals $F$ for which the sum
$$\sum_{i=1}^\infty \lambda_i^{-1} |\langle F,\psi_i \rangle |^2 $$
is finite. Moreover,
$$\|F\|_{H^{-1}({\omega})} = \sup_{\theta\in H^1_0({\omega})} \frac {\langle
F,\theta \rangle} {\|\theta\|_{H^1({\omega})}}= \bigg
(\sum_{i=1}^\infty \lambda_i^{-1} |\langle F,\psi_i \rangle |^2 \bigg
)^{1/2}$$
for all $ F\in H^{-1}({\omega})$.
This implies that for $r\in [-1,0]$,
$$\dH r({\omega}) = \{ F\in \dH {-1}\ : \ \sum_{i=1}^\infty \lambda_i^{r}
|\langle F,\psi_i \rangle |^2<\infty\}$$
and
$$ \|F\|_{\dH r(\omega)} = \bigg(\sum_{i=1}^\infty \lambda_i^{r}
|\langle F,\psi_i \rangle |^2\bigg)^{1/2}.$$
\begin{remark}[Norm equivalence for Lipschitz domains]\label{r:beta-equiv}
For $r\in(1,3/2)$, it is known that ${\tH^r}(\omega)=H^r(\omega)\cap H^1_0(\omega)$.
On the other hand, we note that when $\partial\omega$
is Lipschitz, $-\Delta$ is an isomorphism
from $H^r(\omega)\cap H^1_0(\omega)$ to $\dot H^{r-2}(\omega)$; see Theorem 0.5(b) of \cite{JK95}.
We apply this regularity result into Proposition 4.1 of \cite{BP13}
to obtain $H^r(\omega)\cap H^1_0(\omega)=\dot H^r(\omega)$. So
the norms of ${\tH^r}(\omega)$ and $\dot H^r(\omega)$ are equivalent for $r\in[0,3/2)$
and the equivalence constant may depend on $\omega$. In what follows,
we use ${\tH^r}(D)$ to describe the smoothness of functions defined on $D$.
When functions defined on a larger domain (see Section~\ref{s:truncated} and \ref{s:FEM}),
we will use these interpolation spaces separately so that we can investigate
the dependency of constants.
\end{remark}
We end the section with the following lemma:
\begin{lemma} \label{Tshift}
Let $a$ be in $[0,2]$ and $b$ be in $[0,1]$
with $a+b\le 2$. Then for
$\mu\in (0,\infty)$, we have
$$\| (\mu I + T_{\omega})^{-1} \phi \|_{\dH {-b}(\omega) } \le \mu^{(a+b)/2-1}
\|\phi \|_{\dH a({\omega})},\qquad \hbox{for all } \phi\in \dH a({\omega}).$$
\end{lemma}
\begin{proof} Let $\phi$ be in $\dH a({\omega})=D(L_\omega^{a/2})$.
Setting $\theta:=L_{\omega}^{a/2} \phi \in L^2(\omega)$, it suffices to prove that
\begin{equation}
\| (\mu I + T_{\omega})^{-1} T_{\omega}^{a/2} \theta \|_{\dH {-b}(\omega) } \le \mu^{(a+b)/2-1}
\|\theta \|_{L^2({\omega})},\qquad \hbox{for all } \theta\in L^2(\omega).\label{rewrite}
\end{equation}
The operator $T_{\omega}$ and its fractional powers are symmetric in
the $L^2({\omega})$ inner product.
Therefore, we have
$$\begin{aligned} \| (\mu I + T_{\omega})^{-1} T_{\omega}^{a/2} \theta \|^2_{\dH {-b}(\omega) } &=
\sum_{i=1}^\infty |( (\mu I + T_{\omega})^{-1} T_{\omega}^{a/2}
\theta,\psi_i)_{\omega}|^2 \lambda_i^{-b}\\
&=\sum_{i=1}^\infty \frac {\lambda_i^{-a-b}} {(\mu +
\lambda_i^{-1})^2} |(\theta,\psi_i)_{\omega}|^2.
\end{aligned}
$$
Inequality \eqref{rewrite} follows from Young's inequality
$$\lambda_i^{-(a+b)/2} \mu^{1-(a+b)/2} {(\mu +
\lambda_i^{-1})}\le 1.
$$
\end{proof}
\section{An Alternative Integral Representation of the Bilinear Form}\label{s:alternative}
The goal of this section is to derive the integral expression
\eqref{intid} and some of its properties.
\begin{theorem}[Equivalent Representation] \label{t:alternate1}
Let $s\in(0,1)$ and $0\le r\le s$. For $\eta\in H^{s+r}({\RR^d})$ and
$\theta\in H^{s-r} ({\mathbb R}^d)$,
\begin{equation}
((-\Delta)^{(s+r)/2} \eta,(-\Delta)^{(s-r)/2}\theta)= c_s \int _0^\infty t^{2-2s} (-\Delta (I-t^2\Delta)^{-1}\eta,\theta)\,
\frac {dt}t,
\label{is0}
\end{equation}
where
\begin{equation}\label{e:cs}
c_s:=\bigg( \int_0^\infty \frac {y^{1-2s}} {1+y^2}\, dy\bigg)^{-1}
=\frac{2\sin(\pi s)}{\pi}.
\end{equation}
\end{theorem}
\begin{proof}
Let $I(\eta,\theta)$ denotes the right hand side of \eqref{is0}.
Parseval's theorem implies that
\begin{equation}
(-\Delta (I-t^2\Delta)^{-1}\eta,\theta)=
\int_{\RR^d} \frac {|\zeta|^2}{1+t^2 |\zeta|^2} {\mathcal F}(
\eta)(\zeta)\overline{{\mathcal F}(
\theta)(\zeta)} \, d\zeta.
\label{ftid}
\end{equation}
and so
\begin{equation} I(\eta,\theta)
=c_s \int _0^\infty t^{1-2s}
\int_{\RR^d} \frac {|\zeta|^2}{1+t^2 |\zeta|^2} {\mathcal F}(
\eta)(\zeta) \overline{{\mathcal F}(
\theta)(\zeta)}\, d\zeta\, {dt}.
\label{firsti}
\end{equation}
In order to invoke Fubini's theorem, we now show that
$$
c_s
\int_{\RR^d} \int _0^\infty t^{1-2s} \frac {|\zeta|^2}{1+t^2 |\zeta|^2} |{\mathcal F}(
\eta)(\zeta)|\ | {\mathcal F}(
\theta)(\zeta)|\, d\zeta\, {dt} < \infty.
$$
Indeed, the change of variable $y=t|\zeta|$ and the definition \eqref{e:cs} of $c_s$ implies that the above integral is equal to
\begin{equation*}
\begin{split}
c_s\int_{\RR^d} |{\mathcal F}(
\eta)(\zeta)||{\mathcal F}(
\theta)(\zeta)| \int _0^\infty t^{1-2s}
\frac {|\zeta|^2}{1+t^2 |\zeta|^2} \, dt\, d\zeta = \int_{\RR^d}|\zeta |^{2s} |{\mathcal F}(
\eta)(\zeta)|\, |{\mathcal F}(
\theta)(\zeta)| \, \, d\zeta,
\end{split}
\end{equation*}
which is finite for $\eta\in H^r({\RR^d})$ and
$\theta\in H^{s-r} ({\mathbb R}^d)$.
We now apply Fubini's theorem and the same change of variable $y=t|\zeta|$ in \eqref{firsti} to arrive at
$$\begin{aligned} I(\eta,\theta)
&=\int_{\RR^d} |\zeta| ^{2s} {\mathcal F}(
\eta)(\zeta) \overline{{\mathcal F}(
\theta)(\zeta)} \, d\zeta=
((-\Delta)^{(s+r)/2} \eta,(-\Delta)^{(s-r)/2}\theta).
\end{aligned}$$
This completes the proof.\end{proof}
Theorem~\ref{t:alternate1} above implies that for $\eta,\theta$ in ${\tH^s}(D)$,
\begin{equation}\begin{aligned} a(\eta,\theta) &= c_s \int _0^\infty t^{-2s} (w(\widetilde\eta,t),\theta)_D\,
\frac {dt}t,
\end{aligned}
\label{is}
\end{equation}
where for $\psi \in L^2({\RR^d})$
$$w(t):=w(\psi,t):= -t^2 \Delta
(I-t^2\Delta)^{-1}\psi .$$
Examining the Fourier transform of $w(\psi,t)$, we realize that
$w(t):=w(\psi,t):=\psi+v(\psi,t)$
where $v(t):=v(\psi,t) \in H^1({\RR^d}) $ solves
\begin{equation}
(v(t),\phi)+t^2 d(v(t),\phi)=-(\psi,\phi),\qquad \hbox{for all } \phi \in
H^1({\RR^d}) .
\label{vt}
\end{equation}
The integral in \eqref{is} is the basis of a numerical method for
\eqref{mainWP}. The following lemma, instrumental in our analyze, provides an alternative
characterization for the inner product appearing on the right hand side
of \eqref{is}.
\begin{lemma}\label{l:K} Let $\eta$ be in $L^2({\RR^d})$. Then,
\begin{equation}
(w(\eta,t),\eta)= \inf_{\theta\in H^1({\RR^d}) } \{ \|\eta-\theta\|^2 +t^2
d(\theta,\theta)\}=: K(\eta,t).
\label{kmin1}
\end{equation}
\end{lemma}
\begin{proof}
Let $\eta$ be in $L^2({\RR^d})$.
We start by observing that for any positive $t$ and $\zeta\in {\RR^d}$,
$$\hat \phi(\zeta):= \frac{ {\mathcal F}(\eta) (\zeta)} {1+t^2|\zeta|^2} $$
solves the minimization problem
$$\inf_{z\in {\mathbb C}} \{|{\mathcal F}(\eta)(\zeta)-z|^2+t^2 |\zeta|^2 |z|^2\}$$
and so
\begin{equation}
\inf_{z\in {\mathbb C}} \{|{\mathcal F}(\eta)(\zeta)-z|^2+t^2 |\zeta|^2 |z|^2\}=
\frac {t^2 |\zeta|^2} {1+t^2|\zeta|^2} |{\mathcal F}(\eta)(\zeta)|^2.
\label{point}
\end{equation}
We
denote $\phi$
to be the inverse Fourier transform of $\hat \phi$.
Note that $\phi$ is in
$H^1({\RR^d})$ (actually, $\phi$ is in $ H^2({\RR^d})$).
Applying the Fourier transform, we find that
\begin{equation}
K(\eta,t)=\inf_{\theta\in H^1({\RR^d}) } \int_{\RR^d}
(|{\mathcal F}(\eta)(\zeta)- {\mathcal F}(\theta)(\zeta)|^2 +t^2 |\zeta|^2 |{\mathcal F}(\theta)(\zeta)|^2)\,
d\zeta.
\label{kmin2}
\end{equation}
Now, $\phi$ is the pointwise minimizer of the integrand in
\eqref{kmin2} and since $\phi \in H^1({\RR^d})$, it is also the minimizer of \eqref{kmin1}.
In addition, \eqref{point}, \eqref{kmin2} and \eqref{ftid}
imply that
$$ K(\eta,t)= \int_{\RR^d} \frac {t^2 |\zeta|^2} {1+t^2|\zeta|^2} |{\mathcal F}(\eta)(\zeta)|^2
\, d\zeta = (w(\eta,t),\eta).
$$
This completes the proof of the lemma.
\end{proof}
\begin{remark}[Relation with the vanishing Dirichlet boundary condition case] \label{r:bounded_K}
The above lemma implies that for $\eta\in {\tH^s}(D)$,
$$ a(\eta,\eta)
= c_s \int _0^\infty t^{-2s} K(\widetilde \eta,t)\,
\frac {dt}t.
$$
It is observed in the Appendix of \cite{00BZ} that for any bounded domain
$\omega$, and $\eta\in
(L^2(\omega),H^1_0(\omega))_{s,2}$, the real interpolation space between $L^2(\omega)$ and $H^1_0(\omega)$, we have
$$\|\eta \|_{(L^2(\omega),H^1_0(\omega))_{s,2}}^2= c_s \int
_0^\infty t^{-2s} K_{\omega}^0 (\eta,t)\,
\frac {dt}t$$
where
\begin{equation}
K_{\omega}^0(\eta,t) :=\inf_{\theta\in H_0^1(\omega) } \{
\|\eta-\theta\|_{L^2(\omega)}^2 +t^2
d_{\omega}(\theta,\theta)\}.
\label{infOmega}
\end{equation}
Let $\{\psi_i^0\}\subset H^1_0(\omega)$ denote the $L^2(\omega)$-orthonormal basis of eigenfunctions satisfying
$$d_{\omega}(\psi_i^0,\theta)= \lambda_i (\psi_i^0,\theta)_{\omega},\qquad \hbox{for all } \theta\in
H^1_0(\omega).$$
As the proof in Lemma~\ref{l:K} but using the expansion in the above
eigenfunctions,
it is not hard to see that
\begin{equation}
(w_{{\omega}}(\eta,t),\eta)_{\omega} = K_{\omega}^0(\eta,t)
\label{womegaid}
\end{equation}
with $w_{{\omega}}(\eta,t)=\eta+v$ and $v
\in H^1_0({\omega})$ solving
$$(v,\theta)_{\omega} +t^2 d_{\omega}(v,\theta)=-(u,\theta)_{\omega},\qquad \hbox{for all }
\theta \in H^1_0({\omega}).$$
This means that if $\eta\in L^2(\omega)$, $K(\widetilde\eta,t)\le K_\omega^0(\eta,t)$ and hence
$$
(w(\widetilde\eta,t),\eta)_\omega\le (w_{{\omega}}(\eta,t),\eta)_{\omega}.$$
\end{remark}
\section{Exponentially Convergent Sinc Quadrature}\label{s:quad}
In this section, we analyze a sinc quadrature scheme applied to the integral \eqref{is}.
Notice that the analysis provided in \cite{BLP18} does not strictly apply in the present context.
\subsection{The Quadrature Scheme}
We first use the change of variable $t^{-2} = e^y$ so that \eqref{is} becomes
$$
a(\eta,\theta)= \frac{c_s}{2} \int _{-\infty}^\infty e^{sy} (w(\widetilde \eta,t(y)),\theta)_D \, dy.
$$
Given a quadrature spacing $k>0$ and two positive integers ${N^-}$ and ${N^+}$, set
$y_j:=j k$ so that
\begin{equation}
\label{e:tj}
t_j = e^{-y_j/2} = e^{-jk/2}
\end{equation}
and define the approximation of $a(\eta,\theta)$ by
\begin{equation}
a^k(\eta,\theta):=\frac{c_s k}{2} \sum_{j=-{N^-}}^{N^+} e^{sy_j} (w(\widetilde \eta,t_j),\theta)_D.
\label{ik}
\end{equation}
\subsection{Consistency Bound}
The convergence of the sinc quadrature depends on the properties of the integrand
\begin{equation}
g(y;\eta,\theta):=e^{s y} (w(\widetilde \eta,t(y)),\theta)_D=e^{sy}\left( -\Delta(e^yI-\Delta)^{-1}\widetilde \eta,\widetilde\theta \right) .
\label{integrant}
\end{equation}
More precisely, the following conditions are required:
\begin{enumerate}[(a)]
\item $g(\cdot;\eta,\theta)$ is an analytic function in the band
$$
B=B(\mathsf{d}):=\left\{z=y+iw \in \mathbb C : \ |w|< \mathsf{d}\right\} ,$$
where $\mathsf{d}$ is a fixed constant in $(0, \pi)$.
\item There exists a constant $C$ independent of $y\in{\mathbb R}$ such that
$$
\int_{-\mathsf{d}}^{\mathsf{d}} |g(y+iw;\eta,\theta)|\, dw\leq C;
$$
\item
$$N(B):=\int_{-\infty}^\infty \left(|g(y+i\mathsf{d};\eta,\theta)|+|g(y-i\mathsf{d};\eta,\theta)| \right) dy < \infty .$$
\end{enumerate}
In that case, there holds (see Theorem 2.20 of \cite{sinc_int})
\begin{equation}
\label{inf}
\bigg|\int_{-\infty}^\infty g(y;\eta,\theta)\, dy-k\sum_{j=-\infty}^\infty g(kj;\eta,\theta)\bigg|\le \frac{N(B)}{e^{2\pi \mathsf{d}/k}-1}.
\end{equation}
In our context, this leads to the following estimates for the sinc quadrature error.
\begin{theorem}[Sinc quadrature]\label{t:quad_error}
Suppose $\theta \in{\tH^s}(D)$ and $\eta \in \widetilde{H}^\delta(D)$ with $\delta\in (s,2-s]$.
Let $a(\cdot,\cdot)$ and $a^k(\cdot,\cdot)$ be defined by \eqref{mainWP} and \eqref{ik}, respectively. Then
we have
\begin{equation}
\begin{aligned}
|a(\eta,\theta)-a^k(\eta,\theta)|
&\le \frac{2c(\mathsf{d})}{\delta-s} \left( \frac{2}{e^{2\pi\mathsf{d}/k}-1} + e^{(s-\delta){N^+} k/2}\right)\|\eta\|_{\widetilde{H}^\delta(D)} \|\theta\|_{\widetilde{H}^s(D)} \\
& + \frac{c(\mathsf{d})}{s} \left( \frac{2}{e^{2\pi\mathsf{d}/k}-1} + e^{-s {N^-} k}\right) \|\eta \|_{L^2(D)}\|\theta\|_{L^2(D)},
\end{aligned}
\label{quadbound}
\end{equation}
where $c(\mathsf{d}):=\frac{1}{\sqrt{(1+\cos \mathsf{d})/2}}$.
\end{theorem}
\begin{proof}
We start by showing that the conditions (a), (b) and (c) hold.
For (a), we note that $g(\cdot;\eta,\theta)$ in analytic on $B$ if and only if the operator mapping
$z \mapsto (e^{z}I-\Delta)^{-1}$ is analytic on $B$.
To see the latter, we fix $z_0 \in B$ and set $p_0 := e^{z_0}$.
Clearly,
$p_0 I-\Delta$ is invertible from $L^2({\RR^d})$ to $L^2({\RR^d})$. Let $M_0:= \| (p_0 I -\Delta)^{-1} \|_{L^2(\mathbb R^d) \to L^2(\mathbb R^d)}$.
For $p \in \mathbb C$, we write
$$
pI-\Delta = (p-p_0)I + (p_0I -\Delta) = (p_0I-\Delta) \left( (p-p_0)(p_0I-\Delta)^{-1} + I \right),
$$
so that the Neumann series representation
$$
(pI-\Delta)^{-1} = \left(\sum_{j=0}^\infty (-1)^j (p-p_0)^j (p_0I-\Delta)^{-j}\right)(p_0I-\Delta)^{-1}
$$
is uniformly convergent provided $\| (p-p_0) (p_0I-\Delta)^{-1}\|_{L^2(\mathbb R^d) \to L^2(\mathbb R^d)} < 1$ or
$$
| p - p_0 | < 1/M_0.
$$
Hence $(pI-\Delta)^{-1}$ is analytic in an open neighborhood of $p_0=e^{z_0}$ for all $p_0 \in B$ and (a) follows.
To prove (b) and (c), we first bound $g(z;\eta,\theta)$ for $z$ in the band $B$.
Assume $\eta\in \widetilde{H}^\beta(D)$ and $\theta \in{\tH^s}(D)$ with $\beta>s$.
For $z\in B$, we use the Fourier transform and estimate $|g|$ as follows
$$
\begin{aligned}
|g(z;\eta,\theta)|&= \left| e^{sz}\int_{{\RR^d}} \frac{|\zeta|^2}{e^z+|\zeta|^2} \mathcal F(\widetilde \eta) \overline{ \mathcal F(\widetilde \theta)} \, d\zeta \right|\\
&\le c(\mathsf{d}) e^{s\Re z} \int_{\mathbb{R}^d}\frac{|\zeta|^2}{e^{\Re z}+|\zeta|^2}|\mathcal F(\widetilde \eta)| | \mathcal F(\widetilde \theta)| \, d\zeta ,\\
\end{aligned}
$$
where $c(\mathsf{d})=\frac{1}{\sqrt{(1+\cos \mathsf{d})/2}}$ and upon noting that
$$
|e^z+|\zeta|^2|\ge c(\Im z)^{-1}(e^{\Re z}+|\zeta|^2)\ge c(\mathsf{d})^{-1}(e^{\Re z}+|\zeta|^2).
$$
If $\Re z<0$, we deduce that
\begin{equation}
|g(z;\eta,\theta)|\le c(\mathsf{d})e^{s\Re z}\|\eta \|_{L^2(D)}\|\theta\|_{L^2(D)}.
\label{yl0}
\end{equation}
Instead, when $\Re z\geq 0$, we write
$$
|g(z;\eta,\theta)|\le c(\mathsf{d})e^{(s-\delta)\Re z/2} \int_{\mathbb{R}^d} \frac{(|\zeta|^{2})^{1-(\delta+s)/2}(e^{\Re z})^{(\delta+s)/2}}{e^{\Re z}+|\zeta|^2}|\zeta|^{\delta+s}|\mathcal F(\widetilde \eta)| | \mathcal F(\widetilde \theta)| \, d\zeta.
$$
Whence, Young's inequality guarantees that
\begin{equation}
|g(z;\eta,\theta)|\le c(\mathsf{d})e^{(s-\delta)\Re z/2}\|\eta\|_{\widetilde{H}^\delta(D)} \|\theta\|_{\widetilde{H}^s(D)}.
\label{yg0}
\end{equation}
Gathering the above two estimates \eqref{yl0} and \eqref{yg0} gives
\begin{equation}
\int_{-\mathsf{d}}^{\mathsf{d}} |g(y+iw;\eta,\theta)|\, dw\leq 2\mathsf{d} c(\mathsf{d})
\left\{ \begin{aligned}
\|\eta \|_{L^2(D)}\|\theta\|_{L^2(D)}, & \qquad y<0,\\
\|\eta\|_{\widetilde{H}^\delta(D)} \|\theta\|_{\widetilde{H}^s(D)}, &\qquad y\ge 0,
\end{aligned} \right.
\label{b-bound}
\end{equation}
and $N(B)$ in \eqref{inf} satisfies
\begin{equation}
N(B)\le c(\mathsf{d})(\frac{4}{\delta-s}\|\eta\|_{\widetilde{H}^\delta(D)} \|\theta\|_{\widetilde{H}^s(D)}
+\frac{2}{s}\|\eta \|_{L^2(D)}\|\theta\|_{L^2(D)}).\label{c-bound}
\end{equation}
Estimates \eqref{b-bound} and \eqref{c-bound} prove (b) and (c) respectively.
Having established (a), (b), and (c), we can use the sinc quadrature estimate \eqref{inf}.
In addition, from \eqref{yl0} and \eqref{yg0} we also deduce that
\begin{equation}\label{tail}
\begin{aligned}
k\sum_{j\le -{N^-}-1}^{-\infty} |g(kj;\eta,\theta)| &\le
\frac{c(\mathsf{d})}{s}e^{-s {N^-} k}\|\eta \|_{L^2(D)}\|\theta\|_{L^2(D)} \qquad\text{and }
\\ k\sum_{j\ge{N^+}+1}^\infty |g(kj;\eta,\theta)| &\le
\frac{2c(\mathsf{d})}{\delta-s}e^{(s-\delta){N^+} k/2}\|\eta\|_{\widetilde{H}^\delta(D)} \|\theta\|_{\widetilde{H}^s(D)} .
\end{aligned}
\end{equation}
Combining \eqref{inf} with \eqref{c-bound} and \eqref{tail} shows \eqref{quadbound} and completes the proof.
\end{proof}
\begin{remark}[Choice of $N^-$ and $N^+$]\label{r:balanced} Balancing the three exponentials in \eqref{quadbound} leads to the following choice
$$2\pi\mathsf{d}/k\approx (\delta-s){N^+} k/2 \approx s{N^-} k .$$
Hence, for given the quadrature spacing $k>0$, we set
\begin{equation}\label{e:choiceN}
{N^+} := \bigg\lceil\frac{4\pi\mathsf{d}}{k^2(\delta-s)}\bigg\rceil\qquad\hbox{and}\qquad {N^-}:=\bigg\lceil\frac{2\pi\mathsf{d}}{sk^2}\bigg\rceil.
\end{equation}
With this choice, \eqref{quadbound} becomes
\begin{equation}\label{e:quad_cons}
|a(\eta,\theta)-a^k(\eta,\theta)| \leq \gamma(k) \| \eta \|_{\widetilde{H}^\delta(D)} \| \theta \|_{\widetilde H^s(D)}
\end{equation}
where
\begin{equation}\label{e:gammak}
\gamma(k) := C\left(\frac{1}{\delta-s},\frac{1}{s},\mathsf{d}\right)e^{-2\pi\mathsf{d}/k}.
\end{equation}
\end{remark}
\section{Truncated Domain Approximations}\label{s:truncated}
To develop further approximation to problem \eqref{mainWP} based on the sinc quadrature approximation \eqref{ik}, we replace \eqref{vt}
with problems on bounded domains.
\subsection{Approximation on Bounded Domains} \label{s:truncated_setting}
Let $\Omega$ be a convex bounded domain containing $D$ and the origin.
Without loss of generality, we assume that the diameter of $\Omega$ is 1.
This auxiliary domain is used to generate suitable truncation domains to approximate the solution of \eqref{vt}.
We introduce a domain parameter $M>0$ and define the dilated domains
\begin{equation}\label{e:omega_Mt}
\Omega^M(t):= \left\lbrace
\begin{array}{ll}
\left\lbrace y = (1+t(1+M))x \ : \ x \in \Omega \right\rbrace, &\qquad t\geq 1,\\
\left\lbrace y = (2+M)x \ : \ x \in \Omega \right\rbrace, &\qquad t < 1.
\end{array}\right.
\end{equation}
The approximation of $a^k(\cdot,\cdot)$ in \eqref{ik} reads
\begin{equation}
a^{k,M}(\eta,\theta):= \frac{c_s k}{2}\sum_{j=-{N^-}}^{N^+} e^{\beta y_j}(w^M(\widetilde\eta, t_j),\theta)_{ {D}},
\label{isM}
\end{equation}
with $t_j :=t(y_j)=e^{-y_j/2}$, according to \eqref{e:tj}, and
\begin{equation}\label{wMt}
w^{M}(t):=w^M(\widetilde \eta,t)=\widetilde\eta|_{\Omega^M(t)}+v^M(\widetilde\eta,t),
\end{equation}
where $v^M(t):=v^M(\widetilde\eta,t)$ solves
\begin{equation}
(v^M(t),\phi)_{\Omega^M(t)}+t^2 d_{\Omega^M(t)}(v^M(t),\phi)=-(\eta,\phi)_D,\qquad \hbox{for all } \phi \in
H^1_0(\Omega^M(t));
\label{vtM}
\end{equation}
compare with \eqref{vt}.
The domains $\Omega^M(t_j)$ are constructed for the truncation error to be exponentially decreasing as a function of $M$. This is the subject of next section.
\subsection{Consistency}\label{s:consistency_truncated}
The main result of this section provides an estimate for $a^k-a^{k,M}$.
It relies on decay properties of $v(\widetilde \eta,t)$ satisfying \eqref{vt}.
In fact, Lemma 2.1 of \cite{auscher2002solution}
guarantees the existence of universal constants $c$ and
$C$ such that
\begin{equation}\label{ineq:kato_esti}
t\|\nabla v(\widetilde \eta,t)\|_{L^2(B^M(t))}+\|v(\widetilde \eta,t)\|_{L^2(B^M(t))}\leq Ce^{-\max(1,t)cM/t }\|\eta\|_{L^2(D)},
\end{equation}
provided $\eta \in L^2(D)$ and $v(t):= v(\widetilde \eta,t)$ is given in \eqref{vt}.
Here
$$
B^M(t):= \{x\in \Omega^M(t) \ :\
\hbox{dist}(x,\partial\Omega^M(t))<t\}
$$
so that the minimal distance between points in $D\subset \Omega$ and $B^M(t)$ is greater than $M\max(1,t)$.
An illustration of the different domains is provided in Figure~\ref{f:domains}.
\begin{figure}[ht!]
\begin{center}
\scalebox{1.0}{
\includegraphics[width=0.5\textwidth]{domains-eps-converted-to.pdf}
\begin{picture}(0,0)
\put(0,25){$B^M(t)$}
\put(-100,75){$D$}
\put(-75,110){$\Omega$}
\put(-20,150){$\Omega^M(t)$}
\end{picture}
}
\caption{Illustration of the different domains in $\mathbb R^2$. The domain of interest $D$ is a L-shaped domain, $\Omega \subset \Omega^M(t)$ are interior of discs, and $B^M(t)$ is the filled portion of $\Omega^M(t)$.}\label{f:domains}
\end{center}
\end{figure}
\begin{lemma}[Truncation error] \label{l:deday}
Let $\eta \in L^2(D)$, $e(t):=v(\tilde \eta, t)-v^M(\tilde \eta,t)$ and $c$ be the constant appearing in \eqref{ineq:kato_esti}.
There is a positive constant $C$ not depending
on $M$ and t satisfying
\begin{equation}\label{domainbound2}
\|e(t)\|_{L^2(\Omega^M(t))}\leq Ce^{-\max(1,t)cM/t }\|\eta\|_{L^2(D)} .
\end{equation}
\end{lemma}
\begin{proof}
In this proof $C$ denotes a generic constant only depending on $\Omega$.
Note that $e(t)$ satisfies the relations
\begin{equation}
\begin{aligned}
(e(t),\phi)+t^2d_{\Omega^M(t)}(e(t),\phi)&=0, \quad \forall \phi\in H^1_0(\Omega^M(t)),\\
e(t)&=v(t),\quad\hbox{on }\partial\Omega^M(t).
\end{aligned}
\label{etform}
\end{equation}
Let $\chi(t) \geq 0$ be a bounded cut off function
satisfying $\chi(t)=1$ on $\partial \Omega^M(t)$ and $\chi(t)=0$ on
$\Omega^M(t)\setminus B^M(t)$.
Without loss of generality,
we may assume that
$\|\nabla \chi(t)\|_{L^\infty({\RR^d})} \le C/t$.
This implies
\begin{equation*}
\begin{split}
\|\chi(t) v(t)\|_{L^2(B^M(t))}&+t\|\nabla(\chi(t)
v(t))\|_{L^2(B^M(t))} \\
&\leq C(\|v(t)\|_{L^2(B^M(t))}+t\|\nabla v(t)\|_{L^2(B^M(t))})\\
&\leq Ce^{-\max(1,t)cM/t }\|\eta\|_{L^2(D)}.
\end{split}
\end{equation*}
Here we use the decay estimate~\eqref{ineq:kato_esti} for last inequality above.
Now, setting $e(t):=\chi(t) v(t)+\zeta(t)$, we find that
$\zeta(t)\in H^1_0(\Omega^M(t))$
satisfies
$$
(\zeta(t),\phi)_{{\Omega^M(t)}}+t^2d_{\Omega^M(t)}(\zeta(t),\phi)=-(\chi(t) v(t),\phi)_{{\Omega^M(t)}}-t^2d_{\Omega^M(t)}(\chi(t) v(t),\phi)
$$
for all $\phi\in H^1_0(\Omega^M(t))$. Taking $\phi=\zeta(t)$, we deduce that
$$
\begin{aligned} \|\zeta(t)\|^2_{L^2(\Omega^M(t))}+t^2\|\nabla\zeta(t)\|^2_{L^2(\Omega^M(t))}&\leq\|\chi(t) v(t)\|^2_{L^2(B^M(t))}+t^2\|\nabla(\chi(t) v(t))\|^2_{L^2(B^M(t))}\\
&
\leq Ce^{-2\max(1,t)cM/t }\|\eta\|^2_{L^2(D)}.\end{aligned}$$
Thus, combining the estimates for $\zeta(t)$ and $\chi(t)v(t)$ completes the proof.
\end{proof}
Lemma~\ref{l:deday} above is instrumental to derive exponentially decaying consistency error as $M\to \infty$. Indeed,
we have the following theorem.
\begin{theorem}[Truncation error]\label{t:truncate}
Let $c$ be the constant appearing in \eqref{ineq:kato_esti} and assume $M>2(s+1)/c$. Then,
there is a positive constant $C$ not depending
on $M$ nor $k$ satisfying
\begin{equation}
|a^k(\eta,\theta)-a^{k,M}(\eta,\theta)|\le
Ce^{-cM}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)},\qquad \hbox{for all } \eta,\theta\in
L^2(D).\label{truncateb}
\end{equation}
\end{theorem}
\begin{proof}
In this proof $C$ denotes a generic constant only depending on $\Omega$. Let $\eta,\theta$ be in
$L^2(D)$. It suffices to bound
$$
\begin{aligned}
E&:= \left| \frac{c_s k}{2}\sum_{j=-{N^-}}^{N^+} e^{s y_j}(w(t_j)-w^M(t_j),\theta)_D\right|\\
&\le C\left(k\sum_{j=-{N^-}}^{-1} e^{s y_j}|(v(t_j)-v^M(t_j),\theta)_D|+k\sum_{j=0}^{{N^+}} e^{s y_j}|(v(t_j)-v^M(t_j),\theta)_D|\right)\\
&=: E_1+E_2
\end{aligned}
$$
with $v(t)=v(\widetilde\eta,t)$ defined by \eqref{vt} and $v^M(t)=v^M(\widetilde\eta,t)$ defined by \eqref{vtM}.
We estimate $E_1$ and $E_2$ separately, starting with $E_1$.
From the definition $t_j=e^{-y_j/2}$, we deduce that when $j<0$, $t_j>1$ so that \eqref{domainbound2} gives
$$
\begin{aligned}
E_1&\leq Cke^{-cM}\sum_{j=-{N^-}}^{-1}e^{s y_j}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)} \\
&\leq Ce^{-cM}\frac{ke^{-s k}}{1-e^{-s k}}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)} \leq Ce^{-cM}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)} .
\end{aligned}
$$
Similarly, for $j\geq 0$, i.e. $t_j<1$, using \eqref{domainbound2} again, we have
$$
\begin{aligned}
E_2&\leq Ck\sum_{j=0}^{{N^+}}e^{s y_j}e^{-{c M}/t_j}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)} \\
&\leq Ck\sum_{j=0}^{{N^+}}e^{s y_j}e^{-cM(1+y_j/2)}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)}\\
&= Cke^{-cM}\sum_{j=0}^{{N^+}}e^{(s-cM/2) y_j}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)}\\
&\leq Ce^{-cM}\frac{k}{1-\exp(k(s-cM/2))}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)}\\
& \leq \frac{Ce^{-cM}}{cM/2-s}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)}\le C e^{-cM}\|\eta\|_{L^2(D)}\|\theta\|_{L^2(D)},
\end{aligned}
$$
where we have also used the property $cM/2-s>1$ guaranteed by the assumption $M> 2(s+1)/c$.
\end{proof}
\subsection{Uniform Norm Equivalence on Convex Domains}
Since the domains $\Omega^M(t)$ are convex, we know that the norms in $\dH r(\Omega^M(t))$ are equivalent to those in $H^r(\Omega) \cap H^1_0(\Omega^M(t))$ for $r \in [1,2]$, see e.g. \cite{BP13}.
However, as we mentioned in Remark~\ref{r:beta-equiv}, the equivalence
constants depend a-priori on $\Omega^M(t)$ and
therefore on $M$ and $t$. We show in this section that they can be bounded
uniformly independently of both parameters.
To simplify the notation introduced in Section~\ref{s:dotted_space}.
We shall denote $T_{\Omega^M(t)}$ by $T_t$,
$L_{{\Omega^M(t)}}$ by $L_t$ and $\dH s({\Omega^M(t)})$ by $\dH s$.
We recall that $\Omega^M(t)$ is a dilatation of the convex and bounded domain $\Omega$ containing the origin, see \eqref{e:omega_Mt}. We then have the following lemma.
\def{{\mathbb V}_h^M(t)}{{{\mathbb V}_h^M(t)}}
\begin{lemma}[Ellipitic Regularity on Convex Domains] \label{h2omt} Let $f\in L^2({\Omega^M(t)})$. Then $\theta:=T_t f$ is in
$H^2(\Omega^M(t))\cap H^1_0(\Omega^M(t))$
and satisfies
\begin{equation}
\|\theta \|_{H^2({\Omega^M(t)})}\le C\|f\|_{L^2({\Omega^M(t)})},
\label{uh2}
\end{equation}
where $C$ is a constant independent of $t$ and $M$.
\end{lemma}
\begin{proof} It is well known that the convexity of $\Omega$ and hence
that of ${\Omega^M(t)}$ implies that the unique solution $\theta$ of
\eqref{varscale} with $\omega$ replaced by ${\Omega^M(t)}$
is in $H^2(\Omega^M(t))\cap
H^1_0(\Omega^M(t))$.
Therefore, the crucial point is to show that the constant in \eqref{uh2} does not depend on $M$ or $t$.
To see this, the $H^2$ elliptic regularity on convex domains implies that for $\hat \theta \in H^1_0(\Omega)$ with $\Delta \hat \theta \in L^2(\Omega)$ then $\hat \theta \in H^2(\Omega)$ and there is a constant $C$ only depending on $\Omega$ such that
\begin{equation} |\hat \theta|_{H^2(\Omega)} \le C \|\Delta \hat \theta\|_{L^2(\Omega)}.
\label{h2sreg}
\end{equation}
Here $|\cdot|_{H^2(\Omega)}$ denotes the $H^2(\Omega)$ seminorm.
Let $\gamma$ be such that ${\Omega^M(t)}=\{\gamma x,\
x\in \Omega\} $ (see \eqref{e:omega_Mt}) and $\hat \theta(\hat x) = \theta(\gamma \hat x)$ for $\hat x\in
\Omega$.
Once scaled back to $\Omega^M(t)$, estimate \eqref{h2sreg} gives
\begin{equation}
| \theta |_{H^2({\Omega^M(t)})} \le C\|\Delta \theta \|_{L^2({\Omega^M(t)})} =
C\|f-\theta\|_{L^2({\Omega^M(t)})}.\label{h2sreg1}
\end{equation}
Now \eqref{varscale} immediately implies that
$\|\theta \|_{H^1({\Omega^M(t)}))}\le \|f\|_{L^2({\Omega^M(t)})}$ and \eqref{uh2} follows by
the triangle inequality and obvious manipulations.
\end{proof}
\begin{remark}[Intermediate Spaces]\label{r:h2omt} Lemma~\ref{h2omt} implies that
$D(L_t)=\dH 2=H^2({\Omega^M(t)})\cap H^1_0({\Omega^M(t)})$ with norm equivalence constants independent of $M$ and $t$. As
$D(L_t^{1/2}) = \dH 1 =H^1_0({\Omega^M(t)})$, for $s \in [1,2]$
$$\dH s = (H^1_0({\Omega^M(t)}),H^2({\Omega^M(t)})\cap H^1_0({\Omega^M(t)}))_{s-1,2} {=H^s({\Omega^M(t)})\cap H^1_0({\Omega^M(t)})}$$
with with norm equivalence constants independent of $M$ and $t$.
\end{remark}
\begin{lemma}[Norm Equivalence] \label{hdbound} For $\beta\in [1,3/2)$, let $\theta$ be in $\dH \beta$
and $\widetilde \theta$ denote its extension by zero outside of ${\Omega^M(t)}$. Then
$\widetilde \theta$ is in $H^\beta({\RR^d})$ and
$$\|\theta \|_{\dH \beta } \le C \|\widetilde
\theta \|_{H^\beta({\RR^d})}$$
with $C$ not depending on $t$ or $M$.
\end{lemma}
\begin{proof}
Given $\theta \in H^1({\Omega^M(t)})$, we denote $R\theta $ to be the
elliptic projection of $\theta$ into $H^1_0({\Omega^M(t)})$, i.e., $R\theta \in
H^1_0({\Omega^M(t)})$ is the solution of
$$\begin{aligned} (R\theta,\phi)_{{\Omega^M(t)}} &+d_{\Omega^M(t)}(R\theta,\phi)\\
&=(\theta,\phi)_{{\Omega^M(t)}}
+d_{\Omega^M(t)}(\theta,\phi)
,\qquad \hbox{for all } \phi \in
H^1_0({\Omega^M(t)}).\end{aligned}
$$
It immediately follows that
$$\|R\theta\|_{\dH1} = \|R\theta \|_{H^1({\Omega^M(t)})} \le \|\theta\|_{H^1({\Omega^M(t)})}.
$$
Also, if $\theta \in H^2({\Omega^M(t)})$, Lemma~\ref{h2omt} (see also Remark~\ref{r:h2omt}) implies
$$\|R\theta\|_{\dH2} \le C\|R\theta\|_{H^2({\Omega^M(t)})} \le C \|\theta\|_{H^2({\Omega^M(t)})}$$
with $C$ not depending on $t$ or $M$. Hence, it follows by interpolation that
\begin{equation}
\|R \theta \|_{\dH \beta} \le C_\beta \|\theta\|_{(H^1({\Omega^M(t)}),H^2({\Omega^M(t)}))_{\beta-1,2}}.
\label{feq}
\end{equation}
Now when $\theta \in \dH \beta \subset H^1(\Omega^M(t))$, $R\theta = \theta$ so that in view of \eqref{feq}, it remains to show that
$$
\|\theta\|_{(H^1({\Omega^M(t)}),H^2({\Omega^M(t)}))_{\beta-1,2}} \leq C \| \tilde \theta \|_{H^\beta({\RR^d})},
$$
for a constant $C$ independent of $M$ and $t$.
To see this, note that $\widetilde \theta$ is in
$H^1({\RR^d})$ and the extension of $\nabla \theta$ by zero is in $H^{\beta-1}({\RR^d})$ for $\beta < 3/2$.
We refer to Theorem 1.4.4.4 of \cite{grisvard} for a proof when $d=1$ and the techniques used in Lemma 4.33
of \cite{DD12} for the extension to the higher dimensional spaces.
This implies that $\widetilde \theta$ belongs to $H^\beta({\RR^d})$.
Moreover, the restriction operator is
simultaneously bounded from $H^j({\RR^d})$ to $H^j({\Omega^M(t)})$ for $j=1,2$.
Hence, by interpolation again, we have that
$$\|\theta \|_{(H^1({\Omega^M(t)}),H^2({\Omega^M(t)}))_{\beta-1,2}} \le \|\widetilde \theta\|_{H^\beta({\RR^d})}.$$
This completes the proof of the lemma.
\end{proof}
\section{Finite Element Approximation}
\label{s:FEM}
In this section, we turn our attention to the finite element approximation of each subproblems \eqref{vtM} in $a^{k,M}(\cdot,\cdot)$.
Throughout this section, we omit when no confusion is possible the subscript $j$ in $t_j$, i.e. we consider a generic $t$ keeping in mind that the subsequent statements only hold for $t=t_j$ with $j=-N^-,...,N^+$.
We also make the additional unrestrictive assumption that $\Omega$ used to define $\Omega^M(t)$ (see \eqref{e:omega_Mt}) is polygonal.
In turn, so are all the dilated domains $\Omega^M(t)$.
\subsection{Finite Element Approximation of $a^{k,M}(\cdot,\cdot)$}
For any polygonal domain $\omega$, let $\{\mathcal{T}_h(\omega) \}_{h>0}$ be a sequence of conforming subdivisions
of $\omega$ made of simplices of maximal size diameter $h\le1$. We
use the notation $\mathcal T^M_h(t):=\mathcal{T}_h(\Omega^M(t))$
for $t=t_j$, $j=-N^-,...,N^+$, given by \eqref{e:tj}. We assume that the subdivisions
on $D$
are shape-regular and quasi-uniform. This
means that there exist universal constants $\sigma,\rho >0$ such that
\begin{equation}\label{a:shape-regular}
\sup_{h>0}\max_{T\in\mathcal T_h(D)}\left(\frac {\text{diam}(T)}{r(T)}\right)
\leq \sigma,
\end{equation}
\begin{equation}\label{a:quasi-uniform}
\sup_{h>0} \left(\frac{\max_{T \in \mathcal T_h(D)} \text{diam}(T)} {\min_{T \in \mathcal T_h(D)} \text{diam}(T)}\right) \leq \rho,
\end{equation}
where $\text{diam}(T)$ stands for the diameter of $T$ and $r(T)$ for the
radius of the largest ball contained in $T$. We also assume that these
conditions hold as well for $\mathcal T^M_h(t_j)$ with constants
$\sigma,\rho$ not depending on $j$.
We finally require that all the subdivisions match on $D$, i.e.
\begin{equation}
\mathcal
T_h(D)\subset \mathcal T^M_h(t_j)
\label{a:mesh}
\end{equation}
for each $j$.
We discuss in Section~\ref{s:numerical} how to generate subdivisions meeting these requirements.
Define $\mathbb{V}_h(\omega )\subset H^1_0(\omega)$ to be the space of
continuous piecewise linear finite element functions associated with
$\mathcal{T}_h(\omega)$ with $\omega =D$ or $\Omega^M(t)$.
Also, we use the short notation $\mathbb V_h^M(t):=\mathbb V_h(\Omega^M(t))$.
We are now in position to define the fully discrete/implementable problem.
For $\eta_h$ and $\theta_h$ in $\mathbb V_h {(D)}$, the finite element approximation of $a^{k,M}(\cdot,\cdot)$ given by \eqref{isM} is
\begin{equation}
a_h^{k,M}(\eta_h,\theta_h):=\frac{c_s k}{2}\sum_{j=-{N^-}}^{N^+} e^{s y_j}(w_h^M(\widetilde\eta_h,t_j),\theta_h)_{ {D}}
\label{isMh}
\end{equation}
with
\begin{equation}\label{e:wMh}
w^M_h(\widetilde \eta_h,t):=\widetilde \eta_h|_{\Omega^M(t)} +v^M_h(t)
\end{equation}
and where $v^M_h(t) \in \mathbb V_h^M(t)$ solves
\begin{equation}\label{e:vhM}
(v^M_h(t),\phi_h)_{\Omega^M(t)}+t^2d_{\Omega^M(t)}(v^M_h(t),\phi_h)=-(\widetilde \eta_h,\phi_h)_{\Omega^M(t)},\quad\forall \phi_h\in \mathbb V^M_h(t).
\end{equation}
\begin{remark}[Assumption \eqref{a:mesh}]
Two critical properties follow from \eqref{a:mesh}.
On the one hand, our analysis below relies on the fact that the extension by zero $\widetilde
v_h$ of $v_h \in \mathbb V_h {(D)}$ belongs to all $\mathbb V_h^M(t)$.
This property greatly simplifies the computation of
$(w_h^M(\widetilde\eta_h,t_j),\theta_h)_{ {D}}$ in \eqref{isMh}.
\end{remark}
The finite element approximation
of the problem \eqref{mainP} is to find $u_h\in \mathbb{V}_h {(D)}$ so that
\begin{equation}
a^{k,M}_h(u_h,\theta_h)=(f,\theta_h)_{ {D}}\qquad\hbox{for all }
\theta_h\in\mathbb{V}_h {(D)} .
\label{finaldiscrete}
\end{equation}
Analogous
to Lemma~\ref{l:K}, we have the following representation using K-functional. The proof of the lemma is similar to that of
Lemma~\ref{l:K} and is omitted.
\begin{lemma}[K-functional formulation on the discrete space]\label{l:K_discrete}
For $\eta_h \in \mathbb V_h {(D)}$, there holds
$$
(w_h^M(\widetilde \eta_h,t), \eta_h)_{D} = (w_h^M(\widetilde \eta_h,t), \widetilde \eta_h)_{\Omega^M(t)} = \ K_h(\widetilde \eta_h,t),
$$
where
$$
K_h(\widetilde \eta_h,t):= \min_{\varphi_h \in \mathbb V_h^M(t)} \left( \| \widetilde \eta_h - \varphi_h \|_{L^2(\Omega^M(t))}^2 + t^2 d_{{\Omega^M(t)}}(\varphi_h,\varphi_h) \right).
$$
\end{lemma}
We emphasize that for $v_h \in \mathbb V_h^M(t)$, its extension by zero $\tilde \eta_h$ belongs to $H^1({\RR^d})$ and therefore
\begin{equation}\label{e:monotonK}
K_h(\tilde v_h,t) \geq K(\tilde v_h,t).
\end{equation}
This property is critical in the proof of next theorem, which ensures the $\mathbb V_h(D)$-ellipticity of the discrete bilinear for $a_h^{k,M}$.
Before describing this next result, we recall that according to \eqref{e:quad_cons}
$$
|a(\eta_h,\theta_h)-a^k(\eta_h,\theta_h)|\le\gamma(k) \|\eta_h\|_{\widetilde H^\delta(D)}\|\theta_h\|_{{\tH^s}(D)}
$$
with
$\delta$ between $s$ and $\min(2-s,3/2)$
(since $\mathbb V_h(D)\subset \widetilde H^{3/2-\epsilon}(D)$ for any $\epsilon>0$) and $\gamma(k) \sim Ce^{-2\pi\mathsf{d}/k}$.
Also, we note that from the quasi-uniform \eqref{a:shape-regular} and shape-regular \eqref{a:quasi-uniform} assumptions, there exists a constant $c_I$ only depending on $\sigma$ and $\rho$ such that for $r^- \leq r^+ < 3/2$, there holds
\begin{equation}\label{e:inverse}
\| v_h \|_{\widetilde H^{r^+}(D)} \leq c_I h^{r^{-}-r^+}\| v_h \|_{\widetilde H^{r^-}(D)}, \qquad \forall v_h \in \mathbb V_h^M(t).
\end{equation}
\begin{theorem}[$\mathbb V_h {(D)}$-ellipticity]\label{l:ellipticity}
Let $\delta$ in Theorem~\ref{t:quad_error} between $s$ and
$\min(2-s,3/2)$, $k$ be the quadrature spacing and
$c_I$ be the inverse constant in \eqref{e:inverse}. We assume that the quadrature parameters $N^-$ and $N^+$ are chosen according to \eqref{e:choiceN}. Let $\gamma(k)$ be given by \eqref{e:gammak} and assume that $k$ is chosen sufficiently small so that
$$
c_I \gamma(k) h^{s-\delta} < 1.
$$
Then, there is a constant $c$ independent of $h,k$ and $M$ such that
$$
a_h^{k,M}(\eta_h,\eta_h) \geq c\| \eta_h \|_{{\tH^s}(D)}^2, \qquad \hbox{for all } \eta_h \in \mathbb V_h(D).
$$
\end{theorem}
\begin{proof}
Let $\eta_h \in \mathbb V_h^M(t)$ so that $\widetilde \eta_h\in H^1({\RR^d})$.
We use the equivalence relations provided by Lemmas~\ref{l:K} and~\ref{l:K_discrete} together with the monotonicity property \eqref{e:monotonK} to write
$$
a_h^{k,M}(\eta_h,\eta_h) = \frac{c_s k}{2}\sum_{j=-{N^-}}^{N^+} e^{s y_j}K_h(\widetilde \eta_h,t_j) \geq
\frac{c_s k}{2}\sum_{j=-{N^-}}^{N^+} e^{s y_j}K(\widetilde \eta_h,t_j) = a^k(\eta_h,\eta_h).
$$
The quadrature consistency bound \eqref{e:quad_cons} supplemented by an inverse inequality~\eqref{e:inverse} yields
$$
a_h^{k,M}(\eta_h,\eta_h) \geq a(\eta_h,\eta_h)-\gamma(k) \|\eta_h\|_{\widetilde H^\delta(D)}\|\eta_h\|_{{\tH^s}(D)}
\geq a(\eta_h,\eta_h)- c_I \gamma {(k)} h^{s-\delta}\|\eta_h\|_{{\tH^s}(D)}^2.
$$
The desired result follows from assumption $c_I \gamma(k)
h^{s-\delta}<1
$ and the coercivity of $a(\cdot,\cdot)$, see \eqref{e:coercivity_a}.
\end{proof}
\subsection{Approximations on $\Omega^M(t)$}
The fully discrete scheme \eqref{finaldiscrete} requires approximations by the finite element methods on domains $\Omega^M(t)$.
Standard finite element argumentations would lead to estimates with constants depending on $\Omega^M(t)$ and therefore $M$ and $t$.
In this section, we exhibit results where this is not the case due to the particular definition \eqref{e:omega_Mt} of $\Omega^M(t)$.
We can use interpolation to develop approximation results for functions
in the intermediate spaces with constants independent of $M$ and $t$. The
Scott-Zhang interpolation construction {\cite{SZ90}} gives rise to an approximation
operator $\pi^{sz}_h:H^1_0({\Omega^M(t)})\rightarrow \mathbb V_h^M(t)$ satisfying
$$ \|\eta- \pi^{sz}_h \eta \|_{H^1({\Omega^M(t)})} \le C \|
\eta\|_{H^1({\Omega^M(t)})},
$$
for all $\eta\in H^1_0({\Omega^M(t)})=\dH 1$
and
$$\| \eta-\pi^{sz}_h \eta \|_{H^1({\Omega^M(t)})} \le C h \|
\eta\|_{H^2({\Omega^M(t)})},$$
for all $\eta\in H^2({\Omega^M(t)})\cap H^1_0({\Omega^M(t)})=\dH 2$.
The Scott-Zhang argument is local so the constants appearing above depend
on the shape regularity of the triangulations but not on $t$ or $M$.
Interpolating the above inequalities shows that for all $r\in [0,1]$
\begin{equation}
\inf_{\chi\in {{\mathbb V}_h^M(t)}} \|\eta-\chi\|_{H^1({\Omega^M(t)})} \le C h^r \|\eta\|_{\dH
{1+r}}, \qquad \hbox{for all } \eta\in \dH {1+r}
\label{app}
\end{equation}
with $C$ not depending on $t$ or $M$.
Let $T_{t,h}$ denote the finite element approximation to $T_t$ given by \eqref{e:Tomega}, i.e.,
for $F\in \dH {-1}$, $T_{t,h}F:=w_h$ with $w_h\in \mathbb V_h^M(t)$ being the unique
solution of
$$
(w_h,\phi)_{\Omega^M(t)} +d_{\Omega^M(t)} (w_h,\phi) = \langle F,\phi\rangle, \qquad \hbox{for all } \phi\in \mathbb V_h^M(t).
$$
The approximation result \eqref{app} and standard finite element analysis
techniques implies that for any $r\in [0,1]$,
\begin{equation}
\|T_tF - T_{t,h}F\|_{L^2({\Omega^M(t)})} \le C h^{1+r} \|T_tF\|_{\dH {1+r}}\le C h^{1+r} \|F\|_{\dH {-1+r}},
\label{Tdifapp}
\end{equation}
where the last inequality follows from interpolation since $\|T_t F\|_{H^1({\Omega^M(t)})}\le \|F\|_{H^{-1}({\Omega^M(t)})}$ and \eqref{uh2} hold.
For $f\in L^2(\Omega^M(t))$, we define the operator
\begin{equation}\label{e:Sop}
S_tf:=\eta \in
H^1_0({\Omega^M(t)})
\end{equation}
satisfying,
$$d_{\Omega^M(t)}(\eta,\phi)= (f,\phi)_{\Omega^M(t)},\qquad \hbox{for all } \phi\in H^1_0({\Omega^M(t)})$$
and let $S_{t,h} f\in \mathbb V_h^M(t)$ denote its finite element
approximation; compare with $T_t$ and $T_{h,t}$. Although the Poincar\'e constant depends on the diameter of ${\Omega^M(t)}$, we still have the following lemma.
\begin{lemma} \label{tzer} There is a constant $C$ independent of $h$, $t$, or $M$
satisfying
$$\|S_t f-S_{t,h} f\|_{L^2({\Omega^M(t)})} \le Ch^2 \|f\|_{L^2({\Omega^M(t)})}.$$
\end{lemma}
\begin{proof}
For $f\in L^2({\Omega^M(t)})$, set $e_h:=(S_t -S_{t,h}) f$.
The elliptic regularity estimate \eqref{h2sreg1} on convex domain and Cea's Lemma imply
\begin{equation*}
\begin{split}
|e_h |_{H^1({\Omega^M(t)})} &= \inf_{\chi_h\in {{\mathbb V}_h^M(t)}}
|S_t f -\chi_h |_{H^1({\Omega^M(t)})}
\le Ch |S _t f|_{H^2({\Omega^M(t)})} \\
&\le Ch \|\Delta S _t f\|_{L^2({\Omega^M(t)})} = Ch \| f\|_{L^2({\Omega^M(t)})},
\end{split}
\end{equation*}
where $C$ is a constant independent of $h$, $t$ and $M$.
Galerkin orthogonality and the above estimate give
$$\begin{aligned} \|e_h\|_{L^2({\Omega^M(t)})} ^2&=d_{\Omega^M(t)}(e_h,S_t e_h)=
d_{\Omega^M(t)}(e_h,(S_t-S_{t,h})e_h )\\
&\le |e_h|_{H^1({\Omega^M(t)})} |(S_t-S_{t,h})e_h|_{H^1({\Omega^M(t)})} \\& \le Ch
|e_h|_{H^1({\Omega^M(t)})} \|e_h\|_{L^2({\Omega^M(t)})} .
\end{aligned}
$$
Combining the above two inequalities and obvious manipulations completes
the proof of the lemma.
\end{proof}
We shall also need norm equivalency on discrete scales. Let
$({{\mathbb V}_h^M(t)},\|\cdot\|_{L^2({\Omega^M(t)})})$ and $({{\mathbb V}_h^M(t)},\|\cdot\|_{H^1({\Omega^M(t)})})$ denote
${{\mathbb V}_h^M(t)}$ normed with the norms in $L^2({\Omega^M(t)})$ and $H^1({\Omega^M(t)})$, respectively.
We define $\|\cdot\|_{\dH r_h(\Omega^M(t))}$, or simply $\|\cdot\|_{\dH r_h}$, to be the norm
in the interpolation space
$$
\big({{\mathbb V}_h^M(t)},\|\cdot\|_{L^2({\Omega^M(t)})}),({{\mathbb V}_h^M(t)},\|\cdot\|_{H^1({\Omega^M(t)})})\big)_{r,2}.
$$
For $r\in[0,1]$, as the natural injection is a bounded map (with bound 1) from
${{\mathbb V}_h^M(t)}$ into $L^2({\Omega^M(t)})$ and $H^1_0({\Omega^M(t)})$, respectively,
$\|v_h\|_{\dH r} \le \|v_h\|_{\dH r_h}$, for all $v_h\in {{\mathbb V}_h^M(t)}$.
For the other direction, one needs a projector into ${{\mathbb V}_h^M(t)}$ which is simultaneously
bounded on $L^2({\Omega^M(t)})$ and $H^1_0({\Omega^M(t)})$. In the case of a globally quasi uniform
mesh, it was shown by Bramble and Xu \cite{BX91} that the $L^2({\Omega^M(t)})$ projector
$\pi_h$ satisfies this property. Their argument is local, utilizing the
inverse inequality \eqref{e:inverse} and hence leads to constants depending on those
appearing in \eqref{a:shape-regular} and \eqref{a:quasi-uniform}
but not $t$, $h$, or $M$. Interpolating
these results gives, for
$r\in [0,1]$,
\begin{equation}
c\|v_h\|_{\dH r_h} \le \|v_h \|_{\dH r} \le \|v_h\|_{\dH r_h} , \qquad \hbox{for all }
v_h\in {{\mathbb V}_h^M(t)},
\label{dsplus}\end{equation}
where $c$ is a constant independent of $h$, $M$ and $t$.
The spaces for negative $r$ are defined by duality and the stability of the $L^2(\Omega^M(t))$-projection $\pi_h$ yields
\begin{equation}
c\|v_h\|_{\dH{-r}} \le \|v_h\|_{\dH {-r}_h} \le
\|v_h\|_{\dH {-r}}.\qquad \hbox{for all } v_h\in {{\mathbb V}_h^M(t)}.
\label{dsneg}
\end{equation}
We finally note that a discrete version of Lemma~\ref{Tshift} holds.
Its proof is essentially the same and is omitted for brevity.
\begin{lemma} \label{Tshifth} Let $a$ be in $[0,2]$ and $b$ be in $[0,1]$
with $a+b\le 2$.
Then for
any $\mu\in (0,\infty)$,
$$\| (\mu I + T_{t,h})^{-1} \eta_h\|_{\dH{-b}_h } \le \mu^{(a+b)/2-1}
\|\eta_h\|_{\dH{a}_h},\qquad \hbox{for all } \eta_h\in {{\mathbb V}_h^M(t)}.$$
\end{lemma}
\subsection{Consistency}
The next step is to estimate the consistency error between
$a^{k,M}(\cdot,\cdot)$ and $a_{h}^{k,M}(\cdot,\cdot)$ on $\mathbb V_h(D)$.
Its decay depends on a parameter $\beta \in (s,3/2)$, which will be related later to the regularity of the solution $u$ to \eqref{mainP}.
\begin{theorem}[Finite Element Consistency]\label{t:discrete_consistency}
Let $\beta\in(s,3/2)$.
We assume that the quadrature parameters $N^-$ and $N^+$ are chosen according to \eqref{e:choiceN}. {There} exists a constant
$C$ independent of $h$, $k$ and $M$ satisfying
\begin{equation}\label{h-estimate}\begin{aligned}
|a^{k,M}(\eta_h,\theta_h)&-a^{k,M}_h(\eta_h,\theta_h)| \\
&\le
C(1+\ln(h^{-1}))h^{\beta-s}\|\eta_h\|_{\widetilde
H^\beta(D)}\|\theta_h\|_{\widetilde H^{s}(D)}
\end{aligned}
\end{equation}
for all $\eta_h, \theta_h\in \mathbb{V}_h {(D)}$.
\end{theorem}
\begin{proof}
In this proof, $C$ denotes a generic constant independent of $h$, $M$,
$k$ and $t$.
Fix $\eta_h \in \mathbb V_h(D)$ and denote by $\tilde \eta_h$ its extension by zero outside $D$.
We first observe that for $\theta_h\in \mathbb{V}_h(D)$ and $\tilde \theta_h$ its extension by zero outside $D$, we have
\begin{equation}\begin{aligned}
(w^M(\widetilde \eta_h,t_j),\theta_h)_D&= ( \pi_h
w^M(\widetilde \eta_h,t_j),\widetilde \theta_h)_{\Omega^M(t)},
\end{aligned}
\end{equation}
where $\pi_h$ denotes the $L^2$ projection onto $\mathbb V_h(\Omega^M(t))$.
Using the above identity and recalling that $t_j=e^{-y_j/2}$, we obtain
\begin{equation*}
\begin{split}
a^{k,M}(\eta_h,\theta_h)-a^{k,M}_h(\eta_h,\theta_h)
= \frac{c_s }{2} \underbrace{k\sum_{t_j\le \frac 12} e^{sy_j}
( \pi_hw^M(\widetilde \eta_h,t_j)-w^M_h(\widetilde
\eta_h,t_j),\widetilde \theta_h)_{\Omega^M(t)}}_{=:E_1
} \\ +
\frac{c_s }{2}\underbrace{k\sum_{t_j> \frac 12} e^{sy_j} (
\pi_hw^M(\widetilde \eta_h,t_j)-w^M_h( \widetilde \eta_h,t_j),\widetilde \theta_h)_{\Omega^M(t)}}_{=: E_2}.
\end{split}
\end{equation*}
We bound the two terms separately and start with the latter.
$\boxed{1}$ In view of the definitions \eqref{wMt} of $w^M(t)$ and \eqref{e:wMh} of $w^M_h(t)$, we have
\begin{equation} \label{e:usev_v_h}
\pi_h w^M(\widetilde \eta_h,t)-w^M_h(\widetilde \eta_h,t)
= \pi_h v^M(\widetilde \eta_h,t)-v^M_h(\widetilde \eta_h,t).
\end{equation}
We recall that $T_t = T_{\Omega^M(t)}$ and $S_t$ are defined by \eqref{e:Tomega} and \eqref{e:Sop} respectively.
Using these operators and the relations satisfied by $v^M(t)$ and $v_h^M(t)$ (see \eqref{vtM} and \eqref{e:vhM}), we arrive at
\begin{equation}
\begin{aligned}
\pi_h w^M(\widetilde \eta_h,t)&-w^M_h(\widetilde \eta_h,t)
= [S_{t,h} (S_{t,h}+t^2I)^{-1}- \pi_hS_t(S_t+t^2I)^{-1}]\widetilde \eta_h\\
& = t^{2}(S_{h,t} +t^{-2}I)^{-1} \pi_h(S_{t,h}-S_t)(S_t+t^{2}I)^{-1} \widetilde \eta_h.
\end{aligned}
\label{vdif}
\end{equation}
Thus,
\begin{align*}
\| \pi_h w^M(\widetilde \eta_h,t)&-w^M_h(\widetilde \eta_h,t)\|_{L^2(\omt)}\\
&\leq t^{2}\|(S_{t,h}+t^{2}I)^{-1}
\pi_h(S_{t,h}-S_t)(S_t+t^{2}I)^{-1}\|\ \| \widetilde \eta_h \|_{L^2(\omt)}\\
&\le t^{2}\|(S_{t,h}+t^{2}I)^{-1} \pi_h\|\
\|S_{t,h}-S_t\|\
\|(S_t+t^{2}I)^{-1}\|\ \| \widetilde \eta_h \|_{L^2(\omt)}.
\end{align*}
Here we have used $\|\cdot\|$ to denote the operator norm of operators from $L^2({\Omega^M(t)})$ to $L^2({\Omega^M(t)})$.
Combining
$$
\| (S_{t,h}+t^2I)^{-1}\pi_h\|\le t^{-2},\quad
\| (S_t+t^{2}I)^{-1}\|
\leq t^{-2}
$$
and Lemma~\ref{tzer} gives
$$
\|\pi_h w^M(\widetilde \eta_h,t)-w^M_h(\widetilde \eta_h,t)\| \leq C t^{-2} h^2 \|\widetilde \eta_h\|_{L^2({\Omega^M(t)})}.
$$
Whence,
$$
\begin{aligned}
|E_2| &\le Ch^2k\sum_{t_j> \frac 12} e^{(s+1)jk} \| {\widetilde\eta_h} \|_{L^2({\Omega^M(t)})} \| {\widetilde\theta_h} \|_{L^2({\Omega^M(t)})}\\
&\leq C h^2 \| \eta_h \|_{L^2(D)} \| \theta_h \|_{L^2(D)}\left(k\sum_{jk<2\ln2} e^{(s+1)jk}\right)\le C h^2 \| \eta_h \|_{L^2(D)} \| \theta_h \|_{L^2(D)}.
\end{aligned}
$$
\def\HHp#1{\dH {#1}}
$\boxed{2}$ We now focus on $E_1$ which requires a finer
analysis using
intermediate spaces. Also, we argue differently for
$\beta\in(1,3/2)$ and for $\beta\in(s,1]$. In either case, we define
$$\epsilon:=\min\{1-s,1/\ln(1/h)\}$$
and note that
\begin{equation}
\epsilon^{-1} \le c(1+\ln(1/h)) \quad\hbox{and}\quad
h^{-\epsilon}\le c
\label{epsilon}
\end{equation}
with $c$ depending on $s$ but not $h$.
When $\beta\in(1,3/2)$, we invoke \eqref{e:usev_v_h} again to deduce
\begin{equation}
|E_1| \leq k\sum_{t_j\le\frac 12} e^{sy_j}\|\pi_h v^M(\widetilde
\eta_h,t_j)-v^M_h( \widetilde \eta_h,t_j) \|_{\dH {-s}_h } \| \widetilde\theta_h \|_{\dH
{s}_h}.
\label{e1est}
\end{equation}
We set $\mu(t) := t^{-2}-1$ and compute
\begin{equation}\label{e:three-parts}
\begin{aligned}
\pi_h v^M(\widetilde \eta_h,t)&-v^M_h(\widetilde \eta_h,t) =
t^{-2} [(I+\mu(t)T_{t,h})^{-1}T_{t,h}
-\pi_h T_t (I+\mu(t)T_t)^{-1}]\widetilde \eta_h\\
& = (t\mu(t))^{-2} (\mu(t)^{-1}I+T_{t,h})^{-1}\pi_h(T_{t,h}-T_t)(\mu(t)^{-1}I+T_t)^{-1}\widetilde \eta_h,
\end{aligned}
\end{equation}
which is now estimated in three parts.
Lemma~\ref{Tshift} guarantees that
$$
\|(\mu(t)^{-1}I+T_t)^{-1}\|
_{\dH \beta \rightarrow \dH {\beta-2}}\le 1,
$$
where we recall that $\dH s$ stands for $\dH s(\Omega^M(t))$.
For the second part, the error estimate \eqref{Tdifapp} with $1+r = \beta$ reads
$$
\| T_{t,h}-T_t \|_{\dH {\beta-2} \rightarrow
L^2(\Omega^M(t))}\leq Ch^{\beta}.
$$
We estimate the last term of the product in the right hand side of \eqref{e:three-parts} by
$$\begin{aligned} \|(\mu(t)^{-1} &+T_{t,h})^{-1}
\pi_h\|_{L^2(\Omega^M(t))\rightarrow \dH {-s}_h }\\
&\le C \|(\mu(t)^{-1} +T_{t,h})^{-1}\|_{ \dH {s+\epsilon}_h\rightarrow \dH {-s}_h} \|
\pi_h\|_{L^2(\Omega^M(t))\rightarrow \dH {s+\epsilon}_h}.
\end{aligned}$$
Thus, Lemma~\ref{Tshifth}, the inverse estimate and \eqref{epsilon} yield
$$
\|(\mu(t)^{-1} +T_{t,h})^{-1} \pi_h \|_{L^2(\Omega^M(t))\rightarrow \dH
{-s}_h }\le C h^{-s-\epsilon} t^{(2s+\epsilon-2)}\le C h^{-s}
t^{(2s+\epsilon-2)}.
$$
Note that for $t\in (0,1/2]$, $0<t^2\leq \mu(t)^{-1} \leq \frac 4 3 t^2 \leq \frac 1 3$ so
that
$$ (t\mu(t))^{-2}\le \frac {16t^2} 9.
$$
Combining the above estimates with
\eqref{e:three-parts} gives
\begin{equation}
\| \pi_h v^M(\widetilde \eta_h,t)-v^M_h(\widetilde \eta_h,t)\|_{\dH{-s}_h} \leq
Ct^{2s+\epsilon}h^{\beta-s} \|\widetilde{\eta}_h\|_{\dH\beta},
\label{vmest}
\end{equation}
Since $t_j=e^{-y_j/2}$,
$$e^{sy_j} t_j^{2s+\epsilon}= e^{-\epsilon y_j/2} .$$
Estimates \eqref{e1est}, \eqref{vmest} and \eqref{dsplus} then yield
\begin{equation}\label{e1-sum-bound}\begin{aligned}
|E_1| &\le Ch^{\beta-s}k\sum_{ky_j\ge 2\ln2} e^{-\epsilon y_j/2}
\|\widetilde \eta_h\|_{\dH
\beta}\|\widetilde\theta_h\|_{\dH{s}}\\
&\le Ch^{\beta-s}\epsilon^{-1}
\|\widetilde \eta_h\|_{\dH
\beta}\|\widetilde\theta_h\|_{\dH{s}}.\end{aligned}
\end{equation}
\boxed{3}
We bound the norms on $\Omega^M(t)$ by norms on $D$ using \eqref{extenhs} with $r=s$ and Lemma~\ref{hdbound} to arrive at
$$
\begin{aligned}
|E_1| &\le Ch^{\beta-s}\epsilon^{-1} \|\widetilde \eta_h\|_{H^\beta({\RR^d})}\|\theta_h\|_{\dH {s}(D)}.
\end{aligned}
$$
Applying the norm equivalence \eqref{e:dot-tilde} gives
\begin{equation}\label{e:e1:final}
|E_1| \leq C h^{\beta-s} \epsilon^{-1} \| \eta_h\|_{\widetilde H^\beta(D)}\|\theta_h\|_{\widetilde H^{s}(D)}.
\end{equation}
\boxed{4} When $\beta\in(s,1]$, we bound \eqref{e:three-parts} using different norms. In fact, we have
$$
\|(\mu(t)^{-1}I+T_t)^{-1}\|
_{\dH \beta \rightarrow \dH {-1}}\le t^{\beta-1},
\quad
\| T_{t,h}-T_t \|_{\dH {-1} \rightarrow
L^2(\Omega^M(t))}\leq Ch,
$$
and by Lemma~\ref{Tshifth},
$$\begin{aligned}
\|(\mu(t)^{-1} +T_{t,h})^{-1}& \pi_h \|_{L^2(\Omega^M(t))\rightarrow
\dH {-s}_h }\\
&\le \|(\mu(t)^{-1} +T_{t,h})^{-1} \pi_h
\|_{\dH {1-\beta+s+\epsilon} \rightarrow
\dH {-s}_h }
\|\pi_h\|_{L^2({\Omega^M(t)})\rightarrow \dH {1-\beta+s+\epsilon}}\\
&\le
C h^{-1+\beta-s} t^{(2s+\epsilon-\beta-1)}.
\end{aligned}
$$
These estimates lead \eqref{vmest} and hence \eqref{e1-sum-bound}
as well when $\beta \in (s,1]$.
The remainder of the proof is the same as in the case $\beta \in (1,3/2)$ except that the norm equivalence
\eqref{e:dot-tilde} is invoked in place of Lemma~\ref{hdbound}.
\boxed{5} The proof of the theorem is complete upon combining the estimates for $E_1$ and $E_2$.
\end{proof}
\subsection{Error Estimates}
Now that the consistency error between $a(\cdot,\cdot)$ and $a_h^{k,M}(\cdot,\cdot)$ is obtained, we can apply Strang's lemma to deduce the convergence of the approximation $u_h$ towards $u$ in the energy norm.
To achieve this, we need a result regarding
the stability and approximability of the Scott-Zhang interpolant
$\pi^{sz}_h$ \cite{SZ90} in the fractional spaces $\widetilde
H^\beta(D)$.
This is the subject of the next lemma. Its proof is somewhat technical and given in Appendix~\ref{a:lagrange}.
\begin{lemma}[Scott-Zhang Interpolant] \label{fe-interpolant} Let $\beta \in (1,3/2)$. Then, there is a constant $C$ independent of $h$ such that
\begin{equation}\label{e:stability_sc}
\|\pi^{sz}_h v\|_{\widetilde H^\beta(D)} \le C \|v\|_{\widetilde H^\beta(D)}
\end{equation}
and for $s\in [0,1]$,
\begin{equation}\label{e:approx_sc}
\|\pi^{sz}_h v -v\|_{\widetilde H^s(D)}\le C h^{\beta-s} \|v\|_{\widetilde H^\beta(D)},
\end{equation}
for all $v \in \widetilde H^\beta(D)$.
\end{lemma}
We note that the above lemma holds for $\beta\in (0,1)$ and $s\in
(0,\beta)$ provided that $\pi^{sz}_h$ is replaced by $\pi_h$, the $L^2$ projection onto $\mathbb V_h(D)$; see e.g.
Lemma 5.1 of \cite{BP15}.
In order to consider both case simultaneously in the following proof, we
set $\Pi_h = \pi_h $
when $\beta\in [0,1]$ and $\Pi_h =\pi^{sz}_h $ when $\beta\in (1,3/2)$.
\begin{theorem}\label{t:total_error}
Assume that the solution $u$ of \eqref{mainWP} belongs to $\widetilde H^\beta(D)$ for $\beta \in (s,3/2)$.
Let $\delta:=\min(2-s,\beta)$ be as in Theorem~\ref{t:quad_error}, $k$ be the quadrature spacing and
$c_I$ be the inverse constant in \eqref{e:inverse}. We assume that the quadrature parameters
$N^-$ and $N^+$ are chosen according to \eqref{e:choiceN}. Let $\gamma(k)$ be given by \eqref{e:gammak}
and assume that $k$ is chosen sufficiently small so that
$$
c_I \gamma(k) h^{s-\delta} < 1.
$$
Moreover,
let $u_h \in \mathbb V_h(D)$ be the solution of \eqref{finaldiscrete}.
Then there is a constant $C$
independent of $h$, $M$ and $k$ satisfying
\begin{equation}\label{energy-bound}
\|u-u_h\|_{\widetilde H^s(D)}\leq C(\gamma(k)+e^{-cM}+(1+\ln{(h^{-1})}) h^{\beta-s})\|u\|_{\widetilde H^\beta(D)}.
\end{equation}
\end{theorem}
\begin{proof}
In our context, the first Strang lemma (see e.g. Theorem 4.1.1 in \cite{ciarlet}) reads
$$
\| u - u_h \|_{{\tH^s}(D)} \le C\inf_{v_h \in \mathbb V_h {(D)}} \left(\| u - v_h \|_{{\tH^s}(D)} + \sup_{w_h \in \mathbb V_h {(D)}} \frac{|(a-a_{h}^{k,M})(v_h,w_h)|}{\| w_h \|_{{\tH^s}(D)}}\right),
$$
where $C$ is a constant independent of $h$, $k$ and $M$.
From the consistency estimates \eqref{e:quad_cons}, \eqref{truncateb} and \eqref{h-estimate}, we deduce that
$$\begin{aligned}
\|u-u_h\|_{\widetilde H^s(D)}&\leq C\| u - \Pi_h u\|_{\widetilde H^s(D)} \\
&\qquad + C(\gamma(k)+e^{-cM}+(1+\ln{(h^{-1})}) h^{\beta-s})\|\Pi_h u\|_{\widetilde H^\beta(D)}
\end{aligned}
$$
The desired estimate follows from the approximability and stability of $\Pi_h$.
\end{proof}
\section{Numerical implementation and results}\label{s:numerical}
In this section, we present detailed numerical implementation to solve the following model problems.
\subsection{Model Problems.}\label{s:analytical}
One of the difficulties in developing numerical approximation to
\eqref{mainWP} is that there are relatively few examples where analytical
solutions are available. One exception is the case when $D$ is the
unit ball in ${\RR^d}$. In that case, the solution to the variational
problem
\begin{equation}
a(u,\phi) = (1,\phi)_D,\qquad \hbox{for all } \phi\in {\tH^s}(D)
\label{var1}
\end{equation}
is radial and given by, (see \cite{DG13})
\begin{equation}
u(x)= \frac {2^{-2s}\Gamma(d/2)} {\Gamma(d/2+s) \Gamma(1+s) }
(1-|x|^2)^s
\label{anals}
\end{equation}
It is also possible to compute the right hand side corresponding to the
solution $u(x)=1-|x|^2$ in the unit ball. The corresponding
right hand side can be derived by first computing the Fourier transform
of $\widetilde u$, i.e.,
$${\mathcal F}(\widetilde u)=2J_2(|\zeta|)/|\zeta|^2,$$
where $J_n$ is the Bessel function of the first kind.
When $0<s<1$, we obtain
\begin{equation}\label{rhs_smooth}
f(x)={\mathcal F}^{-1}(2|\zeta|^{2s-2}J_2(|\zeta|))=\frac{2^{2s}\Gamma(d/2+s)}{\Gamma(d/2)\Gamma(2-s)}
{}_2F_1\left(d/2+s, s-1,d/2,|x|^2\right) ,
\end{equation}
where ${}_2F_1$ is the Gaussian or ordinary hypergeometric function.
\begin{remark} [Smoothness]\label{singularf}
Even though the solution $u(x)=1-|x|^2$ is infinitely differentiable on the
unit ball, the right hand side $f$ has limited smoothness. Note that
$f$ is the restriction of $(-\Delta)^s \widetilde u$ to the unit ball.
Now $\widetilde u\in H^{3/2-\epsilon}({\RR^d})$ for $\epsilon>0$ but
is not in $H^{3/2}({\RR^d})$. This means that $(-\Delta)^s \widetilde u$
is only in $H^{3/2-2s-\epsilon}({\RR^d})$ and hence $f$ is only
in $H^{3/2-2s-\epsilon}(\Omega)$. This is in agreement with the
singular behavior of $ {}_2F_1\left(d/2+s, s-1,d/2,t\right)$ at $t=1$ (see
\cite{NIST}, Section 15.4).
In fact,
$$\begin{aligned}
{}_2F_1\left(d/2+s, s-1,d/2,1\right)&=\frac {\Gamma(d/2+s)
\Gamma(1-2s)}{\Gamma(d/2+1-s) \Gamma(-s)}\quad \hbox{when }0<s<1/2,\\
\lim_{t\rightarrow 1^-} \frac{{}_2F_1\left(d/2+s, s-1,d/2,t\right)}
{-\log(1-t)} &= \frac {\Gamma(d/2)}{\Gamma(-1/2)\Gamma(1/2)} \quad
\hbox{when }s=1/2,\\
\lim_{t\rightarrow 1^-} \frac {{}_2F_1\left(d/2+s, s-1,d/2,t\right)}
{(1-t)^{-2s+1}} &= \frac {\Gamma(d/2)}{\Gamma(-1/2)\Gamma(1/2)} \quad
\hbox{when }1/2<s<1.
\end{aligned}
$$
This implies that for $s\ge 1/2$, the trace on $|x|=1$ of $f(x)$ given by \eqref{rhs_smooth} fails to exist (as for generic functions in
$H^{3/2-2s}({\RR^d})$).
This singular behavior affects the convergence rate of the finite element
method when the finite element data vector
is approximated using standard numerical quadrature (e.g. Gaussian quadrature).
\end{remark}
\subsection{Numerical Implementation} \label{ss:num_impl}
Based on the notations in Section~\ref{s:truncated},
we set $\Omega=D$ to be either the unit disk in ${\mathbb R}^2$ or $D=(-1,1)$ in ${\mathbb R}$. Let $\Omega^M(t)$ be corresponding dilated domains.
In one dimensional case, we consider $\mathcal T_h(D)$ to be a uniform mesh and $\mathbb V_h(D)$ to be the continuous piecewise linear
finite element space.
For the two dimensional case, $\mathcal T_h(D)$ a regular (in the sense of page 247 in \cite{ciarlet}) subdivision made of quadrilaterals.
In this case, $\mathbb V_h(D)$ is the set of continuous piecewise bilinear functions.
\subsubsection*{Non-uniform Meshes for ${\Omega^M(t)}$}
We extend $\mathcal T_h(D)$ to non-uniform meshes $\mathcal T^M_h(t)$, thereby violating the quasi-uniform assumption.
For $t\le1$, we use a quasi-uniform mesh on $\Omega^M(t)=\Omega^M(1)$ with the same mesh size $h$.
When $t>1$ and $D=(-1,1)$, we use
an exponentially graded mesh outside of $D$,
i.e. the mesh points are $\pm e^{ih_0}$ for $i=1,\ldots,\lceil M/h\rceil$ with $h_0=h(\ln{\gamma})/M$,
where $\gamma$ is the radius of ${\Omega^M(t)}$ (see \eqref{e:omega_Mt}).
Therefore, we maintain the same number of mesh points for all $\Omega^M(t)$.
When $D$ is a unit disk in ${\mathbb R}^2$, we start with a coarse subdivision of ${\Omega^M(t)}$ as in the left of Figure~\ref{f:grid} (the coarse mesh of $D$ in grey). Note that all vertices of a square have the same radial coordinates.
We also point out that the position of the vertices along the radial direction and outside of $D$ follow the same exponential
distribution as in the one dimensional case.
Then we refine each cell in $D$ by connecting the midpoints between opposite edges.
For the cells outside of $D$, we consider the same refinement
in the polar coordinate system $(\ln r,\theta)$ with $r>1$ and $\theta\in[0,2\pi]$.
This guarantees that mesh points on the same radial direction still follows the exponential distribution after global refinements and the number of mesh points in $\mathcal T^M_h(t)$ is unchanged for all $t>0$. The figure on the right of Figure~\ref{f:grid} shows the
exponentially graded mesh after three times global refinement.
\begin{figure}[ht!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=.45]{coarse-eps-converted-to.pdf} & \includegraphics[scale=.32]{refine-eps-converted-to.pdf}\\
\end{tabular}
\end{center}
\caption{Coarse gird (left) and three-times-refined non-uniform
grid (right) of $\Omega^M(t)$ with $M=4$ and $t=1$. Grids of $D$ are in grey.}
\label{f:grid}
\end{figure}
\subsubsection*{Matrix Aspects}
To express the linear system to be solved, we denote by $U$ to be the coefficient vector of $u_h$ and $F$ to be the coefficient vector of the $L^2$
projection of $f$ onto $\mathbb{V}_h(D)$.
Let $M_h(t)$ and $A_h(t)$ be the mass and stiffness matrix in $\mathbb{V}^M_h(t)$.
Denote $M_{D,h}$ to be the mass matrix in $\mathbb V_h(D)$. The linear system is given by
\begin{equation}\label{system_matrix}
\frac{\sin{(\pi \beta)k}}{\pi}\sum_{i=-{N^-}}^{{N^+}}e^{s y_i}M_{D,h}(e^{y_i}M_h(t_i)+A_{h}(t_i))^{-1}A_{h}(t_i)U=F
\end{equation}
with $y_i=ik$ and $t_i=e^{-y_i/2}$.
Here ${M_{D,h}},\ U$ and $F$ are all extended by zeros so that the dimension
of the system is equal to the dimension of $\mathbb{V}^M_h(t)$.
\subsubsection*{Preconditioner}
Since the linear system is symmetric, we apply the Conjugate Gradient method
to solve the above linear system. Due to the norm equivalence between
$( L^2(D),H^1_0(D))_{s,2}$ and $\widetilde{H}^s(D)$, the condition number of the system matrix
is bounded by $Ch^{-2s}$. In order to reduce the number of iterations in one dimensional space,
we use fractional powers of the discrete Laplacian $L_{D,h}$
as a preconditioner, where $L_{D,h} :H^1_0(\Omega)\to L^2(D)$ is defined by
$$d_D(L_{D,h} w,\phi_h)=d_D(w,\phi_h), \qquad \hbox{for all } \phi_h\in \mathbb V_h(D) .$$
This can be computed by the discrete sine transform similar to the implementation
discussed in \cite{BLP16}.
More precisely, the matrix representation of
$L_{D,h}$ is given by ${(M_{D,h})}^{-1}{A_{D,h}}$, where $A_{D,h}$ is stiffness matrix in $\mathbb V_h(D)$.
The eigenvalues of $A_{D,h}$ and $M_{D,h}$ (for the same eigenvectors) are $a_j:=(2+\cos(j\pi h))/h$ and $m_j:=h(4+2\cos(j\pi h))/6$ for $j=1,\ldots,\dim(\mathbb V_h(D))$, respectively.
Therefore, the eigenvalues of $L_h$ are given by $\lambda_{j,h}:=a_j/m_j$.
We use
$$
B:= S\Lambda S
$$
as a preconditioner, where $S_{ij}:= \sqrt{ 2 h}\sin({ij \pi h})$ and $\Lambda$ is the diagonal matrix whose diagonal entries are $\lambda_{j,h}^{-s}/m_j$. We also note that $S^{-1}=S$.
In two dimensional space, we use the multilevel preconditioner advocated in \cite{bramble2000computational}.
\subsection{Numerical Illustration for the Non-smooth Solution}
We first consider the numerical experiments for the model problem \eqref{var1}
and study the behavior of the $L^2(D)$ error.
\subsubsection*{Influence from the Sinc Quadrature and Domain Truncation.}
When $D=(-1,1)$, we approximate
the solution on the fixed uniform mesh with the mesh size $h=1/8192$. The domain truncation parameter $M$ is also fixed
to be $20$. Thus, $h$ is small enough and $M$ is large enough so that the $L^2(D)$-error
is dominant by the sinc quadrature spacing $k$. The left part of
Figure~\ref{f:quad_domain} shows that the $L^2(D)$-error quickly converges
to the error dominant by the Galerkin approximation when $k$ approaches zero. Similar results are observed from
the right part of Figure~\ref{f:quad_domain} when the domain truncation parameter $M$ increases. In this case,
the mesh size $h=1/8192$ and the quadrature step size $k=0.2$.
\subsubsection*{Error Convergence from the Finite Element Approximation}
We note that we implement the numerical algorithm for the two dimensional case using the deal.ii Library
\cite{bangerth2007deal} and we invert matrices in \eqref{system_matrix} using the direct solver from UMFPACK \cite{davis2007umfpack}.
Figure~\ref{f:sol} shows the approximated solutions for $s=0.3$ and $s=0.7$, respectively.
Table~\ref{table:l2err} reports errors $\|u-u_h\|_{L^2(D)}$ and rates of convergence with $s=0.3, 0.5$ and $0.7$. Here the quadrature spacing
($k=0.25$) and the domain truncation parameter ($M=4$) are fixed so that the finite element discretization dominates the error.
We note that
Theorem 7.1 together with Theorem 5.4 in \cite{GG15} (see also Proposition 2.7
in \cite{BPM17}) guarantees that when $\partial D$ is of class $C^\infty$ and $f$ is
in $L^2(D)$, the solution of \eqref{mainWP} is in
$\widetilde H^{s+\alpha^-}(D)$ where
\begin{equation}\label{e:alpha}
\alpha:=\min\{s,1/2\}
\end{equation}
and $\alpha^-$ denotes any number strictly smaller that $\alpha$.
This indicates that the expected rate of convergence in $L^2(D)$ norm should be $\beta+\alpha^--s$
if the solution $u$ is in $\widetilde H^\beta(D)$.
Since the solution $u$ is in $H^{s+1/2-\epsilon}(D)$ (see \cite{AB15} for a proof),
Table~\ref{table:l2err} matches the expected rate of convergence $\min(1,s+1/2)$.
\begin{figure}[ht!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=.48]{quadeffect-eps-converted-to.pdf} & \includegraphics[scale=.48]{domaineffect-eps-converted-to.pdf}\\
\end{tabular}
\end{center}
\caption{The above figures report the $L^2(D)$-error behavior when $D=(-1,1)$.
The left one shows the error as a function of the quadrature spacing
$k$ for a fixed mesh size ($h=1/8192$) and domain truncation parameter ($M=20$).
The right plot reports the error as a function of the domain truncation parameter
$M$ with fixed mesh size ($h=1/8192$) and quadrature spacing
$(k=0.2)$. The spatial error dominates when $k$ is small (left)
and
$M$ is large (right).}
\label{f:quad_domain}
\end{figure}
\begin{table}[h]
\begin{tabular}{|l| l |l| l |l| l |l| l l}
\hline
$\#$DOFS & \multicolumn{2}{c|}{$s=0.3$}& \multicolumn{2}{c|}{$s=0.5$} & \multicolumn{2}{c|}{$s=0.7$} \\
\hline
345 & $2.69\times 10^{-1} $ & - & $1.63\times 10^{-1} $ & - & $1.03\times 10^{-1} $ & - \\\hline
1361 & $1.59\times 10^{-1} $ & 0.7575& $9.07\times 10^{-2} $ & 0.8426 & $5.55\times 10^{-2} $ & 0.8918 \\\hline
5409 & $9.56\times 10^{-2} $ & 0.7323& $5.05\times 10^{-2} $ & 0.8438 & $2.95\times 10^{-2} $ & 0.9091 \\\hline
21569 & $5.71\times 10^{-2} $ & 0.7447& $2.78\times 10^{-2} $ & 0.8633 & $1.54\times 10^{-2} $ & 0.9366 \\\hline
86145 & $3.38\times 10^{-2} $ & 0.7547& $1.51\times 10^{-2} $ & 0.8832 & $7.91\times 10^{-3} $ & 0.9641 \\\hline
344321 & $1.99\times 10^{-2} $ & 0.7644& $8.07\times 10^{-3} $ & 0.9004 & $3.97\times 10^{-3} $ & 0.9936 \\
\hline
\end{tabular}
\caption{$L^2(D)$-errors for different values of $s$ versus the number
of degree of freedom used for the 2-D nonsmooth computations.
$\#$DOFS denotes the dimension of the finite element space ${{\mathbb V}_h^M(t)}.$}
\label{table:l2err}
\end{table}
\begin{figure}[ht!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=.25]{sol03-eps-converted-to.pdf} & \includegraphics[scale=.25]{sol07-eps-converted-to.pdf}\\
\end{tabular}
\end{center}
\caption{Approximated solutions of \eqref{anals} for $s=0.3$ (left) and $s=0.7$ (right) on the unit disk.}
\label{f:sol}
\end{figure}
\subsection{Numerical Illustration for the Smooth Solution}
When the solution is smooth, the finite element error (assuming
the exact computation of the stiffness entries, i.e. no consistency error) satisfies
$$
\|u-u_h\|_{L^2(D)}\le c h^{2-s+\alpha^-},
$$
where $\alpha$ is given by \eqref{e:alpha}.
In contrast, because of the inherent consistency error, our method
only guarantees (c.f.,
Theorem~\ref{t:total_error})
\begin{equation}
\|u-u_h\|_{L^2(D)}\le c h^{3/2-s+\alpha^-} .
\label{expected}
\end{equation}
Table~\ref{t:l2err_dealii} reports $L^2(D)$-errors and rates for the problem \eqref{mainWP} with the smooth solution $u(x)=1-|x|^2$ and
the corresponding right hand side data \eqref{rhs_smooth} in the unit disk. To see the error decay, here we choose the quadrature step
size $k=0.2$ and the domain truncation parameter $M=5$. The observed decay in the error does not match the expected rate \eqref{expected}.
We think this loss of accuracy may be due either to the deterioration of the shape regularity constant in generating the subdivisions of $\Omega^M(t)$ (see Section~\ref{ss:num_impl}) or to the imprecise numerical integration of the singular right hand side in \eqref{rhs_smooth}.
To illustrate this, we
consider the one dimensional problem. Instead of using
\eqref{rhs_smooth} to compute the right hand side vector,
similar to \eqref{onedstiff}, we
compute
\begin{equation} (f,\phi_j)= a(u,\phi_j) = \frac{(\partial_L^{2s-1}\phi_j,u^\prime)_D
+(\partial_L^{2s-1}u,\phi_j^\prime)_D}{2\, \cos(s\pi)} \label{onedRHS}
\end{equation}
with $D=(-1,1)$. We note that when $s< 1/2$, the fractional derivative with the negative power $2s-1$ still makes sense for the local basis function $\phi_j$.
The right hand side of \eqref{onedRHS} can now be computed exactly.
\begin{table}[H]
\centering
\begin{tabular}{|l| l |l| l |l| l |l| l l}
\hline
$\#$DOFS & \multicolumn{2}{c|}{$s=0.3$}& \multicolumn{2}{c|}{$s=0.5$} & \multicolumn{2}{c|}{$s=0.7$} \\
\hline
409 & $6.24\times 10^{-2} $ & - & $9.55\times 10^{-2} $ & - & $1.35\times 10^{-1} $ & - \\\hline
1617 & $2.90\times 10^{-2} $ & 1.10& $4.33\times 10^{-2} $ & 1.14 & $6.27\times 10^{-2} $ & 1.10 \\\hline
6433 & $1.44\times 10^{-2} $ & 1.01& $1.94\times 10^{-2} $ & 1.15 & $2.81\times 10^{-2} $ & 1.16 \\\hline
25665 & $7.21\times 10^{-3} $ & 1.00& $8.55\times 10^{-3} $ & 1.19 & $1.20\times 10^{-2} $ & 1.23 \\\hline
102529 & $3.56\times 10^{-3} $ & 1.02& $3.67\times 10^{-3} $ & 1.22 & $4.78\times 10^{-3} $ & 1.32 \\\hline
409857 & $1.74\times 10^{-3} $ & 1.04& $1.54\times 10^{-3} $ & 1.25 & $1.73\times 10^{-3} $ & 1.47 \\
\hline
\end{tabular}
\caption{$L^2(D)$-errors and rates for $s=0.3$, $0.5$ and $0.7$ for the problem
\eqref{mainWP} with the right hand side \eqref{rhs_smooth}.
$\#$DOFS denotes the number of degree of freedoms of ${\Omega^M(t)}$.}
\label{t:l2err_dealii}
\end{table}
We illustrate the convergence rate for the one dimensional case in Table~\ref{oned1}
when the $L^2(D)$-projection of right hand side is computed from \eqref{onedRHS}.
In this case, we compute at $s=0.3,0.4,0.7$ as the
expression in \eqref{onedRHS} is not valid for $s=0.5$. We also fix $k=0.2$ and $M=6$. In all cases,
we observe the predicted rate of convergence $\min(3/2,2-s)$, see \eqref{expected}.
\begin{table}[H]
\centering
\begin{tabular}{|l| l |l| l |l| l |l| l l}
\hline
$h$ & \multicolumn{2}{c|}{$s=0.3$}& \multicolumn{2}{c|}{$s=0.4$} & \multicolumn{2}{c|}{$s=0.7$} \\
\hline
1/16 & $4.51\times 10^{-4} $ & & $3.47\times 10^{-4} $ & & $9.27\times 10^{-4} $ & \\\hline
1/32 & $1.42\times 10^{-4} $ & 1.58& $1.02\times 10^{-4} $ & 1.77 &$4.16\times 10^{-4} $ & 1.16 \\\hline
1/64 & $4.25\times 10^{-5} $ & 1.63& $3.31\times 10^{-5} $ & 1.62 & $1.80\times 10^{-4} $ & 1.21 \\\hline
1/128 & $1.34\times 10^{-5} $ & 1.66& $1.14\times 10^{-5} $ & 1.54 & $7.66\times 10^{-5} $ & 1.23 \\\hline
1/256 & $4.43\times 10^{-6} $ & 1.59& $4.06\times 10^{-6} $ & 1.49 & $3.21\times 10^{-5} $ & 1.25 \\\hline
1/512 & $1.50\times 10^{-6} $ & 1.56& $1.46\times 10^{-6} $ & 1.48 & $1.33\times 10^{-5} $ & 1.27 \\\hline
\end{tabular}
\caption{$L^2(D)$-errors and rates for $s=0.3$, $0.4$ and $0.7$ for the
one dimensional problem when right hand
side of the discrete problem is computed by \eqref{onedRHS}.}
\label{oned1}
\end{table}
|
\section{Introduction}
Discs and rings are common products of the evolution of different astronomical systems at different size scales, from galaxies
to stars and planets. Circumstellar discs are found around stars at different evolutionary stages and they are commonly detected
as an infrared (IR) excess over the photospheric value, produced by their thermal emission. Debris discs represent the final
stage of disc evolution, they may orbit mature systems and can coexist with a fully-formed planetary system, as in our solar case.
They are formed by second-generation solid material, ranging from sub-mm dust particles to sub-planetary bodies, and, in a
few cases, a small amount of gas has also been observed \citep[e.g.,][]{kospal2013}.
The grain size distribution is determined by the simultaneous processes that form (e.g., collisional grinding) or remove
(e.g. Poynting-Robertson and stellar wind drag) solid particles \citep{wyatt2011}. Thus, this kind of disc continues to
evolve collisionally and dynamically.
In the past decade, several surveys have explored debris discs around stars of spectral types from A to M in the mid- and far-IR,
probing dust at different locations, due to the positive correlation between the wavelength of the peak of the dust emission and the distance from the star\footnote{For a solar luminosity star, observations at 10~$\mu$m detect dust at a distance of $\sim$1~au, at 22~$\mu$m the distance is about 4.5~au, at 60~$\mu$m the distance increases to $\sim$33~au, while at 100~$\mu$m observations probe dust at about 90~au. These distances scale as the square root of the star luminosity.} \citep[e.g.,][]{wyatt2008}.
The most relevant surveys are FEPS\footnote{Formation and Evolution of Planetary Systems.} \citep{rieke2005}, using the {\em Spitzer}
observatory, DEBRIS \citep{sibthorpe2013} and DUNES\footnote{DUst around NEarby Stars.} \citep{eiroa2013}, carried out with the
{\em Herschel} satellite. \citet{matthews2014} made a comparison of detection rates of debris discs among the different
surveys and found that, for A stars, the detection rates are 33$\%$ and 25$\%$ at 70~$\mu$m and 100~$\mu$m, respectively,
while for solar type stars (FGK) it ranges from 10$\%$, in the case of FEPS survey, to 17$\%$, reported in the DEBRIS survey.
DUNES increased the detection rate up to 20$\%$ \citep{eiroa2013}, while the recent work by \citet{montesinos2016} gives a value
of 22\%. \citet{trilling2008} interpreted the apparent decrease of excess rates from A to K spectral types as due to an age effect,
as expected by circumstellar disc evolution.
A new kind of discs was detected thanks to extensive surveys in the Mid-IR regime. These objects are called warm debris discs
and the emission in excess found at wavelengths between 5~$\mu$m and 35~$\mu$m suggests the presence of material close to the star. \citet{chen2005}, \citet{uzpen2007}, and \citet{hales2009} used {\it Spitzer} data, from 8 to 70~$\mu$m, to identify a significant number of flux excesses attributable to the presence of circumstellar warm dust.
More recent studies, that used Wide Infrared Survey Explorer ({\it WISE}) data, have found stars with IR excesses in
the Mid-IR regime \citep{kennedy2012,cruz2014,patel2014,kuchner2016,costa2016}. We might expect that the presence of
planets at few au from the stars (in the ``terrestrial" zone) and warm debris discs are
correlated, since planets are the final stage of agglomeration of planetesimals and the dust in warm debris discs is also
formed from planetesimals. In this work we want to assess if the correlation between the Mid-IR excesses and planets does
exist or not. In this context, the work presented here complements previous analyses, in particular those of \citet{krivov2011} and
\citet{morales2012}, who looked for IR excesses using, respectively, a preliminary {\it WISE} release and the All-Sky {\it WISE} catalogue.
While \citet{krivov2011} found no evidence for IR excesses in 52 transiting exoplanetary systems, \citet{morales2012} found a
prevalence of 2.6$\%$ in a large sample (591) of planetary systems around both main sequence (MS) and
evolved stars; the incidence decreases to 1$\%$ if only MS stars are considered. At longer wavelengths (100~$\mu$m), \citet{moromartin2015} did not find significant difference in the occurrence of dust around stars hosting planets in selected samples of DUNES and DEBRIS objects.
In this paper we analyze a sample of stars with planets and another comparison sample of stars where planets have been
searched for but not found. Infrared data of the four {\it WISE} bands is used to search for IR excesses at 22~$\mu$m by comparing the observed and expected photospheric flux ratio ratio $f_{22}/f_{12}$. Furthermore, we also explore the
issue by analysing different procedures used by other authors
to obtain more results on the existence (or not) of a correlation
between planets and warm debris discs \citep{morales2012,kennedy2012,patel2014,cruz2012}.
\section{{\it WISE}}
{\it WISE} was a NASA mission (2009-2011) whose main objective was to map the entire sky in four Mid-IR passbands at
3.4~$\mu$m (W1), 4.6~$\mu$m (W2), 12~$\mu$m (W3) and 22~$\mu$m (W4). It consisted of a cryogenically cooled
40~cm aperture telescope. It used four focal planes that simultaneously took images of a 47$\times$47 arcmin field of view.
The full-width-at-half- maximum (FWHM) of the point spread function (PSF) were 6.1'' for W1, 6.4\arcsec for
W2, 6.5\arcsec for W3 and 12.0\arcsec for W4.
{\it WISE} photometry of point sources was conducted using PSF profile fitting and the background level was estimated in a
ring at a fixed distance (form 50 to 70\arcsec) from the source. Two comprehensive photometric catalogues have been delivered.
The first one, called All-Sky \citep{cutri2012}, covered more than 90$\%$ of the sky. The second data release, named
All{\it WISE} \citep{cutri2013}, was delivered in November 2013 and included combined data of previous epochs of {\it WISE}, such as
NEO{\it WISE}, which also provided photometry measurements of asteroids and comets, and Cryogenic {\it WISE}, which obtained data
before their cooling systems ended operations. Taking into account these combinations, the sensitivity and the
photometric precision was improved as compared to the previous All-Sky catalogue.
{\it WISE} achieved sensitivities of $5\sigma$ for point sources with flux of 0.08, 0.11, 1 and 6~mJy (16.5, 15.5, 11.2
and 7.9 Vega magnitudes, respectively) in the four bands, from W1 to W4. This allowed to detect, in W1, faint
objects that do not appear in the 2MASS $K_s$ survey, while at 12~$\mu$m and 25~$\mu$m the {\it WISE} survey is $\sim$100
times more sensitive than IRAS \citep{neugebauer1984}.
\section{Homogeneising {\it WISE} and synthetic photometry.}
\label{sec:fluxcorr}
Prior to the search of IR excesses, we tested the agreement between the observed {\it WISE}
fluxes, provided by the All{\it WISE} catalogue, and the ATLAS9 theoretical photospheric flux \citep{castelli2003} that
we used as reference, using a very large sample of stars.
\subsection{The sample selection.}
\label{selectioncriteria}
We selected stars with spectral types from A0 to K9 and luminosity classes V, IV, or IV/V from the SIMBAD
Astronomical Database\footnote{http://simbad.u-strasbg.fr/simbad/}. The Johnson V magnitude of these stars ranges
from 5 to 12, which make them prone to have a good signal-to-noise ratio (S/N $\geq$ 5) in the {\it WISE} band W4. We
then discarded all objects in multiple systems and pre-main sequence stars (T-Tauri, Ae/Be Herbig).
We assembled an initial sample of 53\,349 objects.
We collected their {\it WISE} magnitudes from the All{\it WISE} catalogue, considering only objects that comply the following
constraints: a) they must have a S/N$\geq$5 in W4 band; b) their W1, W2, W3 and W4 photometry must include an
uncertainty measurement; c) they must be free from any artifact or contamination by nearby objects (i.e. we excluded those sources considered as spurious detections or contaminated by diffraction spikes, scattered light halos or optical ghosts produced by a bright star on the same image, according to the criteria of allWISE catalogue);
d) the saturation in W3 and W4 must be zero.
We also extracted the J, H, and $K_S$ magnitudes from the 2MASS All-Sky catalogue. The restrictions applied on this
catalogue are as follows: a) J, H and $K_S$ photometry must be accompanied by the corresponding uncertainties;
b) the quality flag of the photometry must be A or B, with a S/N$\geq$10 and S/N$\geq$7, respectively.
Finally, we obtained the B and V photometry from SIMBAD. After applying the previous criteria, we obtained a
sample of 24,117 stars.
\subsection{Correction for extinction and magnitude to flux conversion}
Due to their brightness range, we expect most of the stars in our sample to be nearby objects; however,
we considered appropriate to apply a flux correction for those objects that show a considerable colour excess E(B-V).
Correction for interstellar extinction will reduce systematic errors in the determination of stellar parameters for
which we use optical and near-IR photometry.
\begin{table}
\begin{center}
\caption{Zero magnitude flux density for B, V \citep{bessel1979}, J, H, $K_S$ (from http://ipac.caltech.edu/2mass),
W1, W2, W3, W4 (from http://wise2.ipac.caltech.edu) bands.}
\label{zeropoint}
\begin{tabular}{l r@{}l}
\hline
Band & \multicolumn{2}{c}{\bf $F_0$} \\
& \multicolumn{2}{c}{(Jy)} \\
\hline
B (Johnson) & 4266.&7 \\
V (Johnson) & 3836.&3 \\
J (2MASS) & 1594.&0 \\
H (2MASS) & 1024.&0 \\
$K_S$ (2MASS) & 666.&7 \\
W1 ({it WISE}g) & 306.& 7 \\
W2 ({it WISE}g) & 170.& 7 \\
W3 ({it WISE}g) & 29.& 0 \\
W4 ({it WISE}g) & 8.& 3\\
\hline
\end{tabular}
\end{center}
\end{table}
We first obtained the intrinsic colour $(B-V)_0$ for each object, based on its spectral type, from the colour calibration of
\citet{peacut2012}, and derived the colour excess $E(B-V)$. We computed the mean $\mu_{\rm E(B-V)}$ and the stardard deviation
$\sigma_{\rm E(B-V)}$ of the $E(B-V)$ distribution, after applying a $3\sigma$-clipping to the data. We only applied a
reddening correction, to all wavelength bands considered in this work, for the stars with
$E(B-V)>\mu_{\rm E(B-V)} + 1 \sigma_{\rm E(B-V)}$= 0.078~mag. The mean value of the colour excess distribution is 0.026~mag.
A total of 9230 stars were corrected for extinction through the equation:
\begin{equation}
f_\lambda = f_0 10^{ 0.4 (-m_\lambda+A_\lambda) } \, ,
\end{equation}
that also allows to convert apparent magnitudes $m_\lambda$ to fluxes $f_\lambda$. We used the
zero magnitude flux $f_0$ in the Vega scale reported in Table~\ref{zeropoint} and the $A_\lambda$ values provided by
the \citet{rieke1985} extinction curve. No colour correction was applied to {\it WISE} data, since the reddening at micron
wavelengths is negligible.
\subsection{Photospheric fitting and synthetic photometry.}
\label{photfitting}
Since the search of IR excesses requires a good estimation of the photospheric contribution of each star, synthetic spectra are
needed. In the near and mid IR, the choice of a specific model atmosphere library is not critical: \citet{sinclair2010} found
that ATLAS9 \citep{castelli2003} and MARCS \citep{gustafson2008} present discrepancies smaller than $2\%$ at 24~$\mu$m for FGK stars,
and \citet{cruz2012} obtain similar figures comparing ATLAS9 and NEXTGEN SEDs. The stellar spectra in the selected parameter space
of the sample are well reproduced by the LTE plane-parallel, low resolution ATLAS9 model atmospheres
\citep[see, e.g.,][]{bertone2004,martins2007}. In order to compute the synthetic photometry that we needed to determine the
photospheric fluxes, we first estimated the atmospheric parameters of each star through a match with a grid of ATLAS9 library.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=9cm]{w11n.eps} &
\includegraphics[width=9cm]{w22n.eps} \\
\includegraphics[width=9cm]{w33n.eps} &
\includegraphics[width=9cm]{w44n.eps} \\
\end{tabular}
\caption{Flux ratio $f_{\rm Wi,obs}/f_{\rm Wi,syn}$ vs. $f_{Wi,obs}$ for all 24,117 stars (dots). The starred symbols represent
the median for each flux bin with their respective uncertainty, the thick line shows the best linear fit while the dashed line indicates the reference flux ratio equal to one.}
\label{w2med}
\end{center}
\end{figure*}
We derive the effective temperature ($T_{\rm eff}$) by identifying the theoretical spectrum that best match the observed
B, V, J, H, and $K_s$ fluxes\footnote{We extract filter transmission functions from
http://voservices.net/filter/ and http://www.ipac.caltech.edu/2mass.}, that we assume as purely photospheric,
using the MPFIT $\chi^2$ minimization algorithm \citep{markwidth2009}. The procedure also provides the multiplying
factor $c_n$ that takes into account the flux density dependence on distance. Since the surface gravity ($\log{g}$) and
overall metallicity ([M/H]) generate small flux changes in the Rayleigh-Jeans regime \citep[see][]{cruz2012}, we fixed
their values to $\log{g}$=4.0 for main sequence (MS) A and F stars and for stars with luminosity class IV or IV-V, and
$\log{g}$=4.5 for MS G and K stars. We assumed solar metallicity in all cases.
We used the theoretical photospheric SED of each star to compute the {\it WISE} synthetic
photometry\footnote{http://http://wise2.ipac.caltech.edu/docs/release/allsky/expsup}. In Fig.~\ref{w2med} we plot
the results in the form of the flux ratio between the observed and the photospheric value as a function of the flux
value for W1, W2, W3, and W4 bands for all 24,117 stars. We note a significant discrepancy in all cases. It is smaller for the W1 band and
it decreases towards brighter objects, apart for the case of W2 band, where it steeply increases with increasing flux.
To quantify this discrepancy, we divided the sample in flux bins of 0.1~Jy or wider, when the number of objects per
bin fell short of 40; we then computed the median flux ratio for each bin and we estimated its uncertainty as half the
difference between the 25$\%$ and 75$\%$ quartiles. We used these data to search for the best linear fit using the MPFIT
algorithm, with the slope $m$ and the constant $b$ as free parameters in the linear
equation $f_{\rm i, obs}/f_{\rm Wi,syn}=mf_{\rm i, obs}+b$, where $i$ indicates the {\it WISE} band. The use of the median value
minimizes the impact of the objects with a real IR excess on the fit, that therefore should be representative of the
majority of stars whose flux emerges from the photosphere. The results, reported in Table~\ref{param}, show a similar
behavior of the two redder bands: they have a 10$\%$ flux overestimation over the synthetic value at low fluxes and
this discrepancy gradually diminishes for the brigther objects. Conversely, for the case of the W2 band, the
theoretical and observed flux match well for the faint objects, but increases to about a 40$\%$ overestimation
at the brightest end.
Since, this discrepancy may significantly affect the determination of IR excesses, we decided to transform the {\it WISE}
observed fluxes to the theoretical system, by means of the following equation:
\begin{equation}
f^\prime_{\rm i, obs} = \frac{ f_{\rm i, obs} }{ mf_{\rm i, obs}+b },
\end{equation}
where $f^\prime_{\rm i, obs}$ is the transformed flux that we use to explore the {\it WISE} database in search of IR excesses
in stars with and without planets.
\begin{table}
\begin{center}
\caption{Linear fit parameters.}
\label{param}
\begin{tabular}{c c c }
\hline
{\it WISE} band & {\bf $m$} &{\bf $b$} \\
\hline
W1 & -0.008$\pm$0.004 & 1.045$\pm$ 0.009\\
W2 & 0.105$\pm$0.012 & 1.017$\pm$ 0.010 \\
W3 & -0.145$\pm$0.045 & 1.098$\pm$ 0.008 \\
W4 & -0.672$\pm$0.206 & 1.107$\pm$ 0.016 \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Searching for IR excesses in stars with and without planets }
\label{irno}
\subsection{Sample of stars with and without planets}
We constructed the sample of stars hosting planets from ``The Extrasolar Planet Encyclopedia" web page\footnote{http://exoplanet.eu}.
As of November 2016, the database included 2609 stars with confirmed planets. We applied to this sample the same criteria, on {\it WISE}
and 2MASS photometry, described in Sec.~\ref{selectioncriteria} and we obtained a final sample of 236 objects. In order to avoid the inclusion of spurious detections, we also proceeded to perform a careful visual inspection of the W4 images of all these stars.
We should mention that
some well known sources hosting debris discs like Fomalhaut and $\beta$~Pic were not considered in the sample because these
stars are very bright sources and presented saturated photometry, thus they could not pass the selection criteria defined in section
\ref{selectioncriteria}.
To build a comparison sample of stars without planets, we considered objects that were included in planet-search projects but
for which no planetary companions were found. We considered the works by \citet{sousa2008, sousa2011, adibekyan2012, santos2011,
bertran2015}, that used the High Accuracy Radial velocity Planet Searcher (HARPS) spectrograph, the sample of \citet{valenti2005},
that made use of observations with the Keck, the Anglo-Australian and the Lick telescopes, and the study of \citet{wittenmyer2011},
who also used the Anglo-Australian telescope as part of the Pan-Pacific Planet Search programme. We eliminated the duplicate entries
and those stars for which a planet has been discovered after the date of publication. To this initial list of 1850 objects we also
applied the same selection criteria as before and the sample of stars without planets ended with 986 stars.
In Fig.~\ref{fig24} we plot the distribution of the two samples with respect to spectral type, distance, apparent V magnitude and
metallicity ([Fe/H]). We applied a two-sample Kolmogorov Smirnov (KS) test to evaluate whether the samples of stars with and
without planets share the same parent population, considering a confidence level threshold of 99$\%$ \citep{massey1952}. This
hypothesis was verified in the case of the spectral type and V magnitude distributions, while discarded in the distance and
metallicity cases. As far as the distance is concerned, the distribution of Fig.~\ref{fig24} shows that the sample of stars
without planets is composed on average by more nearby stars with a mean of 50~pc while the other sample has a mean distance of
75~pc. The cut at the larger distance is due to the photometry criteria that we imposed. In the case of the metallicity, the
discrepancy between the two samples is more evident: the average metallicity of the stars without planets is [Fe/H]=$-$0.09,
in agreement with the typical value in the solar vicinity \citep[e.g.,][]{casagrande2011,haywood2001}, while the stars with
planets are more metal-rich, with an average [Fe/H]=+0.04. This results concords with the positive correlation that exists
between the presence of giant planets and the metal content of the host stars \citep[e.g.,][]{gonzalez1997,valefisch2005}.
Since about 60$\%$ of stars in our sample have planets with mass $\geq 1$~M$_J$, we expected to find this metal abundance
difference in our samples.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=8cm, height=5.5cm]{espectrali.eps} &
\includegraphics[width=8cm, height=5.5cm]{distance.eps} \\
\includegraphics[width=8cm, height=5.5cm]{vmagnit.eps} &
\includegraphics[width=8cm, height=5.5cm]{metali.eps} \\
\end{tabular}
\caption{Distributions of the 236 stars with planets and of the 986 stars without planets with respect to spectral type, distance, apparent V magnitude and metallicity.}
\label{fig24}
\end{center}
\end{figure*}
\subsection{Looking for IR excesses.}
In this section, we describe the method we devised for identifying stars with IR excess from the All{\it WISE} data. To date,
there have been several works that used {\it WISE} data with the same goal. They differ from our work and among them in the
definition of IR excess. We therefore considered convenient to also incorporate the procedures used by \citet{cruz2014},
\citet{kennedy2012}, \citet{morales2012} and \citet{patel2014} in order to understand how their differences affect the
number of detected IR excesses.
\subsubsection{Excess definition 1: this work.}
We define a IR excess value $E$ as the difference of the $f_{22}$/$f_{12}$ observed flux ratio and the respective stellar
photospheric ratio, divided by the associated uncertainty:
\begin{equation}
E = \frac{\bigl(\frac{f_{22}}{f_{12}} \bigr)_{obs}-\bigl(\frac{f_{22}}{f_{12}}\bigr)_{syn}}{\sigma_{tot}},\quad \sigma_{tot}=\sqrt{\sigma_{obs}^2+\sigma_{cal}^2+\sigma_{syn}^2},
\end{equation}
where $(\frac{f_{22}}{f_{12}})_{obs}$ is the observed flux ratio between W4 and W3 bands, $(\frac{f_{22}}{f_{12}})_{syn}$ is the synthetic flux ratio.
$\sigma_{tot}$ is the quadratic sum of 3 error sources: 1) the photometric contribution $\sigma_{obs}$, obtained
from the {\it WISE} catalogue; 2) the calibration uncertainty of {\it WISE} bands $\sigma_{cal}$ of 4.5$\%$ and 5.7$\%$ for
W3 and W4, respectively \citep{jarret2011}; 3) the uncertainty due to the stellar parameters $\sigma_{syn}$.
This latter value is obtained by first constructing a family of synthetic spectra, whose parameters were a
combination of $T_{\rm eff}\pm\sigma_{\rm Teff}$ and $c_n\pm\sigma_{c_n}$ (see Sec.~\ref{photfitting}). Then,
the synthetic flux is calculated for each spectrum and the standard deviation of the flux distribution is
considered as $\sigma_{syn}$. An infrared excess is considered significant when $E$ is larger than a chosen threshold.
We adopted the $f_{22}$/$f_{12}$ ratio because these 2 bands show a quite similar trend in the flux ratio between observed and
photospheric flux (see Fig.~\ref{w2med}). However, this choice may affect the detection of the hotter discs, if
they emit a non negligible amount of flux at 12~$\mu$m.
\subsubsection{Excess definition 2: \citet{cruz2014}.}
The authors searched for IR excesses in a sample of dwarf stars brighter than V=15~mag, using the
{\it WISE} All-Sky survey database for over 9000 sources. They compared the observed $f_{22}$/$f_{4.6}$ flux ratio with the
expected photospheric level and identified 197 significant excesses at 22~$\mu$m. Assuming that the excess
is caused by thermal emission of a circumstellar dust disc, they found more than 80$\%$ of the sample with
dust temperatures higher than 120~K. The definitions of flux excess and its error used by \citet{cruz2014}
is the same as those adopted in this work, with the only difference that the flux ratio they considered is $f_{22}$/$f_{4.6}$.
\subsubsection{Excess definition 3: \citet{kennedy2012}.}
They analyzed 180\,000 stars observed by {\it WISE} in the Kepler field. After carefully taking
into account contamination by background sources, they identify only one bonafide excess at 22~$\mu$m due
to a debris disc around the A-type star KIC\,7345479. They defined an excess as the difference of the observed
flux in any {\it WISE} band and the correspondent photospheric flux, divided by the total error:
\begin{equation}
E=\frac{f_{i,obs}-f_{i,syn}}{\sigma_{tot}}
\end{equation}
where $\sigma_{tot}$ is the quadratic sum of the observed photometric uncertainty $\sigma_{obs}$ and the
uncertainty of the expected stellar flux, equivalent to the error produced by the uncertainty on stellar
parameters $\sigma_{syn}$. They adopted a $4\sigma$ threshold.
\subsubsection{Excess definition 4: \citep{morales2012}.}
This work explored exoplanet-host stars in sear\-ching for the presence of warm dust.
They analyzed a sample of 591 stars from the Extrasolar Planet Encyclopedia (January 2012) and found
9 planet-bearing stars with excesses at 12~$\mu$m and 22~$\mu$m in young MS and giant stars.
To detect IR excess, after obtaining the synthetic spectrum that better fit the photosphere, they
calculated the significance of the excess using the total signal-to-noise ratio (S/N) defined as:
\begin{equation}
\biggl(\frac{S}{N}\biggr)=\frac{f_{obs}}{\sigma}=\biggl(\frac{S}{N}\biggr)_{syn}+\biggl(\frac{S}{N}\biggr)_d,\quad \quad \quad \sigma=\sqrt{\sigma_{obs}^2+\sigma_{cal}^2} \, ,
\end{equation}
with $f_{obs}=f_{syn}+f_d$, where $f_{syn}$ is the stellar synthetic flux of the best fit, $W_d$ is the flux of a
debris disc in any of the {\it WISE} bands, and $\sigma$ takes into account the uncertainty of the {\it WISE} photometry
($\sigma_{obs}$) and calibration ($\sigma_{cal}$).
The authors considered an infrared excess when $(S/N)_d\geq 3$.
\subsubsection{Excess definition 5: \citep{patel2014}.}
The last work we analyzed is \citet{patel2014}, where they searched for excesses, in the {\it WISE} database at 12~$\mu$m and 22~$\mu$m,
of Hipparcos MS stars within 75~pc from the Sun. Since they were dealing with many bright objects, they apply
corrections for properly handling with saturated objects in W1 and W2 bands. They found 379 stars with flux excesses at 22~$\mu$m.
The procedure they devised for detecting infrared excesses is quite different than the previous ones.
\citet{patel2014} based their searching process on empirical relations found in colour-colour diagrams of
Tycho-2 $B-V$ vs. $(W_i-W_j)$, with $i,j$ representing two different {\it WISE} bands. They did not carry out a
fitting on the star SED like the previous methods to determine the photospheric flux level, however, they
fit the star SED to confirm the validity of the detected excesses.
The authors have calibrated the relations $W_i-W_j$ vs. $B-V$ by iteratively removing the largest colours
$W_i-W_j$, for each $B-V$ bin of 0.1~mag, until half of the data points in the bin are rejected. The different
relations $W_i-W_j$ vs. $B-V$ are traced in steps of 0.02~mag in $B-V$. Thus, the mean $W_i-W_j$ corresponding
to a certain $B-V$ colour is referred as $W_{ij}(B-V)$. The list of the mean $W_{ij}(B-V)$, with its respective
standard error, for all colour combinations is given also in \citet{patel2014}. Taking into account the colour
dependency, an excess $E[W_i-W_j]$ in the colour $W_i-W_j$ for a certain $B-V$ is defined as:
\begin{equation}
E[W_i-W_j]=W_i-W_j-W_{ij}(B-V) \, ,
\end{equation}
while, the S/N of this excess is:
\begin{equation}
\frac{S}{N}=\frac{E[W_i-W_j]}{\sigma_{ij}}=\frac{W_i-W_j-W_{ij}(B-V)}{\sigma_{ij}}
\end{equation}
where $\sigma_{ij}$ is the propagation of the photometric uncertainties of $W_i$ and $W_j$,
together with the standard error $W_{ij}(B-V)$; thus,
\begin{equation}
\sigma_{ij}=\sqrt{\sigma_{Wi}^2+\sigma_{Wj}^2+\sigma_{Wij}^2} \, .
\end{equation}
They calculated the ratio of detecting false positives (FPR) by using the empirical S/N distribution and defined FPR as the number of stars beyond the threshold in which the outliers were considered as reliable excesses.
This enabled to detect excesses as a function of the threshold beyond which the redder objects can be considered
as reliable excesses. Nevertheless, it was not possible to determine empirically the FPR beyond the limit where
the number of false positives drops to zero; for this, an upper limit was needed. To define this limit, the
distributions of S/N, which involved W4 as $W_j$, were constructed and the threshold was located between
99.8$\%$ and 99.9$\%$ for considering a possible excess.
\section{Results}
We applied all five procedures described in the previous section to the search for IR excess in our
two stellar samples of stars with and without planets, using the All{\it WISE} catalogue \citep{cutri2013}.
We present the results in Table~\ref{tab41}; for sake of homogeneity, we used the same threshold
for all methods and we assumed both $3\sigma$ and $4\sigma$ as detection limits. We report the
number and percentage of significant IR excesses detected using both the
modified {\it WISE} flux (as explained in Sec.~\ref{sec:fluxcorr}), in columns 2--3 and 6--7, and the original All{\it WISE}
photometry, in columns 4--5 and 8--9.
\begin{table*}
\begin{center}
\caption[Number and percentage of IR excesses in stars with and without planets.]{\footnotesize Number and percentage of
IR excesses in stars with and without planets, using 3$\sigma$ (top) and 4$\sigma$ (bottom) threshold.}
\label{tab41}
\begin{tabular}{l r@{}l c c c r@{}l c c c}
\noalign{\smallskip} \hline \noalign{\smallskip}
{ IR excess definition} &\multicolumn{5}{|c|}{ Stars with planets}&\multicolumn{5}{|c|}{ Stars without planets}\\
\hline \noalign{\smallskip}
& \multicolumn{3}{|c|}{corrected flux} & \multicolumn{2}{|c|}{All{\it WISE} flux} & \multicolumn{3}{|c|}{corrected flux} & \multicolumn{2}{|c|}{All{\it WISE} flux} \\
\hline \noalign{\smallskip}
& \multicolumn{2}{|c|}{\#} & $\%$ & \# & $\%$ & \multicolumn{2}{|c|}{\#} & $\%$ & \# & $\%$ \\
\hline \noalign{\smallskip}
\multicolumn{11}{|c|}{3$\sigma$ threshold}\\
\hline \noalign{\smallskip}
This work & 2&$^{a,b}$ & 0.85 & 2 & 0.85 & 1&$^d$ & 0.10 & 1 & 0.10 \\
Cruz-Saenz de Miera et al. 2014 & 1&$^a$ & 0.42 & 1 & 0.42 & 1&$^e$ & 0.10 & 1 & 0.10 \\
Kennedy $\&$ Wyatt 2012 & 2&$^{a,b}$ & 0.85 & 6 & 2.54 & 4&$^{d,e,f,g}$ & 0.41 & 35 & 3.55 \\
Morales et al. 2012 & 2&$^{a,b}$ & 0.85 & 2 & 0.85 & 3&$^{d,e,f}$ & 0.30 & 5 & 0.51 \\
Patel et al. 2014 & 3&$^{a,b,c}$ & 1.27 & 3 & 1.27 & 5&$^{d,e,f,g,h}$ & 0.51 & 5 & 0.51 \\
\noalign{\smallskip} \hline \noalign{\smallskip}
\multicolumn{11}{|c|}{4$\sigma$ threshold}\\
\hline \noalign{\smallskip}
This work & 1&$^a$ & 0.42 & 1 & 0.42 & 0& & 0 & 0 & 0 \\
Cruz-Saenz de Miera et al. 2014 & 1&$^a$ & 0.42 & 1 & 0.42 & 0& & 0 & 0 & 0 \\
Kennedy $\&$ Wyatt 2012 & 2&$^{a,b}$ & 0.85 & 2 & 0.85 & 4&$^{d,e,f,g}$ & 0.41 & 9 & 0.91 \\
Morales et al. 2012 & 2&$^{a,b}$ & 0.42 & 2 & 0.85 & 0& & 0 & 3 & 0.30 \\
Patel et al. 2014 & 2&$^{a,b}$ & 0.85 & 2 & 0.85 & 4&$^{d,e,f,g}$ & 0.41 & 4 & 0.41 \\
\noalign{\smallskip} \hline
\multicolumn{11}{|l|}{$^a$HD106906, $^b$HD218396, $^c$BD-10\,3166, $^d$HD85301, $^e$HD136544, $^f$HD107146, $^g$HD60491, $^h$HD11938} \end{tabular}
\end{center}
\end{table*}
The flux homogeneisation between observed and theoretical fluxes (Sec.~\ref{sec:fluxcorr}) decreases the number
of detected IR excesses in those cases where the excess is defined through a difference between observed and
photospheric value, as in \citet{kennedy2012} and \citet{morales2012}.
It has a lower effect on the other methods, based on colours, and, even though the significance of the IR excesses
changes, it is not enough to modify the number of objects whose excess surpasses the detection threshold.
In Table~\ref{tab41}, we also report the identification of the stars with IR excess detected with each definition.
The one adopted in this work allowed to detect 2 excesses over the photosphere at 22~$\mu$m in stars with planets (0.85$\%$ of the sample) and just 1 excess (0.1$\%$) in stars that do not harbour a detected planet, assuming a 3$\sigma$ threshold. The excess definition of \citet{cruz2014} provides even lower numbers, while
the results from the method by \citet{patel2014} include all objects detected with the other definitions (with a detection rate of 1.27$\%$ and 0.51$\%$ in stars with and without planets, respectively) and adds
two unique stars (BD-10\,3166 and HD11938), if the lower significance threshold of 3$\sigma$ is considered.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=15cm, height=10cm]{ratiosn.eps}
\end{tabular}
\caption{Flux ratio diagram of {\it WISE} photometry at 22~$\mu$m and 12~$\mu$m versus
the 2MASS J and $K_s$ fluxes for the 1,222 stars with and without planets, as explained in the inset. Objects with IR excess refer to the stars reported in Table 3. The location of HD 106906 is indicated in the upper inset.}
\label{figcolor}
\end{center}
\end{figure*}
In Fig.~\ref{figcolor} , we show the flux ratio diagram using the 2MASS J and K$_s$ fluxes for the 1,222 stars with and without planets used in this work and the {\it WISE} fluxes at 12~$\mu$m and 22~$\mu$m. We highlight the 8 stars with Mid-IR excess reported in Table~\ref{tab41} with circles.
The SED of all 8 objects found by the method of \citet{patel2014}, using a 3$\sigma$ threshold, are depicted in Figs.~\ref{fig:sed1} and \ref{fig:sed2}.
{\em Spitzer} photometry \citep{chen2005} confirms the presence of an IR excess for 5 stars
[$^a$HD106906, $^b$HD218396 (HR8799), $^d$HD85301, $^f$HD107146, $^g$HD60491], with a clear indication that the flux
peak is located at wavelengths longer than 22~$\mu$m. In fact, apart from the case of HD106906 and HD60491,
{\em Spitzer} data indicate that the W4 band coincides with the wavelength range where the IR excess begins to appear above the photosphere, with a $>$3$\sigma$ significance, according to the blackbody modeling by \citet{chen2014}.
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[width=5.7cm, height=4.4cm]{hd106906n.eps} &
\includegraphics[width=5.7cm, height=4.4cm]{hd218396n.eps} &
\includegraphics[width=5.7cm,height=4.4cm]{bd-103166n.eps} \\
\end{tabular}
\caption{SEDs of the stars with planets for which a significant flux excess has been detected in the {\it WISE} data.
HD106906, HD218396, BD-10\,3166. We show in black dots the {\it WISE} photometry while in grey dots optical and
{\it Spitzer} photometry. Continuous line represents the best atmospheric model fit. }
\label{fig:sed1}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[width=5.7cm, height=4.4cm]{hd85301n.eps} &
\includegraphics[width=5.7cm, height=4.4cm]{hd136544n.eps} &
\includegraphics[width=5.7cm, height=4.4cm]{hd107146n.eps} \\
\includegraphics[width=5.7cm, height=4.4cm]{hd60491n.eps} &
\includegraphics[width=5.7cm, height=4.4cm]{hd11938n.eps} \\
\end{tabular}
\caption{SEDs of the stars without planets for which a significant flux excess has been detected in the {\it WISE}
data. HD85301, HD136544, HD107146, HD 60491, HD11938. We show in black dots the {\it WISE} photometry while in grey
dots optical and {\it Spitzer} photometry. Continuous line represents the best atmospheric model fit.}
\label{fig:sed2}
\end{figure*}
The excess definitions used in this work, in \citet{cruz2014} and in \citet{morales2012} provide the lower number
of detections, which do not include all objects whose excesses are supported by {\em Spitzer} observations. These
are the only approaches that include the {\it WISE} flux calibration error in the computation of the uncertainty.
As an example, in Fig.~\ref{fig:error} we illustrate the impact that this source has on the total uncertainty
on the flux ratio $f_{22}$/$f_{12}$ for the 1,222 stars with and without planets: calibration error is dominant for objects with $f_{22}\gtrsim$20~mJy, which is the case
for all stars with {\em Spitzer} data. This fact seems to point out that the {\it WISE} calibration error may be
overestimated.
In general, the stars which present IR excess in this work share similarities. In the case of HD218396 (HR8799),
a well known planetary system with 5 planets, a debris disc with a cold and warm component is
found \citep{ballering2013}. The same happens to the stars without planets HD85301 and recently to HD107146,
where \citet{macgregor2016} add a cold component (30K) to its disc, in agreement to the {\em Spitzer} photometry
that also shows the excesses for these stars \citep{chen2005}. Some other debris disc were already found by
\citet{patel2014} (HD136544, HD60491, HD11938) where discs within 95K and 190 K are modeled. HD106906 is a
spectroscopic binary, in the Lower Centaurus Crux stellar association, that hosts a well known debris disc,
recently observed with adaptive optics systems \citep{wu2016,lagrange2016}. Its strong
excess in the W4 band is in perfect agreement with the {\it Spitzer} photometry of \citet{chen2005}, whose
blackbody fit of the excess provides a temperature of 90~K \citep{chen2005}. Only BD-10\,3166 is not catalogued
previously as a debris disc host. Furthermore, \citet{lodieu2014} report this star with a M5 star as a large
proper motion companion.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{err.eps}
\caption{Sources of uncertainty in the flux ratio $f_{22}$/$f_{12}$ as a function of the flux at 22~$\mu$m for the complete sample of 1,222 stars with and without planets. The errors due to {\it WISE}
photometry (green dots), flux calibration (red) and uncertainty on the stellar photospheric parameters (blue)
are shown separately, along with the total error (yellow).}
\label{fig:error}
\end{center}
\end{figure}
We performed a Welch's t-test \citep{welch1947}, suitable for data populations with different variance, to
quantify the statistical significance of the difference in the number of IR excess detections between the
two sample of stars, with or without planets. We consider the results obtained by the five excess definitions
analyzed in this work, using the ``corrected" {\it WISE} photometry. We also take into account two detection
thresholds of 3$\sigma$ and 4$\sigma$. In no case the test indicates that the difference in the percentage of
IR excess detections is statistically significant, assuming a confidence level of 95\%.
Our analysis therefore indicates that the presence of warm debris desks (assuming that all detected excesses
are caused by such an object) are not correlated to the presence of planets.
Finally, it is important to note that the detection rate of warm debris desks of $\lesssim$1\% is much lower
than that of cold debris discs, making them a rare phenomena. As \citet{wyatt2008} mentioned, in fact, the
rare finding of warm debris discs in FGK MS stars is due to a combination of their long lifetimes and the
short periods of time in which dust is produced by a transient event.
\section*{Acknowledgements}
The authors thank the referee Antonio Hales for comments that
improved the presentation of the paper. We would like to thank also Grant Kennedy for useful comments.
RFM, MC, EB and FC want to thank CONACyT for financial support through grants CB-2011-169554 and CB-2015-256961.
This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
\bibliographystyle{mnras}
|
\section{introduction}
Let $R$ be a commutative Noetherian local ring with maximal ideal $\mathfrak{m}$ and residue field $k = R / \mathfrak{m}$. Let $M$ be a finitely generated module over $R$.
The Poincar\'e series $P^R_M(t)$ of $M$ over $R$ is a formal power series in ${\NZQ Z}[|t|]$ defined as follows:
$$P^R_M(t) = \sum_{i \geq 0} \beta^R_i(M) t^i \in {\NZQ Z}[|t|],$$
where $\beta^R_i(M) = \dim_k \Tor^R_i(M, k)$ denotes the $i$-th Betti number of $M$.
In the 1950's Serre and Kaplansky asked whether the Poincar\'e series $P^R_k(t)$ is a rational function. This question ties in closely with an analogous question in Algebraic topology attributed to Serre on the rationality of the Poincar\'e series of loop spaces of finite simply connected CW complexes (see \cite{roos1}). Several authors made an attempt to find an affirmative answer. However in 1982 Anick found a counterexample answering the question in negative \cite{anick}. Research on this topic intensified since the appearance of Anick's example. B\o gvad reworked Anick's example and produced an Artinian Gorenstein local ring whose residue field has an irrational Poincar\'e series \cite{bogvad}. Jacobsson constructed a local ring $R$ with residue field $k$ and a cyclic R-module M such that $P^R_k (t)$ is rational, but $P^R_M(t)$ is irrational \cite[Corollary C]{jacob}. Roos defined a ring $R$ to be good if
there exists a polynomial $d_R(t) \in {\NZQ Z}[t]$ such that $d_R(t)P_M^R (t) \in {\NZQ Z}[t]$ for every finitely generated $R$-module $M$ \cite[Definition 2.1]{roos2}. He constructed bad (i.e. not good) affine rings defined by quadratic monomial relations \cite[Theorem 2.4]{roos2}.
Nevertheless there are an abundance of good rings, e.g. Golod rings \cite[Theorem 1]{gul2}, complete intersections \cite[Corollary 4.2]{gul1}, rings R with $\edim(R) - \depth(R) \leq 3$, Gorenstein rings $R$ with $\edim(R) - \depth(R) \leq 4$ \cite[Theorem 6.4]{av4}, etc. (see \cite{av} for more examples). Mccullough and Peeva mentioned in \cite[Example 6.4]{mcc} that to date we do not have a good understanding of how likely it is that rings with residue fields having rational or irrational Poincar\'e series occur.
One of the main objectives of the present article is to show that at low multiplicities Gorenstein local rings are always good.
A rational expression for a Poincar\'e series $P^R_M(t)$ has practical applications. If $P^R_M(t)$ is rational, we write the denominator as $d_R(t) =1 - \sum_{j = 1}^a h_j t^j \in {\NZQ Q}[t]$. Then from $d_R(t)P^R_M(t) \in {\NZQ Q}[t]$ it follows that the Betti sequence $\{\beta^R_i(M)\}$ eventually satisfies a linear recurrent relation :
\[ \beta^R_n(M) = \sum_{j = 1}^a h_j \ \beta^R_{n-j}(M),\ h_j \in {\NZQ Q}, n \gg 0. \]
Precise information on the coefficients $h_j$ allows for efficient estimates of the asymptotic behaviour of the Betti sequence $\{\beta^R_n(M)\}$ \cite{av6}.
Information on the common denominator is also useful to prove results on vanishing of co-homologies (see \cite{sega1}). We refer the reader to \cite{av2}, for a detailed list of applications.
Stretched Cohen-Macaulay local rings were introduced by Sally in \cite{sally}. Later Elias and Valla
introduced almost stretched Cohen-Macaulay local rings in \cite{elias}.
\begin{definition*}\label{intro0}
An Artinian local ring $R$ with maximal ideal $\mathfrak{m}$ is called stretched if $\mathfrak{m}^2$ is a principal ideal and almost stretched if $\mathfrak{m}^2$ is minimally generated by two elements.
Let $R$ be a Cohen-Macaulay local ring of dimension $d$ with maximal ideal $\mathfrak{m}$. Then $R$ is called stretched (almost stretched) if there exists a minimal reduction $\underline{x} = x_1, \ldots, x_d$ of $\mathfrak{m}$
such that $R/(\underline{x})$ is a stretched (almost stretched) Artinian ring. Here by minimal reduction we mean that $\underline{x}$ satisfies $\mathfrak{m}^{r+1} = (x_1, \ldots, x_d)\mathfrak{m}^r$ for some non-negative integer $r$.
\end{definition*}
Sally proved that $P^R_k(t)$ is rational for a stretched Cohen-Macaulay local ring $R$ if the characteristic of its residue field $k$ is not two \cite[Theorem 2]{sally}.
Elias and Valla imitating the proof of Sally, proved that $P^R_k(t)$ is rational if $R$ is an almost stretched Gorenstein ring and the residue field $k$ of $R$ has characteristic zero \cite[Theorem 1.1]{elias1}.
A common step in the proof of both the results is to show that $P^{R/ \soc(R)}_k(t)$ is rational after reducing to the case when $R$ is in addition Artinian and Gorenstein.
Rationality of $P^R_k(t)$ then follows by a theorem of Avramov and Levin which implies that if $R$ is an Artinian, Gorenstein rin
, then proving rationality of $P^R_k(t)$ is tantamount to proving rationality of $P^{R/ \soc(R)}_k(t)$
(see Theorem \ref{prl6.5}).
We raise the following question.
\begin{customqn}{I}\label{intro1}
Let $R$ be an Artinian Gorenstein local ring. Is $R$ good if $R/ \soc(R)$ is so and vice versa?
\end{customqn}
Tate described an iterated process of killing cycles by adjunctions of sets of $\Gamma$-variables in different degrees starting from a local ring $R$ to construct a DG algebra resolution $\epsilon : R\langle X \rangle \twoheadrightarrow k$, $X = \{ X_i : i \geq 1\}$ of its residue field $k$ \cite[\S 2]{tate}. The ring $R$ is called Golod if $R\langle X_i : \deg(X_i) \leq 1 \rangle$ which is same as the Koszul complex of $R$ admits a trivial Massey operation (\S \ref{prl4}). In \cite{av2} Avramov generalised this notion by defining $R$ to be a generalised Golod ring of level $l$ if the DG algebra $R \langle X_i : \deg(X_i) \leq l \rangle$ admits a trivial Massey operation (see \S \ref{prl6} for details).
A local ring $R$ is said to have the Backelin-Roos property if its completion $\hat{R}$ with respect to its maximal ideal is a homomorphic image of a complete intersection under a Golod homomorphism (see \S \ref{prl6}). A classical result of Levin states that if a ring $R$ has the Backelin-Roos property, then $R$ is good. Avramov observed that rings with the Backelin-Roos property are generalised Golod rings of level $2$.
In particular, any complete intersection is a generalised Golod ring of level $2$.
In this article we prove the following two results
\begin{customthm}{II}{\rm(Theorem \ref{sbr2})} \label{intro2}
Let $R$ be an Artinian Gorenstein local ring of embedding dimension $n \geq 2$ such that $R/ \soc(R)$ is a Golod ring. Let $\eta : Q \twoheadrightarrow R$ be a minimal Cohen presentation and $I = \ker(\eta)$. Then the following hold.
\begin{enumerate}
\item
For any $f \in I \setminus \mathfrak{n} I$, the induced map $Q/ (f) \twoheadrightarrow R$ is a Golod homomorphism, i.e. $R$ has the Backelin-Roos property.
\item
Let $d_R(t) =1 - t(P^Q_R(t) - 1) + t^{n+1}(1+t)$. Then for any $R$-module $M$ we have $d_R(t) P^R_M(t) \in {\NZQ Z}[t]$.
\end{enumerate}
\end{customthm}
\begin{customthm}{III}{\rm(Theorem \ref{ggc4})}\label{intro3}
Let $R$ be an Artinian Gorenstein local ring. Let $l \geq 2$. Then $R$ is a generalised Golod ring of level $l$ if and only if $R/ \soc(R)$ is so.
\end{customthm}
Avramov proved that generalised Golod rings are good (see Theorem \ref{prl9.5}). It is not known if the converse holds true (see \cite[Problem 10.3.5]{av}). So our Theorems \ref{intro2}, \ref{intro3} provide a satisfactory answer to Question \ref{intro1}.
We find several interesting consequences of our theorems. In a recent article \cite[Corollary 5.6]{amenda}, it is proved that stretched Cohen-Macaulay local rings are good.
In Theorem \ref{ds4}, we show that quotients of stretched or almost stretched Artinian Gorenstein rings by non zero powers of maximal ideal are Golod rings. So from Theorem \ref{intro2}, it follows that such rings are good without any assumption on characteristic of residue fields. More generally we prove that stretched Cohen-Macaulay and almost stretched Gorenstein rings are good (see Theorem \ref{sbr3}, Corollary \ref{sbr4}, \ref{sbr4.6}).
Using a result of \c Sega, we prove that the Auslander-Reiten conjecture holds true for such rings. We state our result in the Artinian Gorenstein case below where $\mu(-)$ denotes the minimal number of generators.
\begin{customthm}{IV}{\rm(Theorems \ref{sbr3}, Corollary \ref{ar2})}\label{intro3.5}
Let $R$ be an Artinian Gorenstein local ring with maximal ideal $\mathfrak{m}$ such that $\mu(\mathfrak{m}) = n$ and $\mu(\mathfrak{m}^2) \leq 2$. Let $M$ be a finitely generated $R$-module. Then the following hold.
\begin{enumerate}
\item
If $n = 1$, then $P^R_k(t) = \frac{1}{1 - t}$ and $(1- t)P^R_M(t) \in {\NZQ Z}[t]$,
\item
If $n \geq 2$, then $P^R_k(t) = \frac{1}{1 -nt + t^2}$ and $(1+t)^n(1 - nt + t^2)P^R_M(t) \in {\NZQ Z}[t]$,
\item
If $\Ext^i(M, M \oplus R) = 0$ for all $i \geq 1$, then $M$ is a free $R$-module.
\end{enumerate}
\end{customthm}
In \cite[Theorem 5.1]{rossi}, Rossi and \c Sega proved that compressed Artinian Gorenstein local rings with socle degrees not equal to three have the Backelin-Roos property. We recover this result in a stronger form in Corollary \ref{sbr5} (see the discussion preceding Corollary \ref{sbr5}). To prove this result, we identify using Theorem \ref{intro2} above a certain quotient $C$ of the Koszul algebra $K^R$ of an Artinian Gorenstein ring $R$ such that $R$ has the Backelin-Roos property whenever $C$ is a Golod DG algebra (\S \ref{prl4}). We show that for such rings this quotient is a Golod algebra. By a similar approach, we prove that a Gorenstein local ring of codimension at most three has the Backelin-Roos property (see Corollary \ref{sbr6}). Our version is a stronger form of a result of Avramov, Kustin and Miller \cite[Theorem 6.4]{av4} when $R$ is Gorenstein.
Motivated by the description of cohomology algebras of connected sums of compact connected oriented $n$-manifolds, authors in \cite{an3} introduced the connected sum $R \# S$ of Gorenstein local rings $R$, $S$ with a common residue field $k$ (Definition \ref{prl10}). They obtained a formula for the Poincar\'e series of $k$ over $R\#S$ in terms of the Poincar\'e series of $k$ over $R$ and $S$ \cite[Corollary 7.4]{an3}. In particular if both $P^R_k(t)$, $P^S_k(t)$ are rational functions, so is $P^{R \# S}_k(t)$. This naturally raises the question of whether connected sums of good Gorenstein local rings are good. We prove the following.
\begin{customthm}{V}{\rm(Theorem \ref{ggc5})}\label{intro4}
Let $R$, $S$ be Artinian Gorenstein local rings with a common residue field $k$, $l \geq 2$ and $T = R \# S$ denote the connected sum of $R$ and $S$. Then $T$ is a generalised Golod ring of level $l$ if and only if both $R$, $S$ are so.
\end{customthm}
To prove the above theorem we consider the corresponding problem for fibre products of local rings. In \cite[Theorem 4.1]{les1}, Lescot proved that a fibre product (see Definition \ref{prl10}) of local rings is Golod if and only if the constituent rings are so. We prove a counterpart of Lescot's theorem more generally for generalised Golod rings (see Theorem \ref{ggfl4}). The above theorem then follows from Theorem \ref{intro3}.
Rationality of Poincar\'e series of residue fields of Gorenstein local rings of low multiplicities have been an active topic of research for a long time. The Poincar\'e series $P^R_k(t)$ is known to be rational when $R$ is a Gorenstein local ring of multiplicity at most $n$ with residue field $k$ in the cases : $n = 5$, $\Ch(k) \neq 2$ (\cite{sally}); $n = 7$, $\Ch(k) = 0$( \cite[Corollary 2.1]{elias1}); $n = 10$, $k = \bar{k}$, $\Ch(k) = 0$ (\cite{casnati}). Our motivation on this topic comes from a different perspective. In \cite{eisenbud} Eisenbud conjectured that modules with bounded Betti numbers are eventually periodic of period two.
Several authors proved the conjecture for modules over various good local rings using concrete information on the common denominator (see \cite[Theorem 1.6]{av6}, \cite[Corollary 5.2]{kustin}).
However Gasharov and Peeva in \cite{gash} found an example of a graded Artinian Gorenstein ring of multiplicity twelve over which the conjecture fails. Using a completely different technique they were able to prove the conjecture for
Gorenstein local rings of multiplicity at most eleven. This observation motivates us to investigate whether such rings are good. Statement (1) of the theorem below is a generalisation of \cite[Corollary 2.2]{rossi1}.
\begin{customthm}{VI}{\rm(Theorem \ref{ggc6})}\label{intro5}
Let $R$ be a Gorenstein local ring. Then $R$ is good, i.e. the Betti sequence of modules over $R$ eventually satisfies a linear recurrent relation in the following cases.
\begin{enumerate}
\item
The square $\mathfrak{m}^2$ of maximal ideal $\mathfrak{m}$ of $R$ is minimally generated by at most four elements and $\mathfrak{m}^4 = 0$.
\item
The multiplicity of $R$ is at most $11$.
\end{enumerate}
\end{customthm}
To prove the above result, it is enough to assume further that $R$ is Artinian. We decompose $R$ as a connected sum of Gorenstein Artinian rings either of embedding dimension at most four or with the cube of maximal ideal being zero unless $R$ is already one of these types. The result then follows from Theorem \ref{intro4} as rings of such types are good.
\noindent
{\bf Structure of the paper} :
In \S2 we review some background material that we will require. Readers familiar with \cite{av}, \cite{gul} may skip this section and start reading from \S3. In \S3 we prove two important results viz. Theorems \ref{acg5}, \ref{ac3} both of which give criteria for connected sum decompositions. We use Theorem \ref{acg5} to prove Theorem \ref{ds4} whereas Theorem \ref{ac3} is used to decompose Gorenstein rings concerned in the Theorem \ref{intro5} above. The next two sections \S4, \S6 consist of the main technical part of the article. The underlying idea is very simple. Suppose $S = R/ I$ is a quotient of a local ring $R$. A priori there is no known method to compare the acyclic closures of $R$ and $S$. The key new idea of the present article is to show that when $R$ is an artinian Gorenstein ring of embedding dimension at least 2 and $I = \soc(R)$, then the acyclic closure of $S$ is a semi-free extension of the tensor product of $S$ and the acyclic closure of $R$ (see Proposition \ref{ggc3}). Proposition \ref{ggfl3} concerns construction of the acyclic closure of fibre products and is similar in spirit. The reader may read both the propositions in one go. The proofs of our main results in these two sections extensively use a characterisation theorem for Golod algebras (see Theorem \ref{prl7}) and derivations on acyclic closures (see Lemma \ref{ggc2}) whose construction dates back to the work of Gulliksen. The final section \S 6 contains applications of our Main results, viz. Theorems \ref{intro3.5} and \ref{intro5}.
All rings in this article are Noetherian local rings with $1\not=0$. All modules are nonzero and finitely generated.
Throughout this article, the expression ``local ring $(R,\mathfrak{m}, k)$" refers to a commutative Noetherian local ring $R$ with maximal ideal $m$ and residue field $k = R/\mathfrak{m}$. When information on the residue field is not necessary, we denote a local ring $R$ with maximal ideal $\mathfrak{m}$ simply by $(R, \mathfrak{m})$.
\section{preliminaries}
In this section we set notations and discuss some background. Let $(R, \mathfrak{m})$ be a local ring. The embedding dimension of $R$ is denoted by $\edim(R)$. By a minimal Cohen presentation, we mean a surjection $\eta : Q \twoheadrightarrow \hat{R}$ from a regular local ring $(Q, \mathfrak{n})$ onto the $\mathfrak{m}$-adic completion $\hat{R}$ of $R$ such that $\edim(Q) = \edim(R)$ equivalently $\ker(\eta) \subset \mathfrak{n}^2$. Existence of such a map is guaranteed by Cohen's structure theorem.
For an $R$-module $M$, the length of $M$ is denoted by $l(M)$ and the minimal number of generators of $M$ is denoted by $\mu(M)$. Moreover if $R$ is an Artinian ring, the socle of $R$ is defined as $\soc(R) = (0 : _R \mathfrak{m})$ and Loewy length of $R$ as $\lo(R) = \mx \{ n \ : \ \mathfrak{m}^n \not=0\}$. If $I$ is an ideal of $R$, then the residue class $\frac{(x) + I}{I}$ of an element $x \in R$ in the quotient ring $R/ I$ is denoted by $\bar{x}$. By the phrase ``a natural quotient map $q : R \twoheadrightarrow R/I$" we mean the obvious surjective map defined by $q(x) = \frac{(x) + I}{I}$. Throughout the article unless otherwise stated an over-line is used to denote either a homology class or a residue class modulo an ideal, which should be understood from the context.
We briefly recall some definitions and state results needed in latter sections. For all other unexplained notations and terminology we refer the reader to \cite{av}, \cite{gul}.
\begin{subsection}{DG algebras, DG$\Gamma$ algebras}\label{prl1}
Let $(R, \mathfrak{m}, k)$ be a local ring. A DG algebra $(A, \partial)$ over the ring $R$ consists of a non-negatively graded strictly skw-commutative $R$-algebra $A = \oplus_{i \geq 0} A_i$ such that $ A_0 = R/ I$ for some ideal $I$ of $R$ and an $R$-linear differential map $\partial$ of degree $-1$ satisfying Leibniz rule (see \cite[Chapter 1, \S 1]{gul}). The DG algebra $A$ is called minimal if $\partial(A) \subset \mathfrak{m} A$.
In this article we say a DG algebra $(A, \partial)$ augmented if it is equipped with a surjective algebra homomorphism $\epsilon : A \twoheadrightarrow k$ such that $\epsilon(A_{\geq 1}) = 0$ and $\epsilon \circ \partial = 0$. The map $\epsilon$ is called an augmentation map.
The augmentation map $\epsilon$ induces a map on homology $\tilde{\epsilon} : \Ho(A) \rightarrow k$. We set $\I(A) = \ker {\epsilon}$ called an augmentation ideal, $\I\Ho(A) = \ker \tilde{\epsilon}$ and $\I \Zi(A) = \I A \cap \Zi(A)$ .
If the DG algebra $(A, \partial)$ over $R$ satisfies $\partial(A_1) \subset \mathfrak{m} A_0$, then $A$ is augmented naturally with the surjective map $\epsilon = q \circ pr$ where $pr : A \twoheadrightarrow A_0$ is the projection and $q: A_0 \twoheadrightarrow k$ is the natural quotient map.
A skew commutative non-negatively graded algebra is called a $\Gamma$-algebra if to every element $x$ of even positive degree there is associated a sequence of elements $x^{(k)}$, $k \geq 0$ called divided powers of $x$ satisfying usual axioms (see \cite[Chapter, \S 7]{gul}).
A DG$\Gamma$ algebra is a DG algebra which is a $\Gamma$-algebra in its own right and admits divided powers compatible with its differential, i.e. the differential $\partial$ satisfies $\partial(x^{(n)}) = \partial(x)(x^{(n - 1)})$, $\deg(x)$ being even. We treat the local ring $R$ as an augmented DG$\Gamma$ algebra concentrated at degree zero with the natural quotient map $\epsilon : R \rightarrow k$ as an augmentation map.
Let $f : A \rightarrow B$ be a homorphism of DG$\Gamma$ algebras. An $A$-linear derivation of degree $n$ on the DG$\Gamma$ algebra $B$ is a $R$-linear map $\eta : B \rightarrow B[n]$ such that $\eta(A) = 0$ ($A$-linearity); $\eta$ satisfies Leibniz rule, i.e. $\eta(xy) = \eta(x)y + (-1)^{n\deg(x)}x\eta(y)$ for $x, y \in A$, $\eta(x^{(i)}) = \eta(x) x^{(i-1)}$ for $x \in A$, $\deg(x)$ being even and $\eta$ commutes with the differential $\partial$ of $B$ in the graded sense, i.e. $\eta \circ \partial = (-1)^n \partial \circ \eta$.
\end{subsection}
Unless otherwise stated, we shall always assume that DG algebras of the form $(A, \partial)$ are piecewise Noetherian, i.e. $A_0$ is Noetherian and all $A_i$ are finitely generated $A_0$ modules.
\subsection{Acyclic closures, distinguished DG$\Gamma$ algebras, Tate resolutions} \label{prl2}
Assume that $(R, \mathfrak{m}, k)$ is a local ring and $(A, \partial)$ is an augmented DG$\Gamma$ algebra over $R$ with augmentation map $\epsilon_A : A \rightarrow k$. Tate constructed an augmented semi-free extension $ i : A \hookrightarrow A^* = A \langle X \rangle$, $X = \{X_i\}$ by a recurrent process of adjoining sets of variables (called $\Gamma$-variables) to kill cycles such that the augmentation map $\epsilon_{A^*} : A^* \rightarrow k$ satisfies $\epsilon_{A^*} \circ i = \epsilon_A$ and $A^*$ is acyclic, i.e. $\I \Ho(A^*) = 0$ (see \cite[\S 6.3.1]{av} and \cite[Theorem 1.2.3]{gul}). We call $A^*$ the acyclic closure of $A$.
The set $X$ of variables adjoined is countable since by our assumption $A$ is piecewise Noetherian.
We order the set $X = \{X_n : n \in {\NZQ N} \}$ of variables to assume that $\deg X_i \leq \deg X_j$ for $i \leq j$.
Henceforth we assume that the set of adjoined variables in acyclic closures of DG$\Gamma$ algebras are ordered in this manner .
The acyclic closure satisfies $\Bo(A^*) \subset \I(A) A^*$ \cite[Theorem 1.6.2]{gul}. Following \cite[\S 1.3]{av2}, the augmented DG$\Gamma$ algebras $A\langle X_i : \deg(X_i) \leq d \rangle, d \geq 1$ obtained in intermediate steps of adjunction of variables to construct the acyclic closure $A^* = A\langle X \rangle$ are called distinguished DG$\Gamma$ algebras over $A$. Thus a distinguished DG$\Gamma$ algebra $B$ over $A$ is an augmented semi-free extension $B = A\langle X_1, \ldots, X_n \rangle$, $\deg X_i \leq \deg X_j$ for $i \leq j$, $\deg(X_n) = d$ such that $\partial(X_j)$, $\deg(X_j) = i + 1$ minimally generate $\I \Ho_i(A \langle X_j : \deg(X_j) \leq i \rangle)$ for $0 \leq i \leq d-1$.
Since any cycle of $B$ is in the boundary of the acyclic closure $A^*$ extending $B$ we have $\Zi (B) \subset (\I A )B$.
If $R\langle X \rangle$ is the acyclic closure of the local ring $R$
, then its augmentation map $\bar{\epsilon} : R\langle X \rangle \rightarrow k$ forms a minimal DG$\Gamma$ algebra resolution of $k$. In the literature it is called the Tate resolution. Since the augmentation ideal $I(R) = \mathfrak{m}$, the Tate resolution and distinguished DG$\Gamma$ algebras over $R$ are minimal. We denote the Koszul algebra of $R$ on a minimal set of generators of $\mathfrak{m}$ by $K^R$. Note that $K^R$ can be identified with the distinguished DG$\Gamma$ algebra $R \langle X_i : \deg(X_i) \leq 1\rangle$.
\subsection{Homotopy Lie algebras}\label{prl3}
Let $(R, \mathfrak{m}, k)$ be a local ring and $(A, \partial)$ be an augmented DG$\Gamma$ algebra with augmentation map $\epsilon : A \twoheadrightarrow k$. We define $\Tor^A(k, k) = \Ho(k \otimes_A A^*)$, $A^*$ being the acyclic closure of $A$ and $k \otimes_A A^* = \frac{k \otimes_R A^*}{\langle 1 \otimes a : a \in \ker \epsilon \rangle}$. We have a Hopf $\Gamma$-algebra structure on $\Tor^A(k, k)$ and the graded $k$-vector space dual $\Tor^A(k, k)^{\vee}$ is the universal enveloping algebra of a uniquely defined graded Lie algebra over $k$ (see \cite[Theorem 1.1, 1.2]{av3}). This Lie algebra is called the homotopy Lie algebra of $A$ and denote by $\pi(A)$.
Set $\I\Tor^A(k, k)= \ker\{\tilde{\epsilon} : \Tor^A(k, k) \twoheadrightarrow k\}$. Let $\I\Tor^A(k, k)^{(2)}$ be the ideal of $\Tor^A(k, k)$ generated by products $ab$ for $a, b \in \I\Tor^A(k, k)$ and divided powers $x^{(2)}$ of elements $x \in \I\Tor^A(k, k)$ of even positive degrees. Then $\pi(A) = \Hom_k(\frac{\I\Tor^A(k, k)}{\I\Tor^A(k, k)^{(2)}}, k)$ (see \cite[\S 3.1]{av1}). Thus $\pi(A)$ is the graded vector space dual of the space of $\Gamma$-indecomposable elements in $\I\Tor^A(k, k)$.
\subsection{Golod algebras}\label{prl4}
If $W = \oplus_{i \geq 0}W_i$ is a graded $k$-vector space, we define the Hilbert series $\Ho_W(t) = \sum_{i \geq 0} \dim_k W_it^i \in {\NZQ Z}[|t|]$. Let $(R, \mathfrak{m}, k)$ be a local ring. Let $A$ be a minimal DG algebra over $R$ such that $\Ho_0(A) = k$ and each $A_i$ is a free $R$-module of finite rank. Then there is a term-wise inequality of power series $P^R_k(t) \prec \frac{\Ho_{A \otimes k}(t)}{1 - t (\Ho_{\Ho(A)}(t) - 1)}$. Following Levin \cite[Chapter 1, \S 4]{lev3}, we define $A$ to be a Golod algebra when equality holds.
In general an augmented DG algebra $A$ over $R$ with augmentation map $\epsilon : A \twoheadrightarrow k$ is called a Golod algebra if $A$ admits a trivial Massey operation, i.e.
there is a graded $k$-basis $\mathfrak{b}_R = \{h_{\lambda}\}_{\lambda \in \Lambda}$ of $\I \Ho(A)$, a function $\mu : \sqcup_{i = 1}^{\infty} \mathfrak{b}_R^i \rightarrow A$ such that $\mu(h_\lambda) \in \I \Zi(A)$ with $cls(\mu(h_\lambda))= h_\lambda$ and setting $\bar{a} = (-1)^{i+1}a$ for $a \in A_i$ one has
\[ \partial \mu(h_{\lambda_1}, \ldots ,h_{\lambda_p}) = \sum_{j = 1}^{p-1}\overline{\mu(h_{\lambda_1}, \ldots ,h_{\lambda_j})}\mu(h_{\lambda_{j + 1}}, \ldots ,h_{\lambda_p}).\]
\subsection{Golod homomorphisms, Backelin-Roos property and generalised Golod rings}\label{prl6}
Let $f : (R, \mathfrak{m}) \twoheadrightarrow (S, \mathfrak{n})$ be a surjective local homomorphism. Let $R\langle X \rangle$ be the acyclic closure of $R$. Then $f$ is called a Golod homomorphism if $S \otimes_R R\langle X \rangle$ is a Golod algebra. If $f$ is Golod homomorphism, we have $P^S_k(t) = \frac{P^R_k(t)}{1 - t(P^R_S(t) - 1)}$. Let $\eta : Q \twoheadrightarrow \hat{R}$ be a minimal Cohen presentation of $R$. Then $R$ is called a Golod ring if $\eta$ is a Golod homomorphism, i.e. the Koszul complex $K^R$ is a Golod algebra.
The ring $R$ is said to have the Backelin-Roos property if there is a $Q$-sequence $a_1, \ldots, a_m$ in $\ker(\eta)$ such that the induced map $Q/ (a_1, \ldots, a_m) \twoheadrightarrow \hat{R}$ is a Golod homomorphism.
The ring $R$ is called a generalised Golod ring of level $n$ if the distinguished DG$\Gamma$ algebra $R\langle X_i : \deg(X_i) \leq n \rangle$ is a Golod algebra. A generalised Golod ring of level one is a Golod ring.
If $R$ has the Backelin-Roos property, then $R$ is a generalised Golod ring of level $2$ \cite[\S1.8]{av2}.
For more examples we refer the reader to \cite[\S1.8]{av2}.
The following theorem is proved in \cite[Theorem 2]{lev1}.
\begin{theorem}\label{prl6.5}
Let $(R, \mathfrak{m}, k)$ be an Artinian Gorenstein local ring such that $\edim(R) \geq 2$. Then the quotient map $R \twoheadrightarrow \frac{R}{\soc(R)}$ is a Golod homomorphism and $P^{R/ \soc(R)}_k(t) = \frac{P^R_k(t)}{ 1 - t^2 P^R_k(t)}$.
\end{theorem}
We recall the definition of graded Lie algebras and other necessary terminologies from \cite[\S 1]{av3}. The following result follows from a blend of \cite[Theorem 1.3]{lev3}, \cite[Theorem 2.3]{av1}.
\begin{theorem}\label{prl7}
Let $(R, \mathfrak{m}, k)$ be a local ring and $A$ be a minimal DG$\Gamma$ algebra over $R$ such that $\Ho_0(A) = k$ and each $A_i$ is a free $R$-module of finite rank. Let $A^*$ be the acyclic closure of $A$. Consider $A$ augmented naturally.
Then the following are equivalent.
\begin{enumerate}
\item
The DG algebra $A$ is a Golod algebra.
\item
We have $\Zi_{\geq 1}(A) \subset \mathfrak{m} A$ and the inclusion $i : \mathfrak{m} A \hookrightarrow \mathfrak{m} A^*$ induces an injective map $\Ho(\mathfrak{m} A) \hookrightarrow \Ho(\mathfrak{m} A^*) = \Tor^R(\mathfrak{m}, k)$.
\item
The map $\Ho(A \otimes_R k) \rightarrow \Ho(A^* \otimes_R k)=\Tor^R(k, k)$ is injective and $A$ admits a trivial Massey operation.
\item
The map $\Ho(A \otimes_R k) \rightarrow \Tor^R(k, k)$ is injective and the homotopy Lie algebra $\pi(A)$ is a free Lie algebra.
\end{enumerate}
\end{theorem}
The condition that the map $\Ho(A \otimes_R k) \rightarrow \Tor^R(k, k)$ is injective, is freely satisfied when $A$ is a distinguished DG$\Gamma$ algebra.
The following theorem was proved by Avramov in \cite[Theorem A]{av2}.
\begin{theorem}\label{prl9.5}
A local ring $(R, \mathfrak{m}, k)$ is a generalised Golod ring of level $n$ if and only if $\pi^{>n}(R)$ is a free Lie algebra. If $R$ is a generalised Golod ring of level $n$, then there exists a polynomial $d_R(t) \in {\NZQ Z}[t]$ such that for any finitely generated $R$-module $M$, the product $d_R(t) P^R_M(t) \in {\NZQ Z}[t]$. Furthermore, in this case $d_R(t)P^R_k(t)$ divides $\prod_{1 \leq 2i + 1 \leq n}(1 + t^{2i + 1})^{e_{2i +1}}$, where $e_{2i + 1}$ denotes the number of exterior variables of degree $2i + 1$ in the distinguished DG$\Gamma$ algebra $R\langle X_i : \deg(X_i) \leq n \rangle$.
\end{theorem}
\begin{definition}\label{prl10}
Let $(R, \mathfrak{m}_R, k)$, $(S, \mathfrak{m}_S, k)$ be local rings with a common residue field $k$. Let $\pi_R : R \rightarrow k$, $\pi_S : S \rightarrow k$ be natural quotient maps from $R$, $S$ onto $k$ respectively. The fibre product of $R$ and $S$ is defined as the ring $R \times_k S = \{(r, s) \in R \times S : \pi_R(r) = \pi_S(s) \}$. The ring $R \times_k S$ is local with maximal ideal $\mathfrak{m}_R \oplus \mathfrak{m}_S$. The rings $R$, $S$ are modules over $R \times_k S$ by left actions defined by projection maps $(r, s) \mapsto r$, $(r, s) \mapsto s$ respectively.
Now assume that both $R$, $S$ are Artinian Gorenstein local rings with one dimensional socles $\soc(R) = \langle \delta_R \rangle$, $\soc(S) = \langle \delta_S \rangle$. Then the connected sum of $R$ and $S$ is defined as $R \# S = \frac{R \times_k S}{ \langle (\delta_R, - \delta_S)\rangle}$.
We say a Gorenstein local ring $Q$ is decomposable as a connected sum if there are rings $R$, $S$ such that $Q = R \# S$, $l(R) < l(Q)$ and $l(S) < l(Q)$.
\end{definition}
We recall the following
without specific reference.
\noindent
{\bf Facts :}
\begin{enumerate}
\item
The fibre product $T = R \times_k S$ is a local ring and the connected sum $Q = R \# S$ is an Artinian Gorenstein local ring when both $R$, $S$ are so.
\item
Let $\mathfrak{m}_T, \mathfrak{m}_Q$ be the unique maximal ideals of $T$, $Q$ respectively. Then $\mathfrak{m}^i_T = \mathfrak{m}^i_R \oplus \mathfrak{m}^i_S$ for $i \geq 1$. The maps $\mathfrak{m}_R \rightarrow \mathfrak{m}_Q$, $m \mapsto (\bar{m}, 0)$ and $\mathfrak{m}_S \rightarrow \mathfrak{m}_Q$, $n \mapsto (0, \bar{n})$ are injective.
We have $\lo(Q) = \max\{\lo(R), \lo(S)\}$ and $Q/ \mathfrak{m}_Q^n \cong R/ \mathfrak{m}_R^{\min\{n, \lo(R)\}} \times_k S/ \mathfrak{m}^{\min\{n, \lo(S)\}}_S$ for $2 \leq n \leq \lo(Q)$.
\item
$l(R \times_k S) = l(R) + l(S) - 1$ and $l(R \# S) = l(R) + l(S) - 2$.
\end{enumerate}
\begin{theorem}{\rm \cite[Satz 2]{dress}}\label{prl10.6}
Let $(S, \mathfrak{m}_S, k)$, $(T, \mathfrak{m}_T, k)$ be two local rings and $R = S \times_k T$. Then $\frac{1}{P^R_k(t)} = \frac{1}{P^S_k(t)} + \frac{1}{P^T_k(t)} - 1$. If $M$ is an $S$-module, then $\frac{1}{P^R_M(t)} = \frac{P^S_k(t)}{P^S_M(t)}(\frac{1}{P^S_k(t)} + \frac{1}{P^T_k(t)} - 1)$.
\end{theorem}
\section{criteria for connected sum decompositions}
In this section, we develop criteria to decompose an Artinian Gorenstein local ring as a connected sum.
The following result follows from \cite[Proposition 4.1]{an1}
and Fact (2) of Definition \ref{prl10}.
\begin{theorem}\label{acg8}
Let $(Q, \mathfrak{n}, k)$ be a regular local ring and $I \subset \mathfrak{n}^2$ be an ideal such that $R = Q/ I$ is an Artinian Gorenstein local ring. Let $\mathfrak{n}$ be minimally generated by $x_1, \ldots, x_m, y_1, \ldots, y_n$ such that $(x_1, \ldots, x_m)(y_1, \ldots, y_n) \subset I$. Let $\max\{ i : (x_1, \ldots, x_m)^i \not\subset I\} = s$, $\max\{ i : (y_1, \ldots, y_n)^i \not\subset I\} = t$. Then there are ideals $I_1$, $I_2$ in $Q$ containing $(x_1, \ldots, x_m)$, $(y_1, \ldots, y_n)$ respectively such that the following hold.
\begin{enumerate}
\item
The rings $S = Q/ I_1$, $T = Q/ I_2$ are Gorenstein rings. $\edim(S) = n$, $\edim(T) = m$, $\lo(S) = t$, $\lo(T) = s$.
\item
$R = S \#_k T$.
\end{enumerate}
\end{theorem}
\begin{lemma}\label{acg2}
Let $(R, \mathfrak{m}, k)$ be a local ring and $I$ be an ideal of $R$ such that $\mu(xI) \leq 1$ for all $x \in I \setminus \mathfrak{m} I$. Then $I^2 = yI$ for some $y \in I \setminus \mathfrak{m} I$. In particular $\mu(I^2) \leq 1$.
\end{lemma}
\begin{proof}
Let $I$ be minimally generated by $x_1, x_2, \ldots, x_n$. We prove the lemma by induction on $n= \mu(I)$. When $n = 1$, $I^2 = yI = (y^2)$ for any $y \in I \setminus \mathfrak{m} I$.
If we can show that $I^2 = yI + \mathfrak{m} I^2$ for some $y \in I \setminus \mathfrak{m} I$, then by Nakayama's lemma we have $I^2 = yI$. So it is enough to prove that $\frac{I^2}{ \mathfrak{m} I^2} = \frac{yI + \mathfrak{m} I^2}{ \mathfrak{m} I^2}$ for some $y \in I \setminus \mathfrak{m} I$. We have $\mu(\frac{xI + \mathfrak{m} I^2}{ \mathfrak{m} I^2}) \leq 1$ for all $x \in I \setminus \mathfrak{m} I$. Therefore, there is no loss of generality in assuming
that $\mathfrak{m} I^2 = 0$, i.e. $I^2$ is a $k$-vector space. We consider the following two cases. We assume that $I^2 \not=0$ to exclude the trivial case.
\noindent
{\bf Case - 1} : $x_i^2 = 0$ for $1 \leq i \leq n$.
After reordering the generators if necessary we may assume that $x_1x_2 \not= 0$. Then $x_1I = ( x_1x_2 ) = x_2 I$ since $\dim_k(x_1I) = \dim_k(x_2I) = 1$. We have $x_1x_i = \lambda_ix_1x_2$ and $x_2x_i = \lambda'_ix_1x_2$ for $3 \leq i \leq n$ and $\lambda_i, \lambda_i' \in R$. Note that $x_1(x_i - \lambda_ix_2 - \lambda_i'x_1) = x_2(x_i - \lambda_ix_2 - \lambda_i'x_1) = 0$. Let $J$ be the ideal generated by $x_i - \lambda_ix_2 - \lambda_i'x_1, 3 \leq i \leq n$. Then $I = (x_1, x_2) + J$ and $(x_1, x_2)J = 0$. For any $x \in J \setminus \mathfrak{m} J$ we have $x \in I \setminus \mathfrak{m} I$, $xJ = xI $ and $\mu(xJ) = \mu(xI) \leq 1$. Since $\mu(J) = \mu(I) - 2$, by induction hypothesis we find $y' \in J \setminus \mathfrak{m} J$ such that $y'J = J^2$. Let $y = x_1 + y'$. Then $y \in I \setminus \mathfrak{m} I$ and $I^2 = (x_1x_2) + J^2 = (x_1 + y')((x_1, x_2) + J) = yI$.
\noindent
{\bf Case - 2} : Not all $x_i^2 = 0$ for $1 \leq i \leq n$.
Without loss of generality we assume that $x_1^2 \not=0$ and $x_1I = ( x_1^2 )$. Let $x_1x_i = \lambda_i x_1^2$, $\lambda_i \in R$ and $2 \leq i \leq n$. Then $x_1(x_i - \lambda_i x_1) = 0$. Let $J$ be the ideal generated by $x_i - \lambda_i x_1$, $2 \leq i \leq n$. Then $I = (x_1) + J$, $x_1J = 0$. For any $x \in J \setminus \mathfrak{m} J$ we have $x \in I \setminus \mathfrak{m} I$, $xJ = xI $ and $\mu(xJ) = \mu(xI) \leq 1$. Since $\mu(J) = \mu(I) - 1$, by induction hypothesis we find $y' \in J \setminus \mathfrak{m} J$ such that $y'J = J^2$. Let $y = x_1 + y'$. Then $y \in I \setminus \mathfrak{m} I$ and $I^2 = (x_1^2) + J^2 = (x_1 + y')((x_1) + J) = yI$.
\end{proof}
\begin{lemma}\label{acg3}
Let $(R, \mathfrak{m}, k)$ be a local ring such that $\mu(\mathfrak{m}^2) \leq 2$. Then there exists an $x \in \mathfrak{m} \setminus \mathfrak{m}^2$ such that $\mathfrak{m}^2 = x\mathfrak{m}$. If $\lo(R) \geq 3$, then $x^2 \not\in \mathfrak{m}^3$.
\end{lemma}
\begin{proof}
If $\mathfrak{m}^2 = x \mathfrak{m} + \mathfrak{m}^3, x \in \mathfrak{m} \setminus \mathfrak{m}^2$, then by Nakayama's lemma we have $m^2 = x \mathfrak{m}$. So it is enough to assume that $\mathfrak{m}^3 = 0$, i.e. $\mathfrak{m}^2$ is a $k$-vector space. If $\mathfrak{m}^2 = 0$, then $\mathfrak{m}^2 = x \mathfrak{m} = 0$ for all $x \in \mathfrak{m} \setminus \mathfrak{m}^2$. If $\mu(\mathfrak{m}^2) = 1$, then for any $x \in \mathfrak{m} \setminus \mathfrak{m}^2$ such that $x \mathfrak{m} \not= 0$, we have $\mathfrak{m}^2 = x \mathfrak{m}$. If $\mu(\mathfrak{m}^2) = 2$, by Lemma \ref{acg2} we have $x \in \mathfrak{m} \setminus \mathfrak{m}^2$ such that $\mu(x\mathfrak{m}) > 1$. Then $2 \leq \dim_k x \mathfrak{m} \leq \dim_k \mathfrak{m}^2 = 2$. Therefore, $\mathfrak{m}^2 = x \mathfrak{m}$. This proves the first part of the lemma.
If $x^2 \in \mathfrak{m}^3$, then $\mathfrak{m}^3 = x^2 \mathfrak{m} \subset \mathfrak{m}^4$. So by Nakayama's lemma $\mathfrak{m}^3 = 0$ which is a contradiction. Therefore, $x^2 \not\in \mathfrak{m}^3$ if $\lo(R) \geq 3$.
\end{proof}
\begin{theorem}\label{acg5}
Let $(R, \mathfrak{m}, k)$ be a local Artinian Gorenstein ring. Let $\edim(R) = n$, $\lo(R) \geq 3$ and $\dim_k \mathfrak{m}^2/ \mathfrak{m}^3 = m < n$.
Assume that $\mathfrak{m}$ admits a generator $x_1$ such that $\mathfrak{m}^2 = x_1 \mathfrak{m}$. Then there exists a minimal generating set $\{x_1, x_2, \ldots, x_n\}$ of $\mathfrak{m}$ extending $x_1$ such that the following hold.
\begin{enumerate}
\item
$\mathfrak{m}^2 = (x_1^2, x_1x_2, \ldots, x_1x_m)$,
\item
$(x_1, \ldots, x_m)(x_{m +1}, \ldots, x_n) = 0$,
\item
$(x_{m +1}, \ldots, x_n)^2 = \soc(R)$.
\end{enumerate}
The ring $R$ decomposes as connected sum $R = S \# T$ such that $\edim(S) = m$, $\edim(T) = n - m$, $\lo(S) = \lo(R)$ and $\lo(T) = 2$.
\end{theorem}
\begin{proof}
Since $\lo(R) \geq 3$, $x_1^2 \not\in \mathfrak{m}^3$. So we can choose a minimal generating set $\{x_1, x_2, \ldots, x_n\}$ of $\mathfrak{m}$ such that $\mathfrak{m}^2 = (x_1^2, x_1x_2, \ldots, x_1x_m)$.
We have $x_1x_j = \alpha_{1j}x_1^2 + \alpha_{2j} x_1x_2 + \ldots + \alpha_{mj}x_1x_m$, $\alpha_{ij} \in R$ for $1 \leq i \leq m$ and $m +1 \leq j \leq n$. This gives $x_1(x_j - \alpha_{1j}x_1 - \alpha_{2j}x_2 - \ldots - \alpha_{mj}x_m) = 0$. So replacing $x_j - \alpha_{1j}x_1 - \alpha_{2j}x_2 - \ldots - \alpha_{mj}x_m$ by $x_j$, we assume that $x_1x_j = 0$ for $m + 1 \leq j \leq n$.
If $m = 1$, the property (2) is satisfied. So we assume that $\mu(\mathfrak{m}^2) = m \geq 2$.
Since $\mathfrak{m}^2$ is minimally generated by $\{x_1^2, x_1x_2, \ldots, x_1x_m\}$ and $x_1(x_{m+1}, \ldots, x_n) = 0$, we have $(0 : x_1) \subset \mathfrak{m} (x_1, \ldots, x_m) + (x_{m+1}, \ldots, x_n)$. This implies that residue classes of elements $x_{m + 1}, \ldots, x_n$ form a $k$-basis of $\frac{(0 : _R x_1) + \mathfrak{m}^2}{\mathfrak{m}^2}$. Therefore, $\dim_k \frac{(0 : _R x_1) + \mathfrak{m}^2}{\mathfrak{m}^2} = n - m$.
\noindent
{\bf Claim - 1} : $\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2] = \soc(R)$.
Note that $\mathfrak{m}^2(0 : _R x_1) = x_1\mathfrak{m} (0 : _R x_1) = 0$. So $\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2] \subset \soc(R)$. Therefore, it is enough to prove that $\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2] \not= 0$. We have $\mathfrak{m}^{i +1} = x_1\mathfrak{m}^i$ for $i \geq 1$. So $\mu(\mathfrak{m}^{i +1}) \leq \mu(\mathfrak{m}^i), i \geq 1$. Let $t = \mx\{i : \mu(\mathfrak{m}^i) = m\}$. Since $m \geq 2$, we have $\lo(R) \geq t + 1$. The map $\mathfrak{m}^t/ \mathfrak{m}^{t + 1} \xrightarrow{\cdot x_1} \mathfrak{m}^{t + 1}/ \mathfrak{m}^{t + 2}$ is not injective since $\dim_k(\mathfrak{m}^t/ \mathfrak{m}^{t + 1}) > \dim_k(\mathfrak{m}^{t +1}/ \mathfrak{m}^{t + 2})$. So we find $y \in \mathfrak{m}^t \setminus \mathfrak{m}^{t +1}$ such that $yx_1 \in \mathfrak{m}^{t + 2}$. Note that $\mathfrak{m}^{t+2} = x_1 \mathfrak{m}^{t +1}$. So $yx_1 = x_1m$ for some $m \in \mathfrak{m}^{t +1}$. This implies that $x_1(y - m) = 0$. Therefore, $y - m \in [(0 : _R x_1) \cap \mathfrak{m}^2]$ and $y - m \not\in \mathfrak{m}^{t +1}$. Since $\soc(R) \subset \mathfrak{m}^{t + 1}$, we have $y - m \not\in \soc(R)$. So $[(0 : _R x_1) \cap \mathfrak{m}^2] \not \subset \soc(R)$. Consequently $\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2] \not=0$ and the claim is proved.
\noindent
{\bf Claim - 2} : $\dim_k \frac{[(0 : _R x_1) \cap \mathfrak{m}^2]}{\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2]} = m - 1$.
We know $(0 : _R x_1) = \Hom_R(R/ (x_1), R)$. We have $\mathfrak{m}^2 = x_1\mathfrak{m} \subset (x_1)$. So by the Matlis duality $l(0 : _R x_1) = l(R/ x_1R) = l(R/ \mathfrak{m}^2) - l(\frac{x_1R + \mathfrak{m}^2}{\mathfrak{m}^2}) = 1 + n - 1 = n$. Now we have
\begin{equation*}
l[(0 : _R x_1) \cap \mathfrak{m}^2] = l(0 :_R x_1) + l(\mathfrak{m}^2) - l[(0 : _R x_1) + \mathfrak{m}^2]
= l(0 :_R x_1) - l [\frac{(0 : _R x_1) + \mathfrak{m}^2}{\mathfrak{m}^2}]
= n - (n - m) = m.
\end{equation*}
Therefore, $\dim_k \frac{[(0 : _R x_1) \cap \mathfrak{m}^2]}{\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2]} = l(\frac{[(0 : _R x_1) \cap \mathfrak{m}^2]}{\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2]}) = l[(0 : _R x_1) \cap \mathfrak{m}^2] - 1 = m - 1$.
\noindent
{\bf Claim - 3} : The pairing $\frac{(x_2, \ldots, x_m)}{\mathfrak{m}(x_2, \ldots, x_m)} \times \frac{[(0 : _R x_1) \cap \mathfrak{m}^2]}{\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2]} \rightarrow \soc(R)$ given by $(\bar{x}, \bar{y}) \rightarrow xy$ is well defined and non-degenerate.
We have $\mathfrak{m} (x_2, \ldots, x_m)(0 : _R x_1) = 0$ since $\mathfrak{m}^2 = x_1\mathfrak{m}$. So $(x_2, \ldots, x_m)[(0 : _R x_1) \cap \mathfrak{m}^2] \subset \soc(R)$. Therefore, the above pairing exists. Note that $ \mathfrak{m}^2( x_{m + 1}, \ldots, x_n) = 0$. So if $y \in [(0 : _R x_1) \cap \mathfrak{m}^2]$ and $y(x_2, \ldots, x_m) = 0$, we have $y \in \soc(R) = \mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2]$. We also have $\dim_k \frac{(x_2, \ldots, x_m)}{\mathfrak{m}(x_2, \ldots, x_m)} = \dim_k \frac{[(0 : _R x_1) \cap \mathfrak{m}^2]}{\mathfrak{m}[(0 : _R x_1) \cap \mathfrak{m}^2]} = m - 1$. Therefore, the pairing is non-degenerate and the claim follows.
Note that $\mathfrak{m}(x_2, \ldots, x_m)(x_{m +1}, \ldots, x_n) = 0$. So $(x_2, \ldots, x_m)(x_{m +1}, \ldots, x_n) \subset \soc(R)$. For $m+1 \leq j \leq n$, multiplication by $x_j$ defines a map $\frac{(x_2, \ldots, x_m)}{\mathfrak{m}(x_2, \ldots, x_m)} \rightarrow \soc(R)$. Since the pairing in Claim - 3 is non-degenerate, we have $\hat{x}_j \in [(0 : _R x_1) \cap \mathfrak{m}^2]$ such that $x_jx_i = \hat{x}_jx_i$ for $2 \leq i \leq m$ and $m + 1\leq j \leq n$. We also have $x_1x_j = x_1\hat{x}_j = 0$.
So $(x_j - \hat{x}_j)(x_1, \ldots, x_m) = 0$, $m + 1\leq j \leq n$. Therefore, replacing $(x_j - \hat{x}_j)$ by $x_j$ we have $(x_1, \ldots, x_m)(x_{m +1}, \ldots, x_n) = 0$ and the property (2) is satisfied.
Let $\soc(R) = \langle\delta\rangle$ and $K = (x_{m +1}, \ldots, x_n)$. We have $\mathfrak{m}^2K = x_1K\mathfrak{m} = 0$. So $K^2 \subset \mathfrak{m} K \subset \soc(R)$. Note that $K^2 \not= 0$ for otherwise each of $x_j, m + 1 \leq j \leq n$ is in $\soc(R) \subset \mathfrak{m}^2$, a contradiction. Therefore, $K^2 = \mathfrak{m} K = \soc(R)$ and the property (3) is satisfied.
The last statement follows from Theorem \ref{acg8}.
\end{proof}
A careful study of \cite{an2} yields the following result.
\begin{theorem}\label{ac3}
Let $(R, \mathfrak{m}, k)$ be a Gorenstein local ring with $\lo(R) \geq 3$. Then the following are equivalent.
\begin{enumerate}
\item
There are Gorenstein local rings $(S, \mathfrak{p})$, $(T, \mathfrak{q})$ with $\lo(S) = \lo(R)$, $\lo(T) = 2$ such that $R = S \# T$.
\item
$(0 : \mathfrak{m}^2) \not \subset \mathfrak{m}^2$.
\end{enumerate}
\end{theorem}
\begin{proof}
First we assume that (1) holds. Since $\lo(S), \lo(T) \geq 2$, we have $\mathfrak{m}/ \mathfrak{m}^2 = \mathfrak{p}/ \mathfrak{p}^2 \oplus \mathfrak{q} / \mathfrak{q}^2$. We choose $y \in \mathfrak{q} \setminus \mathfrak{q}^2$. Since $\mathfrak{q}^3 = 0$, we have $(0, y) \mathfrak{m}^2 = 0$. But $(0, y) \in \mathfrak{m} \setminus \mathfrak{m}^2$. So (2) follows.
Conversely we assume that (2) holds. We choose $y_1, \ldots, y_n \in (0 : \mathfrak{m}^2)$ such that there images in $\frac{(0 : \mathfrak{m}^2)}{\mathfrak{m}^2 \cap (0 : \mathfrak{m}^2)}$ form a basis of $k$-vector space. We have $(y_1, \ldots, y_n) \cap \mathfrak{m}^2 \subset \soc(R)$. Let $I = (y_1, \ldots, y_n) \subset (0 : \mathfrak{m}^2)$. If $z \in I \cap (0 : I)$, then $z$ kills both $I$ and $\mathfrak{m}^2$. So $z$ kills $I + [\mathfrak{m}^2 \cap (0 : \mathfrak{m}^2)] = (0 : \mathfrak{m}^2)$. This implies that $z \in (0 : (0 : \mathfrak{m}^2)) = \mathfrak{m}^2$ since $R$ is a Gorenstein ring. So $z \in \mathfrak{m}^2 \cap I \subset \soc(R)$. Thus we have $I \cap (0 : I) \subset \soc(R)$. Note that $I \cap (0 : I) \neq 0$ because otherwise $l(I + (0 : I)) = l(I) + l(0 : I) = l(R)$, a contradiction since $I + (0 : I)$ is a proper ideal. So $I \cap (0 : I) = \soc(R)$.
Now $l(I + (0 : I)) = l(I) + l(0 : I) - l[I \cap (0 : I)] = l(R) - 1 = l(\mathfrak{m})$. Also $I + (0 : I) \subset \mathfrak{m}$. Therefore, $I + (0 : I) = \mathfrak{m}$ and $\frac{\mathfrak{m}}{ \soc(R)} = \frac{I}{ \soc(R)} \oplus \frac{(0 : \I)}{\soc(R)}$.
Note that $\mathfrak{m}^2 I = 0$. So $\mathfrak{m} I \subset \soc(R)$. If $\mathfrak{m} I = 0$, then $I \subset \soc(R) \subset \mathfrak{m}^2$. This implies that $(0 : \mathfrak{m}^2) = I + \mathfrak{m}^2 \cap (0 : \mathfrak{m}^2) \subset \mathfrak{m}^2$, a contradiction. So $\mathfrak{m} I = \soc(R)$. If $\mathfrak{m} (0: I) = 0$, then $\mathfrak{m}^2 = \mathfrak{m}[I + (0 : I)] = \mathfrak{m} I = \soc(R)$ which is again a contradiction since $\lo(R) \geq 3$. So $\soc(R) \subset \mathfrak{m} (0 : I)$. Therefore, we have $\frac{\mathfrak{m}}{ \mathfrak{m}^2} = \frac{I}{ \mathfrak{m} I} \oplus \frac{(0 : \I)}{\mathfrak{m} (0 : \I)}$.
The ideal $I$ is minimally generated by $y_1, \ldots, y_n$.
We choose a minimal generating set $x_1, \ldots, x_m$ of $(0 : I)$ such that $x_1, \ldots, x_m, y_1, \ldots, y_n$ is a minimal generating set of $\mathfrak{m}$.
Note that $(x_1, \ldots, x_m)(y_1, \ldots, y_n) = 0$. We have $I^3 = 0$. If $I^2 = 0$, then $ \mathfrak{m} I = 0$, a contradiction. So $\max\{ i : (y_1, \ldots, y_n)^i \neq 0\} = 2$. We have $\mathfrak{m} I = \soc(R) \subset \mathfrak{m} (0 : I)$. So $\mathfrak{m}^2 = \mathfrak{m} (0 : I) = (0 : I)^2$. By an easy inductive argument $\mathfrak{m}^n = (0 : I)^n$ for all $n \geq 2$. Therefore, $\max\{ i : (x_1, \ldots, x_m)^i \neq 0\} = \lo(R)$. Now the result follows from Theorem \ref{acg8}.
\end{proof}
\begin{lemma}\label{ds3}
Let $(R, \mathfrak{m})$ be an Artinian Gorenstein local ring such that $\edim(R) \geq 2$ and $\lo(R) = s$. Then the quotient ring $R/ \mathfrak{m}^i$ can never be a Gorenstein ring for $2 \leq i \leq s$.
\end{lemma}
\begin{proof}
If possible assume that $R/ \mathfrak{m}^i$ is a Gorenstein ring for some $i$ satisfying $2 \leq i \leq s$. Then the injective hull of $k$ over $R/ \mathfrak{m}^i$, viz. $E_{R/ \mathfrak{m}^i}(k) \cong R/ \mathfrak{m}^i$. We know that $E_{R/ \mathfrak{m}^i}(k) \cong \Hom_R(R/ \mathfrak{m}^i, R) = (0 :_R \mathfrak{m}^i)$. So $(0 :_R \mathfrak{m}^i) = (x)$, a principal ideal for some $x \in R$. Now $\mathfrak{m}^{s - i +1} \subset (0 :_R \mathfrak{m}^i)$ and $\mathfrak{m}^{s - i +1} \not \subset \mathfrak{m} (0 :_R \mathfrak{m}^i)$ for otherwise $\mathfrak{m}^{s - i +1} \subset \mathfrak{m} (0 :_R \mathfrak{m}^i) \subset (0 :_R \mathfrak{m}^{i-1})$ which implies that $\mathfrak{m}^s = \mathfrak{m}^{s-i+1} \mathfrak{m}^{i-1} = 0$, a contradiction. Since $(0 :_R \mathfrak{m}^i)$ is principal we have $(0 :_R \mathfrak{m}^i) = \mathfrak{m}^{s-i+1} = (x)$. This implies that $\mathfrak{m}^j$ is a principal ideal for $s-i+1 \leq j \leq s$. So $l(0 :_R \mathfrak{m}^i) = l( \mathfrak{m}^{s-i+1}) = i$. By Matlis duality, $l(R/ \mathfrak{m}^i) = l \Hom(R/ \mathfrak{m}^i, R) =(0 :_R \mathfrak{m}^i) = i$. We have $\sum_{j = 0}^{i-1}[l(\mathfrak{m}^j/ \mathfrak{m}^{j+1}) - 1] = l(R/ \mathfrak{m}^i) - i = 0$ and each summand is non-negative. This shows that $\edim(R) = l(\mathfrak{m}/\mathfrak{m}^2) = 1$, a contradiction.
\end{proof}
\begin{theorem}\label{ds4}
Let $(R, \mathfrak{m}, k)$ be an Artinian Gorenstein local ring such that $\mu(\mathfrak{m}) = n$, $\lo(R) \geq 2$ and $\mu(\mathfrak{m}^2) \leq 2$. Let $i$ be an integer satisfying $2 \leq i \leq \lo(R)$. Then $R/ \mathfrak{m}^i$ is a Golod ring and $P^{R/ \mathfrak{m}^i}_k(t) = \frac{1}{1 - nt}$.
\end{theorem}
\begin{proof}
Fix $i$ satisfying $2\leq i \leq \lo(R)$ and set $\bar{R} = \frac{R}{\mathfrak{m}^i}$. The result is clear when $n=1$ or $\lo(R) = 2$. We assume that $n = 2$ and $\lo(R) > 2$. A result of Scheja \cite{scheja} states that a co-depth 2 local ring is either a Gorenstein or a Golod ring. The ring $\bar{R}$ cannot be a Gorenstein ring by Lemma \ref{ds3}. So $\bar{R}$ is a Golod ring. Since $R$ is a complete intersection, the defining ideal of $\bar{R}$ is minimally generated by $3$ elements. So, $\kappa^{\bar{R}}(t) = \sum_{i \geq 0} \dim_k \Ho_i(K^{\bar{R}})t^i = 1 + 3t + 2t^2$ and $(1 - t(\kappa^{\bar{R}}(t) - 1)) = 1 - t(1 + 3t + 2t^2 - 1) = (1 + t)^2(1 - 2t)$. We have $P^{\bar{R}}_k(t) = \frac{(1+t)^2}{1 - t(\kappa^{\bar{R}}(t) - 1)} = \frac{1}{1 - 2t}$.
Now we assume that $n > 2$ and $\lo(R) > 2$. By Lemma \ref{acg3}, Theorem \ref{acg5}, it follows that $R = S \# T$ where $(S, \mathfrak{p})$, $(T, \mathfrak{q})$ are Gorenstein local rings, $\edim(S) = \mu(\mathfrak{m}^2) \leq 2$, $\edim(T) + \edim(S)$ = n, $\lo(S) = \lo(R)$ and $\lo(T) = 2$. So $\bar{R} = \bar{S} \times_k \bar{T}$ where $\bar{S} = S/ \mathfrak{p}^i$ and $\bar{T} = T/ \mathfrak{q}^2$ (see Facts, Definition \ref{prl10}). Both $\bar{S}$, $\bar{T}$ are Golod rings. We have $P^{\bar{S}}_k(t) = \frac{1}{1- \edim(\bar{S})t}$, $P^{\bar{T}}_k(t) = \frac{1}{1-\edim(\bar{T})t}$. The result follows from Theorem \ref{prl10.6} and the fact that fibre products of Golod rings are Golod \cite[Theorem 4.1]{les1}.
\end{proof}
\section{Main results}
Our goal in this section is to prove our main Theorems \ref{intro2}, \ref{intro3} as stated in the introduction.
Suppose $(A, \partial)$ is an augmented DG$\Gamma$ algebra and $A^* = A \langle W\rangle$, $W = \{W_i\}_{i \geq 1}$ is the acyclic closure of $A$. We recall the notion of (normal) $\Gamma$-monomials from \cite[Remark 6.2.1]{av}.
Pick two $\Gamma$-monomials $M$, $M'$ in $A^*$. Then there exists an integer $n$ large enough such that $M = W_1^{(c_1)} \ldots W_n^{(c_n)}$ and $M' = W_1^{(c'_1)} \ldots W_n^{(c'_n)}$ where $c_j, c'_j \geq 0$. By our assumption $\deg(W_1) \leq \ldots \leq \deg(W_n)$. We define a total order on the set of $\Gamma$-monomials as $M \prec M'$ if the last nonzero entry of $\big(c_1' - c_1, \ldots, c_n' - c_n, \deg(M') - \deg(M)\big)$ is positive. If $\beta$ is an element of $A^*$, then we can express $\beta$ as $\beta = \sum f_M M$, $f_M \in A$ where the sum is taken over finitely many $\Gamma$-monomials $M$.
We define the initial $\Gamma$-monomial of $\beta$ as $\init(\beta) = \max \{M : f_M \neq 0 \}$.
The coefficient of $\init(\beta)$ in $\beta$ is called the leading coefficient of $\beta$.
The following lemma is implicit in \cite[Theorem 1.6.2]{gul}.
\begin{lemma}\label{ggc2}
Let $(A, \partial)$ be an augmented DG$\Gamma$ algebra over a local ring $(R, \mathfrak{m}, k)$ with augmentation map $\epsilon : A \twoheadrightarrow k$ and $A^* = A \langle W\rangle$, $W = \{W_i\}_{i \geq 1}$ be the acyclic closure of $A$. Let $F^n(A^*) = A\langle W_i : \deg(W_i) \leq n \rangle$ denote distinguished DG$\Gamma$ algebras.
Then the following hold.
\begin{enumerate}
\item
There are $A$-linear derivations $\nu_j$, $1 \leq j \leq n$ such that $\nu_j(W_j) = 1$ and $\nu_j(W_i) = 0$ for $i < j$.
\item
$\nu_j(F^k(A^*)) \subset F^k(A^*)$ for $j \leq k$.
\item
For a $\Gamma$-monomial $M = W_1^{(c_1)} \ldots W_n^{(c_n)}$, define $\nu_M = \nu_1^{c_1} \circ \ldots \circ \nu_n^{c_n}$. Then $\nu_M(M) = \pm 1$ and $\nu_M(M') = 0$ for $M' \prec M$.
\end{enumerate}
\end{lemma}
If $(R, \mathfrak{m})$ is a Gorenstein local ring of embedding dimension $n \geq 2$, the Koszul homology algebra $\Ho(K^R)$ is a Poincar\'e algebra \cite{av5}. The multiplication map $\Ho_1(K^R) \times \Ho_{n-1}(K^R) \rightarrow \Ho_n(K^R)$ is a non-degenerate pairing. So for any $z \in \Ho_1(K^R)$, there is a $z' \in \Ho_{n-1}(K^R)$ such that $z z'$ is a generator in $\Ho_n(K^R)$. The following lemma lays the groundwork for main results of this article.
\begin{lemma}\label{sbr1}
Let $(R, \mathfrak{m})$ be an Artinian Gorenstein local ring of embedding dimension $n \geq 2$.
Let $R\langle X \rangle$ be the acyclic closure of $R$ and $ \mathfrak{X} = R \langle X_j : 1 \leq j \leq n + i \rangle, i \geq 1$ be a distinguished DG$\Gamma$ algebra over $R$.
Let $\soc(R) = (t)$, $\bar{R} = \frac{R}{tR}$ and $\bar{ \mathfrak{X}} = \bar{R} \otimes_R \mathfrak{X}$.
Let $z \in \Zi_1(K^R)$, $z' \in \Zi_{n-1}(K^R)$ be cycles such that $\partial{X_{n+1}} = z$ and $zz' = tX_1 \ldots X_n$.
Define $\mathfrak{Y} = \{ x \in \mathfrak{X} : z' \partial(x) \in \soc(R) X_1\ldots X_n\mathfrak{X}\}$ and $\bar{\mathfrak{Y}} = \bar{R} \otimes_R \mathfrak{Y}$.
Then the following hold.
\begin{enumerate}
\item
$\soc(R)K^R_i \subset (0 : \mathfrak{m}^2) \Bo_{i+1} (K^R) \subset \mathfrak{m} \Bo_{i+1} (K^R)$ for $ i = 1, \ldots, n - 1$.
\item
$\soc(R)X_1\ldots X_n \mathfrak{X} = z' \Bo (\mathfrak{Y})$.
\item
$\soc(R) \mathfrak{X} = (0 : \mathfrak{m}^2)\Bo (\mathfrak{X}) + z' \Bo (\mathfrak{Y}) \subset \mathfrak{m} \Bo (\mathfrak{X})$.
\item
$\Zi(\bar{\mathfrak{X}}) = (0 : \mathfrak{m}^2) \bar{\mathfrak{X}} + z' \bar{\mathfrak{Y}} + \bar{R} \otimes_R \Zi(\mathfrak{X})$
\item
The quotient map $\mathfrak{X} \twoheadrightarrow \bar{\mathfrak{X}} $ induces injective maps $\Ho(\mathfrak{X}) \rightarrow \Ho(\bar{\mathfrak{X}})$ and $\Ho(\mathfrak{m} \mathfrak{X}) \rightarrow \Ho(\mathfrak{m} \bar{\mathfrak{X}})$ on homology.
\item For any two cycles $v \in \I \Zi(\mathfrak{X})$, $w \in \Zi(z' \mathfrak{X})$, there is an element $v' \in z' \mathfrak{X}$ such that $vw = \partial(v')$.
\item For any cycle $\bar{v} \in \I \Zi(\bar{\mathfrak{X}}), \bar{w} \in \Zi(z' \bar{\mathfrak{X}})$, there is an element $\bar{v'} \in z' \bar{\mathfrak{X}} + (0 : \mathfrak{m}^2)\bar{\mathfrak{X}}$ such that $\bar{v}\bar{w} = \partial(\bar{v'})$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\partial(X_i) = x_i$, $1 \leq i \leq n$. Then the maximal ideal $\mathfrak{m}$ is minimally generated by $\{x_1, \ldots, x_n\}$.
The statement (1) is proved in \cite[Lemma 1.2]{lev1}.
It is proved in the line no 10, page no 80 in the proof of \cite[Theorem 2]{lev1} that $tX_1 \ldots X_n R\langle X \rangle = z' \partial(X_{n+1}) R\langle X \rangle \subset z' \Bo(R\langle X \rangle)$. By a very similar argument, we have $tX_1 \ldots X_n \mathfrak{X} = z' \partial(X_{n+1}) \mathfrak{X} \subset z' \Bo(\mathfrak{X})$. By definition of $\mathfrak{Y}$, it follows that $tX_1 \ldots X_n \mathfrak{X} = z' \Bo(\mathfrak{Y})$. So (2) is proved.
It is proved in the equation 2.3, page no 79 in the proof of \cite[Theorem 2]{lev1} that $t R\langle X \rangle \subset (0 : \mathfrak{m}^2) \Bo(R\langle X \rangle) + tX_1 \ldots X_n R \langle X \rangle$. The same reasoning applies here and we have $t \mathfrak{X} \subset (0 : \mathfrak{m}^2) \Bo(\mathfrak{X}) + tX_1 \ldots X_n \mathfrak{X}$. So from (2), we have $t \mathfrak{X} \subset (0 : \mathfrak{m}^2) \Bo(\mathfrak{X}) + z' \Bo(\mathfrak{Y})$. The reverse inclusion is clear and (3) follows.
Both (4) and (5) follow from (3). The argument is implicit in \cite[Theorem 2]{lev1}. We skip the details.
We prove (6). We may assume that $\deg(v) \geq 1$ as the result is obvious when $v \in \mathfrak{m}$. Let $v = \sum_{i = 1}^nx_iv_i$, $v_i \in \mathfrak{X}$. Then we have
$$v w = \sum_{i = 1}^n x_iv_iw = \sum_{i = 1}^n[ \partial(X_iv_i)w + X_i \partial(v_i) w ]
=\partial (v'_1)w + \sum_{i = 1}^n X_i \partial(v_i) w \ \text{for} \ v'_1 = \sum_{i = 1}^n X_i v_i.
$$
Note that $ X_i z' \in X_1 \ldots X_n \mathfrak{X}, 1 \leq i \leq n$ and $w \in \Zi(z' \mathfrak{X})$. So $\sum_{i = 1}^n X_i \partial(v_i) w = X_1\ldots X_n u$ for $u \in R \langle X_{n+1}, \ldots, X_{n+i} \rangle$. Now $vw = \partial(v_1')w + X_1\ldots X_n u$ and $v, w $ are cycles. So the element $X_1\ldots X_n u$ is also a cycle. Expanding $\partial(X_1\ldots X_n u)$ and comparing coefficients of both sides of $\partial(X_1\ldots X_n u) = 0$, we conclude that coefficients of $u$ are in $\soc(R)$. Thus $u \in \soc(R) R \langle X_{n+1}, \ldots, X_{n+i} \rangle$ and $X_1 \ldots X_n u \in \soc(R) X_1\ldots X_n \mathfrak{X} \subset z' \Bo(\mathfrak{X})$ by (2). Let $X_1 \ldots X_n u = \partial (v_2') z', v_2' \in \mathfrak{X}$. Then we have $vw = \partial(v_1')w + \partial(v_2')z' = \partial(v') $ for $v' = v'_1 w + v'_2 z'$. Clearly $v' \in z' \mathfrak{X}$. So (6) follows.
Finally we prove (7). We assume that $\deg(\bar{v}) \geq 1$ to exclude the obvious case. We observe from (4) that $ \Zi(\bar{\mathfrak{X}}) \bar{w} = \bar{R} \otimes_R \Zi(\mathfrak{X})\bar{w}$. So we choose $v_0 \in \Zi_{\geq 1}(\mathfrak{X})$ such that $\bar{v} \bar{w} = \bar{v}_0\bar{w}$. Now $\bar{w} \in \Zi(z' \bar{\mathfrak{X}})$. Again by (4), we may choose the lift
\begin{equation}\label{sbreq1}
w = z'w' = w_1 + z' w_2 + w_3 \ \text{for} \ w_1 \in (0 : \mathfrak{m}^2)\mathfrak{X}, w_2 \in \mathfrak{Y}\ \text{and} \ w_3 \in \Zi(\mathfrak{X}).
\end{equation}
Note that $v_0 \in \Zi_{\geq 1}(\mathfrak{X}) \subset \mathfrak{m} \mathfrak{X}$. So $v_0 w_1 \in \soc(R)\mathfrak{X}$. By (6), $v_0z' = \partial(v_1z')$ for some $v_1 \in \mathfrak{X}$. So $v_0 z' w_2 = \partial( v_1 z')w_2 = \partial( v_1 z' w_2) + (-1)^{n + \deg(v_1)} v_1 z' \partial(w_2)$. Now $z' \partial(w_2) \in \soc(R)X_1\ldots X_n \mathfrak{X}$ since $w_2 \in \mathfrak{Y}$. This implies that $v_0 z' w_2 = \partial( v_1 z' w_2) + w_4$ for $w_4 = (-1)^{n + \deg(v_1)} v_1 z' \partial(w_2) \in \soc(R) \mathfrak{X}$.
We write $v_0 = \sum_{i=1}^n x_ia_i$ for $a_i \in \mathfrak{X}$. So we have
\[v_0 w_3 = \sum_{i=1}^n x_i a_iw_3 = \sum_{i=1}^n[\partial(X_ia_i)w_3 + X_i\partial(a_i)w_3] = \partial(v_2)w_3 + \sum_{i=1}^nX_i\partial(a_i)w_3 \ \text{where} \ v_2 = \sum_{i=1}^nX_ia_i.\]
Now from equation \ref{sbreq1} it follows that $w_3 = z'(w' - w_2) - w_1$. So $X_iw_3 \in X_1 \ldots X_n \mathfrak{X} + (0 : \mathfrak{m}^2)\mathfrak{X}$.
We have $\partial(a_i) \in \mathfrak{m} \mathfrak{X}$.
Therefore, $\sum_{i=1}^nX_i\partial(a_i)w_3 = X_1\ldots X_nu + w_5$ for some $u \in R\langle X_{n+1}, \ldots, X_{n+i} \rangle$ and $w_5 \in \soc(R)\mathfrak{X}$.
Now $v_0 w_3= \partial(v_2)w_3 + X_1\ldots X_nu + w_5$ and $v_0, w_3, w_5$ are cycles. So $u \in \soc(R)R\langle X_{n+1}, \ldots, X_{n+i} \rangle$ and $X_1\ldots X_nu = \partial(v_3)z'$ for $v_3 \in \mathfrak{Y}$ by (2). Therefore, we have $v_0w_3 = \partial(v_2 w_3 + v_3z') + w_5$.
Finally we write
$v_0w = v_0(w_1 + z' w_2 + w_3) = (v_0w_1 + w_4 + w_5) + \partial(v_1z'w_2 + v_2 w_3 + v_3z')$.
We have $v_0w_1 + w_4 + w_5 \in \soc(R)\mathfrak{X}$. Define
$$v' = v_1z'w_2 + v_2 w_3 + v_3z' = v_1z'w_2 + v_2 [z'(w' - w_2) - w_1] + v_3z'
= [v_1z'w_2 + v_2 z'(w' - w_2) + v_3z'] - v_2w_1.$$
Since $w_1 \in (0 : \mathfrak{m}^2) \mathfrak{X}$, we have $v ' \in z' \mathfrak{X} + (0 : \mathfrak{m}^2)\mathfrak{X}$.
Therefore, going modulo socle, we have $\bar{v}\bar{w} = \overline{v_0w} = \partial(\bar{v'})$ and $v' \in z' \bar{\mathfrak{X}} + (0 : \mathfrak{m}^2)\bar{\mathfrak{X}}$. Thus (7) follows.
\end{proof}
\begin{theorem}\label{sbr2}
Let $(R, \mathfrak{m}, k)$ be an Artinian Gorenstein local ring of embedding dimension $n \geq 2$ such that $R/ \soc(R)$ is a Golod ring. Let $K^R = R\langle X_i : 1 \leq X_i \leq n, \partial(X_i) = x_i \rangle$ be the Koszul complex of $R$ on a minimal set of generators $x_1, \ldots, x_n$ of $\mathfrak{m}$.
Let $\eta : (Q, \mathfrak{n}) \twoheadrightarrow (R, \mathfrak{m})$ be a minimal Cohen presentation. Then the following hold.
\begin{enumerate}
\item
Let $z \in \Zi_1(K^R) \setminus B_1(K^R)$ and $K^R\langle T : \partial(T) = z \rangle$ be the extension of $K^R$ by adjoining divided powers variable $T$ of degree two to kill the cycle $z$. Then $K^R\langle T : \partial(T) = z \rangle$ is a Golod algebra.
\item
Let $I = \ker(\eta)$. Then for any $f \in I \setminus \mathfrak{n} I$, the induced map $Q/ (f) \twoheadrightarrow R$ is a Golod homomorphism, i.e. $R$ has the Backelin-Roos property.
\item
Let $d_R(t) =1 - t(P^Q_R(t) - 1) + t^{n+1}(1+t)$. Then for any $R$-module $M$ we have $d_R(t) P^R_M(t) \in {\NZQ Z}[t]$.
\end{enumerate}
\end{theorem}
\begin{proof}
The DG$\Gamma$ algebra $K^R\langle T \rangle$ can be extended to the acyclic closure $\mathfrak{X}$ of $R$. We have a derivation $v : \mathfrak{X} \rightarrow \mathfrak{X}$ of degree $-2$ such that $v(T) = 1$ by Lemma \ref{ggc2}.
Set $\bar{R} = R / \soc(R)$. Then $K^{\bar{R}} = \bar{R} \otimes_R K^R$ is the Koszul complex of $\bar{R}$ and $K^{\bar{R}}$ can be extended to the acyclic closure $\mathfrak{Y}$ of $\bar{R}$. Now
the augmentation $\epsilon_{\mathfrak{Y}} : \mathfrak{Y} \twoheadrightarrow k$ is an algebra homomorphism over $K^R$. The acyclic closure $\mathfrak{X}$ is semi-free over $K^R$. So the augmentation $\epsilon_{\mathfrak{X}} : \mathfrak{X} \twoheadrightarrow k$ lifts to a DG$\Gamma$ algebra homomorphism $\beta : \mathfrak{X} \rightarrow \mathfrak{Y}$ over $K^R$ \cite[Proposition 2.1.9]{av}. Let $\alpha : K^R \rightarrow K^{\bar{R}}$ denote the quotient map. We have the following commutative diagram.
\[\xymatrix{
\mathfrak{m} K^R \ar[d]^{\alpha} \ar[r]^i &\mathfrak{m} \mathfrak {X}\ar[d]^{\beta}\\
\mathfrak{m} K^{\bar{R}} \ar[r]^j &\mathfrak{m} \mathfrak{Y}}\]
To show that $K^R\langle T \rangle$ is a Golod algebra, we need to prove that the inclusion $i : \mathfrak{m} K^R \langle T \rangle \rightarrow \mathfrak{m} \mathfrak{X}$ induces an injective map $i_* : \Ho(\mathfrak{m} K^R \langle T \rangle) \rightarrow \Ho(\mathfrak{m} \mathfrak{X})$ on homology (see Theorem \ref{prl7}). A cycle $y$ in $\mathfrak{m} K^R \langle T \rangle$ can be written as $y = \sum_{k=0}^m a_k T^{(m-k)}$, $a_i \in \mathfrak{m} K^R$. Suppose $i(y)$ is in the boundary of $\mathfrak{m} \mathfrak{X}$. We prove by induction on $m$ that $y$ is in the boundary of $\mathfrak{m} K^R \langle T \rangle$.
First assume that $m = 0$. Then $y \in \mathfrak{m} K^R$. Since $i(y)$ is in the boundary of $\mathfrak{m} \mathfrak{X}$, $j \circ \alpha(y)$ is in the boundary of $\mathfrak{m} \mathfrak{Y}$ by the commutative diagram. Now $\bar{R}$ is a Golod ring. So $j$ induces an injective map $j_* : \Ho(\mathfrak{m} K^{\bar{R}}) \rightarrow \Ho(\mathfrak{m} \mathfrak{Y})$. Therefore, $\alpha(y)$ is in the boundary of $\mathfrak{m} K^{\bar{R}}$. So we can write $y = sy_1 + \partial(y_2)$ for $y_1 \in K^R$, $y_2 \in \mathfrak{m} K^R$ and $\soc(R) = (s)$. If $\deg(y) = \deg(y_1) < n$, then $sy_1 \in (0 : \mathfrak{m}^2)\Bo(K^R)$ by (1) of Lemma \ref{sbr1}. So $y \in \mathfrak{m} \Bo(K^R)$. If $\deg(y) = n$, then $y_2 = 0$ and $y = s y_1 = a s X_1\ldots X_n$, $a \in R$. Since $\Ho(K^R)$ is a Poincar\'e algebra \cite{av5}, there is a $z' \in \Zi_{n-1} (K^R)$ such that $zz' = s X_1\ldots X_n$. So $y = azz'= \partial (aTz') \in \mathfrak{m} \Bo(K^R\langle T \rangle)$. Therefore, the induction step for $m = 0$ follows.
Now we assume that $m > 0$. Note that $a_0 = v^m (i(y))$. Since $i(y) \in \mathfrak{m} \Bo(\mathfrak{X})$ and the derivation $v$ commutes with differentials, we have $a_0 \in \mathfrak{m} K^R \cap \Bo(\mathfrak{m} \mathfrak{X})$.
Suppose $\deg(a_0) = n$. Then $a_0 = a s X_1\ldots X_n$, $a \in R$. So $a_0T^{(m)} = azz'T^{(m)} = \partial(az' T^{(m + 1)}) \in \mathfrak{m} \Bo(K^R\langle T \rangle)$. Now $ i(\sum_{k=0}^{m-1} a_k T^{(m-k)}) = y - a_0T^{(m)}$ is in the boundary of $\mathfrak{m} \mathfrak{X}$ and therefore is in the boundary of $\mathfrak{m} K^R\langle T \rangle$ by the induction hypothesis. So $y$ is in the boundary of $\mathfrak{m} K^R\langle T \rangle$.
Suppose $\deg(a_0) < n$. Then by the argument in the induction step $m = 0$, $a_0 = \partial(y_3)$ for $y_3 \in \mathfrak{m} K^R$. We can write $y = \partial(y_3) T^{(m)} + \sum_{k=0}^{m-1} a_k T^{(m-k)} = \partial(y_3 T^{(m)}) + [(-1)^{\deg(a_0)} y_3 z T^{(m-1)} + \sum_{k=0}^{m-1} a_k T^{(m-k)}]$. The first summand is in $\mathfrak{m} \Bo(K^R\langle T \rangle)$. So the second summand is in the boundary of $\mathfrak{m} \mathfrak{X}$ and therefore is in the boundary of $\mathfrak{m} K^R\langle T \rangle$ by induction hypothesis. So $y$ is in the boundary of $\mathfrak{m} K^R\langle T \rangle$. This completes the induction step and the proof of (1) follows.
It is well known that generators of $I$ in $I \setminus \mathfrak{n} I$ are in one-one correspondence with the generators of $\Ho_1(K^{R})$. Let the maximal ideal $\mathfrak{n}$ of $Q$ be minimally generated by $y_1, \ldots, y_n$. The Koszul complex of $Q$ is $K^Q = \langle X_i : \partial(X_i) = y_i, 1 \leq i \leq n \rangle$. Let $f = \sum_{i=1}^n a_iy_i$ and $z = \sum_{i=1}^n a_iX_i \in K^Q_1$. Let $P = Q/ (f)$. Then $K^P = K^Q \otimes_Q P$, $K^{R} = K^Q \otimes_Q {R}$ denote Koszul complexes of $P$, ${R}$ respectively.
Since $f \in I \setminus \mathfrak{n} I$, we have $\bar{z} \in \Zi_1(K^{R}) \setminus B_1(K^{R})$.
By (1), $U = K^{R}\langle T : \partial(T) = \bar{z} \rangle$ is a Golod algebra. Now $V = K^P\langle T : \partial(T) = \bar{z} \rangle$ is the acyclic closure of $P$ and $U = V \otimes_R R$. Therefore, the map $P \twoheadrightarrow R$ is a Golod homomorphism and (2) follows.
The ring $\bar{R}$ is Golod. In \cite[Proposition 6.2]{rossi}, the Poincar\'e series of $R$ is computed as $P^{R}_k = \frac{(1 + t)^n}{1 - t(P^Q_R(t) - 1) + t^{n+1}(1+t) }$. So the result follows from Theorem \ref{prl9.5}.
\end{proof}
The following result is proved in \cite[Theorem 1]{lev1}.
\begin{theorem}\label{ggc1}
Let $(R, \mathfrak{m})$ be an Artinian Gorenstein local ring of embedding dimension $n$ and $K^R$ be the Koszul complex on a minimal set of generators of the maximal ideal $\mathfrak{m}$. Set $\soc(R) = (s)$, $\bar{R} = R/ sR$ and $K^{\bar{R}} = \bar{R} \otimes_R K^R$, the Koszul complex of $\bar{R}$. Define a DG algebra structure on $K^R \oplus \frac{K^R}{\mathfrak{m} K^R}[-1]$ with multiplication : $(k, \bar{l})(k', \bar{l'}) = (kk', \overline{lk'} + (-1)^{\deg(k)}\overline{kl'})$ and differential : $\partial(k, \bar{l}) = (\partial(k) + sl, 0)$. Then the chain map $K^R \oplus \frac{K^R}{\mathfrak{m} K^R}[-1] \rightarrow K^{\bar{R}}$, $(k, \bar{l}) \mapsto \bar{k}$ is a quasi isomorphism. If $\bar{H} = \frac{\Ho(K^R)}{\Ho_n(K^R)}$, $\bar{K} = \frac{K^R \otimes_R k}{K_n^R \otimes_R k}[-1]$, then $\Ho(K^{\bar{R}}) = \bar{H} \lJoin \bar{K}$.
\end{theorem}
A typical element $(\bar{\alpha}, \bar{\beta})$ of degree $i$ in $\bar{H} \lJoin \bar{K}$ is given by a pair $(\alpha, \beta)$ such that $\alpha \in \Zi_i(K^R)$, $\beta \in K_{i-1}^R$, $\alpha$ is a lift of $\bar{\alpha}$ and $\beta$ is a lift of $\bar{\beta}$. By (1) of Lemma \ref{sbr1} choose $\gamma \in K_i^R$ such that $\partial(\gamma) = s \beta$ where $s$ is a generator of $\soc(R)$. Then the residue class of $\alpha - \gamma$ in $K^{\bar{R}}$ is a cycle and its homology class corresponds to $(\bar{\alpha}, \bar{\beta})$ by the above isomorphism.
\begin{proposition}\label{ggc3}
Let $(R, \mathfrak{m}, k)$ be an Artinian Gorenstein local ring with $\edim(R) = m \geq 2$ and $R\langle X \rangle$ be the acyclic closure of $R$. Let $\bar{R} = R/ \soc(R)$. Then $\bar{R}$ has an acyclic closure of the form $\bar{R} \langle X, Y, Z \rangle$ with the following properties.
\begin{enumerate}
\item
$Y = \{Y_i : i \geq 1\}$, $Z=\{Z_j : j \geq 1\}$, $\partial(Y_i) \in (0 : \mathfrak{m}^2) \bar{R} \langle X, Y, Z \rangle$ and $\partial(Z_j) \in z' \bar{R} \langle X, Y, Z \rangle$ for some $z' \in \Zi_{m-1}(K^R) \setminus \Bo_{m-1}(K^R)$.
\item
The $R$-linear map $R\langle X \rangle \rightarrow \bar{R}\langle X, Y, Z \rangle$ sending $X_i$ to $X_i$ induces injective maps on homology viz. $\Ho(R\langle X_i : \deg(X_i) \leq n \rangle) \hookrightarrow \Ho(\bar{R}\langle X_i, Y_j, Z_k : \deg(X_i), \deg(Y_j), \deg(Z_k) \leq n \rangle)$ and $\Ho(\mathfrak{m} R\langle X_i : \deg(X_i) \leq n \rangle) \hookrightarrow \Ho(\mathfrak{m} \bar{R}\langle X_i, Y_j, Z_k : \deg(X_i), \deg(Y_j), \deg(Z_k) \leq n \rangle)$ for all $n \geq 2$.
\end{enumerate}
\end{proposition}
\begin{proof}
Set $\bar{R} = R/ \soc(R)$, $U^n = R \langle X_i : \deg(X_i) \leq n \rangle$ and $\bar{U}^n = \bar{R} \otimes_R U^n$ for $n \geq 1$. The DG$\Gamma$ algebras $U^n, \bar{U}^n$ admit natural augmentation maps. By construction, we have $\I \Ho_i(U^n) = 0$ for $0 \leq i \leq n - 1$ and the homology classes of the cycles $\partial(X_i)$, $\deg(X_i) = n +1$ minimally generate $\I \Ho_n(U^n)$. Now $\deg(X_i) = 1$ for $1 \leq i \leq m = \edim(R)$. Let $\partial(X_i) = x_i$. Then $\mathfrak{m}$ is minimally generated by $\{x_i :1 \leq i \leq m\}$. Clearly $U^1$ is the Koszul complex $K^R$ on $x_1, \ldots, x_m$. Let $z \in \Zi_1(K^R) \setminus \Bo_1(K^R)$ be such that $\partial(X_{m+1}) = \bar{z}$. We choose $z' \in \Zi_{m-1}(K^R)$ such that $zz' = sX_1\ldots X_m$ (see \cite{av5}).
We construct inductively a sequence $V^n$ of distinguished DG$\Gamma$ algebras over $\bar{R}$ with the following properties.
\begin{enumerate}
\item $V^1 = K^{\bar{R}}$, the Koszul complex of $\bar{R}$.
\item
$\I \Ho_i(V^n) = 0$ for $0 \leq i \leq n -1$.
\item $V^n = \bar{U}^{n} \langle Y_{p_{n-1} + 1}, \ldots, Y_{p_{n}}, Z_{q_{n-1} + 1}, \ldots, Z_{q_{n}} \rangle$,
$n \geq 2$ for some nondecreasing sequences $\{p_i\}$, $\{q_i\}$ of integers. Homology classes of $\overline{\partial(X_i)}$ for $\deg(X_i) = n$ and $\partial(Y_j)$, $\partial(Z_k)$ for $p_{n-1} + 1 \leq j \leq p_n$ and $q_{n-1} + 1 \leq k \leq q_n$ minimally generate $\Ho_{n-1}(V^{n-1})$, $n \geq 2$ .
\item
$\partial(Y_j) \in (0 : \mathfrak{m}^2) V^{n-1}$ and $\partial(Z_k) \in z' V^{n-1}$ for $p_{n-1} + 1 \leq j \leq p_n$ and $q_{n-1} + 1 \leq k \leq q_n$.
\item
The composition $U^n \twoheadrightarrow \bar{U}^n \hookrightarrow V^n$ induces injective maps $\Ho(U^n) \hookrightarrow \Ho(V^n)$ and $\Ho(\mathfrak{m} U^n) \hookrightarrow \Ho(\mathfrak{m} V^n)$ for $n \geq 2$.
\end{enumerate}
Note that $\bar{U}^1$ is a Koszul complex of $\bar{R}$. We set $V^1 = \bar{U}^1$ with natural augmentation. By Theorem \ref{ggc1}, $\Ho_1(V^1) = \Ho_1(U^1) \oplus k$. The pairing $\frac{\mathfrak{m}}{\mathfrak{m}^2} \times \frac{(0 : \mathfrak{m}^2)}{(0 : \mathfrak{m})} \rightarrow \soc(R), (\bar{x}, \bar{y}) \mapsto xy$ is non-degenerate. We choose $y_i \in (0 : \mathfrak{m}^2)$, $1 \leq i \leq m$ such that $x_iy_j = 0$ for $i \not= j$ and $x_iy_j= s$ for $i = j$ where $s$ is a generator of $\soc(R)$. Then homology classes of $\overline{\partial(X_i)}$, $\deg(X_i) = 2$ and $\bar{y}_1X_1$ minimally generate $\Ho_1(V^1)$. Set $V^2 = \bar{U}^2 \langle Y_1 : \partial(Y_1) = \bar{y}_1X_1 \rangle$ with natural augmentation.
Here $p_0 = p_1 = 0$, $p_2 = 1$, $q_0 = q_1 = q_2 = 0$.
Clearly $\Ho_1(U^2) = \Ho_1(V^2) = 0$. We prove that the map $\Ho_i(U^2) \rightarrow \Ho_i(V^2)$, $i \geq 2$ is an injective map. Note that $\Zi_1(U^2) = \Bo_1(U^2)$.
We choose $\alpha \in \Zi_i(U^2), i \geq 2$ such that the image $\bar{\alpha}$ is in the boundary of $V^2$. Then $\bar{\alpha} = \partial(\beta)$, $\beta \in \bar{U}^2\langle Y_1 \rangle$. Let $\beta = a_0 + a_1Y_1 + \ldots + a_mY_1^{(m)}$, $a_i \in \bar{U}^2$. By Lemma \ref{ggc2} we have a $\bar{U}^2$-linear derivation $\nu$ on $V^2$ such that $\nu(Y_1) = 1$. We have $\nu^m(\beta) = a_m$. Since $\bar{\alpha}$ is free from $Y^{(i)}$ terms, $\nu(\alpha) = 0$. So $\partial(a_m) = \partial(\nu^m(\beta)) = \nu^m(\partial(\beta)) = \nu^m(\alpha) = 0$. This shows that $a_m \in \Zi(\bar{U}^2)$. If $\deg(a_m) = 0$, then $m \geq 2$ as $\deg(\beta) \geq 3$. We have $\partial(a_{m-1} + a_m Y) = \partial(\nu^{m-1}(\beta)) = \nu^{m-1}(\partial(\beta)) = 0$. So $a_{m-1} + a_m Y \in \Zi_2(V^2)$. Since $V^2$ is a distinguished DG$\Gamma$ algebra over $\bar{R}$, $\Zi_{\geq 1}(V^2) \subset \mathfrak{m} V^2$. So $a_m \in \mathfrak{m}$ if $\deg(a_m) = 0$. Therefore, $a_m \in \I \Zi(\bar{U}^2) \in \mathfrak{m} \bar{U}^2$ by (4) of Lemma \ref{sbr1}.
This implies that $\partial(a_m Y_1^{(m)}) = 0$ since $y_1 \mathfrak{m} \subset \soc(R)$. So $\bar{\alpha} = \partial(\beta_1)$, $\beta_1 = \beta - a_mY^{(m)}$. Repeating this argument $m$ times we have $\bar{\alpha} = \partial(a_0)$. So $\alpha$ is in the boundary of $\bar{U}^2$. By (5) of Lemma \ref{sbr1}, $\alpha$ is in the boundary of $U^2$. This shows that the map $\Ho(U^2) \hookrightarrow \Ho(V^2)$ is injective. The proof of that the map $\Ho(\mathfrak{m} U^2) \hookrightarrow \Ho(\mathfrak{m} V^2)$ is injective follows similarly (in fact more easily as $a_i$ are already in $\mathfrak{m} \bar{U}^2$).
Now we assume that $V^n, n \geq 2$ is constructed with aforesaid properties and construct $V^{n+1}$. Since $V^n$ is a distinguished DG$\Gamma$ algebra over $\bar{R}$, we have $\I \Zi(V^n) \subset \mathfrak{m} V^n$. \\
\noindent
{\bf Claim - 1} : {\it $\Zi_n(V^n) = \Bo_n(V^n) + \bar{R} \otimes_R \Zi_n(U^n) + (0 : \mathfrak{m}^2) V_n^n + z'V^n \cap \Zi_n(V^n), n \geq 2.$}
Set $S = \Bo_n(V^n) + \bar{R} \otimes_R \Zi_n(U^n) + (0 : \mathfrak{m}^2) V_n^n + z'V^n \cap \Zi_n(V^n)$. Clearly $S \subset \Zi_n(V^n)$.
To show the reverse inclusion, choose $\alpha \in \Zi_n(V^n)$.
Then $\alpha$ is a finite linear combination of $\Gamma$-monomials in $Y_j$ and $Z_k$ with coefficients in $\bar{U}^n$. Let the total degree of $\alpha$ in $Y_i$ and $Z_k$ be $d$. We prove $\alpha \in S$ by induction on $d$. If $d = 0$, then $\alpha$ has no terms involving $Y_j$ and $Z_k$. So $\alpha \in \Zi_n(\bar{U}^n)$. By (4) of Lemma \ref{sbr1}, $\alpha \in \bar{R} \otimes_R \Zi_n(U^n) + (0 : \mathfrak{m}^2) \bar{U}_n^n + z' \bar{U}^n \cap \Zi_n(\bar{U}^n)$. Since $\bar{U}^n \subset V^n$, we have $\alpha \in S$.
Now we assume that $d > 0$. We choose a $\Gamma$-monomial term in $Y_j, Z_k$ of the form $f M$ for $f \in \bar{U}^n$, $M= Y_{l_1}^{(c_1)}\ldots Y_{l_s}^{(c_s)}Z_{m_1}^{(d_1)} \ldots Z_{m_t}^{(d_t)}$ in the expression of $\alpha$ such that $\deg(M) = d$, $\deg(f) = e$ and $n = d + e$.
Since $\alpha$ is a cycle and $M$ is a $\Gamma$-monomial of highest total degree in $Y_j, Z_k$, $f$ is a cycle in $\bar{U}^n$. So by (4) of Lemma \ref{sbr1}, we have $f \in \bar{R}\otimes_R \Zi_e(U^n) + (0 : \mathfrak{m}^2)\bar{U}_e^n + \Zi_e(\bar{U}^n) \cap z' \bar{U}^n$.
If $1 \leq e < n$, we have $\Zi_e(U^n) = \Bo_e(U^n)$. If $e = 0$, then $f \in \mathfrak{m} \bar{R}$ since $\alpha \in \Zi_n(V^n) \subset \mathfrak{m} V^n$, $n \geq 2$. We write $f = \partial(g) + f' + f''$ for $g \in \bar{U}^n_{e+1}$, $f' \in (0 : \mathfrak{m}^2)\bar{U}_e^n$ and $f'' \in \Zi_e(\bar{U}^n) \cap z' \bar{U}^n$. So $fM = \partial(gM) + f'M + f''M + (-1)^eg\partial(M)$. Note that $f'M \in (0 : \mathfrak{m}^2)V_n^n$ and $f''M \in z'V^n \cap \Zi_n(V^n)$ since $z' \partial(M) = 0$. Also $g \partial(M)$ has total degree in $Y_j$ and $Z_k$ strictly less than $d$. If we write each $\Gamma$-monomial term of total degree $d$ in $Y_j$, $Z_k$ appearing in the expression of $\alpha$ in this manner, we have $\alpha = \alpha_1 + \alpha_2$ where $\alpha_1 \in S$, $\alpha_2 \in \Zi_n(V^n)$ and total degree of $\alpha_2$ in $Y_j$, $Z_k$ is strictly less than $d$. By the induction hypothesis, $\alpha_2 \in S$. So $\alpha \in S$ and our claim follows.
The homology classes of cycles $\partial(X_i)$, $\deg(X_i) = n+1$ minimally generate $\Ho_n(U^n)$. Because of Claim - 1 and the fact that $\Ho(U^n) \hookrightarrow \Ho(V^n)$ is injective,
we can adjoin $\Gamma$-variables $X_i$ for $\deg(X_i) = n+1$ and
$Y_j$, $Z_k$ for $p_n + 1 \leq j \leq p_{n+1}$, $q_n + 1 \leq k \leq q_{n+1}$ to $V^n$ such that homology classes of $\overline{\partial(X_i)}$, $\partial(Y_j)$, $\partial(Z_k)$ minimally generate $\Ho_n(V^n)$ and $\partial(Y_j) \in (0 : \mathfrak{m}^2)V^n$, $\partial(Z_k) \in z'V^n$. We define
\begin{multline*}
V^{n+1} = V^n \langle X_i, Y_j, Z_k : \deg(X_i) = n+1, p_n + 1 \leq j \leq p_{n+1}, q_n + 1 \leq k \leq q_{n+1} \rangle \\= \bar{U}^{n+1} \langle Y_j, Z_k : 1 \leq j \leq p_{n+1}, 1 \leq k \leq q_{n+1} \rangle
\end{multline*}
with the natural augmentation. Clearly $\I \Ho_{i}(V^{n+1}) = 0$ for $0 \leq i \leq n$.
\noindent
{\bf Claim - 2}:
{\it If $f \in \I\Zi(\bar{U}^{n+1})$ and $M = Y_{l_1}^{(c_1)}\ldots Y_{l_s}^{(c_s)} Z_{m_1}^{(d_1)} \ldots Z_{m_t}^{(d_t)}$ is a $\Gamma$-monomial of total degree $d$
, then $\partial(f M) = \partial(\beta)$ for some $\beta \in (0 : \mathfrak{m}^2)V^{n+1} + z'V^{n+1}$ and the total degree of $\beta$ in $Y_j, Z_k$ is strictly less than $d$.}
Let $\partial(M) = \sum_{i = 1}^m g_iM_i$ where $g_i \in (0 : \mathfrak{m}^2)\bar{U}^{n+1} + z'\bar{U}^{n+1}$ and $M_i$ are $\Gamma$-monomials in $Y_j$, $Z_k$ of total degree strictly less than $d$. One observes that $g_i \partial(Y_j)$, $g_i \partial(Z_k)$ are zero. So $\partial^2(M) = \sum_{i = 1}^m \partial(g_i) M_i = 0$. This shows that $g_i \in \Zi(\bar{U}^{n+1}) \cap [(0 : \mathfrak{m}^2)\bar{U}^{n+1} + z'\bar{U}^{n+1}]$. So we can write $g_i = g'_i + g''_i$ for $g'_i \in (0 : \mathfrak{m}^2)\bar{U}^{n+1}$ and $g''_i \in \Zi(\bar{U}^{n+1}) \cap z'\bar{U}^{n+1}$.
Since $f \in \mathfrak{m} \bar{U}^{n+1}$, we have $f g_i' = 0$.
By (7) of Lemma \ref{sbr1}, we have $fg_i = fg''_i = \partial(h_i)$ for some $h_i \in (0 : \mathfrak{m}^2)\bar{U}^{n+1} + z'\bar{U}^{n+1}$. Therefore, $\partial(fM) = (-1)^{\deg{f}} \sum_{i = 1}^m f g_iM_i = (-1)^{\deg{f}} \sum_{i = 1}^m \partial(h_i) M_i = \partial(\beta)$ where $\beta = (-1)^{\deg{f}} \sum_{i = 1}^m h_i M_i \in (0 : \mathfrak{m}^2)V^{n+1} + z'V^{n+1}$. So the claim follows.
\noindent
{\bf Claim - 3}: The composition map $U^{n+1} \twoheadrightarrow \bar{U}^{n+1} \hookrightarrow V^{n+1}$ induces injective maps $\Ho(U^{n+1} ) \hookrightarrow \Ho(V^{n+1})$ and $\Ho(\mathfrak{m} U^{n+1} ) \hookrightarrow \Ho(\mathfrak{m} V^{n+1})$.
We only show that $\Ho(U^{n+1} ) \hookrightarrow \Ho(V^{n+1})$ is injective. The injective property of the second map will follow similarly. Note that $\Zi_i (U^{n+1}) = \Bo_i (U^{n+1})$ for $1 \leq i \leq n$.
Let $\alpha \in \Zi_i (U^{n+1})$ for $i \geq n + 1$ be such that $\bar{\alpha} \in \bar{U}^{n+1}$ is in the boundary of $V^{n+1}$. Let $\bar{\alpha} = \partial(\beta)$, $\beta \in V_{i+1}^{n+1}$. We rename $Y_j$, $Z_k$ to $W_l$, $1 \leq l \leq p_{n+1} + q_{n+1}$ such that $\deg(W_l) \leq \deg(W_{l'})$ for $l < l'$. This is possible because for any positive integer $d$, there are only finitely many $Y_j$, $Z_k$ such that $\deg(Y_j) = \deg(Z_k) = d$.
Now $V^{n+1} = \bar{U}^{n+1}\langle W_l : 1 \leq l \leq p_{n+1} + q_{n+1}\rangle$ is a distinguished DG$\Gamma$ algebra over $\bar{U}^{n+1}$ since by the construction
the homology classes of $\partial(W_l)$, $\deg(W_l) = d$ minimally generate $\I \Ho_{d-1}(\bar{U}^{n+1}\langle W_l : \deg(W_l) \leq d - 1 \rangle) = \I \Ho_{d-1}(V^{d-1})/ \langle cls(\partial(X_i)) : \deg(X_i) = d \rangle$ for $1 \leq d \leq n + 1$.
So for each $W_l$, we have a $\bar{U}^{n+1}$-linear derivation on $V^{n+1}$ such that $\nu_j(W_j) = 1$ and $\nu_j(W_i) = 0$ for $i < j$ by Lemma \ref{ggc2}.
Let $\init(\beta) = M$ where $M = W_{1}^{(c_1)} \ldots W_{r}^{(c_r)}$, $c_l \geq 0, c_r \geq 1$. Let the leading coefficient of $\beta$ be $f$. Then by Lemma \ref{ggc2}, we have $\nu_M(\beta) = \pm f$. Now $\pm \partial(f) = \partial(\nu_M(\beta)) = \nu_M(\partial(\beta))= \nu_M(\bar{\alpha}) = 0$ since $\bar{\alpha}$ is free from $W_l$. So, we have $f \in \Zi(\bar{U}^{n+1})$. If $\deg(f) = 0$, then $\sum_{l = 1}^r c_l > 1$ since in this case $\deg(\beta) = \deg(M) \geq n + 2$ and $\deg(W_l) \leq n + 1$.
We have $\nu_r(\beta) \in \Zi_{\geq 1}(V^{n+1}) \subset \mathfrak{m} V^{n+1}$, $\init(\nu_r(\beta)) = \nu_r(M)$ and the leading coefficient of $\nu_r(\beta)$ is $\pm f$. So $f \in \mathfrak{m}$ if $\deg(f) = 0$. Therefore, we conclude that $f$ is in $\I \Zi(\bar{U}^{n+1})$.
By Claim - 2, $\partial(fM)$ can be written as the boundary of an element in $V^{n+1}$ of total degree in $W_l$ (equivalently in $Y_j$, $Z_k$) strictly less than that of $\deg(M)$. So we can write $\bar{\alpha} = \partial(\beta_1)$, $\init(\beta_1) \prec \init(\beta)$. If we repeat this argument, we have after finitely many steps a $\beta_L \in V_{i+1}^{n+1}$ free from $W_l$ (equivalently from $Y_j, Z_k$) such that $\bar{\alpha} =\partial( \beta_L)$. This means that $\bar{\alpha}$ is in the boundary of $\bar{U}^{n+1}$. By (5) of Lemma \ref{sbr1}, $\alpha$ is in the boundary of $U^{n+1}$.
Therefore, our induction step of construction $V^{n+1}$ is complete. We define the acyclic closure of $\bar{R}$ as $\bar{R} \langle X, Y, Z \rangle = \projlim V^{n}$. The DG$\Gamma$ algebras $\{V^n\}$ define a filtration of $\bar{R} \langle X, Y, Z \rangle$. The properties (1) and (2) stated in the theorem, are satisfied by $\bar{R} \langle X, Y, Z \rangle$ by the construction.
\end{proof}
\begin{theorem}\label{ggc4}
Let $(R, \mathfrak{m}, k)$ be an Artinian Gorenstein local ring. Let $l \geq 2$. Then $R$ is a generalised Golod ring of level $l$ if and only if $R/ \soc(R)$ is so.
\end{theorem}
\begin{proof}
Any Artinian local ring of embedding dimension one is a Golod ring. Therefore, we may assume that $\edim(R) = n \geq 2$.
Let $R\langle X \rangle$ be the acyclic closure of $R$ and $\bar{R}\langle X, Y, Z\rangle$ be the acyclic closure of $\bar{R} = R/ \soc(R)$ as constructed in Proposition \ref{ggc3}.
We use notations as set in the proof of Proposition \ref{ggc3}. We have the following commutative diagram. The vertical maps are injective by (2) of Proposition \ref{ggc3}.
\[\xymatrix{
\Ho(\mathfrak{m} U^l) \ar@{^{(}->}[d] \ar[r] &\Ho(\mathfrak{m} R\langle X \rangle) \ar@{^{(}->}[d]\\
\Ho(\mathfrak{m} V^l) \ar[r] &\Ho (\mathfrak{m} \bar{R}\langle X, Y, Z \rangle) }\]
First we assume that $\bar{R}$ is a generalised Golod ring of level $l$. Then $V^l$ is a Golod algebra and the inclusion $\mathfrak{m} V^l \hookrightarrow \mathfrak{m} R\langle X, Y, Z \rangle$ induces an injective map $\Ho(\mathfrak{m} V^l) \hookrightarrow \Ho(\mathfrak{m} R\langle X, Y, Z \rangle)$ by Theorem \ref{prl7}. So, the top horizontal map is injective. Therefore, $U^l$ is a Golod algebra and the ring $R$ is a generalised Golod ring of level $l$ by Theorem \ref{prl7}.
Now we assume that $R$ is a generalised Golod ring of level $l$. So $U^l$ is a Golod algebra and the inclusion $\mathfrak{m} U^l \hookrightarrow \mathfrak{m} R\langle X \rangle$ induces an injective map $\Ho(\mathfrak{m} U^l) \hookrightarrow \Ho(\mathfrak{m} R\langle X \rangle)$ by Theorem \ref{prl7}. Let $\alpha \in \Zi(\mathfrak{m} V^l)$ be a boundary in $\mathfrak{m} \bar{R}\langle X, Y, Z \rangle$. Then $\alpha = \partial(\beta)$, $\beta \in \mathfrak{m} \bar{R}\langle X, Y, Z \rangle$. Since $V^n, n \geq 1$ defines a filtration of $\bar{R}\langle X, Y, Z \rangle$, we have $\beta \in \mathfrak{m} V^N$ for large $N$. We show that $\alpha$ is in the boundary of $\mathfrak{m} V^l$ in two cases.
\noindent
{\bf Case - 1} : $\alpha$ is free from $Y_j$, $Z_k$, i.e. $\alpha \in \mathfrak{m} \bar{U}^l$.
By the argument as in Claim - 3, Proposition \ref{ggc3}, we may assume that $\beta$ is free from $Y_j$ and $Z_k$, i.e. $\beta \in \mathfrak{m} \bar{U}^N$. Let $\tilde{\alpha}$, $\tilde{\beta}$ be lifts of $\alpha$, $\beta$ in $\mathfrak{m} U^l$, $\mathfrak{m} U^N$ respectively. Then $\tilde{\alpha} = \partial(\tilde{\beta}) + w$, $w \in \soc(R) R \langle X \rangle$. This shows that $\tilde{\alpha} \in \Zi(\mathfrak{m} U^l)$. We see that $\alpha$ is in the boundary of $\mathfrak{m} \bar{U}^N$. So $\tilde{\alpha}$ is in the boundary of $\mathfrak{m} U^N$ by (5) of Lemma \ref{sbr1}. The top horizontal map in the diagram is injective. So $\tilde{\alpha}$ is in the boundary of $\mathfrak{m} U^l$. This shows that $\alpha$ is in the boundary of $\mathfrak{m} \bar{U}^l$ in particular in $\mathfrak{m} V^l$.
We rename the set of variables $Y_j$, $Z_k$, $\deg(Y_j), \deg(Z_k) \leq N$ to $W_i, 1 \leq i \leq p_N + q_N$ such that $\deg(W_i) \leq \deg(W_{i'})$ for $i < i'$.
This is possible because for any positive integer $d$, there are only finitely many $Y_j$, $Z_k$ such that $\deg(Y_j) = \deg(Z_k) = d$.
We write $V^N = \bar{U}^N\langle W_i; 1 \leq i \leq p_N + q_N \rangle$. For each $W_i$, we have a $\bar{U}^{N}$-linear derivation on $V^N$ such that $\nu_j(W_j) = 1$ and $\nu_j(W_i) = 0$ for $i < j$ by Lemma \ref{ggc2}.
\noindent
{\bf Case - 2} : $\alpha$ is arbitrary.
We have $V^l = \bar{U}^l \langle W_i : 1 \leq i \leq p_l + q_l \rangle$.
We write $\alpha$ as a linear combination of $\Gamma$-monomials in $W_i$ with coefficients in $\bar{U}^l $.
Let $\init(\alpha) = M$ and the leading coefficient of $\alpha$ be $f \in \mathfrak{m} \bar{U}^l$.
Note that $\pm f = \nu_{M}(\alpha) = \nu_{M}(\partial(\beta)) = \pm \partial(\nu_{M}(\beta))$, $\nu_{M}(\beta) \in \mathfrak{m} V^N$. So $f$ is in the boundary of $\mathfrak{m} V^l$ by Case - 1, i.e. $f = \partial(g)$ for some $g \in \mathfrak{m} V^l$. The initial $\Gamma$-monomial of $\alpha - \partial(gM)$ is less than $M$.
If we repeat the above argument for $\alpha - \partial(gM)$ in place of $\alpha$, we have after a finite number of steps a $\beta' \in \mathfrak{m} V^l$ such that $\alpha - \partial(\beta') = \alpha' \in \mathfrak{m} \bar{U}^l$. The element $\alpha'$ is in the boundary of $\mathfrak{m} V^N$ since $\alpha$ is so. So by Case - 1, $\alpha'$ is in the boundary of $\mathfrak{m} \bar{U}^l$. This shows that $\alpha$ is in the boundary of $\mathfrak{m} V^l$ and thus the lower horizontal map is injective. Therefore, $\bar{R}$ is a generalised Golod ring of level $l$.
\end{proof}
\section{The generalised Golod property of fibre products and connected sums}
The following lemma gives a construction of a semi-free DG$\Gamma$ algebra over $R$ from semi-free DG$\Gamma$ algebras over $S$ and $T$.
\begin{lemma}\label{ggfl1}
Let $(S, \mathfrak{p}, k)$, $(T, q, k)$ be two local rings and $R = S \times_k T$. Let $(S\langle X \rangle, \partial_1)$, $(T\langle Y \rangle, \partial_2)$ be two semi-free minimal DG$\Gamma$ algebra extensions of $S$, $T$ obtained by adjoining graded sets of variables $X$, $Y$ respectively.
Then there are semi-free DG$\Gamma$ $R$-algebra extensions $(R\langle X \rangle, \tilde{\partial}_1)$, $(R\langle Y \rangle, \tilde{\partial}_2)$ such that the following hold.
\begin{enumerate}
\item
The differential map satisfies $ \tilde{\partial}_{1}[R\langle X \rangle] \subset \mathfrak{p} R\langle X \rangle$ and $ \tilde{\partial}_2 R\langle Y \rangle \subset \mathfrak{q} R\langle Y \rangle$.
\item
There are isomorphisms of DG$\Gamma$ algebras $S \otimes_R R\langle X \rangle = S\langle X \rangle$ and $T \otimes_R R\langle Y \rangle = T\langle Y \rangle$.
\end{enumerate}
\end{lemma}
\begin{proof}
The maximal ideal of $S \times_k T$ is $\mathfrak{p} \oplus \mathfrak{q}$. There are inclusions $i : \mathfrak{p} \hookrightarrow \mathfrak{p} \oplus \mathfrak{q}$ and $j : \mathfrak{q} \hookrightarrow \mathfrak{p} \oplus \mathfrak{q}$.
We consider the skew-commutative graded algebras $R\langle X \rangle$ and $R\langle Y \rangle$. We define differential maps $ \tilde{\partial}_1$ on $R\langle X \rangle$ and $ \tilde{\partial}_2$ on $R\langle Y \rangle$ as follows:
Pick $x \in X$ and $y \in Y$. Let $ \partial_1(x) = \sum p_{\underline{x}} \, \underline{x}$ and $ \partial_2(y) = \sum q_{\underline{y}} \, \underline{y}$. Here $p_{\underline{x}} \in \mathfrak{p}$, $q_{\underline{y}} \in \mathfrak{q}$; the terms $\underline{x}$, $\underline{y}$ denote $\Gamma$-monomials on variables in $X$, $Y$ respectively and the summation is taken over only finitely many terms. Set $ \tilde{\partial}_1(x) = \sum i(p_{\underline{x}}) \underline{x}$ and $ \tilde{\partial}_2(y) = \sum j(q_{\underline{y}}) \underline{y}$. Extend $ \tilde{\partial}_1$, $ \tilde{\partial}_2$ to degree $(-1)$ maps on $R\langle X \rangle$, $R\langle Y \rangle$ respectively by $R$-linearity and the Leibnitz rule. The DG$\Gamma$ algebras $(R\langle X \rangle, \tilde{\partial}_1)$, $(R\langle Y \rangle, \tilde{\partial}_2)$ satisfy properties (1), (2) by the construction.
\end{proof}
\begin{lemma}\label{ggfl2}
With the notations as above, assume further that $\Ho_0(S \langle X \rangle) = \Ho_0(T \langle Y \rangle) = k$ and $\Zi_{\geq 1}(S \langle X \rangle) \subset \mathfrak{p} S \langle X \rangle$, $\Zi_{\geq 1}(T \langle Y \rangle) \subset \mathfrak{q} T \langle Y \rangle$.
Set $R\langle {X \sqcup Y} \rangle = R\langle {X} \rangle \otimes_R R\langle Y \rangle$. Then the
DG$\Gamma$ algebra $R\langle X \sqcup Y\rangle$ is Golod if and only if both $S\langle X \rangle$, $T\langle Y \rangle$ are Golod algebras.
\end{lemma}
\begin{proof}
We have the following exact sequence of $R$-modules.
\[0 \rightarrow R \xrightarrow{\alpha} S \times T \xrightarrow{\beta} k \rightarrow 0.\]
Here $\alpha(s, t) = (s, t)$ and $\beta(s, t) = s - t$. Taking tensor product with the algebra $R \langle X \sqcup Y \rangle$, we have the following short exact sequence of complexes of $R$-modules.
\[0 \rightarrow R \langle X \sqcup Y \rangle \xrightarrow{\alpha \otimes id} S \otimes_R R \langle X \sqcup Y \rangle \oplus T \otimes_R R \langle X \sqcup Y \rangle \xrightarrow{\beta \otimes id} k \otimes_R R\langle X \sqcup Y \rangle \rightarrow 0.\]
Here $id$ denotes the identity map on the complex $R \langle X \sqcup Y \rangle$. Note that $S \otimes_R R\langle {X \sqcup Y} \rangle = S\langle {X} \rangle \otimes_S S\langle Y \rangle$. The differential $S \otimes_R \tilde{\partial}_2$ is zero on $S\langle Y \rangle$. Therefore, $\Ho(S \otimes_R R \langle X \sqcup Y \rangle) = \Ho (S \langle X \rangle) \otimes_k k\langle Y \rangle$. Similarly we have $\Ho(T \otimes_R R \langle X \sqcup Y \rangle) = k\langle X \rangle \otimes_k \Ho (T \langle Y \rangle)$. The differential on $k \otimes_R R\langle X \sqcup Y \rangle$ is zero. So the long exact sequence of homology has the form
\[ k \langle X \sqcup Y \rangle[+1] \xrightarrow{\partial} \Ho(R \langle X \sqcup Y\rangle) \xrightarrow{\Ho(\alpha \otimes id)} \Ho (S \langle X \rangle) \otimes_k k\langle Y \rangle \oplus k\langle X \rangle \otimes_k \Ho (T \langle Y \rangle) \xrightarrow{\Ho(\beta \otimes id)} k \langle X \sqcup Y \rangle.\]
Let $\ker (\Ho(\beta \otimes id)) = K$ and $\coker (\Ho(\beta \otimes id)) = C$. Then $\Ho_{R\langle X \sqcup Y\rangle}(t) = \frac{1}{t}\Ho_C(t) + \Ho_K(t)$ where $\Ho_C(t) = \sum_{i \geq 0} \dim_k \Ho_i (C)$ and $\Ho_K(t) = \sum_{i \geq 0} \dim_k \Ho_i (K)$.
The quotient maps $\pi_R : S \rightarrow k$ and $\pi_S : T \rightarrow k$ induce maps $\pi_{S\langle X \rangle} : S\langle X \rangle \rightarrow k\langle X \rangle$ and $\pi_{T\langle Y \rangle} : T\langle Y \rangle \rightarrow k\langle Y \rangle$ respectively. Choose cycles $z \in \Zi(S\langle X \rangle)$, $z' \in \Zi(T\langle Y \rangle)$ and elements $u \in k\langle X \rangle$ and $v \in k\langle Y \rangle$. Let $[z]$, $[z']$ denote homology classes of $z$, $z'$ respectively. Note that $\Ho(\beta \otimes id)([z]\otimes v, u \otimes [z']) = \pi_{S\langle X \rangle}(z)v - u \pi_{T\langle Y \rangle}(z')$.
By the given hypothesis $\Zi_{\geq 1}(S \langle X \rangle) \subset \mathfrak{p} S \langle X \rangle$ and $\Zi_{\geq 1}(T \langle Y \rangle) \subset \mathfrak{q} T \langle Y \rangle$. So $\Ho(\beta \otimes id)$ is zero on $\Ho_{\geq 1} (S \langle X \rangle) \otimes_k k\langle Y \rangle \oplus k\langle X \rangle \otimes_k \Ho_{\geq 1} (T \langle Y \rangle)$. The restriction of $\Ho(\beta \otimes id)$ on $\Ho_0 (S \langle X \rangle) \otimes_k k\langle Y \rangle \oplus k\langle X \rangle \otimes_k \Ho_0 (T \langle Y \rangle) = k\langle Y \rangle \oplus k \langle X \rangle$ has one dimensional kernel generated by $(1, 1)$.
So we have $K = \Ho_{\geq 1} (S \langle X \rangle) \otimes_k k\langle Y \rangle \oplus k\langle X \rangle \otimes_k \Ho_{\geq 1} (T \langle Y \rangle) \oplus k(1, 1)$ and $C = \frac{k\langle X \sqcup Y \rangle}{k \oplus k\langle X \rangle_{\geq 1} \oplus k\langle Y \rangle_{\geq 1}}$. Note that $k\langle {X \sqcup Y} \rangle = k\langle {X} \rangle \otimes_k k\langle Y \rangle$. Therefore, we have ${\Ho}_{k\langle X \sqcup Y \rangle}(t) = \Ho_{k\langle X \rangle}(t) \Ho_{k\langle X \rangle}(t)$ and
\begin{align*}
{\Ho}_{\Ho(R\langle X \sqcup Y\rangle)}(t) &= \frac{1}{t}{\Ho}_C(t) + {\Ho}_K(t)\\
&=\frac{1}{t}[{\Ho}_{k\langle X \rangle}(t){\Ho}_{k\langle Y \rangle}(t) - ({\Ho}_{k\langle X \rangle}(t) -1) - ({\Ho}_{k\langle Y \rangle}(t) - 1) -1]\\
& + ({\Ho}_{\Ho(S\langle X \rangle)}(t) -1){\Ho}_{k\langle Y \rangle}(t) + {\Ho}_{k\langle X \rangle}(t)({\Ho}_{\Ho(T\langle Y \rangle)}(t) -1) +1\\
&= 1 + {\Ho}_{k\langle Y \rangle}(t)({\Ho}_{\Ho(S\langle X \rangle)}(t) -1) + {\Ho}_{k\langle X \rangle}(t)({\Ho}_{\Ho(T\langle Y \rangle)}(t) -1) + \frac{1}{t}({\Ho}_{k\langle X \rangle}(t) -1)({\Ho}_{k\langle Y \rangle}(t) - 1).
\end{align*}
The above equality gives
\[1 - t({\Ho}_{\Ho(R\langle X \sqcup Y \rangle)} (t) - 1) = {\Ho}_{k\langle X \rangle}(t)[ 1 - t ({\Ho}_{\Ho(T\langle Y \rangle)}(t) - 1)] + {\Ho}_{k\langle Y \rangle}(t)[ 1 - t ({\Ho}_{\Ho(S\langle X \rangle)}(t) - 1)] - {\Ho}_{k\langle X \rangle}(t){\Ho}_{k\langle Y \rangle}(t).\]
We have term wise inequality of power series
\begin{equation}\label{ggfle1}
P^S_k(t) \prec \frac{{\Ho}_{k\langle X \rangle}(t)}{1 - t ({\Ho}_{\Ho(S\langle X \rangle)}(t) - 1)} \ \text{and} \ P^T_k(t) \prec \frac{{\Ho}_{k\langle Y \rangle}(t)}{1 - t ({\Ho}_{\Ho(T\langle Y \rangle)}(t) - 1)}.
\end{equation}
\begin{align*}
\text{We have}\quad &\frac{1 - t({\Ho}_{\Ho(R\langle X \sqcup Y\rangle)} (t) - 1)}{{\Ho}_{k\langle X \sqcup Y \rangle}(t)}\\
&= \frac{{\Ho}_{k\langle X \rangle}(t)[ 1 - t ( {\Ho}_{\Ho(T\langle Y \rangle)}(t) - 1)] + {\Ho}_{k\langle Y \rangle}(t)[ 1 - t ( {\Ho}_{\Ho(S\langle X \rangle)}(t) - 1)] - {\Ho}_{k\langle X \rangle}(t){\Ho}_{k\langle Y \rangle}(t)}{{\Ho}_{k\langle X \rangle}(t){\Ho}_{k\langle Y \rangle}(t)}\\
&=\frac{1 - t ({\Ho}_{\Ho(T\langle Y \rangle)}(t) - 1)}{{\Ho}_{k\langle Y \rangle}(t)} + \frac{1 - t ({\Ho}_{\Ho(S\langle X \rangle)}(t) - 1)}{{\Ho}_{k\langle X \rangle}(t)} - 1\\
&\prec \frac{1}{P^T_k(t)} + \frac{1}{P^S_k(t)} - 1 \ \ \text{from \ref{ggfle1}}\\
&= \frac{1}{P^R_k(t)}.
\end{align*}
The DG$\Gamma$ algebra $R\langle X \sqcup Y \rangle$ is Golod if and only if the above inequality is an equality. This is equivalent to that both inequalities in \ref{ggfle1} are equalities, i.e. both $S\langle X \rangle$, $T\langle Y \rangle$ are Golod algebras.
\end{proof}
\begin{proposition}\label{ggfl3}
Let $(S, \mathfrak{p}, k)$, $(T, \mathfrak{q}, k)$ be local rings. Let $(S\langle X \rangle, \partial_1)$, $(T\langle Y \rangle, \partial_2)$ be acyclic closures of $S$, $T$ respectively. Set $R = S \times_k T$. Let $(R\langle X \rangle, \tilde{\partial}_1)$, $(R\langle Y \rangle, \tilde{\partial}_2)$ be semi-free DG$\Gamma$ extensions of $R$ as described in Lemma \ref{ggfl1}. Let $p : R\langle X \rangle\twoheadrightarrow S \otimes_R R\langle X \rangle = S\langle X \rangle$, $q : R\langle Y \rangle\twoheadrightarrow T \otimes_R R\langle Y \rangle = T \langle Y \rangle$ be natural quotient maps.
Then $R$ admits an acyclic closure of the form $(R\langle X, Y, Z \rangle, \partial)$ which satisfies the following properties.
\begin{enumerate}
\item
The inclusions $\alpha : R\langle X \rangle \hookrightarrow R\langle X, Y, Z \rangle$, $\beta : R\langle Y \rangle \hookrightarrow R\langle X, Y, Z \rangle$ are chain maps of complexes.
\item
There are surjective homorphisms of DG$\Gamma$ algebras $\phi : R\langle X, Y, Z \rangle \twoheadrightarrow S\langle X\rangle$ and $\psi : R\langle X, Y, Z \rangle \twoheadrightarrow T\langle Y\rangle$ such that $\phi \circ \beta = 0$, $\psi \circ \alpha = 0$ and $\phi \circ \alpha = p $, $\psi \circ \beta = q$
\end{enumerate}
\end{proposition}
\begin{proof}
Set $U^n = S \langle X_i : \deg(X_i) \leq n)$ and $V^n = T \langle Y_i : \deg(Y_i) \leq n)$, $n \geq 1$. Then $U^n$, $V^n$, $n \geq 1$ define filtrations of the acyclic closures $S\langle X \rangle$, $T\langle Y \rangle$ respectively. Augmentation maps of $S\langle X \rangle$, $T\langle Y \rangle$ restrict to augmentation maps of $U^n$, $V^n$ respectively.
From the construction of acyclic closures, it follows that $\I \Ho_i(U^n) = \I \Ho_i(V^n) = 0$ for $0 \leq i \leq n - 1$, the homology classes of cycles $\partial_1(X_i)$ for $\deg(X_i) = n + 1$ minimally generate $\I \Ho_n(U^n)$ and the homology classes of cycles $\partial_2(Y_i)$ for $\deg(Y_i) = n + 1$ minimally generate $\I \Ho_n(V^n)$. Set $\tilde{U}^n = R \langle X_i : \deg(X_i) \leq n\rangle$ and $\tilde{V}^n = R \langle Y_i : \deg(Y_i) \leq n\rangle$, $n \geq 1$.
Let $p^n : \tilde{U}^n \twoheadrightarrow \tilde{U}^n \otimes_R S = U^n$, $q^n : \tilde{V}^n \twoheadrightarrow \tilde{V}^n \otimes_R S = V^n$ be natural quotient maps.
We construct inductively a sequence of augmented $DG\Gamma$ algebras $(F^nR, \partial^n)$ with natural augmentation $\epsilon^n : F^nR \twoheadrightarrow k$ satisfying the following properties.
\begin{enumerate}
\item
$F^1R = K^R$, the Koszul complex of $R$.
\item
$\I\Ho_i(F^nR) = 0$ for $0 \leq i \leq n - 1$.
\item
There are inclusions $\alpha^n : \tilde{U}^n \hookrightarrow F^nR$ and $\beta^n : \tilde{V}^n \hookrightarrow F^nR$ of complexes.
\item
There are surjective DG$\Gamma$ algebra homomorphisms $\phi^n : F^nR \twoheadrightarrow U^n$ and $\psi^n : F^nR \twoheadrightarrow V^n$ such that the compositions $\phi^n \circ \alpha^n = p^n$, $\psi^n \circ \beta^n = q^n$ and $\phi^n \circ \beta^n = 0$, $\psi^n \circ \alpha^n = 0$.
\end{enumerate}
When $n = 1$,
we define $F^1R = \tilde{U}^1 \otimes_R \tilde{V}^1 = R \langle X_i, Y_j : \deg(X_i) = \deg(Y_j) = 1 \rangle$, the Koszul complex on a minimal set of generators of $\mathfrak{p} \oplus \mathfrak{q}$ with the natural augmentation $\epsilon^1$. Define $\alpha^1 : \tilde{U}^1 \hookrightarrow F^1R$, $\beta^1 : \tilde{V}^1 \hookrightarrow F^1R$ to be natural inclusions and $\phi^1 : F^1R \twoheadrightarrow \frac{F^1R}{(Y_i : \deg Y_i = 1)F^1 R} \otimes_R S = U^1$, $\psi^1 : F^1R \twoheadrightarrow \frac{F^1R}{(X_i : \deg X_i = 1)F^1 R} \otimes_R T = V^1$ to be natural quotient maps. The properties are satisfied obviously.
Now we assume that $F^nR$, $n \geq 2$ is constructed with the aforesaid properties. Let $U^{n +1} = U^n \langle X_{c+1}, \ldots, X_{c + u} \rangle$ and $V^{n+1} = V^n \langle Y_{d+1}, \ldots, Y_{d+v}\rangle$, $\partial_1(X_{c+i}) = s_i$ for $1 \leq i \leq u$ and $\partial_2(Y_{d+j}) = t_j$ for $1 \leq j \leq v$. Then homology classes of cycles $s_i$ form a basis of $\Ho_n(U^n)$ and homology classes of cycles $t_j$ form a basis of $\Ho_n(V^n)$. Note that cycles $s_i \in \mathfrak{p} U^n$ and $t_j \in \mathfrak{q} V^n$. So we can view cycles $s_i$, $t_j$ as elements in $\tilde{U}^n$ and $\tilde{V}^n$ respectively. We have $\phi^n \circ \alpha^n (s_i) = s_i$, $\psi^n \circ \alpha^n (s_i) = 0$, $\psi^n \circ \beta^n(t_j) = t_j$ and $\phi^n \circ \beta^n(t_j) = 0$. It is now easy to see that homology classes of cycles $\alpha^n(s_i)$, $\beta^n(t_j)$,
$1 \leq i \leq u$, $1 \leq j \leq v$ are linearly independent in $\Ho_n(F^n R)$.
We choose cycles $z_k$, $1 \leq k \leq w$ in $\Zi_n(F^nR)$ such that homology classes of cycles $\alpha^n(s_i)$, $\beta^n(t_j)$, $z_k$, $1 \leq i \leq u$, $1 \leq j \leq v$, $1 \leq k \leq w$ form a basis of $\Ho_n(F^n R)$. We subtract a finite $R$-linear sum of $\alpha^n(s_i)$, $\beta^n(t_j)$ from $z_k$ if necessary to assume that $\phi^n(z_k)$ is in the boundary of $U^n$ and $\psi^n(z_k)$ is in the boundary of $V^n$. We choose $f_k \in U^n$, $g_k \in V^n$ of degree $n + 1$ such that $\partial_1(f_k) = \phi^n(z_k)$ and $\partial_2(g_k) = \psi^n(z_k)$. Since $p^n, q^n$ are surjective, we have $\tilde{f}_k \in \tilde{U}^n$, $\tilde{g}_k \in \tilde{V}^n$ such that $p^n(\tilde{f}_k) = f_k$ and $q^n(\tilde{g}_k) = g_k$.
We define $F^{n +1} R = F^n R \langle X_{c + i}, Y_{d + j}, Z_k : 1 \leq i \leq u, 1 \leq j \leq v, 1 \leq k \leq w \rangle$ and extend the differential $\partial^n$ from $F^n R$ to a differential $\partial^{n+1}$ on $F^{n+1} R$ by setting $\partial^{n+1}(X_{c + i}) = \alpha^n(s_i)$, $\partial^{n+1}(Y_{d + j}) = \beta^n(t_j)$ and $\partial^{n+1}(Z_j) = z_j$.
The natural augmentation $\epsilon^n$ extends to the augmentation $\epsilon^{n+1} : F^{n+1}R \twoheadrightarrow k$. Clearly $\I \Ho_i(F^{n+1}R) = 0$ for $1 \leq i \leq n$.
By constructions the inclusions $\alpha^n$, $\beta^n$ extend to inclusions $\alpha^{n + 1} : \tilde{U}^{n + 1} \hookrightarrow F^nR$ and $\beta^{n + 1} : \tilde{V}^{n + 1} \hookrightarrow F^nR$ of chain complexes respectively.
We only construct the map $\phi^{n+1}$. The construction of $\psi^{n+1}$ will follow similarly. Let $I^n = \ker \phi^n$. Then
\begin{align*}
\phi^n \partial^{n+1}(Z_k - \alpha^n( \tilde{f}_k)) &= \phi^n \partial^{n+1}(Z_k) - \phi^n \partial^{n} \alpha^n (\tilde{f}_k)\\
&= \phi^n (z_k) - \partial_1 \phi^n \alpha^n (\tilde{f}_k) = \phi^n (z_k) - \partial_1 p^n (\tilde{f}_k) = \phi^n (z_k) - \partial_1(f_k) = 0.
\end{align*}
So $\partial^{n+1}(Z_k - \alpha^n (\tilde{f}_k)) \in I^n$. We also have $\partial^{n+1}(Y_{d+j}) = \beta^n(t_j ) \subset I^n$.
Therefore, $I^{n + 1} = I^n + (Y_{d + j} ; 1 \leq j \leq v) + (Z_k - \alpha^n (\tilde{f}_k) ; 1 \leq k \leq w)$ is a DG$\Gamma$ ideal of $F^{n+1}R$.
Define $\phi^{n+1} : F^{n+1}R \rightarrow U^{n + 1}$ as the composition $ F^{n+1}R \rightarrow \frac{F^{n+1}R}{I^{n +1}} \rightarrow \frac{F^nR}{I^n}\langle X_{c+1}, \ldots, X_{c + u} \rangle \rightarrow U^{n +1}$. Here the first map is the quotient map, the second map is induced by $Y_{d +j} \mapsto 0$, $Z_k \mapsto \alpha^n (\tilde{f}_k)$ and the last one is induced by the isomorphism $\frac{F^nR}{I^n} \cong U^n$.
Clearly $\phi^{n+1} \circ \alpha^{n+1} = p^{n+1}$ and $\phi^{n+1} \circ \beta^{n+1} = 0$. Thus the induction step is complete.
Finally we define $R\langle X, Y, Z \rangle = \varprojlim F^nR$. The augmentations $\epsilon^n$, the inclusions $\alpha^n$, $\beta^n$ and surjections $\phi^n$, $\psi^n$ patch together to form augmentation map $\epsilon : R\langle X, Y, Z \rangle \twoheadrightarrow k$, the inclusions $\alpha : R\langle X \rangle \hookrightarrow R\langle X, Y, Z \rangle$, $\beta : R\langle Y \rangle \hookrightarrow R\langle X, Y, Z \rangle$ of chain maps, surjections $\phi : R\langle X, Y, Z \rangle \twoheadrightarrow S\langle X\rangle$, $\psi : R\langle X, Y, Z \rangle \twoheadrightarrow T\langle Y\rangle$ respectively with required properties.
\end{proof}
A surjective homomorphism $f : (R, \mathfrak{m}, k) \twoheadrightarrow (S, \mathfrak{n}, k)$ of local rings is called large if it induces an injective map $f^* : \Ext_S(k, k) \rightarrow \Ext_R(k, k)$ of Hopf algebras (\cite{lev4}). The homotopy Lie algebra of a local ring is the set of primitive elements of its ext algebra. A homorphism of Hopf algebras is injective if and only if it induces an injection on the sets of primitive elements \cite[Proposition 3.9]{gul}.
So $f$ is large if and only if the induced map $f^* : \pi(S) \rightarrow \pi(R)$ is an injective Lie algebra homomorphism.
A graded sub Lie algebra of a free graded Lie algebra on a positively graded vector space is free \cite[Proposition A.1.10]{lem}. A proof in the case of characteristic zero can also be found in the corollary following \cite[Proposition 21.4]{felix}. We use this fact in the next theorem which generalises a result \cite[Theorem 4.1]{les1} of Lescot.
\begin{theorem}\label{ggfl4}
Let $(S, \mathfrak{p}, k)$, $(T, \mathfrak{q}, k)$ be local rings and $R = S \times_k T$. Then $R$ is a generalised Golod ring of level $n$ if and only if both $S$, $T$ are so.
\end{theorem}
\begin{proof}
Assume that $R$ is a generalised Golod ring of level $n$, i.e. $\pi^{>n}(R)$ is a free Lie algebra. The projection maps $pr_1 : R \rightarrow S$, $pr_2 : R \rightarrow T$ are large homomorphisms \cite[Theorem 3.4]{moore}.
This means that $pr_1$, $pr_2$ induce injective Lie algebra homomorphisms $pr_1^* : \pi(S) \hookrightarrow \pi(R)$ and $pr_2^* : \pi(T) \hookrightarrow \pi(R)$. So both $\pi^{>n}(S)$, $\pi^{>n}(T)$ are free as they are subalgebras of the free Lie algebra $\pi^{>n}(R)$. Therefore, $S$, $T$ are generalised Golod rings of level $n$.
Now assume that both $S$, $T$ are generalised Golod rings of level $n$.
Let $S\langle X \rangle$, $T \langle Y \rangle$ be acyclic closures of $S$, $T$ respectively. Then both $S\langle X_i : \deg(X_i) \leq n \rangle$, $T\langle Y_i : \deg(Y_i) \leq n \rangle$ are Golod algebras. So by Lemma \ref{ggfl2}, $U = R\langle X_i, Y_j : \deg(X_i), \deg(Y_j) \leq n \rangle$ is a Golod DG$\Gamma$ algebra, i.e. $\pi(U)$ is a free Lie algebra. The acyclic closure of $R$ is of the form $R\langle X, Y, Z \rangle$ by Proposition \ref{ggfl3}. Set $V = R\langle X_i, Y_j, Z_k : \deg(X_i), \deg(Y_j), \deg(Z_k) \leq n \rangle$.
Consider the inclusion of DG$\Gamma$ algebras $ i : U \hookrightarrow V$. We have a induced map $i_* : \Tor^U(k, k) \rightarrow \Tor^V(k, k)$. Since $R\langle X, Y, Z \rangle$ is a semi-free extension of both $U$, $V$ and is a minimal resolution of $k$, we have $\Tor^U(k, k) = \Ho(k \otimes_U R\langle X, Y, Z \rangle) = k \otimes_U R\langle X, Y, Z \rangle$ and $\Tor^V(k, k) = \Ho(k \otimes_V R\langle X, Y, Z \rangle) = k \otimes_V R\langle X, Y, Z \rangle$. Therefore, $i_*$ is surjective. So $i_*$ is also a surjective map from the space of $\Gamma$-indecomposable elements of $\I \Tor^U(k, k)$ onto that of $\I \Tor^V(k, k)$. Taking $k$-vector space dual, we observe that $i$ induces an injective Lie algebra homomorphism $i^* : \pi(V) \hookrightarrow \pi(U)$.
Since $\pi(U)$ is a free Lie algebra, the Lie algebra $\pi(V)$ is also free. So $V$ is a Golod algebra. This implies that $R$ is a generalised Golod ring of level $n$.
\end{proof}
\begin{theorem}\label{ggc5}
Let $(R, \mathfrak{m}_R, k)$, $(S, \mathfrak{m}_S, k)$ be two Artinian Gorenstein local rings, $T = R \# S$ and $l \geq 2$. Then $T$ is a generalised Golod ring of level $l$ if and only if both $R$, $S$ are so.
\end{theorem}
\begin{proof}
We have $\frac{T}{\soc(T)} = \frac{R}{\soc(R)} \times_k \frac{S}{\soc(S)}$. So the result follows from Theorems \ref{ggc4}, \ref{ggfl4}.
\end{proof}
The following result generalises \cite[Theorem 3.8]{conca}.
\begin{theorem}\label{ggfl5}
Let $A$, $B$ be two local algebras over a field $k$. If both $A$, $B$ have the Backelin-Roos property, then so does the fibre product $A \times_k B$.
\end{theorem}
\begin{proof}
We may assume that both $A$, $B$ are complete with respect to maximal ideals. Let $A = \frac{k[|x_1, \ldots, x_m|]}{I_A}$ and $B = \frac{k[|y_1, \ldots, y_n|]}{I_B}$. Then $A \times_k B = \frac{k[|x_1, \ldots, x_m, y_1, \ldots, y_m |]}{(I_A + I_B + \underline{x}\underline{y})}$; $\underline{x} \underline{y}$ being the ideal generated by products $x_iy_j$, $1 \leq i \leq m$, $1 \leq j \leq n$. Since $A$, $B$ have the Backelin-Roos property, we have a regular sequence $f_1, \ldots, f_r$ in $I_A$ and a regular sequence $g_1, \ldots, g_s$ in $I_B$ such that $f : P \twoheadrightarrow A$ and $g : Q \twoheadrightarrow B$ are Golod homomorphisms from complete intersections $P = \frac{k[|x_1, \ldots, x_m|]}{(f_1, \ldots, f_r)}$ and $Q = \frac{k[|y_1, \ldots, y_n|]}{(g_1, \ldots, g_s)}$ respectively.
We claim that the surjective homomorphism $H : P \otimes_k Q \twoheadrightarrow A \times_k B$ is a Golod homomorphism.
Let $P \langle X \rangle $, $Q \langle Y \rangle$ be acyclic closures of $P$, $Q$ respectively. By the hypothesis, $A\langle X \rangle = A \otimes_P P\langle X \rangle$, $B\langle Y \rangle = B \otimes_Q Q\langle Y \rangle$ are Golod algebras. Now $P\langle X \rangle \otimes_k Q \langle Y \rangle$ is the acyclic closure of $P \otimes_k Q$ and by Lemma \ref{ggfl2}, $[P\langle X \rangle \otimes_k Q \langle Y \rangle] \otimes_{P \otimes_k Q} A \times_k B = (A \times_k B) \langle X \sqcup Y \rangle$ is also a Golod algebra. So $H$ is a Golod homomorphism and the result follows.
\end{proof}
\section{Applications}
\begin{theorem}\label{sbr3}
Let $(R, \mathfrak{m}, k)$ be an Artinian Gorenstein local ring such that $\edim(R) = n$ and $\mu(\mathfrak{m}^2) \leq 2$. Let $M$ be a finitely generated $R$-module. Let $\eta : (Q, \mathfrak{n}) \twoheadrightarrow (R, \mathfrak{m})$ be a minimal Cohen presentation and $I = \ker \eta \subset \mathfrak{n}^2$. Then the following hold.
\begin{enumerate}
\item
If $\edim(R) = 1$, then $P^R_k(t) = \frac{1}{1 - t}$ and $(1- t)P^R_M(t) \in {\NZQ Z}[t]$.
\item
If $\edim(R) \geq 2$, then for any $f \in I \setminus \mathfrak{n} I$, the induced map $Q/ (f) \twoheadrightarrow R$ is a Golod homomorphism. $P^R_k(t) = \frac{1}{1 -nt + t^2}$. Moreover, $(1+t)^n(1 - nt + t^2)P^R_M(t) \in {\NZQ Z}[t]$
\end{enumerate}
\end{theorem}
\begin{proof}
If $n = 1$, $R$ is a Golod ring and $P^{R}_k(t) = \frac{1}{1 - t}$. The ring $R$ is a quotient of a DVR. Using the structure theorem of modules over PID, the module $M$ is a direct summand of quotients of $R$. So $P^R_M(t) = \mu(M) + \mu(\syz_1^{R}(M))\frac{t}{1 - t}$. This implies that $(1- t)P^R_M(t) \in {\NZQ Z}[t]$. So (1) follows.
The first part of (2) follows from Theorems \ref{ds4}, \ref{sbr2}. The Poincar\'e series of $k$ is computed using Theorems \ref{prl6.5} and \ref{ds4}. The last assertion on Poincar\'e series of $M$ follows from the result of Levin \cite[Proposition 5.18]{av4}.
\end{proof}
Stretched Cohen-Macaulay rings are shown to be good in a recent article \cite[Corollary 5.6]{amenda}. We provide a different way to see this.
Let $(R, \mathfrak{m}, k)$ be a stretched Artinian ring such that $\edim(R) = n$, $\mathfrak{m} = (x_1, \ldots, x_n)$, $\lo(R) = s$ and $\dim_k (0 : \mathfrak{m}) = r$. If $\soc(R) \subset \mathfrak{m}^2$, then $\soc(R) = \mathfrak{m}^s$ and $R$ is Gorenstein. Otherwise assume that $x_1 \in \soc(R) \setminus \mathfrak{m}^2$. We see $R = R/(x_1) \times_k R/(x_2, \ldots, x_n)$. Note that $R/(x_1)$ is also a stretched Artinian ring. Therefore, after a finite number of steps, we may write $R = S \times_k T$ where $(S, \mathfrak{m}_S)$ is a stretched Artinian Gorenstein ring and $(T, \mathfrak{m}_T)$ is a local ring with $\mathfrak{m}_T^2 = 0$. Clearly $r = 1 + \edim(T)$.
To prove both the corollaries below we first reduce to the Artinian case by standard arguments. In the stretched case, we use the fibre product decomposition, Theorems \ref{prl10.6} and \ref{sbr3}. The almost stretched case is immediate.
\begin{corollary}\label{sbr4}
Let $(R, \mathfrak{m}, k)$ be a $d$-dimensional stretched Cohen-Macaulay local ring and $M$ be an $R$-module. Let $n = \dim_k \mathfrak{m}/ \mathfrak{m}^2$ and $r = \dim_k \Ext_R^d(k, R)$ denote the type of $R$. Then the following hold.
\begin{enumerate}
\item
If $r = n - d$, then $R$ is a Golod ring, $P^R_k(t) = \frac{(1+t)^d}{1 - (n-d)t}$ and $(1 - (n-d)t) P^R_M(t) \in {\NZQ Z}[t]$.
\item
If $r \neq n - d$, then $P^R_k(t) = \frac{(1+t)^d}{1 - (n-d)t + t^2}$ and $(1 + t)^{n-d}(1 - (n-d)t + t^2) P^R_M(t) \in {\NZQ Z}[t]$.
\end{enumerate}
\end{corollary}
\begin{corollary}\label{sbr4.6}
Let $(R, \mathfrak{m}, k)$ be a $d$-dimensional almost stretched Gorenstein local ring and $M$ be an $R$-module. Let $n = \dim_k \mathfrak{m}/ \mathfrak{m}^2$. Then $P^R_k(t) = \frac{(1+t)^d}{1 - (n-d)t + t^2}$ and $(1+t)^{n-d}(1 - (n-d)t + t^2) P^R_M(t) \in {\NZQ Z}[t]$.
\end{corollary}
The following is due to Sega \cite[Proposition 1.5]{sega1}.
\begin{proposition}\label{ar1}
Let $R$ be a local ring such that there is a $d_R(t) \in {\NZQ Z}[t]$ satisfying $d_R(t) P^R_M(t) \in {\NZQ Z}[t]$ for each finitely generated $R$-module $M$. Let $d_R(t) = p(t)q(t)r(t)$ where $p(t)$ is $1$ or irreducible; $q(t)$ has non-negative coefficients; $r(t)$ is $1$ or irreducible and has no positive real root among its complex roots of minimal absolute value.
Then the following hold for each pair of $R$-modules $M$, $N$.
\begin{enumerate}
\item
If $\Tor^R_i(M, N) = 0$ for $i \gg 0$, then either $M$ or $N$ has finite projective dimension.
\item
If $\Ext^i_R(M, N) = 0$ for $i \gg 0$, then either $M$ has a finite projective dimension or $N$ has a finite injective dimension.
\end{enumerate}
\end{proposition}
Araya and Yoshino \cite[4.2]{araya} proved that if M has finite CI-dimension, then the projective dimension of $M$ is same as $\sup \{i : \Ext^i(M, M) \neq 0\}$. Therefore, the following corollary follows from Proposition \ref{ar1}.
\begin{corollary}\label{ar2}
Let $R$ be either a stretched Cohen-Macaulay ring or an almost stretched Gorenstein ring. Then $R$ satisfies Auslander-Reiten conjecture, i.e. $\Ext^{\geq 1}(M, M \oplus R) = 0$ for a finitely generated $R$ module $M$ implies that $M$ is free.
\end{corollary}
\begin{theorem}\label{sbr4.5}
Let $(R, \mathfrak{m})$ be an Artinian Gorenstein local ring of embedding dimension $n \geq 2$. Let $K^R$ denote the Koszul complex on a minimal set of generators of $\mathfrak{m}$ and $C$ denote the quotient of $K^R$ defined by $C_i = K^R_i$ for $0 \leq i \leq n - 2$, $C_{n-1} = \frac{K^R_{n -1}}{\Bo_{n -1}(K^R)}$ and $C_n = 0$. Assume that $C$ is a Golod DG algebra with natural augmentation. Then $R/ \soc(R)$ is a Golod ring and $R$ satisfies assertions (1), (2), (3) of Theorem \ref{sbr2}.
\end{theorem}
\begin{proof}
Let
$q : K^R \twoheadrightarrow C$ be the quotient map and $\soc(R) = (s)$.
Let $G : \frac{K^R}{\mathfrak{m} K^R} \rightarrow K^R$, $H : \frac{C}{\mathfrak{m} C} \rightarrow C$ denote chain maps induced by multiplications by $s$ on $K^R$, $C$ respectively.
We have $q \circ G = H \circ \bar{q}$ where $\bar{q} : \frac{K^R}{\mathfrak{m} K^R} \twoheadrightarrow \frac{C}{\mathfrak{m} C}$ is the map induced by $q$.
Let $\cn(G) = K^R \oplus \frac{K^R}{\mathfrak{m} K^R}[-1]$, $\cn(H) = C \oplus \frac{C}{\mathfrak{m} C}[-1]$
be the cones of $G$, $H$ respectively.
Define $\alpha : \cn(G) \twoheadrightarrow \cn(H)$ by $\alpha(k, \bar{l}) = (q(k), \bar{q}(\bar{l}))$. Both $\cn(G)$, $\cn(H)$ have DG algebra structure and $\alpha$ is a surjective DG algebra homomorphism. The kernel of $\alpha$ is the complex $0 \rightarrow \frac{K^R_n}{\mathfrak{m} K^R_n} \xrightarrow{\cdot s} K^R_n \xrightarrow{\partial} \Bo_{n-1}(K^R) \rightarrow 0$ which is exact. Therefore, $\alpha$ is a quasi-isomorphism of DG algebras. By Theorem \ref{ggc1}, $\cn(G)$ is quasi isomorphic to $K^{R/ \soc(R)}$. Therefore, to show that $R/ \soc(R)$ is a Golod ring it is enough to prove that $\cn(H)$ is a Golod algebra.
Since $C$ is a Golod algebra, there is a $k$-basis $\mathfrak{b}_C = \{h_{\lambda}\}_{\lambda \in \Lambda}$ of $\Ho_{\geq 1}(C)$ and a function (trivial Massey operation) $\mu : \sqcup_{i = 1}^{\infty} \mathfrak{b}_C^i \rightarrow C$ such that $\mu(h_\lambda) \in \Zi_{\geq 1}(C)$ with $cls(\mu(h_\lambda))= h_\lambda$ and
$ \partial \mu(h_{\lambda_1}, \ldots ,h_{\lambda_p}) = \sum_{j = 1}^{p-1}\overline{\mu(h_{\lambda_1}, \ldots ,h_{\lambda_j})}\mu(h_{\lambda_{j + 1}}, \ldots ,h_{\lambda_p})$. By (2) of \cite[Lemma 4.1.6]{av}, $\{x \in C : \partial(x) \in \mathfrak{m}^2 C\} \subset \mathfrak{m} C$. Since $\mu(h_\lambda) \in \mathfrak{m} C$, by induction on $p$ we conclude that $\mu(h_{\lambda_1}, \ldots ,h_{\lambda_p}) \in \mathfrak{m} C$.
By (1) of Lemma \ref{sbr1}, $sC \subset (0 : \mathfrak{m}^2)\Bo(C)$. Let $z \in \Zi(C)$ be such that $(z, 0) = \partial(y_1, \bar{y}_2)$ for $(y_1, \bar{y}_2) \in \cn(H)$. Then $z = \partial(y_1) + s y_2 \in \Bo(C)$. So the inclusion $C \hookrightarrow \cn(H)$ induces an injective map $\Ho(C) \hookrightarrow \Ho(\cn(H))$. By abuse of notation we write $(c, 0)$ as $c$. So $\mathfrak{b}_C = \{h_{\lambda}\}_{\lambda \in \Lambda}$ is a linearly independent set in $\Ho_{\geq 1}(\cn(H))$ and $ \mu(h_{\lambda_1}, \ldots ,h_{\lambda_p}) \in \mathfrak{m} \cn(H)$ satisfy properties above.
We extend $\mathfrak{b}_C$ to a basis $\mathfrak{b}_{\cn(H)} = \{h_{\lambda}\}_{ \lambda \in \Lambda \sqcup \Lambda'}$ of $\Ho_{\geq 1}(\cn(H))$. Let $h_{\lambda'}$, $\lambda' \in \Lambda'$ be the homology class of $(c_{\lambda'}, \bar{d}_{\lambda'}) \in \Zi_{\geq 1}(\cn(H))$. Now $\partial(c_{\lambda'}, \bar{d}_{\lambda'}) = 0$ implies that $\partial(c_{\lambda'}) + s d_{\lambda'} = 0$. We have $s d_{\lambda'} = \partial(e_{\lambda'})$ for some $e_{\lambda'} \in (0 : \mathfrak{m}^2)C$. We write $(c_{\lambda'}, \bar{d}_{\lambda'}) = (c_{\lambda'} + e_{\lambda'}, 0) + ( - e_{\lambda'}, \bar{d}_{\lambda'})$. Note that $c_{\lambda'} + e_{\lambda'} \in \Zi(C)$. Therefore, after subtracting suitable $R$-linear combinations of $h_{\lambda}$, $\lambda \in \Lambda$ from each $h_{\lambda'}$, $\lambda' \in \Lambda'$ if necessary, we may assume that $h_{\lambda'}$ is a homology class of some cycle in $(0 : \mathfrak{m}^2)C \oplus \frac{C}{\mathfrak{m} C}$.
We define $\mu(h_{\lambda'})$, $\lambda' \in \Lambda'$ to be an element in $(0 : \mathfrak{m}^2)C \oplus \frac{C}{\mathfrak{m} C}$ whose homology class is $h_{\lambda'}$. We extend $\mu$ from $\mathfrak{b}_C^i$ to $\mathfrak{b}_{\cn(H)}^i$, $i > 1$ such that $\mu : \mathfrak{b}_{\cn(H)}^i \setminus \mathfrak{b}_C^i \rightarrow (0 : \mathfrak{m}^2)C$, $i > 1$ by induction on $i$.
Note that $\overline{\mu(h_{\lambda})} \mu(h_{\lambda'}) \in sC \in (0 : \mathfrak{m}^2) \Bo(C)$ for $(\lambda, \lambda') \not \in \Lambda \times \Lambda$. We choose $\mu(h_{\lambda}, h_{\lambda'}) \in (0 : \mathfrak{m}^2) C$ such that $\partial(\mu(h_{\lambda}, h_{\lambda'})) = \overline{\mu(h_{\lambda})} \mu(h_{\lambda'})$.
Now assume that $\mu(h_{\delta_1}, \ldots, h_{\delta_i}) \in (0 : \mathfrak{m}^2)C$, $(h_{\delta_1}, \ldots, h_{\delta_i}) \in \mathfrak{b}_{\cn(H)}^i \setminus \mathfrak{b}_C^i$ satisfying desired relations are constructed for all $i \leq p$. We choose $(h_{\delta_1}, \ldots , h_ {\delta_{p+1}}) \in \mathfrak{b}_{\cn(H)}^{p+1} \setminus \mathfrak{b}_C^{p+1}$ and observe that $\sum_{j = 1}^{p}\overline{\mu(h_{\delta_1}, \ldots ,h_{\delta_j})}\mu(h_{\delta_{j + 1}}, \ldots ,h_{\delta_{p+1}})$ is an element in $sC$. So we can choose $\mu(h_{\delta_1}, \ldots ,h_{\delta_{p+1}}) \in (0 : \mathfrak{m}^2)C$ such that $\partial(\mu(h_{\delta_1}, \ldots ,h_{\delta_{p+1}})) = \sum_{j = 1}^{p}\overline{\mu(h_{\delta_1}, \ldots ,h_{\delta_j})}\mu(h_{\delta_{j + 1}}, \ldots ,h_{\delta_{p+1}})$. Thus by induction $\mu$ extends to a trivial Massey operation on $\mathfrak{b}_{\cn(H)}$. Therefore, $\cn(H)$ is a Golod algebra and the result follows.
\end{proof}
Poincar\'e series of Modules over compressed Gorensteing local rings were studied by Rossi and \c Sega in \cite{rossi}. Under the hypothesis of Corollary \ref{sbr5} below they proved that if $\eta : Q \twoheadrightarrow R$ is a minimal Cohen presentation, $I = \ker \eta$, $f \in I \setminus \mathfrak{n}^{t+1}$, $t = \max\{ i : I \subset \mathfrak{n}^i\}$, then $Q/ (f) \twoheadrightarrow R$ is a Golod homomorphism \cite[Theorem 5.1]{rossi}. They also showed that $R/ \soc(R)$ is a Golod ring \cite[Proposition 6.3]{rossi}. We use the ideas developed by them and our Theorem \ref{sbr4.5} to produce a simple and short proof of their main result \cite[Theorem 5.1]{rossi}. Note that our result is stronger. Our theorem below shows that any choice of $f \in I \setminus \mathfrak{n} I$ is enough.
\begin{corollary}\label{sbr5}
Let $(R, \mathfrak{m}, k)$ be a compressed Gorenstein local ring such that $\edim(R) = n \geq 2$ and $\lo(R) = s$, $s \geq 2, s \neq 3$. Then $R/ \soc(R)$ is a Golod ring and $R$ satisfies assertions (1), (2), (3) of Theorem \ref{sbr2}.
\end{corollary}
\begin{proof}
We follow notations as set in Theorems \ref{sbr2} and \ref{sbr4.5}. Let $t = \max\{ i : I \subset \mathfrak{n}^t\}$. From the structure theorem of compressed Gorenstein local rings, it follows that $t = \ceil{\frac{s+1}{2}}$; the least integer not less than $\frac{s+1}{2}$ \cite[Proposition 4.2]{rossi}.
By \cite[Lemma 1.4]{rossi}, the map $\Ho_{\geq 1}(R/ \mathfrak{m}^t \otimes_Q K^Q) \rightarrow \Ho_{\geq 1}(R/ \mathfrak{m}^{t-1} \otimes_Q K^Q)$ induced by the surjection $R/ \mathfrak{m}^t \twoheadrightarrow R/ \mathfrak{m}^{t-1}$ is a zero map. This implies that
$\Zi_{\geq 1}(K^R) \subset \Bo_{\geq 1}(K^R) + \mathfrak{m}^{t-1}K_{\geq 1}^R$ and therefore $\Zi_{\geq 1}(C) \subset \Bo_{\geq 1}(C) + \mathfrak{m}^{t-1}C$.
It is proved in \cite[Lemma 4.4]{rossi} that the map $\psi: \Ho_{<n}(\mathfrak{m}^{r+1}K^R) \rightarrow \Ho_{<n}(\mathfrak{m}^{r}K^R)$ induced by the inclusion $\mathfrak{m}^{r+1} \hookrightarrow \mathfrak{m}^r$ is zero for $r = s + 1 - t$. Since, $s \geq 2, s \neq 3$, we have $t - 1 \leq r + 1 \leq r \leq 2(t-1)$. This implies that the map $\Ho_{<n}(\mathfrak{m}^{2(t-1)}K^R) \rightarrow \Ho_{<n}(\mathfrak{m}^{t-1}K^R)$ is also zero since it factors through $\psi$. Therefore, we have $\Zi_{< n}(\mathfrak{m}^{2t - 2}K^R) \subset \Bo(\mathfrak{m}^{t-1}K^R)$ which implies $\Zi(\mathfrak{m}^{2t - 2}C) \subset \Bo(\mathfrak{m}^{t-1}C)$. It is worth pointing out that both the lemmas mentioned are independent of all other results in \cite{rossi}.
Thus we find a basis of $\Ho_{\geq 1}(C)$ represented by cycles in $\mathfrak{m}^{t-1}C$ such that the product of any two representatives is in the boundary of $\mathfrak{m}^{t-1}C$. So we can define inductively a trivial Massey operation on $C$. Therefore, $C$ is a Golod DG algebra and the result follows from Theorem \ref{sbr4.5}.
\end{proof}
Poincar\'e series of modules over a Gorenstein ring $R$ with $\edim(R) - \dim(R) \leq 3$ are known to be rational (see \cite[Satz 9]{wiebe}, \cite[Theorem 6.4]{av4}). We prove the following.
\begin{corollary}\label{sbr6}
Let $(R, \mathfrak{m})$ be a Gorenstein local ring but not a complete intersection. Assume that $\edim(R) - \dim(R) \leq 3$. Let $\eta : (Q, \mathfrak{n}) \twoheadrightarrow (\hat{R}, \hat{\mathfrak{m}})$ be a minimal Cohen presentation and $I = \ker \eta$. Then for any $f \in I \setminus \mathfrak{n} I$, the induced map $Q/(f) \twoheadrightarrow \hat{R}$ is a Golod homomorphism. If $R$ is Artinian, then $R/ \soc(R)$ is a Golod ring and assertions of Theorem \ref{sbr2} hold.
Modules over a Gorenstein ring $R$ with $\edim(R) - \dim(R) \leq 3$, have rational Poincar\'e series.
\end{corollary}
\begin{proof}
Let $\mathfrak{m}$ be minimally generated by $x_1, \ldots, x_n$ such that $x_1, \ldots, x_d$, $d = \dim(R)$ form an $R$-sequence. Now $S = R/ (x_1, \ldots, x_d)$ is an Artinian Gorenstein local ring and $\edim (S) \leq 3$. The Koszul complexes $K^R$ and $K^S$ are quasi-isomorphic \cite[Lemma 4.1.6]{av}. So the adjunction of a degree two variable to kill a degree one cycle in $K^R$ yields a Golod algebra if and only if the same holds true for $K^S$. Therefore, we may assume that $R$ is an Artinian Gorenstein local ring.
We construct $C$ as in Theorem \ref{sbr4.5}. By \cite[Satz 7]{wiebe}, $\Ho_1(K^R)^2 = 0$. So $\Ho_1(C)^2 = 0$. Now $C$ is a DG algebra of length $2$. So any basis of $\Ho_{\geq1}(C)$ admits a trivial Massey operation and the result follows by Theorem \ref{sbr4.5}. The last assertion follows since modules over complete intersections have rational Poincar\'e series \cite[Corollary 4.2]{gul1}.
\end{proof}
To prove the following Theorem, we use the fact that Gorenstein rings of co-depth at most $4$ have the Backelin-Roos property and therefore are generalised Golod rings of level 2 \cite[Theorem 6.4]{av4}.
\begin{theorem}\label{ggc6}
Let $(R, \mathfrak{m})$ be a Gorenstein local ring. Then $R$ is a generalised Golod ring of level $2$, and therefore good in the following cases.
\begin{enumerate}
\item
$\mathfrak{m}^4 = 0$, $\mu(\mathfrak{m}^2) \leq 4$.
\item
The multiplicity of $R$ is at most $11$.
\end{enumerate}
\end{theorem}
\begin{proof}
First we prove (1). We assume that $\lo(R) \geq 3$ because otherwise the result follows from Theorem \ref{sbr3}. It is enough to assume that the ring is indecomposable as a connected sum because of Theorem \ref{ggc5}. If $\mu(\mathfrak{m}) \geq 5$, then $l(0 : \mathfrak{m}^2) = l(R/ \mathfrak{m}^2) = 1 + l(\mathfrak{m}/ \mathfrak{m}^2) \geq 6$ and $l(\mathfrak{m}^2) = l(\mathfrak{m}^2/ \mathfrak{m}^3) + l(\mathfrak{m}^3) \leq 5$. So $(0 : \mathfrak{m}^2) \not\subset \mathfrak{m}^2$, a contradiction due to Theorem \ref{ac3}.
Therefore, $\mu(\mathfrak{m}) \leq 4$ and the result follows from \cite[Theorem 6.4]{av4}.
Now we prove (2). If we go modulo a nonzero divisor in $\mathfrak{m} \setminus \mathfrak{m}^2$, then the acyclic closure of $R$ does not change upto quasi-isomorphism. So we may assume by a standard argument that $R$ is Artinian. If we can show that either $\edim(R) \leq 4$ or $\lo(R) \leq 2$ or else $R = S \# T$ for Gorenstein local rings $(S, \mathfrak{p})$, $(T, \mathfrak{q})$ such that $\edim(S) \leq 4$ and $\lo(T) = 2$, then we are done in view of \cite[Theorem 6.4]{av4}, Theorems \ref{ggc5}, \ref{sbr3}.
If possible assume that there is an Artinian Gorenstein local ring $R$ of multiplicity at most $11$ such that $R$ is not of any of the types mentioned above. We choose $R$ of the least possible multiplicity. Then in view of Theorem $\ref{ac3}$ we must assume that $(0 : \mathfrak{m}^2) \subset \mathfrak{m}^2$. This implies that $l(R/ \mathfrak{m}^2) = l(0 : \mathfrak{m}^2) \leq l(\mathfrak{m}^2) = l(R) - l(R/ \mathfrak{m}^2)$. So $l(R/ \mathfrak{m}^2) \leq \frac{1}{2}l(R) = \frac{11}{2}$ which gives that $\edim(R) \leq l(R/ \mathfrak{m}^2) - 1 \leq 4$, a contradiction. So the result follows.
\end{proof}
We conclude this section with the following question.
\begin{question}
What is the least possible integer $n$ such that there is a Gorenstein bad local ring of multiplicity $n$?
\end{question}
\begin{acknowledgement}
Work on this article started as a result of discussions with Liana M. \c Sega during the conference in honour of Craig Huenke held in Ann Arbor in July 2016. I would like to thank her for clarifications on her results about Auslander-Reiten conjecture.
I owe many thanks to my mentor Matteo Varbaro for his inspiration and encouragement throughout. This research is supported by the INdAM-COFUND-2012 Fellowship cofounded by Marie Curie actions, Italy.
\end{acknowledgement}
|
\section*{Extended abstract}
Machine learning plays a role in many aspects of modern IR systems, and deep learning is applied in all of them.
The fast pace of modern-day research has given rise to many different approaches for many different IR problems.
The amount of information available can be overwhelming both for junior students and for experienced researchers looking for new research topics and directions.
Additionally, it is interesting to see what key insights into IR problems the new technologies are able to give us.
The aim of this full-day tutorial is to give a clear overview of current tried-and-trusted neural methods in IR and how they benefit IR research.
It covers key architectures, as well as the most promising future directions.
\input{2.1.0_motivation}
\input{2.2.0_objectives}
\input{2.4.0_schedule}
\input{2.5.0_materials}
\section{Motivation}
Prompted by the advances of deep learning in computer vision research, neural networks have resurfaced as a popular machine learning paradigm in many other directions of research as well, including information retrieval.
Recent years have seen neural networks being applied to all key parts of the typical modern IR pipeline, such core ranking algorithms \cite{szummer2011semi,huang2013learning,Mitra:2016:www}, click models \citep{Borisov:2016:click,Borisov:2016:time}, knowledge graphs \cite{bordes2011learning,lin2015learning}, text similarity \cite{kenter2016siamesecbow,severyn2015learning}, entity retrieval \citep{VanGysel:2016:www,VanGysel:2016:cikm}, language modeling \cite{bengio2003neural}, question answering \cite{weston2015towards,hewlett2016wikireading}, and dialogue systems \cite{li2016deep,vinyals2015neural}.
A key advantage that sets neural networks apart from many learning strategies employed earlier, is their ability to work from raw input data.
E.g., when given enough training data, well-designed networks can become feature extractors themselves, e.g., incorporating basic input characteristics such as term frequency (tf) and term saliency (idf)---that used to be pre-calculated offline---in their initial layers.
Where designing features used to be a crucial aspect and contribution of newly proposed IR approaches, the focus has shifted to designing network architectures instead.
As a consequence, many different architectures and paradigms have been proposed, such as auto-encoders, recursive networks, recurrent networks, convolutional networks, various embedding methods, deep reinforcement and deep q-learning, and, more recently, generative adversarial networks, of which most have been applied in IR settings.
The aim of the \textit{neural networks for IR (NN4IR)} tutorial is to provide a clear overview of the main network architectures currently applied in IR and to show explicitly how they relate to previous work.
The tutorial covers methods applied in industry and academia, with in-depth insights into the underlying theory, core IR tasks, applicability, key assets and handicaps, scalability concerns and practical tips and tricks.
We expect the tutorial to be useful both for academic and industrial researchers and practitioners who either want to develop new neural models, use them in their own research in other areas or apply the models described here to improve actual IR systems.
\section{Objectives}
The material in the tutorial covers a broad range of IR applications.
It is structured as follows:
\input{2.2.1_preliminaries}
\input{2.2.2_semantic-matching-I}
\input{2.2.3_semantic-matching-II}
\input{2.2.4_ranking}
\input{2.2.5_user-interaction}
\input{2.2.6_generating}
\input{2.2.7_outlook}
\medskip\noindent%
Summing up, the \textbf{objectives of the NN4IR tutorial} are as follows:
\begin{itemize}[leftmargin=0.13in]
\item Give an extensive overview of neural network architectures currently employed in IR, both in academia and industry.
\item Provide theoretical background, thereby equipping participants with the necessary means to form intuitions about various neural methods and their applicability.
\item Identify the IR lessons learned by employing neural methods.
\item Give practical tips and tricks, regarding network design, optimization, hyperparameter values, based on industry best practice.
\item Discuss promising future research directions.
\end{itemize}
\noindent%
The \textbf{target audience} consists of researchers and developers in information retrieval who are interested in gaining an in-depth understanding of neural models across a wide range of IR problems.
The tutorial will be useful as an overview for anyone new to the deep learning field as well as for practitioners seeking concrete recipes.
The tutorial aims to provide a map of the increasingly rich landscape of neural models in IR.
By the end of the tutorial, attendees will be familiar with the main architectures of neural networks as applied in IR and they will have informed intuitions of their key properties and of the insights they bring into core IR problems.
We aim to provide an overview of the main directions currently employed, together with a clear understanding of the underlying theory and insights, illustrated with examples.
\section{Format and detailed schedule}
Table~\ref{table:schedule} gives an overview of the time schedule of the tutorial.
Below we provide the details for each session.
\begin{table}[t]
\centering
\caption{Time schedule for NN4IR tutorial}
\label{table:schedule}
\begin{tabular}{l l}
\toprule
Preliminaries & 60 minutes \\
Semantic Matching I & 60 minutes \\
Semantic Matching II & 60 minutes \\
\midrule
\multicolumn{2}{c}{\textit{lunch break}} \\
\midrule
Learning to Rank & 45 minutes \\
Modeling User Behavior & 45 minutes \\
Generating Responses & 45 minutes \\
Outlook & 30 minutes \\
Wrap up & 15 minutes \\
\bottomrule
\end{tabular}
\vspace*{-\baselineskip}
\end{table}
We bring a team team of six lecturers, all with their specific areas of specialization. Each session will have two expert lecturers who will together present the session.
The initials below refer to the lecturers for this tutorial.
\input{2.4.1_preliminaries}
\input{2.4.2_semantic-matching-I}
\input{2.4.3_semantic-matching-II}
\input{2.4.4_ranking}
\input{2.4.5_user-interaction}
\input{2.4.6_generating}
\input{2.4.7_outlook}
\input{2.4.8_wrap-up}
\section{Type of support materials to be supplied to attendees}
\begin{description}
\item[Slides] Slides will be made publicly available on \url{http://nn4ir.com}.
\item[Bibliography] An annotated compilation of references will list all work discussed in the tutorial and should provide a good basis for further study.
\item[Code] Apart from the various open source neural toolkits (Tensorflow, Theano, Torch) many of the methods presented come with implementations released under an open source license. These will be discussed as part of the presentation of the models and algorithms. We provide a list pointers to available code bases.
\end{description}
|
\section*{Introduction}
A fabrication of a controllable source of a spin polarized current that operates without a magnetic
field is one of the most important challenges of semiconductor spintronics\cite{Fabian2007}. Among
spin filters proposed over the years, including these based on carbon nanotubes\cite{Hauptmann2008},
quantum dots\cite{Folk2003}, Y-shaped nanostructures\cite{Wojcik2015,Tsai2013} or resonant tunneling
diodes\cite{Wojcik2012,Wojcik2013,Slobodsky2003}, recently, the special attention is paid to the
quantum point contacts (QPC) with the spin-orbit (SO)
interaction\cite{Debray2009,Das2011,Das2012,Bhandari2012,Bhandari2013,Wan2009,Kohda2012,Nowak2013,Ngo2010}.
In such nanostructures, the spin filtering effect results from the
interplay between the SO coupling\cite{Rashba1984,Desselhaus1955} and the quantum confinement.
Recent papers\cite{Debray2009,Das2011,Das2012,Bhandari2012,Bhandari2013,Wan2009,
Kohda2012} have reported the experimental evidence of the spin filtering in QPCs, which manifests
itself as the plateau of conductance at $0.5 G_0$ $(G_0=2 e^2 /h)$ measured in the absence of the
external magnetic field. The $0.5G_0$ plateau has been explained as resulting from the combination
of the three effects\cite{Ngo2010,Bhandari2013,Wan2009}: an asymmetric lateral confinement, a
lateral Rashba SO interaction, and an electron - electron interaction. More precisely, the asymmetry
in a lateral confinement, induced by the different voltages applied to the side electrodes of QPC,
is a source of a lateral electric field. Due to the SO interaction, this electric field, in the
electron's rest frame, is seen as an effective magnetic field, which initializes an imbalance
between the spin-up and spin-down electrons. As shown by Ngo et al.\cite{Ngo2010}
this so-called lateral Rashba SO interaction leads to the low spin
polarization of the current not exceeding 6 \%, and therefore its presence did not explain the
$0.5G_0$ plateau observed in the experiments. The full explanation has been given by the further
studies, which have shown that the predicted weak spin filter effect\cite{Ngo2010} can be
strengthened by the electron-electron interaction leading to the nearly 100~\% spin polarization in
the regime of the single-mode transport\cite{Bhandari2013}. QPCs with the SO interaction have been
successfully used as the spin injector and detector in the recent experimental realization of the
spin transistor\cite{Chuang2015, Alomar2016}, in which about $10^5$ times greater conductance
oscillations have been observed as compared to the conventional spin-field effect transistor based
on ferromagnets\cite{Koo2009,Wojcik2014}.
The experimental realizations of spin filters based on QPCs, reported so far, are based on a
two-dimensional electron gas (2DEG) confined in the narrow quantum well at the
AlGaAs/GaAs or InAs/InAlAs interface\cite{Das2011,Das2012}, in which the
electrons occupy the lowest-energy state ("single occupancy"). However, recently, wider or coupled
quantum wells with two populated subbands ("double occupancy") have attracted a growing interest
of both experimentalists\cite{Fisher2005,Fisher2006,Bentmann2012,Hernandez2013,Hu1999,Akabori2012}
and
theoreticians\cite{Calsaverini2008,Bernardes2007,Fu2015,Chwiej2016PE,Chwiej2016PB,
Chwiej2016PRB}. In this case, we deal with the vertically coupled nanowires, in which the coupling
strength is determined by the wave function overlap between the ground and first excited states.
The additional orbital degree of freedom in the bilayer nanowires leads to
interesting physical effects such as inter-subband induced band
anticrossing and spin mixing\cite{Bentmann2012}.
The SO interaction in quantum well with double occupancy has been studied by Bernardes et al\cite{Bernardes2007}.
The inter-subband induced SO interaction has been found which
results from the coupling between states with opposite parity. It can give raise to intriguing
physical phenomena, e.g. unusual
Zitterbewegung\cite{Bernardes2007} or intrinsic spin Hall effect in symmetric quantum
well\cite{Hernandez2013}. The influence of the inter-subband SO interaction in
bilayer nanowire on the spin transistor action has been studied in our
recent paper\cite{Wojcik2016_arxiv}. We have shown that the resonant behavior of spin-orbit coupling
constants obtained for zero gate voltage leads to the spin transistor operation, in which
the on/off transition should be realized in the narrow voltage range.
In the present paper, we demonstrate the novel mechanism of spin filtering based on the
lateral inter-subband spin-orbit coupling in the bilayer nanowire with QPC. We find
that for the non-zero inter-subband coupling induced by the lateral Rashba SO interaction, the spin
polarization of the current flowing through the QPC is almost full. We analyze the conditions, under
which this polarization takes place. The observed spin filtering effect is explained as the joint
effect of the Landau-Zener inter-subband transitions caused by the hybridization of states with
opposite spins and the quantum confinement in the QPC region. Our results provide a new
mechanism to implement spin-polarized electron sources in the realistic bilayer nanowires which
can be built from the double quantum well or wide quantum well structure.
\section*{Theoretical model}
\subsection*{Model of the nanostructure}
\label{sec2}
We consider the bilayer nanowire consisting of two coupled conducting channels of width $W$ and
length $L$ (Fig.~\ref{fig1}). Both ends of the nanowire are connected to the reflectionless, ideal
leads denoted as IN and OUT. In the middle of the nanowire, the QPC is located as schematically
presented
in Fig.~\ref{fig1}(a).
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.5, angle=0]{fig1.pdf}
\caption{
(a) Schematic of the bilayer nanowire with the QPC. In the presence of the lateral electric field
$F_y$, the unpolarized current injected from the contact IN, after passing through
the QPC, is almost fully spin polarized. (b) Cross section of the exemplary realization of the
bilayer nanowire based on the Al$_{0.48}$In$_{0.52}$As/Ga$_{0.47}$In$_{0.53}$As double quantum
wells.}
\label{fig1}
\end{center}
\end{figure}
In recent experiments\cite{Fisher2005,
Fisher2006,Bentmann2012,Hernandez2013,Hu1999,Akabori2012}, vertically stacked and coupled nanowires
with QPC's are prepared from 2DEG bilayer systems built from double quantum wells or a wide
quantum well in which the weak Coulomb repulsion gives rise to the "soft" barrier in the middle of
the quantum well.
Figure~\ref{fig1}(b) presents the cross-section of the exemplary double quantum well
heterostructure, which consists of two
Al$_{0.48}$In$_{0.52}$As/Ga$_{0.47}$In$_{0.53}$As quantum wells with a central
Al$_{0.3}$In$_{0.7}$As barrier with width $w_b$, which controls the coupling between the
conduction electron states in the quantum wells. For the sufficiently high central barrier,
the quantum wells are separated and the electron wave functions are localized in one of
the quantum wells. For the symmetric structure each of the states is fourfold degenerate whereby the
twofold degeneracy results from the spin states and twofold degeneracy is related to the
geometric symmetry. If the height of the barrier decreases, the states in the quantum wells become
coupled to each other with the coupling strength determined by the wave function overlap. The
formation of the symmetric and antisymmetric (bounding and anti-bounding) states leads to the
splitting of the previously degenerate
quantum states with the splitting energy in the range $1-10$~meV\cite{Fisher2005,Fisher2006}. For
appropriate electron density only the two spin degenerate subbands (the ground and first excited)
are occupied (double occupancy) leaving the rest of the higher-energy subbands unoccupied since
their energy is considerably higher than the energy of the ground and first excited states.
The spin filtering effect proposed in this paper requires the lateral Rashba
SO interaction i.e. SOI generated by the lateral electric field
$\mathbf{F}=(0,F_y,0)$. In fact, recently, Gvozdic and Ekenberg\cite{Gvozdic2007} pointed out that
in the modulation-doped wide or coupled quantum wells, used for the bilayer nanowires fabrication,
the large intrinsic $F_y$ exists. Alternatively, in the experiment, the lateral electric field
can also be generated by the side gates attached to the channel. Regardless of the origin, the
lateral electric field $F_y$ induces the Rashba SO interaction with the effective magnetic field
directed along the grown $z$-axis. The possibility of using the SO interaction
induced by the lateral electric field has been recently reported in many
experiments\cite{Chuang2015,Debray2009,Das2011}, in which the lateral SO coupling constant has
been reported to vary in the range $0.04$~eV\AA$-0.5$~eV\AA.\\
\subsection*{Numerical methods}
Let us define the four element basis $\{ |1,\uparrow\rangle , |1,\downarrow\rangle ,
|2,\uparrow\rangle , |2,\downarrow\rangle\}$ which consists of
the spin-degenerate ground and first excited eigenstate related to the
confinement in the $z$ direction. In the presence of the SO interaction, we can derive the $4
\times 4$ Hamiltonian in the basis of these states\cite{Calsaverini2008,Wojcik2016_arxiv}.
After introducing a set $\pmb{\tau}$ of Pauli-like matrices in the orbital space, the Hamiltonian of
the system takes on the form (full derivation of Hamiltonian (\ref{eq:H}) can be found in
Supplementary)
\begin{eqnarray}
\label{eq:H}
H&=&\left [ \frac{\hbar ^2 (\hat{k} _x^2+\hat{k} _y^2)}{2m^*} + U_{QPC}(x,y) + |e|F_yy +
\varepsilon _+ \right ] \mathbf{1} \otimes \mathbf{1}
- \varepsilon _{-}\tau _z \otimes \mathbf{1} +\beta |e| F_y \hat{k}_x \mathbf{1} \otimes \sigma _z
- \beta \delta |e| F_y \tau _y \otimes \sigma _x \nonumber \\
&+&\beta_{12} \tau_{x} \otimes \left ( \sigma _x \hat{k}_y - \sigma _y \hat{k}_x \right ),
\end{eqnarray}
where $\mathbf{1}$ is the $2 \times 2$ unit matrix, $m^*$ is the electron effective mass, $\hat{k}
_{x(y)}=-i \partial / \partial x(y)$ is the wave vector operator, $\beta$ is
the lateral Rashba spin-orbit coupling constant, $\beta_{12}$ is the inter-subband lateral
Rashba spin-orbit coupling constant, $\delta=\langle 1, \sigma | \partial / \partial z | 2, \sigma
\rangle $ is the inter-subband coupling constant, $\varepsilon _{\pm}=(\varepsilon _2 \pm
\varepsilon _1)/2$ with $\varepsilon _{1(2)}$ being
the energy of eigenstate $|1, \sigma \rangle$ ($|2, \sigma \rangle$), $U_{QPC}(x,y)$ is the
potential energy of electron from the QPC which is given by
\begin{eqnarray}
&&U_{QPC}(x,y)=V_{QPC} \exp \left [ - \left ( \frac{x-L/2 }{\xi_x/2} \right ) ^2 \right ] \times
\left \{ \exp \left [ - \left ( \frac{y-W/2 }{\xi _y/2} \right ) ^2 \right ] + \exp \left [
- \left ( \frac{y+W/2 }{\xi _y/2} \right ) ^2 \right ] \right \},
\end{eqnarray}
where $V_{QPC}$ is the maximal potential energy in the QPC while $\xi _x$ and $\xi _y$ determine the
extension of the QPC in the $x$ and $y$ directions, respectively.
In Hamiltonian (\ref{eq:H}) we neglect the intra-subband SO coupling assuming that the system is
symmetric in the $z$ direction with respect to the reflection $z\rightarrow
-z$\cite{Wojcik2016_arxiv}. Since the Dresselhaus SO coupling constant $\gamma _D \sim 1/d_{QW}^2$ ($d_{QW}$
is the width of the quantum well in the $z$ direction), for the wide quantum well used in the experimental
realization of bilayer nanowires, the strength of the Dresselhaus SO interaction is a few orders of magnitude smaller than the Rashba SO coupling.
This allows us to neglect the Dresselhaus term in the Hamiltonian (\ref{eq:H}) - detailed discussion of the Dresselhaus SO coupling
in the considered nanostructure and its influence on the presented spin filtering can be found in Supplementary material.
The calculations of the conductance have been performed by the scattering matrix
method using the Kwant package\cite{kwant}. For this purpose we have transformed the
Hamiltonian (\ref{eq:H}) into the discretized form on the grid $(x_{\mu}, y_{\nu})= (\mu dx, \nu
dx$) with $\mu, \nu = 1,2, \ldots$, where $dx$ is the lattice constant.
We introduce the discrete representation of the electron state in the $4 \times 4$ space as follows:
$|\Psi(x_{\mu}, y_{\nu})\rangle
=
\left(|\psi_1^{\uparrow}( x_{\mu},y_{\nu})\rangle
,|\psi_1^{\downarrow}( x_{\mu},y_{\nu})\rangle, |\psi_2^{\uparrow}( x_{\mu},y_{\nu})\rangle,
|\psi_2^{\downarrow}( x_{\mu},y_{\nu})\rangle \right)^T
= |\Psi_{\mu, \nu}\rangle$.
The Hamiltonian (\ref{eq:H}) takes on the discretized two-dimensional form
\begin{eqnarray}
\label{HTB}
\mathcal{H}_{2D}&=& \sum\limits_{\mu\nu} \big [ (4t + \varepsilon _+ + U_{QPC}(x_{\mu},y_{\nu}) +
|e|F_yy_{\nu}) \mathbf{1} \otimes \mathbf{1}
- \varepsilon _- \tau _z \otimes \mathbf{1} - \beta \delta |e| F_y \tau _y \otimes \sigma _x \big ] |
\Psi_{\mu,\nu } \rangle \langle \Psi_{\mu,\nu}| \\
&+& \sum _{{\mu}{\nu}} \left [ t \mathbf{1} \otimes \mathbf{1} + it_{SO} \beta_{12} \tau
_x \otimes \sigma _y - it_{SO} \beta |e| F_y \mathbf{1} \otimes \sigma _z\right ] | \Psi_{\mu+1,\nu }
\rangle \langle \Psi_{\mu,\nu}| + H.c. \nonumber
\\
&+& \sum _{{\mu}{\nu}} \left [ t \mathbf{1} \otimes \mathbf{1} - it_{SO} \beta_{12} \tau
_x \otimes \sigma _x \right ] | \Psi_{\mu,\nu+1 } \rangle \langle \Psi_{\mu,\nu}| +H.c., \nonumber
\end{eqnarray}
where $t=\hbar ^2/(2m dx^2)$ and $t_{SO}=1/(2dx)$. \\
In the calculations we assume the hard-wall boundary
conditions in the $y$ direction i.e. $ |\Psi(x,0)\rangle=|\Psi(x,W)\rangle=0 $. In the $x$-direction,
the boundary conditions are set based on the fact that the electron state in the input is a linear combination of
the state in which the electron is injected into the channel and all possible states in which the electron
can be reflected from the QPC. Therefore
\begin{equation}
|\Psi_{in}(0,y)\rangle = c_{k_{n}} ^{\sigma} \exp (i k_{n}^{\sigma} x) | \varphi_{in,n}^{\sigma} (y) \rangle + \sum _{m,\sigma'}
c_{k_m}^{\sigma'} \exp (-ik_m^{\sigma'} x) |\varphi_{in,m}^{\sigma'}(y)\rangle,
\end{equation}
where $n,m=1,2$, $\sigma,\sigma '=\uparrow, \downarrow$ and $| \varphi_{in,n}^{\sigma} (y) \rangle$ are the eigenstates of the hamiltonian (\ref{HTB})
calculated
by the use of the boundary conditions in the input ($\mu=0$)
\begin{equation}
|\Psi_{\mu \pm 1, \nu } \rangle = \exp (\pm i k dx) |\Psi_{\nu} \rangle.
\end{equation}
Accordingly, the boundary conditions for the output are given by
\begin{equation}
| \Psi _{\mu, \nu+1} \rangle = \exp (ikdx) | \Psi _{\mu, \nu} \rangle.
\end{equation}
Let us assume that the electron with spin $\sigma$ in the subband $n$ ($n=1,2$) is injected from
the input contact into the nanowire. The electron can be transmitted through the QPC
in one of the four possible processes (in parenthesis the symbol of the transmission probability
is given): intra-subband transmission with spin conservation $(T_{nn}^{\sigma \sigma})$,
intra-subband transmission with spin-flip $(T_{nn}^{\sigma \bar{\sigma}})$, inter-subband
transmission with spin conservation ($T_{nm}^{\sigma \sigma}$), and inter-subband transmission with
spin flip ($T_{nm}^{\sigma \bar{\sigma}}$), where $\bar{\sigma}$ denotes the spin opposite to
$\sigma$, while $T_{nm}^{\sigma \sigma'}$ is the probability of the electron transmission between
the subbands $|n,\sigma \rangle \rightarrow |m,\sigma' \rangle $, ($n,m=1,2$, $\sigma,\sigma
'=\uparrow, \downarrow$).
Having determined the transmission coefficients $T^{\sigma \sigma '}_{nm}$
we calculate the conductance in the ballistic regime using the Landauer formula
\begin{equation}
G_{nm}^{\sigma \sigma^{\prime}}=\frac{e^2}{h} \int_0^{\infty} T_{nm} ^{\sigma \sigma ^{\prime}} (E) \left (
\frac{\partial f_{FD}(E,E_F)}{\partial E} \right ) dE,
\end{equation}
where $f_{FD}(E,E_F)=1/[1+\exp(E-E_F)/k_BT]$ is the Fermi-Dirac distribution function, $T$ is
the temperature and $E_F$ is the Fermi energy.
The spin-dependent conductances through the nanostructure are given by
\begin{eqnarray}
G^{\uparrow}&=&\sum _{n,m=1}^{2} (G_{nm}^{\uparrow \uparrow} + G_{nm}^{\downarrow \uparrow}), \\
G^{\downarrow}&=&\sum _{n,m=1}^{2} (G_{nm}^{\uparrow \downarrow} + G_{nm}^{\downarrow \downarrow}).
\end{eqnarray}
The total conductance $G=G^{\uparrow}+G^{\downarrow}$ and the spin polarization of current
\begin{equation}
P=(G^{\uparrow} - G^{\downarrow})/(G^{\uparrow}+G^{\downarrow}).
\end{equation}
The conductance calculations have been performed for $dx=2$~nm,
$L=3000$~nm, $W=92$~nm, $\xi _x=300$~nm, $\xi _y=48$~nm and
$V_{QPC}=12$~meV. We use the material parameters corresponding to In$_{0.5}$Ga$_{0.5}$As,
i.e., $m_e=0.0465m_0$ and the Rashba spin-orbit interaction constant $|e|\beta F_y=10$~meVnm\cite{Das2011}.
The energy difference between the two occupied subbands is taken to be $\Delta \varepsilon =
\varepsilon _2 - \varepsilon_1=1$~meV.
\section*{Results and Discussion}
In this section, we present the results of the conductance calculations and explain the
physical mechanism responsible for the nearly full spin polarization obtained in the
bilayer nanowires with QPC.
First, we assume that the inter-subband induced SO coupling constant $\beta_{12}=0$.
\begin{figure}[ht]
\begin{center}
\vspace{0.5cm}
\includegraphics[scale=0.6, angle=0]{fig2.pdf}
\caption{
(a) Conductance $G$ as a function of Fermi energy $E_F$ for two values of inter-subband coupling
constants $\delta$. Red and blue curves correspond to $G^{\uparrow}$ and
$G^{\downarrow}$, respectively. Black dashed curve shows the results for $\delta=0$, for which
$G^{\uparrow}=G^{\downarrow}$. (b) Spin polarization $P$ of the current as a function of Fermi
energy $E_F$ for $\delta=10^{-2}$~nm$^{-1}$. }
\label{fig2}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.6, angle=0]{fig3.pdf}
\caption{
Transmission probabilities $T_{nm}^{\sigma \sigma'}$ versus Fermi energy $E_F$ for (a)~$\delta=0$
and (b,c)~$\delta=10^{-2}$~nm$^{-2}$ for the electron injected into (b) the first
subband $|1,\sigma \rangle$ and (c) the second subband $|2,\sigma \rangle$,
$\sigma=\uparrow,\downarrow$.}
\label{fig3}
\end{center}
\end{figure}
Fig.~\ref{fig2}(a) displays the conductance $G^{\uparrow \downarrow}$ as a function of the Fermi
energy $E_F$ for two values of the inter-subband coupling constant $\delta$. Red and blue curves
correspond to $G^{\uparrow}$ and $G^{\downarrow}$, respectively. For $\delta=0$, depicted by the
black dashed curve, $G^{\uparrow}=G^{\downarrow}$ hence the spin polarization of current $P=0$ in
entire range of the Fermi energy. The two conductance steps are due to the subsequent
subbands passing through the Fermi level in the QPC region, each contributing to the increase of the
conductance by $e^2/h$.
For the nonzero inter-subband coupling, i.e for $\delta=10^{-2}$~nm$^{-1}$, the conductances
$G^{\uparrow}$ and $G^{\downarrow}$ differ from each other in some range of the Fermi energy, which
leads to the almost full spin polarization of current presented in Fig.~\ref{fig1}(b).
We note that the spin polarization occurs only in the Fermi energy range, which
corresponds to the conductance step for $\delta=0$. The transmission probabilities
$T_{nm}^{\sigma \sigma'}$ for $\delta=0$ and $\delta=10^{-2}$~nm$^{-1}$ are presented in
Fig.~\ref{fig3}.
Comparing results in Fig.~\ref{fig3}(a) and Fig.~\ref{fig3}(b,c) we see that the imbalance
between the spin-up and spin-down conductance, in the Fermi energy range, in which the spin
polarization is observed, results from the suppression of the intra-subband transmission
of spin-down electrons in the first subband $T_{11}^{\downarrow \downarrow}$ and the enhancement of
the intra-subband transmission of spin-up electrons $T_{22}^{\uparrow \uparrow}$. As expected the
inter-subband transmissions reveal the symmetry $T_{12}^{\uparrow \downarrow}=T_{21}^{\downarrow
\uparrow}$ and $T_{12}^{\downarrow \uparrow}=T_{21}^{\uparrow \downarrow}$.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.6, angle=0]{fig4.pdf}
\caption{
(a) Conductance $G$ as a function of Fermi energy $E_F$ for $V_{QPC}=0$. (b) The $z$ component of
the partial spin density distributions $s_{m}^{n\sigma}$, where $m$ is the index of the subband, for
which the spin density distribution is presented, while $n\sigma$ denotes the index of the subband,
including spin, from which the electrons are injected into the nanowire. Results for $E_F=3.15$~meV
and $V_{QPC}=0$. }
\label{fig4}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.6, angle=0]{fig5.pdf}
\caption{ (a) Dispersion relations $E$ vs $k$ in the leads IN and OUT as well as in
the nanowire with the SO interaction. The dashed horizontal line denotes the Fermi
energy $E_F=3.15$~meV, for which the spin density distribution maps are presented in
Fig.~\ref{fig4}(b). (b) Schematic illustration of the possible transmission processes through the
nanowire without QPC.
}
\label{fig5}
\end{center}
\end{figure}
In order to explain the physical mechanism behind the spin filtering effect let us first
consider the nanowire without QPC ($V_{QPC}=0$).
As shown in Fig.~\ref{fig4}(a), which presents the spin-dependent conductance
$G^{\uparrow \downarrow}(E_F)$, the constriction in the form of QPC located in the nanowire is
necessary to obtain the spin filtering. Without QPC, $G_{\uparrow}=G_{\downarrow}$, which leads
to the spin polarization $P=0$. The full understanding of the spin filtering mechanism, which
emerges when we add QPC, requires the understanding of the spin dynamics in the
nanowire without the constriction.
Fig.~\ref{fig4}(b) presents the $z$ component of the partial spin density distributions
$s_{m}^{n\sigma}$, where $m$ is the index of the subband, for which the spin density distribution is
presented, while $n\sigma$ denotes the index of the subband, including spin, from which the
electrons are injected into the nanowire. Explicitly,
\begin{eqnarray}
s_{1}^{n\sigma}(x_{\mu},y_{\nu})&=&|\psi_1^{\uparrow}( x_{\mu},y_{\nu})|^2-|\psi_1^{\downarrow}(
x_{\mu},y_{\nu})|^2, \\
s_{2}^{n\sigma}(x_{\mu},y_{\nu})&=&|\psi_2^{\uparrow}( x_{\mu},y_{\nu})|^2-|\psi_2^{\downarrow}(
x_{\mu},y_{\nu})|^2.
\end{eqnarray}
The dispersion relations for the leads IN and OUT as well as in the central part of the nanowire
with the SO interaction are presented in Fig.~\ref{fig5}(a). The Fermi energy, for which the maps
from Fig.~\ref{fig4}(b) have been calculated, is marked by the dashed horizontal line. We set
$E_F=3.15$~meV which corresponds to the maximal spin polarization of current presented in
Fig.~\ref{fig2}(b).
As shown in Fig.~\ref{fig4}(b) the electrons injected in the states $|1,\uparrow \rangle$ and
$|2, \downarrow \rangle$ flow through the nanowire in the subband into which they were injected,
conserving their spin. Different behavior is observed for the electrons injected in the
states $|1,\downarrow \rangle$ and $|2, \uparrow \rangle$. The electrons
injected into the subband $|1,\downarrow \rangle$ ($|2,\uparrow \rangle$) are transmitted to the
state $|2, \uparrow \rangle$ ($|1,\downarrow \rangle$) in the middle of the nanowire and are back in
their original state before leaving the nanowire through the lead OUT. To explain this
behavior let us consider the electron propagating in the state $|1,\downarrow \rangle$ from the
input channel IN. In the nanowire where the lateral Rashba SO interaction is present, the spin is no
longer a good quantum number, since the orbital and spin degrees of freedom are mixed. For the
appropriate Fermi energy the states $|1,\downarrow \rangle$ and $|2, \uparrow \rangle$ hybridize
giving raise to the
avoided crossing in the spin-split subbands presented in Fig.~\ref{fig5}(a). The probability
of the transition through the avoided crossing depends on the degree of the adiabaticity of the
electron transport and the avoided crossing width as predicted by the
Landau-Zener theory\cite{Landau,Zener}. For the fully diabatic transport the probability of
electron transfer from the state $|1,\downarrow \rangle$ to $|2, \uparrow \rangle$ is equal to $1$.
Therefore, the transfer probability through the avoiding crossing depends on the rate of the energy
level changes when the electron from the contact, in which the SO interaction is absent, flows
through the nanowires with the SO interaction. In our case, this effect is implemented by the
spatially dependent SO coupling constants which vary in accordance with the function
\begin{equation}
f(x)=\exp {\left [ - \left ( \frac{x-L/2}{\xi _{SO} /2 } \right )^{2p} \right ]},
\end{equation}
where $\xi_{SO}\approx L$ and $p$ is the so-called "softness" parameter. We assume $p=10$, which
guarantees the diabatic transport between the leads and the nanowire. \\
The schematic illustration of all possible electron transmission processes through the nanowire in
the $k$-space is presented in Fig.~\ref{fig5}(b). Among them the Landau-Zener transition between the
states $|1,\downarrow \rangle$ and $|2, \uparrow \rangle$ (with spin flip) is a key to understanding
the spin filtering effect, which emerges when we introduce the QPC.
Now, let us explain in detail the spin filtering mechanism, which emerges if we add the QPC.
For this purpose, in Fig.~\ref{fig6} we present the $z$ component of the partial spin density
distribution for $V_{QPC}=12$~meV.
Even a cursory analysis of this figure indicates, that for the chosen Fermi energy, the current
through the QPC is carried mainly by the spin-up electrons transmitted through both the subbands,
in agreement with the spin-dependent transmission probabilities presented in Fig.~\ref{fig3}. For
completeness, we have also calculated the local density of states (LDOS).
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.6, angle=0]{fig6.pdf}
\caption{ The $z$ component of the partial spin
density distributions $s_{m}^{n\sigma}$, where $m$ is the index of the subband in which the spin
density distribution is presented, while $n\sigma$ denote the index of the subband, including spin,
from which the electrons are injected into the nanowire. The gray contours present the QPC region.
Results for $E_F=3.15$~meV and $V_{QPC}=12$~meV. }
\label{fig6}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.8, angle=0]{fig7.pdf}
\caption{ Local density of states (LDOS) calculated for the subbands participating in
the electron transport. The gray contours correspond to the QPC region.}
\label{fig7}
\end{center}
\end{figure}
Figure~\ref{fig7} shows that there are no available states $|2,\sigma \rangle$
($\sigma=\uparrow,\downarrow$) in the QPC region, which means that the electrons reaching QPC in
the state $|2, \sigma \rangle$, independently of spin, are reflected from it. Now, we consider
separately the transport processes for the
electrons injected into the nanowire from the subsequent subbands (i) $|1,\uparrow \rangle$,
(ii) $|1,\downarrow \rangle$, (iii) $|2,\uparrow \rangle$ and (iv) $|2,\downarrow \rangle$.
Our analysis will be conducted on the basis of the partial spin density distributions
(Figs.~\ref{fig2} and \ref{fig6}), LDOS (Fig.~\ref{fig7}) and the dispersion relations in different
parts of the nanowire presented in Fig.~\ref{fig8}(a).
\begin{figure*}[ht]
\begin{center}
\includegraphics[scale=0.8, angle=0]{fig8.pdf}
\caption{ (a) Dispersion relation $E(k)$ in the leads IN and OUT as well as in
the nanowire with the SO interaction and in the QPC region. The dashed horizontal
line denotes the Fermi energy for which the spin density distribution maps are presented in
Fig.~\ref{fig6}. (b) Schematic illustration of the possible transmission processes through the
nanowire.}
\label{fig8}
\end{center}
\end{figure*}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.8, angle=0]{fig9.pdf}
\caption{
(a) Conductance $G^{\uparrow,\downarrow}$ as a function of Fermi energy $E_F$ for different values
of inter-subband coupling constant $\delta$. Red and blue lines correspond to $G^{\uparrow}$ and
$G^{\downarrow}$, respectively. (b) Spin polarization of current $P$ as a function of Fermi energy
$E_F$ and inter-subband coupling constant $\delta$. (c) Conductance
$G^{\uparrow,\downarrow}$ as a function of Fermi energy $E_F$ for different values
of inter-subband induced SO coupling constant $\beta _{12}$ and $\delta=10^{-2}$~nm$^{-1}$. (d)
Spin polarization of current $P$ as a function of Fermi energy $E_F$ and inter-subband induced SO
coupling constant $\beta _{12}$. The values of $\beta_{12}$ marked by the red arrows are taken to
the further analysis. Results for $V_{QPC}=12$~meV.}
\label{fig9}
\end{center}
\end{figure}
First, we focus on the processes (i) and (iv). As shown in Fig.~\ref{fig4}(b),
which presents the spin dynamics in the nanowire without QPC, the electrons injected in the state
$|1,\uparrow \rangle$ flow through the nanowire conserving their state. Due to the large density of
states $|1,\uparrow \rangle$ in the QPC region (see Fig.~\ref{fig7}), the electrons
injected in this subband are transmitted through the QPC giving raise to the spin-up polarized
current in the output (see Fig.~\ref{fig6}). (iv) Although the electrons in the state $|2,\downarrow
\rangle$ injected into the nanowire without QPC also flow through the nanowire conserving the state
[Fig.~\ref{fig4}(b)], if we introduce the QPC, the electrons are backscattered from the
constriction due to the lack of the states $|2,\downarrow \rangle$ in the QPC region as presented in
Fig.~\ref{fig7}.
The spin dynamics occurring for the electrons injected in the states
$|1,\downarrow \rangle$ and $|2,\uparrow \rangle$ is much more complicated. (ii) Due to the
hybridization of the subbands
$|1,\downarrow \rangle$ and $|2,\uparrow \rangle$ induced by the SO interaction (see avoiding
crossing in Fig.~\ref{fig8}(a)), before reaching the QPC the electrons injected in the state $|1,
\downarrow \rangle$ are transmitted to the state $|2,\uparrow \rangle$. This process is similar
to the previously explained spin dynamics, which occur in the nanowire without QPC
[Fig.~\ref{fig4}(b)]. Therefore, when reaching the QPC the electrons initially injected
in $|1, \downarrow \rangle$ are in the state $|2, \uparrow \rangle$ for which LDOS is zero
in the QPC region (Fig.~\ref{fig7}). These electrons are backscattered from the QPC as presented in
Fig.~\ref{fig6}.
The enhancement of the spin polarization of current is mainly related to the electrons
injected in the state $|2, \uparrow \rangle$ (iii). Although there are no available states for the
subband $|2, \uparrow \rangle$ in the QPC region (Fig.~\ref{fig7}), the electrons in the subband
$|2, \uparrow \rangle$ injected from the contact IN into the nanowire are transmitted to the state
$|1,\downarrow \rangle$ before reaching the QPC (see Fig.~\ref{fig6}). The Landau-Zener
transition is possible due to the lateral SO interaction, which causes the hybridization of the
states $|1,\downarrow \rangle$ and $|2,\uparrow \rangle$. Since LDOS in the QPC region for the state
$|1, \downarrow \rangle$ is large (Fig.~\ref{fig7}), the electrons pass through the QPC and, as
show in Fig.~\ref{fig6}, just after passing through the QPC they are transmitted to the state
$|2, \uparrow \rangle$ leaving the nanowire in this subband. This process, together with the
direct transition through the QPC in the state $|1, \uparrow \rangle$ result in the nearly full
spin-up polarization of the current for the chosen Fermi energy. The transport precesses (i)-(iv)
are schematically illustrated in Fig.~\ref{fig8}(b).
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.9, angle=0]{fig10.pdf}
\caption{ The $z$ component of the partial spin
density distributions $s_{m}^{n\sigma}$ calculated for the nanowire without QPC ($V_{QPC}=0$), for
(a)~$\beta _{12}=3.4$~meVnm and (b)~$\beta_{12}=6.1$~meVnm and for the nanowire with QPC
($V_{QPC}=12$~meV), for (c)~$\beta _{12}=3.4$~meVnm and (d)~$\beta_{12}=6.1$~meVnm. The gray
contours in figures (c) and (d) correspond to the QPC region.
Results for $E_F=3.15$~meV. }
\label{fig10}
\end{center}
\end{figure}
Based on the above analysis one can determine three important factors
necessary to achieve the spin polarization of current in the bilayer nanowires: (a) the lateral
Rashba SO interaction, (b) the inter-subband coupling and (c) the constriction, which in our case
has the form of the QPC. Among these factors the inter-subband coupling
resulting from the lateral Rashba SO interaction plays the crucial role. In order to show how the
value of the parameter $\delta$ affects the spin polarization of current, in Fig.~\ref{fig9}(a) we
present $G^{\uparrow \downarrow}(E_F)$ for
different value of $\delta$. The map of the spin polarization $P$ as a function of the Fermi
energy $E_F$ and the inter-subband coupling constant $\delta$ is presented in Fig.~\ref{fig9}(b).
These results show that there is a value of $\delta$, for which the spin polarization is
the largest and reaches $P=1$.
All the results presented so far have been obtained for the inter-subband SO coupling
constant $\beta_{12}=0$. However, the recent theoretical studies\cite{Calsaverini2008}, in which
the self-consistent calculations of the intra- and inter-subband SO coupling constants were
performed for the double quantum wells, reported that the inter-subband SO coupling constant
exhibits the resonant behavior reaching the values comparable to the ordinary (intra-subband) Rashba
SO constant. For this reason we introduce the inter-subband SO
interaction to our model. Figure~\ref{fig9}(c,d) presents the conductance
$G^{\uparrow \downarrow}(E_F)$ for different values of the inter-subband SO coupling constants
$\beta _{12}$ assuming $\delta=10^{-2}$~nm$^{-1}$.
The dependence of the spin polarization $P$ on $E_F$ and $\beta _{12}$
presented in Fig.~\ref{fig9}(d) reveals the damped oscillations of $P$ as a function of $\beta
_{12}$ with the period approximately equal to $6$~meVnm. The inter-subband SO interaction
causes that the electrons flowing through the nanostructure are periodically transfered between
the hybridized states
$|1,\downarrow \rangle$ and $|2,\uparrow \rangle$. Each of these transitions takes place at a
specific distance, that depends on the coupling constant $\beta _{12}$. Depending of the number of
transitions, which occur when the electrons pass the distance from the lead IN to the QPC region
we obtain the maximal (for odd number of transitions) or minimal (for even number of transitions)
spin polarization. In order to illustrate this property, in Fig.~\ref{fig10} we present the $z$
component of the partial spin density distributions $s_{m}^{n\sigma}$ for two chosen inter-subband
induced SO coupling constants: (a) $\beta _{12}=3.4$~meVnm corresponding to the minimum and (b)
$\beta _{12}=6.1$~meVnm corresponding to the maximum of $P$. The chosen values of $\beta_{12}$ are
marked by the red arrows in Fig.~\ref{fig9}(d).
Since our goal is to analyze the
inter-subband spin dynamics under the influence of the inter-subband SO interaction and its impact
on the spin filtering effect, results in Fig.~\ref{fig10} are presented for the two cases, for the
nanowire without QPC ($V_{QPC}=0$) (a,b) and with QPC
($V_{QPC}=12$~meV) (c,d).
Comparing the spin dynamics for the nanowire without QPC presented in Fig.~\ref{fig4}
and Figs.~\ref{fig10}(a,b) one can conclude that the only difference, which appears when we include
the inter-subband induced SO interaction, is the distance that the electron injected in the state
$|1,\downarrow \rangle$ ($|2,\uparrow \rangle$) needs to make a transition to the state
$|2,\uparrow \rangle$ ($|1,\downarrow \rangle$). Moreover, independently of the inter-subband SO
coupling $\beta _{12}$, the electrons injected in the states $|1,\uparrow \rangle$ and
$|2,\downarrow \rangle$ remain in their states flowing through the nanowire.
As explained, the spin filtering mechanism presented in the paper requires that the electrons
injected from the contact IN in state $|1,\downarrow \rangle$ ($|2,\uparrow \rangle$) make
transitions to the other state and reach the QPC region in state $|2,\uparrow \rangle$
($|1,\downarrow
\rangle$). This necessary condition causes that the electrons injected in the state $|1,\downarrow
\rangle$ are backscattered from the QPC while the electrons in the state $|2,\uparrow \rangle$
pass through the QPC giving raise to the high spin polarization of the current. As shown in
Fig.~\ref{fig10}(a), for $\beta _{12}=3.4$~meVnm corresponding to the spin polarization minima, the
electrons in the state $|1,\downarrow \rangle$ ($|2,\uparrow \rangle$) are transmitted
to the state $|2,\uparrow \rangle$ ($|1,\downarrow \rangle$) after traveling the distance $L$. Since
the QPC is located at $x=L/2$, the electrons injected from the lead IN reach the QPC in the state
that is mostly built from the initial state. This means that the spin
filtering mechanism described in the paper will not occur. Instead, as presented in
Fig.~\ref{fig10}(c) the electrons in the states $|1,\uparrow \rangle$ and $|1,\downarrow
\rangle$ simply pass through the QPC while the electrons in the states $|2,\uparrow \rangle$ and
$|2,\downarrow \rangle$ are backscattered due to zero LDOS in the QPC region (see
Fig.~\ref{fig7}). The electron transport in the nanowire through the subbands with
opposite spin leads to the spin polarization nearly equal to zero, i.e. the unpolarized current.
The further increase of the inter-subband induced SO interaction constant causes that the distance
over which the electron are transmitted between the states $|1,\downarrow \rangle$ and $|2,\uparrow
\rangle$ becomes shorter. For $\beta_{12}=6.1$~meVnm [Fig.~\ref{fig10}(b) and (d)] the necessary
condition, for which the spin filtering effect occurs, namely the inter-subband transition before
reaching QPC, is again satisfied giving raise to the high spin polarization of the current.
We can summarize that depending on the number of the inter-subband transitions, which
occur before the electron reaches the QPC acting as the channel selector, the spin polarization
oscillates between the large and small values as presented in Fig.~\ref{fig9}(d).
Our results in general provide a new mechanism to implement spin-polarized electron sources in the
bilayer nanowires. Therefore, it is important to discuss this proposal from the point of view of the possible perturbations,
which can appear in the experimental realization of the proposed spin filter.
Our model assumes three necessary conditions, which have to be satisfied in order to achieve the high spin
polarization of the current: (a) the lateral Rashba SO interaction,
(b) the inter-subband coupling and (c) the constriction, which in the present paper has the form of the QPC.
The lateral Rashba SO interaction is the most important requirement because it generates the coupling between the
subbands, which in the consider nanostructures, is crucial for the spin polarization. However, in the wells made of
materials with the zincblende crystallographic structure (GaAs, InAs, etc.), the SO coupling also originates from the bulk
Dresselhaus term. Although in the GaAs-based structures both the Rashba and Dresselhaus terms are often of the same order
of magnitudes, in the InGaAs-based wells growing in the $[001]$ direction, the Rashba term dominates.
Since the linear Dresselhaus parameter $\gamma _D \sim 1 / d_{QW}^2$, this domination is strengthened for
the wide quantum wells needed for the experimental realization of the bilayer nanowires, e.g
for the gated Al$_{0.48}$In$_{0.52}$As/Ga$_{0.47}$In$_{0.53}$As double quantum well structure, the Dresselhaus SO parameter is
a four orders of magnitude smaller than the Rashba coupling constant - see the Supplementary material. We have checked by performing direct
calculations that the inclusion of such a small term into our model does not affect the spin filtering presented in the paper.
Although the neglect of the Dresselhaus term in our model is fully justified, the assumption of constant, spatially independent
Rashba parameter requires a detailed discussion. In the considerations presented so far, we assume that the Rashba SO parameter is constant and
spatially independent.
However, in the realistic nanostructure any local imperfection leads to the local change in the SO coupling. Since, the carriers in the quantum well
come from donors, the electric field of ionized donors generates the random, spatially dependent electrostatic potential in the quantum well.
This causes that the SO interaction has a random component, which for some structures can be large. Such strong fluctuations of
the Rashba coupling constant have been recently measured by the scanning tunnelling microscopy in the structure with 2DEG fabricated by adsorbing
Cs on the p-type InSb(110)\cite{Bindel2016}. In that case\cite{Bindel2016}, the strong spatial fluctuations of the Rashba parameter between
$0.4$~eV$\AA$ and $1.6$~eV$\AA$ result from the particular, random locations of the dopant ions, which are very close to the inversion layer
where the electron density is located. In the
Al$_{0.48}$In$_{0.52}$As/Ga$_{0.47}$In$_{0.53}$As double quantum well structure suggested to the experimental realization of
the proposed spin filter, the donor-doped layers are located symmetrically on both sides of the well in the distance $R_d$ much larger than the their
width $w_d$ (usually $R_d/w_d \approx 5-10$).The random distribution of the electric field and the Rashba spin-orbit parameter in the main quantum
well originating from the inhomogeneity of the dopant concentration in the donor-doped layers has been discussed in detail in Ref.~\cite{Glazov2010}.
Following this model\cite{Glazov2010} we have studied the role of the random spin-orbit coupling component generated by the donors
on the spin filter effect presented in the paper. For this purpose, we have calculated the $z$-component
of the electric field of the dopant ions with the local concentration $n^d_{2D}(x',y')$
\begin{equation}
F_z(x,y)=-\frac{|e|}{4 \pi \varepsilon _0 \varepsilon} \int n^d_{2D}(x',y')\frac{R_d}{[(x-x')^2+(y-y')^2+R_d^2]^{3/2}} dx' dy',
\end{equation}
where $e$ is the electron charge and $\varepsilon$ is the dielectric constant. We assume that the random
distribution of dopants $n^d_{2D}(x',y')$ obeys the Gaussian statistics with the mean value $\bar{n}^d_{2D}$ and the standard deviation $\delta
n^d_{2D}$. In Fig.~\ref{fig11} we present the spatial distributions of the electric field $F_z(x,y)$ calculated for the different mean values of donor
concentration $\bar{n}^d_{2D}$ and deviations $\delta n^d_{2D}$ as well as different distances $R_d$ of the dopant layers from the main
quantum well. We assume $\varepsilon=14$ corresponding to InAs.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.9, angle=0]{fig11.pdf}
\caption{Spatial distributions of the electric field $F_z(x,y)$ calculated for the different $\bar{n}^d_{2D}$ and $\delta n^d_{2D}$ as well as
different distances $R_d$ of the dopant layers from the main quantum well. Right panels present the spin polarization of the current $P(E_F)$
calculated with the inclusion of the random component of the spin-orbit coupling generated by the electric field distribution presented on the
left panels.}
\label{fig11}
\end{center}
\end{figure}
The spatially dependent electric field $F_z(x,y)$ leads to the random distribution of the intra-subband SO coupling
$\beta_{11}$ and $\beta_{22}$ [see hamiltonian (5) in the Supplementary material] which in the presented model were assumed to be zero due to the $z
\leftrightarrow -z$ symmetry. In order to show how this effect affects the spin-filtering in the considered nanostructures, we include the
random component of the spin-orbit coupling into our model and calculate the spin polarization of the current $P$ as a function of the Fermi energy
$E_F$. The right panels in Fig.~\ref{fig11} present the spin polarization of the current $P(E_F)$ calculated with the inclusion of the spatial
distribution of the intra-subband spin-orbit parameters, generated by the electric field distribution presented on the left panels.
In the calculations we assume $\beta_{11}(x,y)=\beta_{3D}|e|F_z(x,y)$ (we take on $\beta_{3D}=0.572$~nm$^{-2}$ corresponding to InAs) and
$\beta_{22}(x,y)=-\beta_{11}(x,y)$. The second assumption is based on our
recent results\cite{Wojcik2016_arxiv} which show that the intra-subband spin-orbit coupling constants for the ground and first excited state have
nearly the same value but opposite sign. The rest of the parameters are assumed to be the same as used in the calculations presented the
Fig.~\ref{fig2}. Figure \ref{fig11} shows that for the reasonable values of the dopant concentration and its location with respect to the main quantum
well, the random spatial distribution of the spin-orbit coupling does
not affect the spin filtering presented in the paper. Its negligible contribution results from the fact that the $z$-component of the electric field
of the ionised dopants is several orders of magnitude smaller than the lateral electric field needed to obtain the spin filter effect and used in the
experiments for modulation of the Rashba parameter by the QPCs~\cite{Das2011}. In the above discussion we did not consider the screening of the
$z$-component of the electric field due to the $z \leftrightarrow -z$ symmetry\cite{Glazov2010}. As shown in Ref.~\cite{Glazov2010}, the screening is
important for the in-plane components of the electric field, however, the estimated
ratio of the fluctuations of the lateral and transversal components of the electric field~\cite{Glazov2010} in the symmetrically doped quantum well
is $F_{||} / F_{z} \ll 1$. Therefore, the screening strongly reduces the fluctuations of the lateral electric field and (as much smaller
than the transversal one) does not affect the spin filter effect presented in the paper.
\section*{Summary}
The inter-subband SO interaction attracts the growing interest, because it gives
raise to interesting physical effects, e.g., unusual Zitterbewegung\cite{Bernardes2007}. In
the bilayer nanowires, the strength of this specific SO interaction, arising from the coupling
between states with opposite parity, is comparable to the Rashba intra-subband SO coupling.
It makes the bilayer nanowires a good candidate for investigating the inter-subband SO interaction
and the effects related with it.
In the present paper, we have proposed the spin filtering mechanism based on the
inter-subband SO interaction in the bilayer nanowire with QPC. For this purpose we have studied the
electron transport through the nanowire within the two-subband model including the inter-subband SO
interaction induced by the lateral electric field. We have found that for the non-zero inter-subband
coupling, in the presence of the lateral Rashba SO interaction, the current flowing through
the QPC is almost fully spin polarized. In order to explain the spin filtering effect, first we
have considered the bilayer nanowire without QPC. By the use of the partial spin density
distributions calculated for each of the subband participating in the transport, we have shown that
the electrons injected in the state $|1,\downarrow \rangle$ ($|2,\uparrow \rangle$) are transmitted
to the state $|2, \uparrow \rangle$ ($|1,\downarrow \rangle$) in the middle of the nanowire and
again are back in their original state before leaving the nanowire. On the other hand, the
electrons in the states $|1,\uparrow \rangle$ and $|2,\downarrow \rangle$ flow through the nanowire
conserving their state. The observed Landau-Zener transitions are caused by the
hybridization of states $|1,\downarrow \rangle$ and $|2,\uparrow \rangle$ induced by the
lateral Rashba SO interaction, which mixes the orbital and spin degrees of freedom.
The proposed spin filtering mechanism emerges after adding the QPC, which blocks the
electron traveling in the states $|2, \sigma \rangle$ - as shown in Fig.~\ref{fig7}, LDOS for these
states in the QPC region is zero. Therefore, the introduction of the QPC causes that only the
electrons injected into the states $|1,\uparrow \rangle$ and $|2,\uparrow \rangle$ are transmitted
through the QPC giving raise to the high spin polarization of the current - the electrons in the
state $|1,\uparrow \rangle$ pass through the QPC remaining in their state while the electrons
injected in $|2,\uparrow \rangle$ are transmitted to the state $|1,\downarrow \rangle$ before
reaching QPC, pass through the QPC in the state $|1,\downarrow \rangle$ and just behind the QPC
are again transmitted to $|2,\uparrow \rangle$ leaving the nanowire in this state.
Summing up, the proposed spin filtering effect can be explained as the combined effect of the
Landau-Zener inter-subband transitions caused by the
hybridization of states with opposite spin and the
confinement in the QPC region. We
have determined the three important factors necessary to achieve the high spin polarization of the
current in the bilayer nanowire: (a) the lateral Rashba SO interaction,
(b) the inter-subband coupling and (c) the constriction which in our case has the form of the QPC.
Our results in general provide a new mechanism to implement spin-polarized electron sources in the
realistic bilayer nanowires with QPC, which can be realized
experimentally in the double quantum well structure or the wide quantum well. This is especially
interesting in the context of the current research on the spin filtering in QPC with single
occupancy\cite{Debray2009,Das2011,Das2012,Bhandari2012,Bhandari2013,Wan2009,Bhandari2013,
Kohda2012}. In those systems, the lateral Rashba SO interaction causes a small spin
imbalance, which then is gained by the electron-electron interaction. Our proposal, based on the
inter-subband SO interaction in the bilayer nanowires, gives the nearly full spin
polarization even without inclusion of the electron-electron interaction. This allows us to expect
that the proposed spin filtering effect is more efficient and in the near future
can lead to the fabrication of the efficient spin filter.
|
\section*{Introduction}
There are a wealth of results and conjectures linking important arithmetic properties of number-theoretic objects to special values of their $L$-functions. One way of attacking these conjectures is through the theory of $p$-adic $L$-functions, whose study has facilitated significant progress in examples such as the Birch and Swinnerton-Dyer conjecture. Such objects are ($p$-adic) analytic functions that interpolate special values of classical $L$-functions.
$\\ \\$
The interpolation property that $p$-adic $L$-functions satisfy involves a scaling factor that can, at exceptional values, vanish, forcing the $p$-adic $L$-function to vanish independently of the corresponding classical $L$-value. This happens, for example, at the value $L_p(f,0)$ if $f$ is a weight 2 newform attached to an elliptic curve $E/\Q$ at which $p$ has split multiplicative reduction. In this example, the $p$-adic $L$-function may \emph{a priori} have no meaningful relation to the central special value $L(f,1) = L(E,1)$ which plays the leading role in the Birch and Swinnerton-Dyer conjecture. In the case of newforms for $\GLt$ over $\Q$, this phenomenon was first investigated by Mazur, Tate and Teitelbaum in \cite{MTT86}, where they conjectured a relation between the derivative of the $p$-adic $L$-function and the classical $L$-function at an exceptional value. Explicitly, if $f\in S_{k+2}(\Gamma_0(N))$ is a newform with $k$ even and $p$ exactly dividing $N$, then the $p$-adic $L$-function of $f$ has an exceptional zero at the value $(\chi w_{\mathrm{Tm}}^{k/2},k/2)$, where $ w_{\mathrm{Tm}}$ is the Teichm\"{u}ller character and $\chi$ is any Dirichlet character with $\chi(p) = \omega$, where $-\omega$ is the Atkin-Lehner eigenvalue of $f$ at $p$. Mazur, Tate and Teitelbaum conjectured the existence of an $\mathcal{L}$-invariant $\mathcal{L}_p$ -- depending only on $f$ and $p$ -- such that
\[\frac{d}{ds}L_p(f,\chi w_{\mathrm{Tm}}^{k/2},s)\big|_{s = k/2} = \mathcal{L}_p K(\chi,k/2)\Lambda(f,\chi,k/2+1)/\Omega_f^\pm,\]
where $K(\chi,k/2)$ is some explicit non-zero scalar, $\Omega_f^\pm$ is the period of $f$, and $\Lambda$ is the (normalised) $L$-function of $f$.
$\\ \\$
Since this conjecture was made, it has been widely studied and proved in many different ways, whilst there are a number of equivalent descriptions of the $\mathcal{L}$-invariant in terms of arithmetic data. As an example, for weight 2 it was proved by Greenberg and Stevens in \cite{GS93} and in higher weights it has been proved independently by Kato--Kurihara--Tsuji, Stevens, and Emerton using $\mathcal{L}$-invariants defined by Fontaine--Mazur, Coleman, and Breuil respectively. For a fuller description on previous work in this setting, see the introduction of \cite{BDI10}. Of particular interest for this paper are proofs of the conjecture by Darmon in \cite{Dar01} (for weight 2) and Orton in \cite{Ort04} (for higher weights) using the theory of Bruhat--Tits trees.
$\\ \\$
Despite the problem of exceptional zeros being well-studied for $\GLt$ over $\Q$, many of the approaches mentioned above use deep arithmetic results and structures that do not necessarily exist in the case of $\GLt$ over more general number fields, and accordingly this case has received considerably less attention. In particular, for $\GLt$ over an imaginary quadratic field, the symmetric space playing the role of the modular curve has no complex or algebraic structure, making it much harder to attach Galois representations to automorphic forms. The theory of $p$-adic variation of modular forms, which also plays a prominent role in the proof of Greenberg and Stevens, is also comparatively badly behaved; for example, there are classical Bianchi modular forms which do not vary in a $p$-adic family where the classical forms are dense.
$\\ \\$
Over totally real fields, the conjecture has been proved in recent years for certain weight 2 Hilbert modular forms by Mok and Spie{\ss} (see \cite{Mok09} and \cite{Spi14}). The case of higher weights has been addressed by the first author, Dimitrov and Jorza in \cite{BDJ17}. Until recently, there were no analogous results over fields with complex embeddings. The first were results of Deppe in \cite{Dep16}, where he showed that (again for certain ordinary weight 2 forms) in the cases where exceptional zeros occur, they have (at least) the expected order of vanishing. In this paper, we treat the case of arbitrary weight for $\GLt$ over an imaginary quadratic field using the methods of Darmon and Orton, which allows us to circumvent the need for the deep arithmetic structure missing in this case. We do not require an ordinarity condition, and we also make no assumption on the prime $p$. Our main result is summarised below in Theorem \ref{summarythm} of the introduction.
\subsection*{Outline of method and plan of paper}
Section 1 contains classical preliminaries. In particular, let $F$ be an imaginary quadratic field of class number 1.\footnote{This is not a serious restriction; see Section \ref{class number}.} Let $f$ be a cuspidal \emph{Bianchi modular form}, that is, a cuspidal automorphic form $f$ for $\GLt$ over $F$. Let $f$ have level $\Gamma_0(\n) \subset \SLt(\roi_F)$, with $(p)$ dividing $\n$, and weight $(k,k)$ (since cuspidal Bianchi modular forms exist only at parallel weights).\footnote{We normalise the weight in such a way that a modular elliptic curve over $F$ corresponds to a weight $(0,0)$ form. In the literature, this is sometimes referred to as `weight 2'.} Let $\Lambda$ denote the (normalised) $L$-function of $f$. A construction of $p$-adic $L$-functions for such objects is contained in \cite{Wil17} in the case where the form is an eigenform satisfying a small slope property at primes of $F$ above $p$; a summary of the relevant results of \cite{Wil17} is contained in Section 3. In this setting, the interpolation property of the $p$-adic $L$-function involves an exceptional factor for each prime $\pri$ above $p$.
$\\ \\$
Fix such a prime $\pri|p$ and let $\tree_{\pri}$ be the Bruhat--Tits tree of $\GLt(F_{\pri})$. In Section 2, we show how to vary $f$ in a `family' $\mathcal{F}$ over the edges of $\tree_{\pri}$. In particular, to $f$ we associate a form
\[\mathcal{F} :\mathcal{E}(\tree_{\pri}) \times\GLt(\A_F)\longrightarrow V_{2k+2}(\C),\]
where $V_{2k+2}(\C)$ is the space of complex homogeneous polynomials of degree $2k+2$ in two variables, in such a way that for any edge $e$, $\mathcal{F}(e,\cdot)$ is a Bianchi modular form of specified level (depending on $e$) with the same weight as $f$. We further prove (in Theorem \ref{new forms tree}) that this form is harmonic in the first variable $\tree_{\pri}$ if and only if $f$ is new at $\pri$.
$\\ \\$
In Section 4, we define modular symbols on $\tree_{\pri}$ -- or, rather, families of Bianchi modular symbols indexed by $\mathcal{E}(\tree_{\pri})$ -- using the methods of Section 2, and define Hecke operators on the space of such symbols. The study of $\tree_{\pri}$ allows us to recover a classical result on the Hecke eigenvalue at $\pri$ of a $\pri$-new form as an easy corollary, and this result gives a systematic supply of exceptional zeros at central values (see Corollary \ref{exceptional zero existence}). In particular, we obtain:
\begin{mpropnum}Suppose $f$ is $\pri$-new, weight $(k,k)$ and small slope at all the primes above $p$, let $-\omega$ be the Atkin-Lehner eigenvalue of $f$ at $\pri$, and let $\varphi$ be a Hecke character for which $L(f,\varphi)$ is critical (in the sense of Deligne). Then the $p$-adic $L$-function of $f$ has an exceptional zero (for $\pri$) at $\varphi$ precisely when $k$ is even and $\varphi = \chi|\cdot|^{k/2}$, where $\chi$ is a finite order Hecke character with $\chi(\pri) = \omega$.
\end{mpropnum}
In Section 5, we utilise the overconvergent methods of \cite{Wil17} to attach a family of distributions on $\Proj(F_{\pri})$ to the form $f$. In particular, as for modular forms and modular symbols, we vary the overconvergent modular symbol attached to $f$ over the edges of $\tree_{\pri}$. In the case of a modular elliptic curve $E$ over $F$, this distribution has been studied before; in particular, in \cite{Tri06} Mak Trifkovic used it to construct Stark--Heegner points on $E$, whilst in the same setting the method of construction given here has been implemented independently by Xevi Guitart, Marc Masdeu and Haluk Sengun in \cite{GMS15}.
$\\ \\$
In Section 6, we use the distribution to define double integrals on $\uhp_{\pri}\times\uhs$ in the style of Darmon. Drawing the previous sections together, we use this to define two group cohomology classes $\mathrm{lc}_f$ and $\mathrm{oc}_f$ in $\h^1(\Gamma,\Delta_{k,k})$ (see Definition \ref{group gamma} for the definition of $\Gamma$, and Section 6 for the definition of $\Delta_{k,k})$. These classes, which depend on the choice of a complex period $\Omega_f$, are both eigenclasses for the Hecke action with the same eigenvalues as $f$. In Section 7, they are related to $L$-values using invariants attached to embeddings $F\times F \hookrightarrow \mathrm{M}_2(F)$. In particular, we prove the following:
\begin{mthmalph}Let $\chi$ be a finite order Hecke character with $\chi(\pri) = \omega$. Then there exists a map $\mathrm{Ev}_{\chi} : \h^1(\Gamma,\Delta_{k,k}) \longrightarrow \Cp$ such that
\[\mathrm{Ev}_{\chi}(\mathrm{oc}_f) = K(\chi,k/2+1)\Lambda(f,\chi, k/2+1)/\Omega_f\]
for some explicit (non-zero) factor $K(\chi,k/2+1)$, and
\[\bigg[\prod_{\mathfrak{q}}Z_{\mathfrak{q}}\bigg]\mathrm{Ev}_{\chi}(\mathrm{lc}_f) = \frac{\partial}{\partial s_{\pri}}L_p(f,\chiw_{\mathrm{Tm}}^{k/2},\mathbf{s})\big|_{\mathbf{s} = k/2},\]
where the product is over all $\mathfrak{q}|p$ such that $\mathfrak{q} \neq \pri$, and where $Z_{\mathfrak{q}}$ is the exceptional factor at $\mathfrak{q}$.
\end{mthmalph}
In Section 8, we show that in fact, the Hecke eigenspace in which $\mathrm{oc}_f$ and $\mathrm{lc}_f$ both live is in fact one-dimensional. In particular, after proving a non-vanishing result for $\mathrm{oc}_f$, we see that there exists some element $\mathcal{L}_{\pri} \in \Cp$ such that $\mathrm{lc}_f = \mathcal{L}_{\pri}\mathrm{oc}_f$. Combined with the above, this is enough to prove the following, which is the main result of this paper.
\begin{mthmalph}\label{summarythm}
Let $f$ be as above. Let $\chi$ be a finite order Hecke character with $\chi(\pri) = \omega$, so that $L_p(f,\chi w_{\mathrm{Tm}}^{k/2},\mathbf{k/2}) = 0$ is an exceptional zero (where $w_{\mathrm{Tm}}$ is the Teichm\"{u}ller character). Then there exists an $\mathcal{L}$-invariant $\mathcal{L}_{\pri} \in \Cp$, depending only on $f$ and $\pri$, such that
\[\frac{\partial}{\partial s_{\pri}}L_p(f,\chi w_{\mathrm{Tm}}^{k/2},\mathbf{s})\big|_{\mathbf{s} = k/2} = \mathcal{L}_{\pri}\cdot\bigg[\prod_{\mathfrak{q}}Z_{\mathfrak{q}}\bigg]K(\chi,k/2+1)\Lambda(f,\chi,k/2+1)/\Omega_f.\]
\end{mthmalph}
In Section 10, we discuss the case where $f$ is the base-change of a classical modular form $\tilde{f}$, and explicitly relate the $\mathcal{L}$-invariants of $f$ and $\tilde{f}$. This is enough to resolve a conjecture of Trifkovi\'{c} in the base-change case. We conclude in Section 11 by discussing possible further generalisations of these results; in particular, we discuss the case of higher class number (which has been omitted in this paper for expositional and notational reasons), possible results on the second derivative (to remove the remaining exceptional factor in the case where $p$ is split), and the conjectural form $\mathcal{L}_{\pri}$ might take in terms of arithmetic data, a question the authors hope to address in future work.
\subsection*{Comparison to relevant literature} The results of this paper fit into a body of recent work in complimentary directions. The work of Spie{\ss} (in \cite{Spi14}) and the work of the first author, Dimitrov and Jorza (in \cite{BDJ17}) on the Hilbert case has been mentioned above, as has Deppe's work (in \cite{Dep16}) on the case of weight 2 for general number fields. Exceptional zero phenomena have also been studied in the anticyclotomic setting by Molina (in \cite{Mol15}, for $\GLt$ in the totally real case) and Bergunde (in \cite{Ber17}, for non-split quaternion algebras over general number fields). Finally, the behaviour of $\mathcal{L}$-invariants under abelian base-change, again in the case of weight 2 for $\GLt$, has recently also been studied in work of Gehrmann in \cite{Geh17}.
\subsection*{Acknowledgements} The authors would like to thank Henri Darmon, who initially suggested this problem to us. The project has also benefited from helpful discussions with Tobias Berger, Kevin Buzzard, John Cremona, Xevi Guitart, David Loeffler, Marc Masdeu, and Aurel Page. The first author has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (Grant Agreement No.\ 682152). The second author would like to thank Victor Rotger, whose generous financial support enabled him to visit Barcelona to complete this project. Finally, we would like to thank the anonymous referee for their comments and corrections following their careful reading of the submitted version.
\section{Preliminaries}\label{notation}
\subsection{Basic notation}\label{basic definitions}
Let $F$ be an imaginary quadratic field with ring of integers $\roi_F$, different $\mathcal{D}$ and discriminant $-D$. Let $p$ be a rational prime. If $\pri$ is a prime of $F$ above $p$, write $F_{\pri}$ for the completion of $F$ at $\pri$, write $\roi_{\pri}$ for the ring of integers in $F_{\pri}$ and fix a uniformiser $\pi$ at $\pri$. We also write $e_{\pri}$ for the ramification index of $\pri$ over $p$. In the sequel, $F$ will be taken to have class number 1, in which case we assume $\pi$ is an element of $\roi_F$ that generates $\pri$. Denote the adele ring of $F$ by $\A_F = F_\infty \times \A_F^f$, where $F_\infty$ denotes the infinite adeles and $\A_F^f$ the finite adeles. Furthermore, define $\widehat{\roi_F} \defeq \roi_F\otimes_{\Z}\widehat{\Z}$ to be the finite integral adeles. Throughout, we'll fix a prime $\pri$ above $p$ and a level $\n = \pri\m$, where $\m$ is coprime to $\pri$ and divisible by all the other primes above $p$. For an ideal $\ff \subset \roi_F$, let $\mathrm{Cl}_F(\ff)$ denote the ray class group of $F$ modulo $\ff$.
$\\ \\$
Let $k \geq 0$ be an integer, and for any ring $R$, let $V_{2k+2}(R)$ denote the ring of polynomials over $R$ of degree at most $2k+2$. We also define $V_0(R) = R$. Note that $V_{2k+2}(\C)$ is an irreducible complex right representation of $\SUt(\C)$, and denote the corresponding antihomomorphism by $\rho':\SUt(\C) \rightarrow \mathrm{GL}(V_{2k+2}(\C))$. Finally, define an antihomomorphism $\rho : \SUt(\C) \times \C^\times \rightarrow \mathrm{GL}(V_{2k+2}(\C))$ by $\rho(u,z) = \rho'(u)|z|^{-k}$.
\subsection{Bianchi modular forms}\label{BMF}
A \emph{Bianchi modular form} is an automorphic form for $\GLt$ over an imaginary quadratic field. We will give only a very brief description of the theory; for a more detailed exposition of the literature, see \cite{Wil17}. First, we fix a level group.
\begin{mdef}\label{omega}
Define $\Omega_0(\n) \defeq \left\{\matrd{a}{b}{c}{d} \in \GLt(\widehat{\roi_F}): c \in \n\widehat{\roi_F}\right\}.$
\end{mdef}
\begin{mdef}We say a function $f:\GLt(\A_F) \rightarrow V_{2k+2}(\C)$ is a \emph{cusp form of weight $(k,k)$ and level $\Omega_0(\n)$} if it satisfies:
\begin{itemize}
\item[(i)]$f(zgu) = f(g)\rho(u,z)$ for $u \in \SUt(\C)$ and $z \in Z(\GLt(\C)) \cong \C^\times,$
\item[(ii)]$f$ is right-invariant under the group $\Omega_0(\n)$,
\item[(iii)] $f$ is left-invariant under $\GLt(F)$,
\item[(iv)] $f$ is an eigenfunction of the operator $\partial$ with eigenvalue $(k^2/2 + k)$, where $\partial/4$ denotes a component of the Casimir operator in the Lie algebra $\mathfrak{s}\mathfrak{l}_2(\C)\otimes_{\R}\C$ (see \cite{Hid93}, section 1.3), and
\item[(v)] $f$ satisfies the cuspidal condition that for all $g \in \GLt(\A_F)$, we have
\[\int_{F\backslash\A_F}f(ug)du = 0,\]
where we consider $\A_F$ to be embedded inside $\GLt(\A_F)$ by the map sending $u$ to $\smallmatrd{1}{u}{0}{1}$, and $du$ is the Lebesgue measure on $\A_F$.
\end{itemize}
Denote the space of such functions by $S_{k,k}(\Omega_0(\n))$.
\end{mdef}
\begin{longversion}
There is a good theory of Hecke operators on spaces of Bianchi modular forms, and there are multiplicity one theorems for eigenspaces. In particular, the Atkin-Lehner theory of oldforms and newforms passes over to this setting in a natural way; we will make this more precise in the sequel.
\end{longversion}
\begin{mrem}
Note that if $f \in S_{k,k}(\Omega_0(\n))$, then $f(gh) = f(g)$ for all $h\in Z(\GLt(\A_F^f)).$ Hence we can consider $f$ as a function on $\mathrm{PGL}_2^f(\A_F) \defeq \GLt(F_\infty) \times \mathrm{PGL}_2(\A_F^f)$.
\end{mrem}
\subsection{The $L$-function of a Bianchi modular form}
\begin{longversion}
We recap the definitions of Hecke operators, via double coset operators. Let $I \subset \roi_F$ be an ideal, recall the definition of $x_I$ from Section \ref{basic definitions}, and choose a (finite) set of representatives $\delta_j^{I}$ such that
\[\Omega_0(\n)\matrd{1}{0}{0}{x_I}\Omega_0(\n) = \coprod_{j}\Omega_0(\n)\delta_{j}^{I}.\]
\begin{mdef}\label{Hecke}
The Hecke operator at $I$ is defined by
\[f|T_{I}(g) = \sum_j f\left(g\delta_j^{I}\right).\]
When $I$ is not coprime to $\n$, we instead write $U_{I}$ for this operator.
\end{mdef}
\end{longversion}
\begin{shortversion}There is a good theory of Hecke operators (indexed by ideals of $\roi_F)$ on Bianchi modular forms.
\end{shortversion} Let $f$ be a cuspidal Bianchi modular form that is an eigenform for all of the Hecke operators, and for any non-zero ideal $I \subset \roi_F$, write $f|T_{I} = \lambda_{I}f.$
\begin{mdef}
The (normalised) \emph{$L$-function of $f$} is the function
\[
\Lambda(f,\varphi) \defeq \frac{\Gamma(q+1)\Gamma(r+1)}{(2\pi i)^{q+r+2}}\sum_{I \subset \roi_F, I\neq 0} \lambda_I\varphi(I) N(I)^{-1},
\]
where $\varphi$ is a Hecke character of infinity type $(q,r)$ and $\Gamma$ denotes the usual $\Gamma$-function.
\end{mdef}
The `critical' values of this $L$-function can be controlled; in particular, we have the following:
\begin{mprop}\label{period}
There exists a period $\Omega_{f} \in \C^\times$ and a number field $E$ such that, if $\varphi$ is a Hecke character of infinity type $0 \leq (q,r) \leq (k,k)$, with $q,r \in \Z$, we have
\[\frac{\Lambda(f,\varphi)}{\Omega_{f}} \in E(\varphi),\]
where $E(\varphi)$ is the number field over $E$ generated by the values of $\varphi$.
\end{mprop}
\begin{proof}
See \cite{Hid94}, Theorem 8.1.
\end{proof}
In the case of newforms, as treated in the sequel, exceptional zeros occur at very specific Hecke characters of the form $\chi|\cdot|^r$, where $\chi$ is a finite order Hecke character and $|\cdot|$ is the adelic norm. In this case, we will write
\[\Lambda(f,\chi,r+1) \defeq \Lambda(f,\chi|\cdot|^r)\]
for convenience and to ensure consistent notation in later sections.
\section{Forms on the Bruhat--Tits tree}
In this section, we develop the theory of Bianchi modular forms on the Bruhat--Tits tree. We will give the theory in quite general terms, since it is not significantly more difficult to describe the results in adelic language instead of classical language, and the more general results may be of independent interest. In later sections, we will specialise to the case where the class number is 1.
\subsection{The Bruhat--Tits tree}
\begin{mdef}
Let $\pri$ be a prime of $F$ above $p$. We denote by $\tree_{\pri}$ the Bruhat--Tits tree for $\GLt(F_{\pri})$, that is, the connected tree whose vertices are homothety classes of lattices in a two-dimensional $F_{\pri}$-vector space $V$. Two vertices $v$ and $v'$ are joined by an edge if and only if there are representatives $L$ and $L'$ respectively such that
\[\pi_{\pri}L' \subset L \subset L'.\]
Each edge comes equipped with an orientation. Denote the set of (oriented) edges of $\tree_{\pri}$ by $\mathcal{E}(\tree_{\pri})$ and the set of vertices by $\mathcal{V}(\tree_{\pri}).$
\end{mdef}
Define the \emph{standard vertex} $v_*$ to be the vertex corresponding to the lattice $\roi_{\pri}\oplus\roi_{\pri}$, and the \emph{standard edge} $e_*$ to be the edge joining $v_*$ to the vertex corresponding to $\roi_{\pri}\oplus\pi_{\pri}\roi_{\pri}$. There is a natural notion of \emph{distance} between two vertices $v$ and $v'$, and we say a vertex $v$ is \emph{even} (respectively \emph{odd}) if the distance between $v$ and $v_*$ is even (respectively odd). Each (oriented) edge has a source and a target vertex, and we say such an edge is even if its source is. Write $\mathcal{V}^+(\tree_{\pri})$ and $\mathcal{E}^+(\tree_{\pri})$ (respectively $\mathcal{V}^-(\tree_{\pri})$ and $\mathcal{E}^-(\tree_{\pri})$) for the set of even (respectively odd) vertices and edges respectively.
$\\ \\$
There is a natural transitive action of $\mathrm{PGL}_2(F_{\pri})$ on $\tree_{\pri}$. We can extend this action to a larger group.
\begin{mdef}\label{group omega}
Let $\m$ be an ideal of $\roi_F$ coprime to $\pri$.
\begin{itemize}
\item[(i)] For $v$ a finite place of $F$, define
\[R_0(\m)_v \defeq \left\{\matr\in\mathrm{M}_2(\roi_v): c \equiv 0 \newmod{\m}\right\}.\]
\item[(ii)] Let $R = R_0(\m) \defeq \left\{\gamma \in \mathrm{M}_2\left(\A_F^f\right): \gamma_{v} \in R_0(\m)_{v}\text{ for }v\neq \pri, \gamma_{\pri}\in\mathrm{M}_2(F_{\pri}) \right\}.$
\item[(iii)]Let $\widetilde{\Omega}$ denote the image of $R^\times$ in $\mathrm{PGL}_2\left(\A_F^f\right)$.
\item[(iv)] Finally, let $\Omega \subset \widetilde{\Omega}$ denote the intersection $\Omega \defeq \mathrm{PGL}_2^+\left(\A_F^f\right)\cap\widetilde{\Omega},$ where
\[\mathrm{PGL}_2^+\left(\A_F^f\right) \defeq \left\{\gamma \in \mathrm{PGL}_2\left(\A_F^f\right):v_{\pri}(\det(\gamma_{\pri})) \equiv 0\newmod{2} \text{ for all }\pri|p\right\}.\]
\end{itemize}
\end{mdef}
These groups act on $\tree_{\pri}$ via projection to $\mathrm{PGL}_2(F_{\pri})$.
\begin{mprop}\begin{itemize}
\item[(i)]The group $\widetilde{\Omega}$ acts transitively on the sets $\mathcal{V}(\tree_{\pri})$ and $\mathcal{E}(\tree_{\pri})$.
\item[(ii)]
The group $\Omega$ acts transitively on the sets $\mathcal{E}^\pm(\tree_{\pri})$ and $\mathcal{V}^\pm(\tree_{\pri})$.
\end{itemize}
\end{mprop}
\begin{proof}
See \cite{Ser80}, Theorem 2 of Chapter II.1.4.
\end{proof}
A key reason for introducing the Bruhat--Tits tree is the conceptually nice description it gives of open sets in the projective line. We have a (left) action of $\mathrm{PGL}_2(F_{\pri})$ on $\Proj(F_{\pri})$ by
\[\matr \cdot x = \frac{b+dx}{a+cx},\]
and then to $e = \gamma e_*\in\mathcal{E}(\tree_{\pri})$, we attach the open set
\[
U(e) \defeq \gamma^{-1}(\roi_\pri) = \left\{x\in\Proj(F_{\pri}): \gamma x \in \roi_{\pri}\right\} \subset \Proj(F_{\pri}).
\]
\begin{longversion}
\begin{mrem}
The authors apologise for the regrettable nature of this action. In this instance it would be much nicer to consider, as in \cite{Dar01} and \cite{Ort04}, the natural action by Mobius transformations. However, much the theory of modular symbols -- including the theory over imaginary quadratic fields in \cite{Wil17} -- has been developed using this `twisted' action. Since we are unable to maintain notation that is compatible with both approaches, we have opted for this twisted action to avoid conflict when we later consider (classical and overconvergent) modular symbols on the tree. The reader should also note that our convention that $U(e) = \gamma^{-1}\roi_{\pri}$ is inverse to that of Darmon, who defines it to be $\gamma\roi_{\pri}$. Again, this difference is down to the change of action.
\end{mrem}
\end{longversion}
\begin{shortversion}
\begin{mrem}
The authors apologise for the regrettable nature of this action, which has been chosen to be consistent with \cite{Wil17}, whilst unavoidably differing from the more natural definition of \cite{Dar01} and \cite{Ort04}. (In the sequel, $\GLt(F)$ will act on the cusps $\Proj(F)$ in this more natural way).
\end{mrem}
\end{shortversion}
\begin{mprop}\label{basis of opens}
As $e$ ranges over all edges in $\mathcal{E}(\tree_{\pri})$, the sets $U(e)$ form a basis of open sets in $\Proj(F_{\pri})$.
\end{mprop}
In the sequel, we'll use the theory of `modular symbols on the tree' and this fact to construct distributions on $\Proj(F_{\pri})$ that allow us to define analogues of Darmon's double integrals in this setting.
\subsection{Bianchi forms on the tree}
Modular forms on the Bruhat--Tits tree were first defined (for classical weight 2 modular forms) in Chapter 1.1 of Darmon's seminal paper \cite{Dar01}, and the concepts are thoroughly motivated and explained in his account. The reader unfamiliar with these concepts is strongly urged to read Darmon's account, which is beautiful and well-written. In this section, we give natural analogues of his results in the imaginary quadratic setting. \\
\\
Let $R$, $\widetilde{\Omega}$ and $\Omega$ be as above.
\begin{mdef}
We say a function $f: \mathcal{E}(\tree_{\pri}) \rightarrow V_{2k+2}(\C)$ is \emph{harmonic} if
\[f(\overline{e}) = -f(e) \hspace{12pt}\forall e\in\mathcal{E}(\tree_{\pri})\]
and if for all vertices $v\in\mathcal{V}(\tree_{\pri})$, we have
\[\sum_{e: s(e) = v}f(e) = \sum_{e: t(e) = v}f(e) = 0.\]
\end{mdef}
Recall the definition $\mathrm{PGL}_2^f(\A_F) \defeq \GLt(\C)\times\mathrm{PGL}_2(\A_F^f)$, and note that $\Omega$ acts on $\mathrm{PGL}_2^f(\A_F)$ by right multiplication. We define Bianchi cusp forms on $\tree_{\pri}$ as follows.
\begin{mdef}
A \emph{cusp form} on $\tree_{\pri}\times\mathrm{PGL}_2^f(\A_F)$ of weight $(k,k)$ for $\Omega$ is a function
\[\mathcal{F}:\mathcal{E}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F) \longrightarrow V_{2k+2}(\C),\]
such that:
\begin{itemize}
\item[(i)] $\mathcal{F}(\gamma e,g\gamma) = \mathcal{F}(e,g)$ for all $\gamma \in \Omega$.
\item[(ii)] $\f$ is harmonic as a function on $\mathcal{E}(\tree_{\pri})$.
\item[(iii)] For each edge $e \in \mathcal{E}(\tree_{\pri})$, the function
\[\mathcal{F}_{e}(g) \defeq \mathcal{F}(e,g)\]
is a cusp form of weight $(k,k)$ and level $\Omega_{e} \defeq \mathrm{Stab}_\Omega(e).$
\end{itemize}
Denote the space of such forms by $\mathbb{S}_{k,k}(\Omega,\tree_{\pri})$.
\end{mdef}
It is natural to ask: for which $f \in S_{k,k}(\Omega_0(\pri\m))$ does there exist some $\f \in \mathbb{S}_{k,k}(\Omega,\tree_{\pri})$ such that $\f_{e_*} = f$? The answer is:
\begin{mthm}\label{new forms tree}
The association $\f \mapsto \f_{e_*}$ defines an isomorphism
\[\mathbb{S}_{k,k}(\Omega,\tree_{\pri}) \cong S_{k,k}(\Omega_0(\pri\m))^{\pri-\mathrm{new}}.\]
\end{mthm}
We prove this in the next section. In order to do so, we introduce the following auxiliary spaces which are easier to work with, since their definition omits the harmonicity condition.
\begin{mdef}
A \emph{cusp form} on $\mathcal{E}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F)$ of weight $(k,k)$ for $\widetilde{\Omega}$ is a function
\[\f:\mathcal{E}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F) \longrightarrow V_{2k+2}(\C)\]
such that:
\begin{itemize}
\item[(i)] $\f(\gamma e,g \gamma) = \mathcal{F}(e,g)$ for all $\gamma \in \widetilde{\Omega}$.
\item[(ii)] For each $e$, the function $\f_{e}(g) \defeq \f(e,g)$ is a cusp form of weight $(k,k)$ and level $\widetilde{\Omega}_{e} \defeq \mathrm{Stab}_{\widetilde{\Omega}}(e).$
\end{itemize}
Denote the space of such forms by $\mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri}).$
\end{mdef}
Let $f \in S_{k,k}(\Omega_0(\pri\m))$ be a Bianchi newform, and note that $f$ gives rise to a form $\f$ on $\mathcal{E}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F)$ in a natural way. Indeed, since $\widetilde{\Omega}$ acts transitively on $\mathcal{E}(\tree_{\pri})$, any form on $\mathcal{E}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F)$ is uniquely determined by its restriction to $\{e\}\times\mathrm{PGL}_2^f(\A_F)$, where $e$ is any element of $\mathcal{E}(\tree_{\pri})$.\\
\\
An explicit check shows that $\mathrm{Stab}_{\widetilde{\Omega}}(e_*) = \Omega_0(\pri\m).$ Hence, writing $\f(e_*,g) = f(z,t)$ and extending $\f$ to all of $\mathcal{E}(\tree_{\pri})$ using property (i), we get a well-defined map
\[S_{k,k}(\Omega_0(\pri\m)) \longrightarrow \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri}).\]
This map has a natural inverse given by $\f \longmapsto \f_{e_*}.$ We see that we have proved:
\begin{mprop}\label{iso edges}
There is a natural isomorphism $\mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri})\cong S_{k,k}(\Omega_0(\pri\m))$ given by $\f \mapsto \f_{e_*}.$
\end{mprop}
We also need the notion of degeneracy maps at a prime $\pri$. To define these, we use the following spaces.
\begin{mdef}
A \emph{cusp form} on $\mathcal{V}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F)$ of weight $(k,k)$ for $\widetilde{\Omega}$ is a function
\[\f:\mathcal{V}(\tree_{\pri})\times\mathrm{PGL}_2^f(\A_F) \longrightarrow V_{2k+2}(\C)\]
such that:
\begin{itemize}
\item[(i)] $\f(\gamma v,g \gamma) = \f(v,g)$ for all $\gamma \in \widetilde{\Omega}$.
\item[(ii)] For each $v\in\mathcal{V}(\tree_{\pri})$, the function $\f_{v}(g) \defeq \f(v,g)$ is a cusp form of weight $(k,k)$ and level $\widetilde{\Omega}_{v} \defeq \mathrm{Stab}_{\widetilde{\Omega}}(v).$
\end{itemize}
Denote the space of such forms by $\mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{V}_{\pri}).$
\end{mdef}
Using the same ideas as above, replacing $e_*$ with $v_*$ and noting that $\mathrm{Stab}_{\widetilde{\Omega}}(v_*) = \Omega_0(\m),$ we get:
\begin{mprop}\label{iso vertices}
There is a natural isomorphism $\mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{V}_{\pri}) \cong S_{k,k}(\Omega_0(\m))$ given by $\f \mapsto \f_{v_*}.$
\end{mprop}
\begin{longversion}
\begin{mrem}
The lack of a harmonicity condition, which we have dropped in the definitions above, makes the proof of Propositions \ref{iso edges} and \ref{iso vertices} very straightforward. This harmonicity condition will be crucial in the sequel, however, since we will use forms on the tree to define systems of distributions, and these will only be well-defined when the form is harmonic. Hence we really do need the more difficult Theorem \ref{new forms tree} to proceed.
\end{mrem}
\end{longversion}
\subsection{Proof of Theorem \ref{new forms tree}}
We now turn to the proof of Theorem \ref{new forms tree}, which we will prove by setting up two compatible exact sequences; one at the level of forms on the tree, and one at the level of classical forms. First, we focus on the tree.
\begin{mdef}\label{alpha}Let $\alpha \in \mathrm{PGL}_2(F_{\pri})\backslash\mathrm{PSL}_2(F_{\pri})$ be an element of the normaliser of $R_0(\pri)_{\pri}$, so that $\alpha e_* = \overline{e_*},$ and note that $\alpha$ switches the parity of the edges and vertices of $\tree_{\pri}$.
\end{mdef}
\begin{mdef}Define a homomorphism $i : \mathbb{S}_{k,k}(\Omega,\tree_{\pri}) \longrightarrow \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri})$ by
\[i(\f)(e,g) = \left\{\begin{array}{ll}\f(e,g) &: e\text{ even},\\
\f(\alpha e,g\alpha) &: e\text{ odd}.
\end{array}
\right.
\]
\end{mdef}
Note that by definition we've extended the $\Omega$-invariance to $\widetilde{\Omega}$-invariance and that $i$ is injective by the harmonicity condition.
\begin{mdef}
Define two \emph{degeneracy maps} $\pi_s,\pi_t: \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri})\longrightarrow \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{V}_{\pri})$ by setting
\begin{longversion}\begin{align*}\pi_s(\f)(v,g) &\defeq \sum_{e \in \mathcal{E}(\tree_{\pri}): s(e) = v}\f(e,g),\\
\pi_t(\f)(v,g) &\defeq \sum_{e \in \mathcal{E}(\tree_{\pri}): t(e) = v}\f(e,g).
\end{align*}\end{longversion}
\begin{shortversion}
\[\pi_s(\f)(v,g) \defeq \sum_{e \in \mathcal{E}(\tree_{\pri}): s(e) = v}\f(e,g)\]
(and similarly for $\pi_t$ with $t(e)$ replacing $s(e)$).
\end{shortversion}
\end{mdef}
The kernel of $\pi_s\oplus\pi_t$ is precisely the image of $\mathbb{S}_{k,k}(\Omega,\tree_{\pri})$ under the map $i$. To see this, note that $\mathrm{Im}(i) \subset \mathrm{Ker}(\pi_s\oplus\pi_t)$ from the definition of harmonicity; conversely, suppose $\tilde{f}$ is an element of the kernel. We can directly construct $f \in \mathbb{S}_{k,k}(\Omega,\tree_{\pri})$ with $i(f) = \tilde{f}$ by defining $f(e,g) = \tilde{f}(e,g)$ for $e$ even and extending to odd edges using harmonicity. We see that we get an exact sequence
\[0 \longrightarrow \mathbb{S}_{k,k}(\Omega,\tree_{\pri}) \longrightarrow \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri}) \labelrightarrow{\pi_s\oplus\pi_t} \mathbb{S}_{k,k}(\widetilde{\Omega}, \mathcal{V}_{\pri})\oplus \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{V}_{\pri}).
\]
There are classical analogues of the degeneracy maps above. Write $\Omega_0(\m) = \coprod_{i\in I_{\pri}}\Omega_0(\pri\m)\gamma_i$ for a system of coset representatives $\gamma_i \in \Omega_0(\m)$. Then define:
\begin{mdef}
Let $\varphi_s$ and $\varphi_t$ denote the degeneracy maps $\varphi_s,\varphi_t: S_{k,k}(\Omega_0(\pri\m)) \longrightarrow S_{k,k}\left(\Omega_0(\m)\right)$ defined by
\begin{align*}\varphi_s(f)(g) &\defeq \sum_{i\in I_{\pri}}f(g\gamma_i),\\
\varphi_t(f)(g) &\defeq \sum_{i\in I_{\pri}}f(g\gamma_i\alpha).
\end{align*}
\end{mdef}
\begin{mdef}
The subspace of $S_{k,k}(\Omega_0(\pri\m))$ of \emph{$\pri$-new forms} is the kernel of the map
\[S_{k,k}(\Omega_0(\pri\m)) \labelrightarrow{\varphi_s\oplus\varphi_t} S_{k,k}(\Omega_0(\m))\oplus S_{k,k}(\Omega_0(\m)).\]
\end{mdef}
\subsubsection{Relating the exact sequences}
\begin{longversion}The key step of the proof of Theorem \ref{new forms tree} is the following.\end{longversion}
\begin{mlem}
The following diagram of exact sequences commutes:
\[\begin{diagram}
0 &\rTo& \mathbb{S}_{k,k}(\Omega,\tree_{\pri}) &\rTo^{i}& \mathbb{S}_{k,k}(\widetilde{\Omega},\mathcal{E}_{\pri}) &&\rTo^{\pi_s\oplus\pi_t}&& \mathbb{S}_{k,k}(\widetilde{\Omega}, \mathcal{V}_{\pri})^2 \\
&&\dTo&&\dTo&&&&\dTo\\
0 &\rTo& S_{k,k}(\Omega_0(\pri\m))^{p-\mathrm{new}} &\rTo& S_{k,k}(\Omega_0(\pri\m)) &&\rTo^{\varphi_s\oplus\varphi_t}&& S_{k,k}(\Omega_0(\m))^2 \\
\end{diagram},
\]
where the vertical arrows are the maps of the form $\f \mapsto \f_{e_*}$ (for the first two) and $\f \mapsto \f_{v_*}$ (for the third).
\end{mlem}
\begin{proof}
We've already shown that the sequences are exact. The first square commutes by the definition of $i$, since $e_*$ is even, and hence $[i(\f)]_{e_*} = \f_{e_*}$. The second square commutes since the sets
\[\{\gamma_i e_* : i \in I_{\pri}\} \hspace{6pt}\text{and}\hspace{6pt} \{\gamma_i \alpha e_* : i \in I_{\pri}\}\]
form a complete set of edges with source and target $v_*$ in $\tree_{\pri}$.
\end{proof}
Theorem \ref{new forms tree} now follows from the 5-lemma applied to this diagram in conjunction with Propositions \ref{iso edges} and \ref{iso vertices}.
\section{The $p$-adic $L$-function of a Bianchi modular form}
We briefly recap the results of \cite{Wil17}, which will be used heavily in the sequel. For the rest of the paper, we will make the following simplifying assumption (which was \emph{not} made in \cite{Wil17}):
\begin{mass}
The imaginary quadratic field $F$ has class number 1.
\end{mass}
\subsection{Bianchi modular symbols}\label{BMF}
Define
\[\Gamma_0(\n) \defeq \SLt(F) \cap [\GLt(\C)\Omega_0(\n)].\]
\begin{mdef}
Let $\Delta_0 \defeq \mathrm{Div}^0(\Proj(F))$ denote the space of `paths between cusps' in $\uhs$, and let $V$ be any right $\SLt(F)$-module. For a subgroup $\Gamma\subset\SLt(F)$, denote the space of \emph{$V$-valued modular symbols for $\Gamma$} to be the space
\[\symb_{\Gamma}(V) \defeq \mathrm{Hom}_\Gamma(\Delta_0,V)\]
of functions satisfying the $\Gamma$-invariance property that
\[(\phi|\gamma)(D)\defeq \phi(\gamma D)|\gamma = \phi(D)\hspace{12pt} \forall D\in\Delta_0, \gamma\in\Gamma,\]
where $\Gamma$ acts on the cusps by $\smallmatrd{a}{b}{c}{d}\cdot r = (ar+b)/(cr+d).$ For $r,s \in \Proj(F)$, to ease notation we will henceforth write $\phi\{r-s\}$ for $\phi(\{r\}-\{s\})$.
\end{mdef}
\begin{mdef}For a ring $R$, let $V_{k,k}(R) \defeq V_k(R)\otimes_{R}V_k(R)$. (We think of $V_{k,k}$ as polynomials on $\roi_F\otimes_{\Z}\Zp$ that have degree at most $k$ in each variable).
\end{mdef}
This space has a natural left action of $\GLt(R)^2$ induced by the action of $\GLt(R)$ on each factor by
\[\matr\cdot P(z) = \frac{(a+cz)^k}{(ad-bc)^{k/2}} P\left(\frac{b+dz}{a+cz}\right),\]
which is well-defined since we took $k$ to be even. This induces a right action on the dual space $V_{k,k}(R)^* \defeq \mathrm{Hom}(V_{k,k}(R),R).$
\begin{mrems}
\begin{itemize}
\item[(i)]Note the factor of the determinant, which was not needed in \cite{Wil17}; this ensures that the centre of $\GLt(R)$ acts trivially. This difference means that the Hecke operators we consider in this paper are scalar multiples of those in \cite{Wil17}.
\item[(ii)]We can see any subgroup of $\GLt(F)$ as acting on $V_{k,k}(\C)$ via the natural embedding of $F\hookrightarrow \C\times\C$, where the first factor differs from the second by complex conjugation. In particular, we obtain an action of $\Gamma_0(\n)$ on $V_{k,k}(\C)$.
\end{itemize}
\end{mrems}
\begin{longversion}
This also gives a natural Hecke action on the space $\symb_{\Gamma_0(\n)}(V_{k,k}(\C)^*)$, defined essentially as in Definition \ref{Hecke}. Now and for the rest of the paper, we fix more concrete choices of $\delta_{j}^{I}$.
\end{longversion}
\begin{mdef}\label{delta}
Choose $\beta_I$ to be a generator of the ideal $I$, then choose $\delta_{j}^I \in \GLt(F)\cap \mathrm{M}_2(\roi_F)$ such that
\[\Gamma_0(\n)\matrd{1}{0}{0}{\beta_I}\Gamma_0(\n) = \coprod_j\Gamma_0(\n)\delta_j^{I}.\]
Then define the Hecke operator $T_I$ by
\[(\phi|T_I)\{r-s\}(P) \defeq N(I)^{k/2}\sum_j \phi\{\delta_j^I r - \delta_j^I s\}(\delta_j^I\cdot P).\]
We write $U_I$ if $I$ is not coprime to $\n$. In the case $I = \pri = (\pi)$, we can choose the representatives to be $\smallmatrd{1}{a}{0}{\pi}$ as $a$ ranges over classes $\newmod{\pri}$.
\end{mdef}
\begin{mprop}\label{attach modular symbol}Let $f \in S_{k,k}(\Omega_0(\n))$ be a cuspidal Bianchi modular form.
\begin{itemize}
\item[(i)] To $f$, one can associate a canonical element $\widetilde{\phi}_{f} \in \symb_{\Gamma_0(\n)}(V_{k,k}(\C)^*)$ in a way that is equivariant with respect to the action of Hecke operators.
\item[(ii)] The symbol $\phi_f \defeq \widetilde{\phi}_{f}/\Omega_f$, for $\Omega_f$ as in Proposition \ref{period}, takes values in $V_{k,k}(E)^*$ for some number field $E$.
\end{itemize}
\end{mprop}
\begin{longversion}
\begin{mrem}For the reader that is unfamiliar with modular symbols, it helps to keep the following mantra in mind. Bianchi forms are inherently analytic in nature, whilst the definition of modular symbols above is purely algebraic. The above proposition allows us to attach a modular symbol to a Bianchi eigenform in a way that retains the data of its Hecke eigenvalues, and it is often easier to study such systems of eigenvalues via this algebraic interpretation, particularly for the purposes of $p$-adic variation.
\end{mrem}
\end{longversion}
\subsection{Modular symbols and $L$-values}
There is an explicit link between critical $L$-values and modular symbols. Recall the notation from Section \ref{notation}. The following is a special case of Theorem 2.11 of \cite{Wil17}.
\begin{mthm}Let $f \in S_{k,k}(\Omega_0(\n))$ be a cuspidal Bianchi form with associated modular symbol $\phi_f$. Let $\chi$ be a finite order Hecke character of conductor $(c)$ and let $0\leq r\leq k$. Then
\begin{equation}\label{integralformula}
\Lambda(f,\chi,r+1) = \left[\frac{(-1)^{k}2\Omega_f}{D^{r+1}\tau(\chi^{-1})|c|^r \#\roi_F^\times}\right]\sum_{a \in (\roi_F/c)^\times} \chi(a)C_{r}(a/c),\end{equation}
where
\[C_{r}(a/c) \defeq \phi_{f}\{a/c - \infty\}((cz+a)^r(\overline{c}\zbar + \overline{a})^r).\]
Here $\tau(\chi^{-1})$ is the Gauss sum attached to $\chi^{-1}$ (see \cite{Wil17}, Section 1.2.3).
\end{mthm}
\begin{longversion}
\begin{mrem}There is a neater version of this theorem in \cite{BW16} in the case of general number fields using the theory of automorphic cycles. We use this version as it is considerably more explicit.
\end{mrem}
\end{longversion}
\subsection{Overconvergent modular symbols}
Part (ii) of Proposition \ref{attach modular symbol} allows us to see the modular symbols $\phi_{f}$ as having values in $V_{k,k}(L)^*$ for a sufficiently large $p$-adic field $L$. For suitable level groups, one can then replace this space of polynomials with a space of $p$-adic distributions to obtain \emph{overconvergent modular symbols}.
\begin{mdef}Let $\mathcal{A}(L)$ denote the space of locally analytic functions on $\roi_F\otimes_{\Z}\Zp$ defined over $L$. We equip this space with a weight $(k,k)$-action of the semigroup
\[\Sigma_0(p) \defeq \left\{\matr \in M_2(\roi_F\otimes_{\Z}\Zp): p|c, a\in(\roi_F\otimes_{\Z}\Zp)^\times, ad-bc \neq 0\right\}\]
by setting
\[\gamma\cdot \zeta (z) = \frac{(a+cz)^k}{(ad-bc)^{k/2}}\zeta\left(\frac{b+dz}{a+cz}\right).\]
\end{mdef}
\begin{mdef}
Let $\mathcal{D}_{k,k}(L) \defeq \mathrm{Hom}_{\mathrm{cts}}(\mathcal{A}(L),L)$ denote the space of locally analytic distributions on $\roi_F\otimes_{\Z}\Zp$ defined over $L$, equipped with a weight $(k,k)$ right action of $\Sigma_0(p)$ given by $\mu|\gamma(\zeta) = \mu(\gamma\cdot\zeta).$ For $\Gamma\subset \Sigma_0(p)$, define the space of \emph{overconvergent modular symbols of weight $(k,k)$ and level $\Gamma$} to be $\symb_{\Gamma}(\mathcal{D}_{k,k}(L)).$
\end{mdef}
There is a natural map $\mathcal{D}_{k,k}(L) \rightarrow V_{k,k}(L)^*$ given by dualising the inclusion of $V_{k,k}(L)$ into $\mathcal{A}(L)$. This induces a \emph{specialisation map}
\[\rho:\symb_{\Gamma_0(\n)}(\mathcal{D}_{k,k}(L)) \longrightarrow \symb_{\Gamma_0(\n)}(V_{k,k}(L)^*),\]
noting that the source is well-defined since $\Gamma_0(\n)\subset \Sigma_0(p)$.
\begin{mthm}[Control theorem]\label{control theorem}
For each prime $\pri$ above $p$, let $\lambda_{\pri}\in L^\times$. If $v(\lambda_{\pri}) < (k+1)/e_{\pri}$ for all $\pri|p$, then the restriction of the specialisation map
\[\rho:\symb_{\Gamma_0(\n)}(\mathcal{D}_{k,k}(L))^{\{U_{\pri}=\lambda_{\pri}:\pri|p\}} \isorightarrow \symb_{\Gamma_0(\n)}(V_{k,k}(L)^*)^{\{U_{\pri}=\lambda_{\pri}:\pri|p\}}\]
to the simultaneous $\lambda_{\pri}$-eigenspaces of the $U_{\pri}$ operators is an isomorphism. Here recall that $e_{\pri}$ is the ramification index of $\pri|p$.
\end{mthm}
\begin{mdef}If $f \in S_{k,k}(\Omega_0(\n))$ is an eigenform with eigenvalues $\lambda_I$, we say $f$ has \emph{small slope} if $v(\lambda_{\pri})<(k+1)/e_{\pri}$ for all $\pri|p$.
\end{mdef}
Thus if $f$ is small slope, using the above control theorem, we get an associated overconvergent modular symbol $\Psi_f \in \symb_{\Gamma_0(\n)}(\mathcal{D}_{k,k}(L))$ by lifting the corresponding classical modular symbol.
\subsection{The $p$-adic $L$-function of a Bianchi modular form}\label{padic lfunction section}
In \cite{Wil17}, the $p$-adic $L$-function of a Bianchi modular form $f$ is defined as a locally analytic distribution on $\cl_F(p^\infty)$ that interpolates the classical $L$-values of $f$. For our purposes, it is more convenient to use a different description in terms of analytic functions on $\roi_F\otimes_{\Z}\Zp$.
First, we need a slight extension of the results of \cite{Wil17}. Let $\mathfrak{g}$ be any ideal coprime to $(p)$. An explicit study of $\cl_F(\mathfrak{g}p^\infty)$ shows that
\[\cl_F(\mathfrak{g}p^\infty) = \left[(\roi_F/\mathfrak{g})^\times\times(\roi_F\otimes_{\Z}\Zp)^\times\right]/\roi_F^\times.\]
Let $f \in S_{k,k}(\Omega_0(\n))$ be a small slope Bianchi eigenform with associated overconvergent modular symbol $\Psi_f$. Define a distribution $\mu'_{a\newmod{\mathfrak{g}}}$ on $\{[a]\}\times (\roi_F\otimes_{\Z}\Zp) \subset (\roi_F/\mathfrak{g})^\times\times(\roi_F\otimes_{\Z}\Zp)$, which we see as a copy of $\roi_F\otimes_{\Z}\Zp$, by setting
\[\mu'_{a\newmod{\mathfrak{g}}} \defeq (g\overline{g})^{k/2}\left[\Psi_f\left|\matrd{1}{b}{0}{g}\right.\right]\{0-\infty\},\]
where $b\in \roi_F$ is some lift of $a\newmod{\mathfrak{g}}$ and $\mathfrak{g} = g\roi_F$. This is easily seen to be independent of the choices of $b$ and $g$. Combining these for all $a \in (\roi_F/\mathfrak{g})^\times$, we get a distribution $\mu_p$ on $(\roi_F/\mathfrak{g})^\times \times(\roi_F\otimes_{\Z}\Zp)$. Restricting to units in the second variable, and then restricting to functions that are invariant under $\roi_F^\times$, we obtain a distribution on $\cl_F(\mathfrak{g}p^\infty)$ which in an abuse of notation we will also call $\mu_p$.
\begin{mdef}\label{padiclfunctiondefinition}
For the rest of the paper, fix a choice of $p$-adic logarithm
\[\log_p : \Cp^\times \longrightarrow \Cp\]
with $\log_p(p) = 0$. Also, for $\pri|p$, let $r_{\pri}$ denote the smallest positive integer such that the usual $p$-adic exponential map converges on $\pri^{r_{\pri}}\roi_{\pri}$. (Note that $r_{\pri} = 1$ unless $p = 2$ or $p$ is ramified).
\begin{itemize}
\item[(i)]
Let $\pri|p$ and $s \in \roi_{\pri},$ let $\langle\cdot\rangle : \roi_{\pri}^\times \rightarrow 1+{\pri}^{r_{\pri}}\roi_{\pri}$ denote projection, and define
\begin{align*}
\langle\cdot\rangle^s : \roi_{\pri}^\times &\longrightarrow \Cp,\\
z &\longmapsto \mathrm{exp}(s\cdot\log_p(\langle z \rangle)),
\end{align*}
which is well-defined by the definition of $r_{\pri}$.
\item[(ii)] Similarly, for $\mathbf{s} = (s_{\pri})_{\pri|p} \in\roi_F\otimes_{\Z}\Zp \cong \prod_{\pri|p}\roi_{\pri}$ define
\[\langle\cdot\rangle^{\mathbf{s}} \defeq \prod_{\pri|p}\langle\cdot\rangle^{s_{\pri}} : (\roi_F\otimes_{\Z}\Zp)^\times \longrightarrow \Cp,\]
and note that this is invariant under $\roi_F^\times$, so that it induces a map $\langle\cdot\rangle^{\mathbf{s}} : \cl_F(p^\infty) \rightarrow \Cp$.
\item[(iii)]
Let $\chi$ be a finite order Hecke character of conductor $\mathfrak{g}\ff$, where $\mathfrak{g}$ is coprime to $(p)$ and $\ff|p^\infty$ (so that $\chi$ can naturally be seen as a finite order character on $\cl_F(\mathfrak{g}p^\infty)$). Define an analytic function on $\roi_F\otimes_{\Z}\Zp$ by
\[L_p(f,\chi,\mathbf{s}) = \int_{\cl_F(\mathfrak{g}p^\infty)}\langle\mathbf{z}_p\rangle^{\mathbf{s}}\chi(\mathbf{z})d\mu_p(\mathbf{z}),\]
where $\mathbf{s} = (s_{\pri})_{\pri|p} \in \roi_F\otimes_{\Z}\Zp$, and $\mathbf{z}_p =(z_{\pri})_{\pri|p}$ is the projection of $\mathbf{z}$ to $\cl_F(p^\infty)$.
\item[(iv)] Let $w_{\mathrm{Tm},\pri}: \roi_{\pri}^\times \rightarrow (\roi_{\pri}/\pri^{r_{\pri}})^\times \subset \roi_{\pri}^\times$ denote the Teichm\"{u}ller character at $\pri$, so that for $x\in\roi_{\pri}^\times,$ we have $x = w_{\mathrm{Tm},\pri}(x)\langle x \rangle.$ Also let $w_{\mathrm{Tm}} \defeq \prod_{\pri|p}w_{\mathrm{Tm},\pri}$ be the corresponding character of $(\roi_F\otimes_{\Z}\Zp)^\times.$
\end{itemize}
\end{mdef}
Translated into this setting, the main result of \cite{Wil17} was the following:
\begin{mthm}\label{existence of p-adic L-function}
For a finite order Hecke character $\chi$ with conductor $\mathfrak{g}\ff = (c)$, where $\ff|p^\infty$, and for $0\leq r\leq k$, we have
\[L_p(f,\chiw_{\mathrm{Tm}}^r,\mathbf{r}) = \left(\prod_{\pri|p}Z_{\pri}(\chi,r)\right)\left[\frac{D^{r+1}\tau(\chi^{-1})|c|^{r} \#\roi_F^\times}{(-1)^{k}2\lambda_{\ff}\Omega_{f}}\right]\Lambda(f,\chi,r+1),\]
where $\mathbf{r} = (r)_{\pri|p}$ and
\[Z_{\pri}(\chi,r) \defeq 1-\frac{\chi(\pri)N(\pri)^{r}}{\lambda_{\pri}}
\]
recalling that $\chi(\pri)$ is defined to be $\chi(\pi_{\pri})$ (for $\pi_{\pri} \in \roi_{\pri}$ a uniformiser) if $\chi$ is unramified at $\pri$ and $0$ otherwise (so that $Z_{\pri} = 1$ if $\chi$ ramifies at $\pri$). We call $L_p$ the \emph{$p$-adic $L$-function of $f$}.
\end{mthm}
\begin{proof}\emph{(Sketch).} For $\mathfrak{g} = 1$, this is proved entirely in \cite{Wil17}, so we indicate briefly how the more general result follows. Letting $b$ be a unit$\newmod{\mathfrak{g}\ff}$ and using the notation and arguments \emph{op.\ cit.,} Section 7.1, we see that
\begin{align*}\mu_p(P^{r,r}_{b,\mathfrak{g}\ff}) &= (g\overline{g})^{k/2}\Psi_f\left|\matrd{1}{b}{0}{g}\{0-\infty\}(P^{r,r}_{b,\ff})\right.\\
&= \lambda_{\ff}^{-1}(g\alpha)^r(\overline{g\alpha})^r c_{r,r}\left(\frac{b}{g\alpha}\right),
\end{align*}
where $\ff = (\alpha)$. The rest of the proof proceeds as in \cite{Wil17}.
\end{proof}
\begin{mrems}
\begin{itemize}
\item[(i)] In \cite{Wil17}, this corresponds to the evaluation of $\mu_p$ at the character $\varphi_{p-\mathrm{fin}}$, where $\varphi = \chi|\cdot|^r$. We need the Teichm\"{u}ller character since $\varphi_{p-\mathrm{fin}}(\mathbf{z}) = \chi(\mathbf{z})[\langle \mathbf{z}_p\ranglew_{\mathrm{Tm}}(\mathbf{z}_p)]^r.$
\item[(ii)]
There is a slight error in \cite{Wil17}, where the term $Z_{\pri}$ is incorrect in the case where $\pri\nmid \ff$. This stems from a subtlety in the definition of Gauss sums. \begin{shortversion}In particular, the proof quotes results from \cite{BW16}, and the Gauss sums in that paper are normalised differently. The difference is the factor of $\varphi_{p-\mathrm{fin}}$.\end{shortversion}
\begin{longversion}In particular, the proof quotes the corresponding (more general) result in \cite{BW16}, which uses an adelic Gauss sum defined in terms of Deligne's epsilon factors, whilst in \cite{Wil17}, the Gauss sum is defined in global terms, and it turns out that these are equal only after multiplying by a suitable scaling factor. To illustrate this, consider the following example over $\Q$; let $\chi$ be a Dirichlet character of conductor $pq$. Then the `global' Gauss sum is defined as $\tau_g(\chi) \defeq \sum_{a \newmod{pq}}\chi(a)e^{2\pi i a/pq}$, whilst the `adelic' one is defined as $\tau_a(\chi) \defeq \sum_{a \newmod{pq}}\chi(a)e^{2\pi i (a/p + a/q)} = \chi(p+q)^{-1}\tau_g(\chi).$ In the class number 1 case over imaginary quadratic fields, the result on the exceptional factor is to precisely cancel the factor of $\varphi_{p-\mathrm{fin}}(\pi_{\pri})$ from $Z_{\pri}$, leaving the formulation above. (Regrettably, there is another discrepancy between \cite{Wil17} and \cite{BW16} in that the Gauss sum of a character in the former relates to the Gauss sum of the inverse character in the latter. Throughout this paper we use the normalisations of the former).
\item[(iii)]
For $\mathfrak{g} = 1$, the interpolation property proved in \cite{Wil17} is more general, allowing characters of infinity type $(q,r)$ where both $0 \leq q,r \leq k$. For our purposes it suffices to only consider characters of the form $\chi|\cdot|^r$ as in the theorem above.
\end{longversion}
\end{itemize}
\end{mrems}
\section{Modular symbols on the tree}
\label{modular symbols on the tree}
\subsection{Harmonic cocycles attached to Bianchi modular forms}
Working with class number 1 allows us to work classically with forms on the upper half-space, for which we'll need a classical version of the (adelic) group $\Omega$ introduced in Definition \ref{group omega}.
\begin{mdef}\label{group gamma}
Define
\[\Gamma \defeq \mathrm{PGL}_2(F)\cap \Omega.\]
The group $\Gamma$ has an explicit description as the subgroup of $\mathrm{PGL}_2(\roi_F[\pri^{-1}])$ with lower left entry contained in $\pri^{r}\m$, for some (possibly negative) $r\in\Z$, whose determinant is exactly divisible by an even power of $\pri$.
\end{mdef}
Let $f \in S_{k,k}(\Omega_0(\n))$ be a Bianchi modular form that is new at $\pri$, and let $\f$ be the associated form on the tree $\tree_{\pri}$. To $\f$ we can associate a `modular symbol' $\kappa_{\f}$ on $\tree_{\pri}$ by setting
\[\kappa_{\f}\{r-s\}(e,P) \defeq \phi_{\f_{e}}\{r-s\}(P),\]
for $e \in \mathcal{E}(\tree_{\pri}), r,s \in \Proj(F_{\pri})$ and $P\in V_{k,k}(L)$, for $L$ a suitably large finite extension of $\Qp$.
\begin{mprop}\label{modular symbol tree properties}
The symbol $\kappa_{\f}$ is harmonic on $\mathcal{E}(\tree_{\pri})$, linear in $r, s$ and $P$ and $\Gamma$-invariant in the sense that
\[\kappa_{\f}\{\gamma r - \gamma s\}(\gamma e, \gamma\cdot P) = \kappa_{\f}\{r-s\}(e,P)\]
for all $\gamma \in \Gamma$.
\end{mprop}
\begin{proof}(Sketch). These properties all follow from the corresponding properties of $\f$ and $\phi_{\f_{e}}$. The only one that is not obvious is the $\Gamma$-invariance property, which follows from a lengthy technical calculation, of which we give a sketch. Firstly, one can obtain a more classical definition of Bianchi modular forms on $\tree_{\pri}$ as functions $\f : \mathcal{E}(\tree_{\pri})\times\uhs \longrightarrow V_{2k+2}(\C)$ such that
\[\f(\gamma e, \gamma(z,t))\left[\binomc{X}{Y}\right] = \f(e, (z,t))\left[j(\gamma,(z,t))\binomc{X}{Y}\right],\]
for all $\gamma = \smallmatrd{a}{b}{c}{d} \in \Gamma$, where we consider $V_{2k+2}(\C)$ instead as homogeneous polynomials in two variables $X$ and $Y$ and
\[j(\gamma,(z,t)) \defeq \matrd{c}{0}{0}{\overline{c}}\matrd{z}{-t}{t}{\overline{z}} + \matrd{d}{0}{0}{\overline{d}}.\]
This is precisely the transformation property that classical Bianchi modular forms satisfy under the corresponding level group. Then one examines the definition of the modular symbol $\phi_{\f_e}$, via a $V_{k,k}(\C)^*$-valued harmonic differential $\delta_{\f_e}$ on $\uhs$ attached to $\f_e$. In \cite{Hid94}, Chapter 2.5, this differential is shown to be invariant under $\Gamma_e$. Define a differential on $\mathcal{E}\times\uhs$ by $\delta(e,(z,t)) \defeq \delta_{\f_e}(z,t)$. Using exactly the same methods as in \cite{Hid94}, and the classical transformation property above, this is seen to be invariant under $\Gamma$, in the sense that $\delta(\gamma e, \gamma(z,t)) = \delta(e,(z,t))$ for all $\gamma \in \Gamma$. Since by definition $\kappa_{\f}\{r-s\}(e,P)$ is the integral of $\delta(e,(z,t))$ from $r$ to $s$ in $\uhs$, we then obtain the result by a simple transformation.
\end{proof}
\begin{mdef}
We call such a function a \emph{$\Gamma$-invariant harmonic cocycle on $\mathcal{E}(\tree_{\pri})$ with values in $\mathrm{Hom}(\Delta_0,V_{k,k}(L)^*).$} We denote the space of such functions by \linebreak[4]$C^{\mathrm{har}}(\mathrm{Hom}(\Delta_0,V_{k,k}(L)^*)^{\Gamma}.$
\end{mdef}
We have a map
\[\rho_{e_*} : C^{\mathrm{har}}(\mathrm{Hom}(\Delta_0,V_{k,k}(L)^*)^{\Gamma} \longrightarrow \symb_{\Gamma_0(\n)}(V_{k,k}(L)^*)\]
given by restriction to the standard edge $e_*$. We can implicitly define an inverse to this map by using the tree. In fact, we can write this down explicitly without needing to refer to the tree at all. Indeed, note that any \emph{even} edge $e$ can be written as $e = \gamma e_*$ for some $\gamma \in \Gamma$. Let $\phi_{f} \in \symb_{\Gamma_0(\n)}(V_{k,k}(L)^*)$ be a modular symbol. If $\kappa_{\f}$ is a harmonic cocycle such that $\rho_{e_*}(\kappa_{\f}) = \phi_{f},$ then by $\Gamma$-invariance we have
\[\kappa_{\f}\{r - s\}(e, P) = \phi_{f}|\gamma^{-1}\{r-s\}(P).\]
It remains to determine the value of $\kappa_{\f}$ at odd edges, which we will do in the next subsection using the \emph{Atkin--Lehner} operators to describe the action of a larger group $\widetilde{\Gamma}.$
\begin{mrem}
We have brushed over a slight subtlety here; we should really have defined $\kappa_{\f}$ in terms of the complex symbols $\widetilde{\phi}_{\f_e}$, rather than the algebraic analogues $\phi_{\f_e}$, each of which depends on a choice of period $\Omega_{\f_e}$. However, the invariance property above shows that we can use the same choice $\Omega_{\f_e} = \Omega_f$ at each edge to obtain something algebraic on the tree.
\end{mrem}
\subsection{Atkin--Lehner operators and the action of $\widetilde{\Gamma}$}
Let $\widetilde{\Gamma} \defeq \mathrm{PGL}_2(F)\cap\widetilde{\Omega}$. Analogously to $\Gamma$, we see $\widetilde{\Gamma}$ as the subgroup of $\mathrm{PGL}_2(\roi_F[\pri^{-1}])$ with lower left entry contained in $\pri^{r}\m$, for some (possibly negative) $r\in\Z$.
\begin{mdef}Let $\alpha$ be any element of $\widetilde{\Gamma}\backslash \Gamma$ normalising $\Gamma_0(\n)$. The \emph{Atkin--Lehner operator} $W_{\pri}$ on $\symb_{\Gamma_0(\n)}(V_{k,k}(L)^*)$ is defined by
\[\phi|W_{\pri} \defeq \phi|\alpha.\]
If $f$ (and hence $\phi_f$) is an eigenform for the Hecke operators, then it is also an eigenform for $W_{\pri}$. Denote the eigenvalue by $-\omega$, which will be $1$ or $-1$.
\end{mdef}
\begin{mprop}\label{transform}
For $\gamma\in\GLt(F)$, define $|\gamma| \defeq \mathrm{ord}_{\pri}(\det(\gamma))$, so that in particular $|\gamma| \newmod{2}$ is well-defined on $\mathrm{PGL}_2(F)$. Then we have
\[\kappa_{\f}\{\gamma r - \gamma s\}(\gamma e, \gamma\cdot P) = \omega^{|\gamma|}\kappa_{\f}\{r-s\}(e,P)\]
for all $\gamma \in \widetilde{\Gamma}.$
\end{mprop}
\begin{proof}
If $\gamma \in \Gamma$, this is the invariance statement above. If $\gamma \notin \Gamma$, then $\gamma = \gamma'\alpha$ for some $\gamma' \in \Gamma$, and then the $\Gamma$-invariance means we're reduced to proving that
\[\kappa_f\{\alpha r -\alpha s\}(\alpha e,\alpha \cdot P) = \omega \kappa_f\{r-s\}(e,P).\]
\begin{longversion}To see this, suppose first that $e$ is even, so that $e = \gamma e_*$ for some $\gamma \in \Gamma$. Then
\begin{align*}\omega\kappa_f\{r-s\}(e,P) &= \omega\kappa_f\{\gamma^{-1}r-\gamma^{-1}s\}(e_*,\gamma^{-1}\cdot P)\\
&= \omega\phi_f\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot P)\\
&= -(\phi_f|W_{\pri})\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot P) = -\phi_f\{\alpha\gamma^{-1}r-\alpha\gamma^{-1}s\}(\alpha\gamma^{-1}\cdot P)\\
&= -\kappa_f\{\alpha\gamma^{-1}r-\alpha\gamma^{-1}s\}(e_*,\alpha\gamma^{-1}\cdot P)\\
&= \kappa_f\{\alpha\gamma^{-1}r-\alpha\gamma^{-1}s\}(\overline{e_*},\alpha\gamma^{-1}\cdot P)\\
&= \kappa_f\{\gamma\alpha\gamma^{-1}r-\gamma\alpha\gamma^{-1}s\}(\overline{e},\gamma\alpha\gamma^{-1}\cdot P)\\
&= \kappa_f\{\gamma\alpha\gamma^{-1}r-\gamma\alpha\gamma^{-1}s\}(\gamma\alpha\gamma^{-1}e,\gamma\alpha\gamma^{-1}\cdot P).
\end{align*}
But $\gamma\alpha\gamma^{-1} \in \widetilde{\Gamma}\backslash\Gamma$, so it is of the form $\gamma'\alpha$ for some $\gamma' \in \Gamma$. The result follows since we can cancel $\gamma'$ by $\Gamma$ invariance \end{longversion}
\begin{shortversion}This is a simple explicit calculation \end{shortversion}(compare \cite{Dar01}, Lemma 1.4).
\begin{longversion}
$\\ \\$
The case where $e$ is odd is similar, using the existence of some $\gamma$ such that $e = \gamma\alpha e_*$. This concludes the proof.
\end{longversion}
\end{proof}
\subsection{Hecke operators}
There are natural actions of Hecke operators away from $\n$ on the forms and symbols on the tree, and all the associations we've made are equivariant with respect to these operators. In particular, recall Definition \ref{delta} where for each ideal $I$, we picked a finite set of representatives $\delta_j^I$ for the double coset corresponding to $T_I$.
\begin{mdef}
Define the Hecke operator $T_I$ on $C^{\mathrm{har}}(\mathrm{Hom}(\Delta_0,V_{k,k}(L)^*)^{\Gamma}$ by setting $\kappa|T_I \defeq N(I)^{k/2}\sum_j\kappa|\delta_j^I,$ or more concretely,
\[(\kappa|T_I)\{r-s\}(e,P) \defeq N(I)^{k/2}\sum_j\kappa\{\delta_j^I r - \delta_j^Is\}(\delta_j^Ie, \delta_j^I\cdot P).\]
\end{mdef}
\begin{mprop}
The association of $\kappa_{\f}$ to $f$ is equivariant with respect to the Hecke operators $T_I$ for $I$ coprime to $\n$.
\end{mprop}
\begin{proof}
This follows immediately from the definitions and the fact that, if $I$ is coprime to $\n$, we have $\delta_j^I e = e$ for all $j$ and for any edge $e$.
\end{proof}
\subsection{Hecke eigenvalues of newforms}
As a corollary of the above results, we recover the following classical result, which also follows from the theory of automorphic forms and the structure of Iwahori--Hecke algebras. The authors thank Aurel Page for pointing out the previous existence of such a result, which we were not able to find in the literature.
\begin{mcor}\label{eigenvalues of newforms}
Let $f\in S_{k,k}(\Omega_0(\n))$ be an eigenform, with $\pri|\n$. If $f$ is new at $\pri$, then the Hecke eigenvalue at $\pri$ is $\omega N(\pri)^{k/2}$, where $-\omega$ is the eigenvalue at $f$ of the Atkin--Lehner operator $W_{\pri}$.
\end{mcor}
\begin{proof}
We've already shown that $f$ is new at $\pri$ if and only if the associated form $\f$ on the tree is harmonic; suppose this is the case. It suffices to prove the result on modular symbols. Recall that $\pi$ is a generator of $\pri$, and for $j \in \roi_F$, let $\delta_j \defeq \smallmatrd{1}{j}{0}{\pi}$. Then by definition, for any $r,s \in \Proj(F)$ and $P\in V_{k,k}(L),$
\begin{align*}(\phi_f|U_{\pri})\{r-s\}(P) &= N(\pri)^{k/2}\sum_{j \newmod{\pri}}\kappa_{\f}\{\delta_j r - \delta_j s\}(e_*, \delta_j\cdot P)\\
&= \omega N(\pri)^{k/2}\sum_{j \newmod{\pri}}\kappa_{\f}\{r-s\}(\delta_j^{-1}e_*, P),
\end{align*}
using the transformation property above. But by inspection, the set $\{\delta_j^{-1}e_*: j \newmod{\pri}\}$ is precisely the set $\{e \in \mathcal{E}(\tree_{\pri}): t(e) = v_*, e\neq \overline{e_*}\}.$
\begin{longversion}Hence we get
\begin{align*}(\phi_f|U_{\pri})\{r-s\}(P) &= \omega N(\pri)^{k/2}\sum_{\substack{t(e) = v_*\e \neq \overline{e_*}}} \kappa_f\{r-s\}(e,P)\\
&= -\omega N(\pri)^{k/2}\kappa_f\{r-s\}(\overline{e_*},P)\\
&= \omega N(\pri)^{k/2}\kappa_f\{r-s\}(e_*, P),
\end{align*}
where the last two equalities follow from harmonicity. This completes the proof.
\end{longversion}
\begin{shortversion}Hence an easy calculation using harmonicity shows that
\[(\phi_f|U_{\pri})\{r-s\}(P) = \omega N(\pri)^{k/2}\kappa_f\{r-s\}(e_*,P),\]
which completes the proof.
\end{shortversion}
\end{proof}
Whilst we have presented this in only the generality in which we require it, an argument of this type generalises very easily to the case of arbitrary number fields and arbitrary class number.
\begin{mcor}\label{exceptional zero existence}
Let $f$ as above be new at $\pri = (\pi)$. Suppose that $f$ has small slope at all primes above $p$, and let $\chi$ be a finite order Hecke character for $F$. Then the critical value $L_p(f,\chiw_{\mathrm{Tm}}^{r},\mathbf{r})$ of the $p$-adic $L$-function has an exceptional zero at $\pri$ precisely when $k$ is even, $\chi(\pi) = \omega$ and $r = k/2$.
\end{mcor}
\begin{proof}
The exceptional factor $Z_{\pri}(\chi,k/2)$ of Theorem \ref{existence of p-adic L-function} is zero if and only if these conditions are satisfied.
\end{proof}
\section{Distributions attached to Bianchi modular forms}
In the rational set-up, overconvergent modular symbols naturally take values in the space of locally analytic distributions on $\Zp$, and Darmon and Orton use the tree to extend these to distributions on $\Proj(\Qp)$. In the Bianchi case, the values are distributions on $\roi_F\otimes_{\Z}\Zp$. In this section we show how to extend such distributions to be `projective at $\pri$'. To do so in general, we'll need to work with the other primes above $p$, but to ease the exposition, we first treat the case where there is a single prime above $p$.
\subsection{Distributions on $\Proj_{\pri}$: the case $p$ inert or ramified}
Suppose $p$ is inert or ramified in $F$. To ensure consistent notation with the split case, let
\[\Proj_{\pri} \defeq \Proj(F_{\pri}).\]
\subsubsection{Polynomial distributions on $\Proj_{\pri}$}
\begin{longversion}
Recall Proposition \ref{basis of opens}, which said that edges of the tree give a basis of open sets in $\Proj_{\pri}$ by associating to $e \in \mathcal{E}(\tree_{\pri})$ the open set $U(e)$. Let $f\in S_{k,k}(\Omega_0(\n))^{\pri-\mathrm{new}}$ be an eigenform with associated form $\f$ and modular symbol $\kappa_{\f}$ on $\tree_{\pri}$. To this, we associate a system of locally polynomial distributions.
\begin{mdef}
Let $\mathcal{V}_{k,k}(\Proj_{\pri},L)$ be the space of functions $\Proj_{\pri} \rightarrow L$ that are locally polynomial of degree at most $k$ in each variable.
\end{mdef}
\begin{mdef}
For each pair $r,s \in \Proj(F)$, define a locally polynomial distribution $\mu_{\f}^{\mathrm{poly}}\{r-s\}$ on $\Proj_{\pri}$ by setting
\[\int_{U(e)}P(t)d\mu_{\f}^{\mathrm{poly}}\{r-s\}(t) \defeq \kappa_{\f}\{r-s\}(e,P),\]
where $P$ is polynomial of degree at most $k$ in each variable on $U(e)$, and extending linearly.
\end{mdef}
Recall that $-\omega$ is the eigenvalue of $f$ at the Atkin--Lehner operator $W_{\pri}$. The transformation property of Proposition \ref{transform} above immediately gives:
\begin{mlem}\label{poly distribution}
Let $e = \gamma e_*$ with $\gamma \in \widetilde{\Gamma}$. Then
\[\int_{U(e)}P(t) d\mu_{\f}^{\mathrm{poly}}\{r-s\}(t) = \omega^{|\gamma|}\phi_f\{\gamma^{-1}r -\gamma^{-1}s\}(\gamma^{-1}\cdot P).\]
\end{mlem}
In particular, we can express the distribution solely in terms of $\phi_f$.
\begin{mrems}
\begin{itemize}
\item[(i)]This distribution is well-defined by the harmonicity property of $\kappa_{\f}$. In particular, constructions of the sort above are why we need the harmonicity condition.
\item[(ii)] This construction, which is directly analogous to the one in \cite{Ort04} for classical modular forms, is included mainly to motivate future constructions of locally analytic distributions. We're really seeing the modular symbol attached to $\f_e$ to take values in `polynomials on $U(e)$.'
\end{itemize}
\end{mrems}
\end{longversion}
\begin{shortversion}By Proposition \ref{basis of opens}, a basis of open sets in $\Proj_{\pri}$ is given by $\{U(e): e \in \mathcal{E}(\tree_{\pri})\}.$ Let $f\in S_{k,k}(\Omega_0(\n))^{\pri-\mathrm{new}}$ be an eigenform with associated form $\f$ and modular symbol $\kappa_{\f}$ on $\tree_{\pri}$. Let $\mathcal{V}_{k,k}(\Proj_{\pri})$ be the space of functions $\Proj_{\pri} \rightarrow L$ that are locally polynomial of degree at most $k$ in each variable.
\begin{mdef}
For each pair $r,s \in \Proj(F)$, define a locally polynomial distribution $\mu_{\f}^{\mathrm{poly}}\{r-s\}$ on $\Proj_{\pri}$ by setting
\[\int_{U(e)}P(t)d\mu_{\f}^{\mathrm{poly}}\{r-s\}(t) \defeq \kappa_{\f}\{r-s\}(e,P),\]
where $P$ is polynomial on $U(e)$ of degree at most $k$ in each variable, and extending linearly. (Note that this is well-defined by harmonicity).
\end{mdef}
Recall that $-\omega$ is the eigenvalue of $f$ at the Atkin--Lehner operator $W_{\pri}$. The transformation property of Proposition \ref{transform} above then shows that if $e = \gamma e_*$ with $\gamma \in \widetilde{\Gamma}$, then
\begin{equation}\label{poly distribution}\int_{U(e)}P(t) d\mu_{\f}^{\mathrm{poly}}\{r-s\}(t) = \omega^{|\gamma|}\phi_f\{\gamma^{-1}r -\gamma^{-1}s\}(\gamma^{-1}\cdot P).
\end{equation}
In particular, we can express the distribution solely in terms of $\phi_f$.
\end{shortversion}
\subsubsection{Locally analytic distributions on $\Proj_{\pri}$}
By Corollary \ref{eigenvalues of newforms}, $f$ has eigenvalue $\omega N(\pri)^{k/2} = \omega p^{k/e_{\pri}}$ at $\pri$, so $f$ has small slope at $p$. Hence we can use Theorem \ref{control theorem} to give a unique overconvergent lift $\Psi_f\in\symb_{\Gamma}(\mathcal{D}_{k,k}(L))$ of $\phi_f$.
\begin{mdef}
Let $\mathcal{A}_{k,k}(\Proj_{\pri},L)$ denote the space of functions $\Proj_{\pri} = \Proj(F_{\pri}) \rightarrow L$ that are locally analytic except perhaps for a pole at $\infty$ of order at most $k$.
\end{mdef}
Recall the action of $\mathrm{PGL}_2(F_{\pri})$ on $\Proj_{\pri}$ by $\smallmatrd{a}{b}{c}{d}\cdot x = (b+dx)/(a+cx).$ We get an action of $\mathrm{PGL}_2(F_{\pri})$ on $\mathcal{A}_{k,k}(\Proj_{\pri},L)$ by
\[\gamma\cdot \zeta(x) \defeq \frac{(a+cx)^k(\overline{a}+\overline{cz})^k}{(ad-bc)^{k/2}(\overline{ad}-\overline{bc})^{k/2}}\zeta(\gamma\cdot x),\]
directly extending the action of $\Sigma_0(p)$ on locally analytic functions on $\roi_{\pri} = \roi_F\otimes_{\Z}\Zp$ from earlier.\\
\\
Note that if $e = \gamma e_*$, then $\gamma^{-1}$ carries the set $\roi_{\pri}$ to $U(e) = \gamma^{-1}(\roi_{\pri}) \subset \Proj_{\pri}$. Hence if $\zeta$ is supported on $U(e)$, then $\gamma^{-1}\cdot \zeta$ is supported on $\roi_{\pri}$.
\begin{longversion} This, along with Lemma \ref{poly distribution}, motivates the following.
\end{longversion}
\begin{shortversion}
This, along with equation (\ref{poly distribution}), motivates the following.
\end{shortversion}
\begin{mdef}Let $e = \gamma e_* \in \mathcal{E}(\tree_{\pri})$.
\begin{itemize}
\item[(i)]Let $\mathcal{A}_{k,k}(U(e),L)$ be the subspace of functions $\zeta \in \mathcal{A}_{k,k}(\Proj_{\pri},L)$ that are supported on $U(e)$, with the natural weight $k$ action of $\gamma\Sigma_0(p)\gamma^{-1}$, and let $\mathcal{D}_{k,k}(U(e),L)$ be its continuous dual.
\item[(ii)] Define the \emph{overconvergent modular symbol at $e$ associated to $\f$}, denoted by
\begin{longversion}
\[\Psi_{\f_e} \in \symb_{\Gamma_e}(\mathcal{D}_{k,k}(U(e),L)),\]
by
\end{longversion}
\begin{shortversion}
$\Psi_{\f_e}$, to be the element of $\symb_{\Gamma_e}(\mathcal{D}_{k,k}(U(e),L))$ defined by
\end{shortversion}
\[\Psi_{\f_e}\{r-s\}(\zeta) \defeq \omega^{|\gamma|}\Psi_f\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot \zeta).\]
Here $\Gamma_e \defeq \mathrm{Stab}_{\Gamma}(e) = \gamma\Gamma_0(\n)\gamma^{-1}$.
\end{itemize}
\end{mdef}
\begin{mprop}
For each pair $r,s \in \Proj(F)$, we may define a distribution $\mu_{\f}\{r-s\}$ on $\mathcal{A}_{k,k}(\Proj_{\pri},L)$ attached to $\f$ by setting
\[\int_{U(e)} \zeta(t)d\mu_{\f}\{r-s\}(t) \defeq \Psi_{\f_e}\{r-s\}(\zeta)\]
and extending linearly.
\end{mprop}
\begin{longversion}
\begin{proof}
Suppose $p$ is inert (the case $p$ ramified is essentially identical). We need only show that $\mu_{\f}\{r-s\}$ is distributive over open sets. Let $e = \gamma e_*\in \mathcal{E}(\tree_{\pri})$ be an edge, with source $v \in \mathcal{V}(\mathcal{T})$, and let $\overline{e}, e_1,...,e_{p^2}$ be the edges with target $v$, so that $U(e) = \coprod _{i=1}^{p^2}U(e_i).$ We need to show that, for $\zeta$ locally analytic on $U(e)$, we have
\[\Psi_{\f_e}\{r-s\}(\zeta) = \sum_{i=1}^{p^2}\Psi_{\f_{e_i}}\{r-s\}(\zeta|_{U(e_i)}).\]
For each $j \newmod{p}$, recall that $\delta_j = \smallmatrd{1}{j}{0}{p}$. By direct inspection of the tree, we see that
\[\{e_i: 1\leq i \leq p^2\} = \{\delta_j^{-1} \gamma e_*: j \newmod{p}\} = \{\gamma \delta_j^{-1} e_*: j\newmod{p}\}.\]
(We need not have $\delta_j^{-1} \gamma e_* = \gamma \delta_j^{-1} e_*$ for individual edges, but we do get an equality of sets). Hence
\begin{align*}\sum_{i=1}^{p^2}\Psi_{\f_{e_i}}\{r-s\}(\zeta|_{U(e_i)}) &= \omega^{1+|\gamma|}\sum_{j \newmod{p}}\Psi_{f}\{\delta_j\gamma^{-1} r- \delta_j\gamma^{-1} s\}(\delta_j\gamma^{-1}\cdot \zeta)\\
&= \omega^{1+|\gamma|}N(\pri)^{-k/2}(\Psi_{f}|U_{\pri})\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot\zeta)\\
&= \omega^{|\gamma|}\Psi_f\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot\zeta) = \Psi_{\f_e}\{r-s\}(\zeta),
\end{align*}
as required. Here the penultimate equality follows from Corollary \ref{eigenvalues of newforms}.
\end{longversion}
\begin{shortversion}
\begin{proof}
Suppose $p$ is inert (the case $p$ ramified is essentially identical). We need only show that $\mu_{\f}\{r-s\}$ is distributive over open sets. Let $e = \gamma e_*\in \mathcal{E}(\tree_{\pri})$ be an edge, with source $v \in \mathcal{V}(\mathcal{T})$, and let $\overline{e}, e_1,...,e_{p^2}$ be the edges with target $v$, so that $U(e) = \coprod _{i=1}^{p^2}U(e_i).$ We need to show that, for $\zeta$ locally analytic on $U(e)$, we have
\[\Psi_{\f_e}\{r-s\}(\zeta) = \sum_{i=1}^{p^2}\Psi_{\f_{e_i}}\{r-s\}(\zeta|_{U(e_i)}).\]
For each $j \newmod{p}$, recall that $\delta_j = \smallmatrd{1}{j}{0}{p}$. By direct inspection of the tree, we see that $\{e_i: 1\leq i \leq p^2\} = \{\delta_j^{-1} \gamma e_*: j \newmod{p}\} = \{\gamma \delta_j^{-1} e_*: j\newmod{p}\}.$ (We need not have $\delta_j^{-1} \gamma e_* = \gamma \delta_j^{-1} e_*$ for individual edges, but we do get an equality of sets). A calculation using harmonicity and the fact that $\Psi_{f} = \omega p^{-k}\Psi_f|U_{\pri}$ (from Corollary \ref{eigenvalues of newforms}) shows that
\[\sum_{i=1}^{p^2}\Psi_{\f_{e_i}}\{r-s\}(\zeta|_{U(e_i)}) = \omega^{|\gamma|}\Psi_f\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot\zeta) = \Psi_{\f_e}\{r-s\}(\zeta),
\]
as required.
\end{proof}
\end{shortversion}
\subsection{Distributions on $\Proj_{\pri}$: the case $p$ split}
Now suppose $p$ is split as $\pri\pribar$ in $F$. The only additional complication in this case is that to construct $p$-adic $L$-functions, we must also consider behaviour at $\pribar$. Other than this, the construction is closely analogous to the case $p$ inert, and we will omit the details in this case.
\begin{mass}
Henceforth, we will assume that $f$ has small slope at $\pribar$, that is, that $v_p(a_{\pribar}) < k+1$, where $a_{\pribar}$ is the Hecke eigenvalue at $\pribar$. (Note that $f$ automatically has small slope at $\pri$ since it is new, so in this case we obtain a canonical overconvergent symbol $\Psi_f$, as before).
\end{mass}
\begin{mdef}
\begin{itemize}
\item[(i)]Let
\[\Proj_{\pri} \defeq \Proj(F_{\pri})\times \roi_{\pribar} \cong \Proj(\Qp)\times\Zp \supset \roi_F\otimes_{\Z}\Zp.\]
\item[(ii)]Let $\mathcal{A}_k(\Proj(F_{\pri}),L)$ be the space of functions $\Proj(F_{\pri}) \rightarrow L$ that are locally analytic except perhaps for a pole of order at most $k$ at $\infty$, and let $\mathcal{A}_k(\roi_{\pribar},L)$ be the space of locally analytic functions $\roi_{\pribar} \rightarrow L$. Define
\[\mathcal{A}_{k,k}(\Proj_{\pri},L) \defeq \mathcal{A}_k(\Proj(F_{\pri}),L)\widehat{\otimes}_{L}\mathcal{A}_k(\roi_{\pribar},L).\]
\end{itemize}
\end{mdef}
Let $U'(e) \defeq U(e)\times \roi_{\pribar} \subset \Proj_{\pri}$. Note that $\widetilde{\Gamma}$ preserves $\roi_{\pribar}$, so if $e = \gamma e_*$ for $\gamma \in \widetilde{\Gamma}$, then $U'(e) = \gamma^{-1}(\roi_F\otimes_{\Z}\Zp)$. Hence the following definition makes sense.
\begin{mdef}Let $e = \gamma e_* \in \mathcal{E}(\tree_{\pri})$.
\begin{itemize}
\item[(i)]Let $\mathcal{A}_{k,k}(U'(e),L)$ be the subspace of functions $\zeta \in \mathcal{A}_{k,k}(\Proj_{\pri},L)$ that are supported on $U'(e)$ and let $\mathcal{D}_{k,k}(U'(e),L)$ be its continuous dual.
\item[(ii)]Define the \emph{overconvergent modular symbol at $e$ associated to $\f$}, denoted by $\Psi_{\f_e},$ to be the element in $\symb_{\Gamma_e}(\mathcal{D}_{k,k}(U'(e),L))$ defined by
\[\Psi_{\f_e}\{r-s\}(\zeta) \defeq \omega^{|\gamma|}\Psi_f\{\gamma^{-1}r-\gamma^{-1}s\}(\gamma^{-1}\cdot \zeta).\]
\end{itemize}
\end{mdef}
In exactly the same manner as the inert case, we can then show that we may attach a distribution $\mu_{\f}\{r-s\}$ on $\mathcal{A}_{k,k}(\Proj_{\pri},L)$ to $\f$ by defining it on basic open sets as
\[\int_{U'(e)} \zeta(t)d\mu_{\f}\{r-s\}(t) \defeq \Psi_{\f_e}\{r-s\}(\zeta).\]
\subsection{Properties of the distributions}
Regardless of the splitting behaviour of $p$, we've constructed a family of distributions $\mu_{\f}\{r-s\}$ on $\Proj_{\pri}$ attached to $\f$. We have the following easy properties (see \cite{Ort04}, Lemma 3.2).
\begin{mprop}\label{properties of distribution}
\begin{itemize}
\item[(i)] If $P$ is any polynomial function over $L$, then for any $r,s \in \Proj(F)$ we have
\[\int_{\Proj_{\pri}} P(t)d\mu_{\f}\{r-s\}(t) = 0.\]
\item[(ii)]For $\gamma \in \widetilde{\Gamma}$, we have
\[\int_{U(\gamma e)}(\gamma\cdot\zeta)(t)d\mu_{\f}\{\gamma r - \gamma s\}(t) = \omega^{|\gamma|}\int_{U(e)}\zeta(t)d\mu_{\f}\{r - s\}(t).\]
\item[(iii)]The distributions `behave well under Hecke operators' in the sense that for all $I$ coprime to $\n$,
\[N(I)^{k/2}\sum_{j}\int_{U(\delta_{j}^{I} e)}(\delta_{j}^{I}\cdot \zeta)(t)d\mu\{\delta_{j}^{I} r - \delta_{j}^{I} s\}(t) = \lambda_{I}\int_{U(e)}\zeta(t)d\mu_{\f}\{r-s\}(t),\]
keeping the notation of Definition \ref{delta}, and where $\lambda_I$ is the Hecke eigenvalue of $f$ at $I$.
\end{itemize}
\end{mprop}
\section{Double integrals and cohomology classes}
In this section, we use the machinery constructed above to define double integrals in the manner of Darmon, and then attach two cohomology classes $\mathrm{lc}_f, \mathrm{oc}_f \in \mathrm{H}^1(\Gamma,\HomD)$ to $f$. This closely follows the methods of \cite{Ort04}, which was motivated by the methods of \cite{Dar01}. Darmon defines a \emph{multiplicative} double integral and a class $\mathrm{c}_f$ attached to a weight two modular form $f$, and then constructs classes $\log_p(\mathrm{c}_f)$ and $\mathrm{ord}_p(\mathrm{c}_f)$. In our setting, this multiplicative integral doesn't exist, and we instead define the analogues of $\log_p(\mathrm{c}_f)$ and $\mathrm{ord}_p(\mathrm{c}_f)$ by exploiting their properties as given in \cite{Dar01}.
\subsection{Double integrals}
Let $f \in S_{k,k}(\Omega_0(\n))^{\pri-\mathrm{new}}$ with associated system of distributions $\mu_{\f}\{r-s\}$ on $\Proj_{\pri}$, and recall that we fixed a choice of $p$-adic logarithm $\log_p : \Cp^\times \longrightarrow \Cp$.
\begin{mdef}
Let $x,y$ in $\uhp_{\pri}$, let $P \in V_{k,k}(\Cp)$ and let $r,s \in \Proj(F)$. Then define
\[\int_x^y\int_r^s (P) \defeq \int_{\Proj_{\pri}}\log_p\left(\frac{t_{\pri}-x}{t_{\pri}-y}\right)P(t)d\mu_{\f}\{r-s\}(t),\]
where $t_{\pri}$ is the projection of $t \in \Proj_{\pri}$ to $\Proj(F_{\pri})$. (Note that this is well-defined since $\uhp_{\pri} \defeq \Proj(\Cp)\backslash\Proj(F_{\pri})$, so that $t_{\pri}-x$ and $t_{\pri}-y$ cannot vanish and their quotient gives an element of $\Cp^\times$).
\end{mdef}
The following (compare \cite{Ort04}, Lemma 4.1) follow from properties of $\log_p$ and from Proposition \ref{properties of distribution}.
\begin{mprop}[Properties of the double integral]\label{properties of double integral}
\begin{itemize}
\item[(i)] The double integral is additive in $x,y$ and $r,s$ and is linear in $P$.
\item[(ii)] For $\gamma\in\widetilde{\Gamma}$,
\[\int_{\gamma x}^{\gamma y}\int_{\gamma r}^{\gamma s}(\gamma\cdot P) = \omega^{|\gamma|}\int_x^y\int_r^s (P).\]
\item[(iii)]The double integral `behaves well under Hecke operators' in the sense that for all $I$ coprime to $\n$,
\[N(I)^{k/2}\sum_{j}\int_{\delta_j^I x}^{\delta_k^I y}\int_{\delta_j^I r}^{\delta_j^I s} (\delta_k^I\cdot P) = \lambda_I\int_x^y\int_r^s (P),\]
for notation as in Proposition \ref{properties of distribution} (iii).
\end{itemize}
\end{mprop}
\begin{shortversion}
For ease of notation, let $\Delta_{k,k} \defeq \mathrm{Hom}(\Delta_0,V_{k,k}(\Cp)^*).$ We define cohomology classes $\mathrm{lc}_f$ and $\mathrm{oc}_f$ as follows.
\begin{mdef}[Definition of $\mathrm{lc}_f$] \label{define lc}Let $\tau \in \uhp_{\pri}$, and define a function $\Gamma \rightarrow \Delta_{k,k}$ by
\[\widetilde{\mathrm{lc}}_{f,\tau}(\gamma)\{r-s\}(P) \defeq \int_{\tau}^{\gamma \tau}\int_r^s (P).\]
From the properties of the double integral above, we see that this is a cocycle in $\mathrm{Z}^1(\Gamma,\Delta_{k,k})$ and its class $\mathrm{lc}_f \in \mathrm{H}^1(\Gamma,\Delta_{k,k})$ is independent of $\tau$.
\end{mdef}
\begin{mdef}[Definition of $\mathrm{oc}_f$] Let $v \in \mathcal{V}(\tree_{\pri})$, and define a function $\Gamma \rightarrow \Delta_{k,k}$ by
\[\widetilde{\mathrm{oc}}_{f,v}(\gamma)\{r-s\}(P) \defeq \sum_{e \in v\rightarrow \gamma v}\kappa_{\f}\{r-s\}(e,P),\]
where $v \rightarrow \gamma v$ denotes the (unique) path from $v$ to $\gamma v$ in the tree and where $\kappa_{\f}$ is the modular symbol on the tree from Section \ref{modular symbols on the tree}. From the properties of $\kappa_{\f}$ in Proposition \ref{modular symbol tree properties}, we see that this is a cocycle in $\mathrm{Z}^1(\Gamma,\Delta_{k,k})$ and its class $\mathrm{oc}_f \in \mathrm{H}^1(\Gamma,\Delta_{k,k})$ is independent of $v$.
\end{mdef}
We have natural actions of the Hecke operators on $\mathrm{H}^1(\Gamma,\Delta_{k,k})$, as in \cite{HidE}, Chapter 6.3. As in \cite{Ort04}, using Proposition \ref{properties of double integral} and Lemma \ref{modular symbol tree properties}, we have the following.
\begin{mprop}
Both $\mathrm{oc}_f$ and $\mathrm{lc}_f$ are eigensymbols for the Hecke operators away from $\n$ with the same eigenvalues as $f$, that is, they are elements in the $f$-isotypic subspace of $\mathrm{H}^1(\Gamma,\Delta_{k,k})$.
\end{mprop}
\end{shortversion}
\begin{longversion}
\subsection{Definition of $\mathrm{lc}_f$}
For ease of notation, let $\Delta_{k,k} \defeq \mathrm{Hom}(\Delta_0,V_{k,k}(\Cp)^*).$
\begin{mdef}\label{define lc}
Let $\tau \in \uhp_{\pri}$, and define a function $\Gamma \rightarrow \Delta_{k,k}$ by
\[\widetilde{\mathrm{lc}}_{f,\tau}(\gamma)\{r-s\}(P) \defeq \int_{\tau}^{\gamma \tau}\int_r^s (P).\]
\end{mdef}
From the properties of the double integral above, we see that this is a cocycle in $\mathrm{Z}^1(\Gamma,\Delta_{k,k})$ and its class $\mathrm{lc}_f \in \mathrm{H}^1(\Gamma,\Delta_{k,k})$ is independent of $\tau$.
\subsection{Definition of $\mathrm{oc}_f$}
\begin{mdef}Let $v \in \mathcal{V}(\tree_{\pri})$, and define a function $\Gamma \rightarrow \Delta_{k,k}$ by
\[\widetilde{\mathrm{oc}}_{f,v}(\gamma)\{r-s\}(P) \defeq \sum_{e \in v\rightarrow \gamma v}\kappa_{\f}\{r-s\}(e,P),\]
where $v \rightarrow \gamma v$ denotes the (unique) path from $v$ to $\gamma v$ in the tree and where $\kappa_{\f}$ is the modular symbol on the tree from Section \ref{modular symbols on the tree}.
\end{mdef}
From the properties of $\kappa_{\f}$ in Proposition \ref{modular symbol tree properties}, we see that this is a cocycle in $\mathrm{Z}^1(\Gamma,\Delta_{k,k})$ and its class $\mathrm{oc}_f \in \mathrm{H}^1(\Gamma,\Delta_{k,k})$ is independent of $v$.
\subsection{Hecke operators}
We have natural actions of the Hecke operators on $\mathrm{H}^1(\Gamma,\Delta_{k,k})$, as in \cite{HidE}, Chapter 6.3. As in \cite{Ort04}, using Proposition \ref{properties of double integral} and Lemma \ref{modular symbol tree properties}, we have the following.
\begin{mprop}
For $I$ coprime to $\n$, we have
\[\mathrm{lc}_f|T_I = \lambda_I \mathrm{lc}_f\]
and
\[\mathrm{oc}_f|T_I = \lambda_I \mathrm{oc}_f.\]
In particular, both $\mathrm{lc}_f$ and $\mathrm{oc}_f$ are elements of the $f$-isotypic subspace of $\mathrm{H}^1(\Gamma,\Delta_{k,k})$.
\end{mprop}
\end{longversion}
\section{Relating to $L$-values}
\subsection{Optimal embeddings}
In both \cite{Dar01} and \cite{Ort04}, one obtains a link between the cohomology classes defined above and $L$-values of the form $f$ at characters of conductor $c$ by exploiting \emph{optimal embeddings of conductor $c$}, that is, certain embeddings $F\times F\hookrightarrow \mathrm{M}_2(F)$. Heuristically, one should see such an embedding as giving rise to a cycle in the symmetric space corresponding to our level group, and hence it gives rise to a homology class that we can then pair our cohomology classes against (see, for example, \cite{GMS15} for details of this approach).
$\\ \\$
In practice we can work out the details of this rather more explicitly. In particular, each such embedding gives rise to an infinite cyclic subgroup $\langle \gamma\rangle$ of $\Gamma$, which has precisely two fixed points $x,y$ in $\Proj(F_{\pri})$. One then writes down a polynomial $P(t)$ that is invariant under the action of $\gamma$, and evaluates cohomology classes in $\mathrm{H}^1(\Gamma,\HomD)$ at $\gamma, x, y$ and $P$. In this section, we bypass embeddings completely, and simply write down the corresponding values of $\gamma, x, y$ and $P$. Throughout, and for the remainder of the paper, we will assume $k$ is even (in lieu of Corollary \ref{exceptional zero existence}).
\begin{mdef}
Let $c \in \roi_F$ be prime to $\pri = (\pi)$ and $v\in \roi_F$ prime to $c$.
\begin{itemize}
\item[(i)] Define $s'\defeq \text{order of }\pi\text{ in }(\roi_F/c)^\times,$ and let $s \defeq 2\times(\text{order of }\pi^2\text{ in }(\roi_F/c)^\times).$
\item[(ii)] Define
\[\gamma_{c,v} \defeq \matrd{\pi^{-s/2}}{(\pi^{s/2}-\pi^{-s/2})\frac{v}{c}}{0}{\pi^{s/2}}\]
(As $c|(\pi^{s/2}-\pi^{-s/2})$, there is no problem with denominators). Note that the fixed points of this in $\Proj(F_{\pri})$ are $-v/c$ and $\infty$.
\item[(iii)] Let $M_{c,v}(t) \defeq ct + v$, a M\"{o}bius transformation taking $-v/c$ to $0$ and fixing $\infty$.
\item[(iv)] Let $P_{c,v}(t) \defeq (ct+v)^{k/2}(\overline{ct}+\overline{v})^{k/2},$ an element of $V_{k,k}(L)$ that is invariant under the action of $\gamma_{c,v}$.
\end{itemize}
\end{mdef}
\begin{longversion}
\begin{mrem}
Again, the choice of action on modular symbols means we have to be careful about how $\gamma_{c,v}$ is acting on the various spaces involved here. We have an action of $\gamma_{c,v}$ on $\Proj(F)$ by the usual fractional linear transformations $x \mapsto (ax+b)/(cx+d)$, and under this action, $v/c$ and $\infty$ are fixed. We also have an action of $\gamma_{c,v}$ on $\uhp_{\pri}$ by $\tau \mapsto (b+d\tau)/(a+c\tau)$, and under this action, for any choice of $\tau \in \uhp_{\pri}$, we have
\[\lim_{j \rightarrow \infty} \gamma_{c,v}^j\tau = -v/c, \hspace{12pt} \lim_{j\rightarrow \infty}\gamma_{c,v}^{-j}\tau = \infty.\]
\end{mrem}
\end{longversion}
\begin{shortversion}
\begin{mrem}
The actions of $\gamma_{c,v}$ on the cusps and on $\Proj(F_{\pri})$ and $\uhp_{\pri}$ are different. In particular $\gamma_{c,v}$ fixes the cusps $v/c$ and $\infty$, whilst for any $\tau \in \uhp_{\pri}$, we have $\lim_{j \rightarrow \infty} \gamma_{c,v}^j\tau = -v/c$ and $\lim_{j\rightarrow \infty}\gamma_{c,v}^{-j}\tau = \infty.$
\end{mrem}
\end{shortversion}
\begin{longversion}
The following is a collection of well-known and/or easily proved facts and properties that we'll need when relating cohomology classes to $L$-values.
\end{longversion}
\begin{mprop}\label{embedding invariants}
\begin{itemize}
\item[(i)] Let $t\in\Proj(F_{\pri})$. Then we have $M_{c,v}(\gamma_{c,v} t) = \pi^s M_{c,v}(t)$.
\item[(ii)] There is a bijection between doubly infinite (non-repeating) paths in $\tree_{\pri}$ and distinct pairs of elements of $\Proj(F_{\pri})$. (Here by \emph{doubly infinite} we mean that every vertex in the path is connected to precisely two other vertices, that is, the path has no `end').
\item[(iii)] Let $\mathrm{path}(-v/c,\infty)$ denote the infinite path in $\tree_{\pri}$ corresponding to the pair $(-v/c,\infty)$. It is possible to index the vertices and edges of this path such that $t(e_j) = v_j$ and
\[U(e_j) = \{t \in \Proj(F_{\pri}): v_{\pri}(M_{c,v}(t)) \geq -j\}.\]
\item[(iv)] Under the action of $\Gamma$ on $\tree_{\pri}$, we have $\gamma_{c,v}e_j = e_{j+s}.$
\item[(v)] Let $U(v_j) \defeq \{t \in \Proj(F_{\pri}): v_{\pri}(M_{c,v}(t)) = - j\}.$ Then
\[F_{c,v} \defeq \bigcup_{j=0}^{s-1}U(v_{j})\]
is a fundamental domain for the action of $\gamma_{c,v}$ on $\Proj(F_{\pri})\backslash\{-v/c,\infty\}$.
\end{itemize}
\end{mprop}
\begin{proof}
Part (i) is direct calculation. For part (ii), see \cite{DT08}. Parts (iii) to (v) follow from the same arguments as the corresponding statements in \cite{Dar01} over $\Q$, which are given in equations (88) for part (iii), (90) for part (iv) and (87) for part (v).
\end{proof}
\subsection{Relating $\mathrm{oc}_f$ to classical $L$-values}\label{ocf lvalues}
With the above given, we are in a position to compute a special value of the cocycle $\widetilde{\mathrm{oc}}_{f,v}$.
\begin{mprop}
Let $J_v$ denote the coset $v\langle\pi\rangle \subset (\roi_F/c)^\times.$ For each $a \in J_v$, let $j(a)$ be the smallest non-negative integer such that $a \equiv v\pi^{j(a)} \newmod{c}$. We have
\[\widetilde{\mathrm{oc}}_{f,v}(\gamma_{c,v})\{v/c - \infty\}(P_{c,v}) = \beta\sum_{a\in J_v} \omega^{j(a)}\phi_f\{a/c - \infty\}\bigg[(ct+a)^{k/2}(\overline{ct}+\overline{a})^{k/2}\bigg],\]
where
\[\beta = \left\{\begin{array}{ll} 1 &: s = s',\\
2 &: \omega = 1, s = 2s',\\
0 &: \omega = -1, s = 2s'.
\end{array}\right.\]
\end{mprop}
\begin{proof}
We compute directly that
\[\widetilde{\mathrm{oc}}_{f,v}(\gamma_{c,v})\{v/c-\infty\}(P_{c,v}) = \sum_{j=1}^s\kappa_{\f}\{x-y\}(e_j,P_{c,v}).\]
Let $u\in\roi_F$ be any element such that $u \equiv v/c \newmod{p^s}$. Since $U(e_j) = \{t\in \Proj(F_{\pri}): \pi^{j}(ct+v) \in \roi_{\pri}\} = \{t \in \Proj(F_{\pri}): \pi^{j}(t+u)\in\roi_{\pri}\} = U(\gamma_je_*),$ where
\begin{equation}\label{gammaj}
\gamma_j \defeq \matrd{\pi^{-j}}{u}{0}{1} \in \widetilde{\Gamma},
\end{equation}
we see that $e_j = \gamma_j e_*$. Hence $\kappa_f\{v/c-\infty\}(e_j,P_{c,v}) = \omega^j\phi_f|\gamma_j^{-1}\{v/c-\infty\}(P_{c,v}).$ Now direct computation shows that
\[\gamma_j^{-1}(v/c) = \frac{\pi^{j}v - cu\pi^{j}}{c},\hspace{12pt} \gamma_j^{-1}\infty = \infty,\]
and
\[\gamma_j^{-1}\cdot P_{c,v}(t) = (ct + (\pi^{j}v - cu\pi^{j}))^{k/2}(\overline{ct} + (\overline{\pi}^{j}\overline{v} - \overline{cu}\overline{\pi}^{j}))^{k/2}. \]
As $j$ ranges from 1 to $s$, the quantity $a = \pi^{j}v-cu\pi^{j}$ ranges over $J_v$, once if $s = s'$ and twice if $s = 2s'$. In the latter case, the signs $\omega^j$ at each instance are both equal to 1 if $\omega = 1$ and are opposite if $\omega = -1$. Finally, the result follows from the observation that if $a,a' \in \roi_F$ such that $a \equiv a' \newmod{c}$, then the action of $\smallmatrd{1}{(a-a')/c}{0}{1} \in \Gamma_0(\n)$ shows that
\[\phi_f\{a/c-\infty\}[(ct+a)^{k/2}(\overline{ct}+\overline{a})^{k/2})] = \phi_f\{a'/c-\infty\}[(ct+a')^{k/2}(\overline{ct}+\overline{a'})^{k/2})].\qedhere\]
\end{proof}
Henceforth we assume we are not in the case $\beta = 0$, which will be irrelevant to our results. An immediate corollary is the following.
\begin{mcor}\label{oc and L-values}
Suppose $\chi$ is a finite order character of conductor $c$ (that we naturally see as a character on $(\roi_F/c)^\times$), and that $\chi(\pri) = \omega$. Then
\[\sum_{v \in (\roi_F/c)^\times}\chi(v)\mathrm{oc}_f(\gamma_{c,v})\{v/c - \infty\}(P_{c,v}) = s\left[\frac{D^{r+1}\tau(\chi^{-1})|c|^r \#\roi_F^\times}{2\Omega_f}\right] \Lambda(f,\chi,k/2+1).\]
\end{mcor}
\begin{proof}
The corresponding result for the cocycle $\widetilde{\mathrm{oc}}_{f,v}$ follows from equation (\ref{integralformula}). In particular, note that the condition on $\chi(\pi)$ means that $\chi(v)\omega^{j(a)} = \chi(v\pi^{j(a)}) = \chi(a)$, whilst the sum runs over all cosets $v\langle\pi\rangle \subset (\roi_F/c)^\times$ for all $v\in(\roi_F/c)^\times$, which means that every element of $(\roi_F/c)^\times$ is hit $s'$ times. But $\beta s' = s$.
$\\ \\$
To see that the result holds at the level of cohomology classes, let $b \in \mathrm{B}^1(\Gamma,\Delta_{k,k})$ be a coboundary. Then there exists $\phi \in \Delta_{k,k} =\mathrm{Hom}(\Delta_0,V_{k,k}(\Cp)^*)$ such that $b(\gamma)\{r-s\}(P) = (\phi|\gamma)\{r-s\}(P) - \phi\{r-s\}(P).$ But it follows directly that $b(\gamma_{c,v})\{v/c-\infty\}(P_{c,v}) = 0$ using the $\gamma_{c,v}$-invariance of $v/c, \infty$ and $P_{c,v}$.
\end{proof}
\subsection{Relating $\mathrm{lc}_f$ to $p$-adic derivatives}
In this section, we prove a formula analogous to the one above, this time relating special values of the class $\mathrm{lc}_f$ to the derivative of the $p$-adic $L$-function at the corresponding critical value. For clarity, we will assume throughout this section that $p$ is split; the cases where $p$ is inert or ramified are almost identical (and, in fact, slightly easier where they differ). We freely use the notation of Section \ref{ocf lvalues}.
\subsubsection{Rephrasing as an integral over the fundamental domain}
Let $c$ and $v$ be as before. The following lemma will be useful in the sequel.
\begin{mlem} \label{integral zero}We have
\[\int_{F_{c,v}\times \mathcal{O}_{\overline{\mathfrak{p}}}}P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t)=0.\]
\end{mlem}
\begin{longversion}
\begin{proof} Note that from Proposition \ref{embedding invariants} we have $F_{c,v}= U(e_{s-1})\smallsetminus U(e_{-1})$ and $\gamma_{c,v}e_{-1}= e_{s-1}$. Hence $U(e_{s-1})= \gamma_{c,v}^{-1}U(e_{-1})$, so that $U(e_{s-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}= \gamma_{c,v}^{-1}\left[ U(e_{-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}\right]$. From Proposition 5.11 we deduce that
\begin{align*}
\int_{F_{c,v}\times \mathcal{O}_{\overline{\mathfrak{p}}}}P_{c,v}(t)d\mu_{\mathcal{F}}&\left\lbrace \frac{v}{c}- \infty \right\rbrace (t) =
\int_{U(e_{s-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}}P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t) - \int_{U(e_{-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}}P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t)\\
&= \int_{\gamma_{c,v}^{-1}\left[U(e_{-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}\right]}P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t) - \int_{U(e_{-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}}P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t) \\
&= \int_{U(e_{-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}}\gamma_{c,v}^{-1}\cdot P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \gamma_{c,v}^{-1}\left(\frac{v}{c}\right) - \infty \right\rbrace (t) - \int_{U(e_{-1})\times \mathcal{O}_{\overline{\mathfrak{p}}}}P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t)\\
&= 0,
\end{align*}
where the last equality follows from the fact that $P_{c,v}(t)$ and $\frac{v}{c} \in \mathbb{P}^{1}(F)$ are invariant under the action of $\gamma_{c,v}$.
\end{proof}
\end{longversion}
\begin{shortversion}
\begin{proof}
This is a direct calculation using the fact that $F_{c,v} = U(e_{s-1})\backslash U(e_{-1})$ and $U(e_{s-1}) = \gamma_{c,v}^{-1}U(e_{-1})$, combined with the invariance of $P_{c,v}$, $c/v$ and $\infty$ under $\gamma_{c,v}$.
\end{proof}
\end{shortversion}
\begin{longversion}
Recall that by definition,
\[\tilde{\mathrm{lc}}_{f, \tau}(\gamma_{c,v})\{v/c - \infty \}(P_{c,v})= \int_{\mathbb{P}^{1}_{\mathfrak{p}}}\mathrm{log}_{p}\left( \frac{t- \gamma_{c,v}\tau}{t- \tau}\right)P_{c,v}(t)d\mu_{\mathcal{F}}\left\lbrace \frac{v}{c}- \infty \right\rbrace (t).\]
To ease notation, we'll henceforth write $\mu_{\f}^{c,v}$ for $\mu_{\f}\{v/c-\infty\}$. In the following proposition, we simplify the definition.
\end{longversion}
\begin{shortversion}
Recall the definition of $\widetilde{\mathrm{lc}}_{f,\tau}(\gamma_{c,v})\{v/c - \infty\}(P_{c,v})$ from Definition \ref{define lc}. We rephrase this in a way that allows us to see $p$-adic $L$-values. To ease notation, we'll henceforth write $d\mu_{\f}^{c,v}$ for $d\mu_{\f}\{v/c-\infty\}$.
\end{shortversion}
\begin{mprop}\label{integral over fund domain}
We have
$$\widetilde{\mathrm{lc}}_{f,\tau}(\gamma_{c,v})\{v/c - \infty \}(P_{c,v})= \int_{F_{c,v}\times \mathcal{O}_{\overline{\mathfrak{p}}}}\mathrm{log}_{p}(ct+ v)P_{c,v}(t)d\mu_{\mathcal{F}}^{c,v}(t).$$
\begin{longversion}In particular, the left hand side does not depend on the choice of $\tau$.\end{longversion}
\end{mprop}
The proof will require a lemma. We have a decomposition
\[\Proj_{\pri} = \left(U(\overline{e}_{(n+1)s-1})\sqcup U(e_{-ns-1}) \bigsqcup_{i = -n}^n\gamma_{c,v}^i F_{c,v} \right)\times \roi_{\pribar}.\]
We break the integral up into sums over these components and let $n$ tend to $\infty$.
\begin{mlem}\label{endpoints zero}
The integral over the `endpoints' vanishes in the limit. More precisely, let
\[I(n) \defeq \int_{U'(\overline{e}_{(n+1)s-1})\sqcup U'(e_{-ns-1})}\mathrm{log}_{p}\left( \frac{t- \gamma_{c,v}\tau}{t- \tau}\right)P_{c,v}(t)d\mu_{\mathcal{F}}^{c,v}(t).\]
Then $\lim_{n\rightarrow \infty}I(n) = 0.$ In particular,
\[\tilde{\mathrm{lc}}_{f, \tau}(\gamma_{c,v})\{v/c - \infty \}(P_{c,v}) =\lim_{n\rightarrow \infty}\sum_{i= -n}^{n}\int_{\gamma_{c,v}^{i}\left[ F_{c,v}\times \mathcal{O}_{\overline{\mathfrak{p}}}\right] }\mathrm{log}_{p}\left( \frac{t- \gamma_{c,v}\tau}{t- \tau}\right)P_{c,v}(t)d\mu_{\mathcal{F}}^{c,v}(t).\]
\end{mlem}
\begin{proof}
Since $U'(\overline{e}_{(n+1)s-1}) = \gamma_{c,v}^{-n-1}U'(e_{-1})$ and $U'(e_{-ns-1}) = \gamma_{c,v}^nU'(e_{-1}),$ we can write
\begin{align*}
\int_{U'(\overline{e}_{(n+1)s-1})}\mathrm{log}_{p}&\left(\frac{t- \gamma_{c,v}\tau}{t- \tau}\right)P_{c,v}(t)d\mu_{\mathcal{F}}^{c,v}(t) =\\
& -\int_{U'(e_{-1})}\mathrm{log}_{p}\left( \frac{\gamma_{c,v}^{n+1}t- \gamma_{c,v}\tau}{\gamma_{c,v}^{n+1}t- \tau} \right)P_{c,v}(t)d\mu_{\mathcal{F}}^{c,v}(t)
\end{align*}
using the invariance of $P_{c,v}$ and $d\mu_{\f}^{c,v}$ (and similarly for the integral over $U'(e_{-ns-1})$). For each fixed $n$, the expressions
\[\frac{\gamma_{c,v}^{n+1}t- \gamma_{c,v}\tau}{\gamma_{c,v}^{n+1}t- \tau}, \hspace{12pt} \frac{t- \gamma_{c,v}^{-n}\tau}{t- \gamma_{c,v}^{-n-1}\tau}\]
are both M\"{o}bius transformations sending $\gamma_{c,v}^{-n}\tau$ to zero and $\gamma_{c,v}^{-n-1}\tau$ to $\infty$, hence one is a constant scalar multiple of the other. A short calculation using Lemma \ref{integral zero} now shows that
\[I(n) = \int_{U(e_{-1})}\mathrm{log}_p\left(\frac{t- \gamma_{c,v}^{-n}\tau}{t- \gamma_{c,v}^{-n-1}\tau}\cdot \frac{t- \gamma_{c,v}^{n+1}\tau}{t- \gamma_{c,v}^{n}\tau}\right) P_{c,v}(t)d\mu_{\f}^{c,v}(t).\]
In the limit, the expression in the $\mathrm{log}_p$ becomes $M_{c,v}(\gamma_{c,v}t)/M_{c,v}(t).$ But by Proposition \ref{embedding invariants}, this is equal to $\pi^s$, and $\mathrm{log}_p(\pi^s) = 0$. Hence the integrand tends to zero and the first result follows. The second result is a direct consequence of this and the above decomposition.
\end{proof}
\begin{shortversion}
\begin{proof}\emph{(Proposition \ref{integral over fund domain})}.
We use the second part of Lemma \ref{endpoints zero}. A general term of the sum is an integral over $\gamma_{c,v}^i[F_{c,v}\times\roi_{\pribar}],$ and using the transformation property, we can write it as an integral over $F_{c,v}\times\roi_{\pribar}$. In particular, using similar methods to the proof of Lemma \ref{endpoints zero}, we see that
\[\widetilde{\mathrm{lc}}_{f,\tau}(\gamma_{c,v})\left\{\frac{v}{c} - \infty\right\}(P_{c,v}) = \lim_{n\rightarrow\infty} \sum_{i=-n}^n \int_{F_{c,v}\times\roi_{\pribar}} \mathrm{log}_p\left(\frac{t-\gamma_{c,v}^{i+1}\tau}{t-\gamma_{c,v}^i\tau}\right)P_{c,v}(t)d\mu_{\f}^{c,v}(t).\]
This sum telescopes via the expression inside the $\mathrm{log}_p$, and in the limit, the resulting expression tends to $M_{c,v}(t) = ct+v$. This completes the proof.
\end{proof}
\end{shortversion}
\subsubsection{Rephrasing using the overconvergent modular symbol}
We can decompose the fundamental domain of Proposition \ref{integral over fund domain} into a union of sets of the form $U(v_{j})\times\roi_{\pribar}$, and hence as
\[F_{c,v}\times\roi_{\pribar} = \bigsqcup_{j=0}^{s-1}\gamma_j^{-1}\left[\roi_{\pri}^\times\times\roi_{\pribar}\right],\]
where $\gamma_j$ is as defined in equation (\ref{gammaj}). Here we've used the fact that $U(v_j) = U(\gamma_j v_*) = \gamma_j^{-1}\roi_{\pri}^\times,$ which can be seen directly. Hence we have
\begin{align}\label{lcf value}
\widetilde{\mathrm{lc}}_{f,\tau}(\gamma_{c,v})&\{v/c-\infty\}(P_{c,v}) = \sum_{j=0}^{s-1}\int_{\gamma_j^{-1}[\roi_{\pri}^\times\times\roi_{\pribar}]} \mathrm{log}_p(ct+v)P_{c,v}(t)d\mu_{\f}^{c,v}\\
&= \sum_{j=0}^{s-1}\omega^j\int_{\roi_{\pri}^\times\times\roi_{\pribar}}\mathrm{log}_p(ct+a_j)(ct+a_j)^{k/2}(\overline{ct}+\overline{a_j})^{k/2}d\mu_{\f}\{a_j/c-\infty\},\notag
\end{align}
where $a_j = \pi^j v - cu\pi^j$ (using the calculations and notation of the previous section). \begin{shortversion}Now a short explicit calculation shows:\end{shortversion}
\begin{mlem}\label{lcf oms}
Write $c = g\overline{\pi}^r$, for some $r\geq 0$ and $(g)$ coprime to $(p)$. For any $a\in\roi_F$, we have
\begin{align*}\int_{\roi_{\pri}^\times\times\roi_{\pribar}}\mathrm{log}_p(ct+a)&(ct+a)^{k/2}(\overline{ct}+\overline{a})^{k/2}d\mu_{\f}^{c,v}(t) \\
&= (g\overline{g})^{k/2}\lambda_{\pribar}^{r}\Psi_f\left|\matrd{1}{a}{0}{g}\right.\{0-\infty\}\bigg(\mathrm{log}_p(t)(t\overline{t})^{k/2}\mathbf{1}_{a\newmod{\pribar^r}}\bigg)\\
&= \lambda_{\pribar}^r\int_{\{[a]\}\times\roi_\pri^\times\times\big[a+\pribar^r\roi_{\pribar}\big]}\mathrm{log}_p(z_{\pri})\langle\mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z}),
\end{align*}
where $\Psi_f$ is the overconvergent modular symbol attached to $f$, $\mathbf{1}_{a\newmod{\pri^r}}$ is the indicator function of the set $\roi_{\pri}^\times \times[a + \pribar^r\roi_{\pribar}],$ and $\mu_p$ is the distribution on $(\roi_F/\mathfrak{g})^\times\times(\roi_F\otimes_{\Z}\Zp)$ defined in Section \ref{padic lfunction section}.
\end{mlem}
\begin{longversion}
\begin{proof}
Since $\roi_{\pri}^\times\times\roi_{\pribar} \subset \roi_F\otimes_{\Z}\Zp$, we can write the integral as
\[\Psi_f\{a/c-\infty\}(\mathrm{log}_p(ct+a)(ct+a)^{k/2}(\overline{ct}+\overline{a})^{k/2}\mathbf{1}_{\roi_{\pri}^\times\times\roi_{\pribar}}).\]
In the right hand side, replace $\Psi_f$ with $\lambda_{\pribar}^{-r}\Psi_f|U_{\pribar}^r$ and expand out the sum define the Hecke operator. Since $U_{\pri}$ commutes with the action of $\smallmatrd{1}{a}{0}{g}$, this gives
\begin{align*}
\mathrm{(RHS)} &= (g\pi^r\overline{g\pi}^r)^{k/2}\sum_{b \newmod{\pribar^r}}\Psi_f\left|\matrd{1}{a}{0}{g}\matrd{1}{b}{0}{\overline{\pi}^r}\right.\{0-\infty\}\bigg(\mathrm{log}_p(t)(t\overline{t})^{k/2}\mathbf{1}_{a\newmod{\pribar^r}}\bigg)\\
&= (c\overline{c})^{k/2}\Psi_f\left|\matrd{1}{a}{0}{c}\right.\{0-\infty\}\bigg(\mathrm{log}_p(t)(t\overline{t})^{k/2}\mathbf{1}_{a\newmod{\pribar^r}}\bigg),
\end{align*}
since the indicator kills all other terms of the sum. Expanding this out gives the left hand side. The second equality is by definition of $\mu_p$.
\end{proof}
\end{longversion}
\begin{mcor}\label{lcf oms cor}We have
\[\mathrm{lc}_f(\gamma_{c,v})\{v/c-\infty\}(P_{c,v}) = \sum_{a \in J_v}\omega^{j(a)}\lambda_{\pribar}^r\int_{\{[a]\}\times\roi_\pri^\times\times\big[a+\pribar^r\roi_{\pribar}\big]}\mathrm{log}_p(z_{\pri})\langle\mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z}).\]
\end{mcor}
\begin{proof}
This follows directly from equation (\ref{lcf value}), Lemma \ref{lcf oms} and the fact that as $j$ ranges from $0$ to $s-1$, $a_j$ ranges over the set $J_v$ defined in the previous section with multiplicity $\beta$. As in the previous section, this value depends only on the cohomology \emph{class}, so we drop the tilde.
\end{proof}
\subsubsection{Completing the calculation}
Finally, we tie the above results together to relate $\mathrm{lc}_{f}$ to derivatives of the $p$-adic $L$-function attached to $f$. As in the last subsection we suppose that we are not in the case $\beta= 0$.
\begin{mcor}\label{lc and L-values}
Suppose $\chi$ is a finite order character of conductor $c$ (that we naturally see as a character on $(\roi_F/c)^\times$), and that $\chi(\pri) = \omega$. Write $c = g\overline{\pi}^r$. Then
\[\left[1 - \frac{\chi(\pribar)N(\pribar)^{k/2}}{\lambda_{\pribar}}\right]\sum_{v \in \left(\roi_{F}/ c\right)^{\times}}\chi(v)\mathrm{lc}_f(\gamma_{c,v})\{v/c - \infty \}(P_{c,v})= s\lambda_{\pribar}^r \frac{d}{ds_{\mathfrak{p}}}L_{p}(f, \chi, \mathbf{s})\mid_{\mathbf{s}= \mathbf{k/2}}.\]
\end{mcor}
\begin{proof} Note that if $\mathbf{z} \in \cl_F(\mathfrak{g}p^\infty)$ satisfies $\mathbf{z}\equiv \pi^{j}v \newmod{c}$ then $\chi(\mathbf{z})= \omega^{j}\chi(v)$. We use Corollary \ref{lcf oms cor} and the fact that $s = \beta s'$. In particular, we see that
\begin{align*}\sum_{v \in (\roi_F/ c)^{\times}}&\chi(v)\mathrm{lc}_f(\gamma_{c,v})\{v/c - \infty \}(P_{c,v})
= \\
&s\lambda_{\pribar}^r\int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \roi_{\pribar}]}\chi(\mathbf{z})\mathrm{log}_p(z_{\pri})\langle\mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z}).
\end{align*}
We need an integral over the units in $\roi_{\pribar}$ in order to descend to $\cl_F(\mathfrak{g}p^\infty)$. To this end, note that we can break the integral up as
\[\int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \roi_{\pribar}]} = \int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \roi_{\pribar}^\times]} + \int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \pribar\roi_{\pribar}]}.\]
If $\pribar|c$, then extend $\chi$ to $\roi_{\pribar}$ by defining it to be zero on $\pribar\roi_{\pribar}$ (and correspondingly on the larger space in the integral). Then
\begin{align*}\int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \pribar\roi_{\pribar}]}&\chi(\mathbf{z})\mathrm{log}_p(z_{\pri})\langle \mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z}) \\
&= \lambda_{\pribar}^{-1}\int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \roi_{\pribar}]}\chi(\overline{\pi}\mathbf{z})\mathrm{log}_p(\overline{\pi}z_{\pri})\langle \overline{\pi}\mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z})\\
&= \frac{\chi(\overline{\pi})N(\pribar)^{k/2}}{\lambda_{\pribar}}\int_{(\roi_F/\mathfrak{g})^\times \times[\roi_{\pri}^\times \times \roi_{\pribar}]}\chi(\mathbf{z})\mathrm{log}_p(z_{\pri})\langle \mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z}).
\end{align*}
By collecting these terms together, and using the fact that the integrand is invariant under units, we see that
\begin{align*}\left[1 - \frac{\chi(\pribar)N(\pribar)^{k/2}}{\lambda_{\pribar}}\right]\sum_{v \in (\roi_F/ c)^{\times}}\chi(v)&\mathrm{lc}_f(\gamma_{c,v})\{v/c - \infty \}(P_{c,v})\\
&= s\lambda_{\pribar}^r\int_{\cl_F(\mathfrak{g}p^\infty)}\chi(\mathbf{z})\mathrm{log}_p(z_{\pri})\langle\mathbf{z}_p\rangle^{k/2}d\mu_p(\mathbf{z}).
\end{align*}
But the latter integral is precisely the derivative in the $s_{\pri}$ direction of the $p$-adic $L$-function (as can be seen by differentiating the integrand in the definition). This completes the proof.
\end{proof}
\section{The cohomology of $\Gamma$}
We now prove that the $f$-isotypic component of $\mathrm{H}^1(\Gamma,\HomD)$, that is the subspace of classes $\psi \in \mathrm{H}^1(\Gamma,\HomD)$ such that $T_I \psi = \lambda_I \psi$ for all $I$ coprime to $\n$, is one dimensional. We make use of the tree.
\begin{mprop}\label{long exact sequence}
Let $\Delta_{k,k} \defeq \mathrm{Hom}(\Delta_0,V_{k,k}(\Cp)^*)$. There is an exact sequence
\begin{align*}\big[\h^0(\Gamma_0(\m),\Delta_{k,k})\oplus\h^0(\Gamma'_0(\m),\Delta_{k,k}) \big]\longrightarrow &\hspace{2pt}\h^0(\Gamma_0(\n),\Delta_{k,k}) \longrightarrow \h^1(\Gamma,\Delta_{k,k}) \\
&\longrightarrow \big[\h^1(\Gamma_0(\m),\Delta_{k,k})\oplus\h^1(\Gamma'_0(\m),\Delta_{k,k}))\big]\end{align*}
where $\Gamma'_0(\m) \defeq \alpha\Gamma_0(\m)\alpha^{-1}.$
\end{mprop}
\begin{proof}
This is \cite{Ser80}, Chapter II.2.8, Proposition 13, applied to
\[M = \mathrm{Hom}(\Delta_0,V_{k,k}(\Cp)^*)\]
and $G = \Gamma,$ noting that $\mathrm{Stab}_{\Gamma}(e_*) = \Gamma_0(\n)$, $\mathrm{Stab}_{\Gamma}(s(e_*)) = \Gamma_0(\m)$, and $\mathrm{Stab}_{\Gamma}(t(e_*)) = \Gamma'_0(\m)$.
\end{proof}
\subsection{A cohomological lemma}
We can relate these spaces to compactly supported cohomology using a trick of Ash and Stevens (see \cite{AS86}, Proposition 4.2).
\begin{mlem}\label{ash stevens}
For each $i\geq 0$ there is a Hecke-equivariant isomorphism
$$\mathrm{H}^{i}(\Gamma_{0}(\mathfrak{m}), \Delta_{k,k})) \cong \mathrm{H}^{i+1}_{c}(\Gamma_{0}(\mathfrak{m})\backslash \mathcal{H}_{3}, \widetilde{V_{k,k}(\C_{p})^{\ast}}),$$
where the tilde denotes the natural local system attached to the right $\Gamma_{0}(\mathfrak{m})$-module $V_{k,k}(\C_{p})^{\ast}$. We have identical isomorphisms for $\Gamma_{0}'(\mathfrak{m})$ and $\Gamma_{0}(\mathfrak{n})$.
\end{mlem}
\begin{longversion}
In fact, this is a special case of a more general result. We briefly recall the construction of the Borel--Serre compactification of the manifold $\Gamma_0(\mathfrak{m})\backslash \mathcal{H}_{3}$. In \cite{BS74} Borel and Serre construct a space $\overline{\mathcal{H}}_{3}$, containing $\mathcal{H}_{3}$, and prove that it is a manifold with corners and smooth boundary. More precisely we have
\[\overline{\mathcal{H}}_{3}= \mathcal{H}_{3} \sqcup \bigsqcup_{x \in \Proj(F)}e_{x},\]
where for each $x \in \Proj(F)$ we define $e_{x}\cong \mathbb{R}^{2}$. Moreover, $\overline{\mathcal{H}}_{3}$ is a contractible space. The action of $\Gamma_0(\mathfrak{m})$ on $\mathcal{H}_{3}$ extends to an action on $\overline{\mathcal{H}}_{3}$ and the quotient $\Gamma_0(\mathfrak{m})\backslash \overline{\mathcal{H}}_{3}$ is in fact a compact manifold with corners and it is called the \emph{Borel--Serre compactification} of $\Gamma_0(\mathfrak{m})\backslash \mathcal{H}_{3}$.
\end{longversion}
\begin{shortversion}
In fact, this is a special case of a more general result. Recall first the following definition.
\begin{mdef}
The \emph{Borel--Serre compactification} of $\Gamma_0(\mathfrak{m})\backslash\uhs$ is $\Gamma_0(\mathfrak{m})\backslash\overline{\uhp}_3,$ where
\[\overline{\uhp}_3 = \uhs \sqcup \bigsqcup_{x\in\Proj(F)}e_x,\]
with each $e_x$ a copy of $\R^2$ (see the appendix of \cite{Ser70} or \cite{BS74} for details).
\end{mdef}
\end{shortversion}
Let $M$ be a $\GLt(F)$-module.\footnote{More generally, we may replace $\GLt(F)$ with a semigroup $\Lambda\subset\GLt(F)$ containing $\Gamma_0(\m)$ and the matrices defining the Hecke operators.} Restricting the action to $\Gamma_0(\m)$, we obtain natural local systems on the space $\Gamma_0(\m)\backslash\uhs$, its compactification $\Gamma_0(\m)\backslash\overline{\uhs}$, and the boundary $\partial(\Gamma_0(\mathfrak{m})\backslash \overline{\mathcal{H}}_{3})$; in an abuse of notation, we write $\widetilde{M}$ for each of these local systems. For any $i\geq 0$ we have the following \begin{shortversion}Hecke equivariant\end{shortversion} isomorphisms:
\begin{equation} \label{group vs manifold 1}
\mathrm{H}^{i}\bigg(\Gamma_0(\mathfrak{m}), \mathrm{H}^{0}\big(\overline{\mathcal{H}}_{3}, M\big)\bigg) \cong \mathrm{H}^{i}\left(\Gamma_0(\mathfrak{m}) \backslash \mathcal{H}_{3}, \widetilde{M}\right);
\end{equation}
\begin{equation}\label{group vs manifold 2}
\mathrm{H}^{i}\bigg(\Gamma_0(\mathfrak{m}), \mathrm{H}^{0}\big(\partial\overline{\mathcal{H}}_{3}, M\big)\bigg) \cong \mathrm{H}^{i}\left(\partial\big[\Gamma_0(\mathfrak{m}) \backslash \overline{\mathcal{H}}_{3}\big], \widetilde{M}\right);
\end{equation}
\begin{equation}\label{group vs manifold 3}
\mathrm{H}^{i}\bigg(\Gamma_0(\mathfrak{m}), \mathrm{H}^{1}\big(\overline{\mathcal{H}}_{3}, \partial \overline{\mathcal{H}}_{3}, M\big)\bigg) \cong \mathrm{H}_{c}^{i+1}\left(\Gamma_0(\mathfrak{m}) \backslash \mathcal{H}_{3}, \widetilde{M}\right).
\end{equation}
\begin{longversion}
Using the action of $\GLt(F)$ on $M$ we can define Hecke operators on the cohomology groups on the left hand side of each isomorphism above, and we can prove that these isomorphisms are in fact Hecke equivariant.
\end{longversion}
\begin{mlem}
For each $i\geq 0$, the following diagram commutes:
\[\xymatrix@C=0.7em@R=1.7em{ \ar[r]& \mathrm{H}^{i}(\Gamma_0(\mathfrak{m}), \mathrm{Hom}(\Delta_{0}, M)) \ar[r]\ar^{\wr}[d]& \mathrm{H}^{i+1}(\Gamma_0(\mathfrak{m}),M) \ar[r]\ar^{\wr}[d]& \mathrm{H}^{i+1}(\Gamma_0(\mathfrak{m}), \mathrm{Hom}(\Delta, M)) \ar[r]\ar^{\wr}[d]& \\
\ar[r]& \mathrm{H}^{i+1}_{c}(\Gamma_0(\mathfrak{m})\backslash \mathcal{H}_{3}, \widetilde{M}) \ar[r]& \mathrm{H}^{i+1}(\Gamma_0(\mathfrak{m})\backslash \mathcal{H}_{3}, \widetilde{M}) \ar[r]& \mathrm{H}^{i+1}(\partial\left(\Gamma_0(\mathfrak{m})\backslash \overline{\mathcal{H}}_{3}\right) , \widetilde{M}) \ar[r]& .}\]
Here the vertical maps are isomorphisms, the horizontal sequences are exact, and each map is Hecke equivariant. Moreover, we have analogous diagrams for $\Gamma_{0}'(\mathfrak{m})$ and $\Gamma_{0}(\mathfrak{n})$.
\end{mlem}
\begin{proof} The proof follows the same strategy as \cite{AS86}. Let $\Delta\defeq \mathrm{Div}(\Proj(F))$. We know that the variety $\overline{\mathcal{H}}_{3}$ is contractible and its boundary, $\partial\overline{\mathcal{H}}_{3}$, is a disjoint union of copies of $\mathbb{R}^{2}$ indexed by $\Proj(F)$. Hence we have the following isomorphisms:
\[\mathrm{H}_{0}(\overline{\mathcal{H}}_{3}, \Z)\cong \Z \ \ \ \mathrm{and} \ \ \ \mathrm{H}_{0}(\partial\overline{\mathcal{H}}_{3}, \Z)\cong \Delta.\]
Additionally, the boundary map
$\mathrm{H}_{1}(\overline{\mathcal{H}}_{3}, \partial\overline{\mathcal{H}}_{3}, \Z) \rightarrow \mathrm{H}_{0}(\partial\overline{\mathcal{H}}_{3}, \Z)$
induces an isomorphism
$\mathrm{H}_{1}(\overline{\mathcal{H}}_{3}, \partial\overline{\mathcal{H}}_{3}, \Z)\cong\Delta_{0}.$
Moreover, all these isomorphisms are equivariant with respect to the action of $\mathrm{GL}_{2}(F)$, and assemble into a commutative diagram
\[
\xymatrix@R=1.7em{
0 \ar[r]& \mathrm{H}_{1}(\overline{\mathcal{H}}_{3}, \partial\overline{\mathcal{H}}_{3}\, \Z) \ar[r]\ar^{\wr}[d]& \mathrm{H}_{0}(\partial\overline{\mathcal{H}}_{3}, \Z)\ar[r]\ar^{\wr}[d]& \mathrm{H}_{0}(\overline{\mathcal{H}}_{3}, \Z) \ar[r]\ar^{\wr}[d]& 0\\
0 \ar[r]& \Delta_0 \ar[r]& \Delta \ar[r]& \Z \ar[r]& 0.
}
\]
Taking $\mathrm{Hom}(\bullet, M)$ in the diagram above we obtain the following commutative diagram where the horizontal sequences are exact and the vertical arrows are isomorphisms:
\begin{equation}
\label{before cohomology}
\xymatrix@R=1.7em{
0 \ar[r]& \mathrm{H}^{0}(\overline{\mathcal{H}}_{3}, M) \ar[r]\ar^{\wr}[d]& \mathrm{H}^{0}(\partial\left(\overline{\mathcal{H}}_{3}\right), M) \ar[r]\ar^{\wr}[d]& \mathrm{H}^{1}(\overline{\mathcal{H}}_{3}, \partial\overline{\mathcal{H}}_{3}, M) \ar[r]\ar^{\wr}[d]& 0\\
0 \ar[r]& M \ar[r]& \mathrm{Hom}(\Delta, M) \ar[r]& \mathrm{Hom}(\Delta_0 , M) \ar[r]& 0.
}
\end{equation}
The diagram above is compatible with the action of $\GLt(F)$ on each term. Considering the long exact sequence of the cohomology of $\Gamma_0(\mathfrak{m})$ attached to each line in the diagram above and using isomorphisms (\ref{group vs manifold 1}), (\ref{group vs manifold 2}) and (\ref{group vs manifold 3}) completes the proof. Since each diagram respects the action of $\GLt(F)$, the diagram of the lemma is Hecke equivariant.
\end{proof}
\begin{longversion}
\begin{mrem}Note that this also gives an isomorphism
\[\symb_{\Gamma_0(\n)}(V_{k,k}(L)^*) \cong \h^1_{\mathrm{c}}\big(\Gamma_0(\n)\backslash\uhs, \widetilde{V_{k,k}(L)^*}\big).\]
This cohomological setting gives the appropriate way of generalising modular symbols to other settings.
\end{mrem}
\end{longversion}
\subsection{Multiplicity one for group cohomology}
\begin{mthm}[Multiplicity one]\label{multiplicity one}
Let $f \in S_{k,k}(\Omega_0(\n))^{\pri-\mathrm{new}}$. Then the system of Hecke eigenvalues corresponding to $f$ appears in the cohomology groups
\[\h^i_{\mathrm{c}}(\Gamma_0(\n)\backslash\uhs,\widetilde{V_{k,k}(\Cp)^*})\]
(only) for $i = 1,2$, and the $f$-isotypic component of these spaces is one-dimensional. The system of eigenvalues corresponding to $f$ does not appear in the cohomology of $\Gamma_0(\m)\backslash\uhs$ or $\Gamma'_0(\m)\backslash\uhs$.
\end{mthm}
\begin{proof}
For the first statment, see \cite{Hid94}, Proposition 3.1, which is given for cuspidal cohomology. The cuspidal cohomology can be viewed as a subspace of the compactly supported cohomology, and cokernel of the inclusion is an Eisenstein subspace, that is, the $f$-isotypic parts of $\h^i_{\mathrm{cusp}}$ and $\h^i_{\mathrm{c}}$ are the same (see \cite{Har87}, Section 3.2.5, or \cite{Urb95}, Theorem 3.2). The second statement holds simply because $f$ is new at $\pri$, so its system of eigenvalues cannot appear at level $\m$.
\end{proof}
The main result of this section is the following:
\begin{mthm}\label{cohomology dimension}
Let $H$ be any space with an action of the Hecke operators $T_I$ for $I$ coprime to $\n$, and let $H_{(f)}$ denote the $f$-isotypic component.
\begin{itemize}
\item[(i)] We have $\h^i(\Gamma_0(\m),\Delta_{k,k})_{(f)} = \h^i(\Gamma'_0(\m),\Delta_{k,k})_{(f)} = 0$, for $i = 0,1$.
\item[(ii)] The space $\h^0(\Gamma_0(\n),\Delta_{k,k})_{(f)}$ is a one-dimensional $\Cp$-vector space.
\item[(iii)] The space $\h^1(\Gamma, \Delta_{k,k})_{(f)}$ is a one-dimensional $\Cp$-vector space.
\end{itemize}
\end{mthm}
\begin{proof}
Parts (i) and (ii) are immediate from Lemma \ref{ash stevens} and Theorem \ref{multiplicity one}. A direct calculation shows that the maps in the long exact sequence of Proposition \ref{long exact sequence} are equivariant with respect to the Hecke operators, and hence we get a corresponding exact sequence on the $f$-isotypic parts. But from parts (i) and (ii), the end terms are zero whilst the second term is one-dimensional. Part (iii) follows.
\end{proof}
\begin{mrem}
An identical proof would show that the $f$-isotypic component of $\h^2(\Gamma,\Delta_{k,k})$ is also one-dimensional.
\end{mrem}
\section{Exceptional zeros of Bianchi modular forms}
In this section, we draw our previous results together to prove statements about exceptional zeros.
\begin{mlem}
The cohomology class $\mathrm{oc}_f$ is non-zero.
\end{mlem}
\begin{proof}We exploit a result of Rohrlich on the non-vanishing of $L$-values, namely the theorem in the introduction of \cite{Roh91}. If $\mathrm{oc}_f$ vanishes, then by Corollary \ref{oc and L-values} so must the $L$-values $L(f,\chi,k/2+1)$ for all Dirichlet characters $\chi$ with $\chi(\pi) = \omega$, and this contradicts Rohrlich's result.
\end{proof}
\begin{mdef}
Define $\mathcal{L}_{\pri} \in \Cp$ to be the unique scalar such that
\[\mathrm{lc}_f = \mathcal{L}_{\pri}\mathrm{oc}_f \in \h^1(\Gamma,\Delta_{k,k})_{(f)},\]
which is possible by Theorem \ref{cohomology dimension}.
\end{mdef}
The following is the main result of this paper.
\begin{mthm}
Let $f \in S_{k,k}(\Omega_0(\n))$ be a Bianchi eigenform over an imaginary quadratic field of class number 1, and suppose that $f$ is new at $\pri = (\pi)$ and has small slope at every other prime above $p$. Let $-\omega$ be the eigenvalue of the Atkin--Lehner operator $W_{\pri}$, and let $\chi$ be a finite order Hecke character with $\chi(\pri) = \omega$. Let $\ff$ be the $p$-part of the conductor of $\chi$. Then there is an exceptional zero $L_p(f,\chi,\mathbf{k/2}) = 0$, and there exists $\mathcal{L}_{\pri} \in \Cp$ such that
\begin{align*}
\frac{d}{ds_{\pri}}&L_p(f,\chiw_{\mathrm{Tm}}^{k/2},\mathbf{s})\big|_{\mathbf{s} = \mathbf{k/2}} =\\
& \mathcal{L}_{\pri}\left(\prod_{\substack{\mathfrak{q}|p\\ \mathfrak{q}\neq \pri}}Z_{\mathfrak{q}}(\chi,k/2)\right)\left[\frac{D^{k/2+1}\tau(\chi^{-1})|c|^r \#\roi_F^\times}{\Omega_f\lambda_{\ff}}\right] \Lambda(f,\chi,k/2+1).
\end{align*}
Moreover, $\mathcal{L}_{\pri}$ is independent of the choice of $\chi$ satisfying the above conditions.
\end{mthm}
\begin{proof}
The existence of an exceptional zero was Corollary \ref{exceptional zero existence}. The result follows from Corollaries \ref{oc and L-values} and \ref{lc and L-values} and the definition of $\mathcal{L}_{\pri}$.
\end{proof}
\section{An arithmetic description in the base-change case}
Let $f$ be a cuspidal Bianchi modular form of weight $(k,k)$. In the case where $f$ is the base-change of a classical cusp form $\tilde{f}$ (of weight $k+2$), new at $p$, then we are able to relate the $\mathcal{L}$-invariants at $p$ of $f$ and $\tilde{f}$ under a non-vanishing assumption. Indeed, Artin formalism says that there is a relation
\[\Lambda(f,s) = \Lambda(\tilde{f},s)\Lambda(\tilde{f},\chi_{F/\Q},s),\]
where $\chi_{F/\Q}$ is the quadratic character associated to the extension $F/\Q$.
\begin{longversion}We can extend this to be true of the values of $p$-adic $L$-functions at critical values.
\begin{mlem}
Let $\chi$ be a (rational) Dirichlet character of conductor $cp^n$, where $c$ is coprime to $p$ and $n\geq 1$. Then we have
\[\left[\frac{D^{2(r+1)}(cp^n)^{2(r+1)}}{\tau(\chi\cdot N_{F/\Q})\tau(\chi_F)\lambda_{p^n}(f)}\right]\Lambda(f,\chi,r+1) = \left[\frac{(cp^n)^{r+1}}{\tau(\chi)\lambda_{p^n}(\tilde{f})}\right]\Lambda(\tilde{f},\chi,r+1)\left[\frac{(cp^nD)^{r+1}}{\tau(\chi\chi_F)\lambda_{p^n}(\tilde{f})}\right]\Lambda(\tilde{f},\chi\chi_F,r+1).\]
In particular, under suitable normalisations, we have
\[(2/ \#\roi_F^\times)L_p(f,(\chi\cdot N_{F/\Q})w_{\mathrm{Tm}}^r,\mathbf{r}) = L_p(\tilde{f},\chi,\mathbf{r}w_{\mathrm{Tm}}^r, \mathbf{r})L_p(\tilde{f},\chi\chi_Fw_{\mathrm{Tm}}^r,\mathbf{r}).\]
\end{mlem}
\begin{proof}
Most of the constant terms are evidently equal, and the equality of the $L$-function terms is Artin formalism. The only things left to check are the eigenvalues and the Gauss sums. But $\lambda_p(f) = \lambda_p(\tilde{f})^2$ by inspection, whilst we have $\tau(\chi\cdot N_{F/\Q}) = \tau(\chi)^2\tau(\chi_F)$ by an explicit calculation combined with a characteristic zero version of the classical Hasse--Davenport identity for Gauss sums. If we let $\Omega_{\tilde{f}}^\pm$ be the periods of $\tilde{f}$, then $\Omega_f \defeq \Omega_{\tilde{f}}^+\Omega_{\tilde{f}}^-\tau(\chi_F)$ is a period of $f$, and using these periods, the statement about $p$-adic $L$-values follows immediately from the respective interpolation formulae.
\end{proof}
Given an equality at critical values, we can lift this result to $p$-adic $L$-functions:
\end{longversion}
\begin{shortversion}We can lift this to the level of $p$-adic $L$-functions.
\end{shortversion}
\begin{mprop}
Suppose $f$ is the base-change to $F$ of a classical modular form that has small slope at $p$. Given periods $\Omega_{\tilde{f}}^\pm$ for $\tilde{f}$, we may choose the period of $f$ to be $\Omega_f = \Omega_{\tilde{f}}^+\Omega_{\tilde{f}}^-\tau(\chi_{F/\Q})$. The natural inclusion $\Zp \hookrightarrow \roi_F\otimes_{\Z}\Zp$ allows us to restrict the resulting $p$-adic $L$-function of $f$ to an analytic function on $\Zp$ (the \emph{cyclotomic variable}). This restriction factorises as
\[L_p(f,s) = \frac{ \#\roi_F^\times}{2}\cdot L_p(\tilde{f},s)\cdot L_p(\tilde{f},\chi_{F/\Q},s)\]
for all $s \in \Zp$.
\end{mprop}
\begin{proof}
\begin{shortversion}
The $p$-adic $L$-functions of $f$ and $\tilde{f}$ are defined using canonical overconvergent modular symbols $\Psi_f$ and $\Psi_{\tilde{f}}$ respectively, which give rise to locally analytic distributions $\mu_p$ on $\cl_F(p^\infty)$ and $\tilde{\mu}_p$ on $\Zp^\times$ respectively (see \cite{PS11} for the classical case). We can restrict $\mu_p$ to functions that factor through the norm map and hence consider it as a distribution on $\Zp^\times.$ Finally, define a locally analytic distribution $\tilde{\mu}^{\chi_{F/\Q}}_p$ on $\Zp^\times$ as follows: use $\Psi_{\tilde{f}}$ to define a distribution $\tilde{\mu}_{p,D'}$ on $(\Z/D')^\times \times \Zp^\times$ as in Section \ref{padic lfunction section}, where $D'$ is the prime to $p$ part of $D$, and define
\[\tilde{\mu}_p^{\chi_{F/\Q}}(\zeta) = \tilde{\mu}_{p,D'}\left(\chi_{F/\Q}\cdot\zeta\right),\]
where we consider $\chi_{F/\Q}\cdot\zeta$ as a function on $(\Z/D')^\times\times\Zp^\times$ in the natural way.
\end{shortversion}
\begin{longversion}
Recall the definition of $L_p(f,s)$ from Definition \ref{padiclfunctiondefinition}: we defined a distribution $\mu_p$ on $\cl_F(p^\infty)$ and constructed an analytic function on $\roi_F\otimes_{\Z}\Zp$ using it. The $p$-adic $L$-function of $\tilde{f}$ can be constructed similarly (see \cite{PS11}, for example) using a distribution $\tilde{\mu}_p$ on $\cl_{\Q}^+(p^\infty) \cong \Zp^\times$. We can restrict $\mu_p$ to the characters that factor through the norm map $N_{F/\Q} : \cl_F(p^\infty) \rightarrow \cl_{\Q}^+(p^\infty)$ to obtain a distribution on $\Zp^\times$. Finally, define a locally analytic distribution $\tilde{\mu}_p^{\chi_{F/\Q}}$ on $\Zp^\times$ as follows: extend $\tilde{\mu}_p$ to a distribution $\tilde{\mu}_{p,D'}$ on $(\Z/D')^\times \times\Zp^\times$ as in Section \ref{padic lfunction section}, where $D'$ is the prime to $p$ part of $D$, and define
\[\tilde{\mu}_p^{\chi_{F/\Q}}(\zeta) = \tilde{\mu}_{p,D'}\left(\chi_{F/\Q}\cdot\zeta\right),\]
where we consider $\chi_{F/\Q}\cdot\zeta$ as a function on $(\Z/D')^\times\times\Zp^\times$ in the natural way.
\end{longversion}
$\\ \\$
The result follows if we can prove the (stronger) result that for any locally analytic function $\zeta$ on $\Zp^\times$, we have
\[\mu_p(\zeta\circ N_{F/\Q}) = \frac{ \#\roi_F^\times}{2}\cdot\tilde{\mu}_p(\zeta)\cdot\tilde{\mu}_p^{\chi_{F/\Q}}(\zeta).\]
For $\zeta$ of the form $\chi|\cdot|^r$, for $0 \leq r \leq k$ and $\mathrm{cond}(\chi)|p^\infty$, it is a lengthy but elementary\footnote{The most difficult part of the calculation is to show that the respective Gauss sum terms agree when $p$ is inert; this is a characteristic zero analogue of the Hasse--Davenport identity.} check that this holds by the respective interpolation formulae and the relation between the classical $L$-functions of $f$ and $\tilde{f}$ above. But both of these distributions are admissible of order $k < k+1$, and hence equality on this set of points forces equality everywhere, and the proof is complete.
\end{proof}
Now suppose that $f$ is new (hence small slope) at $p$, so $L_p(f,s)$ has an exceptional zero at $s = k/2$, and that $p$ is inert in $F$. In this case, $\chi_{F/\Q}(p) = -1$, and accordingly we see that
$L_p(\tilde{f},s)$ has an exceptional zero at $s = k/2$ whilst $L_p(\tilde{f},\chi_{F/\Q},s)$ does not. Differentiating, we see that
\[\frac{d}{ds}L_p(f,s)\big|_{s=k/2} = \frac{ \#\roi_F^\times}{2}\left[\frac{d}{ds}L_p(\tilde{f},s)\big|_{s=k/2}\right]L_p(\tilde{f},\chi_{F/\Q},k/2).\]
Using the respective interpolation conditions (for the $p$-adic $L$-functions of both $f$ and $\tilde{f}$),
\begin{longversion}collapsing the exceptional factor in the second classical term, and using classical Artin formalism, we find that
\begin{align*}\mathcal{L}_{\pri}\cdot\left[\frac{|D|^{k/2+1}w}{\Omega_f}\right] \Lambda(f,k/2+1) = \mathcal{L}_p(\tilde{f})\cdot\left[\frac{2|D|^{k/2+1}}{\Omega_{\tilde{f}}^+\Omega_{\tilde{f}}^-\tau(\chi_{F/\Q})}\right]\cdot\Lambda(f,k/2+1),
\end{align*}
where $\mathcal{L}_p(\tilde{f})$ denotes the $\mathcal{L}$-invariant of $\tilde{f}$ at $p$. Without loss of generality, we may renormalise so that $\Omega_f = \Omega_{\tilde{f}}^+\Omega_{\tilde{f}}^-\tau(\chi_{F/\Q})$. Then we deduce the following:
\end{longversion}
\begin{shortversion}
we deduce the following:
\end{shortversion}
\begin{mprop}
Suppose $f$ is the base-change of a classical modular form $\tilde{f}$ to $F$, and that $p$ is inert in $F$. If $\Lambda(f,k/2+1) \neq 0$, we have
\[\mathcal{L}_{\pri} = 2\mathcal{L}_p(\tilde{f}),\]
where $\mathcal{L}_p(\tilde{f})$ is the $\mathcal{L}$-invariant of $\tilde{f}$.
\end{mprop}
A similar argument shows that if $p$ ramifies in $F$, then $\mathcal{L}_{\pri} = \mathcal{L}_p(\tilde{f})$ (the difference arising when we consider the exceptional factor in the second classical term, which equals 2 in the inert case and 1 in the ramified case). More generally, since the $\mathcal{L}$-invariant is purely local, similar methods show that these results also hold if there exists a Dirichlet character $\chi$ over $F$, factoring through the norm to $\Q$, such that $\Lambda(f,\chi,k/2+1) \neq 0$.
\begin{mrems}
\begin{itemize}
\item[(i)]In \cite{Tri06}, Trifkovi\'{c} uses local methods to construct conjectural global points on modular elliptic curves over imaginary quadratic fields. He associates a (conjectural) weight $(0,0)$ Bianchi modular form $f$ to such a curve and defines the points using double integrals of the sort considered in this paper. His main conjecture is Conjecture 6, but this is only well-defined given his previous Conjecture 5, which is the statement that a value of the double integral generates a lattice commensurable with $q^{\Z}$ (where $q$ is the Tate period of the elliptic curve). In the analogous case over $\Q$, this is known via the results of \cite{GS93}, but it remains open over imaginary quadratic fields in general. When $f$ is base-change and $p$ is inert or ramified, the above proposition combined with \cite{GS93} gives a proof of Trifkovi\'{c}'s Conjecture 5.
\item[(ii)] Results of a similar nature for abelian base-change in the case of certain weight 2 automorphic forms for $\GLt$ were recently obtained by Gehrmann in \cite{Geh17}.
\end{itemize}
\end{mrems}
\section{Further remarks}
It is natural to expect exceptional zero results of the form above to exist in wider generality. We conclude by commenting on possible extensions of the results of this paper.
\subsection{Higher class number}\label{class number}
The most obvious restriction we've assumed in the work above is that the class number of $F$ is 1. This is not a serious restriction to the method, but rather a choice made for ease of exposition; indeed, the authors carried out a large portion of the work without assuming any restriction on the class number at all, but found that the increase in technicalities made the arguments much longer and harder to follow. To generalise whilst retaining classical methods, one would need to consider functions on $h$ copies of the tree, where $h$ is the class number, and find elements in the direct sum of $h$ group cohomology groups, and the Hecke operators permute these functions/classes.
\subsection{Higher derivatives}
In the case where $p$ is split, it is possible that even $(d/ds_{\pri})L_p(f,\chi,\mathbf{s})|_{\mathbf{s} = \mathbf{k/2}} = 0$ even when the corresponding classical $L$-value is non-zero, since we are left with an exceptional factor at the other prime $\pribar$ that could vanish. The most natural solution to this would be to differentiate again, this time in the $s_{\pribar}$-direction. In considering this approach, the authors (directly) constructed classes $\mathrm{oc}_{f},\mathrm{lc}_f \in \h^2(\Gamma,\Delta_{k,k})$ that again both live in a one-dimensional eigenspace, and hoped to relate their special values to $L$-values. However, whilst it seemed likely that such results would exist at the level of \emph{cocyles}, there was a fundamental obstacle to this approach in that these special values were no longer well-defined on the corresponding cohomology classes. Results of this nature do exist, however, when the weight of $f$ is 2 (parallel weight 0); over totally real fields, this was done by Spie{\ss} (see \cite{Spi14}), and more recently over arbitrary number fields by Deppe and Bergunde (see \cite{Dep16} and \cite{Ber17}). Their approach, which whilst cohomological is quite different to that considered here, is to construct higher cohomology classes by taking cup products of cycles in $\h^1$. It would be interesting to generalise their results to arbitrary weights.
\subsection{Arithmetic descriptions of $\mathcal{L}_{\pri}$ in general}
We have already seen that it is possible to obtain an arithmetic description of the $\mathcal{L}$-invariant in the case that $f$ is the base-change of a classical modular form. Evidently it is desirable to obtain a similar description when $f$ is not base-change. This case, however, appears far harder. The difficulty arises from the absence of nice geometry, which is a fundamental block to generalising classical methods of determining the $\mathcal{L}$-invariant in arithmetic terms. One of the most interesting cases to study is when the form is weight $(0,0)$ (also referred to as weight 2) and corresponds, via the (conjectural) imaginary quadratic version of modularity, to an elliptic curve over $F$. In this situation, the elliptic curve will have multiplicative reduction at $\pri$, so there is a Tate uniformiser $q \in F_{\pri}^\times$ such that $E(\overline{F}_{\pri}) \cong \overline{F}_{\pri}^\times/q^{\Z}.$ By analogy with the case of Hilbert modular forms, one would expect that:
\begin{mcon}\label{exp desc}Let $Q_{\pri} \defeq \mathrm{Norm}_{F_{\pri}/\Qp}(q),$
and let $f_{\pri} \defeq [\roi_{\pri}/\pri:\Fp]$ denote the inertia degree of $\pri$.
Then we have
\[\mathcal{L}_{\pri} = f_{\pri}\frac{\mathrm{log}_{p}(Q_{\pri})}{\mathrm{ord}_{p}(Q_{\pri})}.\]
\end{mcon}
(Compare \cite{Mok09}, Theorem 8.2 and \cite{Hid09}, Theorem 1.2). The computational results of Trifkovi\'{c} in\cite{Tri06}, and later of Guitart, Masdeu and \c{S}eng\"{u}n in \cite{GMS15}, can be seen as providing strong evidence for this conjecture. More generally, for a form $f$ of arbitrary weight, let $\rho_{f}$ be the Galois representation attached to $f$ by Taylor (see \cite{Tay94}). A natural generalisation of Conjecture \ref{exp desc} is:
\begin{mcon} The $\mathcal{L}$-invariant $\mathcal{L}_{\mathfrak{p}}$ is the Benois--Greenberg $\mathcal{L}$-invariant attached to $\rho_{f}$ (see \cite{Ben11} and \cite{Ros15}).
\end{mcon}
The authors hope to address these questions in future work.
\small
\renewcommand{\refname}{\normalsize References}
|
\section{Appendix}
\label{sec:Appendix}
In this appendix we explain some details about the general chemical potential flow theory, which is the basis for the flow equations \eqref{eq:FlowXiFiniteT} and \eqref{eq:FlowChiFiniteT}. Since the theory relies on a reinterpretation of the chemical potential as a flow parameter, the main results can be transfered directly from Ref.~[\onlinecite{FrasdorfMosig2017}].
The starting point for the derivation of an exact flow equation is the $\mu$-dependent partition function $Z_{\mu}[\boldsymbol{\eta}, \boldsymbol{J}]$. It is defined as the functional Fourier transform of the exponentiated bare action $S_{\mu}[\boldsymbol{\psi}, \boldsymbol{\phi}]$,~\cite{BergesMesterhazy2012, KamenevBook, NegeleOrlandBook, KopietzBook}
\begin{align}
Z_{\mu}[\boldsymbol{\eta}, \boldsymbol{J}] = \int\! \mathcal{D}\psi \mathcal{D}\psi^{\dagger} \mathcal{D}\phi \, e^{iS_{\mu}[\boldsymbol{\psi}, \boldsymbol{\phi}] + i \boldsymbol{\eta^{\dagger}} \tau_1 \boldsymbol{\Psi} + i \boldsymbol{\Psi^{\dagger}} \tau_1 \boldsymbol{\eta} + i \boldsymbol{\phi}^{\intercal} \tau_1 \boldsymbol{J}}.
\label{eq:PartitionFunction}
\end{align}
The bare action is a functional of fermionic and bosonic fields, which can be derived from the purely fermionic Hamiltonian~\eqref{eq:GrandCanonicalHamiltonian} by a standard procedure.~\cite{SchwieteFinkelstein2014-1, NegeleOrlandBook, KamenevBook, FrasdorfMosig2017} The bosonic field is introduced by a Hubbard-Stratonovich transformation of the Coulomb interaction term in the density-density channel. The index $\mu$ indicates that both the partition function and the bare action depend on the chemical potential. The chemical potential dependence of the bare action enters explicitly via the quadratic $\mu$-term in the Hamiltonian and implicitly via the background density $\tilde{n}_{\mu}$. In contrast to the conventional fRG there is no additional infrared regulator.~\cite{BergesEtal2003} Furthermore, we work in the real-time Keldysh formalism,~\cite{GezziPruschkeMeden2006, JakobsMedenSchoeller2007, GasenzerPawlowski2008, BergesHoffmeister2009, KarraschEtal2010, KlossKopietz2011, KennesEtal2012, BergesMesterhazy2012} which involves a doubling of degrees of freedom, with classical $(c)$ and quantum $(q)$ component for each field~\cite{SchwieteFinkelstein2014-1, BergesMesterhazy2012, FrasdorfMosig2017, KamenevBook}
\begin{equation}
\boldsymbol{\Psi} \equiv \begin{pmatrix} \Psi_c & \Psi_q \end{pmatrix}^{\intercal} \,, \quad \boldsymbol{\Psi^{\dagger}} = (\boldsymbol{\Psi})^{\dagger} \,, \quad
\boldsymbol{\phi} \equiv \begin{pmatrix} \phi_c & \phi_q \end{pmatrix}^{\intercal} \,.
\label{eq:DefinitionClassicalQuantumFields}
\end{equation}
Lastly, $\boldsymbol{\eta}$ and $\boldsymbol{J}$ are fermionic and bosonic source fields, respectively. In Eq
~\eqref{eq:PartitionFunction}
we employed a condensed vector notation for the source terms, containing integration and summation of continuous and discrete field degrees of freedom implicitly, \textit{e.g.}
\begin{equation}
\boldsymbol{\eta^{\dagger}} \tau_1 \boldsymbol{\Psi} \equiv \int_x \boldsymbol{\eta^{\dagger}}(x) \tau_1 \boldsymbol{\Psi}(x) \,,
\label{eq:CondensedNotation}
\end{equation}
where $\tau_1$ is a Pauli matrix acting in Keldysh space.
The effective action may now be introduced as the modified Legendre transform of the connected functional $W_{\mu}[\boldsymbol{\eta}, \boldsymbol{J}] = -i \textrm{ln} Z_{\mu}[\boldsymbol{\eta}, \boldsymbol{J}]$,\cite{Wetterich2001, BergesTetradisWetterich2002, BergesEtal2003, MetznerHonercampEtal2012, KopietzBook}
\begin{align}
\Gamma_{\mu}[\boldsymbol{\psi}, \boldsymbol{\phi}] = \, &W_{\mu}[\boldsymbol{\eta}_{\mu}, \boldsymbol{J}_{\mu}] - \boldsymbol{\eta^{\dagger}}_{\mu} \tau_1 \boldsymbol{\Psi} - \boldsymbol{\Psi^{\dagger}} \tau_1 \boldsymbol{\eta}_{\mu} - \boldsymbol{\phi}^{\intercal} \tau_1 \boldsymbol{J}_{\mu} \nonumber \\
&- \boldsymbol{\Psi}^{\dagger} \boldsymbol{\hat{R}}_{\mu} \boldsymbol{\Psi} \,,
\label{eq:LegendreTransform}
\end{align}
with
\begin{align}
\boldsymbol{\hat{R}}_{\mu}(x, y) &= \begin{pmatrix} 0 & \mu \, \delta(x - y) \hat{\mathbb{1}} \\ \mu \, \delta(x - y) \hat{\mathbb{1}} & 0 \end{pmatrix} \,.
\label{eq:KeldyshStructureFermionicRegulators}
\end{align}
The term $\boldsymbol{\Psi}^{\dagger} \boldsymbol{\hat{R}}_{\mu} \boldsymbol{\Psi}$ is the explicit chemical-potential term one obtains in the bare action $S_{\mu}[\boldsymbol{\psi}, \boldsymbol{\phi}]$. Its resemblance with an additive infrared regulator in the conventional fRG is the foundation of the chemical-potential flow theory.~\cite{BergesEtal2003} According to the usual definition of the effective flowing action the ``chemical-potential regulator term'' has been subtracted on the right hand side. Consequently the effective action $\Gamma_{\mu}$ involves flowing vertex functions only, and the explicit chemical-potential term - in comparison to the bare action - is absent. Note that some authors prefer to include a finite chemical potential in the fermionic distribution function, rather than in the spectral part of the inverse propagators as we do here.~\cite{KamenevBook} Such an alternative choice would affect the structure of the regulator~\eqref{eq:KeldyshStructureFermionicRegulators} and the vertex expansion of the effective action, but it cannot lead to any observable consequences, since these two choices are connected by a (time dependent) gauge transformation.
The exact chemical potential flow equation follows immediately upon taking the $\mu$-derivative of Eq.~\eqref{eq:LegendreTransform}, keeping the fields $\boldsymbol{\psi}$ and $\boldsymbol{\phi}$ fixed
\begin{align}
\partial_{\mu} \Gamma_{\mu}[\boldsymbol{\psi}, \boldsymbol{\phi}] = \frac{i}{2} \partial \!\!\!/_{\mu} \textrm{STr} \, \textrm{ln} \left( \boldsymbol{\hat{\Gamma}}_{\mu}^{(2)}[\boldsymbol{\psi}, \boldsymbol{\phi}] + \boldsymbol{\hat{\mathcal{R}}}_{\mu} \right) + 2 \phi_q \partial_{\mu} \tilde{n}_{\mu} \,,
\label{eq:ExactFlowEquation}
\end{align}
where $\boldsymbol{\hat{\Gamma}}_{\mu}^{(2)}$ is a Hesse matrix of second derivatives, and
\begin{equation}
\boldsymbol{\hat{\mathcal{R}}}_{\mu} = \textrm{diag} \left( - \boldsymbol{\hat{R}}_{\mu}, \boldsymbol{\hat{R}}_{\mu}^{\intercal}, 0 \right) \,.
\label{eq:RegulatorMatrix}
\end{equation}
The ``single-scale derivative'' $\partial \!\!\!/_{\mu}$ in Eq.~\eqref{eq:ExactFlowEquation} only acts on the regulator $\boldsymbol{\hat{\mathcal{R}}}_{\mu}$. The above flow equation is the Keldysh analog of the original, imaginary-time flow equation proposed by Berges \textit{et al.}~\cite{BergesEtal2003} The flow equations for the one-particle irreducible vertex functions are obtained by expanding the effective action in powers of fields, which needs to be inserted into the above equation, and comparing coefficients. Since we are only interested in thermal equilibrium, where the fluctuation-dissipation theorem holds,~\cite{KamenevBook} we only need to consider the resulting flow equations for the retarded components of the self-energy and polarization function. In a condensed notation, where numerical arguments denote space and time coordinates, $1 \equiv (\vec{r}_1, t_1)$, and latin indices encompass the discrete fermionic degrees of freedom, sublattice, valley and spin, these flow equations read
\begin{widetext}
\begin{align}
\partial_{\mu} \Sigma_{\mu, ij}^R(1, 2) = i \partial \!\!\!/_{\mu} \sum_{k, l} \int\nolimits'\ \Big( &\Gamma_{\mu, ik}^{qcc}(1, 1'; 4') G_{\mu, kl}^K(1', 2') \Gamma_{\mu, lj}^{ccq}(2', 2; 3') D_{\mu}^A(3', 4') \nonumber \\
+ &\Gamma_{\mu, ik}^{qcc}(1, 1'; 4') G_{\mu, kl}^R(1', 2') \Gamma_{\mu, lj}^{qcc}(2', 2; 3') D_{\mu}^K(3', 4') \Big) \,,
\label{eq:FlowEqRetardedSelfenergy}
\end{align}
\begin{align}
\partial_{\mu} \Pi_{\mu}^R(1, 2) = \frac{i}{2} \partial \!\!\!/_{\mu} \sum_{k, l, m, n} \int\nolimits'\ \Big( &G_{\mu, kl}^K(1', 2') \Gamma_{\mu, lm}^{ccq}(2', 3'; 1) G_{\mu, mn}^R(3', 4') \Gamma_{\mu, nk}^{qcc}(4', 1'; 2) \nonumber \\
+ &G_{\mu, kl}^A(1', 2') \Gamma_{\mu, lm}^{ccq}(2', 3'; 1) G_{\mu, mn}^K(3', 4') \Gamma_{\mu, nk}^{cqc}(4', 1'; 2) \Big) \,.
\label{eq:FlowEqRetardedPolarization}
\end{align}
\end{widetext}
The functions $\Gamma_{\mu, ij}^{\alpha \beta \gamma}(1, 2; 3)$, with $\alpha, \beta, \gamma = c, q$ are the Fermi-Bose three-vertices of the theory in the real-time Keldysh formulation. The primed integration sign indicates that all primed arguments have to be integrated. A transformation to Fourier-space is beneficial, due to energy and momentum conservation. The flowing frequency-momentum space propagators that enter the transformed flow equations read
\begin{subequations}
\begin{align}
\hat{G}_{\mu}^{R/A}(\vec{k}, \varepsilon) &= \frac{1}{\sigma_0^s \otimes \left(\varepsilon + \mu - \vec{\Sigma} \cdot \vec{k} - \hat{\Sigma}_{\mu}^{R/A}(\vec{k}, \varepsilon) \right)} \,, \\
\hat{G}_{\mu}^{K}(\vec{k}, \varepsilon) &= \textrm{tanh} \frac{\varepsilon}{2T} \left( \hat{G}_{\mu}^R(\vec{k}, \varepsilon) - \hat{G}_{\mu}^A(\vec{k}, \varepsilon) \right) \,,
\label{eq:FermionPropagators}
\end{align}
\end{subequations}
\begin{subequations}
\begin{align}
D_{\mu}^{R/A}(\vec{q}, \omega) &= \frac{1}{2} \frac{1}{V^{-1}(\vec{q}) + \Pi_{\mu}^{R/A}(\vec{q}, \omega)} \,, \\
D_{\mu}^K(\vec{q}, \omega) &= \textrm{coth} \frac{\omega}{2T} \left( D_{\mu}^R(\vec{q}, \omega) - D_{\mu}^A(\vec{q}, \omega) \right) \,.
\label{eq:BosonPropagators}
\end{align}
\end{subequations}
The single-scale derivative in Eqs.~\eqref{eq:FlowEqRetardedSelfenergy} and \eqref{eq:FlowEqRetardedPolarization} only acts on the flowing fermionic propagators, substituting the latter by a single-scale propagator
\begin{equation}
\partial \!\!\!/_{\mu} \hat{G}_{\mu}^{R/A}(\vec{k}, \varepsilon) = - \left( \hat{G}_{\mu}^{R/A}(\vec{k}, \varepsilon) \right)^2 \,.
\label{eq:SingleScalePropagator}
\end{equation}
Here, the $\mu$-dependence of the flowing self-energy is held constant upon taking the single-scale derivative $\partial \!\!\!/_{\mu}$. By using the approximations mentioned in the main text - that is setting all the three-vertices to unity and neglecting any dynamical effects - we arrive after a straightforward calculation at the flow equations \eqref{eq:FlowXiFiniteT} and \eqref{eq:FlowChiFiniteT}.
|
\section{Introduction and main results}
The graph of a smooth function $ u : \mathbb{R}^N \to \mathbb{R}$ is a minimal surface in $\mathbb{R}^{N+1}$ if and only if $u$ is a solution to the minimal surface equation
\begin{equation} \label{EqSupMin}
- {\rm div} \left( \frac{\nabla u}{\sqrt{1 + \vert \nabla u \vert^2 }} \right) = 0 \qquad {\rm in} \qquad \mathbb{R}^N, \qquad N\ge2.
\end{equation}
In his work \cite{Ber} (see also \cite{H}) S.N. Bernstein proved that any smooth solution $u$ to the minimal surface equation in $\mathbb{R}^2$ must be an affine function.
This result has been extended to $\mathbb{R}^3$ by E. De Giorgi \cite{DG}, to $\mathbb{R}^4$ by J.F. Almgren \cite{Alm} and, up to dimension $N=7$, by J. Simons \cite{Sim}. On the other hand, in the celebrated paper \cite{BDeGG}, E. Bombieri, E. De Giorgi and E. Giusti proved the existence of a non-affine entire solution of the minimal surface equation \eqref{EqSupMin} for any $ N \ge 8$. Nevertheless, J. Moser \cite{Mos} was able to prove that, if $ \nabla u $ is bounded on $\mathbb{R}^N$, then $u$ must be again an affine function, and this result holds true for every dimension $ N\ge2$. Later, E. Bombieri and E. Giusti \cite{BG} generalized Moser's result by assuming that \emph{only} $N-1$ partial derivatives of $u$ are bounded on $\mathbb{R}^N$, $N\ge2$.
In the recent work \cite{F} the author has further improved this result. More precisely, in \cite{F} it is demonstrated that any smooth solution $u$ to the minimal surface equation \eqref{EqSupMin} with $N-1$ partial derivatives bounded on one side (not necessarily the same) is necessarily an affine function.
The main result of the present work is the following
\bigskip
\begin{theorem} \label{Th.N-7} Assume $ N \ge 8$ and let $u$ be a solution of the minimal surface equation
\begin{equation} \label{EqSupMin2}
- {\rm div} \left( \frac{\nabla u}{\sqrt{1 + \vert \nabla u \vert^2 }} \right) = 0 \qquad {\rm in} \qquad \mathbb{R}^N.
\end{equation}
\noindent If $N-7$ partial derivatives of $u$ are bounded on one side (not necessarily the same) then $u$ is an affine function.
\end{theorem}
Theorem \ref{Th.N-7} is sharp. Indeed, it cannot hold true if one assumes that only $N-8$ partial derivatives are bounded on one side. To see this, let us denote by $U=U(x_1,...,x_8)$ the non-affine solution of \eqref{EqSupMin2} in $\mathbb{R}^8$ constructed in \cite{BDeGG} and, for every $ N\ge 8$, set $u_N(x_1,...,x_N) := U(x_1,...,x_8)$. Clearly, $u_N$ solves \eqref{EqSupMin2}, it has $N-8$ partial derivatives that are identically zero, but $u_N$ is not an affine function.
\medskip
The proof of the previous theorem is based on the following result, which is interesting in its own.
\medskip
\begin{theorem} \label{Splitting} Let $u$ be a {\underline {non-affine}} solution of the minimal surface equation
\begin{equation} \label{EqSupMin3}
- {\rm div} \left( \frac{\nabla u}{\sqrt{1 + \vert \nabla u \vert^2 }} \right) = 0 \qquad {\rm in} \qquad \mathbb{R}^N, \qquad N\ge2
\end{equation}
such that \begin{equation}\label{zero-bordo}
u(0) =0.
\end{equation}
Suppose also that there exists an integer $k \in \{1,...,N\}$ such that, for every $ \alpha =1,...,k,$ the partial derivative $\frac{\partial u}{\partial x_{\alpha}}$ is bounded from below on $\mathbb{R}^N$.
\noindent Then, any blow-down $C \subset \mathbb{R}^{N+1}$
associated to the subgraph of $u$ is a minimal cylinder of the form
\begin{equation}\label{conf-cilindro}
C = P \times \mathbb{R} = (\mathbb{R}^{k} \times P') \times \mathbb{R},
\end{equation}
where $P^{'} \subset \mathbb{R}^{N-k}$ is a minimal cone singular at the origin.
\end{theorem}
\section{Proofs}
Let us first consider Theorem 1.2.
{\begin{proof} [Proof of Theorem \ref{Splitting}]
Let $U$ be the subgraph of $u$ and $U_j$ the one of the functions $u_j (x) = \frac{u(jx)}{j}$, where $x \in \mathbb{R}^N$ and $ j \ge 1$.
By assumption, $U$ is a non-trivial set of least perimeter in $\mathbb{R}^{N+1}$ with $0 \in \partial U$.
Therefore, a classical procedure (see e.g. Theorem 17.3 and Theorem 9.3 of \cite{G}) provides a minimal cone $C \subset \mathbb{R}^{N+1}$ (with vertex at the origin), as the limit of a {\it subsequence} of $U_j$ (still denoted by $U_j$) with respect to the $L^1_{loc}$ convergence. The minimal cone $C$ is usually called a blow-down of $U$ and we have that, for almost every $ R>0$,
\begin{equation}\label{stimaC}
\omega_N \le R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{C} \vert.
\end{equation}
where $\omega_N$ denotes de volume of the unit ball of $ \mathbb{R}^N$. Indeed, by the minimality of the sets $U_j$, the monotonicity of the functions $ R \to R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{U_j} \vert $ and the fact that $ 0 \in \partial U_j$, the following well-known density estimates hold (see for instance formula (5.16) on p. 72 of \cite{G} or formula (1.13) on p.2 of \cite{GMM})
\begin{equation}\label{stima1}
\omega_N \le (Rj)^{1 -(N+1)} \int_{B(0,Rj)} \vert D {\bf 1}_U \vert =R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{U_j} \vert.
\end{equation}
Therefore
\begin{equation}\label{stima2}
\omega_N \le R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{U_j} \vert \to R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{C} \vert,
\end{equation}
proving \eqref{stimaC}. From \eqref{stimaC} we get that $0 \in \partial C$ and, by Lemma 16.3 and Proposition 16.5 of \cite{G}, we also know that $C$ is itself a subgraph of a generalized solution to the minimal surface equation $v : \mathbb{R}^N \to [-\infty, +\infty]$ (also called quasi-solution to the minimal surface equation) (see
\cite{Mir}, \cite{MasMir}, \cite{G}).
Also the set
\begin{equation}\label{P}
P = \{ x \in \mathbb{R}^N\, : \, v(x) = + \infty \, \}
\end{equation}
\noindent must be non-empty since $u$ is non-affine. To see this we follow \cite{G}. If $P = \emptyset$ then, by Lemma 17.7 of \cite {G}, the family of functions $u_j$ is equibounded from above on compact sets of $ \mathbb{R}^N$. The latter and the definition of $u_j$ immediately provides that
\begin{equation}\label{equi-u}
\sup_{B(0,j)} u \le {\cal K} j
\end{equation}
for some constant ${\cal K}>0$.
On the other hand, the celebrated gradient estimate of \cite{BDeGMir} tells us that
\begin{equation}\label{stimaGrad}
\forall \, x \in \mathbb{R}^N, \,\, \forall R>0, \qquad \vert \nabla u(x) \vert \le C_1 exp \Big [ C_2 \Big (
\frac{{\sup_{B(x,R)} u - u(x)}}{R} \Big )\Big ]
\end{equation}
where $C_1$ and $C_2$ are positive constant depending only on the dimension $N$.
Now, combining \eqref{equi-u},\eqref{stimaGrad} and letting $j \to +\infty$ we obtain that $\vert \nabla u \vert \in L^{\infty}(\mathbb{R}^N)$. Thus, we can apply the result of Moser \cite{Mos}, already recalled in the introduction, to get that $u$ is an affine function. The latter is impossible since we are supposing that $u$ is not affine.
We also remark that
\begin{equation}\label{PperRdentroC}
P \times \mathbb{R} \subset C
\end{equation}
by construction, and so $ P \not\equiv \mathbb{R}^N$, since $C \not\equiv \mathbb{R}^{N+1}$.
Also, $P$ is a minimal cone in $\mathbb{R}^N$, with vertex at the origin (since $C$ is a minimal cone with vertex at the origin). Combining the two latter pieces of information we also get that the origin of $\mathbb{R}^N$ belongs to $ \partial P$ and so $0 \in \partial (P\times \mathbb{R})$.
Next we observe that $P$ is singular at the origin of $\mathbb{R}^N$. Suppose not, then $P$ would be a half-space of $\mathbb{R}^N$ and thus the minimal cone $C$ would contain the half-space $P\times \mathbb{R}$ which, in turn, would give $C \equiv P\times \mathbb{R}$ (see for instance Theorem 15.5 of \cite{G} or \cite{MirPMax}) and so
\begin{equation}\label{C-semispazio}
\forall \, R >0 \qquad \omega_N = R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{C} \vert
\end{equation}
Therefore, by \eqref {C-semispazio}, the monotonicity of $ R \to R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_U \vert $ and \eqref{stima1}-\eqref{stima2} we obtain
\begin{equation}\label{U-semispazio}
\forall \, R >0 \qquad \omega_N = R^{1 -(N+1)} \int_{B(0,R)} \vert D {\bf 1}_{U} \vert
\end{equation}
which proves that $U$ is a half-space and $ U = C$ (by the construction of $C$).
But $ U = C \equiv P\times \mathbb{R}$ implies that $ \partial U$ is a vertical hyperplane, contradicting the fact $\partial U$ is the graph of the function $u$.
Moreover, by the discussions above, we have that the minimal cone $ P \times \mathbb{R}$ satisfies $P \times \mathbb{R} \subset C$ and $0 \in \partial (P\times \mathbb{R}) \cap \partial C$. Therefore, $ P \times \mathbb{R}$
must coincide with the minimal cone $C$ (cf. \cite{Moschen} or Theorem 2.4 of \cite{GMM}). This provides the first equality in \eqref{conf-cilindro} with $P$ a minimal cone singular at the origin.
\medskip
To conclude we observe that the assumption on the partial derivatives implies the existence of a constant $ K>0$ such that for every $j \ge 1$, every $x \in \mathbb{R}^N$ and every $\alpha = 1,...,k$
\begin{equation}\label{j-monot}
u_j(x + e_{\alpha}) - u_j(x) = \frac{1}{j} \int_0^j \frac{\partial u}{\partial x_{\alpha}} (jx_1,..., jx_{\alpha}+t,...,jx_N) dt \ge -K,
\end{equation}
where $e_{1},...,e_{k}$ denote the first $k$ vectors of the standard (or natural) basis for $ \mathbb{R}^N$.
Now recall that, up to a subsequence, $(u_j)$ converges almost everywhere to $v$ on $\mathbb{R}^N$ (cf. for instance Proposition 16.5 of \cite{G}) so, by letting $j\to +\infty $ in \eqref{j-monot}, we get that for almost every $x \in P$ the point $x +e_{\alpha} $ also belongs to $P$, for every $ \alpha=1,...,k$.
Therefore we have
\begin{equation}\label{Harnack-Bombieri-Giusti}
\forall \, \alpha=1,...,k \qquad P + e_{\alpha} \subseteq P.
\end{equation}
By applying Proposition \ref{LemmaGMM} we deduce that $P$ is a cylinder in the directions $e_1,..., e_k$, for otherwise, $P$ would be a half-space of $\mathbb{R}^N$ contradicting the fact that $P$ is singular at the origin.
Since $P$ is a cylinder in the directions $e_1,..., e_k$ we must have $P =\mathbb{R}^{k} \times P',$ with $P^{'} \subset \mathbb{R}^{N-k}$ minimal (since $P$ is minimal in $\mathbb{R}^N$) and $P^{'}$ singular at the origin of $\mathbb{R}^{N-k}$ (since $P$ is singular at the origin of $\mathbb{R}^N$). This concludes the proof.
\end{proof}
Now we can prove the main result.
{\begin{proof} [Proof of Theorem \ref{Th.N-7}] The function $v(x) = u(x) - u(0) $ is a solution of \eqref{EqSupMin3} with $v(0) =0$ and with $N-7$ partial derivatives bounded on one side (not necessarily the same).
Up to change the variable $ x_j $ in $- x_j$ (if necessary), we can suppose that $v$ has $N-7$ partial derivatives which are bounded from below. Also, we can assume that those partial derivatives are taken with respect to the first $N-7$ variables. If $v$ were not linear then, by applying Theorem \ref{Splitting} with $k=N-7$, we would get
\begin{equation}\label{conf-cilindro-N-7}
C = P \times \mathbb{R} = (\mathbb{R}^{N-7} \times P^{'}) \times \mathbb{R},
\end{equation}
\bigskip
\noindent where $P^{'} \subset \mathbb{R}^7$ would be a minimal cone singular at the origin. But this is is impossible since all the (non-trivial) minimal cones in dimension $ n \le 7$ are half-spaces. This gives that $v$ is a linear function and so $u$ must be affine, as desired.
\end{proof}
\bigskip
Finally we state and proof Proposition \ref{LemmaGMM} used in the last part of the proof of Theorem \ref{Splitting}. This result is essentially contained in (the proof of) Lemma 2.3 of \cite{GMM}.
\bigskip
\begin{proposition} \label{LemmaGMM} Assume $k \ge 2$ and let ${\cal C}$ be a non-trivial minimal cone in $\mathbb{R}^k$ such that $ {\cal C} + v \subseteq {\cal C},$ for some $ v \in \mathbb{R}^k
\setminus \{ 0 \}$. Then, either $ {\cal C} + v \neq {\cal C}$ and ${\cal C} $ is a half-space or,
$ {\cal C} + v = {\cal C}$ and ${\cal C}$ is a cylinder in the direction $v$.
\end{proposition}
\begin{proof}
Same proof of Lemma 2.3 of \cite{GMM}.
\end{proof}
\vspace{0.5cm}
\noindent \textbf{Acknowledgements: } The author thanks E. Valdinoci and L. Mazet for a careful reading of a first version of this article. The author is partially supported by the ERC grant EPSILON ({\it Elliptic Pde's and Symmetry of Interfaces and Layers for Odd Nonlinearities}) and by the ERC grant COMPAT ({\it Complex Patterns for Strongly Interacting Dynamical Systems}).
\vspace{0.5cm}
|
\section{Introduction and overview}\label{intro}
Ultra cold gases in which some atoms have been highly excited to electronic Rydberg states \cite{book:gallagher} are rapidly becoming an important pillar of ultra cold atomic physics. Through strong and long range dipole interactions between atoms in a Rydberg state and interaction blockades \cite{lukin:quantuminfo,urban:twoatomblock,gaetan:twoatomblock,tong:blockade,cenap:manybody,singer:blockade}, they enable, for example, studies of strongly correlated spin systems \cite{Zeiher:ryd_dress_spinchain,Glaetzle:spinice:prx}, glassy relaxation as in soft-condensed matter \cite{lesanovsky:kinetic}, many-body localization \cite{Marcuzzi:manybodyloc} or non-local and non-linear (quantum) optics \cite{Mohapatra:giantelectroopt,sevincli:nonlocopt,peyronel:quantnonlinopt,dudin:singlephotsource}. Apart from these links to quantum many-body phenomena, also the relation of cold Rydberg systems to classical many-body physics has been explored, for example through the creation of strongly coupled ultracold plasmas \cite{Killian:science,Killian:ucolplasma:PhysRevLett.83.4776,heidelberg:plasmaexp,Bannasch:strongcoup:PhysRevLett.110.253003,pohl:plasmacrystal:PhysRevLett.92.155003}. Review articles on cold Rydberg atoms \cite{loew:rydguide:jpbreview,lim:rydbergreview}, their exposure to strong magnetic fields \cite{pohl:rydBfield:physrep}, their role in ultracold plasmas \cite{kilian:ucolplasma:review} as well as their application to quantum information \cite{Saffman:quantinfryd:review} and nonlinear quantum optics \cite{firstenberg:rydberg:NLquantoptreview} are available.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.7\columnwidth]{fig1_review_classification_v1.pdf}
\caption{Rydberg aggregates and their classification. (a) Essential ingredients and level scheme, (green balls) ground state atoms, (blue balls) Rydberg $\ket{s}$ states, (orange balls) Rydberg $\ket{p}$ states. Distances $R_{ij}$ between locations will be crucial later. We classify these aggregates as follows: (b) Static Rydberg aggregate.
(c) Flexible Rydberg aggregate with atoms free in quasi-1D, two-colored balls indicate a superposition state $\sim\ket{sp}\pm\ket{ps}$. (d) Embedded Rydberg aggregate. (e) Embedded flexible Rydberg aggregate.
\label{overview_sketch}}
\end{figure}
This review covers a complementary subfield of Rydberg physics, namely assemblies of Rydberg atoms that share collective excitations through resonant dipole-dipole interactions, e.g.~involving angular momentum $\ket{s}$ and $\ket{p}$ states. For reasons which will become obvious below, we have dubbed those assemblies {\em Rydberg aggregates}.
They can consist of individually trapped atoms \cite{nogrette:hologarrays,barredo:trimeragg}, or Rydberg excitations in a cold atom cloud, which acts as a host to the aggregate, e.g.~\cite{cenap:emergent,maxwell:polaritonstorage}. Both scenarios are depicted in \fref{overview_sketch}. In the latter case, the position of Rydberg atoms can still be controlled through the interplay of interaction blockade and cloud volume \cite{schauss:mesocryst}, background gas induced Rydberg energy shift \cite{middelkamp:rydinBEC,balewski:elecBEC} or using very tight excitation laser foci. The aggregate atoms are in close enough proximity to experience significant dipole-dipole interactions involving resonant states, if more than a single Rydberg angular momentum state is present in the system.
We can then realize two regimes: (i) Static Rydberg aggregates, with excitation transport through long-range interactions occurring fast, before atoms can be set into motion. This corresponds to the frequently cited ``frozen gas regime''. (ii) Flexible Rydberg aggregates, where one deliberately waits until dipole-dipole interactions have also set the atoms themselves into motion. Both, static as well as flexible Rydberg aggregates can be realized in isolation or embedded in a cold gas cloud, leading to the four possible scenarios sketched in \frefp{overview_sketch}{b-e}.
Together, these four types of Rydberg aggregates offer intriguing opportunities to create and study coherent energy-, angular momentum- and entanglement transport as well as, on a more formal level, the formation of Born-Oppenheimer surfaces governing the motion of the Rydberg aggregate atoms, all due to the existence of collective excited states enabled through strong long-range interactions.
Static Rydberg aggregates can be described with the help of exciton theory well known from molecular aggregates, as we will see. Flexible Rydberg aggregates
add qualitatively new features, by taking the mechanism through which stable molecules form (an intricate interplay of bound coherent nuclear and electronic motion) to coherent dynamics in the continuum.
We will review the basic mechanism for exciton formation, exciton transport and inter-atomic forces at work in Rydberg aggregates, summaries existing experimental studies and review theory proposals on their potential uses. These range from understanding quantum transport on static networks, over transport in the presence of decohering environments to directed entanglement transport.
The broader vision that appears ultimately reachable in these systems is to enable a quantum simulation platform for processes in (bio)-chemistry as presented in \fref{motivation}. A static Rydberg aggregate, as sketched in panel (a), shares the fundamental energy transport mechanism with photosynthetic light harvesting complexes \cite{grondelle:review,grondelle:book} or molecular aggregates \cite{saikin:excitonreview,kirstein:Jaggregates,kuehn_lochbrunner:excitonquantdyn}, sketched in (b). However the spatial- and temporal scales of energy transport in the Rydberg atomic and molecular settings are extremely different. Moreover, the biochemical transport process is typically severely affected by decoherence, while the process in ultracold atoms is not. However, since the introduction of controllable decoherence into the Rydberg setting is possible, the easier accessible scales of the Rydberg system should allow for detailed model studies of the interplay of transport and coherence. Similarly we can draw connections between the motion of atoms in a flexible Rydberg aggregate, shown in panel (c) and the motion of nuclei on a molecular Born-Oppenheimer surface (d). Using similar scale arguments as above, flexible Rydberg aggregates hold promise to give experimentally controlled access to the analogue of quintessential processes from nuclear dynamics in quantum chemistry.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.7\columnwidth]{fig2_review_context_v2.pdf}
\caption{Rydberg aggregates in a bigger context. (a) Static atomic aggregate versus (b) molecular aggregate. The underlying physical energy transport principles are identical, but characteristic scales differ markedly. The characteristic spatial (temporal) scales of excitation migration are shown as orange (blue) text.
(c) Flexible Rydberg aggregates of atoms moving in some constrained geometry. We sketch two one-dimensional red detuned optical traps acting as linear chain for atom movement as red shaded cylinders. Scales there can be compared with those in (d), of nuclear motion on the Born-Oppenheimer surfaces of a complex molecule.
\label{motivation}}
\end{figure}
We focus in this review article more strongly on flexible Rydberg aggregates, see \fref{overview_sketch}(c), and their links to molecular physics, see \fref{motivation}(d): Rydberg aggregates with constituents allowed to move enable the tailored creation of Born-Oppenheimer surfaces with characteristic length-scales of $\mu$m, vastly exceeding typical intra-molecular length-scale of Angstroms. This renders paradigmatic quantum chemical effects, such as passage through a conical intersection, accessible much more directly in experiment. Ideas originating from this novel quantum simulation platform may then in turn benefit chemistry.
The exploration of energy transport with tuneable environments, see \fref{motivation}(a,b), is only briefly commented upon here, more details can be found in \cite{schoenleber:immag}.
The eponym of Rydberg aggregates are molecular aggregates \cite{eisfeld:Jagg,saikin:excitonreview}, due to their related energy transport characteristics \cite{eisfeld:Jagg,cenap:motion,muelken:excitontransfer}. In the molecular context, aggregation implies self-assembly in solution with subsequent mutual binding. Rydberg aggregates, in contrast are typically not bound, except in special cases \cite{zoubi:VdWagg}. However, self-assembly of Rydberg aggregates in regular structures can take place in strongly driven dissipative Rydberg systems \cite{lesanovsky:nonequil_structures,gaertner:faciitationagg:theory,malossi:fullcountstat,Simonelli:excitationavalanche:JPB,valado:exp:kineticconstraints,schempp:countingstat,urvoy:vaporaggreg}, by configuring the excitation process to develop a strongly preferred nearest neighbor distance. The term ``Rydberg aggregates'' is hence also used in the literature to refer to these ordered constructs in a dissipative setting. In contrast, our definition implies collective excited states in form of a coherently shared excitation, with direct analogy for the static Rydberg aggregates to transport physics in molecular aggregates.
The majority of cold Rydberg atomic physics experiments to date has explored the so called ``frozen gas'' regime, where thermal motion of the atoms could be neglected for the sub-microsecond duration of electronic Rydberg dynamics of interest. In experiments motivated by quantum information applications \cite{jaksch:dipoleblockade,lukin:quantuminfo,urban:twoatomblock,gaetan:twoatomblock}, motion is then a residual noise source \cite{wilk:entangletwo,mueller:browaeys:gateoptimise}. Experiments with Rydberg aggregates were also performed in this regime \cite{barredo:trimeragg,cenap:emergent,maxwell:polaritonstorage}.
Some experiments instead created unfrozen Rydberg gases, with Rydberg atoms experiencing strong acceleration, primarily to characterize interactions and ionization, e.g.~\cite{Fioretti:longrangeforces,li_gallagher:dipdipexcit,Amthor:mechVdW,amthor:autoioniz,park:dipdipionization,Zhigang:Collisions:PRA,Faoro:VdWexplosion}. Recent experiments that permit a particularly detailed real time observation of such motion will be reviewed below \cite{thaicharoen:trajectory_imaging,thaicharoen:dipolar_imaging,celistrino_teixeira:microwavespec_motion,guenter:EITexpt}. Flexible Rydberg aggregates, in which motion of atoms is fully controlled or guided, instead of randomly seeded, can now be one of the next experimental steps.
The review is organized as follows: Sections 2-4 will deal with isolated Rydberg aggregates, providing the
theoretical background in section 2, and detailing static and flexible aggregates in sections 3 and 4, respectively. Section 5 will introduce embedded Rydberg aggregates and section 6 describes Rydberg dressed aggregates, an important
possibility to extend the parameter regime for realizing Rydberg aggregates. Section 7 provides
a landscape for the existence of static and flexible Rydberg aggregates in the plane of the two most relevant parameters that can be changed: the distance of Rydberg atoms $d$ and their principal quantum number $\nu$.
The reader may want to consult this section for orientation throughout the review to identify relevant parameter regimes. We return to our quantum simulation perspective presented in the introduction in \sref{outlook} and after concluding provide an appendix with additional technical detail.
\section{Theoretical background}\label{models}
We begin with a compact summary of the formal methods used to describe Rydberg aggregates, both static and flexible.
\subsection{Basic model}\label{basicmodel}
Rydberg aggregates are typically treated with an essential state model, in which each atom of the aggregate can be in either of two highly excited Rydberg states $\ket{s}=\ket{\nu,l=0}$ or $\ket{p}=\ket{\nu,l=1}$, with a principal quantum number $\nu$ usually in the range $\nu=20\dots 100$, see the sketch \frefp{overview_sketch}{a}. Any other pair of adjacent angular momentum values such as $\ket{p}$, $\ket{d}$ gives the same qualitative results, at the expense of less energetic separations to states not desired in the model.
Atoms in their ground-state $\ket{g}$ occur prior to excitation, or when employed as an external bath. For most of the phenomena discussed in this review, the electronic state space of an $N$-atom Rydberg aggregate can be further simplified to the single-excitation manifold spanned by $\ket{\pi_n} =\ket{sss\cdots p\cdots sss}$, where only the $n$th atom is in $\ket{p}$ and all others are in $\ket{s}$. For parameter regimes of interest here, the most important interactions are of a dipole-dipole form
%
\begin{equation}
\sub{\hat{H}}{el}(\bv{R})=\sum_{n\neq m}^N H_{nm}(R_{nm}) \ket{\pi_n}\bra{\pi_m}=\sum_{n\neq m}^N \frac{C_3}{|\bv{r}_n - \bv{r}_m|^3} \ket{\pi_n}\bra{\pi_m},
\label{Hdd}
\end{equation}
where $C_3$ is the dipole-dipole dispersion coefficient and $\bv{r}_n$ the position of atom $n$. With $\bv{R}=\{\bv{r}_1,\dots,\bv{r}_N\}$ we denote the collection of all atomic coordinates.
For simplicity we assume isotropic dipole-dipole interactions. In \sref{conical_intersections} we provide a list of references where model extensions and consequences of anisotropies are discussed.
A general electronic state for a static aggregate can be written as $\ket{\Psi(t)} = \sum_n c_n(t) \ket{\pi_n}$. When the aggregate is initialized such that the single excitation resides on a specific site $k$, with $c_n(0)=\delta_{nk}$, the Hamiltonian \bref{Hdd} will cause a coherent wave-like transport of the excitation, typically de-localizing over all atoms in the aggregate.
Without the restriction to a single excitation, one can consider each aggregate atom as a pseudo-spin $1/2$ object, and write
%
\begin{equation}
\sub{\hat{H}}{el}(\bv{R})=\sum_{n\neq m}^N \frac{C_3}{|\bv{r}_n - \bv{r}_m|^3} \hat{\sigma}_n^+ \hat{\sigma}_m^-,
\label{Hddspin}
\end{equation}
where the $\hat{\sigma}_n^\pm$ are the usual Pauli raising and lowering operators.
We discuss corrections to the simple model presented here later (\sref{interactions}).
\subsection{Exciton eigenstates and Born-Oppenheimer surfaces}\label{excitons_surfaces}
In the time-independent Schr\"odinger equation
%
\begin{equation}
\sub{\hat{H}}{el}(\bv{R})\ket{\varphi_n(\bv{R})} = U_n(\bv{R})\ket{\varphi_n(\bv{R})}
\label{eigensystem}
\end{equation}
for the electronic Hamiltonian \eref{Hddspin}, eigenstates and eigenenergies parametrically depend on all co-ordinates contained in $\bv{R}$.
Eigenstates can be written as superpositions
\begin{equation}
\ket{\varphi_n(\bv{R})}=\sum_k c_{nk}(\bv{R}) \ket{\pi_k}
\label{excitons}
\end{equation}
of the so-called \emph{diabatic} basis states $\ket{\pi_k}$, and are referred to as \emph{adiabatic} electronic basis states or (Frenkel) \emph{excitons} \cite{frenkel_exciton,frenkel:footnote}. Hence, in an exciton state the entire aggregate typically shares the p-excitation collectively, see also footnote \cite{frenkel:footnote}. As in molecular physics, we refer to the corresponding eigenenergies $U_n(\bv{R})$ as \emph{Born-Oppenheimer surfaces}. In the context of Rydberg aggregates, these concepts were first explored in \cite{cenap:motion}.
\subsection{Quantum dynamics and surface hopping}\label{dynamics}
When considering motion of the Rydberg aggregate quantum mechanically, we can expand the total state $\ket{\Psi(\bv{R},t)}=\sum_{n=1}^{N} \phi_{n}(\bv{R},t) \ket{\pi_n}$ and obtain the Schr{\"o}dinger equation ($\hbar=1$) \cite{wuester:CI}
\begin{equation}
i\frac{\partial}{\partial t} \phi_{n} (\bv{R})=\sum_{m=1}^N\left[-\frac{ \nabla^2_{\bv{r}_m} }{2 M} \phi_{n}(\bv{R}) + H_{nm}(R_{nm}) \phi_{m}(\bv{R})\right]
\label{fullSE}
\end{equation}
with mass $M$ of the atoms. Since in this equation the dipole-dipole atoms give rise to both, excitation transport between different atoms as well as forces on these atoms, we expect an intimate link between motion and transport.
For up to about 3 spatial degrees of freedom Eq.~\eref{fullSE} can be numerically solved straightforwardly \cite{wuester:CI,leonhardt:switch}, e.g., by using XMDS \cite{xmds:paper,xmds:docu}. For more
degrees of freedom the direct solution of the quantum problem typically becomes too challenging, and one has to resort to quantum-classical methods.
Before describing those, let us first consider \bref{fullSE} expressed in the adiabatic basis, such that $\ket{\Psi(\bv{R},t)}=\sum_{n=1}^{N} \tilde{\phi}_{n}(\bv{R},t) \ket{\varphi_n(\bv{R})}$. We then obtain the Schr{\"o}dinger equation in terms of Born-Oppenheimer surfaces $U_{n}(\bv R)$,
\begin{equation}
i\frac{\partial}{\partial t} \tilde{\phi}_{n} (\bv{R})=\left( \sum_{k=1}^N\left[-\frac{ \nabla^2_{\bv{r}_k} }{2 M}\right] + U_n(\bv{R}) \right)\tilde{\phi}_{n}(\bv{R}) + \sum_m D_{nm}(\bv{R}) \tilde{\phi}_{m}(\bv{R}),
\label{fullSE_adiab}
\end{equation}
where
\begin{equation}
D_{nm}(\bv{R})=-\frac{1}{2M}\left( \bra{\varphi_n(\bv{R})} \nabla^2 \ket{\varphi_m(\bv{R})}+ 2\bra{\varphi_n(\bv{R})} \nabla \ket{\varphi_m(\bv{R})}\cdot \nabla\right)
\label{nonadiab}
\end{equation}
are the non-adiabatic coupling terms.
The kinetic term within square brackets in \eref{fullSE_adiab} is responsible for
motion on the excitonic Born-Oppenheimer (BO) surface $U_n(\bv{R})$. Even in ultra-cold Rydberg atomic physics, this motion is for Rydberg atoms frequently in a regime which can be treated classically, using Newton's equations \cite{Amthor:mechVdW,amthor:modellingmech,thaicharoen:trajectory_imaging,celistrino_teixeira:microwavespec_motion}. As long as the energy separations between adjacent electronic surfaces of the atoms are large compared to kinetic energies, one typically can also ignore non-adiabatic couplings.
However, excitonic BO surfaces $U_n(\bv{R})$ derived from resonant dipole-dipole interactions frequently exhibit points of degeneracies (conical intersections, see \sref{conical_intersections}), near which the energy separation criterion cannot be fulfilled. Then, one has to take into account non-adiabatic transitions \bref{nonadiab}. They generically lead to quantum dynamics that involves coherent superpositions of multiple occupied BO surfaces $n$. A convenient class of methods from quantum chemistry, designed to tackle this problem while retaining simple classical propagation of the nuclei, is quantum-classical surface hopping \cite{drukker:surfacehopping:review}. We focus particularly on Tully's fewest switches algorithm \cite{tully:hopping,tully:hopping2,tully:hopping:veloadjust,barbatti:review_tully}, which has proven a versatile and accurate method for Rydberg aggregates \cite{wuester:cradle,moebius:cradle,moebius:bobbels,leonhardt:switch}.
In Tully's method, the motion of the Rydberg atoms is still modeled classically
\begin{equation}
M\frac{\partial^2}{\partial t^2} \bv{r}_{k}=- \nabla_{\bv{r}_{k}} U_{s(t)}(\bv{R}),
\label{newton}
\end{equation}
on a specific Born-Oppenheimer surface, which may change in time
through random jumps from one surface to another ($n\to m$), creating a path $s(t)$ in accordance with the probability for non-adiabatic transitions.
Simultaneous population of multiple surfaces is made possible by simulating the full dynamics as a stochastic distribution from different paths $s(t)$.
Along with the motion of the Rydberg aggregate atoms, we are interested in the exciton dynamics. It is described by the time-dependent electronic wave function $\ket{\Psi(t)}=\sum_n c_{n}(t) \ket{\pi_n}=\sum_n \tilde{c}_{n}(t) \ket{\varphi_n(\bv{R}(t))}$. Analogously to \bref{fullSE} and \bref{fullSE_adiab} its time-evolution can be determined in the diabatic,
\begin{equation}
i\frac{\partial}{\partial t} c_{n}(t) =\sum_{m=1}^N H_{nm}[R_{nm}(t)] c_{m}(t),
\label{Tully_electron}
\end{equation}
or adiabatic basis
\begin{equation}
i\frac{\partial}{\partial t} \tilde{c}_{n}(t) =U_n[\bv{R}(t)] \tilde{c}_{n}(t) + \sum_{m=1}^N d_{nm} \tilde{c}_{m}(t)
\label{Tully_electron_adiab}
\end{equation}
with
\begin{equation}
d_{nm}=-\frac{1}{M}\bra{\varphi_n(\bv{R}(t))} \nabla \ket{\varphi_m(\bv{R}(t))}\cdot \mathbf{v},
\label{non_adiab_coupl_tully}
\end{equation}
where $\mathbf{v}=\partial \mathbf{R}(t)/\partial t$ is the system velocity.
In practice it is typically simpler to work with \bref{Tully_electron}, while \eref{Tully_electron_adiab} shows more clearly that the probability for a stochastic transition from surface $n$ to surface $m$ must be subject to $d_{nm}$. It is chosen such that the fraction of trajectories on surface $m$ at time $t$ is given by $| \tilde{c}_{m}(t)|^2$.
Further technical details about the implementation can be found in the original work \cite{tully:hopping,tully:hopping2}, the review \cite{barbatti:review_tully} and for the context of Rydberg aggregates in \cite{leonhardt:unconstrained,leonhardt:orthogonal} or \cite{leonhardt:thesis}.
For Rydberg aggregates, solutions with Tully's algorithm have been directly compared with full quantum dynamical solutions of \bref{fullSE} \cite{wuester:cradle,leonhardt:switch}, and showed excellent agreement in the cases tested, even if conical intersections are involved \cite{leonhardt:switch}. The likely reason is that these cases do not involve trajectories with motion visiting the same spatial point twice, that is~$\bv{R}(t+\Delta t) \approx \bv{R}(t)$ does not occur. Therefore, spatial complex phases of the atomic (nuclear) wave function $\phi_{n} (\bv{R})$, which are not represented in quantum classical methods, do not play a role.
\subsection{Rydberg trapping}\label{trapping}
Some of the physics we review below, will require the controlled guidance of the motion of Rydberg atoms, and hence trapping of Rydberg atoms. However, the main tools for
trapping of ultracold ground-state atoms, optical and magnetic trapping \cite{ketterle:review,book:pethik}, essentially rely on quantitative features of the electronic ground-states, and are typically not directly applicable to the trapping of atoms in a Rydberg state.
Nonetheless the tools can be ported. Optical trapping of Rydberg atoms was demonstrated, e.g.,~in \cite{li:lightatomentangle}, even such that the trapping frequencies for ground- and Rydberg states coincided (magic trapping) \cite{zhang:magicwavelength,Topcu:magictrap:PhysRevA.88.043407}. For the Rydberg state the trapping arises through the averaged ponderomotive energy of the electron, see also \cite{Younge:pondlatt:NJP}.
A more flexible option would be the use of earth-alkali atoms, in which one outer electron can be addressed for trapping, while the other is exploited for Rydberg physics \cite{rick:Rydberglattice,gil:latticeclock}. Trapping through static magnetic and electric fields has been considered as well, where additional features may arise due to the significant extension of the Rydberg orbit on the spatial scale where fields vary. Therefore, nucleus and valence electron of the atom
have to be treated separately \cite{mayle:rydbergtrap:dressing_in_trap,mayle:rydbergtrap:gsprobe,mayle:rydtrap_lowL}.
\section{Static Rydberg aggregates}\label{static}
We call Rydberg aggregates ``static'' if the motion of constituent atoms can be neglected. This may be the case because inertia prevents significant acceleration of atoms on the time-scale of interest
or because atoms are trapped (\sref{trapping}). For static aggregates, \bref{Hdd} provides a simple model for the transport of a single energy quantum (the $p$-excitation) through long-range hopping on a set of discrete ``sites'' (the remaining $s$-atoms). The underlying process shares features with energy transport in molecular aggregates or during photosynthetic light harvesting, as first realized in \cite{eisfeld:Jagg} and developed further in \cite{muelken:excitontransfer}.
\subsection{Exciton theory}\label{transport_theory}
If atomic motion is neglected, the Hamiltonian \bref{Hdd} is time-independent and all information about the system is contained in the excitons \bref{eigensystem} and their energies.
These states are shown in \fref{chain_excitons} (top) for the example of a chain of five Rydberg atoms with equidistant spacing $d$. In this case all excitons describe a single $p$-excitation delocalized over the chain, with spatial distribution reminiscent of the eigenstates of a quantum particle in a 1D box.
Dislocations or disorder, however, lead to more localized exciton states \cite{cenap:motion,eisfeld:disorder}, shown in \fref{chain_excitons} (bottom).
\begin{figure}[htb]
\centering
\includegraphics[width= \columnwidth]{fig3_excitons.pdf}
\caption{Frenkel excitons of a regular Rydberg aggregate with $N=5$, adapted from \cite{cenap:motion}. Bars indicate the magnitude and sign of coefficients $c_{nk}=\braket{\pi_k}{\varphi_n(\bv{R})}$ in \bref{excitons}. The colors distinguish different excitons, from left to right their energy is decreasing (assuming $C_3>0$).
\label{chain_excitons}}
\end{figure}
Due to the $|r_n-r_m|^{-3}$ dependence of dipole-dipole interactions, nearest neighbor interactions $J=C_3/d^3$ are dominant in \bref{Hdd}. Neglecting longer range interactions, we can analytically solve the eigenvalue problem \bref{eigensystem} to obtain
\begin{eqnarray}
U_k = -2 J \cos{[\pi k /(N+1)]},\\
\ket{\varphi_k} =\sqrt{2/(N+1)}\sum_{n=1}^N \sin{[\pi k n /(N+1)]} \ket{\pi_n},
\label{analytical_excitons}
\end{eqnarray}
with $k=1,\cdots ,N$.
For large $N$, the energies become densely spaced and form the exciton band, with bandwidth 2J. Eigenstates as shown in \fref{chain_excitons} can be experimentally accessed using appropriately detuned microwave pulses, see \cite{moebius:cradle}. However this works straightforwardly only for the most symmetric states ($n=1$ in the figure), since the microwave will act symmetrically on all atoms and thus provides no transition matrix element towards asymmetric states such as $n=5$.
A static Rydberg aggregate initialized with an excitation localized on just one specific atom is not in one of the eigenstates \bref{analytical_excitons}. The corresponding time-evolution in a superposition of eigenstates then leads to ballistic quantum transport as shown in \fref{ballistic_transport}, using \eref{Tully_electron} with fixed atomic positions.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.8\columnwidth]{fig4_ballistic_transport_v1.pdf}
\caption{Ballistic transport on a large static Rydberg aggregate with $N=100$ Rubidium atoms at $\nu=80$, with atoms numbered by $n$. These are separated by $d=8\mu$m (along the vertical axis). The excitation probability per atom $p_n=|\braket{\pi_n}{\Psi(t)}|^2$ is shown as color shade. Transport is fast and occurs within the overall effective life-time of $2\mu$s and prior to any acceleration.
\label{ballistic_transport}}
\end{figure}
A more natural scenario for Rydberg atom placement in an ultracold gas environment are random positions, with the exclusion of close proximities due to the dipole blockade \cite{jaksch:dipoleblockade,lukin:quantuminfo}. In this case, some excitons naturally localize on the clusters with the most closely spaced atoms, similarly to \fref{chain_excitons} (bottom).
For more complex assemblies of sites (atoms) than the simple linear chain in \fref{ballistic_transport}, the coherent quantum dynamical propagation there can also be viewed as a quantum random walk \cite{muelken:slowtransp,muelken:CTRW_dendrimer,muelken:quantumwalks:review}.
\subsection{Further effects of disorder}\label{disorder}
Even a static Rydberg aggregate will not show the perfectly regular structure assumed in the simulation for \fref{ballistic_transport}. The position of the atoms will be subject to a certain degree of randomness, either because atoms are randomly placed within a gas, or because of quantum or thermal fluctuations even in case of well defined individual trap centers. Since fluctuations in the position affect the interaction strengths in \bref{Hdd}, this leads predominantly to \emph{off-diagonal disorder}.
Another possible disorder is \emph{diagonal disorder}, if excitations located on different sites have slightly different on-site energies $E_n$, captured in the addition to the Hamiltonian
%
\begin{equation}
\sub{\hat{H}}{dis}(\bv{R})=\sum_{n}^N E_n \ket{\pi_n}\bra{\pi_n}.
\label{Hdis}
\end{equation}
In Rydberg aggregates this is typically negligible, but can still be present due to inhomogeneous stray fields which may lead to different differential Stark or Zeeman shifts between $\ket{s}$ and $\ket{p}$ for different members of the aggregate.
It is known that in large systems, disorder typically leads to the localization of excitations \cite{fidder:molagg:disorders}. For the special case of, typically rather smaller, Rydberg aggregates, the consequences of filling disorder have been investigate in small ordered 2D grids \cite{yu:smallgrid:excithopping} and off-diagonal disorder in 2D and 3D lattices in \cite{Robicheaux:randompos:dipdip}.
Diagonal disorder in molecular aggregates often follows a general L{\'e}vy-distribution. The resulting frequent occurrence of large outliers in the energy of certain sites, can give rise to additional localization features \cite{bas:levy,vlaming:disorder,eisfeld:disorder,moebius:exciton_levy:arxiv}.
\subsection{Experimental realizations}\label{static_realisations}
The static Rydberg aggregates described so far have recently been experimentally realized in two configurations.
{\it Regular array: }To create a regular Rydberg array, cold ground-state atoms were first loaded into a lattice of optical dipole-traps, which can be arranged in almost arbitrary 2D geometries through the use of spatial light modulators \cite{nogrette:hologarrays}. These atoms are then promoted to $\ket{62d}$ and $\ket{63p}$ Rydberg states.
Exciton dynamics on a dimer \cite{ravets:foersterdipdip} and a small regular trimer of Rydberg atoms could be recorded precisely \cite{barredo:trimeragg}, as reprinted in \frefp{experimental_aggregates}{c}, with different panels showing the excitation population on aggregate atom $k=1,2$ and $3$. The graphs correspond to cuts of \fref{ballistic_transport} for $N=3$ at the respective site of the atoms.
The traps are disabled during excitation and exciton transport, but short time-scales ensure that Rydberg atom motion is only a small disturbance, see also \sref{interactions}. These techniques have recently been extended to larger arrays studying spin systems \cite{labuhn:rydberg:ising}.
{\it Bulk excitation: }If Rydberg excitation takes place directly in a bulk cold gas, without prior single atom trapping, the positions of the aggregate atoms will be a-priori unknown and disordered. This was the starting point in \cite{cenap:emergent,maxwell:polaritonstorage}, albeit for experiments with a focus on photon-atom interfacing. These also contained small numbers of Rydberg excitations, $N\approx 3$.
In these experiments excitation transport such as displayed in \fref{ballistic_transport} competes with microwave driving between the two coupled atomic states, as shown in \frefp{experimental_aggregates}{e}. When dipole interactions are stronger than the driving ($V_{dd}>\Omega_\mu$ in the figure), they can decohere a photon wavepacket stored in an EIT medium, thus reducing the retrieved photon count plotted. Experiments take place on sub microsecond time-scales, thus atomic motion is negligible. The article \cite{cenap:emergent} contains a theoretical analysis of these experiments in terms of emergent Rydberg lattices (aggregates).
Another experiment with disordered bulk Rydberg aggregates is covered in \rref{guenter:EITexpt}. Here the second Rydberg angular momentum state is introduced through a F{\"o}rster resonance of the type $\ket{ss'}\rightarrow\ket{pp}$. In the resulting assembly of $N\sim 10-100$ Rydberg atoms, containing both $s$ and $p$ states, excitation transport again takes place through resonant dipole-dipole interactions. In that experiment, the resulting spatial diffusion of $p$ states can be monitored and controlled through an imaging scheme exploiting the optical response of the cold gas in the background \cite{guenter:EIT}. Also in this system, competition between microwave drive and dipole exchange can lead to interesting relaxation dynamics \cite{pineiro_orioli_spinrelax}.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.8\columnwidth]{fig5_static_experimental_v1.pdf}
\caption{Recent experimental progress towards Rydberg aggregates. (a) Fluorescence image of trapped ground-state atoms in a ring arrangement, diverse 2D geometries are feasible [from \cite{nogrette:hologarrays}]. This is achieved with optical dipole trap arrays as shown in (b) for a trimer aggregate. (c) Quantum transport, analogous to \fref{ballistic_transport}, on the trimer aggregate in (b) [from \cite{barredo:trimeragg}]. Alternative creation of an emerging Rydberg aggregate (blue/violet balls) within a bulk cold gas (yellow/grey balls) [from \cite{cenap:emergent}]. (e) Interplay of excitation transport and microwave drive, see text [from \cite{maxwell:polaritonstorage}].
\label{experimental_aggregates}}
\end{figure}
{\it Excitation facilitation: }Complementary to coherent Rydberg aggregates on which we focus here, there are aggregates formed incoherently through a competition of Rydberg excited state decay and continuous excitation. On time-scales large compared to the Rydberg life-time, the decoherence introduced by spontaneous excited state decay gives rise to dynamics that
can often be described by classical rate equations \cite{cenap:manybody}. Using these, one can predict the ``aggregation'' (facilitated excitation) of assemblies of Rydberg atoms with nearest neighbor spacing $d$ \cite{lesanovsky:nonequil_structures,gaertner:faciitationagg:theory}, for excitation laser detuning $\Delta$ that causes resonance for double excitation at a preferred distance $d=[C_6/(2\Delta)]^{1/6}$, where $C_6$ is the van-der-Waals coefficient. This further manifestation of excitation anti-blockade \cite{cenap:antiblockade}, was subsequently observed in a variety of experiments \cite{malossi:fullcountstat,Simonelli:excitationavalanche:JPB,valado:exp:kineticconstraints,schempp:countingstat,urvoy:vaporaggreg}. Such systems show
parallels to the physics of glassy systems \cite{lesanovsky:kinetic}. Even these incoherently created Rydberg aggregates can form the basis for coherent excitation transport as reviewed here, if augmented at some moment with the coherent introduction of a single $p$-excitation and then monitored for short times.
\subsection{Motional decoherence}\label{decoherence}
The Rydberg aggregate experiments cited in \sref{static_realisations} do not constrain positions of the atoms, but take place on time scales before significant motion
occurs. Nonetheless residual thermal atomic motion has some effect, as discussed in \cite{barredo:trimeragg} and its supplemental material. In terms of the previous section this motion can be viewed as leading to time-dependent off-diagonal disorder.
Hence, motion may be regarded as a noise source, or if controllable e.g.~through temperature, as a tuneable asset. In the next section we consider an alternative twist, with which atomic motion can be turned from a hindrance into an interesting feature.
\section{Flexible Rydberg aggregates}\label{flexible}
Static Rydberg aggregates reviewed above are an interesting model platform for the study of excitation transport, spin models and the impact of disorder. However,
residual atomic motion represents a source of decoherence or noise. Turning this negative viewpoint around, we now summarize the rich physics arising from
fully embracing directed or controllable motion of all atoms constituting a \emph{flexible} Rydberg aggregate. We imply with the term ``flexible'' that atoms do not have a preferred (equilibrium) position along directions in which they are free to move.
The motion can, and for conceptual simplicity should, still be confined to a restricted geometry, such as provided by a one-dimensional optical trap.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.99\columnwidth]{fig6_taylored_BO_surfaces_v2.pdf}
\caption{Tailoring Rydberg aggregate Born-Oppenheimer surfaces. (a) Atomic dynamics depends strongly on principal Rydberg quantum number $\nu$, chosen distance scale $R$, external field (e.g.~magnetic fields $\mathbf{B}$) for tuning interactions and isolating electronic states as well as the design of the atomic trapping (e.g.~intensity $I$ of a red detuned laser). (b) Together these components constrain and tune atomic motion possible in a flexible Rydberg aggregate. (c) Born-Oppenheimer surfaces can thusly be designed to contain, for example, accessible conical intersections [adapted from \cite{leonhardt:switch}], for geometrical variables $p$, ${\Delta}x_{12}$, $a_2$ see (b).
\label{flexible_tailoring}}
\end{figure}
\subsection{Tailored Born-Oppenheimer surfaces}\label{taylored_surfaces}
We have discussed in \sref{excitons_surfaces}, how a compact essential states model for each of the Rydberg atoms allows simple calculations of the excitonic Born-Oppenheimer surfaces that govern the motion of the aggregates. In contrast to the situation within normal, tightly-bound molecules, the long range Rydberg-Rydberg interactions create potential landscapes with characteristic length scales of micrometers. These scales exceed the minimal ones on which additional external potentials from magnetic or optical trapping can act on the atoms. All together, the ingredients sketched in \fref{flexible_tailoring} allow the tailoring of Born-Oppenheimer surfaces governing a Rydberg aggregate to provide atomic motional dynamics with interesting and diverse properties. In the following we review the properties already studied and provide an outlook on future opportunities.
\subsection{Adiabatic excitation transport}\label{flexible_transport}
\label{cradle}
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\columnwidth]{fig7_cradle.pdf}
\caption{Adiabatic excitation transport in a one-dimensional flexible Rydberg aggregate, as discussed in \cite{wuester:cradle,moebius:cradle}. (a) Visualization of setup, atoms in $\ket{s}$ (blue) and $\ket{p}$ (yellow) are shown with their position uncertainty (shaded line) and optical trap (red). We also sketch the corresponding pseudo-spin representation. (b) Scaled total atomic density showing successive elastic collisions as in Newton's cradle. (c) Corresponding excitation probability for atoms $1$-$7$. White lines show the mean position of each atom, and the color shading and its width indicate the excitation probability $p_n(t) = |\braket{\pi_n}{\Psi(t)}|^2$. Panels (b,c) are adapted from \cite{wuester:cradle}.
\label{newtons_cradle}}
\end{figure}
Already the simplest flexible aggregate, with $N$ atoms free to move in a one-dimensional optical trap, is a promising candidate to study the interplay of excitation transport and motional dynamics expected from the discussion in \sref{models}. Consider the scenario of an equidistant chain of atoms in a one-dimensional optical trap, with a dislocation at the end, visualized in \frefp{newtons_cradle}{a}. The dislocation causes initial close proximity of two atoms. In this situation, one of the BO surfaces of \eref{eigensystem} $\sub{U}{rep}(\bv{R})$ corresponds to the single p-excitation localized on the two dislocated atoms initially. This feature is also evident in the randomly dislocated chain analyzed in \fref{chain_excitons} (bottom).
The surface $\sub{U}{rep}(\bv{R})$ then causes strong repulsion of the dislocated atoms only, leading to motional dynamics with successive repulsive pairwise collisions between all atoms in the chain \cite{wuester:cradle,moebius:cradle}, shown in \frefp{newtons_cradle}{b}. Interestingly, this classical momentum transfer between the atoms is accompanied
by a directed and coherent quantum transport of the single electronic excitation through the chain with high fidelity as shown in \frefp{newtons_cradle}{c}. Further analysis reveals that the underlying exciton state $\ket{\sub{\varphi}{rep}(\bv{R})}$ is adiabatically preserved due to the relatively slow motion $\bv{R}(t)$ of the atoms. In this scenario, the coherent excitation transport also implies entanglement transport, which could be measured through Bell-type inequalities \cite{wuester:cannon}.
An additional combination of repulsive van-der-Waals and attractive dipole-dipole interactions can modify exciton states at short distances and form aggregates of actually up to 6 atoms \emph{bound} in one dimension, enabling studies of their dissociation \cite{zoubi:VdWagg}. With additional Stark tuning of Rydberg levels, such a binding mechanism can even hold in 3D for dimers and trimers\cite{kiffner:dipdipdimer,kiffner:fewbodybound}.
\subsection{Conical intersections}\label{conical_intersections}
In the preceding section we have seen that Rydberg atomic motion on a single BO surface can already lead to intriguing transport physics far from being trivial. Additional and qualitatively different features come into play if multiple surfaces are involved, for example due to the presence of conical intersections (CIs) \cite{yarkony1996diabolical,yarkony2001conical}. They generically occur in dipole-dipole interacting aggregates described by \bref{Hdd} whenever the motion - which still may be geometrically restricted - takes place in two or more spatial dimensions \cite{wuester:CI}. The simplest example is a ring timer. It possesses one well localized conical intersection (seam) of surfaces, as drawn in \fref{conical_intersection}. As a consequence, we expect a Rydberg atom wavepacket, which traverses the CI, to undergo non-adiabatic transitions. This also should lead to optically resolvable manifestations of Berry's phase \cite{wuester:CI}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\columnwidth]{fig8_review_CI_v2.pdf}
\caption{(a) Geometry of Rydberg Ring trimer. Atomic positions are parametrised in terms of the angles $\theta_{12}$, $\theta_{23}$. (b) The three BO surfaces following from \eref{eigensystem}, as a function of the two relative angles indicated in (a). The conical intersection is marked between the upper two. Figure adapted from \cite{wuester:CI}.
\label{conical_intersection}}
\end{figure}
We can study the impact of CIs on excitation transport, as discussed in \sref{cradle}, in a scenario with two intersecting linear chains. Then, the CI can be functionalized as a switch, where minute changes in aggregate geometry have profound impact on the exciton transport direction and coherence properties \cite{leonhardt:switch}.
Finally non-adiabatic features can also be accessed entirely without atomic confinement, by choosing suitable initial Rydberg excitation locations in a 3D cold gas volume, with Rydberg atoms subsequently moving without constraints \cite{leonhardt:unconstrained}. These results, presented in \fref{embedded_intersection}, move an experimental demonstration of non-adiabatic effects due to passage through a conical intersection in Rydberg aggregates firmly within reach of present day technology \cite{balewski:elecBEC,guenter:EITexpt,thaicharoen:trajectory_imaging,celistrino_teixeira:microwavespec_motion}.
For free motion in 3D, a more sophisticated model for dipole-dipole interactions than \bref{Hdd} is needed in general, which takes into account azimuthal quantum numbers $m_l$ of the electronic states and the anisotropy of dipole-dipole interactions.
This was used in \cite{leonhardt:unconstrained}. There are no qualitative changes in the type of features observed, compared to the simple isotropic model described above. In fact, by isolating specific $m_l$-states with Zeeman shifts, one can tune experiments such that the isotropic model directly applies \cite{leonhardt:orthogonal}.
Other interesting viewpoints on conical intersections within dipole-dipole energy surfaces of Rydberg dimers are highlighted in \cite{kiffner:spinorbitcoupl,kiffner:nonabelian}, for dimers bound as in \cite{kiffner:dipdipdimer,kiffner:fewbodybound}: The non-adiabatic coupling terms in \bref{nonadiab} can be assembled into a vector potential, see e.g.~\cite{Zygelman:nonabelian,domke:yarkony:koeppel:CIs}, to which the motion in \bref{fullSE_adiab} is then coupled. The system thus furnishes a platform for synthetic spin-orbit coupling \cite{kiffner:spinorbitcoupl} and non-Abelian gauge fields \cite{kiffner:nonabelian}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.99\columnwidth]{fig9_agg3d_signatures.pdf}
\caption{Manifestation of non-adiabatic motional dynamics for atoms moving freely in three dimensions. (a) Sketch of initial geometry, with atoms one and two in close proximity on the $x$-axis, and three and four somewhat wider spaced on the $y$ axis. Atoms one and two will initially repel strongly along the green arrows.
(b,c) Ensuing two-dimensional column densities of atoms at late times, showing several lobes due to multiple populated BO surfaces. White crosses mark the initial positions of atoms $1-4$ as shown in panel (a). (d) Time-evolution of potential energy distribution. The splitting on two BO surfaces can also clearly be seen here. Figures taken from \cite{leonhardt:unconstrained}.
\label{embedded_intersection}}
\end{figure}
\subsection{Experimental realizations of unfrozen Rydberg gases}\label{flexible_realisations}
Following our summary of theory proposals for the applications of flexible Rydberg aggregates, we now present a brief overview of the current capabilities of experiments in that direction.
Crucial to the implementation is a sufficiently controlled positioning and equally importantly, high resolution and time resolved read-out of atom locations.
Many recent Rydberg atomic experiments are geared towards utilizing the dipole-blockade for quantum computation, quantum optics or simulation of condensed matter systems.
For these objectives motion represents a noise source (see \sref{decoherence}) and hence is suppressed by design.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.9\columnwidth]{fig10_flexible_experimental_v1.pdf}
\caption{Experimental techniques for the observation of flexible Rydberg aggregates. (a) Diagram for spatially resolved field ionization. (b) Rydberg pair correlations observed in this manner as a function of time. Experiment (top) and simulations (bottom) both show the movement of the correlation peak to larger separations due to dipole dipole repulsion. (a,b) taken from
\cite{thaicharoen:trajectory_imaging}. (c) Microwave spectroscopy of Rydberg atoms in motion. The interaction induced broadening of a resonance signal that is present initially (bottom) reduces as atoms repel and interactions diminish (towards top). Reproduced from \cite{celistrino_teixeira:microwavespec_motion}.
\label{flexible_experimental}}
\end{figure}
Early experiments had a stronger fundamental atomic physics focus, studying atomic interactions, the resulting collisions and ionization \cite{Fioretti:longrangeforces,park:dipdipionization,li_gallagher:dipdipexcit} or the autoionization of Rydberg atoms when these are set into motion due attractive van-der-Waals interactions \cite{amthor:autoioniz,Amthor:mechVdW,amthor:modellingmech,Faoro:VdWexplosion}.
Recent experiments demonstrated the continuous observation of Rydberg atoms in cold gases while these undergo motional dynamics due to van-der-Waals interactions. The first technique used to this end, is time- and space resolved atomic field ionization \cite{thaicharoen:trajectory_imaging,thaicharoen:dipolar_imaging}. Spatial resolution is provided by strong electric field in-homogeneities near a tip, which enable the reconstruction of the point of origin of ion-counts on a detector, as sketched in \frefp{flexible_experimental}{a}. The correlation signals observed, see \frefp{flexible_experimental}{b}, show the effect of vdW repulsion, and could similarly give insight into motional aggregate dynamics such as that shown in \fref{newtons_cradle} or \fref{embedded_intersection}.
The second technique relies on microwave spectroscopy \cite{celistrino_teixeira:microwavespec_motion}, monitoring the time-dependence of two-photon microwave transfer rates from one Rydberg state to another. The signal is significantly broadened due to vdW interactions of the Rydberg atoms initially, but narrows considerably as the atom cluster undergoes vdW repulsion and potential energies are reduced, see \frefp{flexible_experimental}{c}. Similar techniques can measure the time-dependent BO surface energy \eref{eigensystem} of a flexible Rydberg aggregate, as shown in \frefp{embedded_intersection}{d}.
\section{Embedded Rydberg aggregates}\label{embedded}
The most common starting point for experiments is to excite few cold atoms within a host cloud to Rydberg states, typically with angular momenta $s$ or $d$ such that they can be reached via a two-photon excitation. Then, the dipole-blockade guarantees a Rydberg assembly with a minimal nearest neighbor distance equal to the blockade radius, roughly $r_{bl}=(C_6/\Omega)^{1/6}$, where $\Omega$ is the Rabi-frequency of excitation. Augmenting the assembly with one or few atoms in an adjacent angular momentum state (e.g.~$p$) creates a Rydberg aggregate.
The augmentation can be achieved via microwave coupling \cite{maxwell:polaritonstorage,pineiro_orioli_spinrelax}, or direct laser excitation \cite{weber:superatoms}.
One might expect that the the embedding cold background gas will strongly modify the
dynamics of flexible or static Rydberg aggregates created in such a manner.
Interestingly, there should be parameter regimes where this is not the case \cite{leonhardt:unconstrained,moebius:bobbels}, which is experimentally corroborated in \cite{celistrino_teixeira:microwavespec_motion,thaicharoen:trajectory_imaging}, where observations on Rydberg atoms moving through a gas could be explained with classical trajectory simulations containing only the Rydberg excited atoms, without direct or indirect influence of the hosting cold gas cloud.
However, for other parameter regimes strong life-time reductions of high Rydberg excitations \cite{balewski:elecBEC,schlagmueller:ucoldchemreact:prx} or their associative ionization have been reported \cite{niederpruem:giantion}. Additional processes that arise in the system are the formation of ultra-long range molecules, see e.g.~\cite{greene:ultralongrangemol,Bendkowsky:boundquantumreflection,Bendkowsky:ultralongmol,Li:homonuc}, polarons \cite{Schmidt:rydpolaron}, and Rydberg molecule induced spin flips \cite{niederpruem:remotespinflip}. If by experimental design these effects can be made sufficiently small, aggregate physics in a cold gas environment proceeds roughly as discussed so far. Individual atoms in a Rydberg assembly can even be replaced by super-atoms (groups of ground-state atoms coherently sharing a Rydberg excitation), without changing the results described \cite{moebius:bobbels}.
If a Rydberg aggregate can be embedded in an ultracold gas environment without detriment, the latter can actually be used to some advantage. It can serve as an atomic reservoir for repeated excitation of Rydberg atoms \cite{wuester:cannon} but also as a gain medium for sensitive position- and state detection of Rydberg excitations \cite{olmos:amplification,guenter:EIT,guenter:EITexpt,schoenleber:immag}. These schemes add to the toolkit that promises sufficient temporal and spatial resolution for the experimental readout of the processes summarized earlier. Even more interestingly, they provide a control knob over decoherence, which is created by the measurement or readout within the system. It can give rise to decoherence between electronic states forming a Born-Oppenheimer surface, causing a complete cessation of intra-molecular forces \cite{wuester:immcrad}. It also decoheres excitation hopping such as shown in \fref{ballistic_transport}, enabling quantum simulations of light harvesting \cite{schoenleber:immag}, switchable Non-Markovianity \cite{genkin:markovswitch}, tuneable open quantum spin systems \cite{schempp:spintransport} and engineered thermal reservoirs \cite{schoenleber:thermal}.
\section{Rydberg dressed aggregates}\label{dressed}
Most phenomena reviewed so far can be ported to different parameter regimes, if the interactions discussed in \sref{basicmodel} are replaced by \emph{dressed dipole-dipole interactions} \cite{wuester:dressing}. A direct mapping of resonant dipole-dipole interactions requires two separate ground states $\ket{g}$ and $\ket{h}$, that are far off resonantly coupled to the two Rydberg states $\ket{s}$ and $\ket{p}$ employed here earlier, as sketched in \fref{dressed_leveldiag}. This can effectively give rise to a state-flip interaction Hamiltonian $\hat{H}\sim (\ket{gh}\bra{hg} + \mbox{c.c})$ in the space of the ground-states \cite{wuester:dressing}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.6\columnwidth]{fig11_dressing_simplified_v1.pdf}
\caption{Level scheme of dressed dipole-dipole interactions. Separate couplings with identical effective Rabi frequency $\Omega$ enable transitions $\ket{g}\leftrightarrow\ket{s}$ and $\ket{h}\leftrightarrow\ket{p}$, respectively. Both are far detuned $\Omega \ll \Delta$. Thus, while atoms predominantly remain in the ground states $\ket{g}$, $\ket{h}$, the small admixture of Rydberg levels allows an effective state-change interaction $\hat{H}\sim \ket{gh}\bra{hg}$.
\label{dressed_leveldiag}}
\end{figure}
While making an experiment more challenging, the use of dressed interactions gives additional flexibility for the choice of parameters. For example the ring-trimer CI shown in \fref{conical_intersection} can be realized with dressed interactions \cite{wuester:CI} to avoid the need for extreme trap strengths. Dressing can also match the energy/time-scales of Rydberg physics with those of ground-state cold atom- or Bose-Einstein condensate physics.
Moreover, dressed interactions can enable entirely novel effects not present in systems with bare interactions. They can give rise to the dipole-dipole induced repulsion of droplets containing many atoms \cite{genkin:dressedbobbles}, even allowing them to be brought into quantum superpositions of different droplet locations \cite{moebius:cat}. The droplets are then described by mesoscopically entangled states with tens of atoms. Dressed dipole-dipole interactions could also be exploited to design quantum simulators of electron-phonon interactions \cite{hague:su:schrieffer:heeger,hague:rydquantsim,hague:supercondquantsim}.
\section{Rydberg aggregate existence domains}\label{interactions}
\begin{figure}[htb]
\centering
\includegraphics[width= 0.99\columnwidth]{fig12_effective_vs_full_spectra_REVIEW_Li_nu40_v2.pdf}
\caption{Rydberg interaction potentials for Lithium in the vicinity of the $\ket{40s0}\ket{40s0}$ asymptote (which is taken as the zero of energy), red curves are magnified in (b,c). (b) When zooming onto this potential at separations $R$ larger than $\sim0.8\mu$m, we recover the $U(R)=C_6/R^{6}$ dependence characteristic of vdW interactions (fit by the blue dashed line). (c) Resonant dipole-dipole interactions form potentials with asymptotic form $U_\pm(R)=\pm C_3/R^{3}$ (blue dashed) and states $\ket{\varphi_\pm}=(\ket{sp} \pm\ket{ps})/\sqrt{2}$.
\label{spaghetti_onset}}
\end{figure}
Approaching the end of this review, we now provide a brief guide to identify those regions in the vast parameter space for Rydberg physics, where either static- or flexible Rydberg aggregates as reviewed above can be practically realized. To this end, let us have a closer look at the origin and limitations of the simple model \bref{Hdd}, by considering the full picture of Rydberg-Rydberg atomic interactions. Most Rydberg potentials can be obtained by exact diagonalization of the Hamiltonian in a restricted Hilbert space with principal quantum numbers $\nu\in[\sub{\nu}{min}\dots \sub{\nu}{max}]$, angular momenta $l\in[0 \dots \sub{l}{max}]$ and the corresponding ranges of total angular momenta $j$ and their azimuthal quantum numbers $m_j$ \cite{fabian:soliton,walker:zeemandeg:pra,weber:rydint:tutorial}. The inter-atomic distance $R$ is varied and interactions can be added through the dipole-dipole interaction Hamiltonian \cite{book:gallagher,noordam:interactions,park:dipdipbroadening}.
Recently, open source packages for this sort of calculation have become available \cite{weber:rydint:tutorial,pairinteraction:webpage,arc:package}.
Alternatively, interactions can also be calculated using perturbation theory \cite{singer:VdWcoefficients}. The details listed are atomic species dependent; we will focus on Lithium atoms, because they are light and therefore of advantage to see motional effects, and on Rubidium atoms, because they represent the most common species in cold atom experiments.
An exemplary set of molecular potentials is shown in \frefp{spaghetti_onset}{a}, in the vicinity of the $\ket{40s0}\ket{40s0}$ asymptote of Lithium. In this particular example, panel (c) shows that for inter-atomic separations $R>2\mu$m, and when a p-excitation is present in the systems, interactions are clearly dominated by resonant dipole-dipole interactions, involving just a pair of sp states. We can thus employ the model \bref{Hdd}. Since this model does not contain any interactions between two atoms in an $s$-state, we can augment it by a van-der-Waals term
%
\begin{equation}
\sub{\hat{H}}{vdW}=\sum_n E_n(\bv{R})\ket{\pi_n}\bra{\pi_n},
\\
E_k(\bv{R})\approx \sum_{\ell \ne n } \frac{C^{(sp)}_6}{R_{n\ell}^6}+\frac{1}{2}\sum_{j\ne n}\sum_{\ell \ne n,j } \frac{C^{(ss)}_6}{R_{j\ell}^6},
\label{HVdW}
\end{equation}
where $R_{n\ell}=|R_n-R_{\ell}|$ and $C^{(ab)}_6$ denotes the coefficient for the vdW interaction between one atom in the state $a$ and the other one in $b$.
This extension also describes the short range repulsion (for $C_6>0$) shown in panel (b) down to $R>1\mu$m. At even shorter distance $R<\sub{R}{mix}$, we enter the highly irregular and dense region of the molecular spectrum shown in (a). Here simple models such as \bref{Hdd} and \bref{HVdW} break down entirely. Theory predictions presented here then no longer hold, and we thus anticipate that this region is best avoided. It might give rise to interesting experimental observations, but the most likely result of close encounters of Rydberg atoms is their ionization.
A rough estimate of the distances at which the spectra become dense is provided by he formula
%
\begin{equation}
\sub{R}{mix}(\nu)=2 \left( \frac{C_3(\nu)}{{\Delta}E_{pd}(\nu)} \right)^{\frac{1}{3}},
\label{Rmix}
\end{equation}
where ${\Delta}E_{pd}(\nu)$ is the energy gap between the angular momentum $p$ and $d$ states at principal quantum number $\nu$, see \aref{parameter_criteria} for further discussion. We empirically find the formula to work well for both species considered here, Li and Rb, based on spectra such as shown in \fref{spaghetti_onset}.
We are now in a position to determine for which parameters Rydberg aggregates described by the models summarized in this review can be formed, applying the following criteria:
\begin{enumerate}
\item The effective state model should be valid, implying $d>\sub{R}{mix}$, where $d$ denotes any separation between Rydberg excited atoms in the system.
\item The dynamics of interest, static excitation transport or atomic motion, should take place within the life-time $\sub{\tau}{eff} =\tau_0/N$ of the system, where $\tau_0$ is the life-time of a single Rydberg state, and $N$ the number of Rydberg atoms in the aggregate. These effective life-times become dependent on temperature $T$ by taking into account black-body redistribution of states \cite{beterov:BBR}.
\end{enumerate}
Visualizing these conditions, we show in \fref{parameter_ranges_T300} the domain in the parameter plane $(\nu,d)$ where static- or flexible aggregates are viable at room temperature.
We focus on the principal quantum number $\nu$ and the dominant inter-atomic spacing $d$, since these two parameters have most impact on the physics.
We consider aggregates with $\sub{N}{agg}$ atoms of the dominant experimental atomic species, $^{87}$Rb, or the lightest Alkali species with quantum defect $^{7}$Li. For simplicity we assumed a linear chain with dominant spacing $d$, and target dynamics such as in \fref{ballistic_transport} (static) or \fref{newtons_cradle} (flexible). The white region at the bottom is excluded due to $d<\sub{R}{mix}$. For the white region at the top, the overall effective life-time of the aggregate including black-body radiation (BBR) \cite{beterov:BBR} is too short for significant excitation transport (defined in more detail in \aref{parameter_criteria}) to take place.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.8\columnwidth]{fig13_regimes_v2.pdf}
\caption{Parameter domains of static (green and red) versus flexible (blue) Rydberg aggregates. We take into account the reduction of effective life-times by black-body radiation at room temperature $T=300$K. Red shade indicates where static aggregates exist, however with atoms that would visibly accelerate during excitation transport. We consider the atomic species Li (a) and Rb (b), but importantly assume different aggregate sizes $\sub{N}{agg}$ for either as indicated. White areas are excluded, either due to to short aggregate life-times (top) or too close proximities to avoid Rydberg state mixing (bottom). See the text and \aref{parameter_criteria} for the precise criteria used.
\label{parameter_ranges_T300}}
\end{figure}
Within these two boundaries we additionally distinguish: (i) The domain of flexible Rydberg aggregates (blue), defined in more detail in \aref{parameter_criteria}. (ii) The domain of purely static Rydberg aggregates (green). Here, for the duration of anticipated excitation transport, atoms would be trapped by their inertia and not significantly set into motion. (iii) Aggregates perturbed by atom acceleration (red). Here, atoms would significantly accelerate during the duration of excitation transport. We see that Lithium atoms can form larger flexible aggregates and be set into motion faster, all due to the much smaller atomic mass.
The domain for flexible Rydberg aggregates is strongly affected by the reduction of effective life-times due to BBR \cite{beterov:BBR}, hence we also show the corresponding domains when suppressing BBR at $T=0$K in \fref{parameter_ranges_T0}.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.8\columnwidth]{fig14_regimes_T0K_v2.pdf}
\caption{The same as \fref{parameter_ranges_T300} but assuming an extremely cold environment at $T=0$K.
\label{parameter_ranges_T0}}
\end{figure}
\section{Perspectives and outlook}\label{outlook}
We have seen that (flexible) Rydberg aggregates present a well defined, controllable and observable system in which to study the interplay of quantum coherent atomic motion and excitation.
Besides the fundamental interest in such a system, we see the following perspectives for their utilization:
\ssection{Quantum simulations of nuclear dynamics in molecules} We have seen in \sref{conical_intersections} that excitonic Born-Oppenheimer surfaces exhibit conical intersections, one of the most fascinating features of molecular dynamics. In contrast to the intra-molecular situation, motional dynamics of Rydberg aggregates ought to be observable on scales relevant for the CI. It thus seems possible to furnish a quantum simulation platform that can either: (i) replicate fundamental dynamical effects, such as vibrational relaxation downwards a CI funnel \cite{Perun-Domcke-Abinitiostudies-2005}, and gain new insight through a more direct connection between theory and observation; or (ii) attempt to mimick as closely as possible computationally known BO surfaces for a certain molecule and thus function as an analogue quantum computer to tackle the formidable computational challenge of multi-dimensional nuclear dynamics. Such a platform would then complement others, enabling the quantum simulation of \emph{electronic} dynamics in molecules with cold atoms \cite{Luehman_molorbquantsim_prx}.
\ssection{Quantum simulations of energy transport in biological systems} While this review has focused more strongly on the description of flexible Rydberg aggregates, we have listed several works that enable the controlled introduction of decoherence to static Rydberg aggregates in \sref{embedded}. Since decoherence, particularly through internal molecular vibrations, is a dominating feature of energy transport through dipole-dipole interactions on a molecular aggregate (such as a light-harvesting complex), this paves the way towards the utilization of these systems for photosynthetic energy transport simulations.
\section{Summary and conclusions}\label{summary}
We have reviewed Rydberg aggregates, interacting assemblies of few Rydberg atoms that can exhibit excitation sharing and transport through mutual resonant dipole-dipole interactions.
Our overview distinguished static aggregates, where atomic motion can be neglected during the times-scale of an experiment, and flexible aggregates, where atomic motion is by design an essential feature of the system. All results where placed into a broader context, considering the exploitation of these aggregates for quantum simulation applications, mainly by making contact with quantum chemistry or biology. In this direction we have summarized a variety of theory proposals. For both, static- and flexible aggregates, we additionally gave an overview of experimental techniques that are closest to the requirements for the implementation of those proposals.
It appears that the full theoretical potential of Rydberg aggregates in either variety discussed here is only just now becoming clear. Since experiments are now capable to provide all tools required for their controlled creation and observation, we expect a versatile and interdisciplinary quantum simulation platform in the near future.
\ack
We would like to thank C.~Ates, A.~Eisfeld, M. Genkin, K. Leonhardt, S.~M\"obius and H.~Zoubi for their part
in our joint publications featured in this review. Additionally, we thank Karsten Leonhardt for help with figures as well as a critical reading of the manuscript, and the authors of articles from which we reprint figures for their kind permission to do so.
|
\section{Introduction}
JavaScript is one of the fundamental technologies underpinning the world wide
web today. From its humble beginnings as a scripting language to support basic
interactive content, it has matured to the point where it powers large applications for
multi-billion dollar businesses. In addition to client-side JavaScript run in
the user's browser, server-side JavaScript is becoming increasingly popular for
high-throughput, low-latency web applications.
The JavaScript language (actually ECMAScript) is one of the most popular languages
today\footnote{\url{https://github.com/blog/2047-language-trends-on-github}}.
The increasing complexity of JavaScript applications and the deployment in
environments with high performance requirements has driven the development of
JavaScript compilers that produce more efficient, highly-optimised code. These
compilers are complex pieces of software themselves and make a plethora of
configurable parameters available to the user -- one size does not fit all, and
how exactly a piece of code should be optimised may depend on the particular
application and execution environment.
While many JavaScript code optimisers exist, these usually focus
on ``compressing'' the source code so that it can be transferred from the
server to the browser more efficiently by means of source-level transformations.
Such optimisations do not affect the semantics of the code nor how it is compiled, and they usually do not improve performance
in terms of running time.
There are many benefits to optimising not only the code but also the JavaScript
compiler for particular applications. On mobile devices, power consumption
is a major issue, and optimised code can help reduce it. On a server, reducing
the running time of a software component means that more transactions can be supported on
the same hardware. On a desktop machine, lag in interactions can be reduced and user
interfaces be made more responsive.
In the vast majority of applications, the JavaScript
compiler is run in its default configuration, which has been chosen by its developers to achieve robust performance
across a broad range of use cases. While these default settings will provide reasonable
performance in most situations, we demonstrate that often, substantial gains can be realised
easily, by searching the configuration space for compiler parameter settings that better optimise the
JavaScript code produced for a particular application. With little effort, applications can
be made more efficient and consume less resources. This effect even holds for the large,
heterogeneous benchmark sets used during development of the JavaScript engines themselves, upon
which the default settings are purportedly based.
Compared to traditional applications of automated algorithm configuration,
JavaScript code often runs for only relatively short periods of time, but very frequently.
The same code can be run millions of times a day, each time a website is loaded or a request is made to a
server. Even small improvements translate to massive aggregate savings in
resources.
We apply state-of-the-art machine learning techniques for automatic parameter
configuration to the two main JavaScript compilers. On a range of popular and
representative benchmarks, we show that performance can be improved by more than
35\% even with relatively modest configuration effort, without any modification to the JavaScript source code
under consideration, or of the JavaScript engine running it, other than the change in parameter configuration.
\section{Background}
The idea of optimising the configuration of a compiler for a particular
application or set of applications is not new. The Milepost GCC
project~\cite{fursin_milepost_2011} is perhaps the most prominent example and
uses machine learning to dynamically determine the best level of optimisation.
In an iterative process, it can improve execution time, code size, compilation
time and other metrics. The approach has been integrated into the widely-used
GCC compiler. Other approaches that optimise the code generation for C programs
include \cite{haneda_automatic_2005,pan_fast_2006,plotnikov_automatic_2013}.
While most of these optimise the GCC compiler, there exists some work
on LLVM as well~\cite{fursin_collective_2014}.
Another focus of research for automatic dynamic optimisation of compiled code
has been the Jikes Java compiler~\cite{alpern_jikes_2005}.
\citet{hoste_automated_2010}~use multi-objective evolutionary search to identify
configurations that are Pareto-optimal in terms of compilation time and code
quality. \citet{cavazos_method-specific_2006}~learn logistic regression models
that predict the best optimisation to apply to a method.
\citet{kulkarni_mitigating_2012}~use artificial neural networks to determine the
order in which a set of optimisations should be applied during compilation.
A major concern with all compiler configuration optimisation approaches is the
computational effort required to determine a good or optimal configuration. If this is too
large, any benefits gained through the optimisation may be negated. One approach
to reducing the initial overhead is to move the configuration process online and
to learn to identify good configurations over successive compilations, but other
approaches have been explored in the literature
(see, e.g.~\cite{thomson_reducing_2010,ansel_siblingrivalry_2012,tartara_continuous_2013}).
Compilers that translate JavaScript to native code are relatively new compared
to compilers for more established languages likes C. While they are also highly
optimised and, in the case of JavaScriptCore through the use of the LLVM
framework, leverage at least some of the benefits decades of optimisation effort
has brought to compilers for other languages, we believe that performance
improvements over the default configuration are to be gained more easily here.
Furthermore, due to the widespread use of JavaScript in applications with hundreds of millions
of end users (such as web browsers), any performance improvements are likely to be impactful.
\subsection{JavaScript Optimisation}
Existing JavaScript optimisers, such as Google's Closure
Tools\footnote{\url{https://developers.google.com/closure/}} and Yahoo's YUI
compressor\footnote{\url{https://yui.github.io/yuicompressor/}}, focus on
source code transformations that do not alter the syntax or semantics of the
code, but compress the representation by shortening identifiers, removing
white space or inlining code. The aim of these optimisations is to reduce the
size of the code that has to be transferred from the server to the user's
browser, thereby reducing the load time of the page. It focuses on efficiency
\emph{before} the code is run, but does nothing to improve performance
\emph{while} the code is running.
Indeed, many of those tools and techniques are not specific to JavaScript, but
are also applied to other resources that are transferred to the client when a
web page is loaded, such as Cascading Stylesheets (CSS). In contrast, what we
propose here leverages the specific configuration options of JavaScript engines
to optimise the actual runtime behaviour and efficiency of the code.
One attractive aspect of our approach is that it naturally complements any
extensions implemented to an existing JavaScript engine (by performing our automated configuration procedure again),
and is able to search for improving engine configurations while consuming commodity
compute cycles, without significant impact on development and engineering effort.
Running an automated configuration procedure on a commodity compute cluster
for a week is significantly cheaper than the salary of even a single engineer for the same period,
and optimising the engine configuration automatically frees up human development resources, which can then be used
to further enhance the JavaScript engine with new or improved optimisation mechanisms.
\todo{CF}{I feel as if there is a small amount more to say here regarding the suitability of JavaScript engine parameters
to automated algorithm configuration. Specifically, that we are in essence solving extremely similar (or truly identical!) problem
instances potentially hundreds of millions of times. This would be extremely strange in, say, the SAT setting. We do mention this a bit
in the introduction, but I think it should be stressed and this might be a good
spot.}
\todo{LK}{I don't think that this makes JS inherently more suitable -- we can't
expect the massive gains seen in other domains. For real-world stuff, this is
probably even worse as a large part will be IO.}
\subsection{Automated Algorithm Configuration}
Most software has switches, flags and options through which the user can
control how it operates. As the software becomes more complex or is used
to solve more challenging and diverse problems, the number of these options also
tends to increase. While some of these parameters control the input/output behaviour
of a given piece of software or algorithm, others merely affect efficiency in terms of
resource use.
The algorithm configuration problem is concerned with finding the best parameter
configuration for a given algorithm on a set of inputs, where the
definition of ``best'' can vary, depending on the given application scenario. In many practical cases, the
goal is to achieve better performance, and this is how we use algorithm configuration here -- we want to achieve the same
functionality, but with reduced resource requirements. Specifically, in this work
we focus on minimizing the CPU time required, but in principle, any scalar measure
of performance can be used.
Finding the best parameter configuration for a given algorithm is a long-standing
problem. Humans tend to be bad at solving it -- evaluating parameter configurations
requires substantial effort, and interactions between parameters may be complex and unintuitive.
\citet{minton_automatically_1996} notes that,
\begin{quote}
``Unlike our human subjects, [the system] experimented with a wide variety of
combinations of heuristics. Our human subjects rarely had the inclination or
patience to try many alternatives, and on at least one occasion incorrectly
evaluated alternatives that they did try.''
\end{quote}
Fortunately, there exist many automated procedures for algorithm configuration.
Perhaps the simplest approach is to try all combinations of parameter values.
This approach is known as a full factorial design in the statistics literature on experimental design
and as grid search in computer science (specifically, in machine learning); its main
disadvantage lies in its high cost -- the number of configurations to be evaluated grows
exponentially with the number of parameters and their values. For most practical
applications, including the ones we consider in the following, complete grid search is infeasible.
A commonly used alternative is simple random sampling:
Instead of evaluating every combination of parameter values, we randomly sample
a small subset. This is much cheaper in practice and achieves surprisingly good
results~\cite{bergstra_random_2012}.
Indeed, in machine learning, random sampling is a widely used method for hyper-parameter optimisation.
Unfortunately, when searching high-dimensional configuration spaces, random sampling is known
to achieve poor coverage and can waste substantial effort evaluating poorly performing
candidate configurations.
A more sophisticated approach to algorithm configuration is provided by so-called racing
methods~\cite{birattari_racing_2002}, which iteratively evaluate candidate configurations
on a series of inputs and eliminate candidates as soon as they can be shown to
significantly fall behind the current leader of this race.
Local search based configurators, on the other hand,
iteratively improve a given configuration by applying small changes and avoid stagnation
in local optima by means of diversification techniques (see, e.g., \cite{hutter_paramils_2009}).
More recently, model-based algorithm configuration methods have gained prominence.
These are based on the key idea of constructing a model of how the parameters affect performance;
this empirical performance model is then used to select candidate configurations to be evaluated
and updated based on the results from those runs.
Arguably the best known model-based configurator (and the current state of the art) is
SMAC~\cite{hutter_sequential_2011}, which we use in the following.
SMAC and the other general-purpose algorithm configuration methods mentioned
above have been applied with great success to a broad range of problems,
including propositional satisfiability~\cite{hutter_automatic_2007}, mixed
integer programming~\cite{hutter_automated_2010}, machine learning
classification and regression~\cite{thornton_auto-weka_2013}, and improving the
performance of garbage collection in Java~\cite{lengauer_taming_2014}.
The existence of effective algorithm configuration procedures has implications for the
design and development of high-performance software. Namely, rather than limiting design
choices and configurable options to make it easier (for human developers) to find good settings,
there is now an incentive to introduce, expose and maintain many design choices,
and to let automated configuration procedures find performance-optimized configurations for specific application contexts.
This is the core idea behind the recent \emph{Programming by Optimization (PbO)} paradigm~\cite{pbo}.
However, if software is not developed using specific tools supporting PbO, the application of automated
configuration procedures requires the manual specification of a \emph{configuration space} based on the definitions of and constraints on all configurable parameters.
For complex and highly parameterised software, such as the JavaScript engines we consider in this work,
this process can be somewhat involved, since it not only involves collecting the names and domains
(i.e., permissible values) for all parameters, but also conditionality relations between them (e.g., parameter $a$'s value only matters if parameter $b$ has value $x$),
and constraints that rule out certain configurations (e.g., configurations known to cause crashes or faulty behaviour).
Furthermore, in typical applications of automated algorithm configuration, developers
need to carefully construct a set of `training' inputs that is representative of those encountered
in the intended application context of the algorithm or software to be configured.
If automated configuration is applied to produce a performance-optimised configuration using training inputs unlike those
seen in typical use, the resulting configuration is unlikely to perform as well in the actual application as on the training set used
as the basis for configuration. (This, of course, also holds for manual configuration, but the effect tends to become more pronounced if more effective optimisation methods are used.)
Interestingly, JavaScript engine parameter optimisation (and, more generally, certain flavours of compiler optimisation)
differs from most other applications of automated algorithm configuration, in that it makes sense to use a training set consisting of a single input in the form of a program source, whose performance is to be optimised by means of compilation and execution with specific engine parameters.
\todo{HH}{Is it just compilation or also execution?CF: Fixed.}
Consider a popular Node.js application running a JavaScript workload that does not
change appreciably for each request it receives. Any performance increases on that particular workload
are of immediate, significant benefit, and performance decreases on other hypothetical workloads are irrelevant.
These situations are ideal for our approach, as they allow for the performance gains achieved in offline performance optimisation to be leveraged across potentially hundreds of millions of future runs of the software thus optimised.
\section{Automated Configuration of JavaScript Engines}
\subsection{JavaScript Engines}
We consider two state-of-the-art JavaScript engines in this work;
JavaScriptCore\footnote{\url{https://trac.webkit.org/wiki/JavaScriptCore}} and
Google's V8\footnote{\url{https://code.google.com/p/v8/}}.
This choice was motivated by the popularity and availability of these engines, rather than absolute performance.
We note that our goal was not to compare the performance of the two engines, but rather to investigate to what extent the default configuration of each can be improved.
JavaScriptCore (or JSC) is an optimising JavaScript virtual machine
developed as the JavaScript engine for WebKit; it is used in Apple's Safari browser on both OS X and iOS,
as well as in many other Apple software projects, web browsers, and in a WebKit extension of Node.js.
It contains a low-level interpreter (LLInt), a simpler baseline just-in-time (JIT) compiler, another JIT compiler
optimizing for low latency (DFG JIT), and a JIT compiler optimizing for high throughput (FTL JIT). All of these components
can be active simultaneously for different blocks of code, based on execution thresholds, and blocks can be optimised (and deoptimised) between them many times. In fact, a recursive function can be executing in different JITs (or the LLInt) simultaneously at
different levels of the recursion. Our JSC parameter space contains 107
parameters (Table~\ref{tab:compiler-parameter-spaces}), where most
of the parameters have numerical domains. These numerical parameters mostly control counters and thresholds for activating
various functionality, and for triggering optimisation/deoptimisation between the LLInt and the various JITs.
\begin{table}
\begin{center}
\begin{tabular}{r@{\hskip 2em}r@{\hskip 1em}rrr}
\input{tables/compiler-parameter-spaces}
\end{tabular}
\caption{
For each of the two JavaScript engines considered in this work, we give the
total number of parameters in the configuration space as well as how many have
Boolean, integer and real-valued domains, respectively. (There are no parameters
with categorical domains in either configuration space.)
}
\label{tab:compiler-parameter-spaces}
\end{center}
\end{table}
The V8 JavaScript engine was initially developed for Google's Chrome browser
and is now used in other web browsers such as Opera, in server-side applications
using projects like Node.js\footnote{\url{https://nodejs.org/}} and as a library
embedded in other software applications.
V8 is somewhat unique in that it does not contain an interpreter, but instead compiles
JavaScript code blocks directly to native machine code when they are first encountered, which is then
optimised continuously over the course of running on a given input. Our interpretation of V8's parameter configuration space contains 173 parameters, primarily
Boolean choices to enable or disable specific functionality. The remaining integer parameters
control various aspects of that functionality, including inlining levels, loop unrolling, garbage collection thresholds and
stack frame sizing.
In order to specify the parameter configuration space for our two JavaScript
engines, JSC and V8,
we determined the name and type of each parameter, based on the documentation and command-line parser source code.
Unfortunately, domains for the numerical parameters are not specified by the developers, and only few conditional dependencies are explicitly described. We therefore had to resort to educated guesses; when in doubt, we aimed to err on the side of larger domains (within reason). Each space was then refined by sampling 100\,000 random configurations and running the engines on a simple problem instance to check for segmentation faults and other abnormal behaviour. For both engines,
many crashing configurations were thus identified, leading to iterative refinement of the configuration spaces through domain reduction as well as by adding forbidden parameter combinations and conditional parameter dependencies.
\subsection{Benchmark Instances}
We have selected four benchmark sets containing heterogeneous JavaScript problem instances,
identified as relevant to the JavaScript engine development community and end
users.
We aimed to avoid bias towards benchmark sets preferred by particular
development teams. In particular, we included the benchmark sets developed by the
developers of JSC and V8.
Our benchmark suite comprises the Octane 2.0~\footnote{\url{https://developers.google.com/octane/}},
SunSpider 1.0.2~\footnote{\url{https://www.webkit.org/perf/sunspider/sunspider.html}},
Kraken 1.1~\footnote{\url{http://krakenbenchmark.mozilla.org}} and Ostrich~\cite{khan_using_2014} benchmark sets.
We created harnesses that allowed us to execute and measure these benchmarks
programmatically, outside of a browser environment. We note that the techniques we
use here readily extends to browser-based settings, albeit the integration effort would be higher.
Octane 2.0 is Google's JavaScript compiler benchmark suite and includes
18 real-world benchmarks that range over different types of tasks, including a 2D
physics engine, a PDF rendering engine, a portable game system emulator,
a regular expression generator as well as instances testing, e.g., node allocation and reclamation.
The SunSpider 1.0.2 benchmark set was developed by the Web\-Kit team and
contains 26 problem instances representing
a variety of different tasks that are relevant to real-world applications,
including string manipulation, bit operations, date formatting and cryptography.
Kraken 1.1 was developed by Mozilla and contains 14 problem instances that were extracted from
real-world applications and libraries. These benchmarks primarily cover web-specific tasks (e.g., JSON parsing),
signal processing (e.g., audio and image processing), cryptography (e.g., AES, PBKDF2, and SHA256 implementations)
and general computational tasks, such as combinatorial search.
Ostrich is based on benchmark suites for important numerical computation tasks, such
as OpenDwarf~\cite{feng_opencl_2012}. While the other benchmarks focus on the
types of computations that are common on the web, Ostrich provides a way to
measure the performance on computations that are becoming increasingly relevant
as JavaScript gains in popularity and is deployed in new contexts.
\subsection{Experimental Setup}
All of the experiments reported in the following were performed using a single
Microsoft Azure Cloud instance of type ``G5'' running a standard installation
of Ubuntu 15.04. This instance type has two 16-core processors with a total of 448GB of RAM;
it is the sole user of the underlying hardware, based on a one-to-one mapping to two Intel Xeon E5-2698A v3 processors.
We use JavaScriptCore r188124 and V8 version 4.6.40, release builds compiled from source using GCC 4.9.2.
Our version of the SMAC configurator is v2.10.03\footnote{\url{http://www.cs.ubc.ca/labs/beta/Projects/SMAC/}}, run using Oracle Java JDK 1.8.0\_25.
For each of our configuration scenarios, we performed 25 independent runs of SMAC with a 1 CPU-day runtime cutoff,
allocating a maximum of 60 CPU seconds to each run on a particular problem instance. The objective value to be minimised by
SMAC is the so-called \emph{Penalised Average Runtime} (PAR) score, which penalises timed-out and crashing runs by assigning them
an objective value of 10 times the runtime cutoff (PAR-10), and otherwise assigns an objective value of the CPU time used.
This greatly disincentivizes bad and invalid configurations, in order to bias the configurator against selecting them.
The incumbent configuration with the best PAR-10 score reported by SMAC after termination was selected as the final result of the configuration process, and a subsequent validation
phase was performed to run both the JSC/V8 default configuration and the optimised configuration selected by our procedure on the
entire problem instance set. For these validation runs, we perform 100 runs per configuration and benchmark instance, and compute the PAR-10 score across all runs for each configuration.
\todo{HH}{Not PAR-10? CF: How's this?}
We require repeated runs to obtain statistically stable results. Individual runs are very
short and subject to susbstantial noise from the environment, e.g.\ operating system
jobs and contention for shared memory. Through repeated runs and averaging, we
achieve more realistic results that are less affected by very short and very
long outlier runs.
\section{Empirical Results}
The purpose of our experiments is twofold. First, we intend to demonstrate that
the performance \emph{across a set of diverse benchmarks} can be improved by
using a different parameter configuration than the default. This would indicate that
compiler developers may want to adjust the default settings with which they ship
their compilers, or that users that focus on particular types of applications may wish to
do so themselves. It also demonstrates the potential for techniques that periodically adjust
the configuration of the JavaScript engine based on the types of JavaScript code run recently.
The second part of our experiments focusses on \emph{specific individual
benchmarks} and shows that performance can be improved significantly by
specialising the compiler configurations to a specific piece of code to run, rather
than being forced to accept tradeoffs due to competing requirements by a heterogeneous set of benchmarks.
This finding can be exploited in two ways: Users who run the same piece of JavaScript code
over and over again (e.g., in a server-side JavaScript application)
can benefit from offline tuning, while at the same time, very
short online configuration runs for code that a user's browser accesses
frequently can potentially optimise its performance.
\todo{LK}{Can we show some
examples of how many iterations SMAC needs before it finds a significantly
better than default config to give some credence to this? Maybe even random
sampling? HH: Good question.}
\todo{CF}{I can still trawl through the SMAC logs if we have time.}
\subsection{Results on Benchmark Sets}
\todo{HH}{I've rearranged the following to make it sound less defensive. If you like this, please rearrange the table to show the JSC results first. We may then also want to rearrange the order in which the two engines are introduced and discussed in 3.1. I like doing this, because it makes the order lexicographic.}
As can be seen in Table~\ref{tab:full-set-configuration-perf}, we obtained substantial performance improvements
for JavaScriptCore (JSC) on the Ostrich, Octane and Sunspider benchmark sets, indicating that the default configuration
of JSC leaves room for optimisation.
This is not the case for V8, where we did not find significant improvements for any of our benchmark sets,
which suggests that the default parameter values are already well optimised.
This may seem disappointing, but needs to be viewed in light of the fact that compiler developers test against these
same benchmarks, and have much incentive, through constant competition, to be successful in their efforts to find the best configurations.
It is therefore remarkable that we achieved sizeable performance gains for JSC, even on the SunSpider benchmark developed by the WebKit team (as noted earlier, WebKit uses the JSC engine).
\begin{table*}
\begin{center}
\begin{tabular}{l@{\hskip 2em}rrrr@{\hskip 1em}rrrr}
\input{tables/full-set-configuration-perf}
\end{tabular}
\caption{
Validation results using 100 runs per problem instance for each of our 4 configuration scenarios using complete instance sets.
We give results for the JavaScriptCore and V8 default configurations, as well as for the best configuration
obtained by SMAC (as identified by training performance).
We note that the configurations found for JSC exhibit significant
performance improvements over the entire instance set, while those
for V8 show only marginal improvement over the defaults.
}
\label{tab:full-set-configuration-perf}
\end{center}
\end{table*}
\subsection{Results on Individual Benchmark Instances}
When configuring the JavaScript engine parameters for individual instances from
our benchmark sets, we obtain much greater improvements than for the complete
sets. We selected the five most promising individual instances for the
experiments in this section to keep the resource requirements moderate.
We chose the instances based on where we observed performance improvements in
the experiments that optimised the configuration across the entire benchmark
sets.
Three of these instances are taken from the Ostrich set: graph-traversal, sparse-linear-algebra, and structured-grid,
and two instances stem from the Octane set: PDFjs and Splay.
Results from these experiments are shown in Table~\ref{tab:individual-instance-configuration-perf}, and we show additional
empirical cumulative distribution functions of running time and scatter plots for the default \emph{vs} optimised configuration in Figure~\ref{fig:ostrich-sparse-rtd-scatter} and Figure~\ref{fig:octane-rtd-scatter}.
On Ostrich graph-traversal or structured-grid, not shown in the table and figures, we have not obtained
significant performance improvements for either of the two engines.
\begin{table*}
\begin{center}
\begin{tabular}{l@{\hskip 2em}rrrr@{\hskip 1em}rrrr}
\input{tables/individual-instance-configuration-perf}
\end{tabular}
\caption{
Validation results using 100 runs per problem instance for 3 configuration scenarios using a single problem instance,
one from our Ostrich set (Sparse Linear Algebra) and two from the Octane set (Splay and PDFjs).
We omit two other experiments on Ostrich instances (Graph Traversal and Structured Grid), where neither compiler showed
any improvement after configuration.
We give results for the JavaScriptCore and V8 default configurations, as well as for the best configuration
obtained by SMAC (as identified by training performance).
}
\label{tab:individual-instance-configuration-perf}
\end{center}
\end{table*}
\begin{figure*}
\begin{center}
\subfigure[JSC - Octane - PDFjs]{
\includegraphics[width=0.4\textwidth]{plots/jsc-octane2-pdfjs-single-8core-rtd}
\label{fig:jsc-octane-pdfjs-rtd}
}
\subfigure[JSC - Octane - PDFjs]{
\includegraphics[width=0.4\textwidth]{plots/jsc-octane2-pdfjs-single-8core-scatter}
\label{fig:jsc-octane-pdfjs-scatter}
}
\\
\subfigure[JSC - Octane - Splay]{
\includegraphics[width=0.4\textwidth]{plots/jsc-octane2-splay-single-8core-rtd}
\label{fig:jsc-octane-splay-rtd}
}
\subfigure[JSC - Octane - Splay]{
\includegraphics[width=0.4\textwidth]{plots/jsc-octane2-splay-single-8core-scatter}
\label{fig:jsc-octane-splay-scatter}
}
\caption{
For the Octane Splay and PDFjs invidual-instance configuration scenarios, we
show empirical CDFs of runtime for 100 runs on the respective problem instance, along
with scatter plots vs. the default configuration.
}
\label{fig:octane-rtd-scatter}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\subfigure[JSC - Ostrich - Sparse LA]{
\includegraphics[width=0.4\textwidth]{plots/jsc-ostrich-sparse-single-8core-rtd}
\label{fig:jsc-ostrich-sparse-rtd}
}
\subfigure[JSC - Ostrich - Sparse LA]{
\includegraphics[width=0.4\textwidth]{plots/jsc-ostrich-sparse-single-8core-scatter}
\label{fig:jsc-ostrich-sparse-scatter}
}
\\
\subfigure[V8 - Ostrich - Sparse LA]{
\includegraphics[width=0.4\textwidth]{plots/v8-ostrich-sparse-single-8core-rtd}
\label{fig:v8-ostrich-sparse-rtd}
}
\subfigure[V8 - Ostrich - Sparse LA]{
\includegraphics[width=0.4\textwidth]{plots/v8-ostrich-sparse-single-8core-scatter}
\label{fig:v8-ostrich-sparse-scatter}
}
\caption{
Considering the Ostrich Sparse Linear Algebra individual-instance configuration scenario, we
show empirical cumulative distribution functions (CDFs) of runtime for 100 runs on the respective problem instance, along
with scatter plots vs. the default configuration. The CDFs show the
probability that a run will complete within a certain amount of time
as a function of the time, as observed empirically. That is, a
finished run of an instance at a particular time increases the
probability.
}
\label{fig:ostrich-sparse-rtd-scatter}
\end{center}
\end{figure*}
Overall, the performance improvements on these individual-instance configuration scenarios are surprisingly pronounced. JavaScriptCore achieves a relative performance improvement of 35.23\%
over the default configuration on the Octane Splay benchmark, and of 14.76\% on Octane PDFjs.
For V8, we observed a 10.13\% improvement over the default on Ostrich sparse-linear-algebra.
\todo{HH}{Add some discussion of the RTDs (mention randomisation in the code, measurement noise) - make it clear why we even need multiple runs on the same instance for evaluation. I don't understand why the instances shown in Figure 1 were chosen. I'd show detailed results for those mentioned above in the text.)}
Overall, these results are remarkable as even new code optimisation methods often only
result in performance improvements by single-digit percentages. We hypothesize that there are some specific aspects of these problem instances
which differ sufficiently from the other instances in their respective benchmark sets, that these configurations cannot
be successfully be used across those entire sets, but are very effective on the individual instance in question. We present
some preliminary results towards identifying the source of these improvements in the following.
\subsection{Time to Find Improving Configurations}
Even when considering the remarkable performance improvements seen in our individual-instance configuration
experiments, there may be some concern about the time required to find these improving configurations,
given that we used 25 independent SMAC runs of 1 CPU day to achieve these.
Upon further investigation, in all of our individual-instance configuration scenarios, the final optimised configuration was
found in less than 3 CPU hours of runtime, with initial improvements over the default configuration typically
found in less than 5 CPU minutes.
Longer runtimes are required for the complete instance set configuration scenarios, but even in those
cases, the final configuration was found in less than 6 CPU hours, with initial improving configurations
typically being found in less than 1 CPU hour.
In practice, a much smaller configuration budget would be sufficient to achieve
qualitatively similar results. In fact, we observed the first improvements after only a
few minutes of configuration.
\subsection{Changed Parameter Values}
In order to better understand the source of our individual-instance performance improvements, we empirically analysed the parameters changed from their default values using \emph{ablation analysis}~\cite{fawcettHoos2015}.
This approach has been previously used successfully to assess the importance of parameter changes observed in applications of
automated algorithm configuration techniques to propositional satisfiability, mixed-integer programming and AI-planning problems.
Ablation analysis greedily constructs a path through the parameter configuration space from
the default to a given target configuration, selecting at each stage the single parameter modification resulting
in the greatest performance improvement. The order of the resulting modifications reflects the relative contribution to the overall performance improvements obtained by the configuration process, where later changes may occasionally achieve bigger improvements that would not have been possible before earlier modifications to the default configuration.
The three parameter modifications resulting in the greatest performance improvement for the Octane Splay and PDFjs
instances are shown in Table~\ref{tab:jsc-octane-splay-ablation} and Table~\ref{tab:jsc-octane-pdfjs-ablation}, respectively.
For JavaScriptCore on Octane Splay, the parameter changes that
achieved the most significant improvements are related to object tracking and garbage collection.
For the Octane PDFjs benchmark instance, the configuration process resulted in modifications to various parameters controlling
memory management and the aggressiveness of the code optimisation. We note that numberOfGCMarkers is
important in both cases, where the value is changed to 1 from a default of 7.
This parameter controls the amount of parallelism in the garbage collector. Here,
reduced parallelism avoids overhead and improves overall performance.
While the portion of the relative improvement indicated in the tables is approximate due to the
nature of the ablation analysis procedure, it appears that in both cases, over 90\% of the observed
relative improvement can be explained by the modification of the three parameters shown. This is
consistent with previous results using ablation analysis, where in many scenarios, the vast majority of the improvement
was observed to be achieved by modifying a small set of parameters.
Of course, identifying these parameters in \emph{post hoc} ablation analysis is much easier than determining them
within the configuration process that gave rise to the optimised configurations thus analysed.
\begin{table*}
\begin{center}
\begin{tabular}{l@{\hskip 2em}r@{\hskip 2em}rrr}
distance from default & parameter modified & from & to & approx. portion of rel. impr. \\
\midrule
1 & numberOfGCMarkers & 7 & 1 & 38\% \\
2 & minCopiedBlockUtilization & 0.9 & 0.196 & 47\% \\
3 & collectionTimerMaxPercentCPU & 0.05 & 0.292 & 6\% \\
\end{tabular}
\caption{
Parameters modified from the respective default settings for JavaScriptCore in order to achieve the three highest
marginal performance gains on the Octane 2.0 instance ``Splay'', as determined by ablation analysis.
Reported marginal improvement is only approximate, as the ablation analysis procedure is not performing the full 100
runs per instance validation as in our configuration experiments.
}
\label{tab:jsc-octane-splay-ablation}
\end{center}
\end{table*}
\begin{table*}
\begin{center}
\begin{tabular}{l@{\hskip 2em}r@{\hskip 2em}rrr}
distance from default & parameter modified & from & to & approx. portion of rel. impr. \\
\midrule
1 & likelyToTakeSlowCaseMinimumCount& 20 & 56 & 41\% \\
2 & numberOfGCMarkers & 7 & 1 & 40\% \\
3 & forceDFGCodeBlockLiveness & false & true & 16\% \\
\end{tabular}
\caption{
Parameters modified from the respective default settings for JavaScriptCore in order to achieve the three highest
marginal performance gains on the Octane 2.0 instance ``pdfjs'', as determined by ablation analysis.
Reported marginal improvement is only approximate, as the ablation analysis procedure is not performing the full 100
runs per instance validation as in our configuration experiments.
}
\label{tab:jsc-octane-pdfjs-ablation}
\end{center}
\end{table*}
\subsection{Performance under Different Loads}
Modern computers have multiple processors, with multiple CPU cores each, and it is desirable to run
multiple processes simultaneously in order to take full advantage of the processing power thus provided.
However, other factors, such as shared caches, memory bandwidth and the I/O
subsystem can affect performance negatively, if too many processes are vying for
resources.
In order to investigate to which extent such factors may impact our experimental setup, we ran
different configurations of workloads. First, we utilized all 32 cores of the
machine used for our experiments by running 32 benchmark experiments in parallel. Second,
we ran only 8 experiments in parallel, leaving the remaining cores for operating system
processes.
\todo{HH}{I've reworded, but still find it unclear what was done. Was it the case that in the first case, 32 runs of the same benchmark instance were done in parallel? If so, we should state it clearly.}
The results show that there are significant differences. The graph-traversal
instance of the Ostrich benchmark set requires a large amount of memory
and sufficient memory bandwidth. With the machine fully loaded, we observe that
we easily find a parameter configuration that performs better than the default.
On the lightly loaded machine we are unable to do so, and the benchmark runs
significantly faster than on the fully loaded machine, even with the improved
configuration. This clearly indicates a memory bottleneck that can be mitigated
through configuration.
The default configuration of JavaScriptCore performs well on the SunSpider,
Kraken and Octane benchmarks on the fully-loaded machine, and we were unable to
find a better configuration of parameter settings. On the lightly loaded machine,
on the other hand, we did find better configurations for SunSpider and Octane. This
may indicate that the JavaScriptCore default configuration is optimised for a
highly-loaded machine, which is unlikely, when the engine is used inside a
browser on a user's desktop or laptop machine.
The fact that JavaScriptCore and V8 and exhibit different behaviour with respect to
how easy it is to improve on their default configurations on machines with
different load suggests that the benchmarking and tuning the respective development teams
perform may use different experimental setups.
This result shows that the optimisation of compiler flags should be done not only
for the machine that the code will be run on, but also for the expected load on
that machine -- configuring for a lightly loaded machine will yield different
results than configuring for a heavily loaded one. Furthermore, there is much promise
in switching between different configurations based on machine load.
\begin{table*}
\begin{center}
\scriptsize
\begin{tabular}{l@{\hskip 2em}r@{\hskip 1em}rrrrrrrrrr}
\input{tables/machine-load-stability}
\end{tabular}
\caption{
Using the Octane pdfjs problem instance, we performed 100 independent runs of the 25 SMAC configurations for JSC, as well as the JSC default configuration.
This was repeated 3 times with the same random seeds, first allowing 32 simultaneous runs and again allowing 8 simultaneous runs.
We give the PAR10 score for the default configurations, as well as for the 10 best configurations by validation score in each experiment (along with the configuration ID for each).
The configuration ID for the ``best training'' configuration of JSC on this instance is 14.
It is clear that the best configurations are quite different in the case of 32 simultaneous runs, even with a fixed instance and seeds.
As this variability disappears in the case of 8 simultaneous runs, we attribute it to noise from the load (and subsequent cache contention, etc.).
}
\label{tab:machine-load-stability}
\end{center}
\end{table*}
\section{Conclusions}
JavaScript is ubiquitous in the modern world wide web and increasingly spreading
into other areas that have traditionally been dominated by other programming languages. It is
used client-side in web browsers as well as server-side in backend
applications. Performance increasingly matters in practical JavaScript, as
applications grow in size and complexity.
In part, the success of JavaScript has been due to the availability of
highly optimised compilers that produce efficient code that can be executed
with minimal overhead. Just-in-time compilation and dynamic optimisations
further increase the performance of the code.
However, contemporary compilers have a large number of parameters, most of which
are only poorly documented. While the default configuration of these parameters
provides good performance in most cases, the parameter values need to be
optimised for the application at hand to get the best performance in all cases.
Exploring this huge and complex parameter space is a daunting task.
We apply a state-of-the-art, general-purpose automated configuration procedure
with an excellent track record in applications in machine learning and combinatorial optimisation
to the problem of
finding the best parameter configuration for JavaScript engines for a particular
set of problem instances. Sequential model-based optimisation leverages state-of-the-art
techniques from statistics, optimisation and machine learning to efficiently and automatically explore the
parameter space of an algorithm and to home in on promising configurations quickly.
Our experimental evaluation shows that notable performance improvements can be achieved
through automated configuration. Specifically, we demonstrate that the
performance of JavaScriptCore can be substantially improved on 3 out of 4
heterogeneous benchmark sets in common use for JavaScript compiler benchmarking.
We also show that JavaScriptCore (and to a lesser extent V8) can be specialised
to obtain runtime gains of up to 35\% on tasks such as PDF rendering.
This is particularly significant as we are optimising code that is run millions
of times. In contrast, algorithm configuration for combinatorial optimisation
problems considers the different setting where each problem instance needs to be
solved only once.
We believe that our results are promising and believe that our approach enables
many interesting applications and follow-up work.
We are currently planning additional work including a broader set of experiments,
additional analysis of the parameter space structure, a deeper investigation
into the effect of machine load on runtime performance and configuration, and an
investigation of the transferability of these configuration results to machines
other than those used for training.
\subsection*{Acknowledgements}
Part of this research was supported by a Microsoft Azure for Research grant.
HH also acknowledges support through an NSERC Discovery Grant.
\clearpage
|
\section{Introduction}
\label{sec:intro}
The study of the first luminous sources and the consequent epoch of reionization (EoR) occupies a central place in modern cosmology. Amongst the various probes, the redshifted 21-cm line is expected to be the most promising one, potentially illuminating our Universe even before the first stars started to shine \citep[see ][for recent reviews]{Furlanetto2016,McQuinn2016}.
Observations of the redshifted 21-cm line have progressed tremendously over the last few years, with improving upper limits being placed over a wide redshift range from both sky--averaged \cite[][]{Bowman2010,Bernardi2016} and fluctuations \cite[][]{Dillon2015,Jacobs2015,Trott2016,Ewall-Wice2016, Beardsley2016,Patil2017} 21-cm observations. The most stringent upper limits on fluctuations to date are the power spectrum measurements at $z = 8.4$ from observations with the Precision Array to Probe the Epoch of Reionization \citep[PAPER,][]{Parsons2010} that provide evidence for heating of the Intergalactic Medium prior to reionization \cite[][]{Parsons2014,Ali2015,Pober2015,Greig2016}.
The biggest challenge in measuring the 21-cm signal is the presence of foregrounds few orders of magnitude stronger than the cosmological signal \citep[e.g.][]{Jelic2008,Bernardi2009,Jelic2010,Zaroubi2012,Pober2013,Dillon2014,Parsons2014,Chapman2015}. A very effective technique in separating smooth--spectrum foregrounds and the 21-cm signal is the so called {\it foreground avoidance}: smooth--spectrum foregrounds occupy a wedge--shaped limited region of the two--dimensional power spectrum space whereas the remaining area -- the so--called EoR window -- is dominated by the 21-cm emission \citep[e.g.][]{Datta2010,Morales2012,Thyagarajan2013,Liu2014a,Liu2014b,Chapman2016}.
Deviations from frequency smoothness can, however spread power into the EoR window, potentially invalidating the avoidance assumptions. One such deviation has been recognized as leakage from Faraday rotated, polarized Galactic and extragalactic synchrotron emission \citep[][]{Bernardi2010,Jelic2010,Geil2011}. In particular, when the 21-cm power spectrum is estimated through the co-addition of redundant visibilities \citep[e.g.][]{Moore2013,Parsons2014}, leakage from polarized foregrounds due to asymmetric widefield primary polarization response cannot be corrected using similar approaches to interferometric imaging \citep[][]{Bhatnagar2008,Morales2009,Ord2010,Tasse2013,Asad2015} and it may ultimately jeopardize the 21-cm signal in the EoR window.
This paper follows the measurement of two dimensional polarized power spectra from \citet{Kohn2016} using PAPER observations. We extend the delay transform approach \citep{Parsons2012} used to avoid foregrounds in visibility--based 21-cm power spectra to the polarized case, defining a formalism that naturally includes the leakage due to widefield polarized primary beams. We develop models of polarized foregrounds to which we apply the formalism with the aim of simulating PAPER--like observations. We compared our simulated polarized power spectra with existing data and predict the leakage expected in 21-cm power spectra.
The paper is organized as follows: the formalism is derived in Section~\ref{sec:stokes}, simulations of actual observations are described in Section~\ref{sec:simulations}, and results discussed in Section~\ref{sec:results}. We conclude in Section~\ref{sec:conclusion}.
\section{Formalism}
\label{sec:stokes}
The relationship between the sky brightness distribution $s$ and the visibility ${\rm v}$ measured by a two--element interferometer is described by \citep[e.g.][]{thompson2008}:
\begin{equation}\label{eq:visibility_1}
{\rm v}({\boldsymbol{b}}, \nu)= \int_\Omega a(\hat{\boldsymbol{r}},\nu) \,
s(\hat{\boldsymbol{r}},\nu) \, e^{-2\pi i \nu \frac{{\boldsymbol{b}} \cdot \hat{\boldsymbol{r}}}{c} }\, d\Omega,
\end{equation}
where $s$ is the Stokes I parameter, ${\boldsymbol{b}} = (u,v,w)$\footnote{Throughout this paper, capital boldface letters are used to indicate matrices, small boldface letters to indicate vectors and small normal letters to indicate scalars.} is the vector representing the coordinates in meters in the plane of the array, $\hat{\boldsymbol{r}} = (l,m,n)$ is the unit vector representing the direction cosines on the celestial sphere \citep[see][for further details]{thompson2008}, $\nu$ is the observing frequency, $a$ describes the telescope primary beam response, $c$ is the speed of the light and $\Omega$ is the solid angle subtended by the source.
Throughout the paper we will assume array coplanarity and simulate zenith observations, leading to ${\boldsymbol{b}} = (u,v)$, $\hat{\boldsymbol{r}} = (l,m)$, $wn = 0$, and reducing equation~\ref{eq:visibility_1} to a two dimensional Fourier transform relationship.
Equation~\ref{eq:visibility_1} does not specify a polarization frame but it can be generalized to include the polarization state of the sky brightness using the measurement equation formalism \citep{Hamaker1996,Smirnov2011a}, where the input radiation field is related to the measured visibilities via $2 \times 2$ Jones matrices. If the intrinsic sky brightness distribution towards a line of sight $\hat{\boldsymbol{r}}$ at the frequency $\nu$ is described by the usual Stokes parameters ${\bf s}=(I,Q,U,V)^T$, with $T$ the transpose operator, the sky brightness distribution observed through the telescope primary beam ${\bf s'}=(I',Q',U',V')^T$ becomes \citep[e.g.][]{Ord2010}:
\begin{eqnarray}\label{eq:intfresp}
{\bf s}' (\hat{\boldsymbol{r}}, \nu) & = & {\bf A}(\hat{\boldsymbol{r}}, \nu) \, {\bf s} (\hat{\boldsymbol{r}}, \nu) \nonumber \\
& = & {\bf S}^{-1} \big[ {\bf J}(\hat{\boldsymbol{r}}, \nu) \otimes {\bf J}^* (\hat{\boldsymbol{r}}, \nu) \big] {\bf S} \, {\bf s} (\hat{\boldsymbol{r}}, \nu)
\end{eqnarray}
where ${\bf{J}}$ is the $2 \times 2$ Jones matrix representing the polarized primary beam response, $\otimes$ is the outer product operator, $^*$ denotes the complex conjugate and ${\bf{S}}$ is the matrix that relates the Stokes parameters to the orthogonal $x-y$ linear feed frame:
\begin{equation*}
{\bf{S}}=\frac{1}{2}\begin{pmatrix} 1 & 1 & 0 & 0 \\
0 & 0 & 1 & i \\
0 & 0 & 1 & -i \\
1 & -1 & 0 & 0 \end{pmatrix}.
\end{equation*}
The $4 \times 4$ matrix ${\bf A}$ can be seen as a mixing matrix between intrinsic and observed (primed) Stokes parameters:
\begin{equation*}
\begin{pmatrix}
I^{\prime} \leftarrow I & I^{\prime} \leftarrow Q & I^{\prime} \leftarrow U & I^{\prime} \leftarrow V\\
Q^{\prime} \leftarrow I & Q^{\prime} \leftarrow Q & Q^{\prime} \leftarrow U & Q^{\prime} \leftarrow V \\
U^{\prime} \leftarrow I & U^{\prime} \leftarrow Q & U^{\prime} \leftarrow U & U^{\prime} \leftarrow V \\
V^{\prime} \leftarrow I & V^{\prime} \leftarrow Q & V^{\prime} \leftarrow U & V^{\prime} \leftarrow V \\
\end{pmatrix}.
\end{equation*}
Equation~\ref{eq:visibility_1} can be extended to the polarized case by defining the four ``Stokes" visibility products ${\bf{v}} = ({\rm v}_{I},{\rm v}_{Q},{\rm v}_{U},{\rm v}_{V})^T$:
\begin{eqnarray}\label{eq:visibility_2}
{\bf v} ({\boldsymbol{b}}, \nu) & = & {\bf S}^{-1} \, {\bf v}_c ({\boldsymbol{b}}, \nu) \nonumber \\
& = & {\bf S}^{-1} \int_{\Omega} \big[ {\bf J}(\hat{\boldsymbol{r}}, \nu) \otimes {\bf J}^* (\hat{\boldsymbol{r}}, \nu) \big] {\bf S} \, {\bf s} (\hat{\boldsymbol{r}}, \nu) \, e^{-2\pi i \nu \frac{{\boldsymbol{b}}.\hat{\boldsymbol{r}}}{c} }\, d\Omega \nonumber \\
& = & \int_{\Omega} {\bf A}(\hat{\boldsymbol{r}}, \nu) \, {\bf s} (\hat{\boldsymbol{r}}, \nu) \, e^{-2\pi i \nu
\frac{{\boldsymbol{b}}.\hat{\boldsymbol{r}}}{c} }\, d\Omega
\end{eqnarray}
where ${\bf v}_c = ({\rm v}_{xx},{\rm v}_{xy},{\rm v}_{yx},{\rm v}_{yy})^T$ is the four cross--polarization correlator output.
Equation~\ref{eq:visibility_2} recasts the full sky formalism developed by \cite{Smirnov2011a} in the Mueller matrix form.
We are now in the position to define four polarization power spectra by applying the delay transform to equation~\ref{eq:visibility_2}.
The delay--transform is the Fourier transform of a single visibility along frequency \citep[][]{Parsons2009,Parsons2012}:
\begin{equation}\label{eq:delay_1}
\tilde{\rm v}({\boldsymbol{b}}, \tau)= \int_B {\rm w}(\nu) \, {\rm v}({\boldsymbol{b}}, \nu) \, e^{-2\pi i \nu\tau}\,d\nu\, ,
\end{equation}
where $B$ is the observing bandwidth, ${\rm w}$ is the window function and $\tau$ represents the geometric delay between antenna pairs:
\begin{equation}\label{eq:delay_2}
\tau = \frac{{\boldsymbol{b}} \cdot \hat{\boldsymbol{r}}}{c}.
\end{equation}
The delay transform is related to the 21-cm power spectrum $p(k)$ as \citep[][]{Parsons2012,Thyagarajan2016}:
\begin{align}\label{eq:power_spectrum}
p(k) &= p \left ( \sqrt{k_\perp^2 + k_\parallel^2} \right ) \nonumber\\
&=\Big(\frac{\lambda^2}{2k_{\rm B}}\Big)^2 \Big(\frac{D^2\Delta D}{B} \Big) \frac{1}{q} \, |{\tilde{\rm v}}(|{\boldsymbol b}|,\tau)|^2,
\end{align}
with
\begin{eqnarray}
k_\perp & = &\frac{2\pi \frac{|\boldsymbol{b}|}{\lambda}}{D}\\
k_\parallel & = &\eta\frac{2\pi f_{21}H_0E(z)}{c \, (1+z)^2}
\end{eqnarray}
where $\lambda$ is the observing wavelength, $k_B$ is the Boltzmann constant, $D$ is the
transverse comoving distance, $\Delta D$ is the comoving depth along the line of sight corresponding to the bandwidth $B$, $f_{21}$ is the 21-cm line rest frequency, $z$ is the redshift, $H_0$ is the Hubble constant and $E(z) = \sqrt{\Omega_M(1+z)^3 +\Omega_k(1+z)^2 + \Omega_\Lambda}$. In this work, we use $H_0=100\, h$~km~s$^{-1}$, $\Omega_M=0.27$, $\Omega_k = 0$ and $\Omega_\Lambda = 0.73$.
The power spectrum normalization volume $q$ is \citep{Thyagarajan2016}:
\begin{equation}
q= \int_{\Omega} \int_B |a(\hat{\boldsymbol{r}},\nu) \, {\rm w}(\nu)|^2 \, d\Omega \, d\nu.
\end{equation}
The polarized case is obtained by substituting equation~\ref{eq:visibility_2} in the delay transform:
\begin{equation}\label{eq:delay_3}
{\tilde{\bf{v}}}({\boldsymbol{b}}, \tau) = \int_B {\rm w}(\nu) \, {\bf v} ({\boldsymbol{b}}, \nu) \, e^{-2\pi i \nu\tau} \, d\nu\,
\end{equation}
and by using the Hadamard (or element--wise) product $\circ$ in order to extend equation~\ref{eq:power_spectrum} to four power spectra ${\bf p} = \big(p_I,p_Q,p_U,p_V)^T$:
\begin{eqnarray}\label{eq:power_spectrum_2}
{\bf p}(k) = \Big(\frac{\lambda^2}{2k_{\rm B}}\Big)^2 \frac{D^2\Delta D}{B} {\bf Q}^{-1} \left \{ {\tilde{\bf{v}}}(|{\boldsymbol b}|,\tau) \circ {\tilde{\bf{v}}}(|{\boldsymbol b}|,\tau)^* \right \}.
\end{eqnarray}
As the off--diagonal elements of the $\bf A$ matrix are much smaller than the diagonal elements, the normalization matrix $\bf Q$ may be written as a diagonal matrix with diagonal elements:
\begin{equation}\label{eq:omegabw}
\textrm{diag}({\bf Q}) \approx \int_{\Omega} \int_B \textrm{diag} \left ( {\bf A} \, {\rm w} \right ) \left [ \textrm{diag} ({\bf A}\, {\rm w}) \right ]^* \, d\Omega\, d\nu.
\end{equation}
Dimensionless polarized power spectra ${\bf \Delta}^2 = (\Delta_I^2,\Delta_Q^2,\Delta_U^2,\Delta_V^2)$ can be defined in analogy with the scalar case:
\begin{equation}\label{eq:power_spectrum_3}
{\bf \Delta}^2 (k) = \frac{k^3}{2 \pi^2} \, {\bf p} (k).
\end{equation}
Equation~\ref{eq:power_spectrum_2} is one of the main results of our work: if the sky emission is unpolarized, $p_I(k)$ reduces to $p(k)$, but appropriately combines the two orthogonal polarizations and their relative primary beams to form the total intensity power spectrum estimator. The remaining terms represent the visibility--based polarization power spectra which are dominated by leaked total intensity foreground emission.
In the presence of a polarized sky, $p_I(k)$ naturally includes the leakage of polarized emission due to widefield primary beams. Equation~\ref{eq:power_spectrum_2} therefore generalizes the approach of \cite{Moore2013} to full polarization and provides a framework to simulate the expected leakage to the 21-cm power spectrum given both a polarized foreground and an instrument model.
\section{Simulations}
\label{sec:simulations}
In order to evaluate equation~\ref{eq:power_spectrum_2} we need three ingredients: a model of the PAPER dipole beam, an array configuration and a polarized sky model.
We used the FEKO\footnote{\url{https://www.feko.info/product-detail}} package in order to obtain a model of the PAPER x--y complex dipole patterns in the $100-200$~MHz range, spaced 10~MHz apart. The model is based on the dipole physical dimensions and includes a reflective mesh positioned above the ground. Examples of the corresponding ${\bf A}(\hat{\boldsymbol{r}}, \nu)$ matrices are shown in Figure~\ref{fig:Mueller_mat} whereas Figure~\ref{fig:beam_freq} shows the frequency behaviour of the first row of ${\bf A}(\hat{\boldsymbol{r}}, \nu)$ (first three elements) for a few selected lines of sight. The $A_{00}(\hat {\bf r},\nu)$ term has a smooth frequency behaviour that decreases slowly and monotonically from zenith to the horizon. This behaviour is qualitatively in agreement with the beam model presented in \cite{Parsons2010}, although we defer a more quantitative comparison to future work. The off-diagonal terms representing the leakage to Stokes~I have a more complex spatial and spectral behaviour: their magnitude is essentially negligible at zenith whereas it becomes more than 10\% already at $70^\circ$ altitude.
We used the configuration of the PAPER 32-element imaging array \citep[][Figure~\ref{fig:config}]{Stefan2013,Jacobs2013,Kohn2016} to simulate a baseline distribution that covers a relatively wide range of $k_\perp$ values while retaining the 30~m baseline sample used in the estimate of the 21-cm power spectrum \citep[][]{Parsons2014,Jacobs2015,Ali2015}.
We simulated a drift scan observation corresponding to actual PAPER 21-cm observations, i.e. spanning the $0^{\rm h} < {\rm LST} < 8.5^{\rm h}$ range with a 10~minute cadence, covering the $120-180$~MHz frequency range with 500~kHz wide channels. For each frequency channel, the simulated ${\bf A}$ matrices were obtained by averaging over the two closest frequencies at which the FEKO simulations were carried out.
\begin{figure*}
\centering
\includegraphics[width=1.\columnwidth]{Mueller_4_X_4_Images_130.pdf}
\includegraphics[width=1.\columnwidth]{Mueller_4_X_4_Images_150.pdf}
\caption{${\bf A}$ matrices simulated at 130 (left) and 150~MHz (right) respectively. From the simulated complex dipole patterns, ${\bf A}$ matrices are computed in the altitude--azimuth coordinate system with the $x$ dipole assumed to be aligned East-West. Here they are resampled on an $(l,m)$ regular grid over a $45^{\circ}$ field of view, centered at an arbitrary right ascension $\alpha=1^{\rm h} 24^{\rm m}$ - the declination remains fixed at the telescope location $\delta = -30^\circ 43' 17''$. Resampling is carried out through a weighted average amongst the three closest points, with weights equals to the inverse distance to the $(l,m)$ grid point.}
\label{fig:Mueller_mat}
\end{figure*}
\begin{figure}[ht]
\centering
\includegraphics[width=1.1\columnwidth]{beam_alt-freq.pdf}
\caption{Spectral behaviour of the ${\bf A}$ matrix (first row only) along a few selected lines of sight ${\bf {\hat r}}$ corresponding to a $20^\circ$ azimuth and $90^\circ$ (blue), $70^\circ$ (pink), $50^\circ$ (orange), $30^\circ$ (brown) and $10^\circ$ (green) altitude respectively. The $A_{00}(\hat {\bf r},\nu)$ term is normalized to unity at zenith by construction. The $A_{03}(\hat {\bf r},\nu)$ term is not included as no Stokes~V sky emission is assumed throughout the paper.}
\label{fig:beam_freq}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\columnwidth]{psa32config.pdf}
\caption{Array layout used for our simulations.}
\label{fig:config}
\end{figure}
For each time sample, frequency channel and baseline we performed the following operations:
\begin{enumerate}
\item generated the sky emission evaluating equation~\ref{eq:intfresp}, where the sky model consists of either a catalogue of point sources or a realization of diffuse Galactic emission (see Sections~\ref{sec:point_sources} and~\ref{sec:diffuse_emission} below);
\item simulated visibilities via a discrete Fourier transform implementation of equation~\ref{eq:visibility_2};
\item generated delay--transformed visibilities via a fast Fourier transform implementation of equation~\ref{eq:delay_3} where we used a Blackmann--Harris window function;
\item computed power spectra using equation~\ref{eq:power_spectrum_2}.
\end{enumerate}
In the following sections we describe the input foreground models to our simulations.
\subsection{Point source all-sky model}
\label{sec:point_sources}
Simulations of polarized point sources are completely defined by a catalogue that includes both the polarization fraction and the rotation measure (RM) values. Stokes $Q$ and $U$ parameters at any frequency $\nu$, for any source $i$ can indeed be computed as:
\begin{eqnarray}\label{eq:stokesIQU}
Q _{\nu,i} = \gamma_i \, I_{\nu,i} \cos(2 \chi_i) = \gamma_i \, I_{\nu_0,i} \left ( \frac{\nu}{\nu_0} \right )^{\alpha_i} \cos{\left ( 2 \, {\rm RM}_i \frac{c^2}{\nu^2} \right )} \nonumber \\
U_{\nu,i} = \gamma_i \, I_{\nu,i} \sin(2 \chi_i) = \gamma_i \, I_{\nu_0,i} \left ( \frac{\nu}{\nu_0} \right )^{\alpha_i} \sin{\left ( 2 \, {\rm RM}_i \frac{c^2}{\nu^2} \right )} \nonumber \\
\end{eqnarray}
where $\gamma$ is the polarization fraction, $I_{\nu_0}$ is the flux density at the $\nu_0$ reference frequency and $\alpha$ is the spectral index. We note that the knowledge of the absolute polarization angle is not necessary for the purpose of estimating the 21-cm power spectrum leakage, therefore we set it to be zero along each line of sight.
For the total intensity properties we used the \cite{HurleyWalker2014} catalogue that lists all the sources brighter than 120~mJy at 150~MHz and covers the Southern Hemisphere at $-58^{\circ}< \delta < -14^\circ$. Although this catalogue is only somewhat deep, it has the advantage of providing the actual source locations, flux densities and spectral indexes in the $100-200$~MHz band.
From the \cite{HurleyWalker2014} catalogue it is possible to generate a polarized catalogue assuming the RM and polarization fraction statistics respectively. Detailed, wide--area information on the low--frequency polarization properties of radio sources are still lacking to date, therefore we constructed a polarized catalogue derived from statical properties measured at higher frequencies.
The RM distribution was taken from the 1.4~GHz catalogue by \cite{Taylor2009}, one of the most comprehensive polarization catalogues to date. For high Galactic latitude sources ($|b| > 20^\circ$), the RM distribution fairly follows a Gaussian profile, with values as high as $\sim |100|$~rad~m$^{-2}$ (Figure~\ref{fig:RM_dist}). As the rotation measure is the integral of the magnetic field along the line of sight weighted by the electron density, we do not expect it to change with frequency, therefore, for each simulated source, we assigned an RM value drawn from the Gaussian best fit to the RM distribution of the high Galactic latitude sources in the \cite{Taylor2009} catalogue.
The statistics of the polarization fraction $\gamma$ is more uncertain as is, instead, expected to decrease with frequency due to internal Faraday dispersion \citep{Burn1966}. It is not, therefore, straightforward to extrapolate the average polarization fraction at 1.4~GHz as Faraday depolarization depends on the specific source physical conditions \citep[e.g., geometry and magnetic field strength,][]{Tribble1991}.
Recent observations \citep[][]{Bernardi2013,Mulcahy2014,Asad2016} have indeed started to show that the average polarization fraction of radio sources decreases from a few percent value at 1.4~GHz to less than 1\% at 150~MHz. \cite{Lenc2016} present the most stringent constraints to date to be $\langle \gamma \rangle \le 0.32\%$ at 154~MHz, derived from a sample of 187 sources. We conservatively treated this result as a measurement rather than an upper limit and assigned a polarization fraction value drawn from a uniform distribution between 0 and 0.32\% to each simulated source.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\columnwidth]{RM_dist.pdf}
\caption{The RM distribution for sources at Galactic latitude $|b| > 20^{\circ}$ from the \cite{Taylor2009} catalogue. The dotted line shows the Gaussian best fit with a 5.6~rad~m$^{-2}$ mean and a 20~rad~m$^{-2}$ standard deviation.}
\vspace{0.1cm}
\label{fig:RM_dist}
\end{figure}
\subsection{Galactic diffuse emission model}
\label{sec:diffuse_emission}
Observations of diffuse, Galactic polarized emission reveal a wealth of spatial structures in the interstellar medium (ISM) that are strongly frequency dependent. In particular, recalling that the Faraday depth $\phi$ of a synchrotron emitting region between two points $l_1$ and $l_2$ along an arbitrary line of sight $\hat{\boldsymbol{r}}$, given the electron density $n_e$ and magnetic field along the line of sight $B_\parallel$ is defined as \citep[][]{Burn1966,Brentjens2005}:
\begin{eqnarray}\label{eq:faraday_depth}
\phi({\hat{\boldsymbol{r}}},l_1,l_2) = 0.81 \int_{l_1}^{l_2} n_e B_\parallel dl,
\end{eqnarray}
observations below a few hundred MHz show multiple polarized structures in the ISM at different Faraday depth values for almost any given line of sight. \cite{Jelic2010} and \cite{Alonso2014} have recently attempted to model this complexity, although a realistic description requires the knowledge of the distribution of the polarized emission both on angular scales and in Faraday depth, which is only partially emerging from recent observations \citep{Bernardi2013,Jelic2014,Jelic2015,Lenc2016}.
In this work, we made the simplifying assumption to ignore the spatial and line--of--sight Faraday depth structure and consider the contribution from two representative Faraday depths integrated all the way to the observer's location. With this approximation, the Faraday depth coincides with the RM and equation~\ref{eq:stokesIQU} can be used to obtain Stokes $Q$ and $U$ all--sky maps too:
\begin{eqnarray}\label{eq:stokesQU-diffuse}
Q ({\hat{\boldsymbol{r}}, \nu}) & = & P(\hat{\boldsymbol{r}}) \cos{\left ( 2 \, \phi \frac{c^2}{\nu^2} \right )} \nonumber \\
U ({\hat{\boldsymbol{r}}, \nu}) & = & P(\hat{\boldsymbol{r}}) \sin{\left ( 2 \, \phi \frac{c^2}{\nu^2} \right )}.
\end{eqnarray}
The all--sky, polarized intensity map $P$ could be, in principle, derived from a total intensity one, however, interferometric observations of polarized Galactic emission at all radio frequencies are known to suffer from the so called ``missing short spacing problem": they filter out the large scale, smooth background emission and retain the small scale, Faraday--rotated foreground structures introduced by local ISM fluctuations in either the electron density or the magnetic field \citep[e.g.,][]{wieringa1993,Gaensler2011}. This effect leads to the lack of correlation between total intensity and polarized diffuse emission, with an apparent polarization percentage exceeding 100\% \citep[e.g.][]{Gaensler2001,haverkorn2003,Bernardi2003,schnitzeler2009,Iacobelli2013,Bernardi2010}, preventing from using a total intensity template for polarization simulations.
In order to overcome this problem, we follow an approach similar to \cite{Alonso2014}, who simulated a polarized foreground map $P(\hat{\boldsymbol{r}})$ at a reference frequency $\nu_0$ from the polarized spatial power spectrum $C_{\ell}^P$:
\begin{equation}\label{eq:PI-pspectrum-def}
\langle \tilde{P} ({\bf l}) \tilde{P} ({\bf l})^{*} \rangle = (2 \pi)^2 C_{\ell}^P \delta^{(2)} ({\bf l} - {\bf l}')
\end{equation}
where $\tilde{P}$ is the Fourier transform of $P$, ${\bf l}$ is the two dimensional coordinate in Fourier space, $\langle \, \rangle$ is the ensemble average, $\ell = \frac{180}{\Theta}$ with $\Theta$ the angular scale in degrees and $\delta^{(2)}$ is the two--dimensional Dirac function.
The synchrotron polarized power spectrum obtained from large--area, GHz--frequency surveys is well described by a power law \citep[e.g.,][]{LaPorta2006,Carretti2010}:
\begin{equation}\label{eq:PI-pspectrum}
C_{\ell}^P = A_{\ell_0}^P \Big( \frac{\ell}{\ell_0} \Big)^{-\beta^P},
\end{equation}
down to $\ell \sim 100-1000$, with $2 < \beta^P < 3$. Due to the strong frequency dependence of the polarized emission \citep[e.g.][]{Carretti2005}, the extrapolation to low frequencies is very uncertain.
\cite{Bernardi2009} and \cite{Jelic2015} measured the polarized spatial power spectrum at 150~MHz and found it to follow a power law with $\beta^P = -1.65$ in the $100 \le \ell \le 2700$ range, somewhat flatter than the higher frequency values. We therefore adopted this value for our simulations. As they both observed a relatively small ($6^\circ \times 6^\circ$) sky patch, their measurement of the power spectrum amplitude $A_{\ell_0}^P$ may be sample variance limited, therefore we
constrained it by using the 2400~square~degree survey carried out at 189~MHz by \cite{Bernardi2013}.
The survey shows significant difference in the levels of polarized emission as a function of Galactic latitude, with maximum emission around the south Galactic pole and mostly concentrated at $|\phi| < 12$~rad~m$^{-2}$.
We identified the $20^\circ \times 20^\circ$ area centered at $(l,b) \sim (200^{\circ},-80^{\circ})$ to be the brightest polarized emission region in the survey and selected the $\phi = 6, 12$~rad~m$^{-2}$ frames as the brightest frame and the representative of a typical high $\phi$ value for diffuse emission, in good agreement with the distribution of Faraday depth peaks recently measured in the Southern Galactic pole area at 154~MHz by \cite{Lenc2016}.
We labeled the models corresponding to these two frames as $D6$ and $D12$ and measured their polarized intensity root--mean--square (rms) to be $P_{{\rm rms},D6} =1$~K and $P_{{\rm rms},D12} = 0.21$~K at $\phi_{D6} = 6$~rad~m$^{-2}$ and $\phi_{D12} = 12$~rad~m$^{-2}$ respectively.
The power spectrum amplitude $A_{D6,D12}$ was obtained through its relationship with the measured rms value \citep[e.g.,][]{zaldarriaga2004}:
\begin{eqnarray}\label{eq:Trms}
P_{{\rm rms},D6,D12} & = & \sqrt{\sum_{\ell_1}^{\ell_2} \frac{(2 \ell + 1)}{4\pi}C_{\ell,D6,D12}^P} \nonumber \\
& = & \sqrt{\sum_{\ell_1}^{\ell_2} \frac{(2 \ell + 1)}{4\pi} A_{D6,D12} \Big(\frac{\ell}{\ell_0} \Big)^{-\beta^P}},
\end{eqnarray}
where we dropped the $\ell_0$ subscript for clarity and the exclusion of short baselines and the angular resolution of the \cite{Bernardi2013} survey set $\ell_1 \sim 100$ and $\ell_2 \sim 680$ respectively.
At this point we have all the necessary ingredients to generate the two $D6$ and $D12$ realizations of the diffuse polarized emission model using equation~\ref{eq:stokesQU-diffuse}. We generated a map $P_{D6} (\hat{\boldsymbol{r}})$ ($P_{D12} (\hat{\boldsymbol{r}})$) from the polarized power spectrum $C_{\ell,{D6}}^P$ ($C_{\ell,{D12}}^P$) using the Healpix \citep{Gorski2005} routine SYNFAST. We chose an ${\rm N}_{\rm side} = 128$ parameter that retains sufficient sampling for the 30~m baseline used for power spectrum estimation. We then substituted $\phi_{D6}$ ($\phi_{D12}$) in equation~\ref{eq:stokesQU-diffuse} to generate the Stokes $Q$ and $U$ full--sky maps corresponding to the $D6$ ($D12$) model realization.
\section{Results}
\label{sec:results}
In this section, we compare our simulations with existing observations \citep{Kohn2016}, following which
we predict the leakage expected in 21-cm power spectrum measurements. We then provide constraints on the average point source polarization fraction based on the data from \citet{Moore2015}.
\subsection{Polarized Power Spectra}
\label{subsec:unpolarized power_spectra}
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{MWA_n1500-25MHz-unpol.pdf}
\caption{Two--dimensional power spectra ${\bf p} = (p_I,p_Q,p_U,p_V)$ (left to right) over a 25~MHz bandwidth centered at 150~MHz, obtained from the brightest 1500 unpolarized total intensity sources from the \cite{HurleyWalker2014} catalogue. The simulation only includes Stokes I; Stokes Q, U and V are leakages due to the instrumental widefield effects. The white line marks the horizon limit and the orange line is 50~ns beyond.}
\label{fig:MWAwedges_n1500-unpol}
\end{figure*}
\begin{figure}
\includegraphics[width=1\columnwidth]{SphericalPSvsSaul-unpol_ratio.pdf}
\caption{Ratios of Stokes~$Q$ (solid green), $U$ (solid red) and $V$ (solid cyan) versus $I$ power spectra averaged over a $0.092<k_{\perp}<0.097\,h$~Mpc$^{-1}$ region, whose centre corresponds to a 175~m baseline. The dotted lines represent the corresponding power spectrum ratios from Figure~7 in \cite{Kohn2016}. The left and right vertical dashed lines mark the horizon limit and 50~ns beyond it respectively.}
\label{fig:SphericalPSvsSaul-unpol}
\end{figure}
We first tested our simulation framework with a sky model composed of total intensity sources from the \cite{HurleyWalker2014} catalogue. In this case we expected to reproduce the well established two--dimensional wedge--like total intensity power spectrum $p_I$ observed, for example, in \cite{Pober2013}, \cite{Thyagarajan2015b} and \cite{Kohn2016}. Figure~\ref{fig:MWAwedges_n1500-unpol} displays power spectra from this simulation using only the 1500 brightest sources down to a flux density threshold of 2~Jy. Power spectra change by only a few percent with the inclusion of all the sources down to the 120~mJy catalogue threshold, therefore the decreased computing load in using only the 1500 brightest sources merits the small loss in accuracy for the aim of this paper.
The simulated total intensity power spectrum $p_I$ shows the wedge--like morphology and power levels similar to \cite{Pober2013}, confirming that our formalism is consistent with previous works that are limited to the total intensity case. Polarized power spectra display, in this case, the leakage from total intensity due to widefield polarized primary beams. All polarized power spectra have a similar behavior, with emission confined in a wedge--like shape very similar to total intensity. The power ratio between the emission inside and outside the wedge is at the $10^9-10^{10}$ level, indicating that very little
chromatic structure has been introduced by the primary beam outside the wedge.
In order to provide a first order validation of the beam models used in simulations, we compared the ratio $R_i$ of polarized versus total intensity power spectra defined as
\begin{equation}\label{eq:beamerror}
R_i (k) = \frac{p_i (k)}{p_I (k)}, \, \, \, \rm{with } \, \, i = Q, U, V
\end{equation}
calculated from our simulations and from power spectra measured in a 5~hour transit observation with PAPER \citep{Kohn2016}. The ratio is insensitive to a possible different absolute normalization between the simulation and data. Figure~\ref{fig:SphericalPSvsSaul-unpol} shows that simulated $R_i$ ratios are generally fainter than the measured ones, although the simulated $R_U$ and $R_V$ substantially agree with the measured ones. The measured power spectra show emission that extends up to 50~ns beyond the horizon limit, after which they appear to be noise-dominated \citep{Kohn2016}. As our simulations essentially have no emission beyond the horizon limit, this explains why both simulated and measured $R_i$ approach unity outside the wedge.
The largest difference between simulations and data appears in $R_Q$ at $k < 0.2 \, h$~Mpc$^{-1}$, where our model underpredicts the measured value by more than one order of magnitude. This is the consequence of an excess of power found in $p_Q$ at small $k$ values by \citet{Kohn2016} that may be attributable to a calibration mismatch of the two orthogonal polarizations and, possibly, to intrinsic sky polarization not included in our simulations.
\subsection{Predictions of polarization leakage}
\label{subsec:polarized_spectra}
The simulation framework developed in this paper eventually aims to predict the amount of leakage expected in the measured 21-cm power spectra. In order to do so, we carried out two sets of simulations for the models described in Sections~\ref{sec:point_sources} and \ref{sec:diffuse_emission} respectively, where we explicitly set the point source total intensity model to zero, so that $p_I$ directly measures the leakage from polarized foregrounds. The resulting power spectra at $z = 8.5$ (150~MHz) for a fiducial 30~m baseline are shown in Figure~\ref{fig:SphericallPS_150MHz} and \ref{fig:SphericalPS-Diffuse} respectively.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{SphericallPS_f150MHz-bl30m-scaled.pdf}
\caption{Polarized power spectra $\Delta^2_{Q,U}$ (orange and green lines respectively) for a 30~m baseline and an 8~MHz bandwidth centered at 150~MHz ($z = 8.5$) from polarized point source sky model with the total intensity intentionally set to zero (see text for details). The magenta line represents $\Delta_I^2$, i.e. the predicted leakage to total intensity. The magenta dashed line shows the leakage when we assumed the polarization fraction to be distributed between 0 and 0.14\%. The shaded gray region represents power spectra of the 21-cm fiducial model from \cite{Lidz2008} with HI neutral fractions ranging between 0.21 and 0.82.}
\label{fig:SphericallPS_150MHz}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=1\columnwidth]{DiffuseStokesQU-8MHz-pk.pdf}
\includegraphics[width=1\columnwidth]{DiffuseStokesQU-8MHz-delta.pdf}
\caption{\textit{Left panel:} Polarized power spectra $P_{Q,U}$ (orange and green lines respectively) for a 30~m baseline and an 8~MHz bandwidth centered at 150~MHz ($z = 8.5$) from the diffuse polarized foreground $D12$ (solid) and $D6$ (dotted). The magenta line represents the leaked power spectra $P_I$. \textit{Right panel:} Same as left panel but for $\Delta^2_{I,Q,U}(k)$. The shaded gray region represents power spectra of the 21-cm fiducial model from \cite{Lidz2008} with HI neutral fractions ranging between 0.21 and 0.82.}
\label{fig:SphericalPS-Diffuse}
\end{figure*}
Polarized power spectra generated from the point source simulation exhibits significant super--horizon emission, i.e. almost constant power outside the $k \sim 0.06\,h$~Mpc$^{-1}$ horizon limit for a 30~m baseline (Figure~\ref{fig:SphericallPS_150MHz}). This behavior is expected from Faraday rotated foregrounds whose leakage is not confined in $k_\parallel$ space. Power spectra also appear featureless in $k$ space and this can be intuitively understood as a single RM value corresponds to a specific $k_\parallel$ value \citep{Moore2015}:
\begin{eqnarray}\label{eq:kpRM}
k_{\parallel} = \frac{4\, \lambda^2\,H(z)}{c \, (1+z)}\,\textrm{RM},
\end{eqnarray}
therefore, a population of point sources distributed over a broad range of RM values essentially displays power at any $k_\parallel$ value. This is the reason why diffuse emission power spectra show a characteristic knee--shape as a function of $k$ (Figure~\ref{fig:SphericalPS-Diffuse}): by construction, they only have structure at $\phi = 6$ and 12~rad~m$^{-2}$ corresponding to $k_\parallel \approx 0.02$ and $0.06 \, h$~Mpc$^{-1}$ respectively and their power, therefore, falls off at higher $k_\parallel$ values. Noticeably, the $D12$ model power spectrum is brighter than the $D6$ one at $k > 0.2\, h$~Mpc$^{-1}$ despite its normalization being five times smaller (see Section~\ref{sec:diffuse_emission}), showing that the super--horizon contamination depends more on the $\phi$ value rather than the intrinsic foreground brightness.
In terms of contamination to the 21-cm power spectrum, our predictions should be regarded as worst case scenarios, due to our conservative model assumptions. In both foreground models, leaked power spectra approximately behave as scaled versions of polarized power spectra, however, in the diffuse emission case, the leaked power spectrum is $\sim 0.03$~(mK)$^2$ for $0.3 < k < 0.5 \, h$~Mpc$^{-1}$, a reasonably negligible contamination to the 21-cm power spectrum. Bright, diffuse polarized foregrounds are therefore not a concerning contamination to the 21-cm power spectrum as long as their emission is confined at low Faraday depths as all the existing observations are showing \citep[e.g.][]{Bernardi2009,Bernardi2010,Bernardi2013,Jelic2014,Jelic2015,Lenc2016}.
From a pure avoidance perspective, therefore, knowledge of the polarized point source distribution is more relevant as it may contaminate high $k$ modes too: its leakage magnitude is a strong function of the average point source polarization fraction, becoming one order of magnitude smaller if a uniform distribution with a maximum value of 0.14\% is assumed (Figure~\ref{fig:SphericallPS_150MHz}, magenta dashed line). We will return on this point in the next section.
One natural by product of our formalism is the predicted fractional leakage $f$ per $k$ mode defined as the reciprocal of $R_i$:
\begin{equation}\label{eq:fractional-leakage}
f_i(k)=\frac{1}{R_i} = \frac{p_I (k)}{p_i (k)}
\end{equation}
Figure~\ref{fig:RatioI2QU} shows that the fractional leakage contributed by Stokes $Q$ and U is less that 3\% for $k < 0.5 \, h$~Mpc$^{-1}$ in the $8 < z < 10$ range and tends to increase with redshift. \cite{Moore2015} gave a simplified estimate of the fractional leakages that is consistent with ours within a factor of two.\\
Finally, we note that $f_i$ has a rather different behaviour as a function of $k$ than $R_i$, while they should be, in first approximation, similar, due to the fact that the ${\bf A}$ matrices are nearly symmetrical (Figure~\ref{fig:Mueller_mat}). There is an intrinsic difference due to the fact that the simulations presented here pertain to two different baselines, however, most of their difference is due to the input model. The unpolarized point source model (Figure~\ref{fig:SphericalPSvsSaul-unpol}) is very smooth in frequency by construction, leading to a very bright Stokes~$I$ power spectrum at small $k_\parallel$ values and, therefore, a corresponding $R_i$ at those modes; conversely, as mentioned above, the polarized point source model has power at essentially any $k_\parallel$ value by construction, leading to an almost flat $f_i$ as a function of $k_\parallel$.
\begin{figure}
\centering
\includegraphics[width=1.05\columnwidth]{RatioQU2I.pdf}
\caption{Stokes~$Q$ (top panel) and $U$ (bottom panel) fractional leakage per $k$ mode for a 30~m baseline estimated from the polarized point source model at $z=8$ (blue line), $z=9$ (green line) and $z=10$ (red line) respectively. Error bars were calculated as the standard deviation of ten random realizations of the model.}
\label{fig:RatioI2QU}
\end{figure}
\subsection{Constraining the polarization fraction}
\label{subsec:polarization_fraction}
We compared our predictions for the point source model - the worst expected contamination - with the polarized power spectrum measurements from a 30~m baseline deep integration with the PAPER--32 array \citep{Moore2015}. Their $\Delta^2_{Q_m}$ and $\Delta_{U_m}^2$ at 126 and 164~MHz (reported here in Figure~\ref{fig:SphericallPS-David}) are essentially consistent with the noise level in the $\Delta k = [0.2,0.45] \, h$~Mpc$^{-1}$ range. We used these results to constrain our polarized power spectrum $\Delta_{Q}^2$ to $\Delta_{Q'}^2$ at the same frequencies to be:
\begin{equation}
\Delta_{Q'}^2 = \frac{\langle \Delta^2_{Q,U} \rangle_{\Delta k}}{\langle \Delta_{Q_m,U_m}^2 \rangle_{\Delta k}} \Delta_{Q}^2 = r \, \Delta_{Q}^2
\end{equation}
where $\langle \, \rangle_{\Delta k}$ indicates the average over the $\Delta k$ range for both Stokes parameters. Similarly, $\Delta_{U}^2$ is constrained to $\Delta_{U'}^2$ at the corresponding frequencies. In order to be consistent with the data, the simulated power spectra need to be scaled down by, at least, $r \sim 0.1$ at 126~MHz (Figure~\ref{fig:SphericallPS-David}; left panel), whereas they are already consistent (i.e. fainter) with the measurements at 164~MHz (Figure~\ref{fig:SphericallPS-David}; right panel).
These results can be used to improve our assumptions on the point source polarization fraction, allowing it to evolve with frequency. Defining $\gamma_{126}$ the polarization fraction at 126~MHz and recalling that
\begin{equation}
\Delta_{Q,U}^2 \, \propto \, \langle \gamma^2 \rangle
\end{equation}
and
\begin{equation}
\langle \gamma^2 \rangle = \frac{2b}{3} \langle \gamma \rangle
\end{equation}
if $\gamma$ follows a uniform distribution between 0 and $b$, the comparison with the \cite{Moore2015} power spectra yields $\langle \gamma_{126} \rangle \le 0.1\%$, approximately a factor of three smaller than our model assumption.
It is interesting to note that such constraint qualitatively meets the expectations of Faraday depolarization models \citep[][]{Burn1966,Tribble1991} that predict the polarization fraction to decrease at longer wavelengths. Although the estimated leaked power spectra may still remain above the expected 21-cm power spectra, they are now more than one order of magnitude fainter than the initial model predictions.
One caveat of our comparison with real data is related to the role of ionospheric Faraday rotation. \cite{Moore2015} already pointed out that averaging visibilities over many days of observations to form polarized power spectra leads to significant depolarization due to time variable ionospheric Faraday rotation. Without any correction, polarized power spectra measured in an actual observation \citep[e.g.][]{Moore2015} can still be used to predict the leakage as we showed above, although they cannot be used to model the intrinsic sky properties and, therefore, straightforwardly predict the leakage contamination in a different 21-cm observation. In this respect, the constraints we placed on the average polarization fraction of extragalactic radio sources should be seen as constraints on the {\it effective} (i.e. modulated by ionospheric Faraday rotation) rather than the intrinsic fraction.
The effect of ionospheric Faraday rotation could be directly included in our formalism in equation~\ref{eq:intfresp} but we leave this for future work \citep{Aguirre2016}.
\begin{figure*}
\centering
\includegraphics[width=1\columnwidth]{SphericallPS_David_126Mhz-scaled.pdf}
\includegraphics[width=1\columnwidth]{SphericallPS_David_164Mhz-scaled.pdf}
\caption{\textit{Left panel:} Polarized power spectra $\Delta^2_{Q'}$ and $\Delta_{U'}^2$ (orange and green respectively) from our point source simulation scaled down to match the observed $\Delta^2_{Q_m}$ and $\Delta_{U_m}^2$ (blue and gray circles respectively) from \citet{Moore2015} at 126~MHz ($z=10.3$). \textit{Right panel:} Polarized power spectra $\Delta^2_{Q'}$ and $\Delta_{U'}^2$ predicted from our point source simulation at 164~MHz ($z = 7.7$) compared with the observed $\Delta^2_{Q_m}$ and $\Delta_{U_m}^2$ (blue and gray circles respectively) from \citet{Moore2015} at the same frequency. We note that our predictions are compatible with the observed upper limits.}
\label{fig:SphericallPS-David}
\end{figure*}
\section{Discussion and Conclusions}
\label{sec:conclusion}
We have presented a formalism to extend the delay--spectrum, visibility--based power spectrum estimator to full polarization, including the effect of polarized foreground leakage due to widefield primary beams. We applied our formalism to simulate power spectra from PAPER--like observations. We first used a total intensity source catalogue, predicting polarized power spectra in general agreement with observations in \cite{Kohn2016}. We then modeled polarized (Galactic and extragalactic) foregrounds using recent low frequency observations and predicted the corresponding power spectrum leakage, particularly focusing on the contamination for a 30~m baseline.
We found that an ``EoR window" can be defined in terms of polarization leakage from diffuse Galactic foreground as its contamination falls quickly below $\sim 1$~(mK)$^2$ at $k > 0.3 \, h$~Mpc$^{-1}$, i.e. significantly below the fiducial range of 21-cm models. The existence of such EoR window is due to the fact that current observations find significant diffuse polarization only at low Faraday depths, i.e. $\phi \lesssim 12$~rad~m$^{-2}$ corresponding to $k_\parallel \lesssim 0.06 \, h$~Mpc$^{-1}$. Bright, diffuse emission found at high Faraday depth values would appear at proportionally higher $k_\parallel$ modes, narrowing (or jeopardizing) the EoR window. Current deep observations, however, set the presence of polarized diffuse emission to be below $\sim 0.1$~K at $\phi > 5$~rad~m$^{-2}$ \citep{Jelic2015}, supporting our model assumptions.
In the case of point source leakage, an EoR window cannot be identified because point sources show emission essentially at any $k_\parallel$ value due to their broad RM distribution, making polarized point sources a potentially more serious contamination than diffuse emission.
The magnitude of such leakage depends, however, significantly on the average point source polarization fraction for which only upper limits are currently available in the $100-200$~MHz range. By treating such upper limits as actual measurements, our model predicts a worse case scenario where point source polarization leakage is higher than the contamination due to Galactic emission at any $k$ mode for a 30~m baseline.
The comparison with polarized power spectra from \cite{Moore2015} constrains the observed (i.e. uncorrected for ionospheric Faraday depolarization) average polarization fraction at $\nu = 126$~MHz to be $\langle \gamma \rangle < 0.1\%$, leaving upper limits to the 21-cm leakage that are between one and two orders of magnitude greater than the 21-cm signal in the $7.7 < z < 10.3$ range.
Our current simulations do not include the depolarization effect due to ionospheric Faraday rotation average over multiple nights of observations, therefore all our predictions should be regarded as worst cases in terms of contamination to the 21-cm power spectrum.
Finally, our work provides a tool to predict the level of leakage expected in actual 21-cm observations by {\it forward modeling} the polarized foreground emission through the instrument model \citep[see][for relevant examples]{Pindor2011,Bernardi2011,Sullivan2012}: in the case of polarized point sources, for example, the observed average polarization fraction needs to be known in order to predict the leakage. We indicate three ways to determine the observed average polarization fraction:
\begin{itemize}
\item by best fitting the predicted polarized power spectra to the polarized power spectra measured in actual observations as we showed here with the \cite{Moore2015} data;
\item by imaging the polarized sky without correcting for ionospheric Faraday rotation: this directly provides a measurement of the average point source polarization fraction;
\item by applying an ionospheric Faraday rotation model to a polarized point source model realization whose average polarization fraction is provided by independent observations.
\end{itemize}
Although future predictions might require further corrections to this first order picture, our model offers a way to account for polarization leakage in 21-cm power spectrum observations to be applied to future observations with PAPER, the Hydrogen Epoch of Reionization Array \citep{DeBoer2016} and, potentially, the Square Kilometre Array \citep{Koopmans2015}.
\section*{Acknowledgements}
We thank an anonymous referee for helpful comments that improved the manuscript. The authors also thank Tobia Carozzi for his help with the FEKO simulations and Adrian Liu for comments on the manuscript. CDN is supported by the SKA SA scholarship program. GB acknowledges support from the Royal Society and the Newton Fund under grant NA150184. This work is based on the research supported in part by the National Research Foundation of South Africa (Grant Numbers 103424). This research was supported by NSF grant AST-1440343. This research is supported by the South African Research Chairs Initiative of the Department of Science and Technology and National Research Foundation.
\bibliographystyle{apj}
\nocite{*}
|
\section{Introduction}\label{sec:intro}
The clustering of galaxies is often used to constrain models of the background cosmology and galaxy formation. In many studies, the Halo Model \citep{Cooray2002} plays an important role. In the simplest (and most widely used) version of the approach, the clustering of galaxies is determined by a combination of how galaxies populate halos, and the clustering of the halos which host galaxies, and for both ingredients, halo mass is assumed to be the only halo property which matters. \citet{Abbas2007} describe one of the first tests of this assumption; they classified galaxies by the number of neighbours within $\sim 8h^{-1}$Mpc; measured the clustering signal as a function of environment; and showed that the environmental dependence of clustering was similar to that in a mock catalog in which the galaxy content of a halo was determined completely by halo mass and not environment.
However, they also showed that they were able to model the strength of the clustering signal as a function of environment alone. This finding has recently been confirmed by \citet{Pujol2017}. Provided that the environment is defined on a scale that is substantially larger than a typical halo, the clustering signal is a function of environment, and not of halo mass. I.e., at fixed environment, the clustering is independent of halo mass, whereas at fixed mass, the clustering is a strong function of environment. The main goal of the present paper is to provide a more careful derivation of the expression in \citet{Abbas2007}. A final section discusses how this particular clustering signal is related to what has come to be called Assembly Bias \citep{Sheth2004}, and makes the point that it is useful to distinguish between halo--environment correlations, and whether or not the way galaxies populate halos requires additional correlations.
\section{Bias of constrained regions}\label{Lag}
In what follows, it is important to distinguish clearly between the scale associated with halo formation, that on which the environment is defined, and the (typically much larger) scale on which the bias factor is measured. We will use $R_h$, $R_e$ and $R_0$ to denote these scales.
\subsection{Large scale environment as a constraint}
Suppose that we identify those positions in the initial (Gaussian) field which, when smoothed on scale $R_e$ have overdensity $\Delta_e$. Let $S_e\equiv\langle\Delta_e^2\rangle$. The probability of being centred on such a region is
\begin{equation}
p(\Delta_e) = \frac{{\rm exp}(-\Delta_e^2/2S_e)}{\sqrt{2\pi S_e}}.
\end{equation}
The conditional probability that the overdensity, when smoothed on some other scale $R_0$, is $\Delta_0$, given that it is $\Delta_e$ on scale $R_e$, is
\begin{equation}
p(\Delta_0|\Delta_e)
= \frac{{\rm exp}^{-(\Delta_0-\mu_{0|e})^2/2S_{0|e}}}{\sqrt{2\pi S_{0|e}}},
\end{equation}
where
\begin{equation}
\mu_{0|e} = \langle\Delta_0|\Delta_e\rangle
= \frac{\langle\Delta_0\Delta_e\rangle}{\langle\Delta_e^2\rangle}\,\Delta_e
\end{equation}
and
\begin{equation}
S_{0|e} \equiv S_0\,(1 - \langle\Delta_0\Delta_e\rangle^2/S_0S_e).
\end{equation}
Now, $\langle\Delta_0\Delta_e\rangle$ is the correlation between $\Delta$ on the two scales, whereas $\langle\Delta_0|\Delta_e\rangle$ is the cross correlation between the two $\Delta$s subject to the constraint that $\Delta=\Delta_e$ on scale $R_e$. Hence, it is natural to define
\begin{equation}
\langle\Delta_0|\Delta_e\rangle \equiv b_e\,\langle\Delta_0\Delta_e\rangle
\qquad{\rm where}\quad
b_e \equiv \Delta_e/S_e.
\label{bKaiser}
\end{equation}
The expression above shows that we should think of the constrained cross-correlation as biasing the unconstrained correlation; the bias is linearly proportional to the constraint. At the risk of belaboring the point,
\begin{align}
\langle\Delta_0\Delta_e\rangle
&= \int {\rm d}\Delta_e\,\int {\rm d}\Delta_0\,p(\Delta_e)\,p(\Delta_0|\Delta_e) \Delta_0\Delta_e \nonumber\\
&= \int {\rm d}\Delta_e\,p(\Delta_e)\,\Delta_e \langle\Delta_0|\Delta_e\rangle \nonumber\\
&= \int {\rm d}\Delta_e\,p(\Delta_e)\,\Delta_e \,b_e \langle\Delta_0\Delta_e\rangle \nonumber\\
&= \langle\Delta_0\Delta_e\rangle \int {\rm d}\Delta_e\,p(\Delta_e)\,\frac{\Delta_e^2}{S_e} ,
\end{align}
where the final integral equals unity. The second expression shows that the unconstrained correlation $\langle\Delta_0\Delta_e\rangle$ is a weighted sum over the constrained cross-correlations $\langle\Delta_0|\Delta_e\rangle$.
Equation~(\ref{bKaiser}) for the bias is familiar in cosmology from \citet{Kaiser1984}. However, there it was introduced in the context of the bias associated with regions which exceed a high threshold in a Gaussian field, though it is often referred to as the `high peak' limit. The expression above shows that this bias expression is actually associated with a much simpler constraint than either thresholds or peaks: simply that the height {\em equals} a certain value. The correspondence with peaks is a consequence of the fact that if $\Delta_e\gg \sqrt{S_e}$, then the additional constraints which define a peak do not matter for the bias (because the highest positions in the field are almost certainly also local peaks). While this reason was clear in early work, some more recent papers -- arguing that equation~(\ref{bKaiser}) is particular to peaks -- have got the logic backwards.
\subsection{Small scale overdensity as an additional constraint}\label{sharpk}
The analysis of the previous section shows what one should expect if the constraints are more complicated. E.g., if we add a constraint on a third scale $R_h$, then
\begin{align}
\langle\Delta_0|\Delta_e,\Delta_h\rangle
&= \langle\Delta_0|\Delta_e\rangle + \langle\Delta_0|\Delta_{h|e}\rangle
\label{orth}
\end{align}
where
\begin{equation}
\Delta_{h|e}\equiv \Delta_h - \langle\Delta_h|\Delta_e\rangle.
\end{equation}
Notice that, as there are now two constraints, the bias is the sum of two terms; the form of the expression above suggests that we should think of the prefactors of the correlations with $\Delta_e$ and $\Delta_{h|e}$ as being two bias factors. However, $\Delta_{h|e}$ involves both $\Delta_h$ and $\Delta_e$, whereas we are typically interested in keeping the effects of these two terms separate. I.e., we seek the coefficients of the terms proportional to $\langle\Delta_0\Delta_e\rangle$ and $\langle\Delta_0\Delta_h\rangle$, respectively. If we define $\nu_e\equiv \Delta_e/\sqrt{S_e}$, and similarly for $\nu_h$, then a little algebra shows that
\begin{equation}
\langle\Delta_0|\Delta_e,\Delta_h\rangle =
b_e\, \langle\Delta_0\Delta_e\rangle + b_h\,\langle\Delta_0\Delta_h\rangle
\end{equation}
where
\begin{align}
b_e &= \frac{\nu_e - \langle\nu_e\nu_h\rangle\, \nu_h}{\sqrt{S_e}(1 - \langle\nu_e\nu_h\rangle^2)} \\
b_h &= \frac{\nu_h - \langle\nu_h\nu_e\rangle\, \nu_e}{\sqrt{S_h}(1 - \langle\nu_e\nu_h\rangle^2)}
.
\end{align}
Note that $b_e\to \Delta_e/S_e$ only when $\langle\nu_e\nu_h\rangle\to 0$ and, in this limit, $b_h\to \Delta_h/S_h$ as well. I.e., when $\langle\nu_e\nu_h\rangle\to 0$, both $b_e$ and $b_h$ have the form of equation~(\ref{bKaiser}) in their respective variables. Typically, this will happen when $R_e\gg R_h$. The approach to zero will be faster if $R_e$ and $R_h$ are centred on different positions. Furthermore, if $R_0\gg R_e$ and $R_e\gg R_h$, then we expect $\langle\Delta_0\Delta_h\rangle\ll \langle\Delta_0\Delta_e\rangle$. In this limit $\langle\Delta_0|\Delta_e,\Delta_h\rangle \to \langle\Delta_0|\Delta_e\rangle$: the constraint on $\Delta_h$ is irrelevant.
In the present context, equation~(\ref{orth}) is the more transparent expression because it shows that the constraint on $\Delta_h$ will be irrelevant if $\langle\Delta_0|\Delta_{h|e}\rangle=0$. This happens if $\langle\Delta_0\Delta_h\rangle = \langle\Delta_0\Delta_e\rangle\langle\Delta_e\Delta_h\rangle/S_e$; i.e., if the $\Delta_0$-$\Delta_h$ correlation is entirely a consequence of the $\Delta_0$-$\Delta_e$ and $\Delta_e$-$\Delta_h$ correlations. This holds true for the special case when the smoothing filter used to define $\Delta$ on the different scales is sharp in $k$-space. Such a filter was used extensively in the past, as it leads to Markovian walks with uncorrelated steps, which renders many questions of interest analytically tractable \citep{Bond1991}. For this filter, $\langle\Delta_r\Delta_R\rangle = \langle\Delta_R^2\rangle$ where $R\ge r$. Hence, for this filter $\langle\Delta_0\Delta_h\rangle = S_0$ and $\langle\Delta_0\Delta_e\rangle\langle\Delta_e\Delta_h\rangle/S_e = S_0S_e/S_e = S_0$. So, if $\Delta_e$ is fixed, then the constraint from $\Delta_h$ is {\em completely} irrelevant.
More generally, the constraint on $\Delta_h$ will be irrelevant if $\langle\Delta_0|\Delta_{h|e}\rangle\ll \langle\Delta_0|\Delta_e\rangle$, i.e., if the amount of correlation between $\Delta_0$ and $\Delta_h$ which is not due the $\Delta_0$-$\Delta_e$ and $\Delta_e$-$\Delta_h$ correlations is smaller than the $\Delta_0$-$\Delta_e$ correlation. When $R_0\gg R_e\gg R_h$, this is very likely to be the case. Hence, except when $|\Delta_{h|e}|$ is very large the fact that $\Delta_h$ is constrained will not matter; the cross correlation $\langle\Delta_0|\Delta_e,\Delta_h\rangle$ will be dominated by the first term on the rhs of equation~(\ref{orth}). Recalling that our choice of subscripts is not accidental, this discussion implies that when $R_e\gg R_h$ then the cross correlation signal of equation~(\ref{orth}) will be dominated by the correlation with the environment; the halo mass is almost always irrelevant. Halo mass only matters if $|\Delta_{h|e}|$ is large: since $\Delta_h$ is typically of order unity, halo mass matters more if $\Delta_e$ is very negative (i.e. in underdense regions).
\subsection{Small scale overdensity and its derivatives as additional constraints}
Equation~(\ref{orth}) serves mainly to illustrate how cross-correlations with the large scale environment generalize as one adds more constraints. Following \citet{Musso2012}, we are most interested in the case in which the derivatives of $\Delta_h$ also matter. In this case,
\begin{align}
\langle\Delta_0|\Delta_e,\Delta_h,\Delta'_h\rangle
&= \langle\Delta_0|\Delta_e\rangle + \langle\Delta_0|\Delta_{h|e}\rangle
+ \langle\Delta_0|\Delta_{h'|he}\rangle
\end{align}
and we are again faced with the problem of showing when the first term on the right hand side dominates.
In this context, it is interesting to consider a slightly more general problem in which the derivative on scale $R_e$ is also specified. Then we are interested in
$\langle\Delta_0|\Delta_e,\Delta'_e,\Delta_h,\Delta'_h\rangle$.
\citet{Musso2014} describe a family of -- what they call Markov Velocity -- models in which correlations between scales are rather similar to those in $\Lambda$CDM models. They show that, for Markov Velocity models,
\begin{equation}
\langle\Delta_0|\Delta_e,\Delta'_e,\Delta_h,\Delta'_h\rangle
= \langle\Delta_0|\Delta_e,\Delta'_e\rangle ;
\end{equation}
i.e., if both $\Delta_e$ and $\Delta'_e$ are specified, then the smaller scale $R_h$ is irrelevant (see their equation~71). For Markov Velocity models this is an exact, not an approximate, statement. As a result, $\langle\Delta_0|\Delta_e,\Delta_h,\Delta'_h\rangle$ only depends weakly on $R_h$, or depends on $R_h$ only for a rather restricted range of scales.
The similarity of these models to $\Lambda$CDM strongly suggests that $\langle\Delta_0|\Delta_e,\Delta_h,\Delta'_h\rangle$ in $\Lambda$CDM models will also only depend weakly on $R_h$. I.e., if the environment on scale $R_e\ge R_h$ is fixed, then the large scale bias is approximately independent of halo mass.
\subsection{General formulation}
The lesson from the previous explicit models is clear. If the vector ${\bmath h}$ includes all the variables which are important for halo formation, then one should express halos as constraints on these variables in the underlying Gaussian field:
\begin{equation}
n(m) = \int {\rm d}{\bmath h}\,p({\bmath h})\,{\cal C}_m({\bmath h}),
\end{equation}
where ${\cal C}_m({\bmath h})$ specifies the set of constraints on ${\bmath h}$ which must be satisfied to form a halo of mass $m$. Then
\begin{equation}
n(m|\Delta_e) = \int {\rm d}{\bmath h}\,p({\bmath h}|\Delta_e)\,
{\cal C}_m({\bmath h})
\label{nme}
\end{equation}
and
\begin{equation}
\langle\Delta_0|\Delta_e,m\rangle =
\frac{\int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\,{\cal C}_m({\bmath h})\,\langle\Delta_0|\Delta_e,{\bmath h}\rangle}{\int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\,{\cal C}_m({\bmath h})}.
\label{general}
\end{equation}
\subsection{Reconstructing the dependence on mass and environment}
\label{sec:brec}
Fixing $m$ and marginalizing over all $\Delta_e$ yields
\begin{align}
& \frac{\int {\rm d}\Delta_e \int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\, {\cal C}_m({\bmath h})\,\langle\Delta_0|\Delta_e,{\bmath h}\rangle}{\int {\rm d}\Delta_e \int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\,{\cal C}_m({\bmath h})} \nonumber\\
&= \frac{\int {\rm d}{\bmath h}\,p({\bmath h})\,{\cal C}_m({\bmath h})\,\langle\Delta_0|{\bmath h}\rangle}{\int {\rm d}{\bmath h}\,p({\bmath h})\,{\cal C}_m({\bmath h})\int {\rm d}\Delta_e\,p(\Delta_e|{\bmath h})} \nonumber\\
& \quad + \ \frac{\int {\rm d}{\bmath h}\,p({\bmath h})\,{\cal C}_m({\bmath h})\,\int {\rm d}\Delta_e \, p(\Delta_e|{\bmath h})\,\langle\Delta_0|\Delta_{e|{\bmath h}}\rangle}{n(m)} \nonumber\\
&= \frac{\int {\rm d}{\bmath h}\,p({\bmath h})\,{\cal C}_m({\bmath h})\,\langle\Delta_0|{\bmath h}\rangle}{n(m)}.
\label{brecm}
\end{align}
The ratio of the final expression to $\langle\Delta_0\Delta_h\rangle$ is what is usually meant by $b_h(m)$.
On the other hand, marginalizing over all halo masses at fixed environment yields
\begin{align}
\label{brecd}
& \frac{\int {\rm d}m \int {\rm d}{\bmath h}\,p({\bmath h}|\Delta_e)\, {\cal C}_m({\bmath h})\,\langle\Delta_0|\Delta_e,{\bmath h}\rangle}{\int {\rm d}m \int {\rm d}{\bmath h}\,p({\bmath h}|\Delta_e)\,{\cal C}_m({\bmath h})} \\
&= \langle\Delta_0|\Delta_e\rangle + \frac{\int {\rm d}m \int {\rm d}{\bmath h}\,p({\bmath h}|\Delta_e)\, {\cal C}_m({\bmath h})\,\langle\Delta_0|\Delta_{{\bmath h}|e}\rangle}{\int {\rm d}m\,n(m|\Delta_e)} \nonumber.
\end{align}
If ${\bmath h}$ involves $\Delta_h$ (where, typically $R_h^3\propto m$) and its derivatives, then, as we have already discussed, we expect the expression above to be dominated by the first term on the right hand side, especially when $R_h\ll R_e\ll R_0$.
\citet{Pujol2017} show that if one attempts to reconstruct how bias depends on $\Delta_e$ using
\begin{equation*}
\frac{\int {\rm d}m \int {\rm d}{\bmath h}\,p({\bmath h}|\Delta_e)\, {\cal C}_m({\bmath h})\,\langle\Delta_0|{\bmath h}\rangle}{\int {\rm d}m\,n(m|\Delta_e)}
\end{equation*}
(their equation~5) then one gets the wrong answer: almost no predicted dependence of the bias on $\Delta_e$ when the measurements show a strong trend. Comparison with our equation~(\ref{brecd}) shows why; by assuming that halo bias depends only on halo mass, their expression misses the contribution which leads to the first term on the right hand side of our expression -- the term which dominates the answer when $R_m\ll R_e\ll R_0$.
On the other hand, \citet{Pujol2017} found that
\begin{equation*}
\frac{\int {\rm d}\Delta_e \int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\, {\cal C}_m({\bmath h})\,\langle\Delta_0|\Delta_e\rangle}{\int {\rm d}\Delta_e\,n(\Delta_e|m)}
\end{equation*}
(their equation~6) was able to reconstruct the mass dependence of bias rather well. Our analysis shows why this works, even though it too is, formally, incorrect. (The correct expression is our equation~\ref{brecm}.) Namely, the expression above can be written as
\begin{align}
& \frac{\int {\rm d}{\bmath h}\,p({\bmath h})\, {\cal C}_m({\bmath h}) \int {\rm d}\Delta_e \,p(\Delta_e|{\bmath h})\,\langle\Delta_0|\Delta_e\rangle}{n(m)} \nonumber\\
&= \frac{\int {\rm d}{\bmath h}\,p({\bmath h})\, {\cal C}_m({\bmath h}) \,\langle\Delta_0\Delta_e\rangle \langle\Delta_e|{\bmath h}\rangle/S_e}{n(m)} ,
\end{align}
and our discussion of equation~(\ref{orth}) showed that we expect
\begin{equation}
\langle\Delta_0\Delta_e\rangle \langle\Delta_e|{\bmath h}\rangle/S_e \approx
\langle\Delta_0|{\bmath h}\rangle .
\end{equation}
If this approximation were an equality, then their expression would reduce to the correct one, our equation~(\ref{brecm}). That it is only an approximation is why \citet{Pujol2017} only found good, but not perfect agreement with the actual mass dependence of bias, $b_h(m)$.
\section{Evolution}\label{Eul}
The analysis of the previous section was for statistics in the initial conditions, sometimes called Lagrangian space. Since the analysis in \citet{Pujol2017} was for halos and environments defined in the evolved Eulerian space, our assertions in Section~\ref{Lag} are not completely justified until we have shown that they survive nonlinear evolution.
\subsection{Excursion set approach: Analytic}
We use the excursion set approach of \citet{Sheth1998} to model statistics in the evolved Eulerian space. This approach makes use of the spherical evolution mapping between $\delta_{\rm V}$, the Eulerian density on scale $V$, and $\Delta_{\rm M}$, the Lagrangian density on scale $M$:
\begin{equation}
1 + \delta_{\rm V} \equiv M/\bar\rho V
= (1 - \Delta_{\rm M}/\delta_c)^{-\delta_c} ,
\label{scapprox}
\end{equation}
where $V$ is the Eulerian volume, $M$ is the mass in it, and $\delta_c\approx 1.686$ (although $\delta_c = 21/13$ reproduces the monopole of second order perturbation theory). In what follows, we will also make use of the fact that
\begin{equation}
\frac{\Delta_{\rm M}}{\delta_c} = 1 - \left(\frac{\bar\rho V}{M}\right)^{1/\delta_c}
= 1 - (1+\delta_{\rm V})^{-1/\delta_c},
\label{scbarrier}
\end{equation}
which follows from rearranging equation~(\ref{scapprox}).
The gist of the argument is that Eulerian statistics on scale $V$ are related to Lagrangian statistics on scale $M$. While this idea can be traced back to \citet{Bernardeau1994}, the analysis in \citet{Sheth1998} allows one to work down to substantially smaller $V$. \citet{Lam2008} show that it provides a rather good model of what we call the probability distribution of the Eulerian environment here. Our goal is to show that this approach also provides a simple description of the joint distribution of halos and their environment -- i.e. of Eulerian bias -- a point which was made in \citet{Sheth1998}, but has not been followed-up since. This turns out to be straightforward, particularly because of recent advances in our understanding of the excursion set approach \citep{Musso2012}.
In what follows, $\delta$ and $\Delta$ always denote Eulerian and Lagrangian overdensities, and their subscripts always denote the corresponding Eulerian or Lagrangian smoothing scale. E.g., if $\delta_0$ is the Eulerian density on scale $V_0$, then $\Delta_0$ is the Lagrangian density on scale $M_0/\bar\rho = V_0\,(1 + \delta_0)$, and $\delta_0$ and $\Delta_0$ are related by equation~(\ref{scbarrier}).
Our goal is to estimate the mean Eulerian density on scale $V_0$ given that the Eulerian cell is centred on a region with Eulerian density $\delta_e$ on scale $V_e$ which itself is centred on a halo of mass $m$. The Lagrangian version of this quantity is equation~(\ref{general}). It becomes
\begin{equation}
\langle\delta_0|\delta_e,m\rangle =
\frac{\int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\,{\cal C}_m({\bmath h})\,\langle\delta_0|\Delta_e,{\bmath h}\rangle}{\int {\rm d}{\bmath h}\,p({\bmath h},\Delta_e)\,{\cal C}_m({\bmath h})},
\end{equation}
where we have used the fact that the $m$ and $\delta_e$ constraints correspond to simple constraints in Lagrangian space. The main problem is to estimate $\langle\delta_0|\Delta_e,\bmath{h}\rangle$.
On large Eulerian scales $V_0$ we expect $\delta_0\ll 1$, and hence $\Delta_0\approx \delta_0$ almost surely. In this limit, we expect to be able to use the Gaussian expression (equation~\ref{orth}):
\begin{equation}
\langle\delta_0|\delta_e,m\rangle
\approx \langle\Delta_0|\Delta_e\rangle + \langle\Delta_0|\Delta_{h|e}\rangle
\label{2terms}
\end{equation}
where $\langle\Delta_0|\Delta_e\rangle = \Delta_e\,\langle\Delta_0\Delta_e\rangle/\langle\Delta_e^2\rangle$ dominates. This would make
\begin{equation}
b_{\rm e}^{\rm E} = \frac{\Delta_e}{\langle\Delta_e^2\rangle}
= \frac{\delta_c[1 - (1+\delta_e)^{-1/\delta_c}]}{S[\bar\rho V_e (1+\delta_e)]},
\label{kaiserE}
\end{equation}
where $S$ is the the Lagrangian variance on the mass scale $\bar\rho V_e (1+\delta_e)$. Equation~(\ref{kaiserE}) is the expression in \citet{Abbas2007}. Comparison with equation~(\ref{bKaiser}) shows explicitly that, in this limit, the Eulerian bias is like the Lagrangian one provided that one correctly rescales the density {\em and} volume. Figure~3 of \citet{Pujol2017} shows that this simple expression works remarkably well over a wide range of scales.
Before moving on, we note that this expression has been rediscovered by \citet{Uhlemann2017} who appear to be unaware of earlier work. Moreover, as we have spelled out in more detail here, the excursion set approach of \citet{Sheth1998} shows why, even though equation~(\ref{kaiserE}) is quite accurate, it is just an approximation. That is to say, it shows clearly how to go beyond rescaled Kaiser-bias. For example, the top panel of Figure~6 in \citet{Pujol2017} shows the comoving number density of halos in cells of specified overdensity $\delta_e$. They do not remark on it, but this quantity has long been known to be well-approximated by equation~(\ref{nme}), with $\Delta_e$ given by $\delta_c$ times the rhs of equation~(\ref{scbarrier}) when $\delta_e=\delta_V$ in the rhs of equation~(\ref{scbarrier}) \citep{Mo1996, Sheth2002}. This is a limit which our approach is designed to reproduce \cite{Sheth1998}.
\begin{figure}
\centering
\includegraphics[width=1.\hsize]{biasMfixE.eps}
\caption{The Eulerian bias of halos surrounded by large overdensities is larger; however, at fixed overdensity, bias is the same for all except the most massive halos. The thickness of line is proportional to the overdensity value, with the thickest line corresponding to the densest field. Red and magenta lines are for halos in the densest 10\% and the next densest 20\% of the cells in the evolved Eulerian field; blue and green lines show the bias of halos in the least dense 10\% and the next emptiest 20\%. The apparent upper limit in $\nu$, which increases with density, is because massive halos are not present in the least dense cells.}
\label{bMfixE}
\end{figure}
\subsection{Excursion set approach: Monte-Carlo}
We have checked the analysis above explicitly in Monte-Carlo realizations of this process. Namely, we generated $10^5$ random walks, each having a correlation structure appropriate for tophat smoothing of Gaussian field having $P(k)\propto k^{-2}$. We used the algorithm described in \citet{Musso2014} to do this. For each walk, we stored the mass scale on which it first crossed a `constant barrier' of height $\delta_c$, and the mass scale on which it crossed the `moving barrier' of equation~(\ref{scbarrier}), for a range of choices of Eulerian $V$. First crossing of $\delta_c$ is a simple proxy for a halo; by storing first crossings for a range of $V$, we can map out the Eulerian profile around each `halo' (see \citealt{Sheth1998}; indeed, viewed this way, a halo is just the special case in which $V=0$). In addition, we stored the height of the walk on a number of mass scales, which we use to reconstruct Lagrangian profiles of halos or of Eulerian cells.
\begin{figure}
\centering
\includegraphics[width=1.\hsize]{biasEfixM.eps}
\caption{Denser Eulerian cells are more biased, but this bias is independent of the mass of the halo at the cell centre. Cyan dotted and magenta solid lines show results for cells centred on the 10\% lowest and highest mass halos. Thick smooth curve shows $b_e^{\rm E}$ of equation~(\ref{kaiserE}). }
\label{bEfixM}
\end{figure}
Red, magenta, green and blue curves in Figure~\ref{bMfixE} show the Eulerian bias of halos which are centred on patches having Eulerian densities $\langle 1+\delta_{\rm V}\rangle = (7.5, 2.3, 0.5, 0.3)$. The scale $V$ is such that, when smoothed on scale containing mass $M=\bar\rho V$, the rms linear theory overdensity had variance $\langle\Delta_{\rm V}^2\rangle =0.5^2$. (Therefore, a halo of mass $\bar\rho V$ would have $\nu = \delta_c/\sigma = 42/13$. This is why the curves for underdense regions do not extend to larger $\nu$.) Clearly, the Eulerian bias is larger for the halos centred on denser cells; however, except for the densest cells, the bias is the same for all halo masses. I.e., the bias is determined by the environment, and not by halo mass.
Figure~\ref{bEfixM} shows another way of presenting this trend: cyan and magenta curves show how the bias depends on environment for the least and most massive halos. Clearly, the bias is the same strong function of environment whatever the mass of the halo at the centre. (For the moment, we are ignoring the slight tendency for the cells centred on the most massive halos to have slightly smaller bias factors.) The smooth curve shows equation~(\ref{kaiserE}); it provides a good description of the measurements. This shows explicitly that the first term in equation~(\ref{2terms}) really does capture most of the environmental effect. The second term in equation~(\ref{2terms}) must account for the small trend with mass which remains, but note that this is much smaller than the overall trend with environment.
\begin{figure}
\centering
\includegraphics[width=0.9\hsize]{xipmL}
\caption{Initial Lagrangian density profiles around patches which become Eulerian cells with the specified density: red and magenta are for the densest 10\% and the next densest 20\% of the cells in the evolved Eulerian field; blue and green show the least dense 10\% and the next emptiest 20\%. Symbols with error bars show the mean and the rms around the mean -- errors on the mean are smaller than the symbols. Curves show the Lagrangian-space cross correlation; there are no free parameters in this comparison. }
\label{xipmL}
\end{figure}
The trends in Figures~\ref{bMfixE} and~\ref{bEfixM} are remarkably similar to those shown in Figure~4 of \citet{Pujol2017}. Even the slight tendency for environments centred on massive halos to be slightly less biased (Figure~\ref{bEfixM}) is similar. This suggests that our Monte Carlos have captured the essence of the effect.
To show that we really do understand the origin of this effect, on small scales as well, Figure~\ref{xipmL} shows the initial Lagrangian density profiles around the patches which evolve into the densest and least dense cells. Symbols with error bars show the mean and the rms around the mean -- errors on the mean are smaller than the symbols. Curves show the Lagrangian-space cross correlation -- essentially equation~(\ref{orth}); there are no free parameters in this comparison. The agreement justifies the assertions we made in Section~\ref{sec:brec}. Namely, accounting for the joint distribution of mass and environment is straightforward. Doing so shows that, just as in Lagrangian space, the Eulerian bias is also determined primarily by the larger scale environment, and much less so by halo mass. Therefore, analyses which ignore the environmental effect will lead to incorrect conclusions about the nature of halo bias.
\section{Discussion and conclusions}
We discussed how the large scale bias of halos depends on both halo mass and environment, in Lagrangian (Section~\ref{Lag}) and Eulerian (Section~\ref{Eul}) space. We showed that, at fixed environment, the dependence of large scale bias on halo mass should be weak (Figure~\ref{bMfixE}). Indeed, if one defines halos and environment using a filter which is sharp in $k$-space \citep{Bond1991}, and one conditions on large scale environment, then large-scale bias is predicted to be {\em completely} independent of halo mass (see discussion in Section~\ref{sharpk}). Our calculation quantifies the small residual effect which comes from the fact that correlations between scales are more complicated than for sharp-$k$ smoothing.
This has an interesting implication. Following \citet{Sheth2004}, there have been many studies of the dependence of bias on other parameters, if halo mass is held fixed. These are usually called `Assembly Bias' studies, even though the additional parameters may not be explicitly related to halo assembly. The underlying origin of all these signals is nontrivial correlations between scales. Our analysis shows that if bias at fixed environment does show some dependence (presumably weak!) on halo mass, then one has detected the effect of nontrivial correlations between scales. In this sense, one has detected `Assembly Bias' coming from the other way round from what is currently fashionable. To see this explicitly, suppose we order scales as
$$R_{\rm bias} \ge R_{\rm env} \ge R_{\rm halo} \ge R_{1/2}$$
where $R_{1/2}$ (for `half-mass') is a crude proxy for halo assembly. The usual studies fix $R_{\rm halo}$ and look for additional correlation between the more widely separated scales $R_{\rm bias}$ and $R_{1/2}$. But the ranking of scales shows that one could have fixed $R_{\rm env}$ and looked for additional correlation between $R_{\rm bias}$ and $R_{\rm halo}$ (or $R_{1/2}$). This is the sense in which looking for mass dependence at fixed environment is the same as assembly bias.
The analysis of the previous section is particularly relevant to the question of whether or not galaxy properties depend on quantities other than halo mass. The main text shows why, when the environment is constrained, then halo bias is a function of both halo mass and environment. However, this does {\em not} mean that the Halo Occupation Distribution of how galaxies populate halos must also depend on both (it may, but it need not). Indeed, \citet{Abbas2007} showed that mock galaxy catalogs, in which mass is the only variable which determines how galaxies populate halos, automatically exhibit a number of environmental trends that are seen in the data. That is to say, they showed that the data they examined do {\em not} require any additional galaxy-environment effect: the halo-environment correlation which comes for free, and which we have spelled out in some detail in this paper, is sufficient to explain the galaxy-environment correlations.
While this may be true for observables such as luminosity, which are expected to be monotonically related to halo mass, the same may not be true for colors, for which the correlation with halo mass is not as simple. A simple model for galaxy colors, in which galaxy-environment correlations are inherited from the halo-environment correlations, is able to provide a reasonable description of the bright SDSS galaxies considered by Abbas \& Sheth \citep{Skibba2009}. However, it is too simplistic to account for all observed correlations \cite[see][for the current state of the art]{Pahwa2017}, and studies at the faint end have yet to be done. \citet{Pujol2017} show that, in the semi-analytic galaxy formation model they considered, galaxy color appears to correlate more with density than halo mass, and that density appears to be more important than halo mass for faint red central galaxies. The importance of environment over halo mass appears at lower luminosities than Abbas \& Sheth considered in the SDSS. At these lower luminosities, the scatter between halo mass and luminosity becomes larger, so it will be interesting to see if the color-dependent trends in \citet{Pujol2017}'s Figure~8 are reproduced in data. With such studies in mind, we express their results in our formalism in Appendix~\ref{pujol}.
An interesting extension of our work would be to study what happens if the environment of a halo is defined using a measure which does not correlate with the density. E.g., \citet{Paranjape2017} use a measure which is built from the tidal shear. They show that, at fixed mass, halo bias correlates strongly with the morphology of the environment (e.g., `filamentary' versus `isotropic'), {\em and} that bias is also a strong function of mass when the environment is fixed. This `assembly bias' effect appears to be richer than the one with density which we studied here. Again, however, galaxies will inherit the environmental correlations of their host halos, so care must be taken to isolate correlations with environment which are over and above those which come `for free' from the host halo-environment correlation.
Finally, the careful reader will have noticed that our least dense cells have $1+\delta_e\sim 0.2$ (e.g. Figure~\ref{bMfixE}); such cells would be classified as `voids' \citep{ShethWeygaert2004}. These `voids' have Eulerian bias factors which are less than zero (Figure~\ref{bMfixE}), and the associated Lagrangian profiles of these cells are indeed rather underdense, especially on small scales (Figure~\ref{xipmL}). Clearly, then, the excursion set approach allows us to model the evolution of void profiles; this is done in Massara \& Sheth (2017, in preparation).
In addition, study of the redshift-space clustering in the $b=0$ subsample (Figures~\ref{xi3dx} and~\ref{xi3db} suggest this is true of the 30\% underdense sample) may allow simple constraints on the growth rate $f = {\rm d}\ln D/{\rm d}\ln a$, from a comparison of the (projected) real and redshift space clustering signals. Furthermore, subsamples selected using our methodology have a rather wide range of bias factors, making them well-suited for multi-tracer constraints on redshift space distortions and primordial non-Gaussianity \citep{McDonald2009}, and for measuring the gravitational redshift effect from large scale structures \citep{Zhu2017}.
\section*{Acknowledgements}
We thank the ICTP for its hospitality during the summers of 2015, 2016 and 2017, Nordita for its hospitality during the program `Advances in Theoretical Cosmology in Light of Data' in July 2017, and A. Pujol, E. Gazta\~naga and A. Paranjape for helpful discussions.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{zero}
Fourier's law, which relates the macroscopic heat flux to the
temperature gradient in a solid system, was introduced almost two
centuries ago. Nevertheless, its understanding from microscopic basis
is still an important open issue of the non-equilibrium statistical
mechanics~\cite{LLP03}. In particular, among the several theoretical
studies on this subject, important results have been derived mainly
for $1d$ systems. The prototype model is a chain of masses and non
linear springs (Fermi-Pasta-Ulam-like systems), whose ends interacts
with thermal baths at different
temperatures~\cite{GHN01,D01,LLP03}. Other investigated systems are
constituted of $1d$ lattices~\cite{LLP98,GLPV00}, chains of cells,
with an energy storage device, which exchange energy through tracer
particles~\cite{EY04,EY05}, spinning disks~\cite{SLL14},
systems with local thermalization mechanism~\cite{GG08,GG08b}, or
a chain of anharmonic oscillators with local energy conserving
noise~\cite{OS13}. Despite the apparent simplicity of the problem,
and of the considered models, both the analytical approaches and the
numerical studies are rather difficult in this context.
The main aim of the present paper is the study of Fourier's law using
a mechanical model, which is rather crude (but still non-trivial),
allowing for an approach in terms of kinetic theory. More
specifically, we consider a generalized piston model, made of a
certain number of cells, each containing a non-interacting particle
gas. The walls of the cells are mobile, massive objects, that interact
with the particles via elastic collisions. The system at its ends
interacts with thermal baths at fixed temperatures. It is easy to
realize the analogy between such a generalized piston model and the
systems of masses and springs: the pistons and the gas compartments
play the role of masses and springs, respectively.
Our model is an example of partitioning system (as the adiabatic
piston), where previous studies showed that the presence of mobile
walls can induce interesting
behaviours~\cite{CPS96,GP99,GPL03,GL05,BKM04,CPPV07,DCPS11,SGP13,GCSVV,CPV15}.
Basically, in the study of partitioning systems, one can adopt two
approaches: in terms of a Boltzmann equation~\cite{GPL03} or introducing effective
equations (Langevin-like) for suitable observables derived \`a la
Smoluchowski, i.e., from an analysis of the collisions
particles/walls. In our study, we will adopt the latter method.
The paper is organized as follows. Section 1 is devoted to the
introduction of the model; in Section~2, we show how a thermodynamic
approach is not sufficient to determine the values of macroscopic
variables in the steady state. In Section 3, we present a
coarse-grained Smoluchowski-like description of the system, which
provides a good prediction for the main quantities of interest. \mbox{In
Section 4}, we compare the theoretical results with numerical
simulations and discuss the limit of validity of the proposed
approach. In Section 5, some general conclusions on Fourier's Law
in mechanical systems are drawn. In Appendix A, we report the details
of the analytical computations presented in Section 3.
\section{A Generalized Piston Model}
\label{one}
We consider a box of length $L$, partitioned by $n$ mobile
adiabatic walls (also called pistons in the following), with mass $M$,
average positions $z_j$ and velocities $V_j$ ($j=1, 2, \dots n$). The walls
move without friction along the horizontal axis (see Figure~\ref{fig0}). The
external walls are kept fixed in the positions $z_0=0$ and
$z_{n+1}=L$. Each of the $n+1$ compartments separated by the walls
contains $N$ non-interacting point-like particles, with mass $m$,
positions $x_i$ and velocities $v_i$. The particles interact with the
pistons via elastic collisions:
\begin{equation}
\label{eq:urti_elastici}
\begin{aligned}
V' &= V + \frac{2\,m}{m+M} \, (v_i - V), \\
v'_i &= v_i - \frac{2\,M}{m+M} \, (v_i - V),
\end{aligned}
\end{equation}
where primes denote post-collisional velocities.
\begin{figure}[ht!]
\centering
\includegraphics[width=.5\columnwidth,clip=true]{piston-eps-converted-to.pdf}
\caption{Sketch of the system.}
\label{fig0}
\end{figure}
The two external walls in $z=0$ and $z=L$ act as thermal baths at temperatures
$T_A$ and $T_B$. The interaction of the thermostats with the particles is
the following: when a particle collides with the wall, it is
reinjected into the system with a new velocity drawn from the
probability distribution~\cite{TTKB1998}
\begin{equation}
\label{eq:distribuzione_particelle_termostato}
\rho_{A, B}(v') = \frac{m}{k_B T_{A, B}} |v'| \,e^{-\frac{mv'^2}{2k_B T_{A, B}}} \,\Theta(\pm v'),
\end{equation}
with $+$ for the case $A$ and $-$ for $B$, and $k_B$ is the Boltzmann constant, $\Theta(x)$ being the Heaviside step~function:
\begin{equation}
\Theta(x)=
\begin{cases}
1, & \quad\text{if} \quad x>0,\\
0, & \quad\text{if} \quad x\leq 0.
\end{cases}
\end{equation}
For the following discussion, we define the temperature of the $i-$th piston as
\begin{equation}
T_i^{(p)}=M\langle V_i^2\rangle,
\end{equation}
and the temperature of the particle gas in the compartment $j$
as an average on the particles between the $(j-1)$-th and the $j$-th piston
\begin{equation}
T_j=m \frac{1}{N}\sum_{i\in (j-1)N}^{jN}\langle v_i^2\rangle.
\end{equation}
\section{Simple Thermodynamic Considerations}
\label{two}
We expect, and this is fully in agreement with the numerical
computations, that, given a generic initial condition, after a certain
transient, the system reaches a stationary state. The positions of the
walls fluctuate around their mean values $z_j$, in a similar way to
$T_i^{(p)}$ and $T_i$. The first non-trivial problem of the non
equilibrium statistical mechanics is to determine $z_j$, $T_i^{(p)}$
and $T_i$ as function of the parameters of the model, i.e., $n, L, m,
M, T_A$ and $T_B$.
We first analyze the simplest case of a single piston, where
thermodynamics is sufficient to provide a complete description of the
stationary state. Then, we consider the more general case of a multiple
piston; in such a situation, thermodynamic relations are not enough to
univocally determine the steady state: it is necessary to adopt a
statistical mechanics approach. Such an approach will
rely on three main assumptions, discussed in more detail below:
small mass ratio $m/M\ll 1$, local thermodynamic equilibrium in each
compartment, and independence of the collisions particles/pistons.
\subsection{Single Piston}
If $n=1$, using the equation for the perfect gas in each compartment,
we immediately get the~equations
\begin{equation}
\label{eq:gas_perfetti}
\begin{cases}
p z_1 = N k_B T_A,\\
p (L-z_1) = N k_B T_B,
\end{cases}
\end{equation}
where $p$ is the pressure, yielding
\begin{equation}
\label{eq:z_1(TA, TB)}
z_1 = \frac{T_A}{T_A+T_B}L.
\end{equation}
Therefore, in this case, thermodynamics univocally determines the
stationary state of the system.
\subsection{Multiple Piston}
We now consider the generalized piston model with $n>1$. An analysis
of the case $n=2$ is enough to understand how the relations obtained
from thermodynamics can be not sufficient to fully characterize the
non equilibrium steady state. Indeed, we have the following relations:
\begin{equation}\label{eq:gas_perfetti_2}
\begin{cases}
p z_1 = N k_B T_A,\\
p(z_2 - z_1) = N k_B T_1,\\
p(L-z_2) = N k_B T_B,
\end{cases}
\end{equation}
which give the constraints
\begin{equation}
\label{eq:z1z2(TATBT1)}
\begin{aligned}
z_1 = \frac{T_A}{T_A + T_1 + T_B}L,\\
z_2 = \frac{T_A+T_1}{T_A + T_1 + T_B}L.
\end{aligned}
\end{equation}
Therefore, since we have three variables ($z_1,z_2,T_1$) and only two
constraints, thermodynamics is not enough to determine the stationary
state. The computation can be easily extended to an arbitrary value of
$n$, leading to
\begin{equation}
\begin{cases}
pz_1 = N k_B T_A,\\
p(z_2 -z_1) = N k_B T_1\\
...,\\
p(z_n - z_{n-1}) = N k_B T_{n-1},\\
p(L-z_n) = N k_B T_B,
\end{cases}
\end{equation}
that give the conditions
\begin{myequation1}
\label{eq:posizioni_previsione_termodinamica}
\begin{aligned}
&z_1 = \frac{T_A}{T_A+\sum_{k=1}^{n-1}T_k+T_B }L, \quad \cdots \quad z_m = \frac{T_A+\sum_{k=1}^{m-1}T_k}{T_A+\sum_{k=1}^{n-1}T_k+T_B} L, \quad \cdots \quad
z_n = \frac{T_A+\sum_{k=1}^{n-1}T_k}{T_A+\sum_{k=1}^{n-1}T_k+T_B}L,
\end{aligned}
\end{myequation1}
with $m=2,\ldots, n-1$, namely, only $n$ constraints for $2n-1$ variables.
\section{Coarse-Grained Description and Effective Langevin Equations}
\label{three}
In order to obtain a statistical description of the system, we now
derive effective stochastic equations, governing the dynamics of the
relevant variables. Previous theoretical studies based on a Boltzmann
equation approach on similar systems were reported
in~\cite{GP99,GPL03,GL05}. In particular, a generalized piston model
was considered in Reference~\cite{GL05}. In those works, theoretical results
were not compared to numerical simulations, so that the underlying
hypotheses and their range of validity remained~unclear.
Here, we present a different analysis. We will assume that the
evolution of the observables is described by Langevin equations. The
idea, coming back to Smoluchowski, is to integrate out the fast
degrees of freedom of the gas particles by computing conditional
averages, knowing the macroscopic variables: position $z$ and velocity
$V$ of each piston, and temperatures $T$ of the gases. In order to
simplify the notation, let us denote by $\left<\cdot \right>$ this
average, meaning the conditional average $\left<\cdot |z, V,
T\right>$. We will compute the average change of a generic observable
$X$ in a small time interval $\Delta t$, due to the collisions between
the particles of the gas and the pistons.
Let us assume that in the stationary state the particles of the gas, in
each compartment, have uniform space distribution $\rho$ and a
Maxwell--Boltzmann distribution $\phi(v)$ at temperature $T$:
\begin{equation}
\label{eq:maxwell-Boltzmann}
\phi(v)=\sqrt{\frac{m}{2\pi k_{B} T}}e^{-\frac{m}{2k_{ B} T}v^2},\qquad \rho(x) = \frac{1}{\Delta z},
\end{equation}
where $\Delta z$ is the length of the box containing the gas. We can
obtain the rate of the collisions by considering the following
equivalent problem: piston at rest and a particle, which moves with the
relative velocity $v-V$. The point particles which collide against the
piston in $x$ in the time interval $dt$ are:
\begin{equation}
\label{eq:urto_dx_sx}
N\rho(x)(v-V)\Theta(v-V)dt, \quad\quad N\rho(x)(V-v)\Theta(V-v)dt,
\end{equation}
respectively, for particles on the left and on the right with respect
to the piston. The Heaviside function $\Theta$ is necessary to have a
collision.
Let us now consider a generic observable $X_j$, depending in general on the
velocities of the gas particles and of the pistons. We want to write down a Langevin equation:
\begin{equation}
\frac{d X_j}{dt}=D_j({\bf X}) + \textrm{noise},
\end{equation}
where both the drift term $D_j({\bf X})$ (${\bf X}$ being the vector of all relevant macroscopic variables in the system)
and the noise term are due to
collisions with the particles of the gas. We have:
\begin{equation}\label{eq:general_lr}
D_j({\bf X}) =\lim_{\Delta t \rightarrow 0} \frac{\left<\Delta X_j|{\bf X}\right>}{\Delta t} =\lim_{\Delta t \rightarrow 0} \frac{1}{\Delta t}\left[\left<\Delta X_j\right>^L_{coll} + \left<\Delta X_j\right>^R_{coll}\right],
\end{equation}
where
\begin{equation}
\begin{aligned}
\label{eq:general_left}
\left<\Delta X_j\right>^R_{coll}&=N\int^{\mathcal{1}}_{V} dv \, \int_{z_{ L} - (v-V)\Delta t}^{z_{ L}} dx \, X_j\,\rho(x) \phi(v)\Theta(v-V) = \frac{N}{\Delta z_{ L}} \int^{\mathcal{1}}_{V} dv \, X_j\phi(v)(v-V)\Delta t,
\end{aligned}
\end{equation}
\begin{myequation2}
\begin{aligned}
\label{eq:general_right}
\left<\Delta X_j\right>^L_{coll}&=N\int_{-\mathcal{1}}^{V} dv \, \int_{z_{ R}}^{z_{ R} - (v-V)\Delta t} dx \, X_j\,\rho(x) \phi(v) \Theta(V-v) = \frac{N}{\Delta z_{ R}}\int_{-\mathcal{1}}^{V} dv \, X_j\,\phi(v)(V-v)\Delta t.
\end{aligned}
\end{myequation2}
$\left<\Delta X_j\right>^R_{coll}$ is the variation of the observable due to the collisions
with the particles that have the pistons on the left, while $\left<\Delta X_j\right>^L_{coll}$ denotes the variation due to the collisions with the particles
which have the piston on the right.
In the stationary state, the macroscopic variables $T$, $V$ and $z$ do not
depend on time. It means that the time derivative of the conditional
average of a generic observable $X$ is zero, namely:
\begin{equation}
\label{eq:stato_stazionario}
\lim_{\Delta t \rightarrow 0} \frac{1}{\Delta t}\left[\left<\Delta X_j\right>^L_{coll} + \left<\Delta X_j\right>^R_{coll}\right]= 0.
\end{equation}
By using a perturbation development in the small parameter
$\epsilon=m/M$, it is possible to derive the relations between the
average positions and the temperatures of the pistons and the
temperatures of the gas. In order to obtain this result, we need to
average on the piston velocity, which is a stochastic variable. Of
course, in the steady state, the odd momenta have
to be zero:
\begin{equation}
\left< V^{2\alpha+1} \right>=0,\qquad \alpha=0,1,2,...
\end{equation}
The idea is to compute the average force produced by the particles of
the gas, which collide against the piston, by computing the average
momentum exchanged in the collisions.
The evolution of the velocity of the piston is governed by the
following stochastic differential~equation:
\begin{equation}
M \frac{d}{dt} V = F_{coll}(\Delta z_{ L}, \Delta z_{ R}, T_{ L}, T_{ R}, V) + K \eta,
\label{vel}
\end{equation}
where $\eta$ is zero-average white noise, usually $K$ is a constant, and
$F_{coll}$ is the average force which acts on the piston. This force
is due to the collisions with the gas on the left and on the right,
and depends on the average size of the box on the right and on the
left of the piston, $\Delta z_{ L}$ and $\Delta z_{ R}$, respectively, and
on the temperatures of the gas on the left and on the right, $T_{ L}$ and
$T_{ R}$. Now let us set $\Delta X=V'-V$ in Equation~\eqref{eq:general_lr}. As detailed in the Appendix, computing
the left and right contributions to $F_{coll}$ and expanding in the
small mass ratio $\epsilon=\sqrt{m/M}$, we obtain at the lowest
orders:
\begin{flalign}
\label{eq:V_zero}
&O(1):\quad Nk_B\left( \frac{T_{ L}}{\Delta z_{ L}} - \frac{T_{ R}}{\Delta z_{ R}}\right),&\\
\label{eq:V_uno0}
&O(\epsilon): \quad -2N\sqrt{2k_B} \sqrt{\frac{M}{\pi}} \left[ \frac{T_L^{1/2}}{\Delta z_{ L}} + \frac{T_R^{1/2}}{\Delta z_{ R}} \right] V.&
\end{flalign}
In the steady state, when the time derivative vanishes, we have from
the order $O(1)$ the following relation between the temperatures of
the gas and the average length of the boxes:
\begin{equation}
\Delta z_{ R}=\Delta z_{ L} \frac{T_{R}}{T_{L}}.
\end{equation}
This relation is nothing but the one that can be derived from thermodynamics.
From Equation~(\ref{eq:V_uno0}), we obtain the obvious result $\langle V\rangle=0$.
Analogous computations, reported in the Appendix, can be carried out
for the variance of the pistons and the gas temperatures. In
particular, for the variable $\Delta X=M(V'^2-V^2)$, at lowest orders
in $\epsilon$, we obtain the relations:
\begin{flalign}
\label{eq:Vq_zero0}
&O(1): \quad N2k_B \left(\frac{T_L}{\Delta z_{ L}} - \frac{T_R}{\Delta z_{ R}} \right),&\\
\label{eq:Vq_uno0}
&O(\epsilon): \quad \frac{N}{\sqrt{\pi}}\frac{4\sqrt{2}}{M^{1/2}}\left[\left(\frac{(k_BT_L)^{3/2}}{\Delta z_{ L}} + \frac{(k_BT)^{3/2}}{\Delta z_{ R}} \right)-MV^2\left(\frac{(k_BT_R)^{1/2}}{\Delta z_{ R}}+\frac{(k_BT_L)^{1/2}}{\Delta z_{ L}}\right)\right].&
\end{flalign}
In the steady state, by using the thermodynamic relations, the order
$O(1)$ Equation~\eqref{eq:Vq_zero0} is identically zero. After integrating
over all values of the piston velocity and by requiring that the order
$O(\epsilon)$ vanishes, we obtain a relation between the temperature
of the piston $T^{(p)}$ and those of the gases on the left and on the right:
\begin{equation}
T^{(p)}= M\left<V^2\right> = \left(T_{ R} T_{ L}\right)^{1/2}.
\label{Tp}
\end{equation}
Finally, a condition on the temperature of the gas as a function of
the velocity variance of the pistons on its left and on its right can
be obtained by considering the variable $\Delta X=m(v'^2-v^2)/N$ in
Equation~\eqref{eq:general_lr}. This gives (see Appendix) the relations:
\begin{flalign}
\label{eq:T_zero0}
&O(1): \quad \frac{2k_BT}{\Delta z} \left(V_L -V_R\right),\\
\label{eq:T_uno0}
&O(\epsilon): \quad \frac{\sqrt{2k_B}}{\Delta z}\frac{4}{\sqrt{\pi}}\frac{T^{1/2}}{M^{1/2}} \left[MV^2_R+MV^2_L - 2T\right].
\end{flalign}
By integrating over the velocity of the piston, because
$\left<V\right>$ is zero, the order $O(1)$ is identically zero. By
requiring that, in the stationary state, the order $O(\epsilon)$
vanishes, we eventually obtain:
\begin{equation}\
\begin{aligned}
&0 = \frac{\sqrt{2k_B}}{\Delta z}\frac{4}{\sqrt{\pi}}\frac{T^{1/2}}{M^{1/2}} \left[M\left<V^2_R\right>_0+M\left<V^2_L\right>_0 - 2T\right] \qquad \Longrightarrow \qquad T= \frac{M}{2} \biggl(\left<V^2_R\right> + \left<V^2_L\right> \biggr).
\end{aligned}
\end{equation}
As expected from thermodynamics, one can easily show that the gases in
contact with the thermostats reach the temperatures of the thermal
baths (see Appendix for a detailed explanation).
\section{Numerical Simulations}
\label{four}
In order to check the range of validity of the above approach, we have
performed extensive molecular dynamics simulations (using
an event driven algorithm) of the system, varying the relevant
parameters of the model and comparing the analytical prediction of
Section 3 with the actual numerical results. The system is initialized
with the pistons placed at equidistant positions, with zero velocity,
while the gas particles are randomly distributed in each box, with
velocities drawn from a Gaussian distribution at temperature
intermediate with respect to those of the thermostats. We have, however,
checked that the stationary state reached by the system is independent
of the initial conditions. All data presented in the following are
measured in the steady state, after the transient relaxation dynamics
from the initial state.
\begin{figure}[ht!]
\centering
\includegraphics[width=.5\columnwidth,clip=true]{3pistoni_temperaturapistoni.pdf}
\caption{Piston temperatures for a 3-piston system, with parameters $T_A=15$, \mbox{$N = M = 100$}, \mbox{$m = 1$}, \mbox{$L = 1$}.}
\label{fig1}
\end{figure}
In Figure~\ref{fig1}, we show the piston temperatures $T_i^{(p)}$,
$i=1,2,3$, for a 3-piston system, as measured in numerical simulations
(symbols) for a certain choice of the parameters, and compare them
with the theoretical predictions of Equation~(\ref{Tp}) (lines). The
approximation is good for the intermediate piston, while it is not
very accurate for the side pistons, and is the worst in the case of a
large gradient $T_B/T_A\gg 1$. Simulations performed for systems with
more pistons give similar results.
In Figure~\ref{fig2}, we consider a system with a large number of pistons
(\emph{n} = 22), where the temperature profile shows a linear behavior, in
agreement with Fourier's law. We report two sets of data, differing
in the value of the parameter $R=mN/M$. Notice that the linear
behavior extends for a wider range of $z$ in the case $R=10$. Lines
represent linear fits of the data. Let us note that, for large $n$ and
$R\gg 1$, $T^{(p)}$ vs. $z$ is in good agreement with a linear behavior
in the whole space interval (even close to the thermal~baths).
\begin{figure}[ht!]
\centering
\includegraphics[width=.5\columnwidth,clip=true]{22pistoni_profilotemperatura_MN.pdf}
\caption{Temperature profiles for a 22-piston system with $R=1$ and $R=10$. Other parameters are $T_A=15,T_B=30,M=50,m=1,L=1$.}
\label{fig2}
\end{figure}
The derivation of the Langevin equation (see Appendix) is mainly based on the assumptions:
\begin{itemize}
\item $m/M\ll 1$;
\item a local thermodynamic equilibrium, i.e. in each compartment
between the piston $i-1$ and the piston $i$, the spatial
distribution of the particles is uniform and the velocity
probability distribution is a Gaussian function, whose variance is
given by the gas temperature;
\item the collisions particles/pistons are independent.
\end{itemize}
The first assumption is easily checked, while the other ones are more
difficult. We can expect that a necessary condition for their validity
is that $N$ must be large. We have measured the velocity distribution
of the gas particles in the numerical simulations, finding that the
above hypothesis is verified. Moreover, we expect that
the number of recollisions decreases with the ratio $m/M$, and we
have numerically checked that their contribution is negligible (the
fraction of recollisions on the total number of collisions is about
$0.1\%$). Figure~\ref{fig4} shows the probability distribution
functions of particles colliding from left, $\phi_L(v)$, and from
right, $\phi_R(v)$, with a moving wall: the agreement with a
Gaussian assumption is rather evident.
In addition, the numerical computations show that the left- and
right-moving particles in the same compartment have the following probability
distributions:
\begin{eqnarray}
\phi_-(v)&=&\sqrt{\frac{2m}{\pi k_B T_-}}\;\Theta(v)e^{-\frac{mv^2}{2k_BT_-}}, \\
\phi_+(v)&=&\sqrt{\frac{2m}{\pi k_B T_+}}\;\Theta(-v)e^{-\frac{mv^2}{2k_BT_+}}, \\
\end{eqnarray}
where $T_-$ ($T_+$) denotes the temperature of the particles incident
on the piston $i$ ($i-1$) from the left (right), with $T_-\approx T_+$, resulting
in a small but finite heat flow proportional to $(T_-- T_+)$, in agreement
with~\cite{GL05}. Let us note that the above probability distributions are different
from the distributions $\phi_L(v)$ and $\phi_R(v)$, which are computed for particles
actually colliding with the piston; this is the origin of the presence of the factor $v$
appearing in the expressions in the caption of Figure~\ref{fig4}.
\begin{figure}[ht!]
\centering
\includegraphics[width=.5\columnwidth,clip=true]{1pistone_sxdx_dxsx.pdf}
\caption{Probability distribution function of the velocities
$\phi_L(v)$ ($\phi_R(v)$) of the particles colliding with the first
piston from the left (right), for a system with $n=2$. The curves,
red and blue, are respectively: $\phi_L = \frac{m}{k_{ B} T_{ L}}
e^{-v^2/k_{ B} T_{ L}}v\,\theta(-v)$ and $\phi_R = \frac{m}{k_{ B} T_{ R}}
e^{-v^2/k_{ B} T_{ R}}v\,\theta(v)$, where $T_{ L}=15.0$ and $T_{ R}=21.6$ are
the temperatures of the gas on the left and on the right with
respect to the first piston. Other parameters are: $\TTL=15$, $T_{B}=30$,
$N=100$, $M=100$, $m=1$, $n=2$, $L=1$.}\label{fig4}
\end{figure}
The distribution of $v$ in the compartment between pistons $i-1$ and $i$, has the form:
$$
\phi(v) \propto A \Theta(v)e^{- \frac{mv^2}{ 2k_BT_-}} + B \Theta(-v) e^{ - \frac{mv^2}
{2k_BT_+}},
$$
where $A$ and $B$ are constants.
However, since
$$
T_- = T_+ + O(\epsilon^2),
$$
we have a Gaussian distribution for the particle gas in the compartment:
\begin{equation}
\phi_i(v) \propto e^{- \frac{mv^2}{2k_BT_i}}+O(\epsilon^2), \qquad T_i = T_+ + O(\epsilon^2).
\end{equation}
Let us note that all our results are based on an expansion in powers of
$\epsilon$, neglecting $O(\epsilon^2)$. In other words, because the
heat rate is very small, this heat flux perturbs the Gaussian form of the
probability distribution $\phi$ in a negligible way.
In order to better understand the dependence of our theoretical
results on the parameter $R$, we have performed numerical simulations
of the system for different values of $R$. From the previous remark,
we expect to have an improvement of the agreement between the
numerical results and the analytical predictions by increasing $R$.
In Figure~\ref{fig3}, we compare the piston temperature in a 4-piston
system with theoretical predictions from Equation~(\ref{Tp}), finding a
very good agreement for large values of $R$. This shows that the total
mass of the gas contained in a box has to be larger than the mass of
the piston, in order for the kinetic theory presented in the previous
section to be accurate.
\begin{figure}[ht!]
\centering
\includegraphics[width=.5\columnwidth,clip=true]{4pistoni_tuttipistoni_MN.pdf}
\caption{Piston temperatures for a 4-piston system as a function of $R$. Dashed lines are the theoretical predictions. Other parameters are $T_A=15,T_B=30,m=1,L=1$.}
\label{fig3}
\end{figure}
\section{Conclusions}
We have studied a generalized piston in contact with two thermal baths
at different temperatures. This system represents a simple but
interesting case where the emergence of Fourier's law from a
microscopic mechanical model can be studied. We have presented a
kinetic theory treatment inspired by an approach \`a la Smoluchowski,
and we have investigated the range of validity of these results with
molecular dynamics numerical simulations. We have found that, in order
for the theory to be accurate, the ratio $R=mN/M$ should be large
enough, namely the total mass of the gas in each compartment should be
greater than that of the single piston.
We have considered ideal gas in our model, but we do not expect that
the introduction of short-range interactions among gas particles, at
least in not too dense cases, leads to significant differences in the
behavior of the system. An interesting, non-trivial, future line of
research in this model is the study of the relaxation to the
steady state and the dynamical properties of the system, focusing on
correlation and response functions.
|
\section{Conclusion + Future Work}
\label{sec:conclusion}
In this paper, we present a technique to use statistical information extracted from code corpora, encoded as
probabilistic attribute grammars, to guide program synthesis techniques. The method is based on rephrasing expression
enumeration as graph search, and applying classical techniques such as A* to solve these problems. We implemented the
algorithm in the Leon synthesis and verification system, and obtained a much more robust and scalable synthesizer in
comparison to the previous implementation. Our end goal is a predictable one-touch program synthesizer in a rich
programming environment such as Scala: there are several exciting directions of future work, such as investigating other
search methods including swarm search and genetic algorithms, incorporating richer models of statistics such as
probabilistic higher-order grammars (PHOGS)~\cite{Vechev-PHOG}, and combining type- and syntax-driven and statistical
techniques into a highly expressive and performant program synthesizer.
\section{Experimental Results}
\label{sec:experiments}
To experimentally evaluate the above ideas, we integrated them into
the Leon synthesis and repair framework \cite{LeonSynth-OOPSLA}
The Probabilistic Enumeration (PE) of Section~\ref{sec:alg} is deployed as
a closing deductive synthesis rule in the synthesis framework,
in place of Symbolic Term Exploration (STE) \cite{LeonSynth-SYNT}.
For repair, we slightly modify Probabilistic Enumeration to
generate terms similar to the erroneous expression, in the spirit of
\cite{LeonSynth-CAVRepair}.
This can be achieved by automatically modifying the grammar
passed to Enumeration, without any other changes to the algorithm.
This rule is deployed along with plain PE within the synthesis stage
of the repair pipeline,
after test generation and fault localization
In our evaluation, we aimed to measure how well our algorithms
perform in isolation, rather than how well they integrate with the
synthesis framework.
This is straightforward in repair,
as the erroneous implementation already defines the
high-level structure of the function under repair,
and the closing rules can usually be deployed alone,
without instantiating any decomposition rules first.
In the following tables, we compare the efficiency of synthesizing repairs
for a set of erroneous benchmarks taken from \cite{LeonSynth-SYNT},
along with a few new, more challenging ones.
In total, we test 4 tool configurations:
\begin{itemize}
\item Symbolic Term Exploration with built-in grammars and
built-in axioms, i.e. the old version of Leon (\emph{STE})
\item Probabilistic Enumeration with built-in grammars
and built-in axioms (\emph{PB})
\item Probabilistic Enumeration with depth-1 extracted grammars
and built in axioms (\emph{PD1})
\item Probabilistic Enumeration with depth-2 grammars,
without built-in axioms (\emph{PD2})
\end{itemize}
In both the latter cases, the grammars come from a combination
of grammar files extracted from a corpus and the program under repair.
We do not use built-in axioms with depth-2 grammars,
as they are redundant to some degree and they would overcomplicate the
grammar, while yielding diminishing returns.
The detailed results are given in \ref{table:repair}.
The first two columns of the table give an indication of the complexity of the problem
by displaying the total size of the program and the size of the solution
(which might not be generated from scratch, but rather as a variation
of the erroneous expression). The final four columns display
the repair times for the four deployments of the tool described above.
Repair times exclude compilation, test generation, and verification of the solution.
\input{table-repair}
To summarize the results, the new enumerator is slightly faster for the built-in grammars,
whereas the depth-1 grammars are sometimes slightly slower, but manage to solve a couple more benchmarks.
Unfortunately, there is a relatively easy benchmark that depth-1 fails to solve.
The depth-2 grammars, though they show some promise (in the
\lstinline{Compiler.semUntyped} benchmark), they generally perform badly.
One reason for this behavior are the conflicts between annotated nonterminal symbols
generated from the general corpus with those from the file under repair.
Both sources will generate annotated non-terminals that express the same conditional
probability, overloading the synthesizer with duplicated work.
This is a shortcoming we plan to address in the near future.
\section{Term Grammars}
\label{sec:grammar-gen}
In the previous section, we explained our synthesis algorithm
using a PCFG as a black box. In this section, we explain how
the term grammars in our system are defined and extracted.
We designed our grammars with the following goals in mind:
\begin{itemize}
\item \emph{Flexibility.} The system should allow the user to define
and easily deploy a different grammar for each application
or application domain.
\item \emph{Expressivity.} The grammars should be able to express complex
relations between expressions, in a more elaborate way than simply
mapping each type of the target language to a set of production rules.
This allows the user to constrain the space of synthesized programs
to exclude redundant or irrelevant ones,
thus improving the scalability of the synthesizer.
\item \emph{Automation.} Apart from hand-coding the grammars,
it should also be possible to automatically extract them
from a corpus of programs, possibly taken from a specific application domain.
\end{itemize}
To this end, we introduce \textit{Probabilistic Attribute Grammars},
which address all of the above points:
\begin{itemize}
\item A custom grammar is provided by the user along with the synthesis problem
as a set of Scala source files (the \emph{grammar files}),
in which production rules are expressed as plain Scala functions.
Switching between grammars reduces to just including different files in compilation.
It is also trivial to use any desired combination of such files.
If the user does not want to manually provide a specific grammar for the task,
they can always fall back to the system's predefined grammar.
\item The nonterminal symbols of the grammar are not plain types,
but instead are types contained in a wrapper class that annotates them with extra information.
The wrapped types can represent a subtype of the underlying type,
or some other condition under which specific rules are accessible,
which would not be expressible if we used plain nonterminals as types.
While processing the input grammar, our system will desugar
away those wrappers to produce values of the underlying plain type.
The approach of annotating types with attributes has already been introduced
in \cite{LeonSynth-SYNT}, but there the attributes and their interpretation
were predefined. Here we allow the user to provide their own wrapper
classes.
It is of course also possible to omit the wrappers completely
and use plain Scala types as nonterminals.
\item We provide a system to automatically extract grammar files
from a corpus of PureScala programs.
\end{itemize}
The grammar files undergo a series of preprocessing steps before generating
a Probabilistic Attribute Grammar, the exact form of which is
conditional to the synthesis problem. The synthesizer itself is oblivious
to this procedure, which views the grammar just as a function from
nonterminal symbols to production rules.
\subsection{Grammar files, rules and frequencies}
\label{grammarfiles}
As mentioned above, the user can provide a series of grammar files for compilation
along with the synthesis problem.
The format of the files was designed with the aim to be readable and writable by a human user,
but also straightforward to extract from a corpus of programs.
To illustrate the format though an example, let us look at a grammar file describing
a set of simple arbitrary precision integer programs:
\begin{lstlisting}
@production(10)
def plus(a: BigInt, b: BigInt): BigInt = a + b
@production(5)
def minus(a: BigInt, b: BigInt): BigInt = a - b
@production(5)
def o: BigInt = BigInt(1)
@production(10)
def z: BigInt = BigInt(0)
@production(20)
def vBigInt: BigInt = variable[BigInt]
\end{lstlisting}
A function annotated with \lstinline{@production}
is treated as a grammar production rule.
A rule of the form
\lstinline{def pFoo(a: T1, b: T2): T3 = foo(a, b)}
should be interpreted in CFG notation as
$\textsf{T3} \Coloneqq foo(T1, T2)$
The annotation's argument indicates the absolute frequency with which
this rule is encountered. The relative frequencies will be generated
in the end of a processing procedure. In this example, though, they can be computed
straightforwardly from the absolute frequencies.
An invocation of the built-in function \lstinline{variable[T]},
for some type \lstinline{T}, indicates the absolute frequency of generating
any variable of type \lstinline{T}.
This built-in compensates for the fact that we do not know the names of
the variables accessible to the synthesizer a priori.
When they become available, this production will be instantiated
to a concrete production for each variable of the correct type,
with the probability distributed equally among these variables.
If, for example, we are synthesizing a function with two parameters
\lstinline{(x: BigInt, y: BigInt)},
the above grammar file would generate the following grammar:
\[
\begin{array}{rcll}
\textsf{BigInt} & \Coloneqq & \textsf{BigInt} + \textsf{BigInt} & (p_+ = 0.2) \\
& \mid & \textsf{BigInt} - \textsf{BigInt} & (p_- = 0.1) \\
& \mid & \lstinline|0| & (p_0 = 0.1) \\
& \mid & \lstinline|1| & (p_1 = 0.2) \\
& \mid & \lstinline|x| & (p_x = 0.2) \\
& \mid & \lstinline|y| & (p_y = 0.2)
\end{array}
\]
\subsection{Generic productions}
Generic grammar productions offer a concise way of
representing operations on generic data-structures. For
instance, we could describe common operations on a generic
list data-structure with a few user-provided productions:
\begin{lstlisting}
@production(..)
def single[A](a: A): List[A] = List(a, Nil)
@production(..)
def listInsert[A](a: A, l: List[A]): List[A] = List(a, l)
@production(..)
def listTail[A](l: List[A]): List[A] = l.tail
@production(..)
def listHead[A](l: List[A]): A = l.head
@production(..)
def listSize[A](l: List[A]): BigInt = l.size
\end{lstlisting}
When looking for productions of a specific ground type,
we instantiate generic productions that \emph{can}
return that type, and introduce them as normal monomorphic
productions.
For example, if we are looking for productions of
\lstinline{List[BigInt]},
we consider the first three rules above instantiated with
$T \mapsto BigInt$, and the forth with $T \mapsto
List[BigInt]$. The fifth rule is incompatible.
Because the set of types is typically infinite, generic
productions with a type variable absent from the
return type represent an infinite number of ground
productions. This is for instance the case with
\lstinline{listSize}.
Our enumeration procedure described above requires
that each non-terminal symbol has a finite number of ground
productions. To solve this conflict, we abandon the
theoretical completeness of the instantiation by only
considering a set of \emph{reasonable}
types $\mathcal{T}$ during instantiation. This set of types
is initialized with all types returned by ground productions
and then iteratively expanded by discovering new return types of
$\mathcal{T}$-instantiations of generic productions.
For example, starting with $\mathcal{T} = \{\textsf{Int}\}$,
running one iteration yields $\mathcal{T} = \{\textsf{Int},
\textsf{List[Int]}\}$ as \lstinline/single[Int](..)/ returns
a value of type \lstinline/List[Int]/.
The number of discovery iterations we perform has a big
practical impact on the number of final productions, and
thus on performance. On benchmarks with several generic
data-structures, this discovery is typically exponential.
By under-approximating the size of the smallest expression
of any discovered type in $\mathcal{T}$, we bound the
search to reasonable types.
We leave to future work how to better integrate the
instantiation of generic rules within the enumeration
process.
\subsection{Annotated nonterminals}
\label{annotated}
Simple grammars like the one displayed above, where the nonterminal symbols
are plain types, are not expressive enough to handle all constraints that are obvious to a human programmer.
For example, they cannot express that an operand of an addition cannot be 0.
When used as-is, such grammars generate a large number of redundant programs,
which prevent the synthesizer from scaling to larger program sizes.
To amend this, we provide the programmer with a general mechanism
to more accurately express the restrictions of their grammar
by enhancing their nonterminals with attributes
in additional to the plain Scala type.
Nonterminals with the same type but different attributes
are considered distinct, and can have different production rules.
As an example, suppose the user wants to more accurately specify the grammar
of Section~\ref{grammarfiles} to exclude 0 from the operands of \lstinline{+},
as well as the second operand of \lstinline{-}.
This can be captured with the following attribute grammar:
\[
\begin{array}{rcll}
\textsf{BigInt} & \Coloneqq & \textsf{BigInt\{BI\}} & (p = 1.0) \\
\textsf{BigInt\{BI\}} & \mid & \textsf{BigInt\{NZ\}} & (p = 0.8) \\
& \mid & \lstinline|0| & (p = 0.2) \\
\textsf{BigInt\{NZ\}} & \mid & \textsf{BigInt\{NZ\}} + \textsf{BigInt\{NZ\}} & (p_1 = 0.25) \\
& \mid & \textsf{BigInt\{BI\}} - \textsf{BigInt\{NZ\}} & (p_1 = 0.125) \\
& \mid & \lstinline|x| & (p_x = 0.5) \\
& \mid & \lstinline|1| & (p_y = 0.125)
\end{array}
\]
The above grammar contains two annotated nonterminals,
\textsf{BigInt\{NZ\}} and \textsf{BigInt\{BI\}}.
The former expresses nonzero integers and the second one
the unrestricted integer type. Thus, generating \textsf{BigInt} values
reduces to generating values of the annotated type
\textsf{BigInt\{BI\}},
which is captured in the first rule of the grammar.
In Scala source code, this is expressed by defining
implicit Scala classes containing a field of the base type,
and annotating them with the \lstinline{@label} annotation.
The source file gets desugared by the system to the above grammar.
\begin{lstlisting}
@label implicit class NZ(val v: BigInt)
@label implicit class BI(val v: BigInt)
@production(10) def plus(a: NZ, b: NZ): NZ = a.v + b.v
@production(5) def minus(a: BI, b: NZ): NZ = a.v - b.v
@production(5) def o: NZ = BigInt(1)
@production(10) def z: BI = BigInt(0)
@production(20) def vBigInt: NZ = variable[BigInt]
@production(40) def nz2Bi(nz: NZ): BI = nz.v
@production(1) def start(b: BI): BigInt = b.v
\end{lstlisting}
\subsubsection{Built-in axiom system}
\label{axioms}
In case the user does not want to manually implement a custom attribute grammar as above,
they can reside to a set of default built-in axioms described in \cite{LeonSynth-SYNT}.
In case of the default grammars, those axioms are automatically applied to each combination of
nonterminal and production rule.
In case of a manually provided grammar, the user has to guide the system by
tagging the production rules with a set of predefined tags.
A couple of example tagged rules could be the following:
\begin{lstlisting}
@production(10)
@tag("0")
def z: BigInt = BigInt(0)
@production(10)
@tag("commut")
def pFoo(b1: BigInt, b2: BigInt): BigInt = foo(b1, b2)
\end{lstlisting}
\lstinline{@tag("0")} indicates that this rule produces the constant 0
(e.g. it should be treated as neutral element of addition),
whereas \lstinline{@tag("commut")} specifies \lstinline{foo} as a commutative
operation. These rules will undergo a set of predetermined optimizations
described in \cite{LeonSynth-SYNT}.
\subsection{Extracting an attribute grammar form a corpus}
Whereas manually specifying grammars is very useful for specific tasks and small DSLs,
it would be tedious to expect a user-provided grammar for generic programming tasks,
where numerous types and functions are in scope.
To this end, we provide an automated system to extract a grammar from a corpus
of Scala programs. We provide two different grammar extractors,
described in the following paragraphs.
\subsubsection{Unconditional Frequency Extractor}
The simpler Unconditional Frequency Extractor (UFE) uses plain Scala types as nonterminals,
but automatically tags them as in \ref{axioms},
so the built-in axioms can be instantiated on them.
Each production rule generated by the UFE corresponds to a specific \emph{kind}
of expression encountered in the corpus,
whereas the rule's frequency reflects the number of occurrences
of this kind.
By kind of expression, we mean the expression type
along with all information required to reconstruct the expression
from its arguments (except for variables, where the name is not preserved).
More specifically,
\begin{itemize}
\item For variables, only the variable type (the name does not get preserved).
All variables of a specific type will all be summarized by an invocation
to the \lstinline{variable} built-in.
\item For literals, the literal itself (which takes 0 operands).
E.g. \lstinline{0}, \lstinline{BigInt(42)}.
\item For function invocations, the invoked function along with its actual type parameters
(which may be generic).
E.g. \lstinline{foo[BigInt](?, ?)}, \lstinline{foo[T](?, ?)},
where by \lstinline{?} we denote a hole to be filled during synthesis.
\item For type constructors, the type constructor along with its actual
type parameters. E.g. \lstinline{Cons[BigInt](?, ?)},
\lstinline{Nil[A]()}.
\item For every other expression, only the type of the AST of that expression,
and if needed its type instantiation.
E.g. \lstinline{? + ?}, \lstinline{Set[BigInt](?)}.
\end{itemize}
Because the grammars generated by this extractor do not take into
account the parent of the expression,
we will call them \emph{depth-1 grammars}.
For example, a corpus containing only the expression \lstinline{(x * 2)}
would result in the following file:
\begin{lstlisting}
@production(1)
@tag("const")
def pBigIntInfiniteIntegerLiteral0(): BigInt = BigInt(2)
@production(1)
@tag("times")
def pBigIntTimes(v0 : BigInt, v1 : BigInt): BigInt = v0 * v1
@production(1)
@tag("top")
def pBigIntVariable(): BigInt = variable[BigInt]
\end{lstlisting}
\subsubsection{Conditional Frequency Extractor}
The second, more sophisticated extractor, extracts the conditional
probabilities of the occurrence of each expression kind,
given the expression's relation with its parent.
More specifically, along with each expression, it stores a pair
$(par, pos)$, where $par$ is the type of AST of the parent of the expression,
and $pos$ the expression's position in its parent's operands.
This pair is optional and omitted for a top-level expression.
For example, if the extractor encountered the expression \lstinline{(x * 2)},
it would store the triples \lstinline{(? * ?, null, null)},
\lstinline{(<variable>, ? * ?, 0)} and \lstinline{(2, ? * ?, 1)}.
To express these conditional probabilities in a grammar file,
it uses annotated nonterminals as described in \ref{annotated}.
Because these grammars express probabilities which are derived from
an expression along with its parent,
we will call them \emph{depth-2 grammars}.
A corpus containing only the above expression would produce the following
grammar file:
\begin{lstlisting}
@label implicit class BigInt_TOPLEVEL(val v : BigInt)
@label implicit class BigInt_0_Times(val v : BigInt)
@label implicit class BigInt_1_Times(val v : BigInt)
@production(1)
def pBigIntTimes(v0 : BigInt_0_Times, v1 : BigInt_1_Times): BigInt_TOPLEVEL =
v0.v * v1.v
@production(1)
def pBigIntVariable(): BigInt_0_Times = variable[BigInt]
@production(1)
def pBigIntInfiniteIntegerLiteral(): BigInt_1_Times = BigInt(2)
@production(1)
def pBigIntStart(v0 : BigInt_TOPLEVEL): BigInt = v0.v
\end{lstlisting}
\subsection{Biasing grammars for repair}
The flexibility and automation of the system give us a great opportunity
to exploit the local structure of the program.
Especially for program repair, we can ease the task of generating a reasonable
patch for an erroneous program by collecting information about how
the user has chosen to program in the same module.
This is achieved as follows: before repair, we generate an additional grammar file
from the file under repair, and include it for compilation along with
the general-purpose grammar file extracted from the corpus.
This way, we bias the synthesis process towards local definitions,
without completely disregarding the general programming model.
\section{Introduction}
\label{sec:intro}
Program synthesis has attracted significant attention as a promising and ambitious approach
to improve software development productivity \cite{Manna-1971,Sketch,Albarghouthi2013,Feser2015,CVC4-SyGuS,Brahma,Gulwani-etal-2011,InSynth,Osera-Myth-15, Osera-Myth-16,Synquid,LeonSynth-OOPSLA, LeonSynth-SYNT}.
Among possibly deployment scenarios of synthesis is program repair, in which
a tool localized an error repairs the program by synthesizing an alternative
code fragment \cite{PeiETAL15RepairTests,LeonSynth-CAVRepair}.
Synthesis tools typically leverage both \emph{semantic} and \emph{syntactic} specification of the synthesis process. Typical forms of semantic information are specifications in the form of
input-output examples, as well as reference implementations and predicates. These specifications are important to identify the specific intent of the developer. Among the syntactic forms of specification are simple types and probabilistic models of code. Syntactic models of code are appealing for two main reasons: they can improve the efficiency of search, and they can provide implicit, reusable baseline specifications that apply across multiple synthesis problems, reducing the need to provide near-complete semantic specifications of the desired functionality.
Our work aims to unify these two sources of specifications. To this extent, we describe a specification framework and an algorithms that efficiently searches for programs based on:
\begin{itemize}
\item a probabilistic and polymorphic attribute grammar that restricts and prioritizes the relevant space, and
\item dynamically growing set of input examples along with test predicates, which narrows
down the developer's intent.
\end{itemize}
Our attribute grammars describe the generation of trees and can be thought of as
a probabilistic extension of type systems used in previous work. The use of generics
and attributes makes the grammars compact and enables us to naturally describe
constructs such as conditionals that are needed to generate more complex code fragments.
Thanks to the fact
that we use a language supporting generics and value classes (Scala), we show how to describe
such derived grammars in the form of a wrapper library. The developers can therefore
use mostly existing programming language constructs to describe the intended models of
use of their libraries. The use of attributes enables the developers to incorporate
domain knowledge, such as associativity of operators, which can break symmetries and
reduce the search space. In addition to explicitly writing such models, we present an algorithm to automatically derive models from a corpus of code. Our results show that the derived
corpus provides benefits in synthesis of meaningful code fragments.
Our algorithm efficiently supports specifications that include concrete inputs. In the
case where the provided set of inputs is not sufficient to narrow down the search, our system
can also make use of constraint solvers to generate additional inputs.
To make search driven by probabilistic grammars efficient, we adopt a top-down approach for search. We introduce techniques to prune the search for expressions using partial evaluation of incomplete grammars on known input examples. Furthermore, we show how to incorporate equivalence modulo input-output examples in such a search algorithm, which is a technique that has so far been deployed primarily in bottom-up search techniques.
The practical focus of our techniques is on program repair. We use manually supplied
and automatically derived examples to localize the expression likely to be responsible, then synthesize an alternative
expression in its place.
Examples
are particularly relevant because they can be used to localize the error
efficiently. The
support of input examples in our algorithm makes it efficient for synthesizing
candidate repairs.
Furthermore, our support for probabilistic grammars enabled us to adopt the following approach
to repair with probabilistic models:
in addition to using grammars describing general Scala expressions, we bias the search
for repairs towards the constructs specific for a program being repaired, making
the synthesis more effective. We therefore show that our approach is able to produce
program repairs whose size puts them beyond the reach of analogous existing
non-probabilistic repair approaches.
\section{Related Work}
\label{sec:related}
\subsection{Program synthesis}
\label{sec:related:synt}
Program synthesis has been a challenge problem in computer science since its early days~(see, for example,~%
\cite{Manna-1971}). With increasing computing power, and rapid improvements in the performance of SAT and SMT solvers
starting from the early 2000s, the problem of program synthesis has recently received renewed attention. The seminal
Sketch project~\cite{Sketch} was the first to demonstrate the feasibility of using modern constraint solving technology
in synthesis tools, and used the popular counter-example guided inductive synthesis~(CEGIS) framework to solve synthesis
problems~\cite{Gulwani-etal-2011}.
Syntax-guided synthesis~(SyGuS) has emerged as a common formulation and interchange format in which to express many
program synthesis problems~\cite{SyGuS}. Our problem formulation within Leon shares two important similarities with
SyGuS:
\begin{enumerate}
\item \emph{Multi-modal specifications}, which can include any combination of input-output examples~(such as ``$f(2, 3)
= 8$''), logical constraints~(such as ``$\forall x, y, f(x, y) > x + y + 2$''), and even partial implementations~(such
as ``$\forall x, y, x < y \implies f(x, y) = x + y + 3$'').
\item \emph{Expression grammars}, or equivalently, a library of \emph{components}~\cite{Brahma}, are used to
syntactically constrain the space of programs. This is important both to make synthesis feasible for complicated
specifications, and for when the results are to be used in specialized domains, such as assembly code for new computer
architectures~\cite{Chlorophyll}.
\end{enumerate}
Of the various techniques to solve SyGuS problems, ESolver, which enumerates pairwise distinguishable expressions~%
\cite{Transit}, EUSolver, which combines indistinguishability-based enumeration with decision-tree based condition
inference~\cite{EUSolver}, and the refutation-proof-based CVC4 solver~\cite{CVC4-SyGuS} have broadly emerged as the most
efficient SyGuS solvers. Our present work may be viewed as an attempt to bias the enumerative solver with statistical
facts learned from a code corpus.
We would like to highlight that program synthesis and code repair within Leon is a much more challenging problem than
synthesis within SyGuS and related systems: the main difficulties include the rich type system and algebraic data types
of Scala, and the synthesis of recursive functions. In this light, our problem is most similar to
Escher~\cite{Albarghouthi2013} and $\lambda^2$~\cite{Feser2015}: the main differences are that Escher enumerates
programs bottom-up rather than top-down, and our use of PCFGs to improve the performance of the enumeration algorithm.
\subsection{Type-driven synthesis}
\label{sec:related:types}
One appealing approach to program synthesis is to phrase the problem as the type-inhabitation problem in a sufficiently
expressive type system. Examples of such systems include InSynth~\cite{InSynth}, and Synquid~\cite{Synquid}. The Myth
project~\cite{Osera-Myth-15, Osera-Myth-16} extends the type system of the host language with refinement types
corresponding to input-output examples. It can therefore be viewed as an attempt to provide a type-theoretic
interpretation to Escher. The Synquid system~\cite{Synquid} is an elegant extension of this idea, by encoding the
specification as part of the type of the required expression using the liquid type framework.
Previous work on synthesis within Leon~\cite{LeonSynth-OOPSLA, LeonSynth-SYNT} may also be viewed as being implicitly
type-driven. The path conditions and postcondition obtained after each invocation of the deduction rules can be
rephrased as additional typing constraints on the expression being synthesized. As discussed in Section~\ref{sec:alg},
the present PCFG-based enumeration is intended as a drop-in substitute for the symbolic term exploration rule previously
used by Leon.
\subsection{Statistical analysis of code corpora}
\label{sec:related:corpora}
With increasing availability of large open-source code repositories such as GitHub and Bitbucket, the statistical
analysis of code corpora has become an exciting research problem. Code repositories have been used to learn coding
idioms~\cite{AllamanisSutton-Idioms}, to automatically suggest names for program elements~\cite{AllamanisSutton-Names},
and to deobfuscate JavaScript code~\cite{Vechev-BigCode}. Our paper is an attempt to apply similar techniques to
accelerate program synthesis.
The program completion tool Slang~\cite{Vechev-PLDI14} uses $n$-grams and recurrent neural networks to predict missing
API calls in code snippets. There are two main aspects which distinguish our work from Slang:
\begin{inparaenum}[(\itshape a\upshape)]
\item the presence of hard correctness requirements in Leon in the form of pre-/post-conditions, and
\item program synthesis in Leon is fundamentally about synthesizing expressions rather than API call sequences, and
prediction systems such as $n$-grams are insufficient to create the nested recursive structure inherent in the output
we produce.
\end{inparaenum}
The DeepCoder tool~\cite{DeepCoder} uses a recurrent neural network (RNN) to predict the presence of elements in the
program being synthesized. The output of this neural network is then used to guide a more exhaustive search over the
space of possible programs. In this sense, DeepCoder is similar to the PCFG-guided enumeration of Leon, and the use of
neural networks to accelerate the program synthesis problem is an intriguing direction for future research.
Probabilistic context-free grammars (PCFGs) are a classical extension of context-free grammars~\cite{NLP-Jurafsky}. They
may be used both to model ambiguity (for applications in natural language processing), and to model probability
distributions over the generated language, which motivates our application in accelerating code synthesis. The more
recent model of probabilistic higher-order grammars (PHOGs)~\cite{Vechev-PHOG, Vechev-PHOG0} extends PCFGs by allowing
the expansion probabilities of a non-terminal node to depend on attributes such as node siblings and DFS-predecessors.
Experiments indicate that the PHOG model is significantly better at predicting elements of JavaScript programs than
PCFGs. Extending the probabilistic model of our paper to use PHOGs instead of PCFGs is an immediate area of future work.
\subsection{Search algorithms}
\label{sec:related:search}
One of the main technical insights of this paper is to phrase syntax-guided program synthesis as an instance of graph
search. In this space, A*~\cite{AStar} is a classical algorithm used for path finding and graph traversal. As described
in Section~\ref{sec:alg}, A* orders the queue of still-unexplored nodes by their priority values, $f(n) = g(n) + h(n)$,
where $g(n)$ is the cost required to reach the node $n$, and the heuristic function $h(n)$ is the estimated cost from
$n$ to the target node. The historically older Dijkstra's algorithm for shortest paths through graphs is a special case,
where the heuristic function $h(n)$ is identically equal to $0$.
Recall that the graph of partial expansions forms a tree: in this setting, if the cost function $h(n)$ is
\emph{admissible}, i.e. that $h(n) \leq \tilde{h}(n)$, where $\tilde{h}(n)$ is the actual cost of the shortest path from
$n$ to the target node, then A* discovers the shortest path to the target node (\emph{optimality}). The second important
property of A* is that of \emph{optimal efficiency}: of all algorithms using the same heuristic function, A* explores
the fewest nodes to reach the target. In the setting of this paper, optimality implies discovering the expansion with
highest probability. By extending $h(n)$ to include the number of satisfied examples, we have relaxed optimality in
favour of empirical performance improvements.
A* is simply the best known of a large family of search algorithms~(see chapters~3 and~4 of \cite{AIMA}), including
iterative deepening A* (IDA*,~\cite{IDAStar}), simplified memory-bounded A* (SMA*,~\cite{SMAStar}), fringe
search~\cite{FringeSearch}, beam search, swarm search, and various instantiations of genetic
algorithms~\cite{GAMitchell}. Extending the techniques of this paper to these algorithms is an exciting direction of
future work.
\section{Synthesis Algorithm}
\label{sec:alg}
\subsection{Synthesis problem}
Let $\phi = \phi(\seqa, x)$ be the \emph{specification} of a programming task,
specifying a relation between a sequence of input parameters \seqa
and an output variable $x$.
$\phi$ can be given as a logical predicate, a set of input-output examples,
a reference implementation, or a combination thereof.
Also, let $\pcname = \pcname(\seqa)$ be the \emph{path condition} for the problem,
i.e. a relation satisfied by hypothesis by the input parameters.
Let also $t = t(\seqa)$ express a term in the target language.
The \emph{synthesis problem} is then defined as asking for a constructive solution for the formula
$$\forall \seqa. \exists t. \pcname \implies \phi[x \mapsto t]$$
In other words, we want to find an expression $t$ such that replacing every occurrence of the output variable $x$ in the
synthesis predicate $\phi$ causes the formula $\pcname \implies \phi$ to be \emph{valid}. In this section, we will
describe the probabilistic enumeration procedure used to solve such formulas.
\subsection{Counter-example guided inductive synthesis (CEGIS)}
The synthesis algorithm is an instantiation of CEGIS, a widely used framework to describe program synthesis algorithms~%
\cite{Sketch, Gulwani-etal-2011}. The main technical difficulty with finding a $t$ which satisfies rule $\TerminalRule$
is the universally quantified ``$\forall \seqa$'', where the input variables $\seqa$ may range over a large, or even
potentially infinite domain. We present the CEGIS routine in Algorithm~\ref{alg:cegis}. The procedure is parameterized
by two sub-procedures, $\Search$ and $\Verify$. At each step, we maintain a finite set $A$ of concrete input points. In
the search phase, the algorithm finds an expression $t$, such that
\begin{equation}
\Land_{\seqa \in A} \pcname \implies \phi[x \mapsto t].
\label{alg:cegis:search}
\end{equation}
Observe that for a specific choice of $A$, determining whether a given $t$ satisfies requirement~\ref{alg:cegis:search}
reduces to simply evaluating the resulting expression. In the verification phase, we confirm that the candidate
expression $t$ works for all inputs $\seqa$. The procedure $\Verify(\seqa, \pcname, \phi, x, t)$ returns either:
\begin{inparaenum}[(\itshape a\upshape)]
\item $\textsf{valid}$, if $t$ satisfies the synthesis predicate for all values of $\seqa$, or
\item a counter-example input $\seqa_{\text{cex}}$, such that $\pcname[\seqa \mapsto \seqa_{\text{cex}}] \land \lnot
\phi[\seqa \mapsto \seqa_{\text{cex}}, x \mapsto t]$ i.e. $\seqa_{\text{cex}}$ satisfies the path condition $\pcname$, but $t$ fails the
synthesis predicate.
\end{inparaenum}
\begin{algorithm}
\caption{$\CEGIS(\br{\seqa}{\pcname}{\phi}{x})$.}
\label{alg:cegis}
\begin{enumerate}
\item Initialize $A \coloneqq \emptyset$. $A$ is a finite set of concrete instantiations of $\seqa$.
\item Repeat forever:
\begin{enumerate}
\item Let $t = \Search(\seqa, A, \pcname, \phi, x)$.
\item Let $\textsf{res} = \Verify(\seqa, \pcname, \phi, x, t)$.
\item If $\textsf{res} = \textsf{valid}$, then return $t$.
\item Otherwise, if $\textsf{res} = \seqa_{\text{cex}}$, update $A \coloneqq A \union \{ \seqa_{\text{cex}} \}$.
\end{enumerate}
\end{enumerate}
\end{algorithm}
We discharge the calls to $\Verify$ by invoking Leon's verification system~\cite{Leon-Scala13}. The main technical
contribution of this paper is in using probabilistic enumeration to instantiate the $\Search$ subroutine. We devote the
rest of this section to describing this procedure.
\subsection{Probabilistic expression grammars}
The goal of the $\Search(\seqa, A, \pcname, \phi, x)$ subroutine is to find a candidate expression $t$ which satisfies
the synthesis predicate for all input points $\seqa \in A$. The expression $t$ is drawn from the productions of a
probabilistic context-free grammar (PCFG), such as that shown below:
\[
\begin{array}{rcll}
\textsf{IntExpr} & \Coloneqq & \lstinline|0| & (\text{Rule } R_0, p_0 = 0.15) \\
& \mid & \lstinline|1| & (\text{Rule } R_1, p_1 = 0.3) \\
& \mid & \lstinline|x| & (\text{Rule } R_x, p_x = 0.3) \\
& \mid & \textsf{IntExpr} + \textsf{IntExpr} & (\text{Rule } R_+, p_+ = 0.15) \\
& \mid & \lstinline|if ($\textsf{BoolExpr}$) { $\textsf{IntExpr}$ } else { $\textsf{IntExpr}$ }|
& (\text{Rule } R_{\text{ite}}, p_{\text{ite}} = 0.1) \\
\textsf{BoolExpr} & \Coloneqq & \lstinline|$\textsf{IntExpr}$ $\leq$ $\textsf{IntExpr}$|
& (\text{Rule } R_\leq, p_\leq = 0.8) \\
& \mid & \lstinline|$\textsf{BoolExpr}$ && $\textsf{BoolExpr}$|
& (\text{Rule } R_{\land}, p_\land = 0.2)
\end{array}
\]
The productions of $\textsf{IntExpr}$ expand to expressions of type \lstinline|Int|, and the productions of
$\textsf{BoolExpr}$ expand to expressions of type \lstinline|Bool|. Each rule $R$ of each non-terminal symbol $N$ is
annotated with a number $p_R$, which indicates the probability of $N$ expanding to $R$. Naturally, the probability $p_e$
of a production $e$ is then defined as the product of the probabilities of all rules invoked in the derivation tree: for
example, the production \lstinline|x + 1| has probability $p_+ p_x p_1 = 0.0135$.
More generally, a PCFG is a pair $G = (\mathcal{N}, \mathcal{R})$, where:
\begin{enumerate}
\item $\mathcal{N}$ is a finite, non-empty set of non-terminal symbols, where each non-terminal $N \in \mathcal{N}$ is
associated with a type $T_N$,
\item $\mathcal{R}$ maps each non-terminal $N$ to a finite set of production rules $\mathcal{R}(N)$. Each production
rule $R \in \mathcal{R}(N)$ is a well-typed construct of the form $R = t(N_1, N_2, \ldots, N_k)$, where $t$ is the
top-level operator, $N_1$, $N_2$, \ldots, $N_k$ are the child non-terminal symbols, and such that the output types of
$t$ and $N$ coincide: $T_t = T_N$, and
\item each rule $R$ is associated with a probabiltiy $p_R \in [0, 1]$ such that for all non-terminals $N$, $\sum_{R \in
\mathcal{R}(N)} p_R = 1$.
\end{enumerate}
Notice that $\mathcal{N}$ does not necessarily coincide with the set of types of the target language;
we only require that the mapping from non-terminals to types is surjective.
We make two additional requirements of the PCFGs we consider in this paper: first, we require the probability of each
rule, $p_R < 1$ (note the strict inequality), and second, we require that $G$ be \emph{unambiguous}, i.e. that every
expression produced by the grammar have a unique derivation tree. If we relax the first assumption, then there are some
technical difficulties in setting up the probability space, and results such as Theorem~\ref{thm:alg:Dijkstra} will need
additional constraints to be true. The second assumption simplifies the definition of $P(e)$, which would otherwise be
the sum of the probabilities over all derivation trees. Moreover, it reduces the search space to a tree (instead of a
DAG), and thereby allows us to use stronger versions of the guarantees provided by search algorithms.
For simplicity of notation, and whenever the intent is clear, we will conflate productions $e$ of the grammar with the
expressions $t_e$ that they encode. We write $\Expansions_N$ for the set of all productions of a non-terminal $N$. Given
a production $e \in \Expansions_N$, we define its probability, $P(e)$, as the product of the probabilities of all rules
used in $e$. It can be shown that $\sum_{e \in \Expansions_N} P(e) = 1$. The productions of $G$ will constitute elements
of the sample space under consideration, and the space of events is, as usual, the powerset $2^{\Expansions_N}$ of
productions. We will discuss the extraction of PCFGs from a code corpus in Section~\ref{sec:grammar-gen}.
\subsection{Top-down expression enumeration as graph search}
In the search phase of the CEGIS loop, we want to find an expression $t$ which satisfies the requirement in formula~%
\ref{alg:cegis:search}. One approach to implementing $\Search(\seqa, A, \pcname, \phi, x)$ is to enumerate all candidate
expressions, in some order, until a suitable answer is found. Recall that, because $A$ is a finite set, for a given
choice of $A$ and $t$, determining the satisfaction of formula~\ref{alg:cegis:search} reduces to evaluating the
conjunction, and does not involve expensive calls to an SMT solver. The enumerative SyGuS solver~\cite{Transit} uses a
highly optimized form of expression enumeration in the search phase, and is currently among the most competitive SyGuS
solvers. See Algorithm~\ref{alg:search}. In this section, we will show how we utilize a PCFG
to enumerate expressions in order of decreasing probability in order to accelerate the search
process. We will now describe this probabilistic enumeration algorithm.
\begin{algorithm}
\caption{$\Search(\seqa, A, \pcname, \phi, x)$. Implements the search phase of the CEGIS loop by expression
enumeration.}
\label{alg:search}
\begin{enumerate}
\item Let $G$ be the chosen PCFG, and $N$ be the starting non-terminal such that the types coincide, i.e. $T_N = T_x$.
\item For each $e$ emitted by $\Enumerate(G, N)$:
\begin{enumerate}
\item Let $t_e$ be the expression encoded by $e$.
\item If $\Land_{\seqa \in A} \pcname \implies \phi[x \mapsto t_e] = \lstinline|true|$, return $t_e$.
\item Otherwise, discard $t$ and continue enumeration.
\end{enumerate}
\end{enumerate}
\end{algorithm}
\paragraph{Partial productions.}
Engineering the enumeration algorithm turns out to be much simpler if we enumerate expressions top-down rather than
bottom-up. As a first step, we therefore extend the idea of a grammar production into the more general notion of a
\emph{partial} production:
\[
\begin{array}{rcl}
\tilde{e}_N & \Coloneqq & \lstinline|?|_N \mid
t(\tilde{e}_{N_1}, \tilde{e}_{N_2}, \ldots, \tilde{e}_{N_k}),
\end{array}
\]
where $R = t(N_1, N_2, \ldots, N_k)$ is a production rule of $N$ in $G$. In particular, notice that partial productions
may contain incomplete sub-expressions, denoted by $\lstinline|?|_N$. Examples of partial productions include
``\lstinline|x + ?|'' and ``\lstinline|if (x $\leq$ ?) { x } else { ? }|''. We write $\tilde{\Expansions}_N$ for the set
of all partial productions of a non-terminal $N$. We can also speak of the probability of partial productions:
$p_{\tilde{e}}$ of a partial production $\tilde{e}$ is the product of all production rules used to derive $\tilde{e}$.
It will be mathematically more convenient to speak of negative log probabilities: the \emph{cost} of a partial
production $\tilde{e}$,
\begin{equation}
\cost(\tilde{e}) = -\log(p_{\tilde{e}}) = -\sum_{R \in \tilde{e}} \log(p_R).
\label{eq:alg:costdefn}
\end{equation}
Given a pair of partial productions $\tilde{e}_1$ and $\tilde{e}_2$, we say that $\tilde{e}_1$ and $\tilde{e}_2$ are
related by the expansion relation, and write $\tilde{e}_1 \to \tilde{e}_2$, if $\tilde{e}_2$ can be obtained by
replacing the left-most instance of \lstinline|?| in $\tilde{e}_1$ by an appropriate production rule. For example, if
$\tilde{e}_1 = \lstinline|x + ?|$, $\tilde{e}_2 = \lstinline|x + 1|$, and $\tilde{e}_3 = \lstinline|x + (? + ?)|$, then
$\tilde{e}_1 \to \tilde{e}_2$ and $\tilde{e}_1 \to \tilde{e}_3$.
The expansion relation naturally induces a tree $\mathcal{G}$ on the partial productions of a grammar $G$. See Figure~%
\ref{fig:alg:graph}. The initial node is the starting non-terminal, \lstinline|?|. There is an edge from the partial
production $\tilde{e}_1$ to the partial production $\tilde{e}_2$ if they are related by the expansion relation,
$\tilde{e}_1 \to \tilde{e}_2$. The edge $\tilde{e}_1 \to \tilde{e}_2$, produced by an instantiation of the rule $R$, is
annotated with the cost of the rule, $-\log(p_R)$. Then, the resulting graph is a tree because of the unambiguity
assumption on $G$, and the sum of the weights along the path to $\tilde{e}$ is equal to $\cost(\tilde{e})$. The main
insight of this paper is that $\Enumerate(G, N)$ can be thought of as instantiations of various search algorithms on
this tree.
\begin{figure}
\begin{centering}
\begin{tikzpicture}[minimum size=0.5 cm]
\node [draw] (e1) at (-2, -1) {\lstinline|0|};
\node [draw, below=of e1] (e2) {\lstinline|1|};
\node [draw, below=of e2] (e3) {\lstinline|x|};
\node [right=3 of e1] (e4) {\lstinline|? + ?|};
\node [right=of e4] (e5) {\lstinline|if (?) { ? } else { ? }|};
\node (e0) at ($ (e1)!0.5!(e5) + (0, 2.5) $) {\lstinline|?|};
\path [->] (e0) edge [above, sloped] node {$-\log(0.15)$} (e1);
\path [->] (e0) edge [above, sloped] node {$-\log(0.3)$} (e2);
\path [->] (e0) edge [above, sloped] node {$-\log(0.3)$} (e3);
\path [->] (e0) edge [above, sloped] node {$-\log(0.15)$} (e4);
\path [->] (e0) edge [above, sloped] node {$-\log(0.1)$} (e5);
\node (e6) at ({$(e4) + (-1, 0)$} |- {$(e2) + (0, -0.7)$}) {\lstinline|x + ?|};
\path [->] (e4) edge [above, sloped] node {$-\log(0.3)$} (e6);
\node (e7) at ({$(e4) + (1, 0)$} |- {$(e2) + (0, -0.7)$}) {$\cdots$};
\path [->] (e4) edge (e7);
\node [draw] (e8) at ({$(e6) + (-1, 0)$} |- {$(e3) + (0, -1.4)$}) {\lstinline|x + 1|};
\path [->] (e6) edge [above, sloped] node {$-\log(0.3)$} (e8);
\node (e9) at ({$(e6) + (1, 0)$} |- {$(e3) + (0, -1.4)$}) {$\cdots$};
\path [->] (e6) edge (e9);
\node (e10) at ({$(e5) + (1, 0)$} |- e7) {\lstinline|if (? $\leq$ ?) { ? } else { ? }|};
\path [->] (e5) edge [above, sloped] node {$-\log(0.8)$} (e10);
\node [right=of e10] (e11) {$\cdots$};
\path [->] (e5) edge (e11);
\end{tikzpicture}
\end{centering}
\caption{The tree $\mathcal{G}$ of partial productions, connected by the expansion relation. The initial node is the
empty partial production \lstinline|?| of the non-terminal \lstinline|IntExpr|. Each edge indicates the replacement of
the left-most incomplete sub-production by an instance of a production rule $R$. The edges are annotated with the
negative log probabilities, $-\log(p_R)$: because $0 \leq p_R \leq 1$, $\log(p_R) \leq 0$, so the weights $-\log(p_R)$
are non-negative. A production is enclosed in a rectangle if it is complete.}
\label{fig:alg:graph}
\end{figure}
\begin{algorithm}
\caption{$\Enumerate(G, N)$. Emits a sequence of complete productions of $N$.}
\label{alg:enum}
\begin{enumerate}
\item Let $\pi : \tilde{\Expansions}_N \to \R^{\geq 0}$ be a priority function mapping productions to non-negative real
numbers.
\item Let $Q$ be a priority queue of partial productions $\tilde{e} \in \tilde{\Expansions}_N$ arranged in ascending
order according to $\pi$. Initialize $Q \coloneqq \{ \lstinline|?|_N \}$.
\item While $Q$ is not empty:
\begin{enumerate}
\item \label{enum:alg:enum:dequeue} Let $\tilde{e}$ be the element at the front of $Q$. Dequeue $\tilde{e}$.
\item If $\tilde{e}$ is a complete production, emit $\tilde{e}$.
\item Otherwise, for every neighbor $\tilde{e}'$ such that $\tilde{e} \to \tilde{e}'$ in $\mathcal{G}$, insert
$\tilde{e}'$ into $Q$.
\end{enumerate}
\end{enumerate}
\end{algorithm}
See Algorithm~\ref{alg:enum}. It is helpful to visualize the operation of the algorithm as shown in Figure~%
\ref{fig:alg:enum}. The algorithm maintains a priority queue $Q$ of still-unexpanded partial productions. At each step,
the algorithm dequeues the element $\tilde{e}$ at the front of this priority queue and, if it is still incomplete,
expands the leftmost \lstinline|?| with an instance of every applicable production rule $R$. This results in a set of
partial productions $\tilde{e}_1$, $\tilde{e}_2$, \ldots, $\tilde{e}_k$. All of these new partial productions are
inserted back into $Q$. If, on the other hand, $\tilde{e}$ is already complete, it is emitted to be further processed by
$\Search(\seqa, A, \pcname, \phi, x)$. If $\pi(\tilde{e}) = \cost(\tilde{e}) = -\log(P(\tilde{e}))$, i.e. the priority
function is equal to the cost of the partial production, then $Q$ forms the frontier maintained by Dijkstra's algorithm.
\begin{figure}
\resizebox{\textwidth}{!}{
\begin{tikzpicture}
\node (emit) {Emit!};
\node [minimum height=1cm, minimum width=3.5cm, draw, right=of emit] (completeCheck) {Complete?};
\path [->] (completeCheck) edge [above] node {Yes} (emit);
\node [draw, minimum size=1cm, right=of completeCheck] (e1) {$\tilde{e}_1$};
\node [draw, minimum size=1cm, right=0cm of e1] (e2) {$\tilde{e}_2$};
\node [draw, minimum size=1cm, right=0cm of e2] (e3) {$\tilde{e}_3$};
\draw (e3.north east) -- +(2, 0);
\draw (e3.south east) -- +(2, 0);
\path [->] (e1) edge (completeCheck);
\node (edots) at ($ (e3.east) + (1, 0) $) {$\cdots$};
\node [draw, minimum height=1cm, minimum width=3.5cm, below=of completeCheck] (peval) {Partial evaluation};
\path [->] (completeCheck) edge [left] node {No} (peval);
\node [draw, minimum height=1cm, align=left, right=1.2 of peval] (indist1) {Indist. \\ Rewrite 1};
\path [->] (peval) edge (indist1);
\node [draw, minimum height=1cm, right=1.2 of indist1] (expand) {Expand $\tilde{e}_1$};
\path [->] (indist1) edge (expand);
\node [right=1.2 of expand] (echildren) {$\{ \tilde{e}_{11}, \tilde{e}_{12}, \tilde{e}_{13}, \ldots \}$};
\path [->] (expand) edge (echildren);
\node [draw, minimum height=1cm, align=left] (indist2) at ( echildren |- edots) {Indist. \\ Rewrite 2};
\path [->] (echildren) edge (indist2);
\path [->] (indist2) edge ($ (e3.east) + (2, 0) $);
\draw [decorate, decoration={brace}] ($ (e1.north west) + (0, 0.2) $)
-- ($ (e3.north east) + (2, 0.2) $)
node [midway, above] {Deduplicating priority queue};
\end{tikzpicture}
}
\caption{The operation of $\Enumerate(G, N)$ in Algorithm~\ref{alg:enum}. We describe the partial evaluation and
indistinguishability-based optimizations, including production rewriting and queue deduplication, in
Section~\ref{sub:alg:optimizations}.}
\label{fig:alg:enum}
\end{figure}
We begin analysing Algorithm~\ref{alg:enum} with the following invariant: Let $\Expansions_c$ be the set of (complete)
productions emitted by $\Enumerate(G, N)$. Then, $\sum_{\tilde{e} \in Q \union \Expansions_c} P(\tilde{e}) = 1$. We will
now outline the proof of its invariance. First, observe that, for all partial productions $\tilde{e}$ with at least one
occurrence of \lstinline|?|,
\begin{equation}
\sum_{\tilde{e}' \text{s.t.} \tilde{e} \to \tilde{e}'} P(\tilde{e}') = P(\tilde{e}).
\label{eq:alg:inv:sum1}
\end{equation}
Informally, if each partial production $\tilde{e}$ is viewed as an event resulting from the PCFG, then the partial
productions $\tilde{e}'$ obtained by a single application of the expansion relation encode a mutually exclusive and
exhaustive collection of sub-events of $\tilde{e}$. We then have:
\begin{inparaenum}[(\itshape a\upshape)]
\item the base case: $P(\lstinline|?|) = 1$,
\item inductive case \#1: whenever a complete production $e$ is emitted from $Q$, the set $Q \union \Expansions_c$ is
left unchanged,
\item and inductive case \#2: whenever a partial production $\tilde{e}$ is further expanded into partial productions
$\tilde{e}_1$, $\tilde{e}_2$, \ldots, $\tilde{e}_k$, the candidate invariant continues to hold because of
Equation~\ref{eq:alg:inv:sum1}.
\end{inparaenum}
\begin{thm}
\label{thm:alg:dijkstra}
If the priority function, $\pi(\tilde{e}) = \cost(\tilde{e}) = -\log(P(\tilde{e}))$, then $\Enumerate(G, N)$ returns
some sequence $\sigma$ containing all productions of $N$ in $G$. Furthermore, the probability $P(e)$ of productions in
$\sigma$ is monotonically decreasing.
\end{thm}
\begin{proof}
We begin by proving the second part of the claim. Consider a pair of partial productions, $\tilde{e}$ and $\tilde{e}'$
such that $\tilde{e} \to \tilde{e}'$. Then, $P(\tilde{e}) \geq P(\tilde{e}')$. It therefore follows that the sequence of
partial productions, $\tau$, dequeued by $\Enumerate(G, N)$ in line~\ref{enum:alg:enum:dequeue} has monotonically
decreasing probability. Observe that the sequence $\sigma$ of complete productions emitted by $\Enumerate(G, N)$ is a
sub-sequence of $\tau$, and therefore, the probability of productions in $\sigma$ monotonically decreases.
Now, arbitrarily choose a complete production $e \in \Expansions_N$. We will show that $\Enumerate(G, N)$ will
eventually emit $e$. Let $p_{\text{hi}}$ be the highest probability of any production rule in $G$. Recall our
requirement that for each rule $R$, $p_R < 1$, and therefore, $p_{\text{hi}} < 1$ and $\log(p_{\text{hi}}) < 0$.
Before processing $e$, $\Enumerate(G, N)$ can only process those nodes which are at most $\log(P(e)) /
\log(p_{\text{hi}})$ steps away from the root node \lstinline|?| of the search tree $\mathcal{G}$. If there are $N$
rules in the PCFG, then the out-degree of each node is at most $N$, and it follows that there are only finitely many
nodes which can be processed before $\Enumerate(G, N)$ processes $e$. It follows that $e$ is eventually emitted by the
enumerator.
\end{proof}
Unfortunately, when the priority function is instantiated to emulate Dijkstra's algorithm, enumeration does not scale
beyond a few hundred productions. The intuition may be found in the correctness proof above: before emitting $e$, the
enumerator must explore all partial productions with probability bigger than $P(e)$. The number of these productions
rapidly grows with decreasing $P(e)$, and furthermore, most of the processed partial productions are ``very
incomplete'', i.e. containing many instances of \lstinline|?|, and therefore many edges away from turning into complete
productions. This suggests other graph search algorithms, which we will now discuss.
\paragraph{Non-terminal horizons and A* search.}
When considering partial productions $\tilde{e}$, we may speak of two quantities:
\begin{inparaenum}[(\itshape a\upshape)]
\item the cost \emph{already paid}, which we have defined as $\cost(\tilde{e}) = -\log(P(\tilde{e}))$, and
\item the minimum cost \emph{yet to be paid}, before $\tilde{e}$ turns into a complete production.
\end{inparaenum}
We formalize this latter quantity as the \emph{horizon}, defined as:
\begin{equation}
\horizon(\tilde{e}) = \min_{e_f \text{ s.t. } \tilde{e} \to^* e_f} \cost(e_f) - \cost(\tilde{e}),
\label{eq:alg:horizon}
\end{equation}
where $e_f$ ranges over all complete productions reachable from $\tilde{e}$. Recall the intuition for the failure of
Dijkstra's algorithm: before emitting $e$, we have to process every partial production with higher probability, and most
such partial productions $\tilde{e}'$ are themselves many steps away from complete productions. Since the horizon
encodes the distance from the partial production to its nearest completion, it is natural to include it in the priority
function $\pi(\tilde{e})$. If we define $\pi(\tilde{e}) = \cost(\tilde{e}) + \horizon(\tilde{e})$, then we obtain A*
search.
The important property of $\pi(\tilde{e})$ is that it is \emph{admissible}, i.e. that $\pi(\tilde{e})$ is always less
than the cost of all complete productions descendant from $\tilde{e})$. More formally, for all complete productions
$e_c$ reachable from $\tilde{e}$,
\begin{equation}
\pi(\tilde{e}) = \min_{e_c \text{s.t.} \tilde{e} \to^* e_c} \cost(e_c).
\label{eq:alg:astar}
\end{equation}
We then have the following well-known result of the A* algorithm:
\begin{thm}
\label{thm:alg:astar}
If $\pi(\tilde{e}) = \cost(\tilde{e}) + \horizon(\tilde{e})$, then:
\begin{enumerate}
\item The sequence of complete productions emitted by $\Enumerate(G, N)$, $\sigma = e_1, e_2, e_3, \ldots$ contains all
productions in $\Expansions_N$ and has monotonically decreasing probability. (This is called the \emph{optimality}
property.)
\item For all complete productions $e$, every complete optimal algorithm $A'(G, N)$ will, before producing $e$, visit
all strictly less expensive nodes, $\tilde{e}$, such that $\pi(\tilde{e}) < \pi(e)$ . (This is the property of
\emph{optimal efficiency}.)
\end{enumerate}
\end{thm}
\begin{proof}
First, we show that $\Enumerate(G, N)$ produces all complete productions $e \in \Expansions_N$. Similar to Theorem~%
\ref{thm:alg:dijkstra}, we bound the number of nodes that can be processed in line~\ref{enum:alg:enum:dequeue} before
processing $e$. Observe that for all nodes, $\tilde{e}$, $\pi(\tilde{e}) \geq \cost(\tilde{e})$, and for complete
productions $e$, $\pi(e) = \cost(e)$. Therefore, all nodes processed by $\Enumerate(G, N)$ before processing $e$ are
no more than $\log(P(e)) / \log(p_{\text{hi}})$ steps from the root of the search tree $\mathcal{G}$, where, as before,
$p_{\text{hi}} < 1$ is the highest probability appearing in the PCFG.
We will now show that the productions emitted have monotonically decreasing probability. For this, first observe that
for each complete production $e$, and for each of its (necessarily incomplete) ancestors $\tilde{e}'$, $\pi(\tilde{e}')
\leq \pi(e)$. For the sake of contradiction, let there be some pair $e_1$, $e_2$ of complete productions, such that
$P(e_1) < P(e_2)$, but $\Enumerate(G, N)$ emits $e_1$ before $e_2$. Both $e_1$ and $e_2$ are therefore reachable from
the initial node \lstinline|?|. Consider the state of $Q$ when $e_1$ is dequeued from it. It has to be the case that
some ancestor $\tilde{e}'_2$ is present in $Q$. However, it follows that $\pi(\tilde{e}'_2) \leq \pi(e_2) < \pi(e_1)$,
and therefore the priority queue must have made a mistake in its ordering. It follows that the sequence of productions
emitted by $\Enumerate(G, N)$ has monotonically increasing costs, or equivalently, monotonically decreasing
probabilities.
We will now prove the last part of the theorem. Assume otherwise, so for some algorithm $A'(G, N)$ and some node
$\tilde{e}$ such that such that $\pi(\tilde{e}) < \pi(e)$, $A'$ does not expand $\tilde{e}$ before producing $e$.
We know $\tilde{e}$ has some descendant $e'$ such that $\pi(\tilde{e}) = \cost(e') < \pi(e)$. Because the search space
$\mathcal{G}$ is a tree, if $A'$ does not expand $\tilde{e}$, it follows that it did not enumerate $e'$ before $e$,
violating the assumption that $A'$ is a complete optimal enumerator.
\end{proof}
Note that, in a certain sense, the second part of Theorem~\ref{thm:alg:astar} says that this is the best possible
enumeration algorithm among all algorithms that use the same priority function, because $\Enumerate(G, N)$ does not
expand any nodes $\tilde{e}'$ such that $\pi(\tilde{e}') > \pi(e)$. This also captures our intuition for the superiority
of this priority function over the simpler metric used by Dijkstra's algorithm.
\paragraph{Computing the horizon, $\horizon(\tilde{e})$.}
It can be shown that the cost to be paid to complete $\tilde{e}$ is the sum of the costs needed to expand each of its
unexpanded nodes:
\begin{equation}
\horizon(\tilde{e}) = \sum_{\lstinline|?|_N \in \tilde{e}} \horizon(\lstinline|?|_N).
\end{equation}
Before starting the enumerator, we therefore compute the non-terminal horizons, $\horizon(\lstinline|?|_N)$, for each
non-terminal symbol $N$. Observe that, by definition, this simply encodes the probability of the most likely expansion,
$e \in \Expansions_N$:
\[
\horizon(\lstinline|?|_N) = \min_{e \in \Expansions_N} \cost(e) = -\max_{e \in \Expansions_N} \log(P(e)).
\]
The values $\horizon(\lstinline|?|_N)$ can be computed by the following fixpoint computation:
\begin{enumerate}
\item $\horizon^0(\lstinline|?|_N) = \min \{ -\log(p_R) \mid \text{Terminal rule } R \}$.
\item $\horizon^{i + 1}(\lstinline|?|_N) = \min \{ \horizon^i(R) \mid \text{Production rule } R \}$, where the horizon
of a production rule $R$ is defined as $\horizon^i(R) = -\log(p_R) + \sum \{ \horizon^i(\lstinline|?|_{N'}) \mid
\text{child non-terminal } N' \text{ appearing in } R \}$.
\end{enumerate}
\subsection{Optimizations}
\label{sub:alg:optimizations}
\paragraph{Eagerly discarding partial productions.}
Given a single input variable \lstinline|a|, consider the synthesis predicate $\phi$ given by:
\begin{lstlisting}
if (a == 5) { x == 6 }
else if (a == 7) { x == 9 }
else { x == a }
\end{lstlisting}
During CEGIS, say the set of concrete input points, $A = \{ 2, 5, 7 \}$. For this problem, $\Enumerate(G, N)$ will
consider partial productions such as
\[
\tilde{e} = \lstinline|if (x < 6) { x } else { ? }|.
\]
Observe that the conjunction $\Land_{\lstinline|a| \in A} \phi[\lstinline|x| \mapsto e]$ used in the CEGIS loop can only
be evaluated for complete productions $e$. However the partial production $\tilde{e}$ is already incorrect: for the
input point $\lstinline|a| = \lstinline|5|$, regardless of the completion of \lstinline|?|, $\tilde{e}$ will evaluate to
\lstinline|5|, and this fails the requirement that \lstinline|x == 6|.
This pattern of partial productions already failing requirements is particularly common with conditionals and match
statements. Our first optimization is therefore the partial evaluation box shown in Figure~\ref{fig:alg:enum}. For each
input point $\seqa \in A$, we partially evaluate the CEGIS predicate $\pcname \implies \phi[x \mapsto \tilde{e}]$. If
the result is \lstinline|false|, then we discard the partial production $\tilde{e}$ without processing, as we know that
all its descendants will fail the synthesis predicate. Otherwise, if it evaluates to \lstinline|true| or unknown, then
we process $\tilde{e}$ as usual, and insert its neighbors back into the priority queue.
\paragraph{Extending the priority function with scores.}
We just observed that if a partial production fails the synthesis predicate by partial evaluation, then it can be
discarded without affecting correctness. The dual heuristic optimization is to \emph{promote} partial productions which
definitely evaluate to \lstinline|true| on some input points. We do this by modifying the priority function:
\begin{equation}
\pi(\tilde{e}) = \cost(\tilde{e}) + \horizon(\tilde{e}) + c \times \log(\score(\tilde{e})),
\end{equation}
where $\score(\tilde{e})$ is the number of input points on which the partial production evaluates to \lstinline|true|,
and $c$ is a positive coefficient. Small positive values of $c$ results in an enumerator which strictly follows
probabilistic enumeration, while large positive values of $c$ results in the enumerator favoring partial productions
which already work on many points. While the use of this heuristic renders Theorem~\ref{thm:alg:astar} inapplicable, we
have observed improvements in performance as a result of this optimization.
\paragraph{Indistinguishability.}
Consider the partial production $\tilde{e} = \lstinline|if (x < 5) { x - x } else { ? }|$, and focus on the
sub-expression \lstinline|x - x|. This expression is identically equal to \lstinline|0|, and it is therefore possible
to simplify $\tilde{e}$ to \lstinline|if (x < 5) { 0 } else { ? }|. If two expressions $e$ and $e'$ are equivalent, and
$P(e) > P(e')$, then a desirable optimization is to never enumerate $e'$. Observe that, if properly implemented, this
optimization produces exponential savings at each step of the search: if $e$ and $e'$ are equivalent, then, for example,
it follows that both $e + 3$ and $e' + 3$ are equivalent, and only one of them needs to be enumerated to achieve
completeness.
Indistinguishability~\cite{Transit} is a technique to mechanize this reasoning. Given a sequence of concrete input
points $A = \{ \seqa_1, \seqa_2, \ldots, \seqa_n \}$, a complete production $e$ can be evaluated to produce a set of
output values $S_e = \{ x_1, x_2, \ldots, x_n \}$. During enumeration, if an expression $e'$ is encountered such that
for some previous expression $e$, $S_e = S_{e'}$, then $e'$ is discarded as a potential expansion. The original
expression $e$ can thus be regarded as the representative of the equivalence class of all expressions whose signature is
equal to $S_e$.
In this original formulation of indistinguishability-based pruning, the enumerator worked bottom-up, rather than in the
top-down style we consider in this paper. As a result, the enumerator only needs to process complete productions, rather
than the partial productions which populate $Q$ in our setting.
We have extended this optimization to work with partial expressions, and prune expressions in a top-down enumerator. The
idea is to replace the priority queue $Q$ with a ``deduplicating priority queue'', with the additional feature that it
removes duplicate elements from being dequeued. Such a queue can be implemented as a combination of a traditional
priority queue and a mutable set data structure commonly available in standard language libraries. Next, we maintain a
dictionary mapping all previously seen signatures to the representatives of the respective equivalence classes. Every
time we dequeue a partial production $\tilde{e}$ from $Q$, we evaluate every complete sub-production $e_{\text{sub}}$ on
all input points $\seqa \in A$ to obtain its signature $S_{\text{sub}}$, and replace $e_{\text{sub}}$ with the canonical
representative. This forms the box labelled ``Indist. rewrite 1'' in Figure~\ref{fig:alg:enum}.
We also perform a second, fast indistinguishability rewrite after each partial production is expanded. This is done by
maintaining a map from previously seen expressions to their canonical representatives. Consulting this map for every
complete sub-production of the children, $\tilde{e}_1$, $\tilde{e}_2$, \ldots, of the original production $\tilde{e}$ is
much faster than evaluating them on each concrete input point in $A$. There is some freedom in choosing the placement of
various optimizations in Figure~\ref{fig:alg:enum}: our motivation was that the priority queue can get very large, and
many elements enqueued into the queue will never be dequeued, and that it is therefore wise to postpone as much
processing as possible to when partial productions are extracted from $Q$.
|
\section{Introduction}
When learning a policy for playing a game, it helps immensely if the representation of the game state (the input to the controller) is low-dimensional and structured in such a way that it can easily be processed by the controller. However, if the controller does not have access to the internal state of the game, it is forced to rely on the same information a human would have when playing the game, primarily the visual feed. This makes the learning task much harder, as we need to learn to transform the high-dimensional visual input to information that the controller can act on while learning the actual policy.
In general, working in three-dimensional environments, with solely visual input and no easy backend into a high-level representation of the game state, brings us closer to non-game (``real-world'') applications of these AI methods, such as robot control.
This is seen by many as a more relevant task for testing general intelligence in AI agents, which is why we have seen much work on learning from visual input recently.
So far, methods such as Deep-Q learning~\cite{MnihDeepQ} have seen success while working with large-scale visual input, while we have not seen successful applications of evolution to large-scale
visual problems, though it has been suggested that evolutionary methods may provide better results in partially-observable environments~\cite{togelius2009ontogenetic}. In other contexts, evolutionary algorithms have proven effective in generating interesting, unique strategies to succeed in difficult environments, and are capable of coming up with multiple solutions to the same problem; they have also performed well on tasks with delayed rewards and imperfect information~\cite{risi2015neuroevolution}. However, usually such algorithms require a pre-processed, high level representation of the environment, since raw visual data is too high-dimensional for evolutionary algorithms to effectively improve. (the larger the input, the larger the networks to process them must be, and the larger the networks, the larger the genome, and the larger the genome, the slower the evolution). Gradient-based deep learning methods, however, excel at compressing high-dimensional data and extracting relevant high-level, low-dimensional representations~\cite{compressionbasics}. Using these methods to create a compressed representation of the visual data of an environment, and then training a smaller network off of this compressed representation to generate agent behavior, could make it possible to enjoy the advantages of evolutionary algorithms when working with raw visual information.
In this paper, an autoencoder - a particular network architecture described in further detail below - is used to create a compressed representation of a a three-dimensional video game environment, namely Doom. This compressed representation is then fed into a separate, smaller behavior-generating neural network, optimized via evolutionary methods, to demonstrate that the level of compression achieved by the autoencoder is sufficient to enable evolutionary optimization. It should be noted, however, that evolutionary methods are not necessarily vital to this method. For example, evolutionary optimization could be switched out for Deep-Q learning off of the compressed inputs.
\subsection{Autoencoders}
An autoencoder is a network composed of three basic parts: The encoder, the chokepoint, and the decoder~\cite{Autoencoderfeatureextraction}. These networks are trained to reproduce their own input as their output, the idea being that the chokepoint will force the network to learn patterns in the data that allow it to be compressed to fit the chokepoint---because the decoder will have to reconstruct the input, after all, from all the information that something as small as the chokepoint can represent.
In a videogame environment, we can usually depend on a certain uniformity in the environment; different frames look somewhat like each other (the environment will usually be constructed of similar elements, just viewed from different angles). If an autoencoder can successfully encode a single image frame from the environment, perhaps it could generate a diversity of representations of any image frame from the environment, so that it could represent the entire environment in a highly compressed manner. And if we're able to train an autoencoder to compress all frames from a video game environment (even if the decoder's reconstruction of these images is foggy and incomplete), then we could train a smaller network off of the autoencoder's chokepoint compression of these frames, and use it to generate behavior in the gameplay environment. If the resulting chokepoint was low-dimensional enough, we could use standard neuroevolutionary techniques to optimize the weights for this smaller network to play the game well.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{autoencoder_diagram.png}
\caption{Simple autoencoder diagram}
\end{center}
\end{figure}
A simple, high-level way to think of this is that the autoencoder will be converting the raw pixel data from game images into a high-level representation of the environment, like having a picture of how the world looks in your head. Even if the picture is fuzzy or incomplete, it's easier to make decisions based off that small picture rather than by simply considering the large arrays of raw visual information.
\begin{figure*}
\begin{center}
\includegraphics[width = \textwidth]{process_diagram.png}
\caption{Simple diagram of the method used in this paper. Image from the game environment is fed into the autoencoder. The autoencoder generates two outputs: The lower-dimensional representation of the image from the chokepoint, and the reconstructed image at the output layer. The lower-dimensional representation of the image is fed into the behavior-generating network, which generates an action in response to take within the game environment. Once the action step is taken in the game environment, a new image is generated, which is fed back into the autoencoder to generate the next action. In addition, each image input into the autoencoder is compared to its own reconstruction from the output layer. If the reconstruction of the input image isn't very good, the image is added to the training set for the autoencoder. Training of the autoencoder takes place between generations.}
\label{fig:Process Diagram}
\end{center}
\end{figure*}
\section{Previous work}
The idea of compressing visual input so that evolutionary algorithms may be used has been demonstrated before in fairly simple two-dimensional environments~\cite{MountainCar}. In addition, it has been previously shown that autoencoders can reduce high-dimensional visual input to a very small representation at its chokepoint, reducing a simple visual environment down to as small as two values representing the x and y position of the agent in the environment~\cite{Maron}. Autoencoder compression of the environment has been shown to be successful when combined with Q-learning, as well, for successful results in simple Atari environments before the current popularity of Deep-Q learning\cite{Lange}, as well as on simple real-world applications such a slot car control~\cite{lange2012autonomous}. Work has also been done in evolving neural network agents in a more complex 3-dimension FPS environments such as Quake by Parker et al~\cite{parker2009backpropagation}, though this was done by first preprocessing the raw visual input into a much smaller array of pixel values by averaging the value of pixels over a pre-defined block.
A similar method to the one presented in this paper was effective in the TORCS (The Open Racing Car Simulator) environment, demonstrated by Koutnik et al.~\cite{KoutnikTORCS}. However, rather than training their compressor through backpropagation (as in this paper), the weights for the compressing network were evolved to maximize diversity in the feature vector representation of the environment. Success was demonstrated in generating an agent that could remain on a track in a car-driving environment.
While most work with vision-based AI training has taken place in simpler, two-dimensional environments, there have been forays into three-dimensional games, particularly in the Doom environment, using Deep-Q learning methods~\cite{deepQDoom,SLAMdoom,IntelDoom}. Common to all these approaches is that they supplement visual information about the environment with other variables taken from a backend into the game, whether it is information about health, ammo, a depth buffer, etc. While these methods have achieved impressive results, information such as this is not necessarily always something that would be available to a real-world controller.
Outside of the realm of generating agents to play in game environments, combining autoencoders with neuroevolution has been shown to generate interesting results with content creation by Liapis et al with their system DeLeNoX~\cite{liapis2013transforming}. Autoencoders were used to characterize two-dimensional spaceship designs as a low-dimensional array, which was used to refine the exploration of a possibility space for designs of believable-looking spaceships. In addition, the hierarchal representational capabilities of classifier networks have been used in cooperation with neuroevolution to provide a measure of fitness when generating new, unique images, evolving networks to generate images that match the classifier's labels, resulting in interesting representations of various classes that are completely artificially generated~\cite{nguyen2015innovation}.
Evolution strategies are a class of evolutionary algorithms that are particularly well suited to optimization in real-valued space. The weights of a neural network can be seen as a real-valued vector, and optimized with evolution strategies or other similar algorithms; the training of neural networks with evolutionary algorithms is often called neuroevolution~\cite{floreano2008neuroevolution,risi2015neuroevolution}. Neuroevolution has long been known to be an effective reinforcement learning method, in particular when training comparatively small networks~\cite{igel2003neuroevolution,gomez2006efficient}. Methods for scaling up neuroevolution to apply them to larger networks have been developed, focusing on encoding or compressing the genome of the network in lower-dimensional space~\cite{koutnik2013evolving,stanley2009hypercube}. Recently, neuroevolution has also been shown to be a highly scalable alternative to more mainstream reinforcement learning methods when training networks for behavior in high-dimensional environments such as Atari games, achieving competitive results when compared to methods such as Deep-Q learning~\cite{Largescaleevolution}. In addition, it was demonstrated that evolution strategies can also be successful in constructing the topology of deep neural nets as well, with an extension of the popular NEAT~\cite{Stanley} (Neuroevolution of Augmenting Topologies) method into the realm of deep learning~\cite{CoDeepNeat}.
This work distinguishes itself by taking raw visual input, with no preprocessing, and using an autoencoder trained via backpropagation to compress this input to a size more manageable for evolutionary methods. The training set of images for the autoencoder is obtained while playing the game with evolved agents that require no manual coding for the environment, minimizing the amount of necessary human input into the system.
\section{Method}
Our experiments were carried out using \emph{VizDoom}~\cite{VizDoom}, a platform that supports agents in various environments based on the popular 1993 "Doom" video game. VizDoom was chosen because it is a popular implementation of the FPS Doom, a three-dimensional game that provides arrays of RGB visual information every frame, and which implements several different scenarios that concentrate on different goals, such as health-gathering, monster-killing, or navigation through the environment. The autoencoder was created using \emph{Keras}, a high-level deep learning package for python that can run off of either theano or tensorflow. Weights for the behavior-generating network were evolved using a python implementation of CMA-ES~\cite{CMAES,hansen2001completely} (Covariance Matrix Adaptation Evolution Strategy). CMA-ES was chosen as a popular, readily available evolutionary method with broad application, and the purpose of evolution was to demonstrate that the autoencoder had sufficiently compressed the visual input so that evolutionary optimization could be possible. Other evolutionary methods could be substituted in to evolve the behavior-generating network.
In this paper, experiments were run in the VizDoom "health gathering" environment. In this environment, the player is in a room that is filled with acid, slowly deteriorating their health, and must move about to gather health packs to prevent themselves from dying. Small red jars act as mines, and if picked up, will damage the player. The objective is to survive for as long as possible. Three actions are available: turn left, turn right, and move forward. There is a certain element of luck to the environment: health packs spawn randomly, and the player is spawned at random locations in the map.
To generate behavior, an autoencoder is trained to reproduce input images of the environment (environment represented by 120x160x3 RGB images.) Table 1 shows the topology of the autoencoder network. All layers used ReLU (rectified linear unit) activation. ReLU is a simple activation function, where ReLU(x) = x if x is greater than 0, and 0 if x is less than zero. A small fixed-topology
network is then fed the input from the autoencoder's chokepoint (fc2 in the table, 128 float values). All layers in the behavior-generating network used sigmoid activation.
The behavior network was asked for a behavior decision every five frames of gameplay. In the final layer of the behavior network, the first three values were used to indicate which actions to take, with a sigmoid activation of greater than 0.5 being used to indicate that the action should be taken. The final value was multiplied by 5 and rounded up to the nearest integer to indicate how many times the action should be repeated. So, for example, if the network was asked to generate an action, and returned the array [$0.2$, $0.7$, $0.7$, $0.6$], it would repeat action 1 and action 2 (assuming that the first action, represented here by the value 0.2, was referred to as action 0) three times, and then do nothing for the remaining two frames before it was asked for a behavior decision again. There were 2208 weights total being optimized in the behavior-generating network.
\begin{table}[h]
\label{Autoencoder}
\begin{center}
\begin{tabular}{|c||c||c||c||c||c|}
\hline
Layer & Filter dim. & Stride & Filters & Input & Output\\
\hline
conv1 & 8x8x3 & 4 & 32 & 120x160x3 &30x40x32\\
\hline
conv2 & 4x4 & 3 & 64 & 30x40x32 &10x14x64\\
\hline
conv3 & 4x4 & 3 & 64 &10x14x64&4x5x64\\
\hline
fc1 & & & & 1280 & 512\\
\hline
fc2 & & & & 512 & 128\\
\hline
fc3 & & & & 128 & 512\\
\hline
fc4 & & & & 512 & 1024\\
\hline
fc5 & & & & 1024 & 120x160x3\\
\hline
\end{tabular}
\end{center}
\caption{Topology of the autoencoder network. Each row represents a layer in the network. The outputs from one layer feed into the inputs of the next. The first layer is fed the 120x160x3 array of pixel information representing the game environment. \emph{fc} refers to 'fully connected' layers, \emph{conv} refers to two-dimensional convolutional layers.}
\end{table}
\begin{table}[h]
\label{behavior network}
\begin{center}
\begin{tabular}{|c||c||c|}
\hline
Layer & Input & Output\\
\hline
fc1 & 128 & 16\\
\hline
fc2 & 16 & 8\\
\hline
fc3 & 8 & 4\\
\hline
\end{tabular}
\end{center}
\caption{Topology of the behavior-generating network. Table follows similar conventions to Table 1. The first layer is fed the 128 float value representation of the environment from the chokepoint of the autoencoder network.}
\end{table}
The autoencoder network was trained through backpropagation on the images generated by gameplay. The behavior-generating network was trained by evolutionary methods, specifically CMA-ES, as mentioned previously. For the first 30 generations, the behavior-generating network was rewarded for maximizing novelty, where novelty was defined as seeing images that the autoencoder could not accurately reproduce, by comparing the reproduced images to the actual images taken fromt he environment. More specifically, it was calculated by taking the average of the absolute difference between the two image's pixel values. This was done to ensure that, through gameplay in the early stages of training, the images delivered to the autoencoder for training by backpropagation would be a diverse representation of the environment, training the autoencoder to reproduce as much of the environment as possible.
After approximately 1,000,000 frames of gameplay, the autoencoder could reconstruct the environment to a sufficient degree that gameplay decisions could be made from the chokepoint values, so after 30 generations, the fitness function for the behavior-generating network was switched to reward it for actual fitness in gameplay (the number of frames the agent manages to survive in the hostile environment). The autoencoder continues to be trained by backpropagation along with the behavior-generating network after the first 30 generations. The training set for the autoencoder was generated from the actions the behavior-generating network took in the game environment. A simple filtering process was done on these images: Reconstructed images were compared to their input images (by calculating a simple absolute difference between their values) and if the average difference of the pixel values was below a threshold (.05) then the image was not included as part of the training set for the autoencoder to train on. Or, in other words, if the autoencoder already did a good job of reconstructing the input image, that image was excluded from the training set.
\section{Environment reconstruction and performance results}
\begin{figure*}[h]
\begin{center}
\includegraphics[width = \textwidth]{NetworksABC.png}
\caption{Evolutionary progress of agents trained off of alternative autoencoder structure chokepoint output. Left: Network A. Center: Network B. Right: Network C.}
\label{fig:NetworksDE}
\end{center}
\end{figure*}
\subsection{Autoencoder compression of environment}
Below are a couple of images demonstrating the quality achieved by the autoencoder in reconstructing the environment from the chokepoint input. Both of these images are taken from networks that have been trained for at least 300 generations.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{gen_330_medkit_reproduction.png}
\caption{Sample of autoencoder reconstruction of input image, generation 330}
\label{fig:normal medkit detection}
\end{center}
\end{figure}
In Figure~\ref{fig:normal medkit detection}, the original input image is on the left, and the autoencoder reconstruction of the input image is on the right. You can see the network does a decent job of reproducing the walls, floor, and multiple medkits at varying distances - so it would seem that the 128 floating points at the chokepoint can contain enough information to fairly accurately reproduce the images they are fed (composed of 57600 pixel values) even if the images are a bit fuzzy and ghostly.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{gen_312_mine_detection.png}
\caption{Sample of autoencoder reconstruction of input image, generation 312}
\label{fig:normal mine detection}
\end{center}
\end{figure}
Figure~\ref{fig:normal mine detection} demonstrates a weakness of the autoencoder's reconstruction of the environment. It has difficulty properly reproducing the red jars. As mentioned above, the red jars act as a sort of mines - if they are picked up they damage the player. In this image, you can see the red mine closer to the player does seem to be detected by the network - but when reproducing it, it almost seems as if the network is trying to reproduce it as a health pack in the reconstructed image. And the red jar in the background, much further away, doesn't seem to be reproduced by the network at all, even when health packs the same distance away are. The simplest explanation for this is that the red jars are simply seen much less often than the health packs are, so the network has fewer examples to learn from to reproduce them properly.
\subsection{Performance summary}
All of the following networks refer to behavior-generating networks, which use the CMA-ES evolutionary algorithm to optimize their weights. All of these networks (where applicable) used a compressed representation of the environment provided by the same autoencoder, which was trained via backpropagation. So what we have here are multiple behavior-generating networks that were trained off of input from the same autoencoder providing compressed representations of the environment.
A couple of different trials were run as a baseline.
First, a network that was evolved off of input from the autoencoder (identical to Network A, described below) was fed random input. We will call this Network 1. (the idea being that if Network A showed no difference between being fed random input or input from the autoencoder, then obviously it was not successfully gaining useful information from the compressed input from the autoencoder.)
In addition, a behavior-generating network was evolved off random input. We will call this Network 2. (the idea being that if Network 2 reached the same level of performance while being evolved off of random input as networks that were evolved off of autoencoder input, then obviously the networks were not necessarily gaining any benefit by using the autoencoder input.)
There were a few different networks evolved off of autoencoder input:
Network A, whose fitness function was a simple average of its score over 10 games,
Network B, which was fed an additional piece of information in addition to the autoencoder's chokepoint compression - the player's current health, normalized over a value of 0-1.
Network C, which operated in an environment where the red jars were much more deadly (usually resulting in instant death if they were picked up.) This was done to see if the autoencoder could actually distinguish between health kits and mines.
All scores in the below table were calculated over 1000 games, for networks evolved for 400 generations. 'Solved' is defined as a score of 2000 (at which point the environment times out.) 'Good' is defined as a score of 1000-2000. 'Mediocre' is defined as a score of 500-1000. 'Bad' is defined as a score of 0-500 (and is defined because this 'Bad' score basically falls into the range of what we'd expect from a random network.) The score is defined as the number of frames an agent survives within the environment.
\begin{table}[h]
\caption{Performance Results}
\label{performance}
\begin{center}
\begin{tabular}{|c||c||c||c||c||c||c|}
\hline
Network & Mean & std dev & Solved & Good & Mediocre & Bad\\
\hline
1 & 376.16 & 151.95 & 0 & 1 & 113 & 886\\
\hline
2 & 362.7 & 139.69 & 0 & 0 & 108 & 892\\
\hline
A & 515.97 & 288.8 & 1 & 64 & 328 & 607 \\
\hline
B & 466.2 & 265.2 & 1 & 48 & 242 & 709\\
\hline
C & 563.3 & 329.8 & 3 & 74 & 360 & 563 \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Alternative Convolution}
\begin{figure*}[h!]
\begin{center}
\includegraphics[width = \textwidth]{NetworksDE.png}
\caption{Evolutionary progress of agents trained off of alternative autoencoder structure chokepoint output. Left: Network D. Right: Network E.}
\label{fig:NetworksDE}
\end{center}
\end{figure*}
While training the autoencoder to compress the environment, an alternative topology was accidentally stumbled upon that yielded some interesting results.
As described in Section 3, in the first layer, 8x8x3 filters were used on the image on the first layer. This is typically how 2D convolutional nets are used on color images - convolved over the image, looking for patterns in all three RGB channels.
During a couple of the experiments being run, the images were fed into the autoencoder incorrectly. Specifically, the width value was put in place of the channel value. The specific effect this had was, instead of the first convolutional layer being constructed from 8x8x3 filters an 8x8 filter over the 'depth' of the three RGB channels), it was constructed from 3x8x160 filters (so each filter was looking for patterns in a very large space of 3840 pixel values. In other words, the width value was being treated as the channel value, so we had 3x8 size filters convolved over a 'depth' of 160 pixels.)
\begin{table}[h]
\caption{Autoencoder alternative Structure}
\label{Autoencoder}
\begin{center}
\begin{tabular}{|c||c||c||c||c||c|}
\hline
Layer & Filter dim. & Stride & Filters & Input & Output\\
\hline
conv1 & 3x8x160 & 4 & 64 & 120x160x3 &1x30x64\\
\hline
conv2 & 1x4 & 2 & 128 & 1x30x64 &1x15x128\\
\hline
conv3 & 1x4 & 2 & 256 &1x15x128 &1x8x256\\
\hline
fc1 & & & & 2048 & 512\\
\hline
fc2 & & & & 512 & 128\\
\hline
fc3 & & & & 128 & 512\\
\hline
fc4 & & & & 512 & 1024\\
\hline
fc5 & & & & 1024 & 120x160x3\\
\hline
\end{tabular}
\end{center}
\end{table}
What was interesting about this mistake was not only did this odd network architecture manage to reconstruct the environment fairly well, it actually seemed to give better results when it came to evolving behavior networks.
\subsection{Alternative Autoencoder Compression of the Environment}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{exotic_network_gen_374_healthkit_reproduction.png}
\caption{Sample of alternative autoencoder reconstruction of input image, generation 374}
\label{fig:exotic medkit detection}
\end{center}
\end{figure}
Here, in Figure~\ref{fig:exotic medkit detection}, we see the alternative autoencoder's reconstruction of the gameplay environment, including a couple of health packs at varying distances. Again, original image is on the left, reconstruction is on the right. Compare to figure 2: we can see that this alternative structure is capable capturing the walls, floor, and the health packs. However, the reconstruction of the health packs is a bit less detailed than they are in the normal autoencoder's reconstruction in Figure~\ref{fig:normal medkit detection}, mostly being presented as white blobs. Nonetheless, they're still there, so the behavior generating network should be able to make decisions taking them into account.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{exotic_network_gen_418_mine_reconstruction_failure.png}
\caption{Sample of alternative autoencoder reconstruction of input image, generation 418}
\label{fig:exotic mine detection}
\end{center}
\end{figure}
On the other hand, in Figure~\ref{fig:exotic mine detection}, we can see that the alternative autoencoder structure actually fails to reproduce the red jars at all. Generally, this particular network topology could not reproduce the red jars unless they were very close to the screen, and then only very faintly. This could have implications for performance - if the network cannot reproduce them, they may not be reflected at all in the chokepoint compression of the environment, and the network may not be able to make decisions taking their existence into account.
\subsection{Performance Summary}
As with the normal autoencoder structure detailed in section IV, Network 1 in the table below is a baseline network, detailing the results of a behavior generating network that was trained on autoencoder input when it was fed random input. Network 2 is the results of a network trained to optimize on random input. Network D is a behavior generating network trained on top of the alternative autoencoder compression whose fitness function is average performance over 10 games. Because Network D had such a high standard deviation, Network E was also trained, whose fitness function can be described by the following equation:
$$
5\alpha/3 - \beta
$$
Where alpha is the network's score over 10 games, and beta is the standard deviation over those 10 games (so the network was punished for inconsistent performance).
\begin{table}[h]
\caption{Alternative Performance Results}
\label{altperformance}
\begin{center}
\begin{tabular}{|c||c||c||c||c||c||c|}
\hline
Network & Mean & std. dev & Solved & Good & Mediocre & Bad\\
\hline
1 & 369.19 & 147.93 & 0 & 0 & 110 & 890\\
\hline
2 & 364.2 & 137.97 & 0 & 1 & 103 & 896 \\
\hline
D & 646.58 & 397.93 & 12 & 162 & 360 & 466\\
\hline
E & 657.1 & 397.05 & 12 & 170 & 374 & 444\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Video Results and Behavior Discussion}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{video_clip_example.png}
\caption{Clip from video of generated agent play. Bottom left: Actual gameplay footage. Top left: autoencoder reconstruction of the environment. Right: Visual representation of chokepoint activation for this frame.}
\label{fig:video clip}
\end{center}
\end{figure}
Figure~\ref{fig:video clip} is a video clip visualizing the network behavior\footnote{Videos can be found here: https://tinyurl.com/m4df7rn}.
The evolution of networks A and B generated mostly disappointing behavior. While they do manage to navigate and pick up health packs, they evolved to only be able to turn left and right, respectively. This means that sometimes, they get stuck in uninteresting loops around empty rooms and die. Network B is interesting in that its information about the current state of its health seems to provide it an ability to break out of these loops. Networks C, D and E, however, were much more interesting. They all demonstrated an ability to navigate the environment and seek out health packs, to varying degrees of competence. Network C, by being trained in an environment where picking up a mine could immediately kill it, may have been bumped out of a local maxima that A and B fell into.
But why did networks trained on the alternative autoencoder topology get better results, and much faster, than networks trained on an autoencoder doing typical convolution? If we compare the chokepoint activation of the two networks, we can see that the alternative autoencoder had a much sparser activation. Taking the sum of every activation of the chokepoint (so, for an individual frame, the maximum activation would be 128), we see that normal convolution yields an average chokepoint activation sum of 24.3 over 1000 frames of gameplay, while the average sum for the alternative convolution is 12.7. Information about the environment, then, was compressed into fewer values. While the model generated by the unusual convolution wasn't as accurate as the one generated by normal convolution (the red jars in the environment were almost invisible to it) the higher compression allowed evolution to take place at a much faster pace. This suggests that, perhaps, the environment could be compressed even further by normal convolution, down to something less than 128 values. This could aid evolutionary methods by necessitating smaller behavior-generating networks.
\section{Conclusion}
Deep-Q learning off of the raw visual data in this environment outperforms the method presented here, learning to achieve a score between 1000-1500 after 500,000 steps of gameplay~\cite{VizDoom}. However, while it would have been nice to create a method that outperformed Deep-Q, the focus of this paper was on demonstrating that the hierarchal representation capabilities of deep-learning networks could compress the representation of raw visual data in a 3d environment to the degree that evolutionary methods could be effective. Such a compressed representation could be useful in other contexts: as suggested before, Deep-Q learning could be trained off of the compressed inputs as well, and perhaps training off of the compressed data would yield some advantage over training off of the raw visual data. Having a low-dimensional, internal representation of the environment, however fuzzy, could also enable other training methods beyond Deep-Q
While the results from evolution in this paper do not demonstrate a mastery of learning strategy for playing in a three dimensional environment, it has been shown that autoencoder compression can build an internal, low-dimensional model of a three-dimensional FPS environment from the high-dimensional visual data. Moreover, it is not necessary to have a preset database comprising a balanced representation of the environment, as the autoencoder can be trained as the game is played. Evolutionary methods can be used on this compressed representation to improve agent behavior.
\section*{Acknowledgements}
Thanks to to Christopher Dimauro for keeping the machines running and to Gabriella Barros and Christoph Salge for their advice. We also thank Nvidia for their generous hardware donation to the NYU Game Innovation Lab which made this project possible.
\nocite{*}
|
\section{INTRODUCTION}
Materials with nontrivial electronic band structures have been studied intensively in recent years\cite{Hre,TSCTI,Liu,XDPRX}. For instance, topological insulators are bulk insulators but as a consequence of the topology of the electronic bands, the surface states are gapless\cite{Hre,chen2009experimental}, while in Dirac and Weyl semimetals\cite{Liu,XDPRX,CavaCd3As2,huang2015observation,NbAsLYK,WTe2FDL,hirschberger2016chiral}, the symmetry protected bands have linear crossings with massless excitations. Topologically non-trivial band structures can lead to distinctive transport behaviors, such as a $\pi$ Berry phase from the Landau index analysis\cite{mikitik1999manifestation,zhang2005experimental,murakawa2013detection,huang2015observation}, an extremely large magnetoresistance (XMR)\cite{huang2015observation,shekhar2015extremely,liang2015ultrahigh,ali2014large}, and in Weyl semimetals the chiral anomaly can cause a negative longitudinal magnetoresistance \cite{CavaCd3As2,huang2015observation,NbAsLYK,WTe2FDL}.
Recently, the observation of XMR in the $X$(Sb,Bi) series of compounds, where $X$ is a rare earth element, has attracted much attention \cite{LaSbCava,DHLaSb,Zeng2015Topological,LaBiPRB,LaXFDL,LaBiIOP,CeSbWeyl,NdSbPRB,NdSbIOP,NdSbHField,Nayak2017Multiple,Cavatwo,LaSbTheory,YSbSr}. In these materials the temperature dependence of the resistivity has a plateau at low temperatures upon applying a magnetic field. In LaSb, it was suggested that this behavior has a similar origin to that of SmB$ _6 $, where it is proposed to be related to the appearance of the surface state\cite{LaSbCava}. However, angle resolved photoemission spectroscopy (ARPES) measurements provide evidence that the band topology in LaSb is trivial\cite{DHLaSb}. Meanwhile for LaBi, ARPES measurements show multiple Dirac cones near the Fermi level and it is proposed to show a non-trivial band topology with a $\pi$ Berry phase \cite{LaXFDL,Nayak2017Multiple,LaBiIOP}. In CeSb, a negative longitudinal magnetoresistance (MR) was observed, which provides evidence for the presence of the chiral anomaly in the field-induced ferromagnetic state due to time reversal symmetry breaking\cite{CeSbWeyl}. Consequently it is of interest to explore how the topology and Fermi surface evolves upon changing the $f$-electron configuration in this series of compounds. In addition, since $X$(Sb,Bi) materials all have moderately high quantum oscillation frequencies, it is important to extend the MR measurements to higher fields, in order to reliably determine the Berry phases.
PrSb has a simple cubic rock-salt structure similar to other $X$(Sb,Bi) materials and has a paramagnetic ground state without any magnetic transitions \cite{tsuchida1965magnetic,Elastic}. The trivalent Pr$^{3+}$ ions have a $4f^2$ configuration that results in a large local moment at the Pr site\cite{PhysRevLett.39.1028}, which contributes to the large magnetic field induced moments revealed by magnetization measurements\cite{Kido1992De}. Crystal field splitting leads to a singlet ground state and the anisotropic exchange gives rise to singlet-singlet excitations at low temperatures near the $X$ point \cite{NeutronPressure,Elastic}. Previous de Haas-van Alphen effect studies indicate that the extremal cross sectional areas of the electron orbits become larger with increasing field, resulting in a nonlinear relationship between the number of filled Landau levels and the inverse magnetic field.\cite{Kido1992De,Kido1993De}
Here we present a detailed magnetotransport study on PrSb. An extremely large magnetoresistance which reaches $\rho(H)$/$\rho(0)\approx4300$ at 32~T is observed. To probe the Fermi surface via the Shubnikov-de Haas (SdH) effect, angular dependent MR measurements are reported. The angular dependence of the quantum oscillation frequency corresponding to the $\alpha$-branch shows the signature of a two-dimensional nature of the corresponding Fermi surface, while a four-fold symmetric variation of the effective mass is detected. A detailed analysis of the quantum oscillations indicates an enlargement of the Fermi surface with increasing field, which becomes nearly constant when the field exceeds 14 T. The analysis of the Landau index of this enlarged Fermi surface shows that at high fields the Fermi surface of PrSb is topologically trivial.
\section{EXPERIMENTAL DETAILS}
Single crystals of PrSb were synthesized using a flux method described in Ref. \onlinecite{P1992Growth}. The mixture of Pr, Sb, Sn elements in a molar ratio of 1:1:10 were slowly cooled from 1100$^\circ$C down to 780$^\circ$C, before centrifuging to remove the Sn flux. The resulting crystals were cubic with a typical size of 3mm$\times$3mm$\times$3mm.
Resistivity and the angular dependent magnetoresistance measurements were performed using a Quantum Design Physical Property Measurement System (PPMS) from 300 K to 2 K with a maximum applied field of 9~T using a field rotation option. Four Pt wires were attached to the sample by spot welding and the current was applied parallel to the [100] direction. Magnetic susceptibility measurements were performed using both a Quantum Design superconducting quantum interference device (SQUID) magnetometer, and the vibrating sample magnetometer option for the PPMS. The low temperature magnetoresistance were measured in a He$ ^3 $ system with base temperature of 0.27~K and a maximum applied magnetic field of 15~T. The high field magnetoresistance measurements were performed at the High Magnetic Field Laboratory of the Chinese Academy of Sciences.
\section{RESULTS AND DISCUSSION}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{fig1.eps}
\end{center}
\caption{(Color online) (a) Temperature dependence of the magnetic susceptibility of PrSb in an applied field of 0.1 T. (b) Magnetization as a function of applied field at three temperatures, up to a field of 10~T. (c) Temperature dependence of the resistivity of PrSb sample~\#1 from 300~K to 2 K.The inset displays an enlargement of the low temperature region. (d) Temperature dependence of the resistivity of sample \#1 in several applied magnetic fields. }
\label{figure1}
\end{figure}
Figure~\ref{figure1}(a) displays the temperature dependent magnetic susceptibility of PrSb, where there is a broad shoulder at around 10~K, which is consistent with previous reports\cite{monachesi1994optical}. The low temperature behavior can be explained by the crystal field effect commonly observed in Pr-based Van Vleck paramagnets, such as PrPt$_4$Ge$_{12}$\cite{PrPtGe}. No long range magnetic ordering is found down to 2~K and by fitting using the Curie-Weiss law above 50~K, an effective magnetic moment of $ \mu_{eff} $ = 3.3~$\mu_B$ and a Curie Weiss temperature $ \Theta_P $ = -5 K are obtained. These values are close to the previous study\cite{tsuchida1965magnetic}, and suggest a trivalent nature of the Pr atoms with two fully localized $f$-electrons. The field-dependent magnetization measurements shown in Fig.~\ref{figure1}(b) reveal a nearly linear field dependence of the magnetization at different temperatures below 20~K. At 10~T the magnetization reaches 0.8 $ \mu_B$/Pr, consistent with a magnetic moment of 0.9 $ \mu_B $ at 10~T and 15~K \cite{schoenes1990magneto}. The resistivity of PrSb from 300 K to 2 K is presented in Fig.~\ref{figure1}(c), which shows typical metallic behavior in zero applied field. The enlarged view of the low temperature range in the inset shows no anomaly arising from the superconducting transition of Sn, indicating a lack of residual flux in the samples. The residual resistivity is about 0.3~$\mu\Omega$~cm, and the residual resistance ratio is about 110, indicating the high quality of the crystals. Figure~\ref{figure1}(d) displays the temperature dependence of the resistivity when different magnetic fields are applied perpendicular to the current. Unlike in zero field where simple metallic behavior is observed, in the presence of a magnetic field there is an increase of the resistivity upon lowering the temperature, which saturates below 10~K. Similar behavior has also been reported in LaSb\cite{LaSbCava} and YSb\cite{YSbSr}. Although in LaSb this behavior was reported to be analogous to the topological Kondo insulator SmB$_6$, further studies concluded that this plateau is more likely due to electron-hole compensation\cite{DHLaSb}.
\begin{figure*}
\begin{center}
\includegraphics[width=2\columnwidth]{fig2.eps}
\end{center}
\caption{(Color online) (a) Field dependence of the resistivity of sample \#2 at various temperatures with an angle between the magnetic field and current of $ \theta = 86^\circ $ . The inset illustrates the experimental configuration. (b) Magnetoresistance at 2~K for different $\theta$. (c) The fast Fourier transform analysis of the measurements at $\theta$ $\sim$ 86$^\circ$ and 42$^\circ$. (d) Angular dependence of the SdH oscillation frequencies for the $\alpha$-band, where the field is rotated in the (100) plane. (e) Temperature dependence of the quantum oscillation amplitudes of the $\alpha$-band for two values of $\theta$, fitted with the Lifshitz-Kosevich formula. (f) The angular dependence of the effective mass of the $\alpha$-band obtained from fitting the Lifshitz-Kosevich formula at different $\theta$.}
\label{figure2}
\end{figure*}
We also performed magnetoresistance measurements at various temperatures, with fields applied at different angles $\theta$ to the current direction (along [100]), as illustrated in Fig.~\ref{figure2}(a). Figure~\ref{figure2}(a) also displays the MR at $\theta\sim86^\circ$ at several temperature, which shows an XMR effect similar to other $X$Sb compounds, where the MR exceeds 300 at 2 K and 9 T. Upon tilting the magnetic field away from the current direction, the MR decreases but no negative MR is observed and therefore there is no indication of a chiral anomaly, unlike CeSb where such evidence was found \cite{CeSbWeyl}. Upon rotating the sample, we studied the angular dependence of the oscillation frequency and effective mass. As the frequency of the quantum oscillations is proportional to the extremal cross sectional areas of the Fermi surface, the change of oscillation frequency with angle reflects the Fermi surface structure. The fast Fourier transform (FFT) analysis of the SdH quantum oscillations for two values of $\theta$ are shown in Fig.~\ref{figure2}(c). At $\theta$ $\sim$ 86$^\circ $, the FFT analysis reveals a fundamental frequency of $F_\alpha$ $\sim$ 220~T, and two harmonic frequencies. At $\theta$ $\sim$ $ 42^\circ $, the FFT analysis yields two fundamental $\alpha$ band frequencies at $F_\alpha$ $\sim$ 280~T and $F_{\alpha'}$ $\sim$ 320~T, as well as harmonic frequencies. In addition, there are frequencies associated with the $\beta$ and $\gamma$ bands, $F_\beta$ $\sim$ 463~T and $F_{\gamma} \sim 934$~T. Note that the $\beta$ frequency can only be clearly observed at certain angles due to the overlap with $F_{2\alpha}$, and we focus on the anisotropy of $\alpha$-band. The oscillation frequencies of the $\alpha$-band as a function of $\theta$ obtained from the FFT are shown in Fig.~\ref{figure2}(d). The angular dependence of the oscillation frequencies can be fitted by the expression for a two-dimensional (2D) Fermi surface, $F=F_0/$cos$(\theta-(n\pi/2))$, where $ F_0 \sim 219$~T is the first principal frequency, suggesting that the Fermi surface corresponding to the $\alpha$-band has a two-dimensional nature. Meanwhile the temperature dependence of quantum oscillation amplitudes at two angles are shown in Fig.~\ref{figure2}(e). The solid line shows the result of fitting with the Lifshitz-Kosevich (LK) formula \cite{shoenberg2009magnetic}, yielding effective masses of m$^*$ of 0.208~m$_0$ for $\theta$ $\sim$ $ 86^\circ $ and 0.245~m$_0$ for $\theta$ $\sim$ $ 42^\circ $. Moreover, the angular dependence of the effective mass from fitting the LK formula reveals a 4-fold signature, as shown in Fig.~\ref{figure2}(f). Upon rotating the field, there are eight step-like changes of the effective mass, whereas between the steps the value stays at a near constant value. Note that the plateau with a higher effective mass around 45$^ \circ $ coincides to where the two bands cross, which leads to an increase of the oscillation frequency. The overall 4-fold signature is consistent with the 4-fold symmetry of the cubic crystal structure in the (100) plane. The abrupt changes of the effective mass may be related to the structure of the bullet-like Fermi surfaces corresponding to the $\alpha$-band found in other $X$Sb materials \cite{LaBiPRB,YSbSr}, which lie perpendicularly in the Brillouin zone but detailed calculations are required explain this behavior.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{fig3.eps}
\end{center}
\caption{(Color online) The magnetoresistance at different temperatures is shown up to a maximum field of (a) 15 T and (b) 32 T for samples \#1 and \#3 respectively. (c) FFT of quantum oscillations at 1.8~K in different field ranges showing the position of the fundamental frequency of the $\alpha$-band $F_\alpha$. (d ) $F_\alpha$ as a function of applied field, for measurements of two samples.}
\label{figure3}
\end{figure}
To determine whether the large magnetoresistance of PrSb is related to the topological nature of the Fermi surface, we performed high field magnetoresistance measurements on PrSb. Due to the high frequencies of the quantum oscillations, there can be a reasonably large uncertainty in the extrapolated residual Landau index if the measurements are not performed to sufficiently large fields (small $1/B$). Therefore we measured the SdH oscillations in high magnetic fields to accurately determine the Berry phase. Figure~\ref{figure3}(a) displays the transverse magnetoresistance measured at low temperature down to 0.3~K up to 15~T, while Fig.~\ref{figure3}(b) shows the transverse magnetoresistance measured up to 32~T. The current is along the [100] direction while the magnetic field is perpendicular to the current direction. XMR is observed with clear quantum oscillations, where $\rho(H)$/$\rho(0)$ is about 4300 at $B$ = 32~T. The field dependent resistivity shows a quadratic dependence, suggesting that electron hole compensation is the origin of the XMR\cite{pippard1989magnetoresistance}. Figure~\ref{figure3}(c) shows the FFT analysis of the SdH oscillation frequencies of sample \#3 in different field ranges at 1.8~K. It can be seen that in the lower field ranges, the oscillation frequency of the $\alpha$-band increases with increasing magnetic field. The field dependence of these frequencies is displayed in Fig.~\ref{figure3}(d) for both samples, where the field is taken as the midpoint of the field ranges of the FFT. While an increase of the frequency with field was previously reported from a comparison between two field ranges \cite{Kido1992De,Kido1993De}, here we report the detailed evolution. Our results show that there is an increase of the frequency with field up to around 14~T, while at larger fields the frequency is nearly unchanged. This indicates that at low fields there is an enlargement of the Fermi surface of PrSb, which becomes near constant at higher fields.
Since the oscillation frequency increases with increasing field up to 14 T, the residual value cannot be obtained from a linear fit to the inverse field dependence of the Landau index in this field range. Therefore we analyzed the Landau index at higher fields, where the quantum oscillation frequency remains constant. After subtracting the background contribution, the data at 1.8~K above 15~T in Fig.~\ref{figure3}(b) were plotted as a function of $1/B$, as displayed in Fig.~\ref{figure4}(a). Despite the strong harmonic oscillations, the positions of the valleys can be clearly resolved and the positions are marked by the vertical lines. Integer Landau levels were assigned to the valley positions, which are shown in Fig.~\ref{figure4}(b). From fitting the data with a linear relationship, the residual Landau index was determined to be 0.07, which corresponds to a zero Berry phase\cite{mikitik1999manifestation}, and hence the Fermi surface of PrSb in this field range is topologically trivial. Our measurements were measured to sufficiently high fields to measure the ninth Landau index and therefore the uncertainty associated with the fitted residual value is small and the results are reliable.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{fig4.eps}
\end{center}
\caption{(Color online) (a) SdH oscillations of PrSb sample \#3 at 1.8~K as a function of $1/B$, obtained from subtracting the background contribution. (b) The $1/B$ dependence of the Landau index, obtained from the valley positions of the SdH oscillations. The solid line shows a linear fit, which reveals a residual Landau index of $n=0.07$.}
\label{figure4}
\end{figure}
\section{CONCLUSION}
In summary, we report a detailed angle-dependent magnetotransport study of PrSb where we find XMR with $\rho(H)$/$\rho(0)\approx4300$ at 32~T. We probed the structure of the Fermi surface via SdH measurements, where we observe 4-fold symmetry in the angular dependence of the effective mass of the $\alpha$-band, which shows several abrupt changes. In addition, there is an increase of the quantum oscillation frequencies of the $\alpha$-band with magnetic field up to 14 T, indicating an enlargement of the Fermi surface, before reaching a near constant level above 14 T. The analysis of the Landau index at high fields shows that there is zero Berry phase in this field range, and therefore the Fermi surface topology is trivial. Meanwhile the transverse magnetoresistance shows a near quadratic field dependence and therefore these results indicate that the XMR in PrSb is more likely due to electron-hole compensation rather than topological protection of the surface state. Whether the band topology remains trivial when no magnetic field is applied remains to be determined, which needs to be checked using ARPES measurements. Furthermore, the origin of the widely observed XMR behavior in this series of compounds is still under debate, and the role of the lanthanide elements in influencing the properties and topology of the band structure needs to be explored further.
\begin{acknowledgments}
We thank Y.~Liu, C.~Cao and X.~Lu for valuable discussions and helpful suggestions. This work was supported by the National Key R\&D Program of China (No.~2017YFA0303100, No.~2016YFA0300202), the National Natural Science Foundation of China (No.~U1632275, No.~11474251) and the Science Challenge Project of China (No.~TZ2016004).
\end{acknowledgments}
|
\section{Introduction}
In this paper, our interest lies in motives for rational homotopy
types of algebraic varieties.
Rational homotopy theory originated from Quillen
\cite{Q} and Sullivan \cite{S}.
In both approaches,
the main object of interest is
an algebraic invariant associated to
a topological space, that encodes a rational homotopy type of the space
under suitable conditions.
In Quillen's theory, the algebraic invariant
is a differential graded Lie algebra
obtained from a simply connected
topological space.
On the other hand, to a topological space $S$,
Sullivan associated a commutative differential graded (dg) algebra
$A_{PL}(S)$ of polynomial differential forms on $S$ with rational coefficients.
The cohomology ring of $A_{PL}(S)$ is isomorphic to the
graded-commutative ring $H^*(S,\QQ)$ of the singular cohomology.
In his approach, the main algebraic invariants of $S$ are
$A_{PL}(S)$ and its (so-called) Sullivan model.
We now turn to our attention to algebraic varieties.
One of motivating sources of motives is Hodge theory.
When $S$ is a complex algebraic variety,
thanks to the works of Morgan \cite{M} and Hain \cite{Ha},
a suitable model of $A_{PL}(S)$ admits a mixed Hodge structure
in an appropriate setting.
Their work generalized the classical Hodge theory to
Hodge theory for
higher rational
homotopy groups and unipotent fundamental groups, i.e.,
the pro-unipotent completion of fundamental group.
Meanwhile, in 80's, a notion of motivic homotopy type was envisaged by Grothendieck
\cite{G}.
Deligne and Gonchalov developed
a motivic theory for the pro-unipotent completions of fundamental groups
in the setting of mixed Tate (and Artin-Tate) motives over a number field
and its ring of integers \cite{DG}.
Our investigation is an attempt to define and study
a motivic generalization of $A_{PL}(S)$.
In order to get a feeling for invariants we will study,
let us compare the homotopy (triangulated) category arising from topological
spaces and the category of motives.
Let $DM^\otimes(k)$ be the symmetric monoidal
triangulated category of Voevodsky motives over a perfect field $k$,
\cite{MVW}, \cite{Voe} (here $DM^\otimes(k)$ is allowed to admit infinite coproducts).
One of pleasant features of $DM^\otimes(k)$ is
that the construction is given by ``doing homotopy theory'' of schemes,
so that the analogy is quite transparent,
while motivic cohomology groups appear as the hom sets in $DM(k)$.
By analogy with homotopy theory,
$DM^\otimes(k)$ should be thought of as an analogue
of the homotopy category
of module spectra over the Eilenberg-MacLane ring spectrum $H\ZZ$.
The motive $M(X)\in DM(k)$ associated to $X$ \cite{MVW} plays the role of
the singular chain complex of a topological space.
We now work with rational coefficients instead of $\ZZ$,
and take a point of view that
a topological counterpart of $DM^\otimes(k)$ is the derived category of $\QQ$-vector spaces.
Remember that for a topological space $S$,
$A_{PL}(S)$ is a commutative dg algebra with rational coefficients
whereas
the singular cochain complex $C^*(S,\QQ)$ is only a dg algebra that is
not necessarily commutative.
We can think that the commutative dg algebra
$A_{PL}(S)$ amounts to the (underlying) complex $C^*(S,\QQ)$ endowed with
an $E_{\infty}$-algebra structure, that is, a commutative algebra structure
in the operadic or $(\infty,1)$-categorical sense.
This structure is crucial for rational homotopy theory.
(Also, the integral singular
cochain complex $C^*(S,\ZZ)$ admits an $E_{\infty}$-algebra structure \cite{FrC},
\cite{MS}, and it is important to generalizations of rational homotopy theory
such as integral homotopy theory \cite{Man}.)
To incorporate such structures
and to pursue the comparison, we need to replace the derived category
of $\QQ$-vector spaces with its $(\infty,1)$-categorical enhancement,
i.e., the derived $(\infty,1)$-category $\mathsf{D}(\QQ)$
of $\QQ$-vector spaces, that inherits a symmetric monoidal structure
given by the tensor product of complexes.
For the introductions to the $(\infty,1)$-categorical language,
we refer to \cite[Chapter 1]{HTT}, \cite{Ber}, \cite{GrI} for instance.
Then $A_{PL}(S)$ may be viewed as a commutative algebra object
of the symmetric monoidal $(\infty,1)$-category $\mathsf{D}^\otimes(\QQ)$
in the $(\infty,1)$-categorical sense.
Let $\DM^\otimes(k)$ be a symmetric monoidal
$(\infty,1)$-category of motives, that
is an $(\infty,1)$-categorical enhancement of $DM^\otimes(k)$.
Let $\CAlg(\DM^\otimes(k))$ be the $(\infty,1)$-category of
commutative algebra objects of $\DM^\otimes(k)$.
The analogy suggests that it is natural to think that
a motivic generalization of $A_{PL}(-)$ should be defined as an object
of $\CAlg(\DM^\otimes(k))$ whose underlying object in $\DM(k)$
is equivalent to the (weak) dual of $M(X)$.
There are (at least) two approaches to constructing this:
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item If $\Sm_k$ denotes the category of smooth schemes over $k$,
equipped with the symmetric monoidal structure given by
the product $X\times_kY$, then an object $X$ of $\Sm_k$ can be viewed as a
cocommutative coalgebra
object such that the comultiplication is the diagonal $X\to X\times_kX$,
and the counit is the structure morphism $X\to \Spec k$.
If we regard the assignment $X\mapsto M(X)$ as a symmetric monoidal
functor $\Sm_k\to \DM^\otimes(k)$, then $M(X)$ is a cocommutative
coalgebra object in $\DM(k)$. Let $\uni_k$ be a unit object in $\DM(k)$.
Then the internal hom object $\mathsf{Hom}_{\DM(k)}(M(X),\uni_k)$
inherits a commutative algebra structure in the $(\infty,1)$-categorical sense
(i.e., an $E_{\infty}$-algebra structure) from $M(X)$.
\item Let $X$ be an object of $\Sm_k$ and let $f:X\to \Spec k$ be
the structure morphism.
Suppose that a symmetric monoidal $(\infty,1)$-category $\DM^\otimes(X)$
of motives over $X$ is available and there is an adjoint pair
$f^*:\DM(k)\rightleftarrows \DM(X):f_*$.
If $f^*$ is symmetric monoidal, then the right adjoint $f_*$ is a
lax symmetric
monoidal functor, so that $f_*$ sends a commutative algebra object in $\DM(X)$
to a commutative algebra object in $\DM(k)$. We denote by $\uni_X$
a unit object of $\DM(X)$ and think of it as a commutative algebra object.
We then have a commutative algebra object $f_*(\uni_X)$, that is a natural
candidate.
\end{enumerate}
The approach (i) is reminiscent of the setup in topology: singular
chain complexes and singular cochain complexs
(but, the assignment $S\mapsto C_*(S,\ZZ)$ is only oplax monoidal).
We will adopt the approach (ii) since it gives
a clear relationship with the relative situation. We will
use the formalism of motives over $X$, extensively developed by Cisinski and D\'eglise. For a smooth scheme $X$, we define an object $M_X$ of $\CAlg(\DM^\otimes(k))$, which we shall
refer to as the cohomological motivic algebra of $X$.
The definition will be given in Section~\ref{CMAsection}. Actually,
in Section~\ref{CMAsection},
we work with not only rational coefficients but an arbitrary coefficient ring.
The first important property of $M_X$ is that
a (topological) realization of $M_X$
gets identified with the commutative dg algebra $A_{PL}(X^t)$ of polynomial
differential forms on the underlying topological space $X^t$ of $X\times_{k}\Spec\CC$
when $k\subset \CC$.
To Weil cohomology theory such as singular cohomology, analytic/algebraic de Rham cohomology, $l$-adic \'etale cohomology, one can associate a symmetric monoidal functor called a realization functor:
\[
\mathsf{R}:\DM^\otimes(k)\to \mathsf{D}^\otimes(K)
\]
where $K$ is a coefficient field of cohomology theory,
and $\mathsf{D}^\otimes(K)$ is the symmetric monoidal derived $(\infty,1)$-category of $K$-vector spaces. The field $K$ is assumed to be
of characteristic zero.
For example, when $k$ is embedded in $\CC$, the realization functor
$\mathsf{R}:\DM^\otimes(k)\to \mathsf{D}^\otimes(\QQ)$ associated to singular cohomology theory (with rational coefficients) carries $M(X)$ to a complex quasi-isomorphic to the singular chain complex $C_*(X^t,\QQ)$ of the underlying topological space $X^t$. Notice that the realization functor is symmetric monoidal.
It gives rise to a functor
\[
\CAlg(\DM^\otimes(k))\to \CAlg(\mathsf{D}^\otimes(K)),
\]
which we call the multiplicative realization functor,
where $\CAlg(\mathsf{D}^\otimes(K))$ is the $(\infty,1)$-category of commutative algebra objects in $\mathsf{D}^\otimes(K)$. One can naturally identify
$\CAlg(\mathsf{D}^\otimes(K))$ with the $(\infty,1)$-category obtained from the category of commutative dg algebras over $K$ by inverting quasi-isomorphisms (cf. Section~\ref{convention}).
In the case of singular cohomology, we have $\CAlg(\DM^\otimes(k))\to \CAlg(\mathsf{D}^\otimes(\QQ))$.
The commutative dg algebra $A_{PL}(X^t)$ appears as the image of $M_X$
under the multiplicative realization functor. This property is proved in Section~\ref{realization}. Thus, along with merely an analogy,
the multiplicative realization functor
relates $M_X$ with $A_{PL}(X^t)$.
It is worth emphasizing that
it allows one to promote many operations on $A_{PL}(X^t)$ to a motivic level
(even as the elementary nature tends to obscure the significance).
For example,
the multiplicative realization functor preserves (small) colimits.
Suppose that $x$ is a
$k$-rational point on $X$.
Let $\epsilon:A_{PL}(X^t)\to \QQ$ be the augmentation induced by
the point $x$ on $X^t$. The bar construction of the augmented commutative
dg algebra
can be described in terms of (a cosimplicial diagram of)
colimits. Thus, it is possible to promote the bar construction
of $A_{PL}(X^t)\to \QQ$ to a bar construction of $M_X\to \uni_k$
in $\CAlg(\DM^\otimes(k))$.
\vspace{2mm}
{\it Tannakian aspect.}
One of descriptions of ``motivic stuctures'' is
a tannakian formalism.
We discuss a tannakian aspect in Section~\ref{MGA}.
Recall that various ``topological invariants'' of algebraic varieties
are equipped with actions of groups. For instance,
an $l$-adic \'etale cohomology group has an action of the absolute
Galois group, and
a Hodge structure can be described by an action of a Mumford-Tate group.
In our context, the groups will be
the derived motivic Galois group $\MG$, introduced in \cite{Tan},
and
the associated pro-algebraic group $MG$ we call the motivic Galois group
(see the beginning of Section~\ref{MGA}, Section~\ref{underlyinggroup},
and \cite{Tan}).
By using $M_X$ we construct a canonical action of $\MG$ on $A_{PL}(X^t)$
(when $k\subset \CC$).
It is a tannakian representation of the motivic structure
on the rational homotopy type.
When $X$ has a base point,
it is possible to deduce the pro-unipotent completions
$\pi_i(X^t,x)_{uni}$ of homotopy
groups $\pi_i(X^t,x)$ ($i\ge1$) from $A_{PL}(X^t)$
with the augmentation.
We obtain canonical actions of the motivic Galois group $MG$ on pro-unipotent
groups $\pi_i(X^t,x)_{uni}$ from the action of $\MG$ on $A_{PL}(X^t)$
(cf. Theorem~\ref{actiononhomotopy}, Corollary~\ref{actiononhomotopy2}).
Thus, from the tannakian viewpoint, our study may be regarded as
a generalization of motivic structures on (co)homology groups to
motivic structures on
the unipotent non-abelian fundamental groups and
higher rationalized homotopy groups.
\vspace{2mm}
{\it Structure of cohomological motivic algebras.}
In order to understand things more explicitly,
it is natural to attempt to understand the structure of $M_X$,
that is, what a cohomological motivic algebra looks like.
In Section~\ref{Sullivanmodel}, as a first step towards the understanding,
we describe an explicit structure of the cohomological motivic algebra
in several cases such as a projective space over a field.
To do this, we recall an approach that traces back to Sullivan's work.
A (minimal) Sullivan model of $A_{PL}(S)$ is given
by an iterated homotopy pushout of free commutative dg algebras
(see e.g. \cite{He}, \cite{Hin}, \cite{FHT} or the beginning of Section~\ref{Sullivanmodel}).
Based on this idea, we describe $M_X$ as
a colimit of an analogous diagram of
free commutative algebra objects
in $\CAlg(\DM^\otimes(k))$ in an explicit way.
Unlike the classical rational homotopy theory,
the study of $M_X$ is not so simple even in relatively elementary
cases: we need some devices and deep results.
This difference may be regarded as a reflection of the fact that
$M_X$ has rich and interesting structures.
For instance,
suppose that $C$ is a proper smooth curve of genus $g>1$
with a base $k$-rational point $c$.
Let $J_C$ be the Jacobian variety and let $u:C\to J_C$ be
the Abel-Jacobi morphism.
We here take a viewpoint that
the Abel-Jacobi morphism is an ``algebraic abelianization'' of $C$:
when $k=\CC$, the map
$C^t\to J_{C}^t$ of the underlying toplogical spaces
induces an abelianization $\pi_1(C^t,c)\to \pi_1(J_{C}^t,u(c))\simeq \pi_1(C^t,u)^{ab}$.
The Abel-Jacobi morphism $u$ induces
a morphism $u^*:M_{J_C}\to M_C$ of cohomological
motivic algebras.
Then it gives rise to an inductive sequence in $\CAlg(\DM^\otimes(k))$:
\[
M_{J_C}=M_1\to M_2\to \cdots \to M_n\to M_{n+1}\to \cdots \to M_C
\]
that decomposes $u^*:M_{J_C}\to M_C$ such that
$M_C$ is a filtered colimit $\varinjlim_{n\ge 1}M_n$
(cf. Section~\ref{Sull1}, Section~\ref{curveexp}).
One can think of this sequence (or co-tower) starting with $M_{J_C}$
as a structure of $M_C$ or a refined Abel-Jacobi morphism.
It is notable that it does not exist in the category of schemes
and does not arise from $\DM(k)$ .
Roughly speaking, this sequence gives a step-by-step description of
the non-abelian nature of $C$ that starts with
its ``abelian part'' $M_{J_C}$.
From a perspective of the formality,
it is not reasonable to expect a formality of $M_X$
of a smooth projective variety $X$ in general
(even if
one can define a formality by using a motivic $t$-structure).
Actually, there is a counterexample to
the formality at the Hodge level (see \cite{CCM}).
The large class is yet to be explored and remains mysterious, so that
one may expect more to understand
structures of cohomological motivic algebras.
\vspace{2mm}
{\it Cotangent motives.}
In Section~\ref{cotangenthomotopy}, we introduce
a new invariant of a pointed smooth scheme $(X,x)$ over a perfect field,
that lies in $\DM(k)$.
The invariant $LM_{(X,x)}$ in $\DM(k)$ is defined by means of cotangent
complex of $M_X$ endowed with the augmentation induced by $x$.
We shall
call $LM_{(X,x)}$
the cotangent motive of $X$ at $x$ (cf. Definition~\ref{cotangentmotivedef}). For the definition,
we apply the theory of cotangent complexes
in a very general setting, developed by Lurie.
We prove that the rationalized homotopy group appears as the
realiziation
of $LM_{(X,x)}$ (cf. Theorem~\ref{motivichomotopy}, Theorem~\ref{motivichomotopy2}). Namely, when $k$ is embedded in $\CC$ and the underlying
topological space $X^t$ is simply connected,
$H^i(\mathsf{R}(LM_{(X,x)}))$ is the dual of the $i$-th
rationalized homotopy
group of $X^t$. In addition, $H^1(\mathsf{R}(LM_{(X,x)}))$
can be identified with the cotangent space of the origin of
the pro-unipotent completion
of the fundamental group, that is, the
``linear data'' of the fundamental group.
By using Hodge realization of $LM_{(X,x)}$
one can obtain a mixed Hodge structure on the rational homotopy group
(in the simply conneced case).
Intuitively,
we may consider $LM_{(X,x)}$ to be a motive
for (the dual of) rational higher homotopy groups and the linear data of
the fundamental group.
Though $LM_{(X,x)}$
has less information than $M_X$,
the motive $LM_{(X,x)}$ has the relation with
homotopy groups in a more direct way than $M_X$,
and furthermore one can consider motivic cohomology of $LM_{(X,x)}$
since it belongs to $\DM(k)$.
We apply the (explicit) study of $M_X$
in Section~\ref{Sullivanmodel} to compute $LM_{(X,x)}$.
Indeed, one of motivations for it is computation of the cotangent motives.
For instance, if $\mathbb{P}^n$ is the $n$-dimensional projective space (over a perfect field) endowed with a base point $x$, then
\[
LM_{(\mathbb{P}^n,x)}\simeq \uni_k(-1)[-2]\oplus \uni_k(-n-1)[-2n-1],
\]
where ``$(s)$'' and ``$[t]$'' indicate the Tate twist and the shift, respectively. This means that $\uni_k(1)$ is a ``motive for the second rational homotopy group'', and
$\uni_k(n+1)$ is a ``motive for the $(2n+1)$-th rational homotopy group'' (cf. Remark~\ref{motcpsR}).
\vspace{2mm}
{\it Homotopy exact sequence.}
Remember the homotopy exact sequence for \'etale
fundamental groups
\[
1 \to \pi_1^{\textup{\'et}}(X\times_{k}\Spec \bar{k},\bar{x})\to \pi_1^{\textup{\'et}}(X,\bar{x}) \to \Gal(\bar{k}/k)\to 1
\]
where $\bar{k}$ is a separable closure of $k$.
This plays a central role in the theory of \'etale fundamental groups.
In Section~\ref{mhes}, by means of a tannakian theory developed in \cite{DTD},
when $X$ is an algebraic curve we formulate and prove a version of the homotopy exact sequence
in which the derived motivic Galois group (or stack) instead of $\Gal(\bar{k}/k)$ (cf. Proposition~\ref{MHES}).
\vspace{2mm}
From a conceptual point of view,
subjects in Section~\ref{MGA}, \ref{Sullivanmodel}, \ref{cotangenthomotopy},
\ref{mhes} are interconnected with each other.
Nevertheless, it is possible to read these Sections in any order with some exceptions.
It is desirable and important to study various
(multiplicative) realizations of cohomological motivic algebras $M_X$
and $LM_{(X,x)}$ in detail: \'etale, de Rham, Hodge, crystalline realizations,
a relation with Chen's theory of iterated integrals, etc.
These issues remain untouched and are beyond the scope of this paper. We hope to return to subjects in the future.
In Appendix, in the case of mixed Tate motives over a number field
we compare our approach and an approach to motivic fundamental groups
due to Deligne and Gonchalov. We hope that the comparison is helpful
for understanding circle of ideas from the viewpoint of their work.
\section{Notation and Convention}
\label{convention}
\subsection{}
We shall use the theory of {\it quasi-categories} extensively developed by
Joyal and Lurie from the viewpoint of $(\infty,1)$-categories.
This theory provides us with powerful tools and adequate language for
our purpose, though a part of contents might be reformulated
in term of other languages such as model categories or the like.
Following \cite{HTT}, we shall refer to quasi-categories
as {\it $\infty$-categories}.
Our main references are \cite{HTT} and \cite{HA}.
To an ordinary category $\mathcal{C}$,
one can assign an $\infty$-category by taking
its nerve $\NNNN(\mathcal{C})$.
Such simplicial sets $\NNNN(\mathcal{C})$ arising from ordinary categories
naturally constitute a full subcategory of
the simplicial category of $\infty$-categories.
Therefore,
when we treat ordinary categories we often omit the nerve $\NNNN(-)$
and think of them directly as $\infty$-categories.
We often refer to a map $S\to T$ of $\infty$-categories
as a functor. We call a vertex in an $\infty$-category $S$
(resp. an edge) an object (resp. a morphism).
We use Grothendieck universes $\mathbb{U}\in \mathbb{V}\in \mathbb{W}\in\ldots$
and usual mathematical objects such as groups, rings, vector spaces
are assumed to belong to $\mathbb{U}$.
Here is a list of (some) of the convention and notation that we will use:
\begin{itemize}
\item $\Delta$: the category of linearly ordered finite sets (consisting of $[0], [1], \ldots, [n]=\{0,\ldots,n\}, \ldots$)
\item $\Delta^n$: the standard $n$-simplex as the simplicial set represented by
$[n]$,
\item $\sSet$: the category of simplicial sets,
\item $\textup{N}$: the simplicial nerve functor (cf. \cite[1.1.5]{HTT})
\item $\Gamma$: the nerve of the category of pointed finite sets, $\langle 0\rangle=\{*\}, \langle 1\rangle=\{*,1\},\ldots,\langle n\rangle=\{*,1,\ldots,n\},\ldots$
\item $\mathcal{C}^{op}$: the opposite $\infty$-category of an $\infty$-category $\mathcal{C}$. For a functor $F:\mathcal{C}\to \mathcal{D}$, we denote by
$F^{op}:\mathcal{C}^{op}\to \mathcal{D}^{op}$ the induced functor
\item Let $\mathcal{C}$ be an $\infty$-category and suppose that
we are given an object $c$. Then $\mathcal{C}_{c/}$ and $\mathcal{C}_{/c}$
denote the undercategory and overcategory, respectively (cf. \cite[1.2.9]{HTT}).
\item $\CCC^\simeq$: the largest Kan subcomplex (contained) in an $\infty$-category
$\CCC$,
that is, the Kan complex obtained from $\CCC$
by restricting
morphisms (edges) to equivalences.
\item $\operatorname{Cat}_\infty$: the $\infty$-category of small $\infty$-categories, Similarly,
$\wCat$ denotes $\infty$-category of large $\infty$-categories (i.e.,
$\infty$-categories that belong to $\mathbb{V}$),
\item $\SSS$: $\infty$-category of small spaces. We denote by $\widehat{\SSS}$
the $\infty$-category of large $\infty$-spaces (cf. \cite[1.2.16]{HTT})
\item $\textup{h}(\mathcal{C})$: homotopy category of an $\infty$-category (cf. \cite[1.2.3.1]{HTT})
\item $\Fun(A,B)$: the function complex for simplicial sets $A$ and $B$
\item $\Fun_C(A,B)$: the simplicial subset of $\Fun(A,B)$ classifying
maps which are compatible with
given projections $A\to C$ and $B\to C$.
\item $\Map(A,B)$: the largest Kan subcomplex of $\Fun(A,B)$ when $B$ is an $\infty$-category.
\item $\Map_{\mathcal{C}}(C,C')$: the mapping space from an object $C\in\mathcal{C}$ to $C'\in \mathcal{C}$ where $\mathcal{C}$ is an $\infty$-category.
We usually view it as an object in $\mathcal{S}$ (cf. \cite[1.2.2]{HTT}).
If $\mathcal{C}$ is an ordinary category, we write
$\Hom_{\mathcal{C}}(C,C')$ for the hom set.
\item $C^\vee$:
For an object $C$ of a symmetric monoidal $\infty$-category $\mathcal{C}$,
we write $C^\vee$ for a dual of $C$ when $C$ is a dualizable object.
If there are internal objects,
we write $C^\vee$ also for the weak dual, that is,
the internal hom $\mathsf{Hom}_{\mathcal{C}}(C,\uni_{\mathcal{C}})$ with $\uni_{\mathcal{C}}$ a unit object.
\item $\Ind(\mathcal{C})$; $\infty$-category of Ind-objects
in an $\infty$-category $\mathcal{C}$
(see \cite[5.3.5.1]{HTT}, \cite[4.8.1.13]{HA} for the symmetric monoidal setting).
\item $\PR$: the $\infty$-category of presentable $\infty$-categories
whose morphisms are left adjoint functors.
\end{itemize}
\subsection{}
{\it From model categories to $\infty$-categories.}
We recall Lurie's construction
by which one can obtain
$\infty$-categories
from a category (more generally
$\infty$-category) endowed with a prescribed collection of
morphisms (see \cite[1.3.4, 4.1.3, 4.1.4]{HA} for details).
It can be viewed as an alternative of the Dwyer-Kan
hammock localization.
Let $\mathcal{D}$ be a category and let $W$ be a collection
of morphisms in $\mathcal{D}$ which is closed under composition
and contains all isomorphisms.
A typical example of $(\mathcal{D},W)$
which we have in mind is
$(\mathbb{M},W_{\mathbb{M}})$ such that
$\mathbb{M}$ is a model category (see e.g. \cite[Appendix]{HTT}, \cite{Hov})
and $W_{\mathbb{M}}$ is the collection of all weak equivalences.
For $(\mathcal{D},W)$, there is an $\infty$-category $\NNNN(\mathcal{D})[W^{-1}]$
and a functor
$\xi:\textup{N}(\mathcal{D})\to \NNNN(\mathcal{D})[W^{-1}]$
such that for any $\infty$-category $\CCC$ the composition
induces a fully faithful functor
\[
\Map(\NNNN(\mathcal{D})[W^{-1}],\CCC)\to \Map(\textup{N}(\mathcal{D}),\CCC)
\]
whose essential image consists of those functors $F:\textup{N}(\mathcal{D})\to\CCC$ such that $F$ carry morphisms lying in $W$
to equivalences in $\CCC$.
We shall refer to $\NNNN(\mathcal{D})[W^{-1}]$
as the $\infty$-category obtained from $\mathcal{D}$
by inverting morphisms in $W$.
Consider $(\mathbb{M},W_{\mathbb{M}})$ such that
$\mathbb{M}$ is a combinatorial model category
and $W_{\mathbb{M}}$ is the collection of weak equivalences.
The $\infty$-category
$\mathbb{M}^c[W^{-1}]:=\NNNN(\mathbb{M}^c)[(\mathbb{M}^c\cap W_{\mathbb{M}})^{-1}]$ is presentable where $\mathbb{M}^c$ is the full subcategory of cofibrant objects.
(When $\mathbb{M}$ is a monoidal model category,
it is convenient to work with
the full subcategory of cofibrant objects $\mathbb{M}^c\subset \mathbb{M}$
instead of $\mathbb{M}$.)
If $\mathbb{M}$ is a stable model category, then $\mathbb{M}^c[W^{-1}]$ is
a stable $\infty$-category (cf. \cite{Tan}).
The homotopy category of $\mathbb{M}^c[W^{-1}]$ coincides
with the homotopy category of the model category $\mathbb{M}$.
If $\mathbb{M}$ is a symmetric monoidal model
category (whose unit object is cofibrant),
$\mathbb{M}^c[W^{-1}]$ is promoted to a
symmetric monoidal $\infty$-category $\mathbb{M}^c[W^{-1}]^\otimes:=\NNNN(\mathbb{M}^c)[(\mathbb{M}^c\cap W_{\mathbb{M}})^{-1}]^{\otimes}$
(see below for symmetric monoidal
$\infty$-categories).
In addition, there is a symmetric monoidal
functor $\tilde{\xi}:\NNNN(\mathbb{M}^c)^\otimes \to \mathbb{M}^c[W^{-1}]^\otimes$ which has $\xi$ as the underlying functor and
satisfies a similar universal property.
If $\mathbb{M}$ is combinatorial, then
the tensor product $\otimes:\mathbb{M}^c[W^{-1}]\times \mathbb{M}^c[W^{-1}]\to \mathbb{M}^c[W^{-1}]$ preserves
small colimits separately in each variable.
Let $\mathbb{L}$ be another symmetric monoidal model category
and let $\phi:\mathbb{M}\to \mathbb{L}$ be a symmetric monoidal functor.
If $\phi$ carries cofibrant objects to cofibrant objects and preserves
weak equivalences between them (e.g. symmetric monoidal left
Quillen functors), it induces a symmetric monoidal functor
$\mathbb{M}^c[W^{-1}]^\otimes\to \mathbb{L}^c[W^{-1}]^\otimes$
of symmetric monoidal $\infty$-categories.
\subsection{}
{\it Symmetric monoidal $\infty$-categories, modules and algebras.}
We use the theory of (symmetric) monoidal $\infty$-categories
developed in \cite{HA}.
A symmetric monoidal $\infty$-category is a coCartesian fibration
$\mathcal{C}^\otimes\to \Gamma$ that satisfies a ``symmetric monoidal
condition'', see \cite[2.1.2]{HA}.
For a symmetric monoidal $\infty$-category $\mathcal{C}^\otimes\to \Gamma$,
we often write $\mathcal{C}$ for the underlying $\infty$-category.
Also, by abuse of notation, we usually use the superscript in $\mathcal{C}^\otimes$ to indicate a symmetric monoidal structure on an $\infty$-category.
For a symmetric monoidal $\infty$-category $\mathcal{C}^\otimes$,
we write $\CAlg(\mathcal{C}^\otimes)$ (or simply
$\CAlg(\mathcal{C})$) for the $\infty$-category
of commutative algebra objects in $\mathcal{C}^\otimes$.
Let $A$ be a commutative ring spectrum, that is, a commutative algebra
object in the category $\SP$ of spectra.
We write $\Mod^\otimes_A$ for the symmetric monoidal $\infty$-category
of $A$-module spectra, (see e.g. \cite{HA}).
We put $\CAlg_A=\CAlg(\Mod_A^\otimes)$.
For an ordinary commutative ring $K$,
we put $\Mod^\otimes_{K}:= \Mod^\otimes_{HK}$ and
$\CAlg_{K}:=\CAlg_{HK}$
where $HK$ is the Eilenberg-MacLane ring spectrum.
Let $K$ be a field of characteristic zero.
Let $\Comp^\otimes(K)$ be the symmetric monoidal
category of cochain complexes of $K$-vector spaces
(the symmetric monoidal structure is given by the tensor product
of cochain complexes).
This category admits a projective combinatorial symmetric monoidal
model structure, whose weak equivalences are quasi-isomorpisms,
and whose cofibrations (resp. fibrations) are monomorphisms
(resp. epimorphisms), see e.g. \cite[Section 2.3]{Hov} or \cite[7.1.2.8]{HA}.
We shall write $\mathsf{D}^\otimes(K)$ for the symmetric monoidal
stable presentable $\infty$-category obtained from $\Comp^\otimes(K)$
by inverting weak equivalences.
According to \cite[7.1.2.12, 7.1.2.13]{HA}, there is a canonical equivalence
$\mathsf{D}^\otimes(K)\simeq \Mod_K^\otimes$.
We refer to $\mathsf{D}^\otimes(K)$ and $\Mod_K^\otimes$
as the (symmetric monoidal) derived $\infty$-category of $K$-vector spaces.
The equivalence $\mathsf{D}^\otimes(K)\simeq \Mod_K^\otimes$
induces $\CAlg(\mathsf{D}^\otimes(K))\simeq \CAlg_{K}=\CAlg(\Mod_K^\otimes)$. Let $\CAlg_K^{dg}$ be the category of commutative differential graded
$K$-algebras. A commutative differential graded
$K$-algebras is a commutative algebra object in $\Comp^\otimes(K)$.
There is a natural forgetful functor $U:\CAlg_K^{dg}\to \Comp(K)$.
The category $\CAlg_K^{dg}$ admits a combinatorial model structure
such that a morphism $f$ is
a weak equivalences (resp. a fibration)
if and only if $U(f)$ is a quasi-isomorphism
(resp. a epimorphism) (here, we use the assumption of characteristic zero) .
If we write $\NNNN(\CAlg_{K}^{dg})[W^{-1}]$ for the $\infty$-category
obtained from $\CAlg_{K}^{dg}$ by inverting weak equivalences,
then there is a canonical equivalences
$\NNNN(\CAlg_{K}^{dg})[W^{-1}]\simeq \CAlg_K$
(see \cite[7.1.4.10, 7.1.4.11]{HA}, \cite[4.5.4.6]{HA}).
We often use these equivalences
\[
\NNNN(\CAlg_{K}^{dg})[W^{-1}]\simeq \CAlg_K \simeq \CAlg(\mathsf{D}^\otimes(K)).
\]
A variety is a geometrically connected
scheme separated of finite type over a field.
\section{Cohomological motivic algebras}
\label{CMAsection}
Let $K$ be a commutative ring.
\subsection{}
As in our previous works,
we use $\infty$-categories of mixed motives.
They are obtained from the model (dg, etc) categories of motives
or the $\infty$-categorical version of Voevodsky's construction.
In this paper, we adopt symmetric monoidal
model categories constructed by Cisinski and D\'eglise \cite{CD1}, \cite{CDT}.
Let $X$ be a smooth scheme separated of finite type over
a perfect field $k$ (or more generally, a noetherian regular scheme).
Let $\Sm_X$ denote the category of smooth schemes separated of finite type
over $X$.
Let $\mathcal{N}^{tr}(X)$ be the Grothendieck abelian category of
Nisnevich sheaves of $K$-modules with transfers over $X$ (see e.g.
\cite[Example 2.4]{CD1} or \cite{CDT} for this notion).
Let $\Comp(\mathcal{N}^{tr}(X))$ be the symmetric monoidal category
of (possibly unbounded) cochain complexes of $\mathcal{N}^{tr}(X)$.
Then $\Comp(\mathcal{N}^{tr}(X))$
admits a stable symmetric monoidal combinatorial
model category structure,
see \cite[Section 4, Example 4.12]{CD1}.
The construction roughly has two steps: one first defines a certain nice
model structure
whose weak equivalences are quasi-isomorphisms of complexes
of sheaves,
and in the next step one
takes a left Bousfield localization of the model structure
at $\mathbb{A}^1$-homotopy.
Using a generalization of the construction of symmetric spectra,
one can ``stabilize'' the tensor operation with a shifted Tate object
over $X$
and obtains a new category $Sp_{\textup{Tate}}(X)$ from $\Comp(\mathcal{N}^{tr}(X))$
which admits a stable
symmetric monoidal combinatorial model category structure described
in \cite[Proposition 7.13, Example 7.15]{CD1}.
Let $\phi:Y\to X$ be a morphism of smooth schemes.
It gives rise to a Quillen adjunction
\[
\phi^*:Sp_{\textup{Tate}}(X)\rightleftarrows Sp_{\textup{Tate}}(Y):\phi_*
\]
where $\phi^*$ is a symmetric monoidal left Quillen functor.
We suppose further
that $\phi$ is smooth separated of finite type,
then there is a Quillen
adjunction
\[
\phi_\sharp:Sp_{\textup{Tate}}(Y)\rightleftarrows Sp_{\textup{Tate}}(X):\phi^*.
\]
In this case, $\phi^*$ is both a left Quillen functor and a right Quillen
functor. Thus, it preserves (trivial) fibrations and (trivial) cofibrations.
Moreover, by using Ken Brown's lemma we see that $\phi^*$ preserves arbitrary
weak equivalences.
We let $\DM^\otimes_{eff}(X)$ be the symmetric monoidal
stable presentable $\infty$-category, which is obtained from
the full subcategory of cofibrant objects $\Comp(\mathcal{N}^{tr}(X))^c$
by inverting weak equivalences. We refer to it as the symmetric monoidal
$\infty$-category of effective mixed motives over $X$.
Similarly, $\DM^\otimes(X)$
is defined to be the symmetric monoidal
stable presentable $\infty$-category
obtained from $Sp_{\textup{Tate}}(X)^c$
by inverting weak equivalences.
We call $\DM^\otimes(X)$ the symmetric monoidal
stable persentable $\infty$-category of mixed motives over $X$.
We refer to $K$ as the coefficient ring of $\DM^\otimes(X)$.
We write $\uni_X$ for a unit object of $\DM^\otimes(X)$.
We write $\uni_X(n)$ for the Tate object for $n\in \ZZ$.
Given an object $M$ of $\DM(X)$, we usually write $M(n)$ for
the tensor product $M\otimes \uni_X(n)$ in $\DM(X)$.
The tensor product $\DM(X)\times \DM(X)\to \DM(X)$
on $\DM^\otimes(X)$
preserves small colimits separately in each variable.
The detail construction can be found in \cite[Section 5.1]{Tan}
(the notation is slightly different,
and $X$ is assumed to be the Zariski spectrum of
a perfect field in \cite{Tan}, but it works for a noetherian regular scheme $X$). The homotopy category of the
full subcategory of $\DM(\Spec k)$ spanned by compact objects
can be identified with the triangulated category of geometric motives constructed by Voevodsky \cite{Voe}.
Let $f:X \to \Spec k$ be the structure morphism.
Since
we have the restriction of the symmetric monoidal left Quillen functor
$f^*:Sp_{\textup{Tate}}(\Spec k)^c\rightarrow Sp_{\textup{Tate}}(X)^c$
between full subcategories of cofibrant objects,
inverting weak equivalences
we have a symmetric monoidal colimit-preserving functor
\[
f^*:\DM^\otimes(k):=\DM^\otimes(\Spec k)\to \DM^\otimes (X).
\]
By abuse of notation, we use the same notation
for the induced functor between $\infty$-categories.
By relative adjoint functor theorem \cite[7.3.2.6, 7.3.2.13]{HA}, there is
the right adjoint functor $f_*:\DM(X)\to \DM(k)$ that is lax
symmetric monoidal.
It induces an adjunction
\[
f^*:\CAlg(\DM^\otimes(k))\rightleftarrows \CAlg(\DM^\otimes(X)):f_*.
\]
In particular, $f_*$ carries a commutative algebra object $M$ to a commutative
algebra object $f_*(M)$ in $\DM^\otimes(k)$.
For any smooth scheme $X$,
$\CAlg(\DM^\otimes(X))$
is a presentable $\infty$-category (cf. \cite[3.2.3.5]{HA}).
There is another left Quillen functor
$f_\sharp:Sp_{\textup{Tate}}(X)\to Sp_{\textup{Tate}}(\Spec k)$.
The restriction $Sp_{\textup{Tate}}(X)^c\to Sp_{\textup{Tate}}(\Spec k)^c$
to cofibrant objects preserves weak equivalences, and therefore
inverting weak equivalences induces
$f_\sharp:\DM(X)\to \DM(k)$. It determines an adjunction
\[
f_\sharp:\DM(X)\rightleftarrows \DM(k):f^*.
\]
We put $M(X):=f_\sharp f^*(\uni_k)$ where $\uni_k$ is the
unit of $\DM(k)$.
Let us consider the unit object $\uni_X=f^*(\uni_k)$ in $\DM^\otimes(X)$
which we regard as a commutative algebra object in $\DM^\otimes(X)$.
The image $f_*(\uni_X)=f_*f^*(\uni_k)$ is a commutative algebra object in
$\DM^\otimes(k)$, namely, $f_*(\uni_X)$ in $\CAlg(\DM^\otimes(k))$.
\begin{Definition}
We define $M_X$ in $\CAlg(\DM^\otimes(k))$ to be $f_*(\uni_X)$.
We shall refer to $M_X$ as
the {\it cohomological motivic algebra} of $X$ with
coefficients in $K$.
\end{Definition}
\begin{Remark}
This algebra $M_X$ will play a role of a motivic analogue of
the singular cochain complex $C^*(S,K)$ of a topological space $S$
that is endowed with a structure of an $E_\infty$-algebra.
Our principle is that one may consider $M_X$ to be a motivic homotopy
type of $X$ with coefficients in $K$, that occurs in the title of this paper.
On the other hand, $M(X)$ is
a motivic counterpart of the singular chain complex
$C_*(S,K)$.
To our knowledge, contrary to homotopy theory,
there has been little
attention being paid to invariants represented by highly structured algebras
in the theory of motives.
\end{Remark}
\subsection{}
\label{cohomologicalmot}
We consider functoriality of motivic cohomological algebras.
Let $f:X\to \Spec k$ and
$g:Y\to \Spec k$ be two smooth scheme separated of finite type
over $k$.
Let $\phi:Y\to X$ be a morphism over $k$.
As above, there is an adjunction $\phi^*:\CAlg(\DM^\otimes(X))\rightleftarrows \CAlg(\DM^\otimes(Y)):\phi_*$.
If we write $M_Y$ for $g_*(\uni_Y)$ we have a morphism
\[
M_X=f_*(\uni_X)\to f_*\phi_*\phi^*(\uni_X)\simeq g_*(\uni_Y)=M_Y
\]
in $\CAlg(\DM^\otimes(k))$
where the first map is induced by the unit map $\uni_X\to \phi_*\phi^*(\uni_X)\simeq \phi_*(\uni_Y)$. Thus, the assignment $X\mapsto M_X$ is contravariantly functorial
with respect to $X$.
We will write $\phi^*:M_X\to M_Y$ for this morphism in $\CAlg(\DM^\otimes(k))$
or in the underlying category $\DM(k)$.
Unfortunately, the notation
$\phi^*$ in $\phi^*:M_X\to M_Y$ overlaps
with $\phi^*:\DM(X)\to \DM(Y)$ or $\phi^*:\CAlg(\DM^\otimes(X)) \to \CAlg(\DM^\otimes(Y))$ though these have different meanings.
We hope that it causes no confusion.
The assignment $X\mapsto M(X)$ is covariantly functorial.
For $\phi:Y\to X$, consider the unit map $u:\uni_X\to f^*f_\sharp(\uni_X)$.
We then have
\[
M(Y)=g_\sharp (\uni_Y)\simeq g_\sharp \phi^*(\uni_X)\stackrel{g_\sharp\phi^*(u)}{\longrightarrow} g_\sharp \phi^* f^*f_\sharp(\uni_X)\simeq g_\sharp g^* f_\sharp(\uni_X)\to f_\sharp(\uni_X)=M(X)
\]
where the final arrow is induced by the counit $g_\sharp g^*\to \textup{id}$.
Let $\textup{Sm}_k$ be the nerve of the category of smooth schemes
separated of finite type over $k$.
We will give a functorial construction $X\mapsto M_X$
as a functor $\textup{Sm}_k^{op}\to \CAlg(\DM^\otimes(k))$.
The result is summarized as follows:
\begin{Proposition}
\label{cohmotalg}
Let $M(-):\Sm_k\to \DM(k)$ be the functor which carries $X$ to $M(X)$.
We define $\HOM_{\DM(k)}(-,\uni_k):\DM(k)^{op}\to \DM(k)$
to be the functor which carries $M$ to $M^\vee=\HOM_{\DM(k)}(M,\uni_k)$. By $\HOM(-,-)$ we indicate the internal Hom object.
(We will make a construction of these functors below.)
Let $M(-)^\vee:\Sm_k^{op}\to \DM(k)$ be the composite of the above two functors, which carries $X$ to $M(X)^\vee$.
Then there is a functor
$\Xi:\textup{Sm}_k^{op}\to \CAlg(\DM^\otimes(k))$
which makes the diagram commutative
\[
\xymatrix{
& \CAlg(\DM^\otimes(k)) \ar[d] \\
\Sm_k^{op} \ar[r]_{M(-)^\vee} \ar[ru]^{\Xi} & \hhh(\DM(k))
}
\]
where the right vertical arrow is the forgetful functor.
\end{Proposition}
We first construct $\Xi:\textup{Sm}_k^{op}\to \CAlg(\DM^\otimes(k))$.
The busy readers are invited to skip the remainder for the time being
and proceed to Section~\ref{somerem} or~\ref{someex}.
We consider the following general situation.
The functor $\Xi$ will appear in Example~\ref{motivicex}
as an example of the following setup.
Let $I$ be the nerve of a category.
Suppose that $I$ has a final object $\star\in I$.
We are mainly interested in the case $I=\textup{Sm}_k$.
Let us consider a family $\{\mathbb{M}(X)\}_{X\in I}$
of symmetric monoidal model categories
indexed by $I$. More precisely, we assign
a combinatorial symmetric monoidal model category $\mathbb{M}(X)$
to any $X\in I$
(we here assume that a unit is cofibrant)
and assign a symmetric monoidal left Quillen functor
$\phi^*:\mathbb{M}(X)\to \mathbb{M}(Y)$ to any morphism
$Y\to X$ in $I$.
Moreover, suppose that
for $\phi\circ \psi:Z\to Y\to X$ there is
a structural
natural equivalence
$\psi^*\phi^*\simeq (\phi\circ \psi)^*$.
Main example is the family
$\{Sp_{\textup{Tate}}(X)\}_{X\in \Sm_k}$.
Consider the pair $(\mathbb{M}^c(X), W_X^c)$
such that $\mathbb{M}^c(X)$ is the full subcategory of
cofibrant objects in the model category $\mathbb{M}(X)$,
and $W_X^c$ is the collection of weak equivalences in $\mathbb{M}(X)^c$.
We think of this pair as the nerve of a category $\mathbb{M}^c(X)$
endowed with the collection of morphisms, determined by $W_X^c$.
We apply to the assignment $X\mapsto (\mathbb{M}(X)^c, W_X^c)$
the construction in \cite[Section 4.1.3.1, 4.1.3.2]{HA}
of inverting weak equivalences in symmetric monoidal
categories in the functorial way.
We then get a functor
\[
d:I^{op}\to \CAlg(\wCat)
\]
which carries $X$ to $\mathbb{M}_\infty^\otimes(X):=\mathbb{M}^c(X)[(W_X^c)^{-1}]$. Here $\mathbb{M}^c(X)[(W_X^c)^{-1}]$ is the symmetric monoidal
$\infty$-category obtained from $\mathbb{M}(X)^c$
by inverting $W_X^c$.
The symmetric monoidal structure on $\wCat$ is given by cartesian products,
and $\CAlg(\wCat)$ is naturally identified with the $\infty$-category of
symmetric monoidal (large) $\infty$-categories whose morphisms are
symmetric monoidal functors, cf. \cite{HA}.
Recall that the $\infty$-category $\CAlg(\wCat)$
can be realized as the full subcategory of $\Fun(\Gamma,\wCat)$
spanned by commutative monoid objects,
where $\Gamma$ is the nerve of the category of pointed finite sets.
The functor $d:I^{op}\to \CAlg(\wCat)\subset \Fun(\Gamma,\wCat)$
induces a functor $I^{op}\times \Gamma\to \wCat$.
Applying the relative nerve functor to $I^{op}\times \Gamma\to \wCat$
(cf. \cite[3.2.5]{HTT}),
we have a coCartesian fibration
\[
D:\mathcal{E}\to I^{op}\times \Gamma
\]
such that each restriction $\mathcal{E}_X:=D^{-1}(\{X\}\times \Gamma)\to \{X\}\times \Gamma$
is a symmetric monoidal $\infty$-category equivalent to
$\mathbb{M}^\otimes_\infty(X)$.
Let $P:\overline{\CAlg}(\mathcal{E})\to I^{op}$ be a map of simplicial sets
defined as follows.
For $q:K\to I^{op}$, the set of $K\to \overline{\CAlg}(\mathcal{E})$
over $q$ is defined to be the set of maps $K\times \Gamma\to \mathcal{E}$
extending $q\times \textup{id}:K\times \Gamma\to I^{op}\times \Gamma$.
Namely, it is $\Fun(\Gamma,\mathcal{E})\times_{\Fun(\Gamma,I^{op}\times \Gamma)}I^{op}\stackrel{\textup{pr}_2}{\to} I^{op}$
where $I^{op}\to \Fun(\Gamma,I^{op}\times \Gamma)$
is induced by the identity of $I^{op}\times \Gamma$.
By the stability property \cite[3.1.2.1 (1), 2.4.2.3. (2)]{HTT}
of coCartesian fibrations,
$\overline{\CAlg}(\mathcal{E})\to I^{op}$ is a coCartesian fibration.
Let $\CAlg(\mathcal{E})$
be the largest subcomplex of $\overline{\CAlg}(\mathcal{E})$
that consists of those vertices $v\in \overline{\CAlg}(\mathcal{E})$ such that
$\{P(v)\}\times \Gamma \to \mathcal{E}$
determines a commutative algebra object of $\mathcal{E}_{P(v)}$.
According to \cite[3.1.2.1 (2)]{HTT}
the induced map
$\CAlg(\mathcal{E})\to I^{op}$ is also a coCartesian fibration.
Note that by the construction,
for each $X$ in $I$ the fiber over $X$ is $\CAlg(\mathcal{E}_X)\simeq \CAlg(\mathbb{M}_\infty^\otimes(X))$, and for each $\phi:Y\to X$
in $I$ the induced map $\CAlg(\mathbb{M}_\infty^\otimes(X))\to \CAlg(\mathbb{M}_\infty^\otimes(Y))$
is equivalent to the pullback functor $\phi^*$.
Each $(\mathbb{M}(X)^c,W_X^c)$ admits
a symmetric monoidal functor $(\mathbb{M}^c(\star),W_{\star}^c)\to (\mathbb{M}(X)^c,W_X^c)$ induced by the morphism $X\to \star$, which preserves weak equivalences.
If $d_\star:I^{op}\to \CAlg(\wCat)$ denotes the constant
functor taking value $\mathbb{M}_\infty^\otimes(\star)$,
it gives rise to a natural transformation $d_\star \to d$.
By using the relative nerve functor as above,
one has a map between coCartesian
fibrations
\[
\xymatrix{
I^{op} \times \mathbb{M}_\infty^\otimes(\star) \ar[rr]^{F^*_\circ} \ar[rd]_{\textup{id}\times e}& & \mathcal{E} \ar[ld]^{D} \\
& I^{op}\times \Gamma
}
\]
where $e:\mathbb{M}_\infty^\otimes(\star)\to \Gamma$ is a coCartesian fibration
that determines the symmetric monoidal $\infty$-category $\mathbb{M}_\infty^\otimes(\star)$.
The horizontal map preserves coCartesian edges.
Apply the same construction of $\CAlg(\mathcal{E})\to I^{op}$
to $I^{op} \times \mathbb{M}_\infty^\otimes(\star)\to I^{op}\times \Gamma$,
we obtain the constant coCartesian
fibration $I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star))\to I^{op}$
and a map of coCartesian fibrations
\[
F^*:I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star)) \to \CAlg(\mathcal{E})
\]
over $I^{op}$.
For each $f:X\to \star$ in $I$, the fiber $\CAlg(\mathbb{M}_\infty^\otimes(\star)) \to \CAlg(\mathcal{E}_X)\simeq \CAlg(\mathbb{M}_\infty^\otimes(X))$ over $X$ is equivalent to $f^*$.
Thus, each fiber admits the right adjoint functor $f_*:\CAlg(\mathbb{M}_\infty^\otimes(X)) \to \CAlg(\mathbb{M}_\infty^\otimes(\star))$.
In addition, $F^*$ preserves coCartesian edges.
Therefore by the relative adjoint functor theorem \cite[7.3.2.6]{HA}
there is a relative right adjoint $F_*:\CAlg(\mathcal{E})\to I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star))$ over $I^{op}$.
(We refer to \cite[7.3.2]{HA} for the notion of relative adjoint functor.)
For each $f:X\to \star$,
the fiber $\CAlg(\mathbb{M}_\infty^\otimes(X))\to \CAlg(\mathbb{M}_\infty^\otimes(\star))$ is equivalent to
$f_*$.
Now we define a functorial assignment $X\mapsto f_*(\uni_{\mathbb{M}(X)})$
where $\uni_{\mathbb{M}(X)}$ is a unit of $\mathbb{M}(X)$
and $f$ is the natural morphism $X\to \star$.
We let $\iota:I^{op}\to \CAlg(\mathbb{M}_\infty^\otimes(\star))$
be the constant functor whose value is the unit $\uni_\star$
of $\mathbb{M}_\infty^\otimes(\star)$.
It yields a section $\textup{id}\times \iota:I^{op}\to I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star))$.
Composing it with $F^*$,
we obtain a section $S:I^{op}\to \CAlg(\mathcal{E})$ of $\CAlg(\mathcal{E})\to I^{op}$ which carries $X$ to a unit in $\CAlg(\mathcal{E}_X)\simeq \CAlg(\mathbb{M}_\infty^\otimes(X))$ (every edge in $I^{op}$ maps
to a canonical coCartesian edge).
We define $I^{op}\to \CAlg(\mathbb{M}_\infty^\otimes(\star))$
to be the composite
\[
\Phi:I^{op}\stackrel{S}{\to} \CAlg(\mathcal{E})\stackrel{F_*}{\to} I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star)) \stackrel{\textup{pr}_2}{\to} \CAlg(\mathbb{M}_\infty^\otimes(\star)).
\]
\begin{Remark}
\label{sectionadjoint}
We give a little bit more conceptual explanation of $\Phi$.
Let $\mathcal{C}\to \mathcal{O}$ and $\mathcal{D}\to \mathcal{O}$
be categorical fibrations over an $\infty$-category $\OO$. Let
$\alpha:\mathcal{C}\rightleftarrows \mathcal{D}:\beta$ be functors over $\OO$.
Suppose that $\alpha$ is a left adjoint to $\beta$.
Observe that compositions with $\alpha$ and $\beta$ induce an adjoint pair
between functor categories
$\Fun(\OO,\mathcal{C})\rightleftarrows \Fun(\OO,\mathcal{D})$. To see this,
if $\mathcal{M}\to \Delta^1$ is both a coCartesian fibration and a Cartesian
fibration
which represents the adjoint pair $(\alpha,\beta)$ (cf. \cite[5.5.2.1]{HTT}),
the projection
$\Fun(\OO,\mathcal{M})\times_{\Fun(\OO,\Delta^1)}\Delta^1\to \Delta^1$
is both a coCartesian fibration and a Cartesin fibration
that induces an adjoint pair between functor categories,
where $\Delta^1\to \Fun(\OO,\Delta^1)$ is determined by the projection
$\OO\times \Delta^1\to \Delta^1$.
Suppose further that $\alpha$ is a left adjoint to $\beta$ relative to
$\OO$ (cf. \cite[7.3.2.2]{HA}).
The restriction of the above adjunction induces
\[
\textup{Sect}(\alpha):\textup{Sect}_{\OO}(\mathcal{C}):=\Fun_{\OO}(\OO,\mathcal{C}) \rightleftarrows \Fun_{\OO}(\OO,\mathcal{D})=\textup{Sect}_{\OO}(\mathcal{D}):\textup{Sect}(\beta).
\]
We deduce from \cite[7.3.2.5]{HA} that this pair is an adjunction.
We now apply this to
\[
F^*:I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star)) \rightleftarrows \CAlg(\mathcal{E}):F_*
\]
over $I^{op}$.
We then have the induced adjunction
\[
\textup{Sect}(F^*):\Fun(I^{op},\CAlg(\mathbb{M}_\infty^\otimes(\star))) \simeq \textup{Sect}_{I^{op}}(I^{op}\times \CAlg(\mathbb{M}_\infty^\otimes(\star))) \rightleftarrows \textup{Sect}_{I^{op}}(\CAlg(\mathcal{E})):\textup{Sect}(F_*).
\]
If $\iota\in \Fun(I^{op},\CAlg(\mathbb{M}_\infty^\otimes(\star)))$
is the constant functor with value $\uni_\star$,
the unit transformation $\textup{id}\to \textup{Sect}(F_*)\circ \textup{Sect}(F^*)$ induces $\iota \to \textup{Sect}(F_*)\circ \textup{Sect}(F^*)(\iota)=\Phi$.
\end{Remark}
\begin{Example}
\label{motivicex}
Let $I=\Sm_k$ and $\star=\Spec k$.
Let $\mathbb{M}(X)=Sp_{\textup{Tate}}(X)$.
We define
\[
\Xi:\Sm_k^{op}\to \CAlg(\DM^\otimes(k))
\]
to be $\Phi$.
Unfolding our construction we see that $\Xi$ carries $X$ to $M_X$, and $\phi:Y\to X$ maps to
$\phi^*:M_X\to M_Y$.
\end{Example}
\begin{Remark}
Let $I=\Sm_k$ and $\star=\Spec k$.
Let $\mathbb{M}(X)=\Comp(\mathcal{N}^{tr}(X))$.
In this case, the above construction also works.
But we will not consider this setting: $f_*(\uni_X)$ is not an appropriate object
we want to consider (for example, Theorem~\ref{real1} does not hold).
\end{Remark}
\begin{Example}
\label{topcomplex}
Let $I$ be the category $\textup{Sch}$ of
separated and quasi-compact schemes.
For any $X$ in $\textup{Sch}$, we let $\Comp(X)$ be
the symmetric monoidal category of (possibly unbounded) cochain
complexes of quasi-coherent sheaves on $X$.
According to \cite[Example 2.3, 3.1, 3.2]{CD1}, there is a symmetric monoidal
model structure on $\Comp(X)$ such that weak equivalences are
quasi-isomorphisms, and for any $Y\to X$ in $\textup{Sch}$
the pullback functor $\Comp(X)\to \Comp(Y)$ is a
left Quillen functor.
Put $\Comp(X)=\mathbb{M}(X)$.
One can apply to this setting our construction and obtain
$\textup{Sch}^{op} \to \CAlg(\mathbb{M}_\infty^\otimes(\Spec \ZZ))$.
\end{Example}
Next we define a functor $\Sm_k\to \DM(k)$ which carries $X$ to $M(X)$.
In some sense, the construction is the dual of that of $\Xi$ and is easier.
We continue to work with the family $\{\mathbb{M}(X)\}$.
Assume that for each $f:X\to \star$ in $I$,
$f^*:\mathbb{M}(\star)\to \mathbb{M}(X)$ is also right Quillen functor
(therefore, it preserves arbitrary weak equivalences).
We denote by $f_\sharp:\mathbb{M}(X)\to \mathbb{M}(\star)$
the left adjoint.
Applying the ``dual version'' of the relative nerve functor or the unstraightning functor
to $X\mapsto \mathbb{M}_\infty(X)$, we obtain a Cartesian fibration $\mathcal{F}\to I$. For each $X\in I$, its fiber is equivalent to $\mathbb{M}_\infty(X)$.
Notice that it is not a coCartesian fibration but a Cartesian fibration.
As in the case of $\mathcal{E}\to I^{op}$, the
natural pullback functors $\mathbb{M}(\star)\to \mathbb{M}(X)$
induce a morphism of Cartesian fibrations
\[
\xymatrix{
\mathcal{F} \ar[rd] & & I \times \mathbb{M}_\infty(\star) \ar[ld] \ar[ll]_{G^*} \\
& I.
}
\]
where $I \times \mathbb{M}_\infty(\star)\to I$ is the projection
that is regarded as a Cartesian fibration corresponding to the constant functor
$I\to \wCat$ with value $\mathbb{M}_\infty(\star)$.
Each fiber of the horizontal map over $X\in I$ is equivalent to $f^*$
where $f:X\to \star$ is the natural morphism.
Therefore it admits a left
adjoint functor $f_\sharp:\mathbb{M}_\infty(X)\to \mathbb{M}_\infty(\star)$.
Moreover, $G^*$ preserves Cartesian edges.
Thus, by the relative adjoint functor theorem \cite[7.3.2.6]{HA}
there is a left adjoint $G_\sharp:\mathcal{F}\to I \times \mathbb{M}_\infty(\star)$ relative to $I$. (Its fiber over $X\in I$ is equivalent to $f_\sharp$.)
Let $u:I\to I\times \mathbb{M}_\infty(\star)$ be the functor
determined by the identity $I\to I$ and the constant functor
$I\to \mathbb{M}_\infty(\star)$ taking the value $\uni_\star$.
Then $\Psi:I\to \mathbb{M}_\infty(\star)$ is defined to be the composite
\[
I\stackrel{u}{\to} I\times \mathbb{M}_\infty(\star)\stackrel{G^*}{\to}\mathcal{F}\stackrel{G_\sharp}{\to} I\times \mathbb{M}_\infty(\star)\stackrel{\operatorname{pr}_2}{\to}\mathbb{M}_\infty(\star).
\]
\begin{Example}
\label{motivicex2}
Let $I=\Sm_k$ and $\star=\Spec k$.
Let $\mathbb{M}(X)=Sp_{\textup{Tate}}(X)$.
We define $M(-):\Sm_k \to \DM(k)$
to be $\Psi$.
By our construction, it sends $X$ to an object equivalent to
$M(X)$.
\end{Example}
We define a functor
$\HOM_{\DM(X)}(-,\uni_X):\DM(X)^{op}\to \DM(X)$
as follows. We let
\[
\HOM_{Sp_{\textup{Tate}}(X)}(-,\uni_X'):(Sp_{\textup{Tate}}(X)^{c})^{op} \to Sp_{\textup{Tate}}(X)
\]
be the functor given by $M\mapsto \HOM_{Sp_{\textup{Tate}}(X)}(M,\uni_X')$, where $\HOM_{Sp_{\textup{Tate}}(X)}(-,-)$ denotes the
internal Hom object in $Sp_{\textup{Tate}}(X)$, and
$\uni_X'$ is a fibrant model of the unit $\uni_X$.
By the axiom of symmetric monoidal model category,
the functor $\HOM_{Sp_{\textup{Tate}}(X)}(-,\uni_X')$
preserves weak equivalences. We define
$\HOM_{\DM(X)}(-,\uni_X):\DM(X)^{op}\to \DM(X)$
to be the functor obtained from
$\HOM_{Sp_{\textup{Tate}}(X)}(-,\uni_X')$
by inverting weak equivalences.
\vspace{2mm}
{\it Proof of Proposition~\ref{cohmotalg}.}
We have constructed the functor $\Xi:\Sm_k^{op}\to \CAlg(\DM^\otimes(k))$
and $M(-):\Sm_k\to \DM(k)$
in Example~\ref{motivicex} and~\ref{motivicex2}.
For simplicity, we write $\Xi$ also for the composite $\Sm_k^{op}\stackrel{\Xi}{\to} \CAlg(\DM^\otimes(k))\to \DM(k)$.
We first observe that for $f:X\to \Spec k$ in $\Sm_k$
there is a canonical equivalence $M(X)^\vee\stackrel{\sim}{\to} M_X=f_*(\uni_X)$. Actually, this equivalence follows from the equivalences of mapping spaces
\begin{eqnarray*}
\Map_{\DM(k)}(M,\HOM_{\DM(k)}(f_\sharp\uni_X,\uni_k)) &\simeq& \Map_{\DM(k)}(M\otimes f_\sharp\uni_X,\uni_k) \\
&\simeq& \Map_{\DM(k)}(f_\sharp(\uni_X),\HOM_{\DM(k)}(M,\uni_k)) \\
&\simeq& \Map_{\DM(X)}(\uni_X,f^*\HOM_{\DM(k)}(M,\uni_k)) \\
&\simeq& \Map_{\DM(X)}(\uni_X,\HOM_{\DM(X)}(f^*(M),f^*(\uni_k))) \\
&\simeq& \Map_{\DM(X)}(f^*(M),\uni_X) \\
&\simeq& \Map_{\DM(k)}(M,f_*(\uni_X))
\end{eqnarray*}
for any $M\in \DM(k)$.
We here used adjunctions $(f_\sharp,f^*)$, $(f^*,f_*)$ and
$f^*\HOM_{\DM(k)}(M,\uni_k)\simeq \HOM_{\DM(X)}(f^*(M),f^*(\uni_k))$.
If we take $M=\HOM_{\DM(k)}(f_\sharp\uni_X,\uni_k)=M(X)^\vee$, then
the identity of $M$ corresponds to $M(X)^\vee\stackrel{\sim}{\to} f_*(\uni_X)$.
The equivalence $M(X)^\vee=f_\sharp(\uni_X)^\vee\to f_*(\uni_X)$
comes from the dual of $\uni_X\to f^*f_\sharp(\uni_X)$:
\[
f^*(f_\sharp(\uni_X)^\vee)\simeq (f^*f_\sharp(\uni_X))^\vee \to \uni_X
\]
and the composition with
$f_\sharp(\uni_X)^\vee\to f_*f^*(f_\sharp(\uni_X)^\vee)$
where $(-)^\vee$ denotes
the weak dual, that is,
$\HOM_{\DM(-)}(-,\uni_{(-)})$.
By the functoriality of adjoint maps,
it is easy to check that $M(X)^\vee=f_\sharp(\uni_X)^\vee\to f_*(\uni_X)$
is functorial with respect to $X\in \Sm_k$ at the level of
homotopy category, namely,
the functor $\Sm_k^{op}\stackrel{M(-)^{\vee}}{\to}\DM(k)\to \hhh(\DM(k))$
is naturally equivalent to
$\Sm_{k}^{op}\stackrel{\Xi}{\to} \DM(k)\to \hhh(\DM(k))$.
\QED
\begin{Remark}
There should be several approaches to a generalization to singular varieties.
One possible way is to use cubical hyperresolutions of singular varieties \cite{GNPP}
when $k$ is a field of characteristic zero,
so that Hironaka's resolution of singularities is available.
Another one is to adopt a formalism, which works for singular varieties,
such as Beilinson motives \cite{CDT}
when the coefficient ring $K$ is a field of characteristic zero.
But we will not treat singular schemes in this paper.
\end{Remark}
\subsection{}
\label{somerem}
We give some remarks about properties of cohomological motivic algebras.
\begin{Remark}
Since $M_X$ is the weak dual $\Hom_{\DM(k)}(M(X),\uni_k)$ of $M(X)$,
one can observe that $X\mapsto M_X$ satisfies $\mathbb{A}^1$-homotopy
invariance and Nisnevich descent property. Namely, for any
projection $X\times \mathbb{A}^1\to X$ with fiber of the affine line $\mathbb{A}^1=\Spec k[x]$, $M_{X}\to M_{X\times \mathbb{A}^1}$ is an equivalence in $\CAlg(\DM^\otimes(k))$.
For any pullback diagram
\[
\xymatrix{
V\simeq U\times_{X}Y \ar[r] \ar[d] & Y \ar[d]^{f} \\
U \ar[r]^{j} & X
}
\]
in $\Sm_k$ such that $f$ is \'etale, $j$ is an open immersion and
$(Y\backslash V)_{red} \to (X\backslash U)_{red}$
is an isomorphism, the induced morphism
$M_X \to M_U\times_{M_V}M_Y$ is an equivalence in $\CAlg(\DM^\otimes(k))$.
\end{Remark}
The following is the Kunneth formula for cohomological motivic algebras.
\begin{Proposition}
\label{kunneth}
There exist a canonical equivalence
$\uni_k\simeq \Xi(\Spec k)$.
Suppose that $X$ and $Y$ are projective and smooth over $\Spec k$.
Then there exists
a canonical equivalence $\Xi(X)\otimes \Xi(Y)=M_X\otimes M_Y \simeq M_{X\times Y}=\Xi(X\times Y)$.
\end{Proposition}
\Proof
The first assertion is obvious.
Next we prove the second assertion.
Consider the Cartesian diagram
\[
\xymatrix{
X\times Y \ar[r]^q \ar[d]_p & Y \ar[d]^g \\
X \ar[r]_f & \Spec k.
}
\]
We will prove that
$p^*\otimes q^*:M_{X}\otimes M_Y\to M_{X\times Y}$
induced by
$p^*:M_X\to M_{X\times Y}$ and $q^*:M_Y\to M_{X\times Y}$
is an equivalence.
For this purpose,
we apply the projection formula for the smooth proper morphism $f$
and the base change theorem for smooth proper morphism $g$
\cite[Theorem 1]{CDT}: we have the sequence of morphisms
induced by unit maps and counit maps of adjunctions
\begin{eqnarray*}
f_*(\uni_X)\otimes g_*(\uni_Y)&\to& f_*f^*(f_*(\uni_X)\otimes g_*(\uni_Y)) \simeq f_*(f^*f_*(\uni_X)\otimes f^*g_*g^*(\uni_k)) \\
&\to& f_*(\uni_X\otimes f^*g_*g^*(\uni_k)) \simeq f_*(f^*g_*g^*(\uni_k)) \\
&\to& f_*(p_*p^*f^*g_*g^*(\uni_k)) \simeq f_*(p_*q^*g^*g_*g^*(\uni_k)) \\
&\to& f_*(p_*q^*g^*(\uni_k)) \simeq f_*p_*(\uni_{X\times Y}) \\
\end{eqnarray*}
whose composite is an equivalence since the projection formula and the base change
theorem imply that the above sequence induces
$f_*(\uni_X)\otimes g_*(\uni_X) \simeq f_*(\uni_X\otimes f^*g_*(\uni_Y))$ and $f^*g_*(\uni_Y)\simeq p_*q^*(\uni_Y)$. It will suffice to check
that this composite coincides with $p^*\otimes q^*$.
It is straightforward to verify that
\[
f^*(\uni_k)\to f^*g_*g^*(\uni_k)\to p_*p^*f^*g_*g^*(\uni_k)=p_*q^*g^*g_*g^*(\uni_k)\to p_*q^*g^*(\uni_k)=p_*p^*f^*(\uni_k)
\]
is equivalent to $f^*(\uni_k)\to p_*p^*f^*(\uni_k)$
induced by the unit map $\textup{id} \to p_*p^*$.
Then we see that $f_*(\uni_X)\otimes \uni_k\to f_*(\uni_X)\otimes g_*(\uni_X)\stackrel{\sim}{\to}f_*p_*p^*(\uni_{X})$ is equivalent to $p^*:M_X=f_*(\uni_X)\to f_*p_*p^*(\uni_X)=M_{X\times Y}$.
Similarly,
$\uni_k\otimes g_*(\uni_Y) \to f_*(\uni_X)\otimes g_*(\uni_X)\stackrel{\sim}{\to}f_*p_*p^*(\uni_{X})$ is equivalent to $q^*:M_Y\to M_{X\times Y}$.
Thus, $p^*\otimes q^*:M_X\otimes M_Y\to M_{X\times Y}$ is an equivalence.
\QED
\subsection{}
\label{someex}
We will study various objects in $\CAlg(\DM^\otimes(k))$
other than $M_X$:
\begin{Example}
\label{motpath}
Let $X\in \Sm_k$.
Let $x:Y=\Spec k\to X$ and $y:Z=\Spec k\to X$ be two $k$-rational points
on $X$.
Then we have the pushout diagram
\[
\xymatrix{
M_X \ar[r]^{x^*} \ar[d]_{y^*} & M_{\Spec k} \ar[d] \\
M_{\Spec k} \ar[r] & M_{\Spec k}\otimes_{M_X}M_{\Spec k}.
}
\]
in $\CAlg(\DM^\otimes(k))$. Keep in mind that pushouts
in $\CAlg(\DM^\otimes(k))$ do not commute with pushouts in $\DM(k)$
through the forgetful functor.
By Proposition~\ref{kunneth}, $M_{\Spec k}\simeq \uni_k$.
Thus,
$M_{\Spec k}\otimes_{M_X}M_{\Spec k}\simeq \uni_k\otimes_{M_X}\uni_k$ in $\CAlg(\DM^\otimes(k))$. We call
$P_{X}(x,y):=\uni_k\otimes_{M_X}\uni_k$ the {\it motivic algebra of path torsors} from $x$
to $y$.
\end{Example}
\begin{Example}
\label{motloop}
Consider $M_X\otimes_{M_{X}\otimes M_X}M_X$.
Note that $\CAlg(\DM^\otimes(k))$ is presentable, and thus
$\CAlg(\DM^\otimes(k))$ is tensored over $\SSS$.
There is a canonical equivalence $S^1\otimes M_X\simeq M_X\otimes_{M_{X}\otimes M_X}M_X$
where $S^1$ is the circle which belongs to $\SSS$.
Thus, by the functoriality of the tensor operation,
$\Map_{\SSS}(S^1,S^1)\simeq S^1$ naturally acts on
$S^1\otimes M_X$ (it is a version of Connes operator; the precise formulation is left to the reader).
We refer to $HHM_X:=M_X\otimes_{M_{X}\otimes M_X}M_X$ as the {\it motivic algebra of free loop space} of $X$.
\end{Example}
\subsection{}
\label{hopfsection}
In Example~\ref{motpath}, if one supposes $x=y$, then
$P_X(x,y)$ has an additional structure.
The augmentation $M_X\to \uni_k\simeq M_{\Spec k}$,
induced by $x:\Spec k\to X$, gives rise to
\[
\uni_k\otimes_{M_X} \uni_k\simeq \uni_k\otimes_{M_X}M_X\otimes_{M_X} \uni_k\to \uni_k\otimes_{M_X}\uni_k\otimes_{M_X} \uni_k\simeq (\uni_k\otimes_{M_X}\uni_k)\otimes (\uni_k\otimes_{M_X}\uni_k)
\]
and $\uni_k\otimes_{M_X}\uni_k\to \uni_k\otimes_{\uni_{k}}\uni_k\simeq \uni_k$
in $\CAlg(\DM^\otimes(k))$. There is also the flip
$\uni_k\otimes_{M_X}\uni_k\simeq \uni_k\otimes_{M_X}\uni_k$.
Informally, these data define a structure of a cogroup object
on $\uni_k\otimes_{M_X}\uni_k$ in $\CAlg(\DM^\otimes(k))$.
Here $\CAlg(\DM^\otimes(k))$ is endowed with the coCartesian monoidal
structure given by coproducts.
The precise formulation of this structure is as follows.
Let $\Delta_+$ be the category of (possibly nonempty)
linearly ordered finite sets.
Objects are the empty set $[-1]$, $[0]=\{0\}, [1]=\{0,1\},[2]=\{0, 1, 2\},\ldots$. Note that $\Delta_+$ without $[-1]$ is $\Delta$.
The morphism $M_X\to \uni_k$ is described by
$\NNNN(\Delta_+^{\le0})=\NNNN(\{[-1]\to [0]\}) \to \CAlg(\DM^\otimes(k))$.
Since $\CAlg(\DM^\otimes(k))$ has small colimits (in fact, presentable),
the map $\NNNN(\Delta_+^{\le0})\to \CAlg(\DM^\otimes(k))$ admits
a left Kan extension
$\NNNN(\Delta_+)\to \CAlg(\DM^\otimes(k))$.
Namely,
\[
\mathcal{G}(X,x):\NNNN(\Delta_+)^{op}\to \CAlg(\DM^\otimes(k))^{op}
\]
is the Cech nerve associated to
$\NNNN(\{[-1]\to [0]\})^{op}\to\CAlg(\DM^{\otimes}(k))^{op}$ (cf. \cite[6.1.2.11]{HTT}). Its evaluation of $\mathcal{G}(X,x)$ at $[1]$ is equivalent to
$\uni_k\otimes_{M_X}\uni_k$.
The restriction $\NNNN(\Delta)^{op}\to \CAlg(\DM^\otimes(k))^{op}$
is a group object of $\CAlg(\DM^\otimes(k))^{op}$ (i.e.,
a cogroup object in $\CAlg(\DM^\otimes(k))$).
Namely, it determines
a group structure on $\uni_k\otimes_{M_X}\uni_k$
in $\CAlg(\DM^\otimes(k))^{op}$.
We refer to e.g. \cite[7.2.2.1]{HTT} for the notion of group objects.
Next we define an iterated generalization of $\mathcal{G}(X,x)$.
Consider the restriction $\NNNN(\{[1]\to [0]\})^{op} \subset \NNNN(\Delta_+)^{op} \to \CAlg(\DM^\otimes(k))^{op}$
of the above Cech nerve
$\mathcal{G}^{(1)}(X,x):=\mathcal{G}(X,x):\NNNN(\Delta_+)^{op}\to \CAlg(\DM^\otimes(k))^{op}$.
There is a unique isomorphism
$\NNNN(\Delta_+^{\le0})\simeq \NNNN(\{[1]\to [0]\})$.
Consider the composite $\NNNN(\Delta_+^{\le0})^{op} \simeq \NNNN(\{[1]\to [0]\})^{op} \subset \NNNN(\Delta_+)^{op} \to \CAlg(\DM^\otimes(k))^{op}$.
Once again,
take a rigth Kan extension
$\mathcal{G}^{(2)}(X,x):\NNNN(\Delta_+)^{op}\to \CAlg(\DM^\otimes(k))^{op}$
of this composite.
Repeating this process $n$ times, we obtain
\[
\mathcal{G}^{(n+1)}(X,x):\NNNN(\Delta_+)^{op}\to \CAlg(\DM^\otimes(k))^{op}.
\]
By abuse of notation, we write $\mathcal{G}^{(n+1)}(X,x)$ for the
group object defined as the restriction $\NNNN(\Delta)^{op}\subset \NNNN(\Delta_+)^{op}\to \CAlg(\DM^\otimes(k))^{op}$. (Moreover, one can endow $\mathcal{G}^{(n+1)}(X,x)$ with a structure of an $E_{n+1}$-monoid, but we will not use this enhanced structure.)
\section{Realized motivic rational homotopy type}
\label{realization}
We will consider the realizations of algebraic objects
that appears in Section~\ref{CMAsection} such as $M_X$.
The coefficient field $K$ is
a field of characteristic zero.
\subsection{}
There are several mixed Weil cohomology theories: singular (Betti)
cohomology, \'etale cohomology, analytic de Rham cohomology,
algebraic de Rham cohomology, rigid cohomology, etc
(see \cite[17.2]{CDT} for mixed Weil cohomology).
To a mixed Weil cohomology theory $E$ with coefficient field $K$,
one can associate a symmetric monoidal colimit-preserving functor
\[
\mathsf{R}_E:\DM^\otimes(k)\longrightarrow \mathsf{D}^\otimes (K)
\]
(see \cite[Section 5]{Tan} for details of the construction in the $\infty$-categorical setting) which is called the {\it realization functor} associated to $E$.
Here $\mathsf{D}^\otimes(K)$
is the derived $\infty$-category of $K$-vector
spaces (see Section~\ref{convention}).
By the relative adjoint functor theorem \cite[7.3.2.6, 7.3.2.13]{HA},
the realization functor $\mathsf{R}_E$ induces an adjunction
\[
\CAlg(\mathsf{R}_E):\CAlg(\DM^\otimes (k))\rightleftarrows \CAlg(\mathsf{D}^\otimes (K))\simeq \CAlg_{K}:\mathsf{M}_E
\]
where $\CAlg(\mathsf{R}_E)$ is the functor induced by $\mathsf{R}_E$,
and $\mathsf{M}_E$ is a right adjoint.
We shall refer to $\CAlg(\mathsf{R}_E):\CAlg(\DM^\otimes (k))\to \CAlg_{K}$
as the {\it multiplicative realization functor}.
In this Section, we consider
the realization functor associated to singular cohomology theory:
\[
\mathsf{R}:\DM^\otimes(k) \longrightarrow \mathsf{D}^\otimes(\QQ).
\]
We here suppose that the base field $k$ is embedded into the complex number
field $\CC$, and the coefficient field $K$ is $\QQ$.
This functor sends the object $M(X)$ to a complex $\mathsf{R}(M(X))$
that is quasi-isomorphic to the singular chain complex $C_*(X^t,\QQ)$
with rational coefficients. Here $X^t$ stands for the underlying
topological space
of the complex manifold $X\times_{\Spec k}\Spec \CC$.
For ease of notation, when no confusion is likely to arise,
we often write $\mathsf{R}$ for the
multiplicative realization functor $\CAlg(\mathsf{R}):\CAlg(\DM^\otimes(k)) \rightarrow \CAlg_{\QQ}$.
\subsection{}
There are several algebraic models that describe rational homotopy types
of topological spaces.
Quillen \cite{Q} uses {\it differential graded (dg) Lie algebras} whereas
Sullivan \cite{S} adopts {\it commutative differential graded (dg) algebras} as models.
Two approaches are related via Koszul duality between dg Lie
algebras and (augmented) commutative dg algebras.
In this paper, we use cochain
algebras of {\it polynomial differential forms} introduced by Sullivan
as algebraic models of the rational homotopy types of topological spaces.
Let us recall the definition of a cochain algebra of
polynomial differential forms on a topological space $S$,
see \cite[Section 10]{FHT} for the comprehensive reference.
For a simplicial set $P$, we let $A_{PL}(P)$ be the commutative differential
graded (dg) algebra with rational
coefficients of polynomial
differential forms. This commutative dg algebra is defined as follows
(but we will not apparently need this explicit definition).
For each $n\ge 0$,
we let $\Omega_n$ be the commutative
dg algebra of ``polynomial differential forms on the standard $n$-simplex'',
that is,
\[
\Omega_n:=\QQ[u_0,\ldots, u_n,du_0,\ldots, du_n]/(\Sigma_{i=0}^{n}u_i-1, \Sigma_{i=0}^n du_i)
\]
where $\QQ[u_0,\ldots u_n,du_0,\ldots, du_n]$
is the free commutative graded algebra generated by $u_0,\ldots, u_n$
and $du_0,\ldots, du_n$ with cohomological degrees $|u_i|=0$, $|du_i|=1$ for each $i$,
and the differential carries $u_i$ and $du_i$ to $du_i$ and $0$, respectively.
For any map $f:\Delta^n\to \Delta^m$, the pullback morphism
$f^*:\Omega_m\to \Omega_n$ of commutative dg algebras is defined
in a natural way (see e.g. \cite[Section 10 (c)]{FHT}).
An element of $A_{PL}(P)$ of (cohomological) degree $r$
is data that consists of a collection $\{ w_{\alpha}\}$
indexed by the set of all morphisms $\alpha:\Delta^n\to P$ from
standard simplices
such that
\begin{itemize}
\item each $\omega_{\alpha}$ is an element of $\Omega_n$ of degree $r$,
\item $f^*(w_{\beta})=w_{\alpha}$ for any $\alpha:\Delta^n\to P$, $\beta:\Delta^m\to P$,
and $f:\Delta^n\to \Delta^m$ such that $\beta \circ f=\alpha$.
\end{itemize}
The multiplication is given by $\{ w_{\alpha}\}\cdot\{ w'_{\alpha}\}=\{ w_{\alpha}w'_{\alpha}\}$, and the differential is given by $d\{ w_{\alpha}\}=\{ dw_{\alpha}\}$.
If $\phi:P\to P'$ is a map of simplicial sets and $\{\omega_{\alpha}\}_{\alpha:\Delta^n\to P'}$ is an element of $A_{PL}(P')$, then
$\phi^*\{\omega_\alpha\}$ is defined to be $\{\omega_{\phi\circ \beta}\}_{\beta:\Delta^n\to P}$. It gives rise to a map $\phi^*:A_{PL}(P')\to A_{PL}(P)$
of commutative dg algebras.
Let $T$ be a topological space.
If we write $S_*(T)$ for the singuar simplicial complex
whose $n$-th term is the set of singular $n$-simplices,
the commutative dg algebra $A_{PL}(T)$ is defined to be
$A_{PL}(S_*(T))$.
The assignment $P\mapsto A_{PL}(P)$ gives rise to
a functor $A_{PL}:\sSet \to (\CAlg_{\QQ}^{dg})^{op}$ to the category $\CAlg_{\QQ}^{dg}$
of commutative dg algebras over $\QQ$.
There exists a canonical equivalence
between the $\infty$-category $\SSS$ of spaces
and the $\infty$-category obtained
from $\sSet$ by inverting weak homotopy equivalences (cf. \cite[1.3.4.21]{HA}).
As observed below,
the functor $A_{PL}$
sends a weak homotopy equivalence in $\sSet$
to a quasi-isomorphism in $\CAlg_{\QQ}^{dg}$.
Therefore, $A_{PL}:\sSet \to (\CAlg_{\QQ}^{dg})^{op}$ induces
\[
A_{PL,\infty}:\SSS \longrightarrow \NNNN(\CAlg_{\QQ}^{dg})[W^{-1}]^{op}\simeq \CAlg_{\QQ}^{op}.
\]
For a topological space $T$, we shall denote by $A_{PL,\infty}(T)$
the image of $A_{PL}(T)$ in $\CAlg_{\QQ}$.
First we will describe the induced functor $A_{PL,\infty}:\SSS\to \CAlg_{\QQ}^{op}$ in an intrinsic way.
\begin{Proposition}
\label{PLproperty}
The followings hold:
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item The functor $A_{PL}:\sSet \to \CAlg_{\QQ}^{dg}$
sends a weak homotopy equivalence in $\sSet$
to a quasi-isomorphism in $\CAlg_{\QQ}^{dg}$,
\item $A_{PL,\infty}(\Delta^0)\simeq \QQ$,
\item $A_{PL,\infty}:\SSS\to \CAlg_{\QQ}^{op}$ preserves small colimits.
\end{enumerate}
\end{Proposition}
\begin{Remark}
\label{PLremark}
The functor $A_{PL,\infty}$ is uniquely determined by the properties
(2) and (3) in Proposition~\ref{PLproperty}.
Let $\Fun^{\textup{L}}(\SSS,\CAlg_{\QQ}^{op})$ be the full subcategory
of $\Fun(\SSS,\CAlg_{\QQ}^{op})$ spanned by those functors that
preserve small colimits.
Then by left Kan extension \cite[5.1.5.6]{HTT},
the map $p:\Delta^0\to \SSS$ with value $\Delta^0$ (i.e. the
contractible space) induces an equivalence
\[
\Fun^{\textup{L}}(\SSS,\CAlg_{\QQ}^{op}) \stackrel{\sim}{\to} \Fun(\Delta^0,\CAlg_{\QQ}^{op})\simeq \CAlg_{\QQ}^{op}.
\]
Therefore, the colimit-preserving functor $A_{PL,\infty}$ is uniquely
determined by the value $\QQ$ of the contractible space.
Namely, if $u:\Delta^0\to \CAlg_{\QQ}^{op}$ denotes the map
determined by the object $\QQ$ of $\CAlg_{\QQ}$, then
$A_{PL,\infty}:\SSS\to \CAlg_{\QQ}^{op}$
is a left Kan extension of $u:\Delta^0\to \CAlg_{\QQ}^{op}$
along $p:\Delta^0 \to \SSS$.
\end{Remark}
\Proof
We first prove (1).
Let $\CAlg_{\QQ}^{dg} \to \Comp(\QQ)$ be the forgetful functor to the
category $\Comp(\QQ)$ of complexes of $\QQ$-vector spaces.
It is enough to show that the
composite $\sSet\to (\CAlg_{\QQ}^{dg})^{op}\to \Comp(\QQ)^{op}$
preserves quasi-isomorphisms.
According to \cite[Theorem 10.9]{FHT}, there is the zig-zag of quasi-isomorphisms in $\Comp(\QQ)$
\[
C^*(P) \to B(P) \leftarrow A_{PL}(P)
\]
where $C^*(P)$ is the cochain complex associated to a simplicial set $P$ with
rational coefficients, and
$B(P)$ is a certain ``intermediate'' cochain complex
associated to $P$.
These quasi-isomorphisms are functorial in the sense that
for any map $P\to P'$ of simplicial sets,
they commute with $A_{PL}(P')\to A_{PL}(P)$,
$B(P')\to B(P)$, and $C^*(P')\to C^*(P)$.
Thus, it will suffice to observe that
$C^*:\sSet\to \Comp(\QQ)^{op}$ given by $P\mapsto C^*(P)$
sends weak homotopy equivalences to quasi-isomorphisms.
Let $C_*:\textup{Set}_{\Delta}\to \Comp(\QQ)$
be the functor which carries $P$ to the (normalized) chain complex
$C_*(P)$ with rational coefficients.
Since the dual of any quasi-isomorphism $C_*(P)\to C_*(P')$
is a quasi-isomorphism $C^*(P')\to C^*(P)$,
we are reduced to proving that
$C_*$
sends weak homotopy equivalences to quasi-isomorphisms.
Indeed, it is a well-known fact, but we here describe one of the proofs.
Let $\textup{Vect}_{\Delta}$
denote the category of simplicial objects in the category of $\QQ$-vector
spaces, that is, simplicial $\QQ$-vector spaces.
Consider the adjunction
$\QQ[-]:\sSet\rightleftarrows \textup{Vect}_{\Delta}:U$
where $U$ is the forgetful functor, and $\QQ[-]$ is its left adjoint, that is,
the free functor.
Let us consider $\sSet$ as the Quillen model category whose weak equivalences
are weak homotopy equivalence, and whose cofibrations are monomorphisms.
As in the case of simplicial abelian groups, $\textup{Vect}_{\Delta}$
admits a model category structure in which
$f$ is a weak equivalences (resp. a fibration) if $U(f)$ is a weak equivalence
(resp. a Kan fibration).
Then the pair $(\QQ[-],U)$ is a Quillen adjunction.
Let $N:\textup{Vect}_{\Delta}\stackrel{\sim}{\to} \Comp^{\le 0}(\QQ)$ be
the Dold-Kan equivalence
which carries a simplicial vector space to its normalized chain complex,
where $\Comp^{\le 0}(\QQ)$ is the full subcategory
of $\Comp(\QQ)$ spanned by those object $C$ such that $H^i(C)=0$ for $i>0$.
The composite $\sSet\stackrel{\QQ[-]}{\to} \textup{Vect}_{\Delta}\stackrel{N}{\to} \Comp^{\le 0}(\QQ)$
is naturally equivalent to the functor $\sSet\to \Comp^{\le 0}(\QQ)$
which sends $P$ to $C_*(P)$.
The functor $\QQ[-]$ preserves weak equivalences
since every object in $\sSet$ is cofibrant, and $N$ sends weak equivalences
to quasi-isomorphims. Thus, $C_*$ sends weak homotopy equivalences to quasi-isomorphisms.
The equality $A_{PL}(\Delta^0)=\QQ$ is clear from the definition (see \cite[Example 1 in
page 124]{FHT}). Hence (2) follows.
Next we prove (3). Note that the
forgetful functor $\CAlg_{\QQ}\to \Mod_{\QQ}\simeq \mathsf{D}(\QQ)$
preserves limits (cf. \cite[3.2.2.4]{HA}).
Thus, it will
suffice to prove that
$\SSS\stackrel{A_{PL,\infty}}{\to} \CAlg_{\QQ}^{op}\to \Mod_\QQ^{op}$
preserves small coimits; a small colimit diagram in $\SSS$ maps
to a limit diagram in $\Mod_\QQ$.
According to \cite[4.4.2.7]{HTT}, $\SSS\to \Mod_\QQ^{op}$
preserves small colimits if and only if it preserves pushouts
and small coproducts.
It is enough to
show that
$\sSet\stackrel{A_{PL}}{\to} (\CAlg_{\QQ}^{dg})^{op} \to \Comp(\QQ)^{op}$
sends homotopy pushout diagrams and homotopy coproduct diagrams
to homotopy pullback diagrams and homotopy product diagrams in $\Comp(\QQ)$,
respectively.
Here
the second functor is the forgetful functor, and $\Comp(\QQ)$
is endowed with the projective model structure, see Section~\ref{convention}.
As discussed in the proof of (1), we may replace this composite
by $C^*:\sSet\to \Comp(\QQ)^{op}$.
We will observe that $C_*:\sSet \to \Comp(\QQ)$ preserves
homotopy colimits.
We equip $\Comp^{\le0}(\QQ)$
with the projective model
structures (cf. \cite[2.3]{Hov}, \cite[4.1]{SS}).
A morphism $p$ in $\Comp^{\le0}(\QQ)$
is a weak equivalence (resp. a fibration)
if it is a quasi-isomorphism
(resp. surjective in cohomologically
negative degrees).
Cofibrations are monomorphisms
(keep in mind that $\QQ$ is a field).
The free functor $\QQ[-]:\sSet\to \Vect_{\Delta}$
is a left Quillen functor.
The normalization functor
$N:\textup{Vect}_{\Delta}\to \Comp^{\le0}(\QQ)$
is a left Quillen functor (see \cite[Section 4]{SS}).
In addition,
$\Comp^{\le0}(\QQ)\hookrightarrow \Comp(\QQ)$ is a left Quillen functor.
Therefore, we deduce that
$C_*:\sSet \to \Comp(\QQ)$ preserves
homotopy colimits.
Note that $C^*$ is composite $\sSet\stackrel{C_*}{\to} \Comp(\QQ)\to \Comp(\QQ)^{op}$ where the second functor
is given by the hom complex $\mathcal{H}om_{\QQ}(-,\QQ)$.
Then $\mathcal{H}om_{\QQ}(-,\QQ):\Comp(\QQ)\to \Comp(\QQ)^{op}$
preserves homotopy colimits, so that the induces
functor $\Mod_{\QQ}\to \Mod_{\QQ}^{op}$ preserves colimits.
(Indeed, it is enough to check that
it preserves homotopy pushouts and homotopy coproducts.
In $\Comp(\QQ)$, every object is both cofibrant and fibrant.
By the explicit presentation of homotopy pushouts/coproducts cf. \cite[A.2.4.4]{HTT},
we easily see that
$\mathcal{H}om_{\QQ}(-,\QQ)$ sends a homotopy pushout (resp. coproduct)
diagram
to a homotopy pullback (resp. coproduct) diagram.)
Consequently, $\SSS\to \Mod_{\QQ}^{op}$ induced by
$C^*$ preserves small colimits.
\QED
\subsection{}
Let us consider the composite
\[
T:\Sm_k^{op}\stackrel{\Xi}{\to} \CAlg(\DM^\otimes(k))\stackrel{\mathsf{R}}{\to} \CAlg_{\QQ}
\]
See Proposition~\ref{cohmotalg} for $\Xi$. We put $T_X=T(X)=\mathsf{R}(M_X)$.
\begin{Theorem}
\label{real1}
Let $X$ be a smooth scheme separated of finite type over $k\subset \CC$.
Let $X^t$ be the underlying topological space of the complex manifold $X\times_{\Spec k}\Spec \CC$.
There is a canonical equivalence $\mathsf{R}(M_X)=T_X\stackrel{\sim}{\to} A_{PL,\infty}(X^{t})$
in $\CAlg_{\QQ}$.
\end{Theorem}
\Proof
We first introduce some categories.
Let $\mathsf{D}^{\otimes}(X^{t})$ be the symmetric monoidal
presentable $\infty$-category of complexes of sheaves on $\QQ$-vector
spaces on $X^{t}$.
We define this $\infty$-category by the machinery of
model categories as in Example~\ref{topcomplex}.
Let $\Sh(X^t)$ be the Grothendieck abelian category of sheaves of $\QQ$-vector spaces on $X^t$ and let $\Comp(\Sh(X^t))$ be the category of cochain complexes
of $\Sh(X^t)$. It is endowed with the
symmetric monoidal structure by tensor product.
Thanks to \cite[Theorem 2.5, Example 2.3, Proposition 3.2]{CD1}, there is a symmetric monoidal
combinatorial model category structure on $\Comp(\Sh(X^t))$
in which weak equivalences consists of quasi-isomorphisms.
We then obtain the symmetric monoidal
presentable $\infty$-category $\mathsf{D}^{\otimes}(X^{t})$ from $\Comp(\Sh(X^t))^c$ by inverting weak equivalences (cf. Section~\ref{convention}).
By replacing $X^t$ by the one-point space $*$,
we also have a symmetric monoidal combinatorial model category
$\Comp(\Sh(*))$ which coincides with $\Comp(\QQ)$ endowed with
the projective
model structure.
By abuse of notation we denote the associated symmetric monoidal presentable $\infty$-category by $\mathsf{D}^\otimes(\QQ)$.
The canonical map to the one-point space $f^t:X^t\to *$ induces
the symmetric monoidal pullback functor
$\Comp(\Sh(*))\to \Comp(\Sh(X^t))$ that is a left Quillen functor
\cite[Theorem 2.14]{CD1}.
It gives rise to a symmetric monoidal colimit-preserving pullback
functor
$f^{t*}:\mathsf{D}(\QQ)\to \mathsf{D}(X^t)$.
According to relative adjoint functor theorem \cite[7.3.2.6]{HA},
there is a right adjoint functor $f_*^t:\mathsf{D}(X^t)\to \mathsf{D}(\QQ)$
which is lax symmetric monoidal.
We then use the Beilinson motives studied by Cisinski-D\'eglise \cite{CDT}.
Let $\mathbb{M}_{B}(X)$ be a symmetric monoidal combinatorial
model category of Beilinson motives over $X$ with rational
coefficients (see \cite[14.2]{CDT})
and let $\DM_B^\otimes(X)$ be the symmetric monoidal presentable
$\infty$-category obtained from $\mathbb{M}_{B}(X)$.
Since $X$ is regular,
according to \cite[16.1.1, 16.1.4]{CDT}
there is a symmetric monoidal equivalence
$\DM_B^\otimes(X)\stackrel{\sim}{\to} \DM^\otimes(X)$
induced by a symmetric monoidal left Quillen functor
$\mathbb{M}_{B}(X)\to Sp_{\textup{Tate}}(X)$
(by \cite[16.1.4]{CDT} the induced functor between their homotopy categories
is an equivalence, from which the equivalence of
stable $\infty$-categories follows, see e.g. \cite[Lemma 5.8]{Tan}).
The equivalences $\DM_B^\otimes(X)\simeq \DM^\otimes(X)$
and $\DM_B^\otimes(\Spec k)\simeq \DM^\otimes(k)$
commute with pullback functors.
Let $\mathsf{R}_X:\DM^\otimes(X)\simeq
\DM_B^\otimes(X)\to \mathsf{D}^\otimes(X^t)$
be the (relative)
realization functor that is a symmetric monoidal functor.
It is obtained from symmetric monoidal functors
of model categories (cf. Section~\ref{convention}):
as explained in \cite[17.1.7]{CDT} that uses the construction of Ayoub,
there is a diagram of symmetric monoidal functors
$\mathbb{M}_{B}(X)\stackrel{r}{\to} \mathbb{M}(X)\stackrel{p}{\leftarrow} \Comp(\Sh(X^t))$
of model categories where $\mathbb{M}(X)$ is an intermediate
symmetric monoidal model category, $r$ is a symmetric monoidal left Quillen
functor, and $p$ induces an equivalence of symmetric monoidal $\infty$-categories.
Similarly, we have the realization functor $\mathsf{R}:\DM^\otimes(k)\simeq \DM_B^\otimes(\Spec k)\to \mathsf{D}^\otimes(\QQ)$ of singular cohomology theory.
The functors $\mathsf{R}$ and $\mathsf{R}_X$ commute with the pullback
functors (because of the construction).
Therefore, we have the diagram
\[
\xymatrix{
\DM^\otimes(X) \ar[r]^{\mathsf{R}_X} \ar@<5pt>[d]^{f_*} & \mathsf{D}^\otimes(X^{t}) \ar[d]_{f_*^{t}} \\
\DM^\otimes(k) \ar[r]_{\mathsf{R}} \ar[u]^{f^*} & \mathsf{D}^\otimes(\QQ) \ar@<-5pt>[u]_{f^{t*}}.
}
\]
with a canonical equivalence
$\mathsf{R}_X\circ f^*\simeq f^{t*}\circ \mathsf{R}$
of symmetric monoidal functors.
Let $A$ be a commutative algebra object in $\DM^\otimes(X)$,
that is, an object of $\CAlg(\DM^\otimes(X))$.
Consider the canonical exchange map $e:\mathsf{R}(f_*(A))\to f_*^{t}(\mathsf{R}_X(A))$ in $\mathsf{D}(\QQ)$.
This map is the composition of morphisms
\begin{eqnarray*}
\RRR(f_*(A)) &\to& f^{t}_*f^{t*}(\RRR(f_*(A)) \\
&\simeq& f^{t}_*\RRR_Xf^*(f_*(A)) \\
&\simeq& f^{t}_*\RRR_X(f^*f_*)(A) \\
&\to& f^{t}_*\RRR_X(A)
\end{eqnarray*}
where the first arrow is induced by the unit map $\textup{id}\to f^{t}_*f^{t*}$, the second arrow is induced by $\RRR_Xf^*\simeq f^{t*}\RRR$, and the fourth one is induced by the counit map $f^*f_*\to \textup{id}$.
The unit map $\textup{id}\to f^{t}_*f^{t*}$
and the counit map $f^*f_*\to \textup{id}$ are promoted to
a unit map and a counit map for
adjunctions
$\CAlg(\mathsf{D}^\otimes(\QQ))\rightleftarrows \CAlg(\mathsf{D}^\otimes(X^t))$ and $\CAlg(\DM^\otimes(k))\rightleftarrows \CAlg(\DM^\otimes(X))$, respectively.
In particular, $e:\mathsf{R}(f_*(A))\to f_*^{t}(\mathsf{R}_X(A))$ is
promoted to
an morphism in $\CAlg(\mathsf{D}^\otimes(\QQ))\simeq \CAlg_{\QQ}$.
By \cite[17.2.18, 4.4.25]{CDT}, if $A$ is compact in the underlying $\infty$-category
$\DM(X)$, $e$ is an equivalence. In particular, if $A=\uni_X$,
we have a canonocal equivalence
$\mathsf{R}(f_*(\uni_X))=\mathsf{R}(M_X)\stackrel{\sim}{\to} f_*^{t}(\mathsf{R}_X(\uni_X))$ in $\CAlg_{\QQ}$.
Consequently, to prove our assertion it suffices to prove that
$f_*^{t}(\uni_{X^t})$ is equivalent to $A_{PL,\infty}(X^{t})$
where $\uni_{X^t}$ is the unit of $\mathsf{D}^\otimes(X^{t})$,
i.e., the constant sheaf with value $\QQ$.
For this purpose, recall first that since $X\times_{\Spec k}\Spec \CC$ is a complex smooth scheme separated of finite type, the underlying topological space $X^t$ is
a hausdorff paracompact smooth manifold.
Therefore, according to \cite[Theorem 5.1]{BT}, it admits
a good cover $\mathcal{U}=\{U_\lambda\}_{\lambda\in I}$, that is,
an open cover $\mathcal{U}=\{U_\lambda\}_{\lambda\in I}$ such that
every non-empty finite intersection
$U_{\lambda_0}\cap\ldots \cap U_{\lambda_r}$ is contractible.
Take the augmented simplicial diagram of the Cech nerve $U_\bullet \to U_{-1}:=X^{t}$ associated to the cover.
The $n$-th term $U_n$ of $U_\bullet$ is the disjoint union of intersections of $n+1$ open sets in $\mathcal{U}$.
We denote by $j_{U_n}:U_n\to X^t=U_{-1}$ the canonical map.
If we think of $U_\bullet \to X^t$ as an augmented simplicial diagram
in $\mathcal{S}$, then by Dugger-Isaksen
\cite[Theorem 1.1]{DI}, it is a colimit diagram.
According to Proposition~\ref{PLproperty},
the functor $A_{PL,\infty}:\SSS\to \CAlg_{\QQ}^{op}$ commutes with small
colimits. Thus,
the canonical morphism
$ A_{PL,\infty}(X^t)\to \varprojlim_{[n]\in \Delta} A_{PL,\infty}(U_n)$
is an equivalence where $\varprojlim_{[n]\in \Delta} A_{PL,\infty}(U_n)$
is a limit of the cosimplicial diagram in $\CAlg_\QQ$.
Thus, it is enough to show that $\varprojlim_{[n]\in \Delta} A_{PL,\infty}(U_n)\simeq f_*^{t}(\uni_{X^t})$.
For $i\ge -1$, we let $\textup{Comp}(\Sh(U_i))$ be the category of
complexes of sheaves of $\QQ$-vector spaces on $U_i$.
As in the case of $\mathsf{D}(X^t)$,
by the model structure in \cite[2.3, 2.5]{CD1}
we have a symmetric monoidal presentable $\infty$-category
$\mathsf{D}^\otimes(U_n)$ from $\textup{Comp}(\Sh(U_n))$.
For each morphism $U_n\to U_m$,
a symmetric monoidal colimit-preserving functor
$\mathsf{D}^\otimes(U_m)\to \mathsf{D}^\otimes(U_n)$.
It gives rise to a cosimplicial diagram of symmetric monoidal
$\infty$-categories
which we denote simply by $\mathsf{D}^\otimes(U_\bullet)$.
It also has the natural
coaugmentation $\mathsf{D}^\otimes(X^{t})\to \mathsf{D}^\otimes(U_\bullet)$.
Let $\Gamma(U_n,-):\mathsf{D}(U_n)\to \mathsf{D}(\QQ)$ be
the (derived) global section functor,
that is a lax symmetric monoidal right adjoint functor
to the pullback functor $\mathsf{D}^\otimes(\QQ)\to \mathsf{D}^\otimes(U_i)$ of $U_i\to *$.
We denote by $\uni_{U_n}$ the unit of
$\mathsf{D}(U_n)$ that corresponds to the constant sheaf with value $\QQ$.
Note $f^t_*(-)=\Gamma(U_{-1},-)=\Gamma(X^t,-)$, and
$\Gamma(U_{n},\uni_{U_n})$ in $\mathsf{D}(\QQ)$ is a complex computing the sheaf cohomology of $U_i$ with coefficients in $\QQ$.
Remember that $U_n$ is a disjoint union of contractible
spaces for $n\ge0$.
For each connected component $V$ of $U_n$,
$\Gamma(V,\uni_{U_n}|_{V})$ in $\CAlg_{\QQ}$ is
an initial object of $\CAlg_{\QQ}$, i.e., $\QQ$
since the unit map $\QQ\to \Gamma(V,\uni_{U_n}|_{V})$
is an equivalence in $\mathsf{D}(\QQ)$, cf. \cite[Corollary 3.2.1.9]{HA}.
By Proposition~\ref{PLproperty},
the image of a contractible space under $A_{PL,\infty}$ is $\QQ$.
Therefore, $\Gamma(U_n,\uni_{U_n})\in \CAlg_{\QQ}$ is
equivalent to $A_{PL,\infty}(U_n)$, i.e., $\Gamma(U_i,\uni_{U_i})\simeq \prod_{\pi_0(U_n)}\QQ \simeq A_{PL,\infty}(U_n)$
for $i\ge0$ ($\pi_0(-)$ is the set of connected components). We may consider $\{\Gamma(U_i,\uni_{U_i})\}_{[i]\in \Delta}$
to be a cosimplicial diagram of ordinary commutative algebras (arising
from connected components of $U_\bullet$).
We then have $\varprojlim A_{PL,\infty}(U_n)\simeq \varprojlim\Gamma(U_n,\uni_{U_n})$.
It will suffice to prove that
the canonical morphism $\Gamma(X^{t},\uni_{X^{t}})\to \varprojlim_{[n]\in \Delta}\Gamma(U_n,\uni_{U_n})$
in $\mathsf{D}(\QQ)$ is an equivalence (we may and will disregard their commutative algebra structures).
To this end, we use the descent for hypercovers on $X^t$.
Let $j_{U_n!}:\mathsf{D}(U_n)\to \mathsf{D}(X^t)$ be the left adjoint to the restriction $j_{U_n}^*:\mathsf{D}(X^t)\to \mathsf{D}(U_n)$.
According to \cite[Example 2.3, Theorem 2.5]{CD1}, we see that the canonical
morphism $\varinjlim_{[n]\in \Delta^{op}}j_{U_n!}(\uni_{U_n})\to \uni_{X^t}$ is an equivalence in $\mathsf{D}(X^t)$.
For any $F$ in $\mathsf{D}(X^t)$, it induces an equivalence
$\Gamma(X^t,F)\stackrel{\sim}{\to} \varprojlim_{[n]\in \Delta}\Gamma(U_n,F|_{U_n})$.
In particular, we have a canonical equivalence
$\Gamma(X^{t},\uni_{X^{t}})\stackrel{\sim}{\to} \varprojlim_{[n]\in \Delta}\Gamma(U_i,\uni_{U_i})$.
\QED
\begin{Remark}
\label{real1R}
Let $\phi:Y\to X$ be a morphism in $\Sm_k$.
Then $\phi^*:M_X\to M_Y$ induces $\mathsf{R}(\phi^*):\mathsf{R}(M_X)=T_X\to \mathsf{R}(M_X)=T_Y$.
On the other hand, the associated
continuous map $\phi^t:Y^t\to X^t$ of topological
spaces induces $\phi^{t*}:A_{PL,\infty}(X^t)\to A_{PL,\infty}(Y^t)$ induced
by $A_{PL}(X^t)\to A_{PL}(Y^t)$.
The morphism $\mathsf{R}(\phi^*):T_X\to T_Y$ in $\CAlg_{\QQ}$
is equivalent to $\phi^{t*}:A_{PL,\infty}(X^t)\to A_{PL,\infty}(Y^t)$
through equivalences $T_X\simeq A_{PL,\infty}(X^t)$ and $T_Y\simeq A_{PL,\infty}(Y^t)$ in Theorem~\ref{real1}.
To observe this, note first that by the compatibility of the realization
functor with pushforward functors,
$\mathsf{R}(\phi^*)$ can be identified with
$f_*^t(\uni_{X^t})\to g_*^t(\uni_{Y^t})$ induced by $\phi^t:Y^t\to X^t$
where $g^t:Y^t\to *$ is the canonical map to one point space.
Let us unfold the equaivalence given in the proof of
Theorem~\ref{real1}.
As in the proof,
choose a good cover $\mathcal{U}=\{U_\lambda\}_{\lambda\in I}$ of $X^t$
and take the augmented Cech nerve $U_\bullet\to X^t=U_{-1}$.
We know from the proof of Theorem~\ref{real1} that
there are canonical equivalences
$\Gamma(U_n,\uni_{U_n})\stackrel{\sim}{\to} \prod_{\alpha\in \pi_0(U_n)}\Gamma(U_{n,\alpha},\uni_{U_{n,\alpha}})\stackrel{\sim}{\leftarrow}\prod_{\alpha\in \pi_0(U_n)}\QQ$ in $\CAlg_\QQ$
where
each $U_n$ is a disjoint union $\sqcup_{\alpha\in \pi_0(U_n)}U_{n,\alpha}$ of contractible spaces.
Similarly, we have canonical equivalences
$A_{PL,\infty}(U_n)\stackrel{\sim}{\to} \prod_{\alpha\in \pi_0(U_n)}A_{PL,\infty}(U_{n,\alpha})\stackrel{\sim}{\leftarrow}\prod_{\alpha\in \pi_0(U_n)}\QQ$.
Both cosimplicial
objects $\{\Gamma(U_n,\uni_{U_n})\}_{[n]\in \Delta}$
and $\{A_{PL}(U_n)\}_{[n]\in \Delta}$ are equivalent to
the cosimplicial ordinary
commutative $\QQ$-algebra, regarded as a
cosimplicial object in $\CAlg_\QQ$,
that is defined by the assignment
$[n] \mapsto \prod_{\alpha\in \pi_0(U_n)}\QQ=\QQ^{\pi_0(U_n)}$
such that for any $[n]\to [m]$, $\QQ^{\pi_0(U_n)}\to \QQ^{\pi_0(U_m)}$
is induced by the map $\pi_0(U_m)\to \pi_0(U_n)$ (by
the superscript we mean cotensor). It gives
rise to $A_{PL,\infty}(X^t) \simeq \varprojlim A_{PL,\infty}(U_n)\simeq \varprojlim \Gamma(U_n,\uni_{U_n})\simeq f_*^t(\uni_{X^t})$.
Taking account of these steps,
we are reduced to checking
a functoriality of good covers: it suffices to verify
that
if $\mathcal{U}=\{U_\lambda\}_{\lambda\in I}$
is a good cover of $X^t$,
then there is a good over $\mathcal{V}=\{V_{\mu}\}_{\mu \in J}$ of $Y^t$
such that any $V_\mu \to Y^t\to X^t$ factors through some $U_{\lambda}\to X^t$. Actually,
it follows from the proof of the existence of a good cover.
See \cite[Corollary 5.2]{BT} and
the discussion after the proof of \cite[Theorem 5.1]{BT}.
\end{Remark}
It is useful to have a smooth de Rham model of $T_X$.
We will describe $T_X\otimes_{\QQ}\RR$ in terms of smooth differential
forms.
By $X_{\infty}$ we mean the underlying differential manifold of
$X\times_{\Spec k}\Spec \CC$.
Let $A_{X_\infty}$ be the commutative dg algebra
of $C^\infty$ real differential forms on $X_{\infty}$.
We call $A_{X_\infty}$ the smooth de Rham algebra
on $X_\infty$
We think of $A_{X_\infty}$ as an object in $\CAlg_{\RR}$.
\begin{Corollary}
\label{real2}
Consider the base change $T_X\otimes_{\QQ}\RR$ which belongs to $\CAlg_{\RR}$.
There is an equivalence $T_X\otimes_{\QQ}\RR\simeq A_{X_\infty}$
in $\CAlg_{\RR}$.
\end{Corollary}
\Proof
There is a zig-zag of quasi-isomorphisms between
$A_{X_\infty}$ and $A_{PL}(X^t)\otimes_{\QQ}\RR$
(see \cite[Theorem 11.4]{FHT}). Thus, by Theorem~\ref{real1}
we see that $T_X\otimes_{\QQ}\RR\simeq A_{X_\infty}$.
\QED
By using Theorem~\ref{real1} and Remark~\ref{real1R},
we can easily prove the following:
\begin{Proposition}
Let $\CAlg(\DM^\otimes(k))\to \CAlg_{\QQ}$
be the multiplicative realization functor.
Then the image of the motivic algebra of path torsor $P(X,x,y)$ (cf. Example~\ref{motpath}) in $\CAlg_{\QQ}$ is equivalent to
the pushout
$\QQ\otimes_{A_{PL,\infty}(X^t)}\QQ$ associated to two augmentations
$A_{PL,\infty}(X^t)\to \QQ$ and $A_{PL,\infty}(X^t)\to \QQ$ respectively
induced by
points $x$ and $y$ in $X^t$.
(We remark that $\QQ\otimes_{A_{PL,\infty}(X)}\QQ$
can be obtained by a bar construction
of $A_{PL}(X)$ with two augmentations, see \cite{Ol}.)
The image of the motivic algebra of free loop space $HHM_X$ (cf. Example~\ref{motloop}) in $\CAlg_{\QQ}$ is
$A_{PL,\infty}(X^t)\otimes_{A_{PL,\infty}(X^t)\otimes A_{PL,\infty}(X^t)}A_{PL,\infty}(X^t)\simeq S^1\otimes A_{PL,\infty}(X^t)$.
(It might be worth mentioning that if $X^t$ is simply connected, then $A_{PL,\infty}(X^t)\otimes_{A_{PL,\infty}(X^t)\otimes A_{PL,\infty}(X^t)}A_{PL,\infty}(X^t)$
is equivalent to
$A_{PL,\infty}(LX^t)$ where $LX^t$ is the free loop space of $X^t$ \cite[Example 1 in page 206]{FHT}.)
\end{Proposition}
\subsection{}
\label{preDAG}
Before proceeding the next subsection, we introduce some algebro-geometric
notions. Let $K$ be a field of characteristic zero.
Let $\CAlg^{\textup{dis}}_K$ be the full subcategory of $\CAlg_K$ that is spanned by discrete objects $C$,
i.e., $H^i(C)=0$ for $i\neq 0$.
Put another way, we let $\Mod_K^{\DIS}$ be the
(symmetric monoidal) full subcategory of $\Mod_K\simeq \mathsf{D}(K)$
spanned by discrete objects $M$,
i.e., $H^i(M)=0$ for $i\neq 0$.
This full subcategory is nothing else but (the nerve of) the category
of $K$-vector spaces.
Then $\CAlg_K^{\textup{dis}}=\CAlg(\Mod^{\DIS}_K)$.
The $\infty$-category
$\CAlg^{\textup{dis}}_K$ is naturally equivalent
to the nerve of category of ordinary commutative $K$-algebras.
Let $\Aff_K$ be the opposite category of $\CAlg_K$.
We write $\Spec R$ for an object in $\Aff_K$
that corresponds to $R\in \CAlg_K$.
We shall refer to it as a derived affine scheme (or affine scheme) over $K$.
The Yoneda embedding identifies $\Aff_K$ with the full subcategory of
$\Fun(\CAlg_K,\SSS)$. This embedding preserves
small limits.
The functor $\Spec R:\CAlg_K \to \SSS$
corepresented by $R$ satisfies the sheaf condition with respect to flat
topology, see e.g. \cite{DAG}.
We often regard $\Spec R$ as a sheaf $\CAlg_K\to \SSS$.
We remark that in the literature of derived geometry (see e.g. \cite{DAG}
for its $\infty$-categorical theory),
$\Spec R$ with $R\in \CAlg_K$ is usually called a nonconnective
(derived) affine scheme.
Let $\Aff^{\textup{dis}}_K$
be the full subcategory of $\Aff_K$ that
corresponds to $\CAlg^{\textup{dis}}_K$.
One can naturally identify $\Aff^{\textup{dis}}_K$
with the category of ordinary affine schemes over $K$
(keep in mind that the full subcategories
$\Aff_K^{\DIS}$ are not closed under some constructions;
for example, in general,
fiber products in $\Aff_K^{\DIS}$
are not compatible
with those in $\Aff_K$).
For an $\infty$-category $\mathcal{C}$ that has
finite products, we write $\Grp(\mathcal{C})$ for the $\infty$-category
of group objects in $\mathcal{C}$.
We shall
call a group object in $\Aff_K$ a derived affine group scheme over $K$.
There is a canonical Yoneda embedding $\Grp(\Aff_K)\hookrightarrow \Fun(\CAlg_K,\Grp(\SSS))$. Therefore, through this functor we often
think of a derived affine group scheme
as a sheaf $\CAlg_K\to \Grp(\SSS)$.
Put another way, $\Spec R$ in $\Grp(\Aff_K)$ amounts to
a commutative Hopf algebra object $R$ in $\Mod_K^\otimes$.
See \cite[Appendix A]{Tan} for details.
\subsection{}
\label{realhopfsection}
\begin{Definition}
\label{realhopfdef}
In Section~\ref{hopfsection},
for a pointed smooth variety $(X,x)$ and a natural number $n\ge1$,
we have defined the group
object
$\mathcal{G}^{(n)}(X,x):\NNNN(\Delta^{op})\to \CAlg(\DM^\otimes(k))^{op}$.
Since the multiplicative realization functor
$\CAlg(\mathsf{R}_E):\CAlg(\DM^\otimes(k))\to \CAlg_{K}$
preserves
coproducts,
we see that
the composite
\[
G_E^{(n)}(X,x):\NNNN(\Delta^{op})\stackrel{\mathcal{G}^{(n)}(X,x)^{op}}{\longrightarrow} \CAlg(\DM^\otimes(k))^{op} \stackrel{\CAlg(\mathsf{R}_E)^{op}}{\longrightarrow} \CAlg_{K}^{op}=\Aff_{K}
\]
is a group object in $\Aff_{K}$.
Namely, $G_E^{(n)}(X,x)$ is a derived affine group scheme over $K$.
If no confusion is likely to arise,
we often write $G^{(n)}(X,x)$ for $G_E^{(n)}(X,x)$.
\end{Definition}
\begin{Proposition}
\label{cechtop}
Suppose that $k$ is embedded in $\CC$ and
consider the case of singular realization $\mathsf{R}=\mathsf{R}_E$.
The point $x$ on $X$ determines a point of
the associated topological space $X^t$
which we denote also by $x$.
Let $\Spec \QQ\to \Spec T_X$ be a morphism induced by $x$.
Then the derived affine group scheme $G(X,x)=G^{(1)}(X,x)$
is equivalent to the Cech nerve obtained from $\Spec \QQ\to \Spec T_X$.
The iterated group scheme $G^{(n)}(X,x)$ ($n\ge2$) also has
a similar description.
\end{Proposition}
\Proof
By Remark~\ref{real1R},
the map $T_X=\mathsf{R}(M_X)\to \QQ=\mathsf{R}(M_{\Spec k})$
induced by $M_X\to M_{\Spec k}=\uni_k$
can be viewed as the map $T_X\to \QQ$ induced by $x\in X^t$.
Remember that
the opposite of the multiplicative realization functor
$\CAlg(\DM^\otimes(k))^{op}\to \CAlg^{op}_{\QQ}=\Aff_{\QQ}$
preserves small limits.
Therefore, the derived affine group scheme $G(X,x)$ is the Cech nerve
of $\Spec \QQ\to \Spec T_X$ in $\Aff_{\QQ}$.
The second claim is clear from this argument.
\QED
\section{Motivic Galois action}
\label{MGA}
Let $K$ be a field of characteristic zero.
Let $\mathsf{R}_E:\DM^\otimes(k)\to \mathsf{D}^\otimes(K)$ be a
realization functor associated to a mixed Weil cohomology theory $E$.
In \cite{Tan} (see also \cite{Bar}, \cite{DTD}), we constructed
a derived affine group scheme $\mathsf{MG}_E$
over $K$ out of $\mathsf{R}_E$,
which we refer to as the {\it derived motivic Galois group}
with respect to $E$.
It has many favorable properties such as the consistency of motivic conjectures.
The most important property of $\mathsf{MG}_E$ for us is that
it represents the automorphism group
of the symmetric monoidal functor $\mathsf{R}_E$,
see Definition~\ref{realauto} or \cite{Tan} for the formulation.
Besides,
we have the usual affine group scheme
$MG_E$ associated to $\MG_E$
which we call the {\it motivic Galois group} with respect to $E$.
Note that a symmetric monoidal natural equivalence from $\mathsf{R}_E$
to itself induces a natural equivalence from $\CAlg(\mathsf{R}_E):\CAlg(\DM^\otimes(k))\to \CAlg_K$ to itself.
Actually, there is a canonical morphism from the automorphism
group
of $\mathsf{R}_E$ to the automorphism group of $\CAlg(\mathsf{R}_E)$.
Since $M_X$ belongs to $\CAlg(\DM^\otimes(k))$ for a smooth variety $X$,
the automorphism group of $\CAlg(\mathsf{R}_E)$ acts on
the image of $M_X$, e.g. $A_{PL,\infty}(X^t)$ in $\CAlg_{\QQ}$.
Consequently, it gives
rise to an action of the derived affine group scheme $\mathsf{MG}_E$
on the image of $M_X$.
Based on this natural idea,
we will
construct motivic Galois actions,
i.e., actions of $\MG_E$
by using the machinery of $\infty$-categories
in Section~\ref{actionstep1}.
In Section~\ref{diagramaction}, we focus on the case of
a cosimplicial diagram in $\CAlg(\DM^\otimes(k))$. The motivating cases
come from Section~\ref{hopfsection} and Section~\ref{realhopfsection}:
it yields an action of $\MG_E$ on the derived
affine group schemes $G_E^{(n)}(X,x)$ in Definition~\ref{realhopfdef}.
In Sections~\ref{underlyinggroup} to \ref{finalaction},
we turn to study how to obtain an action of the pro-algebraic group
$MG_E$ on the pro-unipotent completions of homotopy groups
and related invariants arising from various cohomology theories.
Our approach is to deduce the actions of $MG_E$ from the actions of $\MG_E$ on $G_E^{(n)}(X,x)$.
For example, under the situation of Proposition~\ref{cechtop},
one can derive the pro-unipotent completion of the fundamental group of $X^t$ from $G^{(1)}(X,x)$: if $G^{(1)}(X,x)=\Spec A$, the pro-unipotent completion
is given by $\Spec H^0(A)$
($A\simeq \QQ\otimes_{A_{PL}(X^t)}\QQ$ in $\CAlg_{\QQ}$).
\subsection{}
\label{actionstep1}
Our first task is to construct motivic Galois actions
on the images of multipilicative realization functors such as $A_{PL}(X^t)$.
\subsubsection{}
\begin{Definition}
\label{autogroup}
Let $\mathcal{I}$ be an $\infty$-category and $D:\mathcal{I}\to \Cat$ a functor. Suppose that $\mathcal{I}$ has an initial object $\xi$.
Let $C$ be an object of $D(\xi)$.
Let $(-)^\simeq:\Cat \to \SSS$ be the functor which carries an $\infty$-category
$\mathcal{C}$ to its largest Kan subcomplex $\mathcal{C}^\simeq$.
Namely, it is the right adjoint to the inclusion $\SSS\to \Cat$.
Let $\mathcal{F}_D\to \mathcal{I}$ be a left fibration
obtained by applying the unstraightening functor or relative nerve functor \cite{HTT} to $\mathcal{I}\to \Cat\to \SSS$.
By \cite[3.3.3.4]{HTT}, a section $\mathcal{I}\to \mathcal{F}_D$
of $\mathcal{F}_D\to \mathcal{I}$ corresponds to an object
in the limit $\varprojlim_{i\in \mathcal{I}}D(i)^\simeq$ in $\SSS$.
We let $s:\mathcal{I}\to \mathcal{F}_D$ be the section
that corresponds to the image of $C$ under the canonoical functor
$D(\xi)^\simeq\to \varprojlim_{i\in \mathcal{I}}D(i)^\simeq$.
Through the correspondence between left fibrations over $\mathcal{I}$
and functors $\mathcal{I}\to \SSS$ (cf. \cite[3.2, 4.2.4.4]{HTT}),
$\mathcal{F}_D\to \mathcal{I}$ endowed with the section $s$
amounts to the functor $(-)^\simeq \circ D:\mathcal{I}\to \SSS$
with a natural transformation $\ast \to (-)^\simeq \circ D$
from the constant functor
$\ast:\mathcal{I}\to \SSS$ taking the value $\Delta^0$.
By the adjunction, the natural transtransformation is described as
a functor $D_*:\mathcal{I}\to \SSS_{*}:=\SSS_{\Delta^0/}\subset \Fun(\Delta^1,\SSS)$
such that the composition $\mathcal{I}\to \SSS_{*}\to \SSS$
with the forgetful functor is $(-)^\simeq \circ D$.
We shall refer to $D_*$ as the functor extended by $C$.
Let $\Grp(\SSS)$ denote the category of group objects in $\SSS$ (see e.g. \cite[7.2.2.1]{HTT}, \cite[Definition A.2]{Tan}).
Let $\Omega_*:\SSS_*\to \Grp(\SSS)$
be the functor which carries the based space $S$ to the based loop space
$\Omega_*(S)$.
We define the automorphism group functor of $C$ over $\mathcal{I}$
to be the composite
\[
\Aut_{\mathcal{I}}(C):\mathcal{I}\stackrel{D_*}{\longrightarrow}\SSS_*\stackrel{\Omega_*}{\longrightarrow} \Grp(\SSS).
\]
We usually write $\Aut(C)$ for $\Aut_{\mathcal{I}}(C)$.
\end{Definition}
\begin{Remark}
\label{autogroupR}
For any object $i$ in $\mathcal{I}$, the composition
$\mathcal{I}\stackrel{\Aut_{\mathcal{I}}(C)}{\to} \Grp(\SSS)\to \SSS$
with the forgetful functor sends $i$ to the $\infty$-groupoid (space)
that is equivalent to the mapping space $\Map_{D(i)}(f(C),f(C))$
where $f:\xi\to i$ is the canonical functor from the initial object.
Indeed, the composite $\mathcal{I} \to \SSS$
sends $i$ to the fiber product $\Delta^0\times_{D(i)^\simeq}\Delta^0$
in $\SSS$,
defined by the map $\Delta^0\to D(i)$ determined by $f(C)$.
The fiber product $\Delta^0\times_{D(i)^\simeq}\Delta^0$
is explicitly given by the fiber product $\{f(C)\}\times_{D(i)^\simeq}\Fun(\Delta^1,D(i)^\simeq)\times_{D(i)^\simeq}\{f(C)\}$ of (genuine) simplicial sets, that is a model of the mapping space (cf. \cite[1.2.2, 4.2.1.8]{HTT}).
\end{Remark}
\begin{Definition}
\label{symfuncauto}
Let $\CAlg_{(-)}:\CAlg_K\to \wCat$ be a functor which carries $A$ to
$\CAlg_A$ where $\CAlg_A$ is the $\infty$-category of commutative
ring spectra over $A$, that is, commutative algebra objects
in $\Mod_A^\otimes$ (a morphism $A\to A'$
maps to $\CAlg_A\to \CAlg_{A'}$ given by the base change $\otimes_{A}A'$,
see Section~\ref{pre2} for the formulation).
Let $C$ be an object of $\CAlg_{K}$.
We apply Definition~\ref{autogroup} to $\CAlg_{(-)}:\mathcal{I}=\CAlg_K\to \wCat$
and $C$ after replacing $\Cat$ and $\SSS$ by $\wCat$ and $\wSSS$, respectively.
We then define $\Aut_{\CAlg_K}(C):\CAlg_{K}\to \Grp(\wSSS)$
to be the automorphism group functor of $C$ over $\CAlg_K$.
Let $L$ be an $\infty$-category.
Let $(-)^L:\wCat\to \wCat$ be the functor which carries
$\mathcal{C}$ to $\Fun(L,\mathcal{C})$.
Namely, it is given by cotensoring with $L$.
Let $h:L\to \CAlg_K$ be a functor (which we will consider to be a diagram
in $\CAlg_K$ indexed by $L$).
Consider the composition
\[
\mu_L:\CAlg_K\stackrel{\CAlg_{(-)}}{\to} \wCat\stackrel{(-)^L}{\to} \wCat.
\]
Applying Definition~\ref{autogroup} to $\mu_L$
and $h$,
we define $\Aut_{\CAlg_K}(h):\CAlg_K\to \Grp(\wSSS)$ to be the automorphism group functor of $h$
over $\CAlg_{K}$.
Notice that $\Aut_{\CAlg_K}(C)$ is the spacial case of $\Aut_{\CAlg_K}(h)$.
We usually write $\Aut(C)$ and $\Aut(h)$
for $\Aut_{\CAlg_K}(C)$ and $\Aut_{\CAlg_K}(h)$, respectively.
\end{Definition}
\begin{Definition}
Let $\Mod_{(-)}:\CAlg_K\to \wCat$ be a functor which carries $A$ to
$\Mod_A$ (a morphism $A\to A'$
maps to $\Mod_A\to \Mod_{A'}$ given by the base change $\otimes_{A}A'$,
see Section~\ref{pre2} for the formulation).
Let $P$ be an object of $\mathsf{D}(K)\simeq \Mod_K$.
Applying Definition~\ref{autogroup} to $\Mod_{(-)}:\CAlg_K\to \wCat$
and $P$,
we define $\Aut(P)=\Aut_{\CAlg_K}(P):\CAlg_{K}\to \Grp(\wSSS)$
to be the automorphism group functor of $P$ over $\CAlg_K$.
\end{Definition}
Let $\mathsf{R}_E:\DM^\otimes(k)\to \mathsf{D}^\otimes(K)=\Mod_K^\otimes$
be the realization functor associate to a mixed Weil cohomology theory $E$
with coefficients in a field $K$ of characteristic zero.
The coefficient field of $\DM(k)$
will be $K$, but one can also
adopt the setting where the coefficient
field of $\DM(k)$ is $\QQ$ (one may choose either one
depending on the purpose).
Let $\mathsf{MG}_E=\Spec B$
be a derived affine group
scheme over $K$
which we call
the derived motivic Galois
group with respect to $E$ (see \cite{Tan}).
Here the fundamental property of $\mathsf{MG}_E$ for us
is that it represents
the automorphism group functor
$\Aut(\mathsf{R}_E):\CAlg_K\to \Grp(\wSSS)$
of the realization functor $\mathsf{R}_E$
(see Definition~\ref{realauto} for its definition).
Namely, if one regards $\mathsf{MG}_E$ as a functor
$\CAlg_K\to \Grp(\wSSS)$, then we have an equivalence
$\mathsf{MG}_E\simeq \Aut(\mathsf{R}_E)$.
\begin{Proposition}
\label{goodaction}
Let $C$ be an object of $\CAlg(\DM^\otimes(k))$.
There is a (canonical) action of $\mathsf{MG}_E$ on $\CAlg(\mathsf{R}_E)(C)$.
(Recall that $\CAlg(\mathsf{R}_E):\CAlg(\DM^\otimes(k))\to \CAlg_K$
is the multiplicative realization functor, Section~\ref{realization}.)
Namely, there is a morphism $\mathsf{MG}_E\to \Aut(\CAlg(\mathsf{R}_E)(C))$
in $\Fun(\CAlg_K,\Grp(\wSSS))$.
In particular, we have a (canonical) action of $\mathsf{MG}_E$ on $\CAlg(\mathsf{R}_E)(M_X)$. Moreover, the following properties hold:
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item The actions are functorial in $\CAlg(\DM^\otimes(k))$:
Namely, if we let $p:L\to \CAlg(\DM^\otimes(k))$ be a functor from an $\infty$-category $L$ and let $h:L\to \CAlg(\DM^\otimes(k))\stackrel{\CAlg(\mathsf{R}_E)}{\to} \CAlg_{K}$
be the composition with the multiplicative realization functor, then there is a morphism $\mathsf{MG}_E\to \Aut(h)$.
For a functor $g:M\to L$ of $\infty$-categories,
the action (morphism) $\mathsf{MG}_E\to \Aut(h\circ g)$
is naturally equivalent to $\mathsf{MG}_E\to \Aut(h)\to \Aut(h\circ g)$
where the the first arrow is given by the action on $h$, and the second arrow
is induced by the composition with $M\to L$.
\item The action is compatible with the formation of colimits:
Let $p:L\to \CAlg(\DM^\otimes(k))$ be a functor from a small $\infty$-category, and $\overline{p}:L^\triangleright \to \CAlg(\DM^\otimes(k))$ a colimit diagram of $p$ (here $(-)^\triangleright$ indicates the right cone
\cite{HTT}). Let $C$ be the colimit
in $\CAlg(\DM^\otimes(k))$, that is, the image of the cone point.
Let
$q:L\to \CAlg_K$ and $\overline{q}:L^\triangleright \to \CAlg_K$
be the composites $\CAlg(\mathsf{R}_E)\circ p$ and $\CAlg(\mathsf{R}_E)\circ \overline{p}$, respectively.
Then the (action) morphism $\mathsf{MG}_E\to \Aut(\CAlg(\mathsf{R}_E)(C))$ factors through
the morphism $\mathsf{MG}_{E}\to \Aut(q)$ in the sense that the
restriction to $L$ induces an equivalence $\Aut(\overline{q})\stackrel{\sim}{\to}\Aut(q)$, and the composite
\[
\mathsf{MG}_{E}\to \Aut(q)\simeq \Aut(\overline{q})\to \Aut(\CAlg(\mathsf{R}_E)(C))
\]
is naturally equivalent to the ``action''
$\mathsf{MG}_E\to \Aut(\CAlg(\mathsf{R}_E)(C))$. Here the final arrow is
induced by the restriction to the cone point of $L^\triangleright$.
\item There is a (canonical) action of $\mathsf{MG}_E$
on $\mathsf{R}_E(C)$, that is a morphism $\mathsf{MG}_E\to \Aut(\mathsf{R}_E(C))$. We here distinguish the underlying module
$\mathsf{R}_E(C)$ in $\mathsf{D}(K)$ from
$\CAlg(\mathsf{R}_E)(C)$ in $\CAlg_K$.
The action on $\CAlg(\mathsf{R}_E)(C)$ is compatible with
that on $\mathsf{R}_E(C)$ in the sense that there is a canonical morphism
$\Aut(\CAlg(\mathsf{R}_E)(C))\to \Aut(\mathsf{R}_E(C))$ induced by the forgetful functor, and $\mathsf{MG}_E\to \Aut(\mathsf{R}_E(C))$
is equivalent to the composite $\mathsf{MG}_E\to \Aut(\CAlg(\mathsf{R}_E)(C))\to \Aut(\mathsf{R}_E(C))$.
\end{enumerate}
\end{Proposition}
\begin{Corollary}
Suppose that $k$ is embedded in $\CC$. Let
$X^t$ be the underlying topological space of $X\times_{\Spec k}\Spec \CC$.
If $\mathsf{MG}$
denotes the derived motivic Galois group with respect to the singular cohomology
theory, there is a canonical
action of $\mathsf{MG}$ on $A_{PL,\infty}(X^t)\simeq T_X$.
\end{Corollary}
\Proof
Combine Proposition~\ref{goodaction} and Theorem~\ref{real1}.
\QED
\begin{Remark}
Let $A\in \CAlg_K$
and let $g:\Delta^0\to \mathsf{MG}_E(A)$ be an ``$A$-valued point''.
Through the equivalence $\mathsf{MG}_E(A)\simeq \Aut(\mathsf{R}_E)(A)$,
$g$ may be viewed as an automorphism of the composite $\DM^\otimes(k)\stackrel{\mathsf{R}_E}{\to} \Mod^\otimes_K\stackrel{\otimes_KA}{\to} \Mod^\otimes_A$.
It gives rise to an automorphism $u$ of the composite
\[
\CAlg(\DM^\otimes(k)) \stackrel{\CAlg(\mathsf{R}_E)}{\to} \CAlg_K\stackrel{\otimes_KA}{\to} \CAlg_A
\]
(see Section~\ref{prepre} below).
The image $\Delta^0\to\Aut(\CAlg(\mathsf{R}_E)(C))(A)$ of $g$
under the ``action'' $\mathsf{MG}_E(A)\to \Aut(\CAlg(\mathsf{R}_E)(C))(A)$
is a class of an equivalence $\CAlg(\mathsf{R}_E)(C)\otimes_KA\stackrel{\sim}{\to}\CAlg(\mathsf{R}_E)(C)\otimes_KA$ in $\CAlg_A$ obtained from
the automorphism $u$
by evaluating at $C$ (composing with the map
$\Delta^0\to \CAlg(\DM^\otimes(k))$ determined by $C$).
\end{Remark}
\begin{Remark}
\label{motsubcategory}
One can replace $\DM^\otimes(k)=\mathcal{C}^\otimes$ by
a stable subcategory $\mathcal{E}^\otimes\subset \DM^\otimes(k)$ that is closed under small colimits
and is generated by a small set of dualizable objects.
Again by the main result of \cite{Tan} there is a derived affine group
scheme $\mathsf{MG}_{E,\mathcal{E}^\otimes}$ that represents
$\Aut(\mathsf{R}_E|_{\mathcal{E}^\otimes})$, and for $C\in \CAlg(\mathcal{E}^\otimes)$,
$\mathsf{MG}_{E,\mathcal{E}^\otimes}$ acts on $\CAlg(\mathsf{R}_E(C))$.
In certain good cases, one can obtain
$\mathsf{MG}_{E,\mathcal{E}^\otimes}$ by means of equivariant
bar constructions, see
\cite{Bar}, \cite{DTD}, \cite{Spi}.
\end{Remark}
\subsubsection{}
\label{prepre}
We start with some $\infty$-categorical preliminary constructions.
To make things elementary, we make some efforts to make extensive use of
the machinery of simplicial categories, i.e., simplicially enriched categories, whereas in
the earlier version of this manuscript in 2016,
many constructions heavily rely on the theory of left/(co)Cartesian fibrations.
\vspace{2mm}
Let $\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}$ be a simplicial category defined
as follows. The objects of $\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}$
are symmetric monoidal small
$\infty$-categories $\mathcal{C}^\otimes\to \Gamma$.
Give two symmetric monoidal $\infty$-categories $\mathcal{C}^\otimes \to \Gamma$ and $\mathcal{D}^\otimes\to \Gamma$, we define
$\Fun_{\Gamma}^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$
to be the full subcategory of $\Fun_{\Gamma}(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$ that consists of symmetric monoidal functors (cf. \cite[2.1.2]{HA}).
We define the mapping simplicial set
$\Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes):=\Map_{\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}}(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$ to be the largest Kan subcomplex of
$\Fun_{\Gamma}^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$.
The composition is defined by the restriction of composition of
function complexes. The $\infty$-category $\operatorname{Cat}_{\infty}^{\textup{sMon}}$ is defined to be the simplicial nerve of $\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}$.
We let $\operatorname{Cat}_{\infty}^{\Delta}$ be the simplicial category
defined as follows. Objects are $\infty$-categories, and given two $\infty$-categories $\mathcal{C}$ and $\mathcal{D}$, the simplicial set $\Map(\mathcal{C},\mathcal{D})$ is the largest Kan subcomplex of $\Fun(\mathcal{C},\mathcal{D})$.
By definition, the simplicial nerve of $\operatorname{Cat}_{\infty}^{\Delta}$ is $\Cat$.
Let $\operatorname{Kan}^{\Delta}$ be the simplicial full subcategory of
$\operatorname{Cat}_{\infty}^{\Delta}$ that consists of Kan complexes.
For a symmetric monoidal $\infty$-category $\mathcal{C}^\otimes$,
the assignment $\mathcal{D}^\otimes \mapsto \Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$ determines a simplicial functor $h^\Delta_{\mathcal{C}^\otimes}:\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta} \to \operatorname{Kan}^{\Delta}$ in the natural way.
Taking the simplicial nerve, we obtain $h_{\mathcal{C}^\otimes}:=\NNNN(h_{\mathcal{C}^\otimes}^\Delta):\operatorname{Cat}_{\infty}^{\textup{sMon}}=\NNNN(\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta})\to \NNNN(\operatorname{Kan}^{\Delta})=\SSS$. We remark that it is equivalent to the functor $\operatorname{Cat}_{\infty}^{\textup{sMon}} \to \SSS$ corepresented by
$\mathcal{C}^\otimes$ defined in \cite[5.1.3]{HTT} (in the dual form).
Similarly, for an $\infty$-category $\mathcal{C}$,
the assigment $\mathcal{D} \mapsto \Map(\mathcal{C},\mathcal{D})$ determines a simplicial functor $h^\Delta_{\mathcal{C}}:\operatorname{Cat}_{\infty}^{\Delta} \to \operatorname{Kan}^{\Delta}$.
Taking the simplicial nerve, we obtain $h_{\mathcal{C}}:=\NNNN(h_{\mathcal{C}}^\Delta):\operatorname{Cat}_{\infty}=\NNNN(\operatorname{Cat}_{\infty}^{\Delta})\to \NNNN(\operatorname{Kan}^{\Delta})=\SSS$.
Next we construct a functor
$\CAlg:\smCat\to \Cat$
from the $\infty$-category of symmetric monoidal (small) $\infty$-categories
to the $\infty$-category of $\infty$-categories, which sends $\mathcal{C}^\otimes$ to $\CAlg(\mathcal{C}^\otimes)$.
For this purpose we construct a simplicial functor
\[
\CAlg^{\Delta}:\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}\longrightarrow \operatorname{Cat}_{\infty}^{\Delta}
\]
which carries $\mathcal{C}^\otimes\to \Gamma$ to $\CAlg(\mathcal{C}^\otimes)=\Fun^{\textup{lax}}_{\Gamma}(\Gamma,\mathcal{C}^\otimes)$ where
$\Fun^{\textup{lax}}_{\Gamma}(-,-)$ indicates the full subcategory of
$\Fun_{\Gamma}(-,-)$ that consists of lax symmetric monoidal functors.
To do this, given two symmetric monoidal $\infty$-categories
we will define a map of simplicial sets
$\Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)\to \Map(\CAlg(\mathcal{C}^\otimes),\CAlg(\mathcal{D}^\otimes))$.
Let $K$ be a simplicial set and $f:K\to \Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$ a map of simplcial sets.
The map amounts to a map of marked simplicial sets $\mathcal{C}^\otimes\times K^{\sharp}\to \mathcal{D}^\otimes$ over $\Gamma$ where $K^\sharp$ denotes
the marked simplicial sets such that all edges are marked.
To the map we associate a map
of simplicial sets $\CAlg(\mathcal{C}^\otimes)\times K\to \CAlg(\mathcal{D}^\otimes)$, equivalently $K\to \Fun(\CAlg(\mathcal{C}^\otimes),\CAlg(\mathcal{D}^\otimes))$ as follows.
Note that for a simplicial set $S$, $S\to \Fun_{\Gamma}(\Gamma,\mathcal{C}^\otimes\times_{\Gamma}(\Gamma\times K))$ corresponds to a pair of maps $S\times \Gamma\to \mathcal{C}^\otimes$ over $\Gamma$ and $S\times \Gamma \to K$.
To $S\to \CAlg(\mathcal{C}^\otimes)\times K$ corresponding to
$\phi:S\times \Gamma\to \mathcal{C}^\otimes$ over $\Gamma$ and $\psi:S\to K$
we associate $S\to \Fun_{\Gamma}(\Gamma,\mathcal{C}^\otimes\times_{\Gamma}(\Gamma\times K))$ corresponding to the pair
$\phi:S\times \Gamma\to \mathcal{C}^\otimes$ over $\Gamma$
and $S\times \Gamma\stackrel{\textup{pr}_1}{\to} S\stackrel{\psi}{\to} K$.
It gives rise to a map
\[
r:\Fun_{\Gamma}^{\textup{lax}}(\Gamma,\mathcal{C}^\otimes)\times K \to \Fun_{\Gamma}(\Gamma,\mathcal{C}^\otimes\times K).
\]
Let $c:\Fun_{\Gamma}(\Gamma,\mathcal{C}^\otimes\times K)\times \Fun_{\Gamma}(\mathcal{C}^\otimes\times K,\mathcal{D}^\otimes)
\to \Fun_{\Gamma}(\Gamma,\mathcal{D}^\otimes)$
be composition. Let $\iota:\Delta^0\to \Fun_{\Gamma}(\mathcal{C}^\otimes\times K,\mathcal{D}^\otimes)$ be the map determined by $\mathcal{C}^\otimes\times K \to\mathcal{D}^\otimes$ over $\Gamma$ that corresponds to $f$.
Consider the following composite
\begin{eqnarray*}
\Fun_{\Gamma}^{\textup{lax}}(\Gamma,\mathcal{C}^\otimes)\times K \simeq (\Fun^{\textup{lax}}_{\Gamma}(\Gamma,\mathcal{C}^\otimes)\times K)\times \Delta^0 &\stackrel{r\times \iota}{\longrightarrow}& \Fun_{\Gamma}(\Gamma,\mathcal{C}^\otimes\times K)\times \Fun_{\Gamma}(\mathcal{C}^\otimes\times K,\mathcal{D}^\otimes) \\
&\stackrel{c}{\longrightarrow}& \Fun_{\Gamma}(\Gamma,\mathcal{D}^\otimes).
\end{eqnarray*}
The image of composition is contained in $\Fun_{\Gamma}^{\textup{lax}}(\Gamma,\mathcal{D}^\otimes)$. Therefore we obtain $\CAlg(\mathcal{C}^\otimes)\times K\to \CAlg(\mathcal{D}^\otimes)$ from $f$. According to
the functoriality with respect to $K$
it yields
\[
\Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)\to \Fun(\CAlg(\mathcal{C}^\otimes),\CAlg(\mathcal{D}^\otimes)).
\]
Since $\Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)$ is a Kan complex,
its image is contained in $\Map(\CAlg(\mathcal{C}^\otimes),\CAlg(\mathcal{D}^\otimes))$. It is straightforward to see that
$\mathcal{C}^\otimes\mapsto \CAlg(\mathcal{C}^\otimes)$ and
$\Map^\otimes(\mathcal{C}^\otimes,\mathcal{D}^\otimes)\to \Map(\CAlg(\mathcal{C}^\otimes),\CAlg(\mathcal{D}^\otimes))$ determine a simplicial
functor
$\CAlg^\Delta:\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}\to \operatorname{Cat}_{\infty}^{\Delta}$.
Taking the simplicial nerves we obtain a functor of $\infty$-categories
\[
\CAlg:\smCat \longrightarrow \Cat.
\]
There is another obvious simplicial functor
$\textup{For}^\Delta:\operatorname{Cat}_{\infty}^{\textup{sMon},\Delta}\longrightarrow \operatorname{Cat}_{\infty}^{\Delta}$ which carries any symmetric monoidal $\infty$-category $\pi:\mathcal{C}^\otimes\to \Gamma$
to the fiber $\pi^{-1}(\langle 1 \rangle)$, i.e., the underlying $\infty$-categories $\mathcal{C}$.
There is the forgetful functor $\CAlg(\mathcal{C}^\otimes)\to \mathcal{C}$
which is defined as
$\Fun_{\Gamma}^{\textup{lax}}(\Gamma,\mathcal{C}^\otimes) \to \Fun_{\Gamma}(\{\langle 1 \rangle\},\mathcal{C}^\otimes)$ induced by composition with $\{\langle 1 \rangle\}\to \Gamma$.
It gives rise to a simplicial natural transformation $\CAlg^\Delta\to \textup{For}^\Delta$.
\subsubsection{}
\label{pre2}
Replacing the universe $\mathbb{U}$ by a larger universe $\mathbb{U}\in \mathbb{V}$,
we define
the $\infty$-category $\wCat$ of $\mathbb{V}$-small $\infty$-categories
, the $\infty$-category $\wsmCat$ of symmetric monoidal $\mathbb{V}$-small $\infty$-categories, and $\widehat{\CAlg}:\wsmCat\to \wCat$ instead of $\Cat$, $\smCat$ and $\CAlg:\smCat\to \Cat$. But for simplicity we write $\CAlg$ for $\widehat{\CAlg}$.
Let $\Theta_K:\CAlg_K\to \wsmCat$ be a functor which carries $A$ to
$\Mod_A^\otimes$ where $\Mod_A^\otimes$ is the symmetric monoidal $\infty$-category of $A$-module spectra (see \cite{HA}, \cite[Appendix A.4]{HA} for the precise construction).
Any morphism $A\to A'$ maps to the symmetric monoidal
functor $\Mod_A^\otimes\to \Mod_{A'}^\otimes$ informally given by the base
change $\otimes_AA'$.
Let $\NNNN(\textup{For}^\Delta):\wsmCat\to \wCat$ be the forgetful functor.
We define $\Mod_{(-)}:\CAlg_K\to \wCat$ to be the composite
of $\Theta_K$ and the forgetful functor.
We define $\CAlg_{(-)}$ to be the composite
$\CAlg_K\stackrel{\Theta_K}{\longrightarrow} \wsmCat\stackrel{\CAlg}{\longrightarrow} \wCat$.
\begin{Definition}
\label{realauto}
Consider the composite $\rho:\CAlg_K \stackrel{\Theta_K}{\to} \wsmCat\stackrel{h_{\DM(k)^\otimes}}{\longrightarrow} \wSSS$, which carries $A$ to $\Map^\otimes(\DM^\otimes(k),\Mod_A^\otimes)$.
Let $\mathsf{R}_E:\DM^\otimes(k)\to \mathsf{D}^\otimes(K)=\Mod_K^\otimes$
be the realization functor.
It may be viewed
as an object of $\Map^\otimes(\DM^\otimes(k),\Mod_K^\otimes)$.
Applying Definition~\ref{autogroup} to $\rho:\CAlg_K \to \wSSS$
and $\mathsf{R}_E$ we define the automorphism group functor
$\Aut(\mathsf{R}_E):\CAlg_K\to \Grp(\wSSS)$ of $\mathsf{R}_E$
over $\CAlg_K$.
\end{Definition}
\begin{Remark}
The definition of $\Aut(\mathsf{R}_E)$ is apparently different from
that in \cite{Tan} because in {\it loc.cit.} we use the full subcategory
$\DM_\vee^\otimes(k)$ spanned by compact (dualizable) objects
instead of $\DM^\otimes(k)$. But this point is neglective.
Since $\DM^\otimes(k)$ is canonically
equivalent to the symmetric monoidal
$\infty$-category $\Ind(\DM_\vee^\otimes(k))$
of Ind-objects, thus by the (symmetric monodial) Kan extension,
we see that there is a canonical equivalence $\Aut(\mathsf{R}_E)\simeq \Aut(\mathsf{R}_E|_{\DM_\vee^\otimes(k)})$ induced by the restriction
to $\DM_\vee^\otimes(k)\subset \DM^\otimes(k)$.
\end{Remark}
\subsubsection{}
{\it Construction of the action/Proof of Proposition~\ref{goodaction}}.
Let $L$ be an $\infty$-category.
Consider the following three simplicial functors:
\begin{itemize}
\item Put $\alpha^\Delta=h_{\DM^\otimes(k)}^\Delta:\widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta}\to \widehat{\operatorname{Kan}}^{\Delta}$.
It sends a symmetric monoidal $\infty$-category
$\mathcal{D}^\otimes$ to the Kan complex
$\Map^\otimes(\DM^\otimes(k),\mathcal{D}^\otimes)$.
\item Let $\beta_{L}^\Delta:\widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta}\to \widehat{\operatorname{Kan}}^{\Delta}$ be a simplicial functor that carries $\mathcal{D}^\otimes$ to $\Map(L,\CAlg(\mathcal{D}^\otimes))$. It is defined as the composite
$\widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta}\stackrel{\CAlg^\Delta}{\longrightarrow} \widehat{\operatorname{Cat}}_{\infty}^{\Delta}\stackrel{h_{L}}{\longrightarrow} \widehat{\operatorname{Kan}}^{\Delta}$.
\item Let $\gamma_{L}^\Delta:\widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta}\to \widehat{\operatorname{Kan}}^{\Delta}$ be a simplicial functor that carries $\mathcal{D}^\otimes$
to $\Map(L,\mathcal{D})$.
It is defined as the composite
$\widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta}\stackrel{\operatorname{For}^\Delta}{\longrightarrow} \widehat{\operatorname{Cat}}_{\infty}^{\Delta} \stackrel{h_{L}}{\longrightarrow} \widehat{\operatorname{Kan}}^{\Delta}$.
\end{itemize}
For each $\mathcal{D}^\otimes$, the simplicial functor $\CAlg^\Delta$ induces a map of simplicial sets
\[
\Map^\otimes(\DM^\otimes(k),\mathcal{D}^\otimes)\to
\Map(\CAlg(\DM^\otimes(k)),\CAlg(\mathcal{D}^\otimes)).
\]
It is easy to check that these maps determine a simplicial natural
transformation $\alpha^\Delta\to \beta_{\CAlg(\DM^\otimes(k))}^\Delta$.
Similarly, $\operatorname{For}^\Delta$ induces a map of simplicial sets
\[
\Map^\otimes(\DM^\otimes(k),\mathcal{D}^\otimes)\to
\Map(\DM(k),\mathcal{D}).
\]
It gives rise to
a simplicial natural
transformation $\alpha^\Delta\to \gamma_{\DM(k)}^\Delta$.
Let $L\to \CAlg(\DM^\otimes(k))$ be a functor.
The composition induces simplicial natural transformation
$\beta_{\CAlg(\DM^\otimes(k))}^\Delta \to \beta_{L}^\Delta$.
Also, $L\to \CAlg(\DM^\otimes(k))\stackrel{\textup{forget}}{\to} \DM(k)$
induces
$\gamma_{\DM(k)}^\Delta \to \gamma_{L}^\Delta$.
Now applying the simplicial nerve functor to $\alpha^\Delta$
we obtain $\alpha=h_{\DM^\otimes(k)}:\wsmCat \to \wSSS$.
Similarly, we obtain $\beta_{L},\gamma_L:\wsmCat \to \wSSS$
from $\beta_{L}^\Delta$ and $\gamma_L^\Delta$.
Consider the simplicial natural transformation
$\alpha^\Delta\to \beta_{\CAlg(\DM^\otimes(K))}^\Delta \to \beta_{L}^\Delta$.
It determines a natural transformation from $\alpha$ to $\beta_L$.
In fact,
we think of $\alpha^\Delta\to \beta_{L}^\Delta$
as a simplicial functor $[1]\times \widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta} \to \operatorname{Kan}^\Delta$
such that $[1]=\{0,1\}$ is the linearly ordered set
regarded as a (simplicial) category, and
the restriction to $\{0\}\times \widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta} \to \operatorname{Kan}^\Delta$
(resp. $\{1\}\times \widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta} \to \operatorname{Kan}^\Delta$)
is $\alpha^\Delta$
(resp. $\beta_L^\Delta$).
Since the simplicial nerve functor preserves products,
$\Delta^1\times \wsmCat\simeq \NNNN(\{0\to 1\}\times \widehat{\operatorname{Cat}}_{\infty}^{\textup{sMon},\Delta}) \to \NNNN(\widehat{\operatorname{Kan}}^\Delta)=\wSSS$
defines a natural transformation from $\alpha$ to $\beta_L$,
that is, $\Delta^1\times \wsmCat\to \wSSS$ such that
$\{0\}\times \wsmCat\to \wSSS$ is $\alpha$, and $\{1\}\times \wsmCat\to \wSSS$
is $\beta_L$.
Similarly, $\alpha^\Delta\to \gamma^{\Delta}_{\DM(k)}\to \gamma_{L}^\Delta$
determines a natural transformation from $\alpha$ to $\gamma_L$.
\vspace{1mm}
Next for $p:L\to \CAlg(\DM^\otimes(k))$ and $h=\CAlg(\mathsf{R}_E)\circ p$,
we construct an action of $\MG_E$ on $\Aut(h)$ (cf. Definition~\ref{symfuncauto}).
If $C$ is an object of $\CAlg(\DM^\otimes(k))$,
the automorphism group functor $\Aut(\CAlg(\mathsf{R}_E(C)))$ of $\CAlg(\mathsf{R}_E(C))$ over $\CAlg_k$
is nothing but $\Aut(h)$ where $L=\Delta^0$, and the functor $p:\Delta^0\to \CAlg(\DM^\otimes(k))$ is determined by $C$.
Let $\Delta^1\times \wsmCat\to \wSSS$
be the natural transformation from $\alpha$ to $\beta_L$
defined above. Composing with $\Theta_K$, we have
$\Delta^1\times \CAlg_K \to \wSSS$, that is a natural transformation
from $\rho=\alpha\circ \Theta_K$ to $(-)^\simeq\circ \mu_L=\beta_L\circ \Theta_K$
(we here use the notation in Definition~\ref{symfuncauto}, \ref{realauto}).
Remember that $\mathsf{R}_E$ is an object of $\Map^\otimes(\DM^\otimes(k),\Mod_K^\otimes)$. Thus, as in Definition~\ref{autogroup},
both $\alpha$ and $\beta_L$ are respectively promoted to functors
$\alpha_*,\beta_{L*}:\wsmCat\to \wSSS_*$ extended by $\mathsf{R}_E$
and $h\in \Map(L,\CAlg_K)$, and
$\Delta^1\times \CAlg_K \to \wSSS$ is promoted to
a natural transformation $\Delta^1\times \wsmCat \to \wSSS_*$
from $\alpha_*$ to $\beta_{L*}$.
Composing $\Omega_*:\wSSS_*\to \Grp(\wSSS)$ and $\Theta_K$,
we obtain
\[
\Delta^1\times \CAlg_K\to \Delta^1\times \wsmCat\to \wSSS_*\to \Grp(\wSSS)
\]
that is a natural tranformation from $\Aut(\mathsf{R}_E)$ to
$\Aut(h)$ (cf. Definition~\ref{symfuncauto}, \ref{realauto}).
Since we have the equivalence $\MG_E\simeq \Aut(\mathsf{R}_E)$,
it defines a morphism $\MG_E\simeq \Aut(\mathsf{R}_E)\to \Aut(h)$
in $\Fun(\CAlg_K,\Grp(\wSSS))$. An action of $\MG_E$ on $h$
is defined to be this morphism.
\vspace{1mm}
We prove the property (1).
For a map $g:M\to L$, there is a simplicial natural
transformation $\beta^\Delta_L\to \beta_M^\Delta$ induced by the composition
with $g$. Therefore, by our construction
the functoriality is obvious.
Next we prove the property (2).
Let $\mathcal{K}$ be the full subcategory
of $\Fun(L^\triangleright,\CAlg_A)$, that consists of
those functors $F:L^\triangleright\to \CAlg_K$
such that the image of the cone point of $L^{\triangleright}$ is a colimit of the restriction $F|_{L}$.
Then by taking account of left Kan extensions \cite[4.3.2.15]{HTT} (keep in mind that $\CAlg_A$
admits small colimits),
the map $\Fun(L^\triangleright,\CAlg_A)\to \Fun(L,\CAlg_A)$ given by
the restriction induces
an equivalence $\mathcal{K}\stackrel{\sim}{\to} \Fun(L,\CAlg_A)$
of $\infty$-categories.
Note that $\overline{p}:L^\triangleright\to \CAlg(\DM^\otimes(k))$ is a colimit diagram
(of $L\to \CAlg(\DM^\otimes(k))$).
The composite
$\overline{q}:L^\triangleright\to \CAlg(\DM^\otimes(k))\stackrel{\CAlg(\mathsf{R}_E))}{\to} \CAlg_K$ is also a colimit diagram
because $\CAlg(\mathsf{R}_E)$ is a left adjoint.
Also, the base change $\otimes_KA:\CAlg_K\to \CAlg_A$ is a left
adjoint.
Thus, the composite $L^\triangleright \stackrel{\overline{q}}{\to} \CAlg_K\to \CAlg_A$ belongs to $\mathcal{K}$.
By these observations, we see that $\Aut(\overline{q})\to \Aut(q)$
induced by the restriction is an equivalence in $\Fun(\CAlg_K,\Grp(\wSSS))$.
By the functoriality (1), we have the desired factorization of the action.
Finally, we prove (3).
On can define $\MG_E\to \Aut(\mathsf{R}_E(C))$
by using
$\alpha^\Delta \to \gamma_{\DM(k)}^\Delta \to \gamma_L^\Delta$
and $\mathsf{R}_E$ in the same way as we constructed
$\MG_E\to \Aut(\CAlg(\mathsf{R}_E)(C))$
from
$\alpha^\Delta \to \beta_{\DM^\otimes(k)}^\Delta \to \beta_L^\Delta$
and $\mathsf{R}_E$.
There is a simplicial natural transformation
$\beta_L^\Delta\to \gamma_L^{\Delta}$
which is given by $\Map(L,\CAlg(\mathcal{D}^\otimes))\to \Map(L,\mathcal{D}))$
induced by the composition with the forgetful functor
$\CAlg(\mathcal{D}^\otimes)\to \mathcal{D}$ for each $\mathcal{D}^\otimes$.
By the simplicial nerve fucntor and the construction in Definiton~\ref{autogroup}, it gives rise to $\Aut(\CAlg(\mathsf{R}_E)(C))\to \Aut(\mathsf{R}_E(C))$.
Note that the simplicial natural transformation
$\beta_L^\Delta\to \gamma_L^{\Delta}$ commutes with $\alpha^\Delta\to \beta_L^\Delta$ and $\alpha^\Delta\to \gamma_L^\Delta$.
By this commutativily we see that $\MG_E\to \Aut(\CAlg(\mathsf{R}_E)(C))\to \Aut(\mathsf{R}_E(C))$ is naturally equivalent to
$\MG_E\to \Aut(\mathsf{R}_E(C))$.
\QED
\begin{Remark}
\label{freemot}
Let $M$ be an object of $\DM(k)$.
Let $\Free_{\DM(k)}(M)$ in $\CAlg(\DM^\otimes(k))$
be the free commutative algebra
object generated by $M$ (see Definition~\ref{freealg}).
Let us observe that the action of $\MG_E$ on $\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M))$ is essentially determined by
the action of of $\MG_E$ on $\mathsf{R}_E(M)$.
Since the realization functor is a left adjoint,
there is a canonical equivalence
$\Free_{K}(\mathsf{R}_E(M))\simeq \CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M))$
where $\Free_K:=\Free_{\Mod_K}$ is the free functor $\Mod_K\to \CAlg_K$,
i.e., the left adjoint to the forgetful functor.
Let $S$ be a space that belongs to $\SSS$.
Let $f:S\to \MG_E(K)\simeq \Aut(\mathsf{R}_E)(K)$ be a morphism (in $\SSS$).
Let $g:S\to \Aut(\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)))(K)\simeq \Map_{\CAlg_K}(\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)),\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)))$
be a class of the map induced by the action of $f$.
The forgetful functor induces morphisms
\begin{eqnarray*}
\Map_{\CAlg_K}(\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)),\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)))\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \to \Map_{\Mod_K}(\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M))^{\sharp},\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M))^{\sharp}) \\
\simeq \Map_{\Mod_K}(\Free_{K}(\mathsf{R}_E(M))^{\sharp},\Free_{K}(\mathsf{R}_E(M))^{\sharp})\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\end{eqnarray*}
in $\SSS$ where $(-)^\sharp$ here indicates the underlying object.
By the compatibility (3) in Proposition~\ref{goodaction},
the image of $g$
is equivalent to the map
\[
h:S\to \Aut(\Free_{K}(\mathsf{R}_E(M))^{\sharp})(K)\simeq \Map_{\Mod_K}(\Free_{K}(\mathsf{R}_E(M))^{\sharp},\Free_{K}(\mathsf{R}_E(M))^{\sharp})
\]
that is determined by the action of $f$ on $\Free_{K}(\mathsf{R}_E(M))^{\sharp}$.
The composition with the canonical (unit) map $\mathsf{R}_E(M)\to \Free_{K}(\mathsf{R}_E(M))^{\sharp}$ yields the morphisms
\begin{eqnarray*}
\Map_{\Mod_K}(\Free_{K}(\mathsf{R}_E(M))^{\sharp},\Free_{K}(\mathsf{R}_E(M))^{\sharp}) &\to& \Map_{\Mod_K}(\mathsf{R}_E(M),\Free_{K}(\mathsf{R}_E(M))^{\sharp}) \\ &\stackrel{i}{\leftarrow}& \Map_{\Mod_K}(\mathsf{R}_E(M),\mathsf{R}_E(M))
\end{eqnarray*}
in $\SSS$. By the functoriality similar to (1) in Proposition~\ref{goodaction},
we see that the image of $h$ in $\Map_{\Mod_K}(\mathsf{R}_E(M),\Free_{K}(\mathsf{R}_E(M))^{\sharp})$ is equivalent to the image of
$r:S\to \Aut(\mathsf{R}_E(M))(K)\simeq \Map_{\Mod_K}(\mathsf{R}_E(M),\mathsf{R}_E(M))$ that is determined by the action of $f$ on $\mathsf{R}_E(M)$.
Note that by the adjunction,
the composition gives an equivalence
\[
\Map_{\CAlg_K}(\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)),\CAlg(\mathsf{R}_E)(\Free_{\DM(k)}(M)))\stackrel{\sim}{\to} \Map_{\Mod_K}(\mathsf{R}_E(M),\Free_{K}(\mathsf{R}_E(M))^{\sharp})
\]
in $\SSS$.
Also, the left arrow $i$ is a fully faithful functor
since $\mathsf{R}_E(M)\to \Free_{K}(\mathsf{R}_E(M))^{\sharp}$ defines a direct summand
of $\Free_{K}(\mathsf{R}_E(M))^{\sharp}$.
The image of $g$ in $\Map_{\Mod_K}(\mathsf{R}_E(M),\Free_{K}(\mathsf{R}_E(M))^{\sharp})$ lies in the essential image of $i$.
The image of $g$ is equivalent to
the image of $r$ under $i$. One can adopt this argument
to not only $K$ but arbitrary $A\in\CAlg_K$.
We remark that any object of $\CAlg(\DM^\otimes(k))$
is constructed from free commutative algebra objects
by forming colimits, see Section~\ref{Sull1}.
\end{Remark}
\subsection{}
\label{diagramaction}
Let $\Fun(\NNNN(\Delta^{op}),\Aff_K)$ be the $\infty$-category
of simplicial diagrams in $\Aff_K$. The $\infty$-category of
group objects in $\Aff_K$, i.e., derived affine group schemes,
is its full subcategory consisting of those simplicial diagram
satisfying the condition of group objects (cf. Section~\ref{preDAG}, see also
\cite[Appendix]{Tan}).
In Section~\ref{diagramaction}
we focus on actions on such objects.
We continue to use the notation in Section~\ref{actionstep1}.
Let $\mathfrak{C}:\NNNN(\Delta) \to \CAlg(\DM^\otimes(k))$ be
a functor which we regard as a cosimplicial diagram of commutative algebra objects in $\DM^\otimes(k)$.
Suppose that $\mathfrak{C}^{op}:\NNNN(\Delta)^{op} \to \CAlg(\DM^\otimes(k))^{op}$ is a group object.
One of our main examples is the opposite of the group object
$\mathcal{G}^{(n+1)}(X,x):\NNNN(\Delta)^{op}\to \CAlg(\DM^\otimes(k))^{op}$ introduced in Section~\ref{hopfsection}.
The multiplicative realization functor $\CAlg(\mathsf{R}_E)$ preserves coproducts and
sends a unit to $K\in \CAlg_K$.
It follows that
the composite
\[
G_\bullet:\NNNN(\Delta)^{op} \stackrel{\mathfrak{C}^{op}}{\to} \CAlg(\DM^\otimes(k))^{op} \stackrel{\CAlg(\mathsf{R}_E)}{\longrightarrow} \CAlg_K^{op}=\Aff_K
\]
is a group object, that is, a derived affine group scheme over $K$.
We denote it simply by
$G$.
Invoking Proposition~\ref{goodaction} (see also Definition~\ref{symfuncauto})
to the opposite of the group object
$G_\bullet^{op}=h:\NNNN(\Delta)=L\to \CAlg_K$,
we get a morphism
\[
\mathsf{MG}_E\to \Aut(G_\bullet^{op})
\]
in $\Fun(\CAlg_K,\Grp(\wSSS))$, that is, an action of $\mathsf{MG}_E$
on $G_\bullet^{op}$.
Put
$\Aut(G):=\Aut(G_\bullet^{op})$.
Thus we have
\begin{Proposition}
\label{actiongroup}
Let $\mathfrak{C}^{op}:\NNNN(\Delta)^{op}\to \CAlg(\DM^\otimes(k))^{op}$
be a group object.
Let $G$ be the derived affine group scheme over $K$ that is induced by
$\mathfrak{C}^{op}$.
Then there is a (canonical) action of $\mathsf{MG}_E$ on $G$, that is a morphism $\mathsf{MG}_E\to \Aut(G)$
in $\Fun(\CAlg_K,\Grp(\wSSS))$.
\end{Proposition}
\begin{Remark}
We remark that informally $\Aut(G_\bullet^{op})$ is the automorphism group
of the cosimplicial object $G_\bullet^{op}$
in $\CAlg_K$.
Therefore, by our convention $\Aut(G)=\Aut(G_\bullet^{op})$, the morphism
$\mathsf{MG}_E\to \Aut(G)$ should be viewed as the ``right action'' on $G$
that corresponds to the ``left action'' on $G_\bullet^{op}$.
\end{Remark}
\begin{Remark}
The action is functorial with repect to a morphism of derived
affine group schemes.
Let
$\mathfrak{C}^{'op}:\NNNN(\Delta)^{op}\to \CAlg(\DM^\otimes(k))^{op}$
be another group object and $G'_{\bullet}:\NNNN(\Delta)^{op}\to \CAlg_K^{op}=\Aff_K$ the derived affine group scheme induced by the composition with
the multiplicative realization functor.
Suppose that there is a morphism (i.e., a natural
transformation) $\mathfrak{C}^{op}\to \mathfrak{C}^{'op}$.
It gives rise to $\theta:\Delta^1\times \NNNN(\Delta)\to \CAlg(\DM^\otimes(k))\to \CAlg_K$, such that $\{0\}\times \NNNN(\Delta) \to \CAlg_K$ is $G_{\bullet}^{'op}$,
and $\{0\}\times \NNNN(\Delta) \to \CAlg_K$ is $G_{\bullet}^{op}$.
By (1) of Proposition~\ref{goodaction}
the actions of $\MG_E$ on $G_{\bullet}^{op}$ and
$G_{\bullet}^{'op}$ are simultaneously promoted to
an action on $\Aut(\theta)$, i.e., $\MG_E\to \Aut(\theta)$.
\end{Remark}
\begin{Example}
\label{exgroup}
Let $(X,x:\Spec k\to X)$ be a pointed smooth variety over $k$.
As discussed in Section~\ref{realhopfsection} it gives rise to
a derived affine group scheme $G_E^{(n)}(X,x):\NNNN(\Delta)^{op}\to \Aff_{K}$.
Therefore, $\mathsf{MG}_E$ acts on $G_E^{(n)}(X,x)$.
\end{Example}
\subsection{}
\label{underlyinggroup}
In \cite{Tan} we defined the motivic Galois group
$MG_E$ of $\DM(k)$ (with respect to $E$)
to be a usual affine group scheme over $K$
(i.e., a pro-algebraic group) obtained from $\MG_E$.
Also, we can construct a usual affine group scheme
$\GG_E^{(n)}(X,x)$ from $G_E^{(n)}(X,x)$, Example~\ref{exgroup}.
In general,
if $G$ is a derived affine group scheme over the field of characteristic zero
$K$,
one can obtain a usual affine group scheme (i.e., pro-algebraic group)
$\overline{G}$
over $K$
from $G$, which we will call the underlying affine
group scheme (cf. \cite{Tan}).
We briefly review the procedure.
\vspace{1mm}
Let $\CAlg_K^{dg}$ be the category of
commutative dg algebras $C$ over $K$ (cf. Section~\ref{convention}).
Let $\CAlg_K^{dg,\ge0}$ be the full subcategory
of $\CAlg_K^{dg}$ that consists of
those objects $C$ such that
$H^i(C)=0$ for $i<0$.
It admits a combinatorial model category structure
such that a morphism $f:C\to C'$
is a weak equivalence (resp.
fibration) if the underlying map is a quasi-isomorphism
(resp. a surjective in each degree), see
\cite[Proposition 5.3]{Ol} or \cite[Theorem 6.2.6]{Fre}. Any object
is fibrant.
Any ordinary commutative algebra over $K$ is
a cofibrant object in $\CAlg_K^{dg,\ge0}$
when it is regarded as a commutative
dg algebra placed in degree zero.
The inclusion $\CAlg_K^{dg,\ge0}\hookrightarrow \CAlg_K^{dg}$
is a right Quillen functor. Its left adjoint $\tau:\CAlg_K^{dg}\to \CAlg_K^{dg,\ge0}$ carries
$C$ to the quotient of $C$ by the differential graded ideal
generated by elements $x\in C^i$ for $i<0$.
Namely, we have a Quillen adjunction $\tau:\CAlg_K^{dg}\rightleftarrows \CAlg_K^{dg,\ge0}$. We shall write
$\CAlg_K^{\ge0}$ for the $\infty$-category obtained from
the full subcatgory of cofibrant objects in $\CAlg_K^{dg,\ge0}$
by inverting weak equivalences.
The Quillen adjunction induces an adjunction of $\infty$-categories
\[
\tau:\CAlg_K\rightleftarrows \CAlg_K^{\ge0}
\]
\cite{Maz} where by ease of notation we write $\tau$ also for
the induced left adjoint functor
$\CAlg_K\rightarrow \CAlg_K^{\ge0}$.
We put $G=\Spec C$ with $C\in \CAlg_K$.
The functor $\tau$ preserves colimits, especially coproducts.
We put $\Aff_K^{\ge0}=(\CAlg_K^{\ge0})^{op}$.
We write $\Spec R$ for the object in $\Aff_K^{\ge0}$ corresponding to
$R\in \CAlg_K^{\ge0}$.
Then $\Spec \tau C$ inherits a group structure from $G=\Spec C$.
Namely, $\Spec \tau C$ is a group object
in $\Aff_K^{\ge0}$.
There is a fully faithful left adjoint $\CAlg_K^{\DIS}\to \CAlg_K^{\ge0}$
induced by the natural inclusion from the category of ordinary
commutative $K$-algebras to $\CAlg_K^{dg,\ge0}$.
Its right adjoint
$\CAlg_K^{\ge0} \to \CAlg_K^{\DIS}$
is given by taking the cohomology $C \mapsto H^0(C)$.
The inclusion $\CAlg_K^{\DIS}\to \CAlg_K$ is canonically
equivalent to
the composite $\CAlg_K^{\DIS}\to \CAlg_K^{\ge0}\to \CAlg_K$.
Also, the left adjoint $\tau$ is compatible with
inclusions
$\CAlg_K^{\DIS}\subset \CAlg_K$ and $\CAlg_K^{\DIS}\subset \CAlg^{\ge0}_K$
(use the fact that any object $C$ in $\CAlg_K^{\ge0}$ is the limit of a
cosimplicial diagram of ordinary $K$-algebras).
Consider $G$ to be the functor $\CAlg_K\to \Grp(\SSS)$.
Its restriction $G^\circ:=G|_{\CAlg_K^{\DIS}}:\CAlg_K^{\DIS}\to \Grp(\SSS)$
is naturally equivalent to the functor given by $A\mapsto \Map_{\CAlg_K^{\ge0}}(\tau C,A)$.
Take the cohomology $H^0(\tau C)$.
The structure of a commutative Hopf ring spectrum on $\tau C$
over $K$ (that is, the ``dual'' of the group structure
on $\Spec \tau C$ in $\Aff_K^{\ge0}$, see \cite[Appendix]{Tan}) gives the structure of a commutative Hopf ring on $H^0(\tau C)$ over $K$.
Namely, the comultiplication $\tau C\to \tau C\otimes_K\tau C$,
the counit $\tau C\to K$ and the antipode give rise to the structure of comultiplication
$H^{0}(\tau C)\to H^0(\tau C\otimes_K\tau C)\simeq H^0(\tau C)\otimes_K H^0(\tau C)$ of $H^0(\tau C)$, etc. We denote the associated affine group scheme
by $\overline{G}=\Spec H^0(\tau C)$.
We shall refer to $\overline{G}$ as the {\it underlying affine group
scheme} of $G$ (or the {\it coarse moduli space} for $G$ as in \cite{Tan}).
The assignment $G\mapsto \overline{G}$
is functorial and we actually have a
functor $\Grp(\Aff_K)\to \Grp(\Aff_K^{\DIS})$ which sends $G$ to
the associated affine group scheme $\overline{G}$.
By the adjunction,
the natural morphism $\pi:\Spec \tau C\to \overline{G}=\Spec H^0(\tau C)$ is universal among morphisms to ordinary
affine schemes over $K$ in $\hhh(\Aff_K^{\ge0})$
(note that $\Aff_K^{\ge0}$ contains $\Aff_K^{\DIS}$ as a full subcategory).
Namely, if $\phi:\Spec \tau C \to H$ is a
morphism to an ordinary affine scheme
$H$ in $\hhh(\Aff_K^{\ge0})$, there is a unique morphism $\psi:\overline{G}\to H$
such that $\phi=\psi\circ \pi$.
In addition, $H$ is an affine group scheme and $\phi:\Spec \tau C \to H$ is a
homomorphism to the affine group scheme over $K$,
then there is a unique homomorphism $\psi:\overline{G}\to H$ in $\hhh(\Grp(\Aff_K^{\ge0})$
such that $\phi=\psi\circ \pi$.
\vspace{2mm}
As mentioned above,
we define $MG_E$ to be the underlying affine group scheme of $\MG_E$.
For the properties of $MG_E$ we refer to \cite{Tan}, \cite{Bar}, \cite{DTD},
\cite{PM}.
We define $\GG^{(n)}(X,x):=\GG_E^{(n)}(X,x)$ to be the underlying affine group scheme of $G_E^{(n)}(X,x)$ (cf. Section~\ref{realhopfsection}).
We consider a geometric interpretation of $\GG^{(n)}(X,x)$.
Suppose that $K=\QQ$ and
the base field $k$ is embeded in $\CC$.
We consider the case when the realization functor
is associated to singular cohomology theory.
\begin{Proposition}
\label{undcomp}
Let $(X,x:\Spec k \to X)$ be a pointed smooth variety over $k$.
Let $\pi_i(X^t,x)$ be the homotopy group of the underlying topological
space $X^t=X\times_{\Spec k}\Spec \CC$.
For any $n\ge1$, the affine group schemes $\overline{G}^{(n)}(X,x)$
is a unipotent group scheme (i.e., a pro-unipotent algebraic group).
Moreover, $\overline{G}^{(1)}(X,x)$ is the pro-unipotent completion of
$\pi_1(X^t,x)$ over $K=\QQ$.
Suppose further that the topological space
$X^t$ is nilpotent and of finite type (e.g. simply connected smooth varieties).
Then $\overline{G}^{(n)}(X,x)$ is a pro-unipotent completion of $\pi_n(X^t,x)$
for $n\ge2$.
\end{Proposition}
Before proceeding the proof,
we briefly recall
the notion of affinization (affination in French) studied in \cite{AS}
(in \cite{AS}, cosimplicial algebras are used instead of dg algebras,
see \cite[6.4]{Fre} for the comparison as a Quillen
equivalence between the model category of cosimplicial algebras
and $\CAlg_K^{dg,\ge0}$).
Let $\widehat{\CAlg}^{\ge0}_K$ be the $\mathbb{V}$-version of
$\CAlg^{\ge0}_K$ (cf. Section~\ref{underlyinggroup}).
Write $\widehat{\textup{Aff}}^{\ge0}_K:=(\widehat{\CAlg}^{\ge0}_K)^{op}$.
We write $\Spec R$ for an object of $\widehat{\textup{Aff}}^{\ge0}_K$
corresponding to $R\in \widehat{\CAlg}^{\ge0}_K$.
There is an adjunction
\[
\mathcal{O}:\Fun(\CAlg^{\textup{dis}}_K,\widehat{\SSS}) \rightleftarrows \widehat{\Aff}^{\ge0}_K
\]
where $\mathcal{O}$ is a left Kan extension of
the inclusion
$\Aff^{\textup{dis}}_K\hookrightarrow \widehat{\Aff}^{\ge0}_K$
along the Yoneda embedding $\Aff^{\textup{dis}}_K\to \Fun(\CAlg^{\textup{dis}}_K,\widehat{\SSS})$
(cf. \cite[Section 2.2]{AS}).
The right adjoint sends $R\in \CAlg_K^{\ge0}$ to the functor $h_R:\CAlg^{\textup{dis}}_K\to \widehat{\SSS}$ informally given by
$A\mapsto \Map_{\widehat{\CAlg}^{\ge0}_K}(R,A)$.
The restriction $\Aff^{\ge0}_K=(\CAlg_K^{\ge0})^{op}\to \Fun(\CAlg^{\textup{dis}}_K,\widehat{\SSS})$ of the right adjoint is fully faithful.
Let $F:\CAlg_K^{\DIS}\to \wSSS$ be a functor.
If $\mathcal{O}(F)$ belongs to $\Aff_{K}^{\ge0}$ (not to $\widehat{\Aff}_{K}^{\ge0}$), we refer to $\mathcal{O}(F)$
as the affinization of $F$.
An object $P$ in $\SSS$
can be viewed as the constant functor $\CAlg_{K}^{\DIS}\to \widehat{\SSS}$
with value $P$.
One can consider the affinization of the space $P\in \SSS$.
The composite
$\SSS=\Fun(\Delta^0,\SSS)\to \Fun(\CAlg_K^{\DIS},\widehat{\SSS}) \to \widehat{\Aff}^{\ge0}_K$
preserves small colimits and sends a contractible space to $\Spec K$,
where the first arrow is the functor given by the composition
with $\CAlg_K^{\DIS}\to \Delta^0$.
Consequently, the composite carries the space $S \in\SSS$ to
$\Spec K^{S}$ where $K^S$ is the cotensor with the space $S$.
By Proposition~\ref{PLproperty} and Remark~\ref{PLremark},
we conclude that $\SSS\to \Aff_K^{\ge0}\to \Aff_K$
is equivalent to $A_{PL,\infty}$.
By Theorem~\ref{real1}, $\Spec T_X$ in $\Aff_{K}$ is the affinization of $X^t$.
\vspace{2mm}
\Proof
There are several ways to prove the assertion, and we will give one
of them.
We treat the case $n=1$.
Let $G^{(1)}(X,x)^{\circ}:\CAlg_{K}^{\DIS}\subset\CAlg_K\to \Grp(\wSSS)$
denote the restriction.
It carries $A$ to $\Omega_*\Spec T_X(A)$, where $\Spec T_X(A)$
is the space of $A$-valued points on $\Spec T_X$, and $\Omega_*\Spec T_X(A)$
is its base loop space (the base point comes from $x_A:\Spec A\to \Spec K\to \Spec T_X$).
We let $G_\circ^{(1)}(X,x):\CAlg_K^{\DIS}\to \Grp(\textup{Set})$
be the sheaf of groups with respect to fpqc topology associated to
the presheaf $A \mapsto \pi_0(\Omega_*\Spec T_X(A))\simeq \pi_1(\Spec T_X(A),x_A)$.
Then according to \cite[2.4.5]{AS}
(or \cite[4.4.8]{DAG}), $G_\circ^{(1)}(X,x)$
is represented by a unipotent affine group scheme
(i.e., a pro-unipotent algebraic group).
(We remark that
there is a canonical equivalence $\Map_{\CAlg_K}(T_X, A)\simeq \Map_{\CAlg_K^{\ge0}}(K^{X^t},A)$ for any $A\in \CAlg_K^{\DIS}$,
see \cite[7.2]{Ol}.)
Note that the natural morphism $G^{(1)}(X,x)^{\circ} \to G_\circ^{(1)}(X,x)$
is universal among morphisms to sheaves of groups on $\CAlg_K^{\DIS}$.
On the other hand, there is the natural map $G^{(1)}(X,x)^{\circ}\to \overline{G}^{(1)}(X,x)$
(recall that if $G^{(1)}(X,x)=\Spec C$, the restriction
$G^{(1)}(X,x)^{\circ}$ is represented by $\Spec \tau C$).
Consequently, by the universal property
there is a natural morphism $G_\circ^{(1)}(X,x)\to\overline{G}^{(1)}(X,x)$ of affine group schemes over $K$.
We wish to show that it is an isomorphism.
Since $K=\QQ$ is characteristic zero and
$G_\circ^{(1)}(X,x)\to \overline{G}^{(1)}(X,x)$
is a morphism as affine group schemes over $K$,
it is enough to prove that
for any algebraically closed field $L$, the induced map
$G_\circ^{(1)}(X,x)(L)\to \overline{G}^{(1)}(X,x)(L)$
of sets of $L$-valued points is bijective.
In fact, according to \cite[Theorem 5.17]{Tan}
(its proof that works also for $G^{(1)}(X,x)$ instead of $\MG_E$) and \cite[VIII 4.4.8]{DAG},
we see that
$G_\circ^{(1)}(X,x)(L)\to \overline{G}^{(1)}(X,x)(L)$ is bijective.
It follows that $\overline{G}^{(1)}(X,x)$
is a unipotent affine group scheme.
By \cite[2.4.11]{AS} and Theorem~\ref{real1}, the group
scheme
$G_\circ^{(1)}(X,x)\simeq \overline{G}^{(1)}(X,x)$ is naturally isomorphic to a pro-unipotent completion of $\pi_1(X^t,x)$ (that is
endowed with the morphism
form the constant functor with value $\pi_1(X^t,x)$).
The case of $n\ge 2$ is similar.
If $G^{(n)}(X,x)^{\circ}:\CAlg_{K}^{\DIS}\subset\CAlg_K\to \Grp(\wSSS)$
denotes the restriction of $G^{(n)}(X,x)$,
it carries $A$ to the $n$-fold
loop space $\Omega_*^{n}\Spec T_X(A)$.
As in the case of $n=1$ (\cite[2.4.5]{AS}),
we observe that
the sheaf associated to
the presheaf $A \mapsto \pi_n(\Spec T_X(A),x_A)$ is
isomorphic to $\overline{G}^{(n)}(X,x)$.
Then the final assertion follows from \cite[2.5.3]{AS}.
\QED
\subsection{}
\label{finalaction}
We will construct an action of the motivic Galois group $MG_E$ on
the affine group scheme $\GG^{(n)}(X,x):=\GG_E^{(n)}(X,x)$.
Unfortunately, if one does not assume motivic conjectures
that imply the existence of a motivic $t$-structure,
it seems difficult to obtain an action of $MG_E$ on $\GG^{(n)}(X,x)$
from that of $\MG_E$ on $G^{(n)}(X,x)$ in a purely categorical way.
To overcome this issue, we use a method of homological algebras,
which yields
a natural action of $MG_E$ on $\GG_E^{(n)}(X,x)$.
For a usual affine group scheme $H$ over $K$,
we let $\Gamma(H)$ be the (ordinary) coordinate ring
on $H$, that is a commutative Hopf ring over $K$.
We let $\Aut(H):\CAlg_K^{\DIS}\to \Grp(\textup{Set})$
be the functor which assigns $A$ to the group of automorphisms
of the commutative Hopf ring
$\Gamma(H)\otimes_KA \stackrel{\sim}{\to} \Gamma(H)\otimes_KA$ over $A$.
\begin{Theorem}
\label{actiononhomotopy}
Let $(X,x)$ be a pointed smooth variety over $k$.
Then there is a (canonical)
morphism $MG_E\to \Aut(\overline{G}^{(n)}(X,x))$ in $\Fun(\CAlg_K^{\DIS},\Grp(\textup{Set}))$, that is,
an action of $MG_E$ on $\GG^{(n)}(X,x)$.
In other words, the action is described as an action on the scheme $\GG^{(n)}(X,x)$
\[
\GG^{(n)}(X,x)\times MG_E \to \GG^{(n)}(X,x)
\]
which is compatible with the group structure. Moreover,
the following properties hold:
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item The action is functorial:
Let $\phi:(X,x)\to (Y,y)$ be a morphism of smooth varieties
over $k$ that sends $x$ to $y$.
Let $\phi_*:\overline{G}^{(n)}(X,x)\to \overline{G}^{(n)}(Y,y)$
be the induced morphism of group schemes.
Then the action of $MG_E$ commutes with $\phi_*$.
\item
The action has a moduli theoretic interpretation in a coarse sense
(see Remark~\ref{coarsemoduli}).
\end{enumerate}
\end{Theorem}
\begin{Corollary}
\label{actiononhomotopy2}
Suppose that $k$ is embedded in $\CC$ and consider the case of singular
realization.
Let $\pi_i(X^t,x)_{uni}$ be the pro-unipotent completion
of $\pi_i(X^t,x)$ over $\QQ$.
Then
we have a canonical action
\[
\pi_1(X^t,x)_{uni} \times MG \to \pi_1(X^t,x)_{uni}.
\]
If $X^t$ is nilpotent and of finite type, there
is a canonical action of $MG$ on $\pi_n(X^t,x)_{uni}$
for $n\ge2$.
\end{Corollary}
\Proof
It follows from Theorem~\ref{actiononhomotopy} and Proposition~\ref{undcomp}.
\QED
{\it Construction of an action/Proof of Theorem~\ref{actiononhomotopy}.}
Let $G^{(n)}(X,x):\NNNN(\Delta)^{op}\to \Aff_K$
be the derived affine group scheme over $K$, associated
to $(X,x)$ (see Section~\ref{realhopfsection}).
Let $\Gamma(G^{(n)}(X,x))$ be the image of
$[1]$ under $G^{(n)}(X,x)^{op}:\NNNN(\Delta)\to \CAlg_K$.
(Namely, $\Gamma(G^{(n)}(X,x))$ is the underlying algebra of commutative Hopf algebra
object $G^{(n)}(X,x)^{op}$ in $\CAlg_K$.)
Let $\MG_E=\Spec C$.
The identity
$\MG_E\to \MG_E$ determines a component of the space
($\infty$-groupoid) $\MG_E(C)$.
The action on $G^{(n)}(X,x)$
(cf. Proposition~\ref{actiongroup} and Example~\ref{exgroup})
induces its image in
$\Aut(G^{(n)}(X,x))(C)$.
The equivalence class of the image in $\Aut(G^{(n)}(X,x))(C)$
gives rise to a morphism
$\Gamma(G^{(n)}(X,x))\otimes_K C\stackrel{\sim}{\to} \Gamma(G^{(n)}(X,x))\otimes_K C$
in $\CAlg_C$ (cf. Remark~\ref{autogroupR}).
Composing with the unit $K\to C$, we have
\[
\theta:\Gamma(G^{(n)}(X,x))=\Gamma(G^{(n)}(X,x))\otimes_K K \to \Gamma(G^{(n)}(X,x))\otimes_K C\stackrel{\sim}{\to} \Gamma(G^{(n)}(X,x))\otimes_K C.
\]
The composite is a coaction of $C$ on $\Gamma(G^{(n)}(X,x))$
at the level of the homotopy category $\hhh(\CAlg_K)$.
Namely, if we think of $C$
as a coalgebra in $\hhh(\CAlg_K)$
determined by the class of comultiplication
$C\to C\otimes_KC$ and the unit $C\to K$,
then
$\Gamma(G^{(n)}(X,x))\to\Gamma(G^{(n)}(X,x))\otimes_K C$
is an (associative) coaction on $\Gamma(G^{(n)}(X,x))$
in the obvious sense. Also, it commutes with the structure of
coalgebra on $\Gamma(G^{(n)}(X,x))$ at the level of homotopy category.
Let $B:=\tau C$ (see Section~\ref{underlyinggroup} for $\tau$.
Applying $\tau$ to $\theta$ we obtain
\[
\rho:\tau(\Gamma(G^{(n)}(X,x)))\to \tau(\Gamma(G^{(n)}(X,x)))\otimes_K B\stackrel{\sim}{\to} \tau(\Gamma(G^{(n)}(X,x)))\otimes_K B.
\]
Taking the cohomology in the $0$-th term we have
\begin{eqnarray*}
\xi:H^0(\tau(\Gamma(G^{(n)}(X,x))))\to H^0(\tau(\Gamma(G^{(n)}(X,x)))\otimes_K B) &\stackrel{\sim}{\to}& H^0(\tau(\Gamma(G^{(n)}(X,x)))\otimes_K B) \\
&\simeq& H^0(\tau(\Gamma(G^{(n)}(X,x)))\otimes H^0(B).
\end{eqnarray*}
Recall that the commutative Hopf
ring $\Gamma(\overline{G}^{(n)}(X,x))$ of the coordinate ring on $\overline{G}^{(n)}(X,x)$
is $H^0(\tau(\Gamma(G^{(n)}(X,x))))$ equipped with the structure of
commutative Hopf ring that comes from the structures on $\Gamma(G^{(n)}(X,x))$.
Moreover, $MG_E=\Spec H^0(B)$.
The morphism $\xi$ is a coaction of $H^0(B)$ on the commutative $K$-algbera $H^0(\tau(\Gamma(G^{(n)}(X,x))))=\Gamma(\overline{G}^{(n)}(X,x))$ which is
compatible with the structure of coalgebra
on $H^0(\tau(\Gamma(G^{(n)}(X,x))))$.
It gives rise to an action
\[
\GG^{(n)}(X,x)\times MG_E \to \GG^{(n)}(X,x).
\]
The functoriality (1) is obvious from the construction.
\QED
\begin{Remark}
\label{coarsemoduli}
The affine group scheme $MG_E$ is a coarse moduli
space for $\MG_E$.
It has a coarse moduli theoretic interpretation:
for any field $L$ over $K$, $\MG_E^\circ \to MG_E$ induces
an isomorphism $\pi_0(\MG_E(L))\stackrel{\sim}{\to} MG_E(L)$
of sets where $\pi_0(\MG_E(L))$ is the set of connected components,
i.e., the set of equivalence classes of $L$-valued points on $\MG_E$
(cf. \cite[Theorem 1.3]{Tan}).
By $\MG_E\simeq \Aut(\mathsf{R}_E)$,
the set $MG_E(K)$ is naturally identified with
the set of equivalence classes of the automorphism
of $\mathsf{R}_E:\DM^\otimes(k)\to\Mod^\otimes_K$.
Suppose that $q\in MG_E(K)$ corresponds to an automorphism $\sigma$
of $\mathsf{R}_E$.
The automorphism of $\GG^{(n)}(X,x)$ induced by $q$
is the automorphism induced by $\sigma$.
Recall that $\sigma$ induces an automorphism of the multiplicative
realization functor
$\CAlg(\mathsf{R}_E):\CAlg(\DM^\otimes(k))\to\CAlg_K$ (cf. Section~\ref{actionstep1}).
It gives rise to an automorphism
on $G^{(n)}(X,x)^{op}:\NNNN(\Delta)\to \CAlg_K$
(cf. Section~\ref{diagramaction}).
The induced automorphism $\Gamma(G^{(n)}(X,x))\stackrel{\sim}{\to}\Gamma(G^{(n)}(X,x))$
gives rise to
$a:H^0(\tau\Gamma(G(X,x)))\stackrel{\sim}{\to}H^0(\tau\Gamma(G(X,x)))$.
By our construction, the action of $q$ is equal to $a$.
This interpretation holds also for any field $L$ over $K$.
\end{Remark}
\section{Sullivan models and computational results}
\label{Sullivanmodel}
In rational homotopy theory, an inductive construction of a Sullivan model
is quite powerful.
Let $S$ be a topological space and $A_{PL}(S)$ the
commutative dg algebra of polynomial differential forms.
As in Section~\ref{realization}
we write $A_{PL,\infty}(S)$ for the image of $A_{PL}(S)$ in $\CAlg_{\QQ}$.
Let $\Free_{\QQ}$ denote the free functor $\Mod_{\QQ}\simeq \mathsf{D}(\QQ)\to \CAlg_{\QQ}$ which is defined to be a left adjoint to the forgetful functor
$\CAlg_{\QQ}\to \mathsf{D}(\QQ)$.
Contrary to genuine commutative dg algebras,
in the setting of $\CAlg_{\QQ}$ it is nonsense to say
what a underlying graded algebra is.
But in the language of $\CAlg_{\QQ}$, the inductive construction describes
$A_{PL,\infty}(S)$ as a colimit of a sequence
\[
\QQ\simeq A_0\to A_1\to \cdots \to A_n\to A_{n+1}\to \cdots
\]
such that for any $n\ge 0$, $A_{n+1}$ fits in the pushout diagram of the form
\[
\xymatrix{
\Free_{\QQ}(V) \ar[r] \ar[d] & A_n \ar[d] \\
\QQ \ar[r] & A_{n+1}
}
\]
in $\CAlg_{\QQ}$ where $V$ is a $\ZZ$-graded vector space over $\QQ$
regarded as an object in $\mathsf{D}(\QQ)$, and the vertical arrow
is $\Free_{\QQ}(V)\to \Free_{\QQ}(0)\simeq \QQ$ induced by $V\to 0$.
Note that $\Free_{\QQ}(V)\to A_n$ is determined by a morphism $V\to A_n$
in $\mathsf{D}(\QQ)$.
Suppose that $V$ is concentrated in a fixed positive degree $n$, i.e.,
$V^i=0$ for $i\ne n$, and the $\QQ$-vector space $V^n$ is finite dimensional.
Then $\Free_{\QQ}(V)$ is
the commutative dg algebra that corresponds to the rational homotopy type
of the Eilenberg-MacLane space $K((V^n)^\vee,n)$.
Informally, the above sequence may be thought of as a presentation
of $A_{PL}(S)$ as a ``successive extension'' of ``simple pieces''
of the form $\Free_{\QQ}(V[1])$.
We will apply this approach to $\CAlg(\DM^\otimes(k))$
and study motivic cohomological algebras.
Free commutative algebra objects in $\DM(k)$ play the role of
free commutative dg algebras.
Actually, from the tannakian viewpoint,
such free objects are quite ``simple'' objects, see Remark~\ref{freemot}.
Put another way,
presentations of successive extensions by free objects
is useful for computations of a motivic counterpart
of rational homotopy groups.
We will inroduce the notion of
cotangent motives in Section~\ref{cotangenthomotopy}.
We then apply
the study of structures of cohomological motivic algebras in this Section
to obtain explicit descriptions of cotangent motives (Theorem~\ref{motivichomotopyexp}).
In this Section, we work with rational coefficients,
but $\QQ$ can be replaced by any field $K$ of characteristic zero.
\subsection{}
\label{Sullivanmodel0}
We will study some ``elementary examples'' such as projective spaces.
We also
hope that the reader will get the feeling of
the idea of the
constructions of ``Sullivan models'' of motivic cohomological algebras
in $\CAlg(\DM^\otimes(k))$.
Recall free commutative algebras in a general setting.
\begin{Definition}
\label{freealg}
Let $\mathcal{C}^\otimes$ be a symmetric monoidal
$\infty$-category
that has small colimits and
the tensor product $\otimes:\mathcal{C}\times \mathcal{C}\to \mathcal{C}$
preserves small colimits separately in each variable.
Let $u_{\mathcal{C}}:\CAlg(\mathcal{C}^\otimes)\to \mathcal{C}$
be the forgetful functor.
By \cite[3.1.3]{HA}, there exists a left adjoint
\[
\Free_{\mathcal{C}}:\mathcal{C} \longrightarrow \CAlg(\mathcal{C}^\otimes)
\]
to $u_\mathcal{C}$, which we shall call the free functor of
$\mathcal{C}^\otimes$ (\cite{HA} treats a broader setting).
Given $C\in \mathcal{C}$ we refer to $\Free_{\mathcal{C}}(C)$
as the free commutative algebra (object) generated by $C$.
We often omit the notation $u_{\mathcal{C}}$.
For $A\in\CAlg(\mathcal{C}^\otimes)$,
by the adjunction,
a morphism $f:\Free_{\mathcal{C}}(C)\to A$
corresponds to the composite
$\alpha:C\stackrel{\textup{unit}}{\to} u_{\mathcal{C}}(\Free_{\mathcal{C}}(C))\stackrel{u_{\mathcal{C}}(f)}{\to} u_{\mathcal{C}}(A)$ in $\mathcal{C}$.
We say that $f:\Free_{\mathcal{C}}(C)\to A$
is classified by $\alpha$.
According to \cite[3.1.3.13]{HA}, the underlying object
$\Free_{\mathcal{C}}(C)$ is equivalent to the coproduct
$\sqcup_{n\ge0} \Sym^n(C)$
in $\mathcal{C}$, where $\Sym_{\mathcal{C}}^n(C)$ is the $n$-fold symmetric product (we usually omit the subscript when the setting is obvious).
If $\mathcal{D}^\otimes$ is a symmetric monoidal
$\infty$-category having the same property and
$F:\mathcal{C}^\otimes\to \mathcal{D}^\otimes$
is a colimit-preserving functor, then there is a canonical
equivalence
$\Free_{\mathcal{D}}(F(C))\stackrel{\sim}{\to} F(\Free_{\mathcal{C}}(C))$
for any $C\in \mathcal{C}$.
\end{Definition}
\vspace{2mm}
\subsubsection{}
We start with results that are useful for computations.
Let $\GL_{d}$ be the general linear algebraic group over $\QQ$.
Let $\Vect^\otimes(\GL_d)$ be the symmetric monoidal
abelian category of (possibly
infinite dimensional) representations of $\GL_d$, that is,
$\QQ$-vector spaces
with action of $\GL_d$.
The symmetric monoidal category
$\Comp(\GL_d):=\Comp(\Vect(\GL_d))$ of (possibly unbounded) cochain complexes
admits a proper combinatorial symmetric monoidal model structure
such that (i) $f:C\to C'$ is a weak equivalence
if a quasi-isomorphism, (ii) every object is cofibrant,
and (iii) $\{\iota_M:S^{n+1}M \hookrightarrow D^nM\}_{M\in I \atop \\ n\in \ZZ}$ is
a set of generating cofibrations consisting of natural inclusions,
where $I$ is the set of irreducible
representations of $\GL_d$, and
$S^nM$ (reps. $D^nM$) in $\Comp(\GL_d)$ defined by $(S^nM)^n=M$ and $(S^nM)^m=0$
for $m\ne n$ (resp. $(D^nM)^{n}=(D^{n}M)^{n+1}=M$, $D^mM=0$
for $m\ne n, n+1$, and $d:(D^{n}M)^{n}\to (D^{n}M)^{n+1}$ is the identity),
see \cite[Section 2.3]{DTD}, \cite[Corollary 3.5]{CD1} for details.
Let $\Rep^\otimes(\GL_d)$ be the symmetric
monoidal $\infty$-category, which
is obtained from $\Comp(\GL_d)$
by inverting quasi-isomorphisms.
Let $\CAlg(\Rep^\otimes(\GL_d))$ be the $\infty$-category
of commutative algebra objects in $\Rep^\otimes(\GL_d)$.
\begin{Lemma}
\label{GLcommalg}
We denote by $\CAlg(\Comp(\GL_d))$ the category of commutative algebra
objects in $\Comp(\GL_d)$. (We may think of an object
as a commutative dg algebra equipped with action of $\GL_d$.)
Then there is a combinatorial model structure on
$\CAlg(\Comp(\GL_d))$ such that
a morphism $f:A\to A'$ in $\CAlg(\Comp(\GL_d))$
is a weak equivalence (resp. a fibration) if
$f$ is a weak equivalence (reps. a fibration) in the underlying category
$\Comp(\GL_d)$.
In addition, if $\CAlg(\Comp(\GL_d))[W^{-1}]$
denotes the $\infty$-category obtained from the full subcategory
of cofibrants in $\CAlg(\Comp(\GL_d))$
by inverting weak equivalences, then the canonical functor
\[
\CAlg(\Comp(\GL_d))[W^{-1}]\to \CAlg(\Rep^\otimes(\GL_d))
\]
is an equivalence of $\infty$-categories.
\end{Lemma}
\Proof
Thanks to \cite[4.5.4.4, 4.5.4.6, 4.5.4.7]{HA}, it is enough
to prove that every cofibration in $\Comp(\GL_d)$ is a power cofibration
in the sense of \cite[4.5.4.2]{HA}.
To this end, we first observe that
a morphism $f:C\to C'$ in $\Comp(G):=\Comp(\Vect(G))$
is a cofibration if and only if $f$ is a monomorphism
when $G$ is either $\GL_d$ or a symmetric group $\Sigma_n$.
Let $M$ be an irreducible representation of $G$.
By the representation theory of $\GL_d$ or $\Sigma_n$,
$\Vect(G)$ is semi-simple and
$\Hom_{\Vect(G)}(M,M)=\QQ$ for any irreducible representation $M$ of
$G$. Let $\xi_M:\Vect(G)\to \Vect$ be the functor to the category of $\QQ$-vector spaces, that is given by $N\mapsto \Hom_{\Vect(G)}(M,N)$.
Taking the product indexed by the set $I(G)$ of irreducible representations
of $G$,
we have $\sqcap_{M\in I(G)} \xi_M:\Vect(G)\to \sqcap_{I(G)}\Vect$.
Note that this functor is an equivalence of categories
and induces an equivalence $\Comp(G)\to \sqcap_{I(G)}\Comp(\QQ)$
in the obvious way.
For an irreducible representation
$P$, $S^{n+1}P \to D^nP$ corresponds to a morphism $\{f_M\}_{M\in I(G)}$ in
$\sqcap_{I(G)}\Comp(\QQ)$ such that
$f_P:S^{n+1}\QQ\to D^n\QQ$ and $f_M=0$ if $M\ne P$ through this equivalence.
Therefore, it will suffice to show that
the smallest weakly saturated class containing $\{S^{n+1}\QQ\to D^n\QQ\}_{n\in \ZZ}$
coincides with a collection of monomorphisms in $\Comp(\QQ)$.
In fact, $\{S^{n+1}\QQ\to D^n\QQ\}_{n\in \ZZ}$ is a set of
generating cofibrations
in the projective model structure of $\Comp(\QQ)$.
Since $\QQ$ is a field, a morphism in $\Comp(\QQ)$ is a cofibration
with respect to the projective model structure exactly when it is
a monomorphism. Thus, we conclude that
a morphism $f:C\to C'$ in $\Comp(G)$
is a cofibration if and only if $f$ is a monomorphism.
Next we prove that a cofibration $f:C\to C'$ of $\Comp(\GL_d)$
is a power cofibration.
We say that $f$ is a power cofibration if
a $\Sigma_n$-equivariant map
$\wedge^n(f):\square^n(f)\to (C')^{\otimes n}$
is a cofibration in $\Comp(\GL_d)^{\Sigma_n}$ for any $n\ge 0$.
Here $\Comp(\GL_d)^{\Sigma_n}$ is the category of objects
in $\Comp(\GL_d)$
endowed with action of the symmetric group $\Sigma_n$,
which is equipped with the projective model structure.
We refer to \cite[4.4.4.1]{HA} for these definitions and notations.
Let $U:\Comp(\GL_d)\to \Comp(\QQ)$ be the forgetful functor, that
is a symmetric monoidal
left adjoint. It follows that
$\wedge^n U(f)\simeq U(\wedge^n(f))$.
Suppose that $f$ is a cofibration.
Then $U(f)$ is a cofibration with respect to the projective model
structure because it is a monomorphism. According to \cite[7.1.4.7]{HA},
$U(f)$ is a power cofibration. Thus
by the above consideration $U(\wedge^n(f))\simeq \wedge^n U(f)$ is a monomorphism.
Then $\wedge^n(f)$ is a monomorphism in $\Comp(\GL_d)$.
Note that there is a canonical
equivalence $(\sqcap_{I}\Comp(\QQ))^{\Sigma_n}
\stackrel{\sim}{\to} \sqcap_{I}(\Comp(\QQ)^{\Sigma_n})=\sqcap_{I}\Comp(\Sigma_n)$.
The image of $\wedge^n(f)$ in $\sqcap_{I}(\Comp(\QQ)^{\Sigma_n})$
is a monomorphism. Again by the above consideration, the image is
a cofibration in $\sqcap_{I}(\Comp(\QQ)^{\Sigma_n})$ endowed with
the projective structure.
Therefore, $\wedge^n(f)$ has the left lifting property
with respect to epimorphic quasi-isomorphisms
in $\Comp(\GL_d)^{\Sigma_n}$, namely, it is a cofibration.
\QED
Let $u:\CAlg(\Comp(\GL_d))\to \Comp(\GL_d)$
be the forgetful functor.
By the definition of the model structure on $\CAlg(\Comp(\GL_d))$
in Lemma~\ref{GLcommalg}, it is a right Quillen functor.
Let $\Free_{\Comp(\GL_d)}:\Comp(\GL_d)\to \CAlg(\Comp(\GL_d))$
be a left Quillen functor to $u$.
It is the free functor
of $\Comp(\GL_d)$. Since every object in $\Comp(\GL_d)$
is cofibrant, thus $\Free_{\Comp(\GL_d)}$ preserves weak equivalences;
that is to say, it is ``derived''.
Let $u_\infty:\CAlg(\Rep(\GL_d))\to \Rep(\GL_d)$ be the forgetful
functor
of $\infty$-categories. We write $\Free_{\Rep(\GL_d)}:\Rep(\GL_d)\to\CAlg(\Rep(\GL_d))$ for the free functor of $\Rep^\otimes(\GL_d)$.
The following Lemma guarantees compatibility between
$\Free_{\Rep(\GL_d)}$ and $\Free_{\Comp(\GL_d)}$.
\begin{Lemma}
\label{freecomp}
Let $C$ be an object in $\Comp(\GL_d)$.
By abuse of notation, we write $C$ (resp. $\Free_{\Comp(\GL_d)}(C)$)
for the images of the cofibrant object $C$ (resp. $\Free_{\Comp(\GL_d)}(C)$) in $\Rep(\GL_d)$
(resp. $\CAlg(\Rep(\GL_d))$).
Then there is a canonical equivalence
$\Free_{\Comp(\GL_d)}(C) \simeq \Free_{\Rep(\GL_d)}(C)$ in $\CAlg(\Rep(\GL_d))$,
which commutes with $C\to u_{\infty}(\Free_{\Comp(\GL_d)}(C))$ and
$C\to u_\infty(\Free_{\Rep(\GL_d)}(C))$.
\end{Lemma}
\Proof
The forgetful functors $u$ and $u_\infty$ commute with
canonical maps $\CAlg(\Comp(\GL_{d}))\to \CAlg(\Rep(\GL_d))$
and $\Comp(\GL_d)\to \Rep(\GL_d)$.
By Lemma~\ref{GLcommalg} we identify the induced functor $\hhh(u_\infty):\hhh(\CAlg(\Rep(\GL_d)))\to \hhh(\Rep(\GL_d))$ of homotopy categories with the right adjoint
\[
\overline{u}:\hhh(\CAlg(\Comp(\GL_d))[W^{-1}])\to \hhh(\Rep(\GL_d))
\]
of homotopy categories
induced by the right Quillen functor $u$.
Thus, we can identify the left adjoint
$\hhh(\Free_{\Rep(\GL_d)}):\hhh(\Rep(\GL_d))\to \hhh(\CAlg(\Rep(\GL_d)))$
with the left adjoint $\hhh(\Rep(\GL_d))\to \hhh(\CAlg(\Comp(\GL_d))[W^{-1}])$
induced by $\Free_{\Comp(\GL_d)}$.
\QED
\begin{Proposition}
\label{hopushout}
Let $A$ be a cofibrant object in $\CAlg(\Comp(\GL_d))$
and let $\alpha:C\to u(A)$ be a morphism in $\Comp(\GL_d)$.
Let $\phi_\alpha:\Free_{\Comp(\GL_d)}(C)\to A$ be the morphism
classified by $\alpha$. Let $\iota:S^{0}\QQ \hookrightarrow D^{-1}\QQ$
be the cofibration in $\Comp(\GL_d)$, where $\QQ$ here denotes the
unit object in $\Comp(\GL_d)$ (we abuse notation).
Let $\Free_{\Comp(\GL_d)}(C)\to \Free_{\Comp(\GL_d)}(C\otimes (D^{-1}\QQ))$ be the morphism induced by $C\otimes\iota:C\simeq C\otimes (S^0\QQ) \to C\otimes(D^{-1}\QQ)$.
Let $A\langle \alpha \rangle$ be the pushout of the following diagram in
$\CAlg(\Comp(\GL_d))$:
\[
\xymatrix
{
\Free_{\Comp(\GL_d)}(C)\ar[r]^{\phi_{\alpha}} \ar[d] & A \ar[d] \\
\Free_{\Comp(\GL_d)}(C\otimes (D^{-1}\QQ)) \ar[r] & A\langle \alpha \rangle.
}
\]
Then this diagram is a homotopy pushout.
See Remark~\ref{explicitdg} for the explicit presentation of $A\langle \alpha \rangle$.
\end{Proposition}
\begin{Remark}
\label{explicitdg}
The commutative algebra object $A\langle \alpha \rangle$
is regarded as a commutative dg algebra endowed with an action of $\GL_d$.
The explicit presentation of $A\langle \alpha \rangle$ is described as follows
(see the proof of Proposition~\ref{hopushout}).
For simplicity, we suppose that differential of $C$ is zero and
we view it as a graded vector space with an action of $\GL_d$.
This assumption is not essential in practice because $\Vect(\GL_d)$ is
semi-simple. Let $\overline{A}$ be the underlying
graded algebra
of $A$ obtained by forgetting the differential.
The underlying graded algebra of $A\langle \alpha \rangle $ is
given by the tensor product
$\overline{A}\otimes \Free_{\Comp(\GL_d)}(C[1])$
of commutative graded algebras with the action of $\GL_d$.
If one forgets the action of $\GL_d$ on $\Free_{\Comp(\GL_d)}(C[1])$,
then it is the free commutative graded algebra generated by
the underlying graded algebra of $C[1]$.
The differential
on $\overline{A}\otimes \Free_{\Comp(\GL_d)}(C[1])$
is given by the differential on $A$ and $\partial|_{C}=\alpha$.
When $\GL_d$ is the trivial, i.e., $d=0$ or one forgets the action of
$\GL_d$,
then the construction of $A\langle \alpha \rangle$ is
classical, see \cite[2.2.2]{Hin}.
\end{Remark}
\begin{Example}
\label{Sullex}
Let $\mathbb{G}_m=\GL_1$ and let $\chi_{i}$ in $\Comp(\mathbb{G}_m)$
be one dimensional representation of $\mathbb{G}_m$ of weight $i$
placed in degree zero.
Let $A=\Free_{\Comp(\mathbb{G}_m)}(\chi_1[-2])$ be the
free commutative algebra generated by $\chi_1[-2]$.
The underlying cochain complex is $\oplus_{i\ge0}\chi_i[-2i]$
with zero differential.
Let $\alpha:\chi_{n+1}[-2n-2]\to \oplus_{i\ge0}\chi_i[-2i]=A$
be the canonical inclusion.
Let us consider $A\langle \alpha \rangle$.
Note that $\Free_{\Comp(\mathbb{G}_m)}(\chi_{n+1}[-2n-1])$ is the trivial square zero extension $\chi_0\oplus \chi_{n+1}[-2n-1]$
by $\chi_{n+1}[-2n-1]$ (since the generator is in the odd degree).
The underlying graded algebra is $(\oplus_{i\ge0}\chi_i[-2i])\otimes (\chi_0\oplus \chi_{n+1}[-2n-1])$. The non-zero part of differential
is given by ``identities'' $\chi_i[-2i]\otimes \chi_{n+1}[-2n-1]\to \chi_{i+n+1}[-2i-2n-2]\otimes \chi_0$ for $i\ge0$.
\end{Example}
The standard consequence of Proposition~\ref{hopushout} is
\begin{Corollary}
The image of the square diagramin Proposition~\ref{hopushout}
in $\CAlg(\Rep(\GL_d))$ is a pushout diagram.
We remark that the image of $\Free_{\Comp(\GL_d)}(C)$
and $\Free_{\Comp(\GL_d)}(C\otimes (D^{-1}K))$ in $\CAlg(\Rep(\GL_d))$ are
equivalent to $\Free_{\Rep(\GL_d)}(C)$ and the unit algebra, respectively
(Lemma~\ref{freecomp}).
\end{Corollary}
{\it Proof of Proposition~\ref{hopushout}.}
Let $B$ be a pushout of $C\otimes D^{-1}\QQ \leftarrow C\to u(A)$ in $\Comp(\GL_d)$, that is the standard mapping cone $(u(A)\oplus C[1], d)$
of $\alpha:C\to u(A)$. Since $u(A)$ is cofibrant
and $C\otimes S^0K\to C\otimes D^{-1}K$
is a cofibration, $B$ is a homotopy pushout, see e.g. \cite[A.2.4.4]{HTT}.
Then we have the commutative diagram
\[
\xymatrix
{
\Free_{\Comp(\GL_d)}(C) \ar[r] \ar[d] & \Free_{\Comp(\GL_d)}(u(A)) \ar[r] \ar[d]& A \ar[d] \\
\Free_{\Comp(\GL_d)}(C\otimes D^{-1}K) \ar[r] & \Free_{\Comp(\GL_d)}(B)\ar[r] & A\langle \alpha \rangle
}
\]
that consists of pushout squares. The upper right horizontal map is the
counit map.
Since $A$ is cofibrant and
the left vertical arrow is an cofibration,
again by \cite[A.2.4.4]{HTT} both left and right
(and the outer) squares are homotopy pushouts,
as claimed. The explicit structure of $A\langle \alpha \rangle$ in Remark~\ref{explicitdg}
can easily be seen from the right pushout.
\QED
\subsubsection{}
\label{cps}
We will consider the $n$-dimensional projective space $\PP^n$
over a perfect field $k$.
We denote by $\Free_{\DM(k)}:\DM(k)\to \CAlg(\DM^\otimes(k))$
the free functor of $\DM^\otimes(k)$.
For ease of notation, we put $\Free:=\Free_{\DM(k)}$.
By projective bundle theorem, there is a decomposition
\[
M_{\PP^n}\simeq M(\PP^n)^\vee\simeq \uni_k\oplus \uni_k(-1)[-2]\oplus \ldots \oplus \uni_k(-n)[-2n]=\oplus_{i=0}^n\uni_k(-i)[-2i]
\]
in $\DM(k)$, see e.g. \cite[Lec.15]{MVW}.
Consider the inclusion $\iota:\uni_k(-1)[-2]\hookrightarrow M_{\PP^n}\simeq \oplus_{i=0}^n\uni_k(-i)[-2i]$ that is a morphism in $\DM(k)$.
It gives rise to a morphism
\[
f:\Free(\uni_k(-1)[-2])\to M_{\PP^n}
\]
in $\CAlg(\DM^\otimes(k))$,
that is classified by $\iota$.
We note that $\Free(\uni_k(-1)[-2])\simeq \oplus_{i\ge0} \uni(-i)[-2i]$ in $\DM(k)$.
Observe that for $j>n$, the composite
\[
\uni_k(-j)[-2j]\hookrightarrow \oplus_{i\ge0} \uni(-i)[-2i]\simeq \Free(\uni_k(-1)[-2])\to M_{\PP^n}
\]
is null homotopic. Indeed, $\uni_k(-j)[-2j]\to \uni_k(-i)[-2i]$ is null homotopic for $0\le i \le n$
since $\uni_k(j)[2j]\otimes(\uni_k(-j)[-2j]\to \uni_k(-i)[-2i])$
corresponds to
an element of motivic cohomology $H_{M}^{2j-2i}(\Spec k,j-i)\simeq \CH^{j-i}(\Spec k)=0$. Here $\CH^p(-)$ denotes the $p$-th
Chow group, and the comparison isomorphism between motivic cohomology
and (higher) Chow groups is due to Voevodsky.
Next
we let
\[
g:\Free(\uni_k(-n-1)[-2n-2])\to \Free(\uni_k(-1)[-2])
\]
be a morphism that is classified by the inclusion $\uni_k(-n-1)[-2n-2]\hookrightarrow \Free(\uni_k(-1)[-2])$. Consider the
morphism $h:\Free(\uni_k(-n-1)[-2n-2])\to \Free(0)\simeq \uni_k$ induced by $\uni_k(-n-1)[-2n-2]\to 0$.
Take a pushout
\[
S_{\PP^n}:=\Free_{\DM(k)}(\uni_k(-1)[-2])\otimes_{\Free_{\DM(k)}(\uni_k(-n-1)[-2n-2])}\uni_k
\]
along $h$ in $\CAlg(\DM^\otimes(k))$.
Note that $f\circ g$ factors through $h:\Free(\uni_k(-n-1)[-2n-2])\to \Free(0)\simeq \uni_k$ because
$\uni_k(-n-1)[-2n-2]\to M_{\PP^n}$ is null homotopic.
Consequently, by the universal property of the pushout
we obtain the induced morphism
\[
S_{\PP^n}\to M_{\PP^n}.
\]
\begin{Proposition}
\label{cpsprop}
The morphism $S_{\PP^n}\to M_{\PP^n}$
is an equivalence in $\CAlg(\DM^\otimes(k))$.
\end{Proposition}
\Proof
We first claim that
$\oplus_{i\ge0} \uni_k(-i)[-2i]\simeq \Free(\uni_k(-1)[-2])\to M_{\PP^n}\simeq \oplus_{i=0}^n \uni_k(-i)[-2i]$ induces an equivalence
$\Free(\uni_k(-1)[-2])\supset \uni_i(-i)[-2i]\stackrel{\sim}{\to} \uni_i(-i)[-2i]\subset M_{\PP^n}$ for $0\le i\le n$.
As discussed before this Proposition,
$\uni_i(-i)[-2i]\subset \Free(\uni_k(-1)[-2])\to M_{\PP^n}$ is null homotopic if $i>n$ because $\Hom_{\hhh(\DM(k))}(\uni_k(a)[2a],\uni_k(b)[2b])$
is $\QQ$ (resp. $0$) if $a=b$
(resp. $a\ne b$).
Consider the dual $M(\PP^n)\simeq \oplus_{i=0}^n \uni_k(i)[2i]$
of the isomorphism $M_{\PP^n}\simeq \oplus_{i=0}^n \uni_k(-i)[-2i]$.
Recall that the Chow ring $\CH^*(\PP^n)$ is isomorphic to
$\ZZ[H]/(H^{n+1})$ where $H\in \CH^1(\PP^n)$ is a class of
a hyperplane.
The projection
$M(\PP^n)\to \uni_k(i)[2i]$ corresponds to a generator of Chow group
$\QQ=\CH^i(\PP^n)\otimes_{\ZZ}\QQ\simeq H_{M}^{2i}(X,i)\simeq \Hom_{\hhh(\DM(k))}(M(\PP^n),\uni_k(i)[2i])$. Using scalar multiplication (if necessary),
we may and will assume that $M(\PP^n)\to \uni_k(i)[2i]$
corresponds to $H^i$.
Now we prove our claim by induction on $i$.
By the construction, the case of $i=1$ is clear.
We suppose that the case $i(< n-1)$ is true. We will show the case $i+1$.
By Lemma~\ref{homotopyfree},
$\Free(\uni_k(-1)[-2])$ in the homotopy category $\hhh(\DM(k))$
is also regarded as the free commutative algebra object lying in $\CAlg(\hhh(\DM^\otimes(k))$
generated by $\uni_k(-1)[-2]$ in $\hhh(\DM^\otimes(k))$.
Thus, the multiplication
map $\Free(\uni_k(-1)[-2])\otimes \Free(\uni_k(-1)[-2])\to \Free(\uni_k(-1)[-2])$ induces an isomorphism from the component $\uni_k(-a)[-2a]\otimes \uni_k(-b)[-2b]$ in the domain to $\uni_k(-a-b)[-2a-2b]$
in the target. Therefore, by the induction hypothesis
and the compatibility of multiplication maps,
if the multiplication $M_{\PP^n}\otimes M_{\PP^n}\to M_{\PP^n}$
induces an isomorphism of the composite
\[
\xi:\uni_k(-1)[-2]\otimes \uni_k(-i)[-2i]\hookrightarrow M_{\PP^n}\otimes M_{\PP^n}\to M_{\PP^n}\to \uni_k(-i-1)[-2i-2],
\]
then $\Free(\uni_k(-1)[-2])\to M_{\PP^n}$ induces an isomorphism
from the component $\uni_k(-i-1)(-2i-2)$ in the domain to $\uni_k(-i-1)[-2i-2]\subset M_{\PP^n}$ (namely, the case $i+1$ holds).
Note that the dual $M(\PP^n)\to \uni_k(i)[2i]$
of $\uni_k(-i)[-2i]\to M_{\PP^n}$
corresponds to the element $H^i\in \CH^i(\PP^n)$ (for any $i$).
Observe that the dual $M(\PP^n)\to \uni_k(i+1)[2i+2]$ of the composite
$l:\uni_k(-1)[-2]\otimes \uni_k(-i)[-2i]\hookrightarrow M_{\PP^n}\otimes M_{\PP^n}\to M_{\PP^n}$ corresponds to the intersection product
$H^{i+1}=H\cdot H^i\in \CH(\PP^n)$.
To see this, recall that the product of motivic cohomology
\begin{eqnarray*}
\Hom_{\hhh(\DM(k))}(M(\PP^n),\uni_k(1)[2])\otimes \Hom_{\hhh(\DM(k))}(M(\PP^n),\uni_k(i)[2i])\ \ \ \ \ \ \ \ \\
\ \ \ \ \ \ \ \ \to \Hom_{\hhh(\DM(k))}(M(\PP^n),\uni_k(i+1)[2i+2])
\end{eqnarray*}
is induced by the composition with $M(\PP^n)\to M(\PP^n)\otimes M(\PP^n)$ defined by the diagonal map.
By Lemma~\ref{multcomult} below, the multiplication
$M_{\PP^n}\otimes M_{\PP^n}\to M_{\PP^n}$ is the dual of
$M(\PP^n)\to M(\PP^n)\otimes M(\PP^n)$.
In addition, the product structure on motivic cohomology
is compatible with that of
(higher) Chow groups via the comparison isomorphism \cite{KY}.
Therefore, we conclude that
the dual of $l$ corresponds to $H^{i+1}\in \CH^{i+1}(\PP^n)$.
It follows that $\xi$ is an isomorphism.
Next, by \cite[Theorem 3.1]{DTD}
there is a colimit-preserving symmetric monoidal
functor $F:\Rep^\otimes(\mathbb{G}_m)\to \DM^\otimes(k)$
which sends one dimensional representation $\chi_1$
of weight one placed in degree zero to $\uni_k(1)$.
Here $\mathbb{G}_m:=\GL_1$ and we denote by $\chi_p$ one dimensional
representation of weight $p$.
Let $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])$
and $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n-1}[-2n-2])$
be the free commutative algebra in $\Comp(\mathbb{G}_m)$
generated by $\chi_{-1}[-2]$ and $\chi_{-n-1}[-2n-2]$, respectively.
Let $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n-1}[-2n-2])\to \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])$ be the morphism classified by the inclusion
$\alpha:\chi_{-n-1}[-2n-2]\hookrightarrow \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])$.
Take a homotopy pushout $ \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])\langle \alpha \rangle$, see Proposition~\ref{hopushout}.
By Remark~\ref{explicitdg}, the easy computation shows that
$ \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])\langle \alpha \rangle \simeq \oplus_{i=0}^n\chi_{-i}[-2i]$ in $\hhh(\Rep(\mathbb{G}_m))$
and the natural map
\[
\Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])\simeq \oplus_{i\ge0}\chi_{-i}[-2i]\to \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])\langle \alpha \rangle\simeq \oplus_{i=0}^n\chi_{-i}[-2i]
\]
is the projection
(cf. Example~\ref{Sullex}).
By abuse of notation,
we will write $\chi_i$, $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])$ and likes
also for their images in $\Rep(\mathbb{G}_m)$ or $\CAlg(\Rep(\mathbb{G}_m))$.
Note that $F$ sends the $\chi_i$ to $\uni_k(i)$ in $\DM(k)$.
The left adjoint functor
$\CAlg(F):\CAlg(\Rep(\mathbb{G}_m))\to \CAlg(\DM^\otimes(k))$
sends $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n-1}[-2n-2])\to \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])$
to $g$.
Then since $\CAlg(F)$ preserves a pushout,
$\Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])\to \Free_{\Comp(\mathbb{G}_m)}(\chi_{-1}[-2])\langle \alpha \rangle$ maps to the canonical morphism
$\Free(\uni_k(-1)[-2]) \to S_{\PP^n}$.
We see that the composite
\[
\oplus_{i=0}^n\uni_k(-i)[-2i]\hookrightarrow \oplus_{i\ge0}\uni_k(-i)[-2i]\simeq \Free(\uni_k(-1)[-2]) \to S_{\PP^n}\simeq \oplus_{i=0}^n\uni_k(-i)[-2i]
\]
is an equivalence.
Taking account of the first claim of this proof, we see that
the underlying morphism $S_{\PP^n}\to M_{\PP^n}$ in $\DM(k)$ is
an equivalence. Thus, $S_{\PP^n}\to M_{\PP^n}$ in $\DM(k)$ is an equivalence
in $\CAlg(\DM^\otimes(k))$.
\QED
\begin{Remark}
Suppose that the base field $k$ is embedded in $\CC$.
Let $\mathsf{R}:\CAlg(\DM^\otimes(k))\to \CAlg_{\QQ}$
be
the multiplicative realization functor considered in Section~\ref{realization}.
The multiplicative realization functor commutes with
free functors and the formulation of colimits.
Then the above construction of $S_{\PP^n}$ and the equivalence $S_{\PP^n}\simeq M_{\PP^n}$ is compatible with the classical construction of a Sullivan
model of $A_{PL}(\CC \PP^n)$ where $\CC \PP^n$ is the complex projective space.
The morphism $\mathsf{R}(\Free(\uni_k(-1)[-2]))\simeq \Free_{\QQ}(\QQ[-2])\to \mathsf{R}(M_{\PP^n})\simeq A_{PL,\infty}(\CC \PP^n)$ induced by $f$
is determined by a morphism $\QQ[-2]\to A_{PL,\infty}(\CC \PP^n)$
defined by a generator of $H^2(\CC\PP^n,\QQ)=\QQ$. This is
the first step of the construction of a Sullivan model.
The subsequent steps are also compatible. See e.g. \cite{He}.
Also, we remark that $\pi_{i}(\CC\PP^n)\otimes_{\ZZ}\QQ=\QQ$ if $i=2, 2n+1$,
and $\pi_{i}(\CC\PP^n)\otimes_{\ZZ}\QQ=0$ if otherwise.
See also Theorem~\ref{motivichomotopyexp} and Remark~\ref{motcpsR}.
\end{Remark}
\begin{Remark}
The object $M_{\PP^n}$ lies in the full subcategory of mixed Tate motives
in $\DM(k)$. But the above argument works for arbitrary perfect base
fields and does not need a (conjectural) motivic $t$-structure.
\end{Remark}
\subsubsection{}
\label{minus1}
Let $\mathbb{A}^n$ denote the $n$-dimensional affine space over a perfect
field $k$.
Let $X=\mathbb{A}^n-\{p\}$ be the open subscheme
of $\mathbb{A}^n$
that is obtained by removing a $k$-rational point
$p$.
Let $j:X\to \mathbb{A}^n$ be the open
immersion.
By the dual of the Gysin triangle \cite[14.5]{MVW},
we have a distinguished triangle
\[
\uni_k(-n)[-2n]\to M_{\mathbb{A}^n}\stackrel{j^*}{\to} M_{X}
\]
in the triangulated category $\hhh(\DM(k))$. Note that
$M_{\mathbb{A}^n}\simeq \uni_k$ and $\uni_k(-n)[-2n]\to M_{\mathbb{A}^n}$
is null homotopic (see the case in \ref{cps}).
Hence we have an equivalence $M_{X}\simeq \uni_k\oplus \uni_k(-n)[-2n+1]$
in $\DM(k)$.
We let $\Free(\uni_k(-n)[-2n+1])\to M_{X}$ be a morphism in
$\CAlg(\DM^\otimes(k))$, that is classified by the inclusion $\uni_k(-n)[-2n+1]\hookrightarrow \uni_k\oplus \uni_k(-n)[-2n+1]\simeq M_{X}$.
\begin{Proposition}
\label{sphere}
The morphism
$\Free_{\DM(k)}(\uni_k(-n)[-2n+1])\to M_{X}$ is an equivalence.
\end{Proposition}
\Proof
We continue to use the notation in the proof of Proposition~\ref{cpsprop}
and the
colimit-preserving symmetric monoidal functor
$F:\Rep^\otimes(\mathbb{G}_m)\to \DM^\otimes(k)$.
Let $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1])$
be the free algebra that belongs to $\CAlg(\Comp(\mathbb{G}_m))$
(keep in mind that it can be viewed as
a commutative dg algebra endowed with an action of $\mathbb{G}_m$).
Since the generator is concentrated in
the odd degree $2n-1$, by the Koszul sign rule
there is an isomorphism $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1])\simeq \chi_0\oplus \chi_{-n}[-2n+1]$ as objects in $\Comp(\mathbb{G}_m)$.
The functor $F$ carries $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1])$
to $\Free(\uni_k(-n)[-2n+1])$ in $\CAlg(\DM^\otimes(k))$.
Thus, the underlying object of $\Free(\uni_k(-n)[-2n+1])$
is equivalent to $\uni_k\oplus \uni_k(-n)[-2n+1]$.
Moreover, the canonical inclusion (unit map)
$\uni_k(-n)[-2n+1]\to \Free(\uni_k(-n)[-2n+1])$ is compatible with $\uni_k(-n)[-2n+1]\hookrightarrow \uni_k\oplus \uni_k(-n)[-2n+1]$.
Using these facts we deduce that $\Free(\uni_k(-n)[-2n+1])\simeq \uni_k\oplus \uni_k(-n)[-2n+1] \to M_{X}\simeq \uni_k\oplus \uni_k(-n)[-2n+1]$
is an equivalence, as desired.
\QED
\begin{Remark}
Suppose that the base field $k$ is embedded in $\CC$.
Then the complex manifold $X\times_{\Spec k}\Spec \CC$
is homotopy equivalent to the $(2n-1)$-dimensional sphere $S^{2n-1}$.
Proposition~\ref{sphere} is a motivic generalization
of the fact that the free commutative dg algebra generated
by one dimensional vector space placed in (cohomological) degree $2n-1$ is
a Sullivan model of $A_{PL}(S^{2n-1})$ (cf. \cite[Example 1 in page 142]{FHT}).
\end{Remark}
\subsubsection{}
\label{minus2}
Proposition~\ref{cpsprop} and~\ref{sphere}
gives explicit ``models'' $S_{\PP^n}$, $\Free_{\DM(k)}(\uni_k(-n)[-2n+1])$
of motivic cohomological algebras.
The constructions of models have only finitely many steps.
As in the classical rational homotopy theory,
an inductive construction often consists of infinite steps.
The following is such an example.
Let $Y=\mathbb{A}^n-\{p\}-\{q\}$ be the open subscheme
of $\mathbb{A}^n$ that is obtained by removing two $k$-rational points
$p,q$. Let $s:Y\to \Spec k$ denote the structure morphism.
\begin{Proposition}
\label{minus2prop}
Let $A_0=\uni_k$ be the unit algebra in $\CAlg(\DM^\otimes(k))$
and let $A_0=\uni_k \to M_{Y}$ be a unique morphism from the initial
object $\uni_k$ in $\CAlg(\DM^\otimes(k))$.
Then there is a refinement of $A_0\to M_{Y}$
\[
\uni_k=A_0\to A_1\to A_2\to \cdots \to A_i\to A_{i+1} \to \cdots\to M_Y
\]
that satisfies the following properties:
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item
The canonical morphism
$\varinjlim_{i\ge0}A_i\to M_Y$ is an equivalence.
Here $\varinjlim_{i}A_i$ be a colimit of the sequence in
$\CAlg(\DM^\otimes(k))$.
\item Let $V_i$ be the kernel (homotopy fiber) of
$A_i\to M_Y$ in $\DM(k)$ for any $i\ge0$. Then
for each $i\ge0$, $A_i\to A_{i+1}$ is of the form
$A_i\to A_i\otimes_{\Free(V_i)}\uni_k$ given by the pushout of
$A_{i}\leftarrow \Free(V_i) \rightarrow \uni_k$
where $\Free(V_i)\to A_i$ is classified by $V_i\to A_i$.
\end{enumerate}
Moreover, for $n\ge2$, one can explicitly compute each $A_i$
in the sense explained below.
\end{Proposition}
The first statement is a consequence of a
more general fact, see Lemma~\ref{Sullivanmodel1} below.
We explain the second statement, that is,
the procedure of an explicit computation.
We will compute the lower degrees $A_1$, $A_2$, $A_3$.
We can apply the same procedure and arguments also to higher
degrees and we leave it to the interested reader.
We continue to use the notation in Section~\ref{cps}, \ref{minus1}.
As in the case of $X=\mathbb{A}^n-\{p\}$,
applying the dual of Gysin triangle to the open immersion
$Y\hookrightarrow \mathbb{A}^n$, we see that
there is an equivalence
$M_Y \simeq \uni_k\oplus \mathsf{1}(-n)[-2n+1]^{\oplus 2}$
in $\DM(k)$.
The morphism $s^*:\uni_k\to \uni_k\oplus
\mathsf{1}(-n)[-2n+1]^{\oplus 2}\simeq M_Y$
induces an equivalence $\uni_k\stackrel{\sim}{\to} \uni_k\hookrightarrow
\uni_k\oplus
\uni_k(-n)[-2n+1]^{\oplus 2}$. Thus,
$V_0\simeq \uni_k(-n)[-2n]^{\oplus 2}$.
We then find that
\[
A_1=\Free(0)\otimes_{\Free(\uni_k(-n)[-2n]^{\oplus 2})}\Free(0) \simeq \Free(0\sqcup_{\uni_k(-n)[-2n]^{\oplus 2}}0)\simeq \Free(\uni_k(-n)[-2n+1]^{\oplus 2}).
\]
The induced morphism $f:A_1=\Free(0)\otimes_{\Free(\uni_k(-n)[-2n]^{\oplus 2})}\Free(0)\simeq \Free(\mathsf{1}_k(-n)[-2n+1]^{\oplus 2})\to M_Y$ is
classified by the inclusion $\iota:\uni_k(-n)[-2n+1]^{\oplus 2}\hookrightarrow
\uni_k\oplus \uni_k(-n)[-2n+1]^{\oplus 2}\simeq M_Y$.
Let $F:\Rep^\otimes(\mathbb{G}_m)\to \DM^\otimes(k)$
be the colimit-preserving symmetric monoidal functor
which carries $\chi_1$ to $\uni_k(1)$ (cf. the proof of Proposition~\ref{cpsprop}).
Consider $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})$.
The underlying object in $\Comp(\mathbb{G}_m)$
is isomorphic to $\uni_k\oplus \chi_{-n}[-2n+1]^{\oplus 2} \oplus \Sym^2(\chi_{-n}[-2n+1]^{\oplus 2})\simeq \uni_k\oplus \chi_{-n}[-2n+1]^{\oplus 2} \oplus \chi_{-2n}[-4n+2]$. The image of $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})$ under $\CAlg(\Rep^\otimes(\mathbb{G}_m))\to \CAlg(\DM^\otimes(k))$ is equivalent to $A_1$.
The composite
$\uni_k\oplus \uni_k(-n)[-2n+1]^{\oplus 2}\hookrightarrow
\uni_k\oplus \uni_k(-n)[-2n+1]^{\oplus 2}\oplus \uni_k(-2n)[-4n+2]\simeq \Free(\mathsf{1}_k(-n)[-2n+1]^{\oplus 2})\to M_Y$
is an equivalence.
Note that
a morphism $\uni_k(-2n)[-4n+2]\to \uni_k\oplus \uni_k(-n)[-2n+1]^{\oplus 2}$
is null homotopic because it corresponds to
an element in $\Hom_{\hhh(\DM(k))}(\uni_k,\uni_k(n)[2n-1])^{\oplus 2})\oplus \Hom_{\hhh(\DM(k))}(\uni_k,\uni_k(2n)[4n-2])\simeq (\CH^n(\Spec k,1)^{\oplus 2}\oplus \CH^{2n}(\Spec k,2))\otimes_{\ZZ}\QQ=0$ (we use the condition $n\ge2$). Here $CH^i(-,j)$ is the Bloch's higher Chow group. Hence $V_1=\uni_k(-2n)[-4n+2]$
and $V_1\to A_1\simeq \uni_k\oplus \uni_k(-n)[-2n+1]^{\oplus 2}\oplus \uni_k(-2n)[-4n+2]$ may be viewed as the canonical inclusion.
We see that
\[
A_2=\Free(\uni_k(-n)[-2n+1]^{\oplus 2})\oplus_{\Free(\uni_k(-2n)[-4n+2])}\uni_k.\]
Consider $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-2n}[-4n+2])\to \Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})$ classified by
the inclusion $\alpha:\chi_{-2n}[-4n+2]\hookrightarrow \Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})$.
Let $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})\langle \alpha \rangle$ be the homotopy pushout, see Proposition~\ref{hopushout}.
Note that the image of $\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})\langle \alpha \rangle$ in $\CAlg(\DM^\otimes(k))$ (under $F$)
is equivalent to $A_2$.
By the computation using Remark~\ref{explicitdg},
we see that
\[
\Free_{\Comp(\mathbb{G}_m)}(\chi_{-n}[-2n+1]^{\oplus 2})\langle \alpha \rangle\simeq \chi_0\oplus \chi_{-n}[-2n+1]^{\oplus 2}\oplus \chi_{-3n}[-6n+2]^{\oplus 2}\oplus \chi_{-4n}[-8n+3]
\]
in $\Rep(\mathbb{G}_m)$. Hence $A_2\simeq \uni_k\oplus \uni_k(-n)[-2n+1]^{\oplus 2}\oplus \uni_k(-3n)[-6n+2]^{\oplus 2}\oplus \uni_k(-4n)[-8n+3]$.
By the argument similar to the case of $V_1$, we see that $V_2=\uni_k(-3n)[-6n+2]^{\oplus 2}\oplus \uni_k(-4n)[-8n+3]$ and $V_2\to A_2$ may be viewed as the canonical inclusion. We thus find
\[
A_3=A_2\otimes_{\Free(\uni_k(-3n)[-6n+2]^{\oplus 2}\oplus \uni_k(-4n)[-8n+3])}\uni_k.
\]
\subsubsection{}
\label{Sull1}
Let $\mathcal{C}^\otimes$ be a stable presentable
$\infty$-category endowed with a symmetric monoidal
structure whose tensor
operation
$\mathcal{C}\times \mathcal{C} \to \mathcal{C}$
preserves small colimits separately in each variable.
Let $\Free_{\CCC}:\CCC\to \CAlg(\CCC^\otimes)$ be the free functor of $\mathcal{C}^\otimes$.
Let $A$ and $B$ be commutative algebra objects in $\CAlg(\CCC^\otimes)$
and $f:A\to B$ be a morphism in $\CAlg(\CCC^\otimes)$.
Let $V$ be the kernel of $f$ in the stable $\infty$-category $\CCC$,
i.e.,
the pullback $A\times_{B}\{0\}$.
Let
$\sigma:\Free_{\CCC}(V)\to A$ in $\CAlg(\CCC^\otimes)$
be the morphism classified by $V\to A$.
Let $\epsilon:\Free_{\CCC}(V)\to \uni_{\CCC}=\Free_{\CCC}(0)$
be the morphism induced by $V\to 0$ where $\uni_{\CCC}$
is the unit algebra in $\CAlg(\CCC^\otimes)$.
We have a pushout diagram
\[
\xymatrix{
\Free_{\CCC}(V) \ar[r]^\sigma \ar[d]_\epsilon & A \ar[d] \\
\uni_{\CCC} \ar[r] & A(f)
}
\]
in $\CAlg(\CCC^\otimes)$.
Note that the composite $\Free_{\CCC}(V)\to A \stackrel{f}{\to} B$ factors
through $\Free_{\CCC}(V)\to \uni_{\CCC}$.
We have a factorization
\[
A\to A(f)\stackrel{f'}{\to} B
\]
of $f$. Applying this procedure to $f':A(f)\to B$
we obtain a refined factorization $A\to A(f)\to A(f,f'):=A(f)(f')\to B$.
Repeating it in the inductive way we have a sequence in $\CAlg(\CCC^\otimes)_{/B}$
described as
\[
A=A_0\to A_1\to A_2\to \cdots \to A_n\to A_{n+1} \to \cdots
\]
where $A_1=A(f)$, $A_2=A(f,f')\ldots$.
We denote by $f_n:A_n\to B$ the structural morphism.
We shall refer to this sequence as the inductive sequence associated to
$A\to B$.
\begin{Lemma}
\label{Sullivanmodel1}
Let $\varinjlim_{n}A_n$ be a colimit of the sequence in $\CAlg(\CCC^\otimes)$.
Then the canonical morphism
$\varinjlim_{n}A_n \to B$ is an equivalence in $\CAlg(\CCC^\otimes)$.
\end{Lemma}
\Proof
According to \cite[3.2.3.1]{HA},
the forgetful functor $\CAlg(\CCC^\otimes)\to \CCC$
preserves filtered colimits.
Hence it is enough to prove that
a colimit $\varinjlim_{n}A_n$ in $\CCC$ (by abuse of notation
we continue to use the same symbol)
is naturally equivalent to $B$ in $\CCC$.
If $V_n$ denotes the kernel of $f_n:A_n\to B$ in $\CCC$, then
$V_n\to \Free_{\CCC}(V_n)\to A_n\to A_{n+1}$ is null-homotopic.
Thus, $A_n\to A_{n+1}$ factors as composition
$A_n\to \Coker(V_n\to A_n)\to A_{n+1}$ in $\CCC$ where
$\Coker(-)$ stands for cokernel (cofiber/cone) in $\CCC$.
The sequence $A\to A_{1}\to A_2\to\cdots$ in $\CCC$
is refined as
\[
A\to A_1\to \Coker(V_1\to A_1)\to A_2\to \Coker(V_2\to A_2)\to A_3\to\cdots.
\]
By cofinality, the colimit of this sequence is naturally equivalent
to $\varinjlim_{n}A_n$.
Notice that $\Coker(V_n\to A_n)\simeq B$ in $\CCC$
for any $n\ge1$.
Hence we deduce that
$\varinjlim_{n}A_n \to B$ is an equivalence in $\CCC$.
\QED
\begin{Remark}
\label{Sull1R}
Let $\mathcal{D}^\otimes$ be another stable presentable
$\infty$-category endowed with a symmetric monoidal
structure whose tensor
operation
$\mathcal{D}\times \mathcal{D} \to \mathcal{D}$
preserves small colimits separately in each variable.
Let $F:\mathcal{C}^\otimes\to \mathcal{D}^{\otimes}$ be
a symmetric monoidal functor that preserves small colimits.
Our main example of interest is
the realization functor $\mathsf{R}:\DM^\otimes(k) \to \mathsf{D}^\otimes(\QQ)$.
Let $A=A_0\to A_1 \to \cdots\to B$
be the inductive sequence
associated to $f:A\to B$.
Note that $\mathcal{C}\to \mathcal{D}$ is an exact functor of stable $\infty$-categories, and $\CAlg(F):\CAlg(\mathcal{C}^\otimes)\to \CAlg(\mathcal{D}^\otimes)$
preserves small colimits.
Then the sequence $F(A_0)\to F(A_1)\to \cdots \to F(B)$
is canonically equivalent to the inductive sequence
associated to $F(A)\to F(B)$ as a diagram in $\CAlg(\mathcal{D}^{\otimes})_{/F(B)}$.
\end{Remark}
\subsection{}
\label{semiab}
Let $G$ be a semi-abelian variety over $k$.
There is a (canonical) equivalence
\[
M(G)\stackrel{\sim}\to \oplus_{n\ge0}M_n(G)
\]
in $\DM(k)$ such that $M_n(G)=\Sym^n(M_1(G))$.
This is a result of Ancona-EnrightWard-Huber \cite{AEH},
which builds upon the works
of Shermenev, Deniger-Murre and K\"unnemann
on a decomposition of the motives of an abelian variety, see \cite{Kun}
and references therein.
If $G$ is an extension of a $g$-dimesional abelian
variety by a torus of rank $r$,
then $\Sym^n(M_1(G))\simeq 0$ for $n> 2g+r$.
The direct summand $M_1(G)$ is represented,
as an object in $\Comp(\mathcal{N}^{tr}(X))$, by
the \'etale sheaf of $\QQ$-vector spaces
given by $S\mapsto \Hom_{\Sm_k}(S,G)\otimes_{\ZZ}\QQ$
which is promoted to a sheaf with transfers (see
e.g. \cite[Section 2.1]{AEH}).
By using their work, we prove the following:
\begin{Theorem}
\label{abelianvar}
Let $M_1(G)^\vee$ be the dual of $M_1(G)$ in $\DM(k)$ ($M_1(G)$
is a dualizable object).
Let $\Free_{\DM(k)}(M_1(G)^\vee)$ be a free commutative algebra object
in $\DM(k)$ generated by $M_1(G)^\vee$.
Then there is an equivalence
\[
\Free_{\DM(k)}(M_1(G)^\vee) \stackrel{\sim}{\longrightarrow} M_G
\]
in $\CAlg(\DM^\otimes(k))$.
\end{Theorem}
\begin{Remark}
Let $\mathbb{G}$ be a connected compact Lie group.
A theorem of Hopf says that
there are elements $x_1,\ldots,x_n$ of odd degrees in
$H^*(\mathbb{G},\QQ)$ such that
$H^*(\mathbb{G},\QQ)$ is a free commutative graded algebra
generated by $x_1,\ldots,x_n$.
One can deduce from this theorem
that a Sullivan model of $A_{PL}(\mathbb{G})$ is
given by a free commutative
graded algebra generated by some graded vector space, see \cite[Example 3 in
page 143]{FHT}.
Theorem~\ref{abelianvar} may be thought of as a generalization of
this homotopical statement to $\CAlg(\DM^\otimes(k))$ for semi-abelian varieties.
\end{Remark}
\begin{Lemma}
\label{homotopyfree}
Let $\CCC^\otimes$ be a symmetric monoidal presentable
$\infty$-category
whose tensor
operation
$\mathcal{C}\times \mathcal{C} \to \mathcal{C}$
preserves small colimits separately in each variable.
Suppose that $\CCC^\otimes$ is $K$-linear, namely,
it is endowed with a colimit-preserving
symmetric monoidal functor $\Mod_K^\otimes\to \CCC^\otimes$
($K$ is a field of characteristic zero).
Let $\hhh(\CCC)^\otimes$ be the homotopy category of $\CCC$
endowed with a symmetric monoidal structure induced by
that of $\CCC^\otimes$.
The canonical functor $\pi:\CCC\to \hhh(\CCC)$
can be promoted to a symmetric monoidal functor.
Let $\pi':\CAlg(\CCC^\otimes)\to \CAlg(\hhh(\CCC)^\otimes)$
be the ``projection'' induced by the symmetric monoidal functor $\pi$.
In this Lemma we use the temporaty notation
$\Free:=\Free_{\CCC}:\CCC\to \CAlg(\CCC^\otimes)$
be a free functor of $\CCC$.
Let
$\Free^h:=\Free_{\hhh(\CCC)}:\hhh(\CCC)\to \CAlg(\hhh(\CCC)^\otimes)$
be a free functor of $\hhh(\CCC)$.
Let $\theta:\CAlg(\CCC^\otimes)\to \CCC$ and $\theta^h:\CAlg(\hhh(\CCC)^\otimes)\to \hhh(\CCC)$ be forgetful functors.
Let $C$ be an object in $\CCC$.
The unit map $C\to \theta(\Free(C))$ induces
$\pi(C)\to \pi(\theta(\Free(C)))=\theta^h(\pi'(\Free(C)))$.
By the adjunction
$(\Free^h,\theta^h)$, it gives rise to $\sigma:\Free^h(\pi(C))\to \pi'(\Free(C))$.
Then the canonical morphism $\sigma$ is an equivalence.
\end{Lemma}
\Proof
Let $A=\theta^h(\pi'(\Free(C)))$.
The $n$-fold multiplication $A^{\otimes n}\to A$
induces $\Sym_{\hhh(\CCC)}^n (A)\to A$ where $\Sym_{\hhh(\CCC)}^n(-)$ is the $n$-fold symmetric
product in the $K$-linear idempotent complete category $\hhh(\CCC)$.
The map $\pi(C)\to A$ induces $\Sym_{\hhh(\CCC)}^n(\pi(C))\to \Sym_{\hhh(\CCC)}^n(A)\to A$.
Taking its coproduct we have
\[
\tau:\oplus_{n\ge0} \Sym_{\hhh(\CCC)}^n(\pi(C))\to A.
\]
Taking account of the canonical
equivalence $\oplus_{n\ge0} \Sym_{\hhh(\CCC)}^n(\pi(C))\simeq \Free^h(\pi(C))$,
it will suffice to show that $\tau$ is an isomorphism in $\hhh(\CCC)$.
By \cite[3.1.3.13]{HA}, there is an equivalence $\oplus_{n\ge 0}\Sym_{\CCC}^n(C)\to \theta(\Free(C))$ where each $\Sym_{\mathcal{C}}^n(C)\to \theta(\Free(C))$
is induced by the composition of
$C^{\otimes n}\to \Free(C)^{\otimes n}$
and the $n$-fold multiplication $\Free(C)^{\otimes n} \to \Free(C)$.
Here $\Sym_{\CCC}^n(C)$ is the symmetric product in $\CCC$.
Therefore,
it is enough to prove that
the natural morphism $\Sym_{\hhh(\CCC)}^n(\pi(C))\to \pi(\Sym_{\CCC}^n(C))$
is an isomorphism.
Note that for any $D$ in $\CCC$,
the set $\Hom_{\hhh(\CCC)}(\Sym_{\hhh(\CCC)}^n(\pi(C)),\pi(D))$
is the invariant part $\Hom_{\hhh(\CCC)}(\pi(C)^{\otimes n},\pi(D))^{\Sigma_n}$
of $\Hom_{\hhh(\CCC)}(\pi(C)^{\otimes n},\pi(D))$
with the permutation action of the symmetric group $\Sigma_n$.
On the other hand, the hom complex
$\Hom_{\CCC}(\Sym_{\CCC}^n(C),D)$ in $\Mod_K$ is a limit $\Hom_{\CCC}(C^{\otimes n},D)^{\Sigma_n}$ of
$\Hom_{\CCC}(C^{\otimes n},D)$ with permutation action of $\Sigma_n$.
(By definition, the hom complex $\Hom_{\CCC}(C,D)$
is given by the image of $D$ under the right adjoint $\Hom_{\CCC}(C,-)$
to the colimit preserving functor $(-)\otimes C:\Mod_K\to\CCC$.)
Since $K$ is a field of characteristic zero (the semi-simplicity of
representations of finite groups), we have
$H^0(\Hom_{\CCC}(C^{\otimes n},D)^{\Sigma_n})=H^0(\Hom_{\CCC}(C^{\otimes n},D))^{\Sigma_n}=\Hom_{\hhh(\CCC)}(\Sym_{\hhh(\CCC)}^n(\pi(C)),\pi(D))$.
Thus, we see that $\Sym_{\hhh(\CCC)}^n(\pi(C))\to \pi(\Sym_{\CCC}^n(C))$
is an isomorphism.
\QED
\begin{Remark}
\label{homotopyfreeR}
By the proof,
if we define the canonical
functor $\pi:\DM(k)\to \hhh(\DM(k))$,
then we have
a canonical isomorphism $\Sym_{\hhh(\DM(k))}^n(\pi(C))\simeq \pi(\Sym_{\DM(k)}^n(C))$.
Namely, $\pi$ commutes with
the formulation of symmetric products.
By this canonical isomorphism, we often abuse notation
by writing $\Sym^n(C)$
for both $\Sym_{\hhh(\DM(k))}^n(\pi(C))$ and $\Sym_{\DM(k)}^n(C)$.
\end{Remark}
{\it Proof of Theorem~\ref{abelianvar}.}
Let $\alpha_G:M(G)\to M_1(G)$ be the morphism described in \cite[2.1.4]{AEH}
(in {\it loc. cit.},
$\alpha_G$ is a morphism $\DM_{eff}(k)$, but we here regard
it as a morphism in $\DM(k)$).
We remark also that in \cite[2.1.4]{AEH} \'etale motives
are empolyed, but
$\DM^\otimes(k)$ agrees with the \'etale version
since $K$ is a field of characteristic zero, cf. \cite{MVW}, \cite[1.6.1]{AEH}.
Let $\alpha^\vee_G:M_1(G)^\vee\to M(G)^\vee$ be a dual of $\alpha_G$.
Since $M_G=M(G)^\vee$ in $\DM(k)$,
$\alpha_G^\vee$ induces
a morphism $\Free(M_1(G)^\vee)\to M_G$ in
$\CAlg(\DM^\otimes(k))$.
We will prove that it is an equivalence.
To see this, it is enough to show that
$\pi'(\Free(M_1(G)^\vee))\to \pi'(M_G)$ is an isomorphism
where $\pi':\CAlg(\DM^\otimes(k))\to \CAlg(\hhh(\DM(k))^\otimes)$
is the canonical functor (we continue to use the notation
in Lemma~\ref{homotopyfree}).
Lemma~\ref{homotopyfree} guarantees that
$\Free^h(\pi(M_1(G)^\vee))\stackrel{\sim}{\to} \pi'(\Free(M_1(G)^\vee))$.
The composite $\Free^h(\pi(M_1(G)^\vee))\to \pi'(M_G)$ is
induced by $\pi(M_1(G)^\vee)\to \pi(\theta(M_G))=\theta^h(\pi'(M_G))$.
The proof is reduced to showing that this morphism
$\Free^h(\pi(M_1(G)^\vee))\to \pi'(M_G)$ is an isomorphism in $\hhh(\DM(k))$.
The each factor $\phi_n:\Sym^n(\pi(M_1(G)^\vee)\to \pi'(M_G)$ of $\oplus_{n\ge0}\Sym^n(\pi(M_1(G)^\vee) \stackrel{\sim}{\to} \Free^h(\pi(M_1(G)^\vee))\to \pi'(M_G)$
is induced by $\pi(M_1(G)^\vee)^{\otimes n} \stackrel{(\alpha_G^\vee)^{\otimes n}}{\to} \theta^h(\pi'(M_G))^{\otimes n}\to \theta^h(\pi'(M_G))$
where the second morphism is the $n$-fold multiplication.
In the following Lemmata, we will observe that $\phi_n$ is a dual of
the projection $\psi_n:M(G)\to \Sym^n(M_1(G))$ of
the equivalence
$M(G)\stackrel{\sim}{\to} \oplus_{0\le n\le 2g+r}\Sym^n(M_1(G))$
proved in \cite[Theorem 7.1.1]{AEH}. It will finish the proof.
\QED
\begin{Lemma}
\label{algcoalg}
$\phi_n:\Sym^n(\pi(M_1(G)^\vee))\to \theta^h(\pi'(M_G))$
is a dual of $\psi_n:M(G)\to \Sym^n(M_1(G))$.
\end{Lemma}
\Proof
We first recall $\psi_n$.
We work with the homotopy category
$\hhh(\DM(k))$.
By abuse of notation,
we put $M(G)=\pi(M(G))$, $M_1(G)=\pi(M_1(G))$,
$\Sym^n(M_1(G)^\vee)=\Sym^n(\pi(M_1(G))^\vee)$,
$\Sym^n(M_1(G))=\Sym^n(\pi(M_1(G)))$, $M_G=\theta^h(\pi'(M_G))$, etc.
These idetifications are harmless (cf. Lemma~\ref{homotopyfree} and
Remark~\ref{homotopyfreeR}).
The morphism $\psi_n:M(G)\to \Sym^n(M_1(G))$
is the composite
$M(G)\to M(G)^{\otimes n}\stackrel{\alpha_G^{\otimes n}}\to M_1(G)^{\otimes n}\to \Sym^n(M_1(G))$ where the first morphism is the $n$-fold comultiplication
and the the third morphism is the canonical projection.
By ease of notation,
we let $f_\sharp:\hhh(\DM(G))\rightleftarrows \hhh(\DM(k)):f^*$
be the adjoint pair induced by $f_\sharp:\DM(G)\rightleftarrows \DM(k):f^*$
where $f:G\to \Spec k$ is the structure morphism.
The colax monoidal functor $f_\sharp$
induces the coalgebra structure on $M(G)=f_\sharp(\uni_X)$ in $\hhh(\DM(k))$:
the comultiplication is given by the composition
\[
f_\sharp(\uni_G)=f_\sharp(\uni_G \otimes \uni_G)\to f_\sharp(f^*f_\sharp(\uni_G)\otimes f^*f_\sharp(\uni_G))\simeq f_\sharp f^*(f_\sharp(\uni_G)\otimes f_\sharp(\uni_G)) \to f_\sharp(\uni_G)\otimes f_\sharp(\uni_G)
\]
where the left arrow is induced by the counit of $(f_\sharp,f^*)$,
and the right arrow
is induced by the unit. The counit $M(G)\to \uni_k$ is
$f_\sharp f^*(\uni_k)\to \uni_k$.
If one regards $\Sym^n (M_1(G))$ as a direct summand of $M_1(G)^{\otimes n}$,
then $M(G)\to M(G)^{\otimes n}\stackrel{\alpha_G^{\otimes n}}\to M_1(G)^{\otimes n}$ factors as
$M(G)\stackrel{\psi_n}{\to} \Sym^n(M_1(G))\to M(G)^{\otimes n}$.
On the other hand,
$\phi_n:\Sym^n(M_1(G)^{\vee})\to M_G$ is induced by
$\Sigma_n$-equivariant morphism $(M_1(G)^{\vee})^{\otimes n}\stackrel{(\alpha_G^\vee)^{\otimes n}}{\to} M_G^{\otimes n } \to M_G$ where $M_G$ has the trivial
action, and the second arrow is the $n$-fold multiplication.
To prove our assertion of this Lemma, it is enough to show
the following general fact:
\begin{Lemma}
\label{multcomult}
Let $X$ be a smooth scheme separated of finite type over $k$.
The multiplication morphism $M_X\otimes M_X\to M_X$ is a
dual of the comultiplication morphism $M(X)\to M(X)\otimes M(X)$
given by the diagonal
in $\hhh(\DM(k))$ through the isomorphism $M_X\simeq M(X)^\vee$.
(We remark that $M(X)$ is also dualizable in $\DM^\otimes(k)$
since we work with coefficients
of characteristic zero.)
\end{Lemma}
\Proof
We here write $\uni:=\uni_X$ and the structure morphism $f:X\to \Spec k$.
Remember that the multiplication $M_X\otimes M_X\to M_X$
is given by the composition
\[
f_*(\uni)\otimes f_*(\uni)\to f_*f^*(f_*(\uni)\otimes f_*(\uni))\simeq f_*(f^*f_*(\uni)\otimes f^*f_*(\uni))
\to f_*(\uni\otimes \uni)\simeq f_*(\uni)
\]
such that the left arrow is induced by the counit of the adjunction
$(f^*,f_*)$, and the right arrow is induced by its counit.
The canonical isomorphism $\eta:M(X)^\vee \stackrel{\sim}{\to}M_X$
is defined as follows (see the proof of Proposition~\ref{cohmotalg}).
For $M \in \DM(X)$, consider the unit $M\to f^*f_\sharp(M)$.
Taking the dual and $f_*$, we have
$f_*f^*(f_\sharp (M))^\vee \simeq f_*(f^*f_\sharp (M))^\vee\to f_*(M^\vee)$.
Composing with the unit
$(f_\sharp (M))^\vee \to f_*f^*(f_\sharp (M))^\vee$
we obtain $\eta_M:(f_\sharp (M))^\vee\to f_*(M^\vee)$
which determines an isomorphism $\eta=\eta_{\uni_X}:M(X)^\vee \stackrel{\sim}{\to}M_X$.
We will check that
the dual of $f_\sharp(\uni)\to f_\sharp(\uni)\otimes f_\sharp(\uni)$ is $f_*(\uni)\otimes f_*(\uni)\to f_*(\uni)$
through $\eta:f_\sharp (\uni)^\vee\simeq f_*(\uni)$.
By using the counit of $(f^*,f_*)$ and its counit-unit equations,
we see that
the dual $f^*f_\sharp(M)^\vee\to M^\vee$ of $M\to f^*f_\sharp(M)$
is $f^*(f_\sharp(M)^\vee)\stackrel{f^*(\eta_M)}{\to} f^*f_*(M^\vee)\to M^\vee$
where the final arrow is the counit of $(f^*,f_*)$.
When $M=\uni$, we deduce that the unit $s:\uni\to f^*f_\sharp(\uni)$ is the dual of the counit $t:f^*f_*(\uni)\to \uni$
through the isomorphism $\eta:f_\sharp(\uni)^\vee\simeq f_*(\uni)$. It follows that its tensor product
$s\otimes s:\uni\otimes \uni \to f^*f_\sharp(\uni)\otimes f^*f_\sharp(\uni)$
is the dual of $t\otimes t:f^*f_*(\uni)\otimes f^*f_*(\uni)\to \uni\otimes \uni$
through the isomorphism through the isomorphism $\eta:f_\sharp(\uni)^\vee\simeq f_*(\uni)$.
Thus the triangle in the following diagram commutes.
\[
\xymatrix{
(f_\sharp(\uni)\otimes f_\sharp(\uni))^\vee \ar[r]^(0.45){a} \ar[d]_{\simeq} & (f_\sharp f^*(f_\sharp(\uni)\otimes f_\sharp(\uni)))^\vee \ar[r]^(0.7)b \ar[d]_{\eta_{f^*(f_\sharp(\uni)\otimes f_\sharp(\uni))}} & f_\sharp(\uni)^\vee \ar[dd]^{\eta} \\
f_\sharp(\uni)^\vee\otimes f_\sharp(\uni)^\vee \ar[r]^(0.45){c} \ar[d]_{\eta\otimes \eta} & f_*f^*(f_\sharp(\uni)^\vee\otimes f_\sharp(\uni)^\vee) \ar[rd]^{f_*(s^\vee \otimes s^\vee)} \ar[d]_{f_*f^*(\eta\otimes \eta)} & \\
f_*(\uni)\otimes f_*(\uni) \ar[r]^(0.45){d} & f_*f^*(f_*(\uni)\otimes f_*(\uni)) \ar[r]_(0.7){f_*(t\otimes t)} & f_*(\uni)
}
\]
Here $a$ is induced by
the dual of $f_\sharp f^*\to \textup{id}$, and $b$ is induced by the
dual of $s\otimes s:\uni \otimes \uni \to f^*f_\sharp (\uni)\otimes f^*f_\sharp(\uni)=f^*(f_\sharp(\uni)\otimes f_\sharp(\uni))$.
Note that the composite of the upper horizontal arrows is the dual of comultiplication
$M(X)\to M(X)\otimes M(X)$.
Both $c$ and $d$ is induced by the unit $\textup{id}\to f_*f^*$.
The composite of lower horizontal arrows is the multiplication
$M_X\otimes M_X\to M_X$.
The commutativity of other squares follow from
the contravariant functoriality of $\eta_M$ with respect to $M$,
the functoriality/naturality of $\textup{id}\to f_*f^*$,
and the counit-unit equations for the adjunction $(f_\sharp,f^*)$.
Thus, we have a commutativity of the outer square, which completes the proof.
\QED
\begin{Remark}
The unit map $\uni_k\to M_X$ is nothing but a dual
of the morphism $M(X)\to M(\Spec k)=\uni_k$ induced by $f$.
\end{Remark}
\subsection{}
\subsubsection{}
\label{curveexp}
We consider a once-punctured smooth proper curve,
that is, $C=\overline{C}-\{p\}$ obtained by removing
a $k$-rational point $p$ from
a connected smooth proper curve $\overline{C}$
over the perfect field $k$.
Let $j:C\to \overline{C}$
be the open immersion.
The genus of $\overline{C}$ is $g\ge1$.
If $k=\CC$, the fundamental
group of the underlying topological space
is the free group generated by $2g$ elements.
Let $M(\overline{C})\simeq M_0(\overline{C})\oplus M_1(\overline{C})\oplus M_{2}(\overline{C})$
be a (Chow-K\"unneth) decomposition of $M(\overline{C})$
such that $M_0(\overline{C})\simeq \uni_k$
and $M_2(\overline{C})\simeq \uni_k(1)[2]$ (see the first paragraph
of the proof of Lemma~\ref{lemmasqzero} for the precise
formulation).
(In this case, $M(C)$ is equivalent to
$M_0(\overline{C})\oplus M_1(\overline{C})$ as an object in $\DM(k)$.)
We put $M^i_{\overline{C}}:=M_i(\overline{C})^\vee$
so that $M_{\overline{C}}\simeq \oplus_{i=0}^2M^i_{\overline{C}}$.
Let
\[
A_0=\uni_k\to A_1\to A_2\to \cdots \to A_n\to A_{n+1}\to \cdots
\]
be the inductive sequence in $\CAlg(\DM^\otimes(k))_{/M_C}$
associated to the unique morphism
$\uni_k \to M_C$ in $\CAlg(\DM^\otimes(k))$ (cf. Section~\ref{Sull1}).
By Lemma~\ref{Sullivanmodel1}, the colimit $\varinjlim_{n\ge0}A_n\simeq M_C$.
\begin{Theorem}
\label{curve3}
The first three terms $A_1$, $A_2$ and $A_3$ are computed as follows:
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item $A_1$ is $\Free(M^1_{\overline{C}})$, and $f_1:A_1\to M_C$ is classified by $M^1_{\overline{C}} \hookrightarrow M^0_{\overline{C}}\oplus M^1_{\overline{C}}\oplus M^2_{\overline{C}}\simeq M_{\overline{C}}\stackrel{j^*}{\to} M_C$.
\item $f_1:A_1\to M_C$ is the composite $M_{\Alb_{\overline{C}}}\stackrel{u^*}{\to} M_{\overline{C}}\stackrel{j^*}{\to} M_C$ up
to an equivalence $\Free(M^1_{\overline{C}})\simeq M_{\Alb_{\overline{C}}}$, where $u:\overline{C}\to \Alb_{\overline{C}}$ is the Albanese (Abel-Jacobi) morphism into the Albanese
variety, which carries $p$ to the origin.
\item Let $W_1$ be $\oplus_{i=2}^{2g}\Sym^i(M_{\overline{C}}^1)$.
Let $\Free(W_1)\to A_1$ be the morphism in $\CAlg(\DM^\otimes(k))$
that is classified by the inclusion
$W_1\to A_1\simeq \oplus_{i\ge0}\Sym^i(M_{\overline{C}}^1)$
in $\DM(k)$. Then $A_2$ is the pushout $A_1\otimes_{\Free(W_1)}\uni_k$,
\item Let $W_2$ be the object in $\DM(k)$
which will be defined just before the proof (Section~\ref{curve3proof}).
Then $A_3$ has the form of the pushout $A_2\otimes_{\Free(W_2)}\uni_k$.
\end{enumerate}
\end{Theorem}
\begin{Remark}
The symmetric product $\Sym^N(M^1_{\overline{C}})$ is zero for $N>2g$
(see the proof of Lemma~\ref{lemmasqzero}).
Thus $A_1=\oplus_{i\ge0}\Sym^i(M^1_{\overline{C}})=\oplus_{i=0}^{2g}\Sym^i(M^1_{\overline{C}})$.
\end{Remark}
\begin{Remark}
As mentioned in Introduction,
the sequence $\{A_n\}_{n\ge0}$ can be
viewed as a step-by-step description of the ``non-abelian structure'' of $C$.
To give a feeling for this, let us make the following observation.
Suppose that $c$ is a $k$-rational point on $C$, and let
$M_C\to \uni_k$ be the induced augmentation. The sequence
$\{A_n\}_{n\ge0}$ is promoted to a sequence
to $\CAlg(\DM^\otimes(k))_{/\uni_k}$ in the obvious way.
By applying the construction in Section~\ref{hopfsection} to $A_n\to \uni_k$,
we obtain the sequence
of cogroup objects in $\CAlg(\DM^\otimes(k))$, which we denote
by $\{\uni_k\otimes_{A_n}\uni_k\}_{n\ge0}$ (we abuse notation
since $\uni_k\otimes_{A_n}\uni_k$
is the underlying object).
Now consider the ``topological aspect'' of this sequence.
For this purpose,
suppose further that $k\subset \CC$. Let $R_n\to \QQ$ be the image of $A_n\to \uni_k$
under the singular realization functor. Let $R\to \QQ$ be the image of $M_C\to \uni_k$.
Then
$\{R_n\}_{n\ge0}$ is the inductive sequence associated to $\QQ\to R$ (cf.
Remark~\ref{Sull1R}), and
the image of $\{\uni_k\otimes_{A_n}\uni_k\}_{n\ge0}$ under the realization
is $\{\QQ\otimes_{R_n}\QQ\}_{n\ge0}$ (we abuse notation again).
Taking the $0$-th cohomology, we have the sequence of the pro-unipotent
algebraic groups
\[
\cdots \to \Spec H^0(\QQ\otimes_{R_n}\QQ)\to \cdots \to \Spec H^0(\QQ\otimes_{R_1}\QQ)\to \Spec H^0(\QQ\otimes_{R_0}\QQ)\simeq \Spec \QQ.
\]
Define $G_n:=\Spec H^0(\QQ\otimes_{R_n}\QQ)$.
In this case, by Theorem~\ref{curve3} (i), $G_1$ is a commutative
unipotent group of rank $2g$ (in fact, $\Free(M_{\overline{C}}^1[1])$ maps to
$\QQ\otimes_{R_1}\QQ\simeq \Free_{\QQ}(\QQ^{\oplus 2g})\simeq H^0(\QQ\otimes_{R_1}\QQ)$).
Recall that $G:=\Spec H^0(\QQ\otimes_{R}\QQ)$ is the pro-unipotent completion
$\pi_1(C^t,c)$ of $C^t$ (cf. Section~\ref{MGA}).
By a standard argument in rational homotopy theory,
each morphism $G_{n+1}\to G_n$ is a surjective morphism
with a commutative kernel, and the canonical morphism
$\pi_1(C^t,c)_{uni}\simeq G\to \varprojlim_{n\ge0}G_n$ is an isomorphism of pro-unipotent algebraic groups.
\end{Remark}
\begin{Example}
Let $\overline{C}$ be an elliptic curve
and let $C=\overline{C}-\{0\}$ be the open curve obtained by removing
the origin $0$.
Then by Theorem~\ref{curve3}, one can easily see that
$A_1=\Free(M^1_{\overline{C}})$, $A_2=\Free(M^1_{\overline{C}})\otimes_{\Free(\uni_k(-1)[-2])}\uni_k$, and $A_2$ is equivalent to $\uni_k\oplus M^1_{\overline{C}}\oplus M^1_{\overline{C}}(-1)[-1]\oplus \uni_k(-2)[-3]$
as an object in $\DM(k)$.
We have $W_2=M^1_{\overline{C}}(-1)[-1]\oplus \uni_k(-2)[-3]$,
and the third term $A_3$ is of the form
$A_2\otimes_{\Free(W_2)}\uni_k$.
\end{Example}
\subsubsection{}
\begin{Lemma}
\label{lemmasqzero}
The multiplication map $M_C\otimes M_C\to M_C$ in the homotopy category
$\hhh(\DM(k))$ is
\[
m:(\uni_k\oplus M^1_{\overline{C}})^{\otimes 2} \simeq \uni_k\otimes \uni_k\oplus \uni_k\otimes M^1_{\overline{C}}\oplus M^1_{\overline{C}}\otimes \uni_k\oplus (M^1_{\overline{C}})^{\otimes 2} \to \uni_k\oplus M^1_{\overline{C}}
\]
defined as a coproduct
of $m|_{(M^1_{\overline{C}})^{\otimes2}}=0$
and ``identities''
$\uni_k\otimes \uni_k\to \uni_k$, $\uni_k\otimes M^1_{\overline{C}}\to M^1_{\overline{C}}$,
$M^1_{\overline{C}}\otimes \uni_k\to M^1_{\overline{C}}$.
Namely, $M_C$ is the trivial square zero extension of $\uni_k$
by $M_{\overline{C}}^1$
in $\hhh(\DM(k))$.
\end{Lemma}
\begin{Remark}
The unit $\uni_k\to M_C$ may be identified with the
morphism $\uni_k=M_{\Spec k}\to M_{C}$
determined by the structure morphism $C\to \Spec k$.
By the construction of the decomposition (see below),
it is the inclusion $\uni_k\hookrightarrow \uni_k\oplus M^1_{\overline{C}}$. Thus the non-trivial part of the Lemma is $m|_{(M^1_{\overline{C}})^{\otimes2}}=0$.
\end{Remark}
In the proof of the above Lemma,
we
discuss decompositions of motives
and use the category $\Chow_k$ of Chow motives
with rational coefficients, cf. \cite[Section 1]{Sc}, \cite[Section 2.2]{MNP}.
We choose the contravariant Chow motives since we will refer to \cite{Sc}
and \cite{MNP} in which the authors adopt the contravariant formulation.
But $\DM(k)$ is a covariant theory in the sense that there is
the canonical covariant functor $\Sm_k\to \DM(k)$ given by $X\mapsto M(X)$
while $\Chow_k$ has a contravariant functor $\textup{SmPr}_k\to \Chow_k$
given by $X\mapsto \Ch(X)$. Here
$\textup{SmPr}_k$ is the category of connected
smooth projective varieties over $k$,
and
following \cite{MNP} we denote by $\Ch(X)$ the Chow motive of
$X$ (that is $h(X)$ in \cite{Sc}). The relation between $\DM(k)$
and $\Chow_k$ is quite well-known, but the difference between
the covariant and the contravariant formulations is likely
to cause unnecessary confusion.
We thus give some remarks.
There is a fully faithful $\QQ$-linear functor
$\Chow_k^{op}\to \hhh(\DM(k))$ that is symmetric monoidal,
\cite[20.1, 20.2]{MVW}, \cite[9.3.6]{MNP}.
It carries $\Ch(X)$ to $M(X)$.
The Lefschetz motive $\mathbb{L}$ maps to $\uni_k(1)[2]$.
As the level of hom sets,
\begin{eqnarray*}
\Hom_{\Chow_k}(\Ch(Y),\Ch(X))=\CH^{d}(Y\times X) &\stackrel{\textup{transpose}}{\to}& \CH^{d}(X\times Y) \\
&\simeq& \Hom_{\hhh(\DM(k))}(M(X\times Y),\uni_k(d)[2d]) \\
&\simeq& \Hom_{\hhh(\DM(k))}(M(X),M(Y)^\vee\otimes \uni_k(d)[2d]) \\
&\simeq& \Hom_{\hhh(\DM(k))}(M(X),M(Y))
\end{eqnarray*}
where $d$ and $e$ are the dimensions of $Y$ and $X$, respectively.
For $f:X\to Y$, we write $f^*:\Ch(Y)\to \Ch(X)$ for the class of the transposed graph
${}^t\Gamma_f$ in $\CH^{d}(Y\times X)$. We also use $f_*:\Ch(X)\to \Ch(Y)\otimes \mathbb{L}^{\otimes e-d}$ that
corresponds to the class of $\Gamma_f$ in $\CH^{e+(d-e)}(X\times Y)$.
The functor $\Chow_k^{op}\to \hhh(\DM(k))$ carries
$f^*:\Ch(Y)\to \Ch(X)$ to $M(X)\to M(Y)$ induced by the graph of $f$,
that is the dual of $f^*:M_Y\to M_X$ (the final $f^*$ is defined in Section~\ref{cohomologicalmot}).
{\it Proof of Lemma~\ref{lemmasqzero}.}
For ease of notation, $X=\overline{C}$.
We first recall the decomposition $\Ch^0(X)\oplus \Ch^1(X)\oplus \Ch^2(X)\simeq \Ch(X)$. We define the retract $\Ch(\Spec k)\stackrel{s^*}{\to} \Ch(X)\stackrel{p^*}{\to} \Ch(\Spec k)$ given by $p:\Spec k=\{p\}\to X$ and the structure morphism $s:X\to \Spec k$.
There is also the retract defined by
$\mathbb{L} \stackrel{p_*\otimes \mathbb{L}}{\to} \Ch(X)\stackrel{s_*}{\to} \mathbb{L}$. The components $\Ch^0(X)=\uni_k$
and $\Ch^2(X)$ arise from the first retract and the second retract, respectively. Here we abuse notation by writing $\uni_k$ for the unit object in $\Chow_k$
because it corresponds to the unit object in $\DM(k)$.
The component $\Ch^1(X)$ can be described as the Picard motives
in the sense of J.P. Murre \cite[Section 6.2]{MNP}, \cite[Section 4]{Sc}.
Let $\Alb_X$ be the Albanese variety of $X$ and let $\Pic_X$ be
the Picard variety of $X$.
Note that
\[
\CH^1(X\times X)\supset \Hom^*((X,p),(\Pic_X,0))\simeq \Hom_{AV}(\Alb_X,\Pic_X)
\]
where $\Hom^*$ indicates the set of morphisms that preserve base points,
and $\Hom_{AV}$ indicates the set of morphisms of abelian varieties.
Here we implicitly use the Albanese morphism $(X,p)\to (\Alb_X,0)$.
The set $\Hom^*((X,p),(\Pic_X,0))$ corresponds to
the subgroup of $\CH^1(X\times X)$,
that consists of those classes of divisors
$D\in \CH^1(X\times X)$ such that $(\textup{id}_X\times p)^*(D)=0$
and
$(p\times \textup{id}_X)^*(D)=0$ in $\CH^1(X)$.
We will call such divisors $p$-normalized divisors
and denote by $\CH^1_{(p)}(X\times X)$ the subgroup of $p$-normalized divisors.
Consider the isomorphism $\theta:\Alb_X\stackrel{\sim}{\to} \Pic_X$
defined by the theta divisor.
By \cite[Lemma 6.2.6]{MNP}
the element $\pi_1$ in $\Hom_{\Chow_k}(\Ch(X),\Ch(X))=\CH^1(X\times X)\otimes_{\ZZ}\QQ$
corresponding to $\theta$ is an idempotent morphism of $\Ch(X)$.
We define $\Ch^{1}(X)$ to be the object corresponding to $\pi_1$,
namely, $\Ch^{1}(X)\to \Ch(X)$ is $\Ker(\textup{id}-\pi_1)\to \Ch(X)$.
Let $M_0(X)\oplus M_1(X)\oplus M_2(X)\simeq M(X)$
be the decomposition that arises from the decomposition
of $\Ch(X)\simeq \Ch^1(X)\oplus \Ch^1(X)\oplus \Ch^2(X)$.
Put $M^i_X=M_i(X)^\vee$ and let
$M^0_X\oplus M^1_X\oplus M^2_X\simeq M_X$ be the decomposition
obtained by taking the dual. We remark that $M^0_X\simeq \uni_k$
and $M^2_X\simeq \uni_k(-1)[-2]$.
Next we construct a Picard motive $\Ch^1(\Alb_X)$ by using the
Albanese (Abel-Jacobi) map $u:X\to \Alb_X$ which carries $p$ to the origin.
Consider the isomorphisms
$\Pic_{\Alb_X}\stackrel{\sim}{\to} \Pic_X\stackrel{\theta}{\leftarrow}\Alb_X\stackrel{\sim}{\to} \Alb_{\Alb_X}$.
The third morphism induced by the functoriality
is an isomorphism because of the universal property of Albanese vatieties,
and the first morphism is its dual.
Let $\sigma:\Alb_{\Alb_X}\to \Pic_{\Alb_X}$
be the inverse of the composite.
If we denote by $\CH^1_{(0)}(\Alb_X\times \Alb_X)$ the subgroup
of $0$-normalized divisors ($0$ is the origin),
we have the canonical isomorphisms $\Hom_{AV}(\Alb_{\Alb_X},\Pic_{\Alb_X})\simeq \Hom_{AV}(\Alb_X,\Pic_{\Alb_X})\simeq \CH^1_{(0)}(\Alb_X\times \Alb_X)$.
Let $Z$ be the divisor that corresponds to $\sigma$
and let $\phi:\Ch(\Alb_X)\otimes \mathbb{L}^{\otimes 1-g}\to \Ch(\Alb_X)$
be the morphism defined by $Z\in \CH^1(\Alb_X\times\Alb_X)$.
Let $\omega_1:\Ch(\Alb_X)\to \Ch(\Alb_X)$ be the composite
\[
\Ch(\Alb_X)\stackrel{u^*}{\to} \Ch(X)\stackrel{u_*}{\to} \Ch(\Alb_X)\otimes \mathbb{L}^{\otimes 1-g}\stackrel{\phi}{\to} \Ch(\Alb_X).
\]
We can apply the proof of \cite[Lemma 6.2.6]{MNP}
and see that $\omega_1$ is an idempotent map.
We define $\Ch^1(\Alb_X)$ to be $\Ker(\textup{id}-\omega_1)$.
We now claim that the composite
\[
\Ch^1(\Alb_X)\to \Ch(\Alb_X)\stackrel{u^*}{\to} \Ch(X)\to \Ch^1(X)
\]
is an isomorphism where the first arrow is the canonical monomorphism
and the final arrow is the ``projection''.
As observed in \cite[proof of Lemma 6.2.6]{MNP},
the equality $(E):\phi\circ u_*\circ u^* \circ \phi \circ u_*=\phi\circ u_*$
holds (indeed, $\omega_1\circ \omega_1=\omega_1$ is a direct consequence
of $(E)$).
Thus, $u^*\circ \phi\circ u_*\circ u^* \circ \phi \circ u_*=u^*\circ \phi\circ u_*$. Namely, $u^*\circ \phi\circ u_*:\Ch(X) \to \Ch(X)$ is an idempotent
morphism. The morphism $\pi_1$ coincides with $u^*\circ \phi\circ u_*$.
Actually, again by the observation in \cite[proof of Lemma 6.2.6]{MNP},
$u^*\circ \phi\circ u_*$ corresponds to the composite
$\Alb_X\stackrel{\sim}{\to} \Alb_{\Alb_X}\stackrel{\sigma}{\to} \Pic_{\Alb_X}\stackrel{\sim}{\to} \Pic_X$, that is $\theta$, through
$\CH^1(X\times X)\supset \Hom^*((X,p),(\Pic_X,0))\simeq \Hom_{AV}(\Alb_X,\Pic_X)$.
Let $F:=(u^*\circ \phi\circ u_*)\circ u^*\circ (\phi \circ u_* \circ u^*)$
and $G:=(\phi \circ u_* \circ u^*)\circ \phi\circ u_*\circ (u^*\circ \phi\circ u_*)$. To prove our claim,
it will suffice to show $F\circ G=u^*\circ \phi \circ u_*$
and $G\circ F=\phi\circ u_*\circ u^*$. These equalities follow from
$(E)$. We also see that $\phi\circ u_*$
induces $\Ch^1(X)\to\Ch(X)\stackrel{\phi\circ u_*}{\to} \Ch(\Alb_X)\to \Ch^{1}(\Alb_X)$ is an isomorphism.
Let $\Free_{\Chow_k}(\Ch^1(\Alb_X))=\oplus_{n\ge0}\Sym^n(\Ch^1(\Alb_X))$
be the free commutative algebra object in $\Chow_k$ and
let $h:\Free_{\Chow_k}(\Ch^1(\Alb_X))\to \Ch(\Alb_X)$ be the morphism
of commutative algebra objects that is classified by $\Ch^1(\Alb_X)\to \Ch(\Alb_X)$. Here $\Ch(\Alb_X)$ admits the commutative algebra structure defined by
$\Ch(\Spec k)\to \Ch(\Alb_X)$ induced by the structure morphism
and $\Ch(X)\otimes \Ch(X)\to \Ch(X)$ induced by the diagonal.
We will show that $h$ is an isomorphism.
Let $R_l:\Chow_k\to \textup{GrVect}$ be the (symmetric monoidal)
$l$-adic realization functor to the category of $\ZZ$-graded $\QQ_l$-vector
space (the symmetric monoidal structure on $\textup{GrVect}$
adopts the Koszul rule). For a projective smooth variety $U$,
it carries $\Ch(U)$ to the $\ZZ$-graded $\QQ_l$-vector space $\textup{H}^*_{\textup{\'et}}(U\times_{k}\overline{k},\QQ_l)$ of \'etale cohomology
($\overline{k}$ is a separable closure).
Then $\textup{H}^*_{\textup{\'et}}(\Alb_{X}\times_{k}\overline{k},\QQ_l)$
is the free commutative graded algebra generated by
$\textup{H}^1_{\textup{\'et}}(\Alb_{X}\times_{k}\overline{k},\QQ_l)$ placed in
degree one.
By \cite[Theorem 6.2.1]{MNP}, $R_l(\Ch^1(\Alb_X))$ is $\textup{H}^1_{\textup{\'et}}(\Alb_X\times_{k}\overline{k},\QQ_l)$ placed in
degree one, and $R_l(\Ch^1(\Alb_X)\to \Ch(\Alb_X))$ is
$\textup{H}^1_{\textup{\'et}}(\Alb_X\times_{k}\overline{k},\QQ_l)\hookrightarrow
\textup{H}^*_{\textup{\'et}}(\Alb_X\times_{k}\overline{k},\QQ_l)$.
We then conclude that $R_l(h)$ is an isomorphism.
Since $\Sym^{N}(\Ch^1(\Alb_X))=\Sym^N(\Ch^1(X))=0$ for $N>2g$ (see e.g. \cite{Kun}),
$\Free_{\Chow_k}(\Ch^1(\Alb_X))=\oplus_{n=0}^{2g}\Sym^n(\Ch^1(\Alb_X))$.
Both $\Ch(\Alb_X)$ and $\Free_{\Chow_k}(\Ch^1(\Alb_X))$
are Kimura finite (see e.g. \cite{MNP} for this notion).
Thanks to Andr\'e-Kahn \cite[Proposition 1.4.4.(b), Theorem 9.2.2]{AK} (explained also in \cite[Theorem 1.3.1]{AEH}),
we deduce from the isomorphism $R_l(h)$ that $h$ is an isomorphism.
(We remark that $h$ is not necessarily compatible with the equivalence
in Theorem~\ref{abelianvar}.)
Next consider the composition
\[
\psi:\Free_{\Chow_k}(\Ch^1(\Alb_1))=\oplus_{i=0}^{2g}\Sym^{i}(\Ch^1(\Alb_X))
\stackrel{h}{\simeq}\Ch(\Alb_{X})\stackrel{u^*}{\to} \Ch(X)\stackrel{\pi_1}{\to} \Ch^1(X).
\]
We will
show that for $i\ne1$, $\Sym^{i}(\Ch^1(\Alb_X))\hookrightarrow \Ch(\Alb_{X})\stackrel{\psi}{\to} \Ch^1(X)$ is zero. Let $\omega_i:\Ch(\Alb_X)\to \Sym^{i}(\Ch^1(\Alb_X))\to \Ch(\Alb_X)$ denote the idempotent map arising from the direct summand
$\Sym^{i}(\Ch^1(\Alb_X))$. Note that $\omega_1:\Ch(\Alb_X)\to \Ch(\Alb_X)$
equals to $\Ch(\Alb_X)\stackrel{u^*}{\to}\Ch(X)\stackrel{\pi_1}{\to}\Ch^1(X)\hookrightarrow \Ch(X)\stackrel{\phi\circ u_*}{\to} \Ch(\Alb_X)$.
Indeed, the equality $(E)$ implies that $\phi\circ u_*\circ \pi_1\circ u^*=\phi \circ u_* \circ u^*=\omega_1$.
Suppose that $\Sym^{i}(\Ch(\Alb_X))\to \Ch^1(X)$ induced by $\psi$ is not
zero. It follows that $w_1\circ w_i$ is not zero
because $\phi\circ u_*$ induces the isomorphism $\Ch^1(X)\simeq \Ch^1(\Alb_X)\subset\Ch(\Alb_X)$.
For $i\ne 1$, this contradicts the orthogonality $w_1\circ w_i=0$.
Hence $\Sym^{i}(\Ch(\Alb_X))\to \Ch^1(X)$ is zero for $i\ne 1$.
Remember that $\Ch(X)$ has a commutative algebra structure (in $\Chow_k$)
that is defined by the structure morphism and the diagonal in the
same way as $\Ch(\Alb_X)$. In addition, $u^*$ is
a homomorphism of commutative algebras.
The homomorphism $u^*$ induces an isomorphism $\Ch^1(\Alb_X)\simeq \Ch^1(X)$.
Taking account of the compatibility of multiplications,
we see that the multiplication $\Ch(X)\otimes \Ch(X)\to \Ch(X)\simeq \Ch^1(X)\oplus \Ch^1(X)\oplus \Ch^2(X)$
sends $\Ch^1(X)\otimes \Ch^1(X)$ to the direct summand $\Ch^2(X)\subset \Ch(X)$.
Namely, the composition $\Ch^1(X)\otimes \Ch^1(X)\hookrightarrow \Ch(X)\otimes \Ch(X)\to \Ch(X)\to \Ch^1(X)$ is zero (by the compatibility of the unit maps
the projection to $\Ch^0(X)$ is also zero).
Now move to $\hhh(\DM(k))$.
The commutative algebra $\Ch(X)$ corresponds to the cocommutative coalgebra
$M(X)$ whose coalgebra structure is determined by the structure morphism
and the diagonal. Take the dual of $M(X)$, that is, $M_X$ in $\hhh(\DM(k))$.
According to Lemma~\ref{multcomult},
the algebra structure of $M_X$ (in $\hhh(\DM(k))$)
given by the coalgebra structure of $M(X)$ coincides with that of
$M_X$ induced by the cohomological motivic algebra.
We have proved that the multiplication $M_X\otimes M_X\to M_X$
sends the component $M_X^1\otimes M_X^1$ to the direct summand $M^2_X$.
If $j:C\to X$ denote the open immersion, then we have
the dual of the Gysin distinguished triangle \cite[14.5]{MVW}
\[
\uni_k(-1)[-2] \stackrel{\eta}{\to} M_X\stackrel{j^*}{\to} M_C\to
\]
in $\hhh(\DM(k))$. By the exact sequence
$\CH^0(\Spec k)\stackrel{p_*}{\to} \CH^1(X)\stackrel{j^*}{\to} \CH^1(C)\to 0$,
the composite
$\uni_k(-1)[-2]\stackrel{e}{\to}M_X\stackrel{j^*}{\to} M_C$ is zero
where $e$ is the dual of the morphism
$M(X)\to \uni_k(1)[2]$ corresponding to the class of $p$ in $\CH^1(X)$.
Since $e$ is non-trivial and $\End_{\hhh(\DM(k))}(\uni_k(-1)[-2])=\QQ$, thus we may suppose that $e=\eta$.
Consequently, $\eta$ is the canonical inclusion
$\uni_k(-1)[-2]\simeq M^2_X\hookrightarrow M^0_X\oplus M_X^1\oplus M_X^2$.
It follows that
$j^*$ is identified with the projection $M_X\simeq M^0_X\oplus M_X^1\oplus M_X^2\to M_X^0\oplus M_X^1\simeq M_C$ (with respect this decomposition).
Therefore, the multiplication $M_C\otimes M_C\to M_C$ sends
the component $M_X^1\otimes M_X^1\subset M_C\otimes M_C$
to zero.
\QED
\begin{Remark}
One can ask
whether or not $M_C$ is a trivial square zero extension of $\uni_k$
by some motive $M$ in $\DM(k)$
(not only at the level of $\hhh(\DM(k))$).
It would be an interesting problem.
We refer to \cite[7.3.4]{HA} for the notion of trivial square zero extensions
in $\infty$-categorical setting.
If $M_C$ is the trivial square zero extension
at the level of $\DM(k)$,
it should be regarded as {\it formality} of $M_C$.
Suppose that we are given a connected affine smooth
curve $C$ over $\CC$.
Write $C^t$ for the underlying topological space of $C$.
Then $A_{PL}(C^t)$ is equivalent to
the square zero extension $\QQ\oplus H^1(C^t,\QQ)[-1]$
of $\QQ=H^0(C^t,\QQ)$ by $H^1(C^t,\QQ)[-1]$ in $\CAlg_{\QQ}$.
(Namely, $A_{PL}(C^t)$ is formal.)
The problem about formality of $M_C$ makes sense for arbitrary
(geometrically connected) affine smooth curves.
\end{Remark}
\subsubsection{}
\label{curve3proof}
Before the proof of Theorem~\ref{curve3}, we will define $W_2$.
Let $K$ be the standard representation of $\GL_{2g}$, that is,
the $2g$-dimensional vector space $V$ endowed with the canonical
action of $\Aut(V)=\GL_{2g}$.
We usually consider $K$ to be the complex
concentrated in degree zero, that belongs to either $\Comp(\GL_2)$
or $\Rep(\GL_2)$.
Let $\Free_{\Comp(\GL_2)}(K[-1])$ is the free commutative algebra object
in $\Comp(\GL_{2g})$, that is isomorphic to
$\oplus_{i=0}^{2g}\Sym^i(K[-1])$ as an object of $\Comp(\GL_{2g})$.
Put $U_1=\oplus_{i=2}^{2g}\Sym^i(K[-1])$ and consider the inclusion
$\alpha:U_1\hookrightarrow \oplus_{i=0}^{2g}\Sym^i(K[-1])$.
Let $\phi_{\alpha}:\Free_{\Comp(\GL_{2g})}(U_1)\to \Free_{\Comp(\GL_{2g})}(K[-1])$ be the morphism classified by $\alpha$
and let $\Free_{\Comp(\GL_{2g})}(K[-1])\langle \alpha\rangle$
be the homotopy pushout (cf. Proposition~\ref{hopushout}).
Consider
$\Free_{\Comp(\GL_{2g})}(K[-1])\langle \alpha\rangle$
as an object in $\Comp(\GL_{2g})$. Then by the explicit presentation
in Remark~\ref{explicitdg}, we find that
its $0$-th cohomogy is the unit, and the first cohomology is $K$.
Thus, $\Free_{\Comp(\GL_{2g})}(K[-1])\langle \alpha\rangle$
is equivalent to $\uni\oplus K[-1]\oplus U_2$ in $\Rep(\GL_{2g})$
where $\uni$ is a unit object in $\Rep(\GL_{2g})$, and $U_2$ is concentrated
in the degrees larger than one.
(We remark that in practice one can compute $U_2$ explicitly
by means of the representation theory of $\GL_{2g}$.)
Note that the wedge product $\wedge^{N}(M^1_{\overline{C}}[1])=\Sym^N(M^1_{\overline{C}})[N]$
is zero exactly when $N>2g$
because $M^1_{\overline{C}}$ is equivalent to
the dual of the direct summand $M_1(\Alb_{\overline{C}})$
arising from $\Ch^1(\Alb_{\overline{C}})$
(see the proof of Lemma~\ref{lemmasqzero}),
and $\Sym^i(M_1(\Alb_{\overline{C}})^\vee)=0$
for $N>2g$, see e.g. \cite{Kun} for this vanishing.
By \cite[Theorem 3.1, Proposition 6.1]{DTD},
there is a
colimit-preserving symmetric monoidal functor $F:\Rep^\otimes(\GL_{2g})\to \DM^\otimes(k)$ which carries $K$ to $M^1_{\overline{C}}[1]$.
Indeed, $F$ is a unique $\QQ$-linear symmetric monoidal functor having
this property (see \cite{DTD} for the detail of the formulation).
We define $W_2$ to be $F(U_2)$.
{\it Proof of Theorem~\ref{curve3}.}
We first prove (1). For simplicity, we put $X=\overline{C}$.
Taking account of the construction of the
decomposition $\Ch(C)\simeq \Ch^0(X)\oplus\Ch^1(X)$ in the proof of Lemma~\ref{lemmasqzero}, $\uni_k=M_{\Spec k}\to M_C$ induced by the
structure morphism $C\to \Spec k$ is
identified with the inclusion $\uni_k=M^0_{X}\hookrightarrow M^0_{X}\oplus M^1_{X}$.
Note that the unit algebra $\uni_k$ is an initial object. Thus,
the kernel of $A_0=\uni_k\to M_{C}$ in $\DM(k)$ is $M^1_{X}[-1]$,
that is, $V_0=M^1_{X}[-1]$.
Therefore $A_1=\uni_k\otimes_{\Free(M^1_{X}[-1])}\uni_k\simeq \Free(0\sqcup_{M^1_{X}[-1]}0)\simeq \Free(M^1_{X})$ ($\sqcup$ indicates the pushout).
The composite $M_X\to \Free(M^1_{X})\to M_C$
is equivalent to the inclusion
$M^1_X\simeq 0\sqcup_{M^1_{X}[-1]}0\to \uni_k\sqcup_{M^1_{X}[-1]}0\simeq \uni_k\oplus M^1_X$ where the second arrow is induced by $0\to \uni_k$.
Next we prove (2).
Let $M(\Alb_X)\to M_1(\Alb_X)$ be the morphism
arising from $\Ch^1(\Alb_X)\to \Ch(\Alb_X)$ (see the second paragraph of
the proof of Lemma~\ref{lemmasqzero}).
If one takes its dual
$M^1_{\Alb_{X}}=M_1(\Alb_X)^\vee\to M_{\Alb_X}$, then
by Lemma~\ref{homotopyfree},
$\Free(M^1_{\Alb_{X}})\to M_{\Alb_X}$ classified by it is an equivalence
in $\CAlg(\DM^\otimes(k))$.
By the isomorphism $\Ch^1(\Alb_X)\simeq \Ch^1(X)$ in the third
paragraph of the proof of Lemma~\ref{lemmasqzero},
the composite $\Free(M^1_{\Alb_{X}})\to M_{\Alb_X}\stackrel{u^*}{\to} M_{X}\stackrel{j^*}{\to} M_C$ induces an equivalence
$M^1_{\Alb_X}\to \Free(M^1_{\Alb_{X}})\to M_C\to M^1_X$.
Also, $M^1_{\Alb_X}\to M_C\to M^0_X$ is null homotopic.
Consider $\Free(M^1_{X})\simeq \Free(M^1_{\Alb_X})$ induced by
$\Ch^1(X)\simeq \Ch^1(\Alb_X)$. Then
$\Free(M^1_{X})\simeq \Free(M^1_{\Alb_X})\simeq M_{\Alb_X}\stackrel{j^*u^*}{\to} M_C$ is equivalent to $A_1\to M_C$.
Next we prove (3).
Let $V_1$ is the kernel of $A_1=\Free(M^1_{X})\to M_C$.
Then $M^1_{X} \to\Free(M^1_{X})\to M_C\simeq \uni_k\oplus M^1_{X}$
may be viewed as the inclusion.
In addition, Lemma~\ref{lemmasqzero} shows that
$M_C\otimes M_C\to M_C$ kills $M^1_X\otimes M^1_X$.
Thus, taking account of the commutative algebra
structure of $\Free(M^1_{X})$ in $\hhh(\DM(k))$
we deduce that $\Free(M^1_{X})=\oplus_{i=0}^{2g}\Sym^i(M^1_X)\to M_C\simeq \uni_k\oplus M^1_{X}$ can be identified with the projection.
Hence $V_1\to \Free(M^1_{X})$ is $\oplus_{i=2}^{2g}\Sym^i(M^1_X)\hookrightarrow \oplus_{i=0}^{2g}\Sym^i(M^1_X)$.
Let $\Free(V_1)\to \Free(M^1_{X})$ is the morphism classified by $V_1\to\Free(M^1_{X})$. Thus $A_2=\Free(M^1_{X})\otimes_{\Free(V_1)}\uni_k$.
Next we prove that (4).
Note that we already defined an ``explicit'' model of $A_2$ before this proof.
Namely, $A_2$ is equivalent to the image of
$\Free_{\Comp(\GL_{2g})}(K[-1])\langle \alpha\rangle$
under $\CAlg(F):\CAlg(\Rep^\otimes(\GL_{2g}))\to \CAlg(\DM^\otimes(k))$.
Thus $A_2\simeq \uni_k\oplus M^1_{X}\oplus W_2$.
Moreover, using the sequence $A_1\to A_2\to M_C$
we find that the composite $r:\uni_k\oplus M^1_{X}\hookrightarrow \uni_k\oplus M^1_{X}\oplus W_2\simeq A_2\to M_C\simeq \uni_k\oplus M^1_{X}$ is an equivalence.
Put $h:W_2\hookrightarrow\uni_k\oplus M^1_{X}\oplus W_2\simeq A_2\to M_C\simeq \uni_k\oplus M^1_{X}$.
Then $H=(-r^{-1}\circ h) \oplus \textup{id}_{W_2}:W_2\to (\uni_k\oplus M^1_{X})\oplus W_2\simeq A_2$
is the kernel of $A_2\to M_C$ (we expect that $h$ is zero).
Let $\Free(W_2)\to A_2$ be the morphism classified by $H$.
Then $A_3=A_2\otimes_{\Free(W_2)}\uni_k$.
\QED
\section{Cotangent complexes and homotopy groups}
\label{cotangenthomotopy}
We introduce a cotangent motive of a pointed
(smooth) scheme $(X,x)$.
Under a suitable condition,
the dual of rationalized homotopy groups will appear as the realization of the cotangent
motive. The notion of cotagent motives is inspired by
Sullivan's description of homotopy groups
in terms of the space of indecomposable elements of a minimal Sullivan model.
We may think of cotangent motive as motives
of (dual of) rationalized homotopy groups.
In this Section, the coefficient ring of $\DM(k)$ is $\QQ$.
\subsection{}
Let $(X,x:\Spec k\to X)$ be a pointed smooth scheme
over $k$. It gives rise to an augmented
object $x^*:M_X\to \uni_k=M_{\Spec k}$.
We will define an object of $\DM(k)$ by means
of cotangent complexes for $\CAlg(\DM^\otimes(k))$.
For this purpose, we use the theory of cotangent complexes
for presentable $\infty$-categories, developed in \cite[Section 7.3]{HA}.
This theory is a vast generalization of cotangent complexes
(topological Andr\'e-Quillen homology) for $E_\infty$-algebras.
Let us briefly recall some definitions about cotangent complexes for
the reader's convenience.
Let $\CCC$ be a presentable $\infty$-category
and let $A$ be an object in $\CCC$.
Let $\SP(\CCC_{/A})$ be the stabilization (stable envelope) of $\CCC_{/A}$ (cf. \cite[1,4]{HA}).
Let $(\CCC_{/A})_{*}$ denote the $\infty$-category of pointed objects
of $\CCC_{/A}$: one may take $(\CCC_{/A})_{*}=(\CCC_{/A})_{A/}$.
Then $\SP(\CCC_{/A})$ is defined to be the limit of the sequence of $\infty$-categories
\[
\cdots \stackrel{\Omega_*}{\to} (\CCC_{/A})_{*} \stackrel{\Omega_*}{\to} (\CCC_{/A})_{*} \stackrel{\Omega_*}{\to} (\CCC_{/A})_{*},
\]
where $\Omega_*$ is informally given by $S\mapsto *\times_{S}*$
($*$ is a final object). The stable $\infty$-category $\SP(\CCC_{/A})$ is also
presentable.
Another presentation of $\SP(\CCC_{/A})$
is the $\infty$-category of spectrum objects of $\CCC$, see \cite[1.4.2]{HA}.
There is a canonical functor $\Omega^{\infty}:\SP(\CCC_{/A})\to (\CCC_{/A})_*\to \CCC_{/A}$ where the first arrow
is the projection to $(\CCC_{/A})_*$ placed in the right end in the above sequence, and the second arrow is the forgetful functor.
Let $\Sigma_{+}^{\infty}:\CCC_{/A} \to \SP(\CCC_{/A})$ be a left adjoint to
$\Omega^\infty$, whose existence is ensured by adjoint functor theorem
since $\Omega^\infty$ preserves small limits and is accessible.
An absolute cotangent complex $L_{A}$ of $A$ is defined to be $\Sigma^{\infty}_+(A\stackrel{\textup{id}}{\to} A)$.
If $A$ is an initial object, then $L_A$ is a zero object.
We now take $\CCC$ to be $\CAlg(\DM^\otimes(k))$.
By \cite[7.3.4.13]{HA}, there is a canonical equivalence
$\SP(\CAlg(\DM^\otimes(k))_{/A})\simeq \Mod_{A}(\DM(k))$
of $\infty$-categories. Here $\Mod_{A}(\DM(k))$ denotes
the $\infty$-category of $A$-module objects in $\DM(k)$.
We refer to \cite[3.3.3, 4.5]{HA} for the notion of module
objects over a commutative algebra object.
We have the adjunction
\[
\Sigma_+^{\infty}:\CAlg(\DM^\otimes(k))_{/A}\rightleftarrows \SP(\CAlg(\DM^\otimes(k))_{/A})\simeq \Mod_{A}(\DM(k)):\Omega^{\infty}.
\]
We regard $L_A=\Sigma_+^{\infty}(A\stackrel{\textup{id}}{\to} A)$ as an object of $\Mod_{A}(\DM(k))$.
Let $\phi:A\to B$ be a morphism in $\CAlg(\DM^\otimes(k))$.
Let $(-)\otimes_AB:\Mod_{A}(\DM(k))\to \Mod_B(\DM(k))$ denote
a left adjoint to the forgetful functor $\Mod_B(\DM(k))\to\Mod_{A}(\DM(k))$
induced by $A\to B$. Then as in the classical theory of cotangent complexes,
there is a canonical morphism $L_A\otimes_{A}B \to L_B$;
indeed, $L_A\otimes_{A}B\simeq \Sigma_+^{\infty}(A \stackrel{\phi}{\to} B)$
when $A \to B$ is thought of as an object of $\CAlg(\DM^\otimes(k))_{/B}$
(see \cite[7.3.2.14,
7.3.3, 7.3.4.18]{HA} and Remark~\ref{basefree}).
We define the relative cotangent complex $L_{B/A}$ of $A\to B$ to be
a cokernel (cofiber) of $L_A\otimes_{A} B \to L_B$ in $\Mod_B(\DM(k))$.
\begin{Definition}
\label{cotangentmotivedef}
Let $(X,x)$ be a pointed smooth scheme separated of finite type over $k$.
Let $x^*:M_X\to \uni_k=M_{\Spec k}$ be the morphism induced by $x$.
We define $LM_{(X,x)}$ to be
$L_{M_X}\otimes_{M_X}\uni_k$ in $\DM(k)$.
Here $L_{M_X}$ belongs to $\Mod_{M_X}(\DM(k))$, and $(-)\otimes_{M_X}\uni_k:\Mod_{M_X}(\DM(k))\to \Mod_{\uni_k}(\DM(k))\simeq \DM(k)$
is induced by $x^*$.
We shall refer to $LM_{(X,x)}$
as the {\it cotangent motive} of $X$ at $x$.
For $i\in \ZZ$
and $j\in\ZZ$,
we define $\prod_{i,j}(X,x):=\Hom_{\hhh(\DM(k))}(LM_{(X,x)},\uni_k(-j)[-i])$.
\end{Definition}
\begin{Remark}
There is a canonical equivalence $LM_{(X,x)}=L_{M_X}\otimes_{M_X}\uni_k\simeq L_{\uni_k/M_X}[-1]$.
Indeed, there is the distinguished triangle (cofiber sequence) arising from
$\uni_k\to M_X\to \uni_k$:
\[
L_{M_X}\otimes_{M_X}\uni_k\to L_{\uni_k/\uni_k}\to L_{\uni_k/M_X}\to
\]
in the homotopy category of $\DM(k)$, see \cite[7.3.3.5]{HA}.
In addition, $L_{\uni_k/\uni_k}\simeq 0$.
It follows that
$L_{M_X}\otimes_{M_X}\uni_k\simeq L_{\uni_k/M_X}[-1]$.
\end{Remark}
\begin{Remark}
The definition of the cotangent motives makes sense also when
we work with an arbitrary coeffiecient ring $K$ of $\DM(k)$.
\end{Remark}
The main result of this Section is the following:
\begin{Theorem}
\label{motivichomotopy}
Let $X$ be a smooth variety over $k$ and let $x$ be a $k$-rational point.
Suppose that $k$ is embedded in $\CC$
and the underlying topological space
$X^t$ of $X\times_{\Spec k}\Spec \CC$ is simply connected.
Then the (singular) realization functor $\mathsf{R}:\DM(k)\to \mathsf{D}(\QQ)$ carries
$LM_{(X,x)}$ to $\oplus_{2\le i }(\pi_i(X^t,x)\otimes_{\ZZ}\QQ)^{\vee}[-i]$
in $\mathsf{D}(\QQ)\simeq \Mod_{\QQ}$.
Namely,
there is an isomorphism
\[
H^i(\mathsf{R}(LM_{(X,x)}))\simeq (\pi_i(X^t,x)\otimes_{\ZZ}\QQ)^{\vee}
\]
for $i\ge 2$, and $H^i(\mathsf{R}(LM_{(X,x)}))=0$ for $i<2$.
Here $(\pi_i(X^t,x)\otimes_{\ZZ}\QQ)^{\vee}$
is the dual $\QQ$-vector space of $\pi_i(X^t,x)\otimes_{\ZZ}\QQ$.
\end{Theorem}
\begin{Remark}
Under the hypothesis of Theorem~\ref{motivichomotopy},
the cohomology $H^i(X^t,\QQ)$ is finite dimensional for any $i\ge 0$.
Indeed, the simply connectedness is not necessary for this finiteness.
In general, if $S$ is a simply connected topological space
whose cohomoogy $H^i(S,\QQ)$ is finite dimensional for any $i\ge0$, then
$\pi_{i}(S,s)\otimes_{\ZZ}\QQ$ is finite dimensional for any $i\ge2$.
\end{Remark}
\subsection{}
The proof proceeds in several steps.
\begin{Lemma}
\label{realmorita}
Let $\mathsf{R}_E:\DM^\otimes(k)\to \mathsf{D}^\otimes(K)$ be the symmetric monoidal realization functor associated to mixed Weil theory $E$ with coefficients in a field $K$ of characteristic zero.
Let $G$ be a right adjoint to $\mathsf{R}_E$, that is lax symmeytric monoidal.
Let $G(K)$ be the commutative algebra object (i.e., an object of $\CAlg(\DM^\otimes(k))$) where $K$ is the unit algebra in $\mathsf{D}(K)$.
Consider the composition of symmetric monoidal functors
\[
\Mod_{G(K)}^{\otimes}(\DM(k))\to \Mod_{\mathsf{R}_E(G(K))}^\otimes(\mathsf{D}(K))\to \Mod_{K}^\otimes(\mathsf{D}(K)) \simeq \mathsf{D}^\otimes(K)
\]
where the first arrow is induced by $\mathsf{R}_E$, and
the second arrow is given by the base change
$(-)\otimes_{\mathsf{R}_E(G(K))}K$ induced by the counit map
$\mathsf{R}_E(G(K))\to K$.
Then the composite is an equivalence, and $\mathsf{R}_E$
is
equivalent to the base change functor
$(-)\otimes_{\uni_k}G(K):\DM^\otimes(k)\to \Mod_{G(K)}^{\otimes}(\DM(k))\simeq \mathsf{D}^\otimes(K)$.
\end{Lemma}
\Proof
If we verify two conditions
\begin{itemize}
\item there is a set $\{M_{\lambda}\}_{\lambda\in \Lambda}$ of compact and
dualizable objects of $\DM(k)$ such that
the whole category $\DM(k)$ is the smallest
stable subcategory which contains $\{M_{\lambda}\}_{\lambda\in \Lambda}$
and is closed under small coproducts
(that is to say, $\{M_{\lambda}\}_{\lambda\in \Lambda}$ is a generator
of $\DM(k)$),
\item each $\mathsf{R}_E(M_{\lambda})$ is compact, and there is some
$\mu\in I$ such that $\mathsf{R}_E(M_{\mu})\simeq K$,
\end{itemize}
then our assertion follows from \cite[Proposition 2.1]{DTD}.
For $X\in \Sm_k$ and $n\in \ZZ$, $M(X)(n)$ is compact in $\DM(k)$,
and the set $\{M(X)(n)\}_{X\in \Sm_k,n\in \ZZ}$ is a generator of $\DM(k)$.
In addition, $M(X)$ is dualizable
because it holds if $X$ is projective,
and
we work with rational coefficients, so that we can use
the standard argument based on de Jong's
alteration (or one can directly apply a very general result in
\cite[4.4.3, 4.4.17]{CDT}).
Since $\mathsf{R}_E$ is symmetric monoidal and $M(X)(n)$
is dualizable, $\mathsf{R}_E(M(X)(n))$ is also dualizable.
In $\mathsf{D}(K)$, an object is compact if and only if it is dualizable.
Finally, $\mathsf{R}_E(M(\Spec k))=\mathsf{R}_E(\uni_k)\simeq K$
since $\mathsf{R}_E$ is symmetric monoidal.
Hence the above two conditions are satisfied.
\QED
According to Lemma~\ref{realmorita}, under the setting of Theorem~\ref{motivichomotopy},
we write $P:=G(\QQ)$ and identify the (singular) realization functor
$\mathsf{R}$ with $(-)\otimes_{\uni_k}P:\DM^\otimes(k)\to \Mod_P^\otimes(\DM(k))\simeq \mathsf{D}^\otimes(\QQ)$.
The multiplicative realization functor
$\CAlg(\mathsf{R}):\CAlg(\DM^\otimes(k))\to \CAlg_{\QQ}$
can naturally be identified with
\[
\CAlg(\DM^\otimes(k))\longrightarrow \CAlg(\Mod_P^\otimes(\DM(k))\simeq \CAlg(\DM^\otimes(k))_{P/}
\]
which sends $A$ to $P\simeq \uni_k\otimes P \to A\otimes P$.
For the right equivalence, see \cite[3.4.1.7]{HA}.
We focus on cotangent complexes of commutative dg algebras,
that is, objects of $\CAlg_{\QQ}$.
Let $C$ be an object of $\CAlg_{\QQ}$.
If we take $\CCC$ to be $\CAlg_{\QQ}$ in the above formalism of cotangent
complexes, we have the adjunction
\[
\Sigma_+^{\infty}:(\CAlg_{\QQ})_{/C}\rightleftarrows \SP((\CAlg_{\QQ})_{/C})\simeq \Mod_{C}(\mathsf{D}(\QQ)):\Omega^{\infty}.
\]
Here we abuse notation by using $\Sigma_+^{\infty}, \Omega^{\infty}$ again.
We define the absolute
cotangent complex
$L_C$ of $C$ to be $\Sigma_+^{\infty}(C\stackrel{\textup{id}}{\to} C)$.
Given a morphism $C\to D$ we define $L_{D/C}$ to be a cokernel of
$D\otimes_CL_C\to L_D$ in $\Mod_{D}(\mathsf{D}(\QQ))$.
\begin{Remark}
\label{basefree}
Let $\mathcal{C}$ be either $\CAlg(\DM^\otimes(k))$ or
$\CAlg_{\QQ}=\CAlg(\mathsf{D}^\otimes(\QQ))$.
More generally, $\mathcal{C}$ could be a presentable
$\infty$-category $\CAlg(\DDD^\otimes)$
such that $\DDD^\otimes$ is a symmetric monoidal stable
presentable $\infty$-category
whose tensor product $\DDD\times \DDD\to \DDD$ preserves small colimits
separately in each variable.
Let $A$ and $B$ be objects of $\CCC$. Let $B\to A$ be a morphism.
Consider the adjunction
\[
\Sigma_+^{\infty}:\CCC_{/A} \rightleftarrows \SP(\CCC_{/A})\simeq \Mod_A(\DDD):\Omega^{\infty}.
\]
If we regard $B\to A$ as an object of $\CCC_{/A}$, then
$\Sigma_+^{\infty}$ sends $B\to A$ to $L_{B}\otimes_BA$,
where $(-)\otimes_BA:\Mod_B(\DDD)\to \Mod_A(\DDD)$ denotes the base change
functor. It is a direct consequence of a functorial construction
of cotangent complexes by using the notion
of a tangent bundle in \cite[7.3.2.14]{HA}
and a presentation of the tangent bundle by a presentable fibration
of module
categories \cite[7.3.4.18]{HA}.
Suppose that $A$ is an initial object (that is, a unit algebra).
The above adjunction is extended to
\[
\DDD \rightleftarrows \CCC_{/A} \rightleftarrows \SP(\CCC_{/A})\simeq \DDD
\]
where the left arrow
$j:\CCC_{/A}=\CAlg(\DDD^\otimes)_{/A} \to \DDD$
is the functor which carries $\epsilon:B\to A$ to the kernel (fiber) $\Ker(\epsilon)$
of $B\to A$ in $\DDD$. The left adjoint $\DDD\to \CCC_{/A}$ to $j$
sends $M\in \DDD$ to
$\Free_{\DDD}(M)\to \Free_{\DDD}(0)\simeq A$ determined by $M\to 0$,
where $\Free_{\DDD}:\DDD\to \CCC$ is the free functor,
see Definition~\ref{freealg}.
By the construction of $\SP(\CCC_{/A})\simeq \DDD$ (cf. \cite[7.3.4.13]{HA}),
the composite $\DDD\simeq \SP(\CCC_{/A}) \stackrel{\Omega^\infty}{\to} \CCC_{/A}\stackrel{j}{\to} \DDD$ is naturally equivalent to the identity functor.
Thus $\Sigma^{\infty}_+$ carries $\Free_{\DDD}(M)\to A$
to $M$. Namely, $L_{\Free_{\DDD}(M)}\otimes_{\Free_{\DDD}(M)}A\simeq M$.
\end{Remark}
\begin{Remark}
If one considers $x^*:M_X\to \uni_k$ to be an object
of $\CAlg(\DM(k))_{/\uni_k}$, then
its image under $\Sigma^\infty_+:\CAlg(\DM(k))_{/\uni_k}\to \DM(k)$
is $LM_{(X,x)}$
(cf. Remark~\ref{basefree}). The right adjoint $\Omega^\infty:\DM(k)\to \CAlg(\DM(k))_{/\uni_k}$
sends $LM_{(X,x)}$ to a square zero extension of $\uni_k$ by $LM_{(X,x)}$,
which is informally given by $\uni_k\oplus LM_{(X,x)}\stackrel{\textup{pr}_1}{\to} \uni_k$, see \cite[7.3.4]{HA} for square zero extensions.
By the adjunction, we have the unit map
$u:M_X\to \uni_k\oplus LM_{(X,x)}$ in $\CAlg(\DM(k))_{/\uni_k}$.
Let $\overline{M}_X$ be the kernel (fiber) of $M_X\to \uni_k$ in $\DM(k)$.
It gives rise to a morphism in $\DM(k)$
\[
h:\overline{M}_X\to LM_{(X,x)}
\]
induced by $u$.
This morphism is a motivic version of dual Hurewicz map.
\end{Remark}
\begin{Lemma}
\label{nicebasechange}
Let $\epsilon:A\to \uni_k$ be a morphism in $\CAlg(\DM^\otimes(k))$, that is,
an augmented commutative algebra object in $\DM(k)$.
Let $B:=\mathsf{R}(A)\to \mathsf{R}(\uni_k)= \QQ$ be a
the image of
$\epsilon$ in $\CAlg_{\QQ}$ under the multiplicative realization functor.
Let $L_{B}$ be the (absolute) cotangent complex of $B$
and let $L_B\otimes_{B}\QQ$ be the base change that lies in $\mathsf{D}(\QQ)$.
Then there is a canonical equivalence
\[
\mathsf{R}(L_{A}\otimes_{A}\uni_k)\simeq L_B\otimes_{B}\QQ
\]
in $\mathsf{D}(\QQ)$.
\end{Lemma}
\Proof
As explained above, Lemma~\ref{realmorita}
allows us to identify the multiplicative realization functor with
$\CAlg(\DM^\otimes(k))\to \CAlg(\DM^\otimes(k))_{P/}\simeq \CAlg_{\QQ}$.
Then we have the pushout diagram
\[
\xymatrix{
\uni_k \ar[r] \ar[d] & P \ar[d] \\
A \ar[r] & A\otimes P
}
\]
in $\CAlg(\DM^\otimes(k))$, and
$B$ corresponds to the right vertical arrow $P \to A\otimes P$
which we regard as an object of $\CAlg(\DM^\otimes(k))_{P/}$.
By \cite[7.3.3.8, 7.3.3.15]{HA},
the absolute cotangent complex of $P \to A\otimes P$ regarded as an object
of $\CAlg(\DM^\otimes(k))_{P/}$ is equivalent to the relative cotangent complex
$L_{A\otimes P/P}$ of the morphism
$P\to A\otimes P$ in $\CAlg(\DM(k))$.
It follows that $L_{A\otimes P/P}\simeq L_B$ under the canonical
equivalence $\Mod_{A\otimes P}(\DM(k))\simeq \Mod_{B}(\mathsf{D}(\QQ))$.
The final equivalence is induced by
\begin{eqnarray*}
\Mod_{A\otimes P}(\DM(k)) &\simeq& \SP(\CAlg(\DM^\otimes(k))_{/A\otimes P}) \\
&\simeq& \SP((\CAlg(\DM^\otimes(k))_{P/})_{/A\otimes P}) \\
&\simeq& \SP((\CAlg_{\QQ})_{/B}) \\
&\simeq& \Mod_{B}(\mathsf{D}(\QQ))
\end{eqnarray*}
where the first and final equivalences follow from \cite[7.3.4.13]{HA},
and the second one follows from \cite[.3.3.9]{HA}.
Since $A\otimes P\to \uni_k\otimes P\simeq P$ corresponds to
$B\to \QQ$, we see that $L_{B}\otimes_{B}\QQ$ corresponds to
$L_{A\otimes P/P}\otimes_{A\otimes P}P$ in $\Mod_{P}(\DM(k))\simeq \mathsf{D}(\QQ)$.
By the base change formula for cotangent complexes \cite[7.3.3.7]{HA},
$L_{A\otimes P/P}\simeq L_{A}\otimes_{A}(A\otimes P)$.
Therefore, we obtain
\[
L_{A\otimes P/P}\otimes_{A\otimes P}P\simeq L_{A}\otimes_{A}(A\otimes P)\otimes_{A\otimes P}P\simeq (L_A\otimes_A\uni_k)\otimes P.
\]
Note that
$\mathsf{R}(L_{A}\otimes_{A}\uni_k)\simeq (L_A\otimes_A\uni_k)\otimes P$
in $\Mod_P(\DM(k))$. Hence our assertion follows.
\QED
The following is a theorem of Sullivan \cite[Section 8]{S},
reformulated
in terms of cotangent complexes.
\begin{Lemma}
\label{cotangSul}
Let $(S,s)$ be a simply connected topological space $S$
with a point $s$.
Assume that
the cohomology $H^i(S,\QQ)$ is a
finite dimensional $\QQ$-vector space for any $i\ge0$.
Let $A_{PL,\infty}(S)$ be the image of $A_{PL}(S)$
in $\CAlg_{\QQ}$ (see Section~\ref{realization}).
Let $A_{PL,\infty}(S)\to \QQ$ be the augmentation induced by $s$.
Then $L_{A_{PL,\infty}(S)}\otimes_{A_{PL,\infty}(S)}\QQ
\simeq \oplus_{2\le i }(\pi_i(S,s)\otimes_{\ZZ}\QQ)^{\vee}[-i]$ in $\mathsf{D}(\QQ)$.
\end{Lemma}
\Proof
For ease of notation, we may assume that
$S$ is a rational space, so that $\pi_i(S,s)$ is a $\QQ$-vector space
for each $i\ge 2$.
Consider a Postnikov tower
\[
S=S_{\infty} \to \cdots \to S_n\to S_{n-1}\to \cdots \to S_2\to S_1.
\]
We first show our assertion in the case of $S_{n}$.
The case of $n=1$ is trivial because $S_1$ is contractible
and $L_{A_{PL,\infty}(S_1)}\simeq 0$. We suppose that
our assertion holds for $S_{n-1}$.
Consider the diagram
\[
\xymatrix{
K(\pi_{n}(S,s),n) \ar[r] \ar[d] & S_n \ar[d] \\
\ast \ar[r] & S_{n-1}
}
\]
where $\ast$ is a contractible space, $K(\pi_{n}(S,s),n)$
is an Eilenberg-MacLane space, and we here think of the diagram with
a pullback square in $\SSS$.
By a computation for the Eilenberg-MacLane space
\cite[Section 15 Example 3, Section 12, Example 2]{FHT},
$A_{PL,\infty}(K(\pi_{n}(S,s),n))\simeq \Free_{\QQ}(\pi_{n}(S,s)^\vee[-n])$
where $\pi_{n}(S,s)^\vee[-n]$ is the dual $\QQ$-vector space placed in cohomological
degree $n$, that we consider to be an object of $\mathsf{D}(\QQ)$,
and
$\Free_{\QQ}:\mathsf{D}(\QQ)\to \CAlg_{\QQ}$ is the free functor, see Definition~\ref{freealg}.
By \cite[Theorem 15.3]{FHT},
$A_{PL,\infty}(K(\pi_{n}(S,s),n))$ is a pushout of $A_{PL,\infty}(S_{n})\leftarrow A_{PL,\infty}(S_{n-1})\to \QQ\simeq A_{PL,\infty}(\ast)$ in $\CAlg_{\QQ}$
(the result found in \cite{FHT} shows that it is a homotopy pushout in
$\CAlg_{\QQ}^{dg}$).
When $S_{n}$ and $S_{n-1}$ are equipped with (compatible) base points,
$A_{PL,\infty}(K(\pi_{n}(S,s),n))$ is promoted to a pushout in $(\CAlg_{\QQ})_{/\QQ}$. Note that $\Sigma_+^\infty:(\CAlg_{\QQ})_{/\QQ}\to \SP((\CAlg_{\QQ})_{/\QQ})\simeq \mathsf{D}(\QQ)$ preserves small colimits.
Taking account of Remark~\ref{basefree} we have a pushout diagram
\[
\xymatrix{
L_{A_{PL,\infty}(S_{n-1})}\otimes_{A_{PL,\infty}(S_{n-1})}\QQ \ar[r] \ar[d] & L_{A_{PL,\infty}(S_{n})}\otimes_{A_{PL,\infty}(S_{n})}\QQ \ar[d] \\
0 \ar[r] & \pi_{n}(S,s)^\vee[-n]
}
\]
in $\mathsf{D}(\QQ)$.
By the assumption, $L_{A_{PL,\infty}(S_{n-1})}\otimes_{A_{PL,\infty}(S_{n-1})}\QQ\simeq \oplus_{2\le i \le n-1}\pi_i(S,s)^{\vee}[-i]$.
Then $L_{A_{PL,\infty}(S_{n})}\otimes_{A_{PL,\infty}(S_{n})}\QQ$ is a cokernel (cofiber) of
$\pi_{n}(S,s)^\vee[-n-1]\to \oplus_{2\le i \le n-1}\pi_i(S,s)^{\vee}[-i]$. Thus the case of $S_{n}$ follows.
Next we show the case of $S$.
For simplicity, $A:=A_{PL,\infty}(S)$ and $A_{n}:=A_{PL,\infty}(S_{n})$.
As the above proof reveals,
$L_{A_{n-1}}\otimes_{A_{n-1}}\QQ \to L_{A_{n}}\otimes_{A_{n}}\QQ$
can be identified with the inclusion
$\oplus_{2\le i \le n-1}\pi_i(S,s)^{\vee}[-i]\to \oplus_{2\le i \le n}\pi_i(S,s)^{\vee}[-i]$.
It will suffice to prove that the canonical morphism
$\varinjlim_{n} L_{A_{n}}\otimes_{A_{n}}\QQ\to L_{A}\otimes_{A}\QQ$
is an equivalence in $\mathsf{D}(\QQ)$.
Since $\Sigma_+^{\infty}$ preserves colimits,
it is enough to show that the canoncial morphism
$\varinjlim_{n}A_n\to A$ is an equivalence $\CAlg_{\QQ}$.
For this we are reduced to proving
the canonical map $\varinjlim_{n}H^{i}(S_n,\QQ)=\varinjlim_{n}H^{i}(A_n)\to H^{i}(S,\QQ)=H^i(A)$
is bijective for $i\ge0$.
By applying Serre spectral sequence
to the fiber sequence $F_{m,n}=*\times_{S_n}S_m \to S_{m}\to S_{n}$
for $n\le m \le \infty$, we see that
$H^{n}(S_n,\QQ)\simeq H^n(S_{n+1},\QQ)\simeq \ldots \simeq H^n(S,\QQ)$,
so that $\varinjlim_{n}H^{i}(S_n,\QQ)\simeq H^{i}(S,\QQ)$.
\QED
\vspace{2mm}
{\it Proof of Theorem~\ref{motivichomotopy}.}
By Theorem~\ref{real1} and Remark~\ref{real1R},
the image of $M_X\to \uni_k$
can be identified with $A_{PL,\infty}(X^t)\to \QQ$
induced by the point $x$ on $X^t$.
Write $B:=A_{PL,\infty}(X^t)$.
Taking account of Lemma~\ref{nicebasechange},
we see that $\mathsf{R}(LM_{(X,x)})\simeq L_{B}\otimes_{B}\QQ$.
Now our assertion follows from Lemma~\ref{cotangSul}.
\QED
We would like to relate cotangent motives with
partial data of fundamental groups.
\begin{Theorem}
\label{motivichomotopy2}
Let $(X,x)$ be a pointed smooth variety over $k$.
Suppose that
$k$ is embedded in $\CC$.
Let $\pi_i(X^t,x)_{uni}$ be the pro-unipotent completion
of the fundamental group $\pi_i(X^t,x)$ of $X^t$ over $\QQ$.
Then
the $\QQ$-vector space
$H^1(\mathsf{R}(LM_{(X,x)}))$ gets identified with
the cotangent space of the unipotent affine scheme $\pi_1(X^t,x)_{uni}$
at the origin.
\end{Theorem}
\Proof
As in the proof of Theorem~\ref{motivichomotopy},
the image of $M_X\to \uni_k$
can be identified with $A_{PL,\infty}(X^t)\to \QQ$
induced by the point $x$ on $X^t$.
Write $B:=A_{PL,\infty}(X^t)$.
The image of $\uni_k\otimes_{M_X}\uni_k$ under the multiplicative
realization functor can naturally be identified with $\QQ\otimes_{B}\QQ$.
According to Hochschild-Kostant-Rosenberg (HKR)
theorem for $B \in \CAlg_{\QQ}$,
we have $\QQ\otimes_{B}\QQ\simeq \QQ\otimes_{B}B\otimes_{B\otimes B}B\simeq \Free_{\QQ}((L_B\otimes_B\QQ)[1])$
(see e.g. \cite[Prop. 4.4]{BN} for HKR theorem: strictly speaking,
the connectivity on $B$ is assumed in loc. cit., but
its proof shows that the nonconnective affine case holds).
It follows that $H^0(\QQ\otimes_{B}\QQ)\simeq H^0(\Free_{\QQ}((L_B\otimes_B\QQ)[1]))$ (keep in mind that
the dual of the base point $H^0(\QQ\otimes_{B}\QQ)\to H^0(\QQ\otimes_{\QQ}\QQ)\simeq \QQ$ is identified with $H^0(\Free_{\QQ}((L_B\otimes_B\QQ)[1]))\to H^0(\Free_{\QQ}(0))\simeq \QQ$ induced by $L_B\otimes_B\QQ\to L_{\QQ}\simeq 0$).
Remember that $H^0(\QQ\otimes_{B}\QQ)$ is isomorphic to
the coordinate ring of the pro-unipotent completion
$\pi_1(X^t,x)_{uni}$ of $\pi_1(X^t,x)$
over $\QQ$, cf. Proposition~\ref{undcomp}.
By Lemma~\ref{nicebasechange},
$\mathsf{R}(LM_{(X,x)})\simeq L_B\otimes_{B}\QQ$.
Let us observe that
$H^0(\Free_{\QQ}(\mathsf{R}(LM_{(X,x)})[1]))\simeq
H^0(\Free_{\QQ}((L_B\otimes_B\QQ)[1]))$
is naturally isomorphic to the free
ordinary commutative $\QQ$-algebra $\Free_{ord}(H^1(L_B\otimes_B\QQ))$
generated by the $\QQ$-vector space
$H^1(L_B\otimes_B\QQ)\simeq H^0(\mathsf{R}(LM_{(X,x)})[1])$.
Taking account of Lemma~\ref{homotopyfree},
we are reduced to showing that $H^i(L_B\otimes_B\QQ)=0$ for $i<1$.
Thus, it will suffice to prove the following Lemma.
\QED
\begin{Lemma}
Let $B$ be a commutative dg algebra over $\QQ$, which
we regard as an object of $\CAlg_{\QQ}$.
Suppose that we are given an augmentation $B\to \QQ$.
Assume that $H^0(B)=\QQ$, and $H^i(B)=0$ for $i<0$.
Then $H^i(L_B\otimes_B\QQ)=0$ for $i<1$.
\end{Lemma}
\Proof
(This fact or equivalent
versions is well-known, but we prove it for the completeness.)
Let $B_0=\QQ\to B_1\to \cdots \to B_n \to\cdots \to B$
be the inductive sequence associated to
the canonical morphism $\QQ\to B$ in $\CAlg_{\QQ}$,
see Section~\ref{Sull1}.
By Lemma~\ref{Sullivanmodel1},
$\varinjlim_{n}B_n\simeq B$. It follows that $\varinjlim_{n}L_{B_n}\otimes_{B_n}\QQ\simeq L_{B}\otimes_B\QQ$.
Therefore, it is enough
to show that $H^i(L_{B_n}\otimes_{B_n}\QQ)=0$ for $i<1$.
We will prove, by induction on $n\ge 0$,
that (i) $H^0(B_n)=\QQ$, $H^i(B_n)=0$ for $i<0$,
(ii) $H^1(B_n)\to H^1(B)$ is injective, and
(iii) $H^i(L_{B_n}\otimes_{B_n}\QQ)=0$ for $i<1$.
For $n=0$, this is obvious.
Assume therefore that all (i), (ii), (iii) hold for $n$.
Let $M$ be the kernel (fiber) of $B_n\to B$ in $\mathsf{D}(\QQ)$.
Then $H^i(M)=0$ for $i<2$ by the inductive assumptions (i) and (ii).
By definition, $B_{n+1}=B_{n}\otimes_{\Free_{\QQ}(M)}\QQ$.
By the explicit presentation of the homotopy pushout
$B_{n}\otimes_{\Free_{\QQ}(M)}\QQ$ (Propsition~\ref{hopushout}
and Remark~\ref{explicitdg}),
(i) holds for $B_{n+1}$.
In addition, again by the explicit homotopy pushout,
we have an exact sequence $0\to H^1(B_n)\to H^1(B_{n+1})\to H^2(M)$.
Comparing it with the exact sequence
$0\to H^1(B_n)\to H^1(B)\to H^2(M)$ induced by the
cofiber sequence $M\to B_n\to B$,
we see that $H^1(B_{n+1})\to H^1(B)$ is injective.
Note that $L_{B_{n+1}}\otimes_{B_{n+1}}\QQ$ is a cokernel (cofiber) of
$L_{\Free_{\QQ}(M)}\otimes_{\Free_{\QQ}(M)}\QQ\to L_{B_{n}}\otimes_{B_{n}}\QQ$. By Remark~\ref{basefree},
$L_{\Free_{\QQ}(M)}\otimes_{\Free_{\QQ}(M)}\QQ\simeq M$.
Taking account of the inductive assumption (iii) for $B_n$,
we conclude that (iii) holds for $B_{n+1}$.
\QED
\subsection{}
We use the explicit computations of
cohomological motivic algebras in Section~\ref{Sullivanmodel}
to obtain explicit presentations of cotangent motives.
\begin{Theorem}
\label{motivichomotopyexp}
We have the following explicit presentations:
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item
Let $\mathbb{P}^n$ be the projective space over a perfect field $k$
and let $x$ be a $k$-rational point, see Section~\ref{cps}.
Then $LM_{(\mathbb{P}^n,x)}\simeq \uni_k(-1)[-2]\oplus \uni_k(-n-1)[-2n-1]$.
\item Let $X=\mathbb{A}^n-\{p\}$ and let $x$ be a $k$-rational
point, see Section~\ref{minus1}.
Then $LM_{(X,x)}\simeq \uni_{k}(-n)[-2n+1]$.
\item Let $Y=\mathbb{A}^n-\{p\}-\{q\}$ ($n\ge2$)
and let $y$ be a $k$-rational
point, see Section~\ref{minus2}.
Then $LM_{(Y,y)}\simeq \uni_{k}(-n)[-2n+1]^{\oplus 2}\oplus \uni_k(-2n)[-4n+3]\oplus \uni_k(-3n)[-6n+3]^{\oplus 2}\oplus \ldots$.
\item Let $G$ be a semi-abelian variety
and let $o$ be the origin, see Section~\ref{semiab}.
Then $LM_{(G,o)}\simeq M_1(G)^{\vee}$.
\end{enumerate}
\end{Theorem}
\Proof
We show (1). We use the notation in Section~\ref{cps}.
By Propsition~\ref{cpsprop}, $M_{\mathbb{P}^n}\simeq \Free(\uni_{k}(-1)[-2])\otimes_{\Free(\uni_k(-n-1)[-2n-2])}\uni_k$. Let $x^*:M_{\mathbb{P}^n}\to \uni_k$
be the morphism induced by the $k$-rational point $x$.
Note that
$\Free(\uni_{k}(-1)[-2])\to M_{\mathbb{P}^n}\stackrel{x^*}{\to}\uni_k$
is equivalent to
$\Free(\uni_{k}(-1)[-2])\to \Free(0)\simeq \uni_k$
determined by $\uni_{k}(-1)[-2]\to 0$. Indeed, the
morphism $\Free(\uni_{k}(-1)[-2])\to \uni_k$ in $\CAlg(\DM^\otimes(k))$
is classified by the composite $\uni_{k}(-1)[-2]\hookrightarrow \Free(\uni_{k}(-1)[-2])\to \uni_k$ in $\DM(k)$, which is null-homotopic because
$\textup{CH}^1(\Spec k)=0$. Similarly, $\Free(\uni_k(-n-1)[-2n-2])\to M_{\mathbb{P}^n}\to \uni_k$ is equivalent to
$\Free(\uni_k(-n-1)[-2n-2])\to \Free(0)$ determined by $\uni_k(-n-1)[-2n-2]\to 0$.
Since $\Sigma_+^{\infty}:\CAlg(\DM^\otimes(k))_{/\uni_k}\to \SP(\CAlg(\DM^\otimes(k))_{/\uni_k})\simeq \DM(k)$ preserves small colimits, thus
we have the pushout diagram
\[
\xymatrix{
L_{\Free(\uni_{k}(-n-1)[-2n-2])}\otimes_{\Free(\uni_{k}(-n-1)[-2n-2])}\uni_k \ar[d] \ar[r] & L_{\Free(\uni_{k}(-1)[-2])}\otimes_{\Free(\uni_{k}(-1)[-2])}\uni_k \ar[d] \\
L_{\uni_k} \ar[r] & LM_{(\mathbb{P}^n,x)}
}
\]
in $\DM(k)$,
cf. Remark~\ref{basefree}.
Moreover, again by Remark~\ref{basefree},
the upper left term (resp. the upper right term)
is equivalent to $\uni_{k}(-n-1)[-2n-2]$ (resp. $\uni_{k}(-1)[-2]$).
The/any morphism $\uni_{k}(-n-1)[-2n-2]\to \uni_{k}(-1)[-2]$
is null-homotpic.
Combining this consideration with $L_{\uni_k}\simeq0$,
we conclude that $LM_{(\mathbb{P}^n,x)}\simeq \uni_k(-1)[-2]\oplus \uni_k(-n-1)[-2n-1]$. The cases (2) and (4) are easier than (1)
(cf. Proposition~\ref{sphere} and Proposition~\ref{abelianvar}), and the case (3) is similar
to (1) (cf. Proposition~\ref{minus2prop}).
\QED
\begin{Remark}
\label{motcpsR}
Let us consider a meaning of the presentation of the case of projective
spaces. In light of Theorem~\ref{motivichomotopy}, if $k\subset \CC$,
we have
\[
\mathsf{R}(\uni_k(-1))\simeq (\pi_2(\CC\mathbb{P}^n,x)\otimes_{\ZZ}\QQ)^\vee, \ \ \ \ \mathsf{R}(\uni_k(-n-1))\simeq (\pi_{2n+1}(\CC\mathbb{P}^n,x)\otimes_{\ZZ}\QQ)^\vee.
\]
Thus, it is natural to think that $\uni_k(1)$ is a motive for $\pi_2(\CC\mathbb{P}^n,x)\otimes_{\ZZ}\QQ$,
and $\uni_k(n+1)$ is a motive for $\pi_{2n+1}(\CC\mathbb{P}^n,x)\otimes_{\ZZ}\QQ$.
\end{Remark}
\begin{Remark}
According to Theorem~\ref{motivichomotopyexp} (4),
the cotangent motives may also be viewed as
a generalization of (the dual of)
$1$-motives of semi-abelian varieties.
\end{Remark}
\section{Motivic homotopy exact sequence for algebraic curves}
\label{mhes}
Let $X$ be a geometrically connected scheme
of finite type over a perfect field $k$
and let $X_{\bar{k}}$ be the base change to a separable closure $\bar{k}$.
Let $G_k$ denote the absolute Galois group $\Gal(\bar{k}/k)=\pi_1^{\textup{\'et}}(\Spec k,\Spec \bar{k})$.
We write $\pi_1^{\textup{\'et}}(-,a)$ for the \'etale fundamental group of ``$(-)$'' with a base point $a$. Let $\bar{x}:\Spec \bar{k} \to X_{\bar{k}}$ be a geometric
point
and let $x:\Spec \bar{k}\to X$ be the composite.
There is an exact sequence of
profinite groups
\[
1\to \pi_1^{\textup{\'et}}(X_{\bar{k}},\bar{x})\to \pi_1^{\textup{\'et}}(X,x) \to G_k \to 1
\]
induced by $X_{\bar{k}}=X\times_{\Spec k}\Spec \bar{k} \to X \to \Spec k$.
It is usually called the homotopy exact sequence
because it can be thought of as a fairly precise analogue of
the long exact sequence that comes from
a homotopy fiber sequence of topological spaces.
The higher homotopy groups of \'etale homotopy type of $\Spec k$
in the sense of Artin-Mazur are trivial, and
the above exact sequence may be
understood as a part of a long exact sequence.
In this Section,
combining the results
of this paper with
the tannakian theory developed in \cite{DTD}
we formulate and prove a motivic counterpart of a homotopy exact sequence
when $X$ is a smooth curve (Proposition~\ref{MHES}).
The coefficient field of $\DM(k)$
and its full subcategories will be $\QQ$,
whereas $K$ will be a coeffcient field of Weil cohomology theory.
Let $C$ be a smooth curve, that is,
a connected one dimensional smooth
scheme
separated of finite type over a perfect field $k$.
Let $j:C\hookrightarrow \overline{C}$ be a smooth compactification
of $C$.
Namely, $\overline{C}$ is a smooth proper curve over $k$, and
$j$ is an open immersion with a dense image.
Let $Z$ denote the complement $\overline{C}- C$, that is
a finite set of
closed points $Z= p_0 \sqcup p_1 \sqcup \ldots \sqcup p_m$.
For simplicity, we assume that $\overline{C}$ admits a $k$-rational point.
We begin by the definition of a symmetric monoidal
full subcategory of $\DM^\otimes(k)$ that is
``smaller'' and more tractable than $\CAlg(\DM^\otimes(k))$.
\begin{Lemma}
\label{subcategory}
Let $A$ be an abelian variety over $k$ and let $l$ be a finite Galois
extension of $k$.
Let $\DM^\otimes(A,l/k)$
be the smallest symmetric monoidal stable
full subcategory of $\DM^\otimes(k)$
which is closed under colimits and contains $M(A)$,
the dual $M(A)^\vee$, $M(\Spec l)$
and Tate objects $\uni_k(n)$ for any $n\in \ZZ$.
(We remark that the symmetric monoidal structure on $\DM^\otimes(A,l/k)$ inherits from that of $\DM^\otimes(k)$, and
$\DM(A,l/k)$ is presentable.)
Let $C$ be a smooth curve over $k$.
Let $k'$ be a Galois field extension of $k$ such that
for any $0\le i \le m$, the residue field $k_i\supset k$
of $p_i$ can be embedded into $k'$.
Let $J_{\overline{C}}$ be the Jacobian variety of $\overline{C}$.
Then $M_C$ lies in $\CAlg(\DM^\otimes(J_{\overline{C}},k'/k))$.
\end{Lemma}
\Proof
Since the underlying object $M_C \in \DM(k)$ is a dual of $M(C)$,
it suffices to prove that $M(C)^\vee$
belongs to $\DM(J_{\overline{C}},k'/k)$.
We note a decomposition $M(J_{\overline{C}})\simeq \oplus_{i=0}^{2g}M_i(J_{\overline{C}})$
for the Jacobian variety $J_{\overline{C}}$ such that
$M_i(J_{\overline{C}})\simeq \Sym^i(M_1(J_{\overline{C}}))$ (see Section~\ref{semiab}).
Here $g$ is the genus of $\overline{C}$.
Also, there is an isomorphism
$M(\overline{C})\simeq \uni_k\oplus M_1(J(\overline{C}))\oplus \uni_k(1)[2]$ in $\DM(k)$. Thus both $M(\overline{C})$
and $M(\overline{C})^\vee\simeq M(\overline{C})\otimes \uni_k(-1)[-2]$ lie in $\DM(J_{\overline{C}},k'/k)$.
By Gysin triangle (see \cite[14.5]{MVW}),
there is a distinguished triangle
\[
M(C) \to M(\overline{C}) \to M(Z)(1)[2]\to
\]
in the triangulated categories
$\hhh(\DM(k))$.
Therefore, we are reduced to showing that
$M(Z)^\vee\simeq \oplus_{0\le i \le m} M(\Spec k_i)^\vee\simeq \oplus_{0\le i \le m} M(\Spec k_i)$ lies in
$\DM(J_{\overline{C}}, k'/k)$.
Using the functoriality with respect to finite correpondences,
we deduce that each $M(\Spec k_i)$ is a direct summand of $M(\Spec k')$
(since we work with rational coefficients).
\QED
The symmetric monoidal stable presentable $\infty$-category
$\DM^\otimes(A,l/k)$ is a nice property: it is an {\it algebraic fine
tannakian $\infty$-category}.
This notion has been introduced and studied in our work \cite{DTD}.
\begin{Proposition}
We follow the notation in Lemma~\ref{subcategory}.
Let $M_1(A)$ be the direct summand of $M(A)$ in the decomposition in
Section~\ref{semiab}.
Then $M=M_1(A)[-1] \oplus \uni_k(1)\oplus M(\Spec l)$
is a wedge-finite object. Namely, there is an natural number $n$
such that the wedge product $\wedge^{n+1} M$ is zero, and $\wedge^n M$
is an invertible object, see \cite[Section 1]{DTD}.
Consequently, the symmetric monoidal $\infty$-category $\DM^\otimes(A,l/k)$ is an algebraic fine tannakian $\infty$-category,
see \cite[Definition 4.4, Theorem 4.1]{DTD}.
\end{Proposition}
\Proof
By \cite[Proposition 6.1]{DTD} and the fact that $\Hom_{\hhh(\DM(k))}(\uni_k,\uni_k)\simeq \QQ$, it is enough to prove that the wedge product
$\wedge^N M$ is
zero for $N>>0$. To this end, we are reduced to proving that
$\wedge^N (M_1(A)[-1])=0$, $\wedge^N \uni_k(1)=0$, and $\wedge^N M(\Spec l)=0$
for $N>>0$.
By the well-known Kimura finiteness
(see \cite{Kun}, \cite[Thereom7.1.1]{AEH}),
$\wedge^{2e+1} (M_1(A)[-1])\simeq (\Sym^{2e+1}M_1(A))[-2e-1]\simeq 0$
where $e$ is the dimension of $A$.
Also, $\wedge^2 \uni_k(1)=0$
and $\wedge^{d+1}M(\Spec l)=0$. Here $d=[l:k]$.
The final claim follows from the definition of $\DM^\otimes(A,l/k)$
and the definition of algebraic fine tannakian $\infty$-category.
\QED
We define a derived stack from $\DM^\otimes(A,l/k)$ and $M=M_1(A)[-1] \oplus \uni_k(1)\oplus M(\Spec l)$.
By a derived stack over a field $K$, we mean a sheaf $\CAlg_{\QQ}\to \wSSS$
which satisfies a certain geometric condition.
The $\infty$-category $\textup{AlgSt}_{K}$ of derived stacks is defined to be
the full subcategory of $\Fun(\CAlg_{K},\wSSS)$ that consists of derived stacks.
A typical example is a derived affine scheme $\Spec R:\CAlg_{K}\to \wSSS$, that is corepresented by $R\in \CAlg_{K}$. Thus
there is a natural fully faithful embedding $\Aff_{K}\subset \textup{AlgSt}_{K}$.
Another main example for us is a quotient stack $[\Spec R/G]$
that arises from an action of an algebraic affine group scheme
$G$ on $\Spec R$.
We refer to \cite[Section 2.1]{DTD}
for conventions and terminology concerning derived stacks.
\vspace{2mm}
Applying \cite[Theorem 4.1]{DTD} to $\DM^\otimes(A,l/k)$ with the wedge-finite object
$M$ we obtain
\begin{Corollary}
\label{MGS}
Let $n$ be the natural number such that $\wedge^{n+1}M\simeq 0$ and $\wedge^nM$
is an invertible object. (Actually, one can see that
$n=2e+d+1$ if $e$ is the dimension of $A$, and $d=[l:k]$.)
There exist a derived stack $\mathcal{X}_{A,l}$ over $\QQ$ such that
$\mathcal{X}_{A,l}$ has a presentation as a quotient stack of the form
$[\Spec V_{A,l}/\GL_n]$ where
$V_{A,l}$ is in $\CAlg_{\QQ}$,
and a symmetric monoidal $\QQ$-linear equivalence
\[
\phi:\QC^\otimes(\mathcal{X}_{A,l})\simeq \DM^\otimes(A,l/k).
\]
Here $\GL_n$ is the general linear group over $\QQ$ that acts on
$V_{A,l}$,
and $\QC^\otimes(\mathcal{X}_{A,l})$ is the symmetric monoidal
$\QQ$-linear presentable $\infty$-category of quasi-coherent complexes
on $\mathcal{X}_{A,l}$.
We shall call $\mathcal{X}_{A,l}$
the {\it motivic Galois stack} associated to $\DM^\otimes(A,l/k)$ and $M$.
For the definition of $\QC(-)$,
we refer to either \cite[Section 2.3]{DTD} or Remark~\ref{MGSR1}.
\end{Corollary}
\begin{Corollary}
\label{trivialcor}
We continue to use the notation in Lemma~\ref{subcategory}.
Then $M_C$ can be naturally regarded as a commutative
object in $\CAlg(\QC^\otimes(\mathcal{X}_{J(\overline{C}),k'}))$.
\end{Corollary}
\Proof
Combine Lemma~\ref{subcategory} and Corollary~\ref{MGS}.
\QED
\begin{Remark}
\label{MGSR1}
For a quotient stack $[\Spec V/G]$ such that $G$ is an algebraic affine group
scheme, $\QC^\otimes([\Spec V/G])$ can be described in the following way.
The action of $G$ on $\Spec V$ can be defined by a simplicial diagram of
derived affine schemes which is informally given by
$[i]\mapsto \Spec V\times G^{\times i}$.
If we put $\Spec R^i=\Spec V\times G^{\times i}$,
then $\QC^\otimes([\Spec V/G])$
is defined to be $\varprojlim_{[i]}\Mod_{R^i}^\otimes$.
The limit of the cosimplicial
diagram $\{\Mod_{R^n}^\otimes\}_{[n]\in \Delta}$
is taken in the $\infty$-category of symmetric monoidal $\infty$-categories.
\end{Remark}
\begin{Remark}
\label{MGSR2}
The stack $\mathcal{X}_{A,l}\simeq [\Spec V_{A,l}/\GL_n]$
is defined as follows (see \cite{DTD} for details):
Let $\Rep^\otimes(\GL_{n})$ be the symmetric monoidal stable $\infty$-category
of representations of $\GL_n$
(cf. Section~\ref{Sullivanmodel}).
There is a canonical equivalence $\QC^\otimes ([\Spec \QQ/\GL_n])\simeq\Rep^\otimes(\GL_{n})$.
Since $[\Spec V_{A,l}/\GL_n]$ is affine over $B\GL_n:=[\Spec \QQ/\GL_n]$,
$\Spec V_{A,l}$ with action of $\GL_n$
can be identified with an object in $\CAlg(\Rep^\otimes(\GL_{n}))$.
By \cite[Theorem 3.1]{DTD}
we have a symmetric monoidal
colimit-preserving functor
$p:\Rep^\otimes(\GL_{n})\to \DM^\otimes(A,l/k)$ which carries
the standard representation of $\GL_n$ placed in degree zero to
$M$.
By the relative adjoint functor theorem, this functor admits
a lax symmetric monoidal right adjoint
$q:\DM^\otimes(A,l/k)\to \Rep^\otimes(\GL_{n})$.
Thus, $q$ carries a unit object $\uni_{\DM^\otimes(A,l/k)}$
to a commutative algebra object $U_{A,l}:=q(\uni_{\DM^\otimes(A,l/k)})\in \CAlg(\Rep^\otimes(\GL_{n}))$.
This object $U_{A,l}$
amounts to $V_{A,l}$ endowed with action of $\GL_n$,
i.e., data of $[\Spec V_{A,l}/\GL_n]$.
The commutative algebra $V_{A,l}$ in $\CAlg_{\QQ}$ is
the image of $U_{A,l}$ in $\CAlg_{\QQ}$.
We remark that there is a canonical equivalence $\QC^\otimes([\Spec V_{A,l}/\GL_n])\simeq \Mod_{U_{A,l}}^\otimes(\Rep^\otimes(\GL_n))$
where $\Mod_{U_{A,l}}^\otimes(\Rep^\otimes(\GL_n))$
is the symmetric monoidal $\infty$-category of $U_{A,l}$-module objects
in $\Rep^\otimes(\GL_n)$. This equivalence
makes the diagram
\[
\xymatrix{
\QC^\otimes (B\GL_n) \ar[r] \ar[d]_{\simeq} & \QC^\otimes([\Spec V_{A,l}/\GL_n])\ar[d]^{\simeq} \ar[rd]^{\phi} & \\
\Rep^\otimes(\GL_n) \ar[r]_{\otimes U_{A,l}} & \Mod_{U_{A,l}}^\otimes(\Rep^\otimes(\GL_n)) \ar[r] & \DM^\otimes(A,l/k).
}
\]
commute up to homotopy, where the top horizontal arrow is the
pullback functor of the projection $[\Spec V_{A,l}/\GL_n]\to B\GL_n$.
The equivalence
$\Mod_{U_{A,l}}^\otimes(\Rep^\otimes(\GL_n)) \to \DM^\otimes(A,l/k)$
is defined to be the composite
\[
\Mod_{U_{A,l}}^\otimes(\Rep^\otimes(\GL_n))\to \Mod_{p(U_{A,l})}^\otimes(\DM^\otimes(A,l/k))\to \Mod_{\uni_{\DM^\otimes(A,l/k)}}^\otimes(\DM^\otimes(A,l/k))\simeq \DM^\otimes(A,l/k)
\]
where the first functor is induced by $p$, and the second functor is induced by the base change along the counit map $p(U_{A,l})=pq(\uni_{\DM^\otimes(A,l/k)})\to \uni_{\DM^\otimes(A,l/k)}$.
The composite of lower horizontal arrows is equivalent to $p$.
\end{Remark}
\begin{Remark}
There is the following uniqueness.
Let $(\mathcal{Y},N)$ be a pair that consists of
a derived stack $\mathcal{Y}$ over $\QQ$, and $N$ is a vector bundle
on $\mathcal{Y}$.
Here by a vector bundle we mean an object $N$
in $\QC(\mathcal{Y})$
such that for any $f:\Spec R\to \mathcal{Y}$,
the restriction $f^*(N)$ is equivalent to
a direct summand of
some finite coproduct $R^{\oplus m}$.
The stack $\mathcal{X}_{A,l}\simeq [\Spec V_{A,l}/\GL_n]$ has a vector bundle $N_{A,l}$
that is defined to be the pullback of the tautological vector bundle
on $BGL_n=[\Spec \QQ/\GL_n]$.
So we have such a pair $(\mathcal{X}_{A,l},N_{A,l})$.
By the diagram in Remark~\ref{MGSR2}, the equivalence
$\phi:\QC^\otimes(\mathcal{X}_{A,l/k})\simeq \DM^\otimes(A,l/k)$
sends $N_{A,l}$ to $M$.
Assume that there
is a symmetric monoidal $\QQ$-linear equivalence
$\QC^\otimes(\mathcal{Y})\simeq \DM^\otimes(A,l/k)$
which sends $N$ to $M$.
Then there is an equivalence
$\mathcal{Y}\simeq \mathcal{X}_{A,l}$ such that
the induced equivalence
$\QC^\otimes(\mathcal{Y})\simeq \QC^{\otimes}(\mathcal{X}_{A,l})$ sends $N$ to $N_{A,l}$.
This uniqueness will not be necessary in this paper,
so that we will not present the proof.
But one can prove it by using arguments in \cite{DTD}.
\end{Remark}
We say that a morphism $\mathcal{X}\to \mathcal{Y}$ of derived stacks over $K$
is affine
if, for any $\Spec R \to \mathcal{Y}$ from a derived affine scheme,
the fiber product $\mathcal{X}\times_{\mathcal{Y}}\Spec R$ belongs to
$\Aff_{K}$.
Let $\Aff_{\mathcal{Y}}$ be the full subcategory of the overcategory $(\textup{AlgSt}_{K})_{/\mathcal{Y}}$ that consists of affine morphisms
$\mathcal{X}\to \mathcal{Y}$.
There is a canonical equivalence $\Aff_{\mathcal{Y}}\simeq \CAlg(\QC^\otimes(\mathcal{Y}))^{op}$
(cf. \cite[Section 2.3]{DTD}, this is a direct generalization of the analogous
fact in the usual scheme theory).
\begin{Definition}
\label{MRHTMGS}
By Corollary~\ref{trivialcor}, let us consider
$M_C$ as an object
in $\CAlg(\QC^\otimes(\mathcal{X}_{J(\overline{C}),k'}))$.
Let $\mathcal{M}_{C}\to \mathcal{X}_{J(\overline{C}),k'}$
be a derived stack affine over $\mathcal{X}_{J(\overline{C}),k'}$
that corresponds to $M_C$ through
the equivalence $\Aff_{\mathcal{X}_{J(\overline{C}),k'}}\simeq \CAlg(\QC^\otimes(\mathcal{X}_{J(\overline{C}),k'}))^{op}$.
\end{Definition}
Let $\mathsf{R}_E:\DM^\otimes(k)\to \mathsf{D}^\otimes(\KK)\simeq \Mod_{\KK}^\otimes$
be the realization functor associated to a mixed Weil Theory $E$
with coefficients in a field $\KK$ of characteristic zero.
By abuse of notation we write $\mathsf{R}_E$ also for the restriction
$\DM^\otimes(A,l/k) \to \mathsf{D}^\otimes(\KK)$.
Suppose that $\mathsf{R}_E(M)$ is concentrated in degree zero $\mathsf{D}(\KK)$
(all known mixed Weil theories satisfy this condition).
As discussed in \cite[Section 4.1]{PM} or \cite[Remark 6.13]{DTD}, it gives rise to a morphism
\[
\rho_E:\Spec \KK\to \mathcal{X}_{A,l}.
\]
We refer to this morphism as the base point of $\mathsf{R}_E$.
We briefly recall the construction of $\rho_E$.
Let $p:\QC^\otimes(B\GL_n)\simeq \textup{Rep}^\otimes(\GL_n)\to \DM^\otimes(A,l/k)$ be the sequence contained in the diagram in
Remark~\ref{MGSR2}.
Note that this functor carries the standard representation of $\GL_n$
placed in degree zero to $M$, and the realization functor carries
$M$ to the $n$-dimensional vector space placed in degree zero in $\mathsf{D}(\KK)$. Therefore, by the universal property of $\textup{Rep}^\otimes(\GL_n)$
\cite[Theorem 3.1]{DTD}, the composite
$\QC^\otimes(B\GL_n)\simeq \textup{Rep}^\otimes(\GL_n)\to \DM^\otimes(A,l/k)\to \mathsf{D}^\otimes(\KK)$
is equivalent to the pullback functor
$\QC^\otimes(B\GL_n)\to \mathsf{D}^\otimes(\KK)\simeq \QC^\otimes(\Spec \KK)$
along $\Spec \KK\to \Spec \QQ\to B\GL_n$.
Let $u:\mathsf{D}^\otimes(\KK)\to \Rep^\otimes(\GL_n) \simeq \QC^\otimes(B\GL_n)$
be the lax symmetric monoidal right adjoint to
$\QC^\otimes(B\GL_n)\to \QC^\otimes(\Spec \KK)\simeq \mathsf{D}^\otimes(\KK)$,
whose existence is ensured by the relative adjoint functor theorem.
Then this right adjoint induces $\CAlg(\mathsf{D}^\otimes(\KK))\simeq \CAlg_{\KK}\to \CAlg(\Rep^\otimes(\GL_n))$ which carries the unit algebra $\KK$
to $u(\KK)\simeq \Gamma(\GL_n)\otimes_{\QQ}\KK \in \CAlg(\Rep^\otimes(\GL_n))\simeq \CAlg(\QC^\otimes(B\GL_n))$. Here, write $\Gamma(\GL_n)$ for
the (ordinary) coordinate ring of the general linear group $\GL_n$
which is endowed with the natural action of $\GL_n$. The symbol $\KK$ in $\Gamma(\GL_n)\otimes_{\QQ}\KK$ is understood as the $\QQ$-algebra $\KK$ with the trivial
action of $\GL_n$.
Note that there is a natural morphism $U_{A,l}\to u(\KK)\simeq \Gamma(\GL_n)\otimes_{\QQ}\KK$ in $\CAlg(\QC^\otimes(B\GL_n))$. In fact, if $v:\CAlg_{K}\to \CAlg(\DM^\otimes(A,l/k))$
denotes the right adjoint to
the restricted multiplicative realization functor
$\CAlg(\DM^\otimes(A,l/k))\to\CAlg_{\KK}$, then
there is a unit map $\uni_{\DM(A,l/k)}\to v(\KK)$
that induces $U_{A,l}=q(\uni_{\DM(A,l/k)})\to qv(\KK)=u(\KK)$, as claimed (for the functor $q$, see Remark~\ref{MGSR2}).
By using the equivalence $\Aff_{B\GL_n}\simeq \CAlg(\QC^\otimes(B\GL_n))^{op}$,
we obtain
$\rho_E:\Spec K\simeq [\Spec \Gamma(\GL_n)\otimes_{\QQ}K/\GL_n]\to \mathcal{X}_{A,l}=[\Spec V_{A,l}/\GL_n]$.
\begin{Remark}
\label{geomreal}
By this construction and Remark~\ref{MGSR2}, we see that the diagram
\[
\xymatrix{
\QC^\otimes (\mathcal{X}_{A,l}) \ar[r]^{\rho^*_E} \ar[d]_{\simeq}^\phi & \QC^{\otimes}(\Spec \KK) \ar[d]^\simeq \\
\DM^\otimes(A,l/k) \ar[r]^{\mathsf{R}_E} & \mathsf{D}^\otimes(\KK)
}
\]
commutes up to homotopy,
where $\rho^*_E$ is the pullback functor (cf. \cite[Section 2.3]{DTD}),
the right vertical arrow is a canonical equivalence.
\end{Remark}
One can associate to
the base point $\rho_E:\Spec \KK\to \mathcal{X}_{A,l}$
a derived affine group scheme over $\KK$.
Namely, we take the Cech nerve $G:\NNNN(\Delta_+)^{op}\to \textup{AlgSt}_\KK$
of $\rho_E\times\textup{id}:\Spec \KK\to \mathcal{X}_{A,l}\times_{\Spec \QQ}\Spec \KK$, which is defined to be the right Kan extension
$\NNNN(\Delta_+)^{op}\to \textup{AlgSt}_{\KK}$ of $\NNNN(\Delta_+^{\le0})^{op}=\NNNN(\{[-1]\to [0]\})^{op}\to \textup{AlgSt}_{\KK}$ determined by $\rho_E\times\textup{id}$.
The evaluation $G([1])$ is equivalent to
$\Spec \KK\times_{\mathcal{X}_{A,l}\times \Spec \KK}\Spec \KK$ which is affine because
the diagonal $[\Spec V_{A,l}/\GL_n]\to [\Spec V_{A,l}/\GL_n]\times [\Spec V_{A,l}/\GL_n]$ is affine.
Thus the restriction of $G$ defines a group object $\NNNN(\Delta)^{op}\to \Aff_K$,
whose underlying derived affine scheme is $\Spec \KK\times_{\mathcal{X}_{A,l}\times \Spec \KK}\Spec \KK$. We write $\Omega_{\rho_E}\mathcal{X}_{A,l}$ for this
derived affine group scheme over $\KK$.
The derived group scheme $\Omega_{\rho_E}\mathcal{X}_{A,l}$
is related to the derived motivic Galois
group:
\begin{Proposition}
\label{loopgp}
Let $\mathsf{MG}_{E,\DM^\otimes(A,l/k)}$ be the derived motivic Galois
group which represents the automorphism group functor $\Aut(\mathsf{R}_E|_{\DM^\otimes(A,l/k)}):\CAlg_{\KK}\to \Grp(\wSSS)$, cf. Remark~\ref{motsubcategory}.
Then $\Omega_{\rho_E}\mathcal{X}_{A,l}$ is naturally equivalent to
$\mathsf{MG}_{E,\DM^\otimes(A,l/k)}$.
\end{Proposition}
\Proof
By Remark~\ref{geomreal}, we have
$\Aut(\mathsf{R}_E|_{\DM^\otimes(A,l/k)})\simeq \Aut(\rho_E^*)$
where
$\rho^*_E:\QC^\otimes (\mathcal{X}_{A,l}) \to \QC^{\otimes}(\Spec \KK)$.
It will suffice to show that $\Omega_{\rho_E}\mathcal{X}_{A,l}\simeq \Aut(\rho_E^*)$.
This equivalence follows from \cite[Proposition 4.6]{Bar}.
\QED
\begin{Remark}
By the representability of automorphism groups,
the restriction to $\DM^\otimes(A,l/k)$
induces $\MG_E\to \mathsf{MG}_{E,\DM^\otimes(A,l/k)}\simeq \Omega_{\rho_E}\mathcal{X}_{A,l}$.
The action of $\MG_E$ on $\mathsf{R}_E(M_C)$ described in Proposition~\ref{goodaction} factors
through
$\MG_E\to \Omega_{\rho_E}\mathcal{X}_{A,l}$.
\end{Remark}
\vspace{2mm}
Now we are ready to prove
the following motivic generalization of homotopy exact sequence.
\begin{Proposition}
\label{MHES}
Let $\mathcal{M}_C\to \mathcal{X}_{J(\overline{C}),k'/k}$ be
the affine morphism defined in Definition~\ref{MRHTMGS}.
Let us consider the pullback diagram of derived stacks
\[
\xymatrix{
\mathcal{F}_E \ar[r] \ar[d] & \mathcal{M}_C \ar[d] \\
\Spec \KK \ar[r]^{\rho_E} & \mathcal{X}_{J(\overline{C}),k'}
}
\]
in $\textup{AlgSt}_{\QQ}$.
(One may think of this diagram as a Cartesian diagram in $\Fun(\CAlg_{\QQ},\wSSS)$.)
Then
the fiber $\mathcal{F}_E$ is naturally equivalent to $\Spec \mathsf{R}_E(M_C)$,
where $\mathsf{R}_E(M_C)$ in $\CAlg_{\KK}$ is the image of $M_C$
under the multiplicative realization functor $\mathsf{R}_E:\CAlg(\DM^\otimes(k))\to \CAlg_{\KK}$.
In particular, when $E$ is the singular cohomology theory,
by Theorem~\ref{real1} we have
a Cartesian diagram
\[
\xymatrix{
\Spec A_{PL,\infty}(C^t) \ar[r] \ar[d] & \mathcal{M}_C \ar[d] \\
\Spec \QQ \ar[r]^{\rho_E} & \mathcal{X}_{J(\overline{C}),k'}.
}
\]
\end{Proposition}
\begin{Remark}
The morphism
$\mathcal{M}_C \to \mathcal{X}_{J(\overline{C}),k'}$
should be thought of as a motivic counterpart of the delooping of
$\pi_1^{\textup{\'et}}(X,x) \to G_k$.
By Proposition~\ref{loopgp}
we can obtain the derived motivic Galois group $\MG_{E,\DM^\otimes(J(\overline{C}),k'/k)}\simeq \Omega_{\rho_E}\mathcal{X}_{J(\overline{C}),k'}$ from the base stack $\Spec \KK\to \mathcal{X}_{J(\overline{C}),k'}$
by using the construction of the base loop space.
The fiber $\mathcal{F}_E$ shoud be understood as a role of
the delooping of $\pi_1^{\textup{\'et}}(X_{\bar{k}},\bar{x})$.
Consider the situation that $k$ is a subfield of $\CC$.
Then
$\pi_1^{\textup{\'et}}(X_{\bar{k}},\bar{x})$ is isomorphic to
the profinite completion of the (topological)
fundamental group $\pi_1(X^t,\bar{x})$
of the underlying topological space $X^t$
of $X\times_{\Spec k}\Spec \CC$.
On the other hand, if we fix a $k$-rational point $c$,
the unipotent group scheme
$\overline{G}^{(1)}(C,c)\simeq \Spec H^0(\QQ\otimes_{A_{PL}(C^t)}\QQ)$ is
the pro-unipotent completion of the topological fundamental group
$\pi_1(C^t,c)$.
\end{Remark}
\Proof
We have already done almost things.
By Remark~\ref{geomreal}, one can identify the multiplicative realization functor
$\CAlg(\DM^\otimes(J(\overline{C}),k'/k)) \to \CAlg_{\KK}$
with
$\CAlg(\QC^\otimes(\mathcal{X}_{J(\overline{C}),k'}))\to \CAlg_{\KK}\simeq \CAlg(\QC^\otimes(\Spec \KK))$ induced by the pullback functor $\rho_E^*$.
Then we use the observation that the canonical equivalences
\[
\CAlg(\QC^\otimes(\mathcal{X}_{J(\overline{C}),k'}))^{op}\simeq \Aff_{\mathcal{X}_{J(\overline{C}),k'}}\ \ \textup{and}\ \ \CAlg(\QC^\otimes(\Spec \KK))^{op}\simeq \Aff_{\Spec \KK}
\]
are compatible
with pullback functors. Namely, through these canonical equivalences,
the opposite functor $\CAlg(\QC^\otimes(\mathcal{X}_{J(\overline{C}),k'}))^{op}\to \CAlg(\QC^\otimes(\Spec \KK))^{op}$ can be identified with
$\Aff_{\mathcal{X}_{J(\overline{C}),k'}}\to \Aff_{\Spec \KK}=\Aff_{\KK}$
given by $\{\mathcal{Z}\to \mathcal{X}_{J(\overline{C}),k'}\} \mapsto \{\textup{pr}_2:\mathcal{Z}\times_{\mathcal{X}_{J(\overline{C}),k'}}\Spec \KK\to \Spec \KK\}$.
Therefore, we see that
$\mathcal{F}_E$ is equivalent to $\Spec \mathsf{R}_E(M_C)$
via these identifications.
\QED
|
\section{Introduction}
Permanent magnets (PM) are widely used in daily life,
e.g. loudspeakers, electric motors in air conditioners, washing
machines, and hybrid cars. PMs are also used in technological
applications, such as traveling wave tubes, Hall-effect sensors, and
sorting or separation
equipment~\cite{mccallum.arms2014,kramer.cms2012}. The PM family is
mainly composed of ferrites, rare-earth based PM (Nd-Fe-B and Sm-Co),
and alnico~\cite{mccallum.arms2014,kramer.cms2012}. Alnico is an alloy
system composed mainly of Al, Ni, Co, Fe and a small amount of Ti and
Cu. Alnico magnets have attracted recent attention due to their good
performance at high temperature and worldwide abundance of their
elemental
components~\cite{mccallum.arms2014,kramer.cms2012,anderson.jap2015}. Alnico
was discovered in the 1930s and had been continuously developed and
optimized until the 1970s. Currently, the commercial alnico alloys
with the highest coercivity and energy product ($BH_\text{max}$) are alnico
8H with coercivity ($H_\text{cj}$) of \SI{2.02}{kOe} and alnico 9 with a
$BH_\text{max}$ of \SI{10.0}{MGOe},
respectively~\cite{kramer.cms2012,zhou.acta2014}. The commercial
alloys consist of periodically distributed and crystallographically
oriented (Fe-Co)-rich ($\alpha_1$) rods embedded in an (Al-Ni)-rich
($\alpha_2$) matrix form by spinodal decomposition (SD) within an applied
magnetic field~\cite{zhou.acta2014}. Microstructural details depend on
specific alloy chemistry and processing~\cite{zhou.acta2014}. Shape
anisotropy of the $\alpha_1$ rods is currently understood to provide a
basic level of coercivity to the alloys~\cite{aquino.msf1999}. The
relatively high magnetic energy product of alnico 9 is derived from a
directional solidification process used to cast the magnet that
provides overall microstructural alignment and magnetic hysteresis
loop ``squareness''. Theoretical simulations performed for similar
systems suggest that alnico may be able to double its energy product
if the $\alpha_1$ rods can be controlled with a volume fraction of $2/3$
and a diameter of
$\sim$\SIrange{5}{15}{\nm}~\cite{zeng.prb2002,skomski.jap2010}.
Alnico alloys achieve their best magnetic properties after a
complicated heat-treatment (HT) process, which includes a high
temperature ($\sim$\SI{1250}{\celsius}) solution treatment, followed
by short-time thermal-magnetic annealing ($\sim$\SI{800}{\celsius})
and a long-time lower temperature stepped drawing
($\sim$\SIrange{550}{700}{\celsius})~\cite{mccurrie.hfm1982}. Previous
studies indicated that the magnetic properties of alnico are sensitive
to the time and temperature of the magnetic field annealing (MA), as
well as to the constituent elements of the alloy, including minor
additions
~\cite{zhou.acta2014,mccurrie.hfm1982,szymura.actapol1975,sergeyev.ieeetm1970,sergeyev.jap1969,iwama.tjim1974,iwama.tjim1970,iwama.tjim1967,tang.ieeetm2015}.
However, a comprehensive investigation on how different HT parameters
affect the microstructure and magnetic property evolution of alnico is
lacking. Better understanding of the effect of the various processing
steps on morphology and chemistry of the SD phase separation process
for alnico is required in order to achieve the alloy’s theoretical
limit. In this study, we have delved further into unraveling the
complex nanostructures that control the alloy properties by careful
microstructural evaluation of an isotropic alnico alloy with different
HT process conditions. In particular, we analyzed the effects of
annealing time, annealing temperature, as well as the presence of an
external magnetic field. A combination of transmission electron
microscopy (TEM) and atom probe tomography (APT) techniques were used
to characterize both the morphology and chemistry of phases in the
alloy.
\section{Experimental details}
Using commercial alnico 8H as a model composition, a powder having a
composition of 7.3Al-13.0Ni-38.0Co-32.3Fe-3.0Cu-6.4Ti
(wt.\si{\percent}) was produced by close-coupled gas-atomization. The
spherical gas atomized powders provide a uniform starting
material. Sieved powders, $<$\SI{20}{\um} in diameter, were loaded
into a stainless steel can for hot isostatic pressing (HIP). Before
sealing, the powders were outgassed at \SI{425}{\celsius} for
\SI{2}{\hour} under vacuum (\SI{1.3d-3}{\pascal}). The HIP can was
sealed under vacuum by electron beam welding. The HIP process was
performed under a pressure of \SI{60}{\mega\pascal} at
\SI{1250}{\celsius} for \SI{4}{\hour}. Center sections of the HIP
consolidated alnico were cut into cylinders (\SI{3}{\mm} diameter and
\SI{8}{\mm} length) using electrical discharge machining. The
resulting alloy was polycrystalline with random grain orientations.
Since the HIP samples undergo slow cooling, the nanostructures were
reset by reheating the cylindrical samples at \SI{1250}{\celsius} for
\SI{30}{\minute} in vacuum (\SI{1.3d-4}{\pascal}) and quenched using
an oil bath. The samples were then annealed between \SI{800}{\celsius}
and \SI{860}{\celsius} for different periods of time (0.5 to
\SI{60}{\minute}) with and without an external applied magnetic field
of \SI{1.0}{\tesla}. Some samples also went through an additional
multi-stepped low temperature annealing or ‘drawing process’ that is
typically used in industry. The first step was at \SI{650}{\celsius}
for \SI{5}{\hour} and the second step was at \SI{580}{\celsius} for
\SI{15}{\hour}. The HT process that includes both \SI{10}{\minute}
magnetic field annealing and multi-stepped drawing is termed ``full
heat treatment (FHT) condition''. To capture the effect of applying an
external field on the morphology of SD phases at the early stage of
phase separation, APT analysis was performed on two samples, which
were annealed at \SI{840}{\celsius} for only \SI{90}{\second} with and
without an external applied field of \SI{1}{\tesla}, followed by a
water quench to room temperature. All other samples were furnace
cooled to room temperature after their HT process. Samples labeled as
$\smpl{A}$ to $\smpl{H}$ were selected for TEM analysis and their
heat-treatment conditions are summarized in \rtbl{tbl:ht-condition}.
\begin{table}[hbtp]
\footnotesize
\centering
\caption{Heat-treatment conditions of samples analyzed by TEM. Draw
was performed for \SI{5}{\hour}@\SI{650}{\celsius} followed by
\SI{15}{\hour}@\SI{580}{\celsius}.}
\label{tbl:ht-condition}
\bgroup
\def\arraystretch{1.2
\begin{tabular}[0.9\linewidth]{ccccccc}
\hline
\\[-1.1em]
\multirow{3}{*}{Samples}& & \multicolumn{3}{c}{MA} & & \multirow{3}{*}{Draw} \\ \cline{3-5}
& & Temperature & Time & Magnetic & & \\
& & (\si{\celsius}) & (\si{\minute}) & Field & & \\
\\[-1.1em]
\hline
\\[-1.2em]
$\smpl{A}$ & & 800 & 10 & Yes & & \\
$\smpl{B}$ & & 840 & 10 & Yes & & \\
$\smpl{C}$ & & 860 & 10 & Yes & & \\
\\[-1.2em]
$\smpl{D}$ & & 800 & 10 & Yes & & Yes \\
$\smpl{E}$ & & 840 & 10 & Yes & & Yes \\
$\smpl{F}$ & & 860 & 10 & Yes & & Yes \\
\\[-1.2em]
$\smpl{G}$ & & 840 & 1.5 & Yes & & \\
$\smpl{H}$ & & 840 & 60 & Yes & & \\
\\[-1.2em]
$\smpl{I}$ & & 800 & 10 & & & \\
$\smpl{J}$ & & 840 & 10 & & & \\
$\smpl{K}$ & & 860 & 10 & & & \\
\\[-1.2em]
\hline
\end{tabular}
\egroup
\end{table}
The magnetic properties were measured using a Laboratorio
Elettrofisico Engineering Walker LDJ Scientific AMH-5 Hysteresis graph
with a \SI{5}{\mm} coil and a maximum applied field of \SI{15.0}{kOe}
at room temperature in a closed-loop setup. TEM samples were prepared
by mechanical wedge-polishing followed by a short time, low-voltage Ar
ion-milling with a liquid-nitrogen-cold stage. TEM analysis was
performed on transverse (observation along the magnetic field
direction during annealing) and longitudinal (observation
perpendicular to the magnetic field direction during annealing)
orientations. An FEI Tecnai F20 (\SI{200}{\kV}, equipped with a
field-emission gun (FEG)) was used for microstructural
characterization. Scanning electron microscopy (SEM) was performed
using an FEI Nova FEG-SEM. For the MA sample, APT tips were lifted
out from a grain with its $[001]$ crystallographic direction parallel
to the external magnetic field direction. The orientation of that
grain was identified using orientation imaging microscopy (OIM) on an
Amray 1845 FEG-SEM. An FEI Nova 200 dual-beam focused ion beam (FIB)
instrument was used to perform lift-outs and annular milling of
targeted grains to fabricate the needle-shaped APT specimens. A wedge
shaped lift-out geometry was used to mount multiple samples on a Si
``microtip'' array to enable the fabrication of several needles from
one wedge lift-out~\cite{thompson.um2007}. APT was performed with a
local electrode atom probe (CAMECA LEAP 4000X HR). Samples were run in
voltage mode with a base temperature of \SI{40}{\K} and
\SI{20}{\percent} pulse fraction at a repetition rate of
\SI{200}{\kHz}. The datasets were reconstructed and analyzed using the
IVAS 3.6.12 software (CAMECA
Instruments)~\cite{miller.sia2007,guo.nanotech2016}.
\section{Experimental results}
\subsection{Effect of magnetic field annealing temperature}
\begin{figure}[!bh]
\centering \includegraphics[width=0.42\textwidth,clip]{fig01.pdf}%
\caption{ Effect of MA temperature on (a) $H_\text{cj}$ and (b) $M_\text{s}$ of
MA-only (blue), MA+draw$@$\SI{650}{\celsius} (red), and FHT (green)
samples.}
\label{fig:01}
\end{figure}
\begin{figure*}[!thb]
\centering
\includegraphics[width=0.70\textwidth,clip]{fig02.pdf}%
\caption{ SEM images of samples (a) $\smpl{D}$, (b) $\smpl{E}$, and
(c) $\smpl{F}$. The bottom row is a corresponding higher
magnification image from the image on top.}
\label{fig:02}
\end{figure*}
\begin{figure*}[!bht]
\centering
\includegraphics[width=0.70\textwidth,clip]{fig03.pdf}%
\caption{HAADF STEM images of samples (a) $\smpl{A}$, (b) $\smpl{B}$,
(c) $\smpl{C}$, (d) $\smpl{D}$, (e) $\smpl{E}$, and (f) $\smpl{F}$
looking along the transverse direction. The TEM images are taken
close to $[100]$ zone axis. A well defined mosaic pattern is only
shown in sample $\smpl{B}$ and $\smpl{E}$.}
\label{fig:03}
\end{figure*}
The effect of MA temperature on the intrinsic coercivity ($H_\text{cj}$) and
saturation magnetization ($M_\text{s}$) is shown in \rfig{fig:01}. The $H_\text{cj}$
and $M_\text{s}$ of MA-only, MA + \SI{650}{\celsius} (\SI{5}{\hour} draw),
and FHT samples are indicated by the blue, red, and green lines,
respectively. For all samples, it is obvious that $H_\text{cj}$ is very
sensitive to the MA temperature. The $H_\text{cj}$ maximizes at \SI{1360}{Oe}
at \SI{840}{\celsius} for MA-only and at \SI{1845}{Oe} after FHT. Note
that low-temperature drawing consistently increases $H_\text{cj}$ and the
peak position remains at \SI{840}{\celsius}. Each of the two
annealing steps results with a nearly constant increase. The first
draw after MA caused a slight decrease of $M_\text{s}$, but there is no
obvious further change of $M_\text{s}$ after the final drawing. To better
understand how the applied magnetic field and various annealing steps
affect magnetic properties, microstructure evolution of select
samples, labeled as $\smpl{A}$ through $\smpl{F}$ in \rfig{fig:01}
were examined in detail.
SEM images of the FHT samples subjected to MA at
\SIlist{800;840;860}{\celsius} (samples $\smpl{D}$, $\smpl{E}$, and
$\smpl{F}$, respectively), are shown in \rfig{fig:02}. For all
samples, an average grain size of $\sim$\SIrange{30}{100}{\um} was
observed. The variation of contrast between grains is due to their
random orientation. At higher magnification, a small phase fraction of
mixed phases at the grain boundaries (GBs) can be observed. The
thickness and, therefore, volume fraction of these GB phases increase
with increasing MA temperature. Energy dispersive X-ray analysis (EDX)
of sample $\smpl{F}$ shows that the bright region in the grain
boundary is (Fe-Co)-rich, while the gray region is (Al-Ni)-rich.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.70\textwidth,clip]{fig04.pdf}%
\caption{HAADF STEM images of samples (a) $\smpl{A}$, (b) $\smpl{B}$,
(c) $\smpl{C}$, (d) $\smpl{D}$, (e) $\smpl{E}$, and (f) $\smpl{F}$
looking along the longitudinal direction. Elongation of the $\alpha_1$
phase is observed in all samples. Small $\alpha_1$ phases are indicated
by yellow dotted arrows.}
\label{fig:04}
\end{figure*}
TEM examination of samples $\smpl{A}$ through $\smpl{F}$ was performed
along both the transverse and longitudinal
directions. High-angle-annular-dark-field (HAADF) scanning
transmission electron microscopy (STEM) imaging ($Z$-contrast imaging)
was used to minimize strain contrast and to differentiate phase
morphology. HAADF-STEM imaging is sensitive to the atomic number $Z$,
i.e., elements with higher atomic number will exhibit brighter
contrast in the image because of greater electron scattering at large
detector collection angles~\cite{williams.book1996}. Samples
$\smpl{A}$, $\smpl{B}$, and $\smpl{C}$ were subjected to MA for
\SI{10}{\minute} at \SIlist{800;840;860}{\celsius}, respectively,
followed by furnace cooling. STEM observations along the transverse
direction showed that diamond shaped ($\sim$\SI{20}{\nm} diameter) and
rod shaped ( $\sim$\SI{20}{\nm} diameter and
$\sim$\SIrange{60}{200}{\nm} length) $\alpha_1$ phases coexisted in sample
$\smpl{A}$ (\rfig{fig:03}a). Also, Cu-enriched particles are observed
at the corners of the two $\{110\}$ facets of the diamond shape $\alpha_1$
phases~\cite{zhou.acta2014}. The longer axis of the rod shape $\alpha_1$
phase is aligned along the $\langle100\rangle$ crystallographic
direction but formed nearly orthogonal to the applied magnetic
field. Sample $\smpl{B}$ has a well-defined ``mosaic'' patterned
structure (\rfig{fig:03}b), similar to commercial alnico
8~\cite{zhou.acta2014}. The $\alpha_1$ phase has a diameter of
$\sim$\SIrange{30}{40}{\nm}, indicating an increase in the $\alpha_1$ rod
diameter with a higher annealing temperature. A further increase of MA
temperature to \SI{860}{\celsius} destroys the well-defined mosaic
pattern, as shown in sample $\smpl{C}$ (\rfig{fig:03}c), and the
morphology of $\alpha_1$ phase becomes irregular. Both large $\alpha_1$
regions with sizes up to \SI{70}{\nm} and small $\alpha_1$ phases with a
diameter of $\sim$\SI{10}{\nm} were observed. Cu-enriched particles
were sometimes observed between two $\alpha_1$ phases. Samples $\smpl{D}$
through $\smpl{F}$ were subjected to an additional lower temperature
drawing process (\SI{5}{\hour}@\SI{650}{\celsius} followed by
\SI{15}{\hour}@\SI{580}{\celsius}), after MA at different
temperatures. Compared to the MA-only samples treated at the same
temperature, no significant morphology modification was observed along
the transverse direction after drawing, as shown in
\rfig{fig:03}d--f. However, there are some subtle contrast changes in
the $\alpha_1$ and $\alpha_2$ phases and morphology changes in the Cu-rich
phases as discussed below.
Observation along the longitudinal direction (\rfig{fig:04}) showed
that the $\alpha_1$ phase is elongated for all samples ($\smpl{A}$ through
$\smpl{F}$), though less consistently for the samples subject to MA at
\SI{860}{\celsius}. For sample $\smpl{A}$, the diameter of the
elongated $\alpha_1$ phase is $\sim$\SI{20}{\nm}. Small $\alpha_1$
precipitates ($\sim$\SIrange{3}{5}{\nm}), as indicated by dotted
yellow arrows in \rfig{fig:04}a and confirmed by EDS point analysis,
were commonly observed inside the $\alpha_2$ phase between two elongated
$\alpha_1$ rods. Small $\alpha_1$ precipitates agglomerate and form slightly
bigger particles after the drawing process, as exhibited by sample
$\smpl{D}$ (\rfig{fig:04}d). For samples treated at
\SI{840}{\celsius}, the $\alpha_1$ rods are $\sim$\SIrange{30}{40}{\nm}
wide and up to \SI{1}{\um} long. Cu-enriched rods with a length of
$\sim$\SIrange{10}{50}{\nm} precipitated along the edge of the $\alpha_1$
rods in sample $\smpl{B}$ (\rfig{fig:04}b). After the lower
temperature drawing process, the length of the Cu-enriched rods
increases to $\sim$\SIrange{200}{300}{\nm}, as shown in sample
$\smpl{E}$ (\rfig{fig:04}e). For the sample subject to MA at
\SI{860}{\celsius} (sample $\smpl{F}$), small $\alpha_1$ particles tended
to coarsen after drawing. The length of $\alpha_1$ rods had a wide size
range between tens of nanometers and up to \SI{1}{\um}, as shown in
\rfig{fig:04}f.
\subsection{Effect of magnetic field annealing time}
\begin{figure}[h]
\centering
\includegraphics[width=0.42\textwidth,clip]{fig05.pdf}%
\caption{Effect of MA time at \SI{840}{\celsius} on $H_\text{cj}$ of MA-only
(blue) and FHT (green) samples.}
\label{fig:05}
\end{figure}
\begin{figure*}[h]
\centering
\includegraphics[width=0.70\textwidth,clip]{fig06.pdf}%
\caption{HAADF STEM images of samples (a) $\smpl{G}$, (b) $\smpl{B}$,
and (c) $\smpl{H}$ looking along the transverse direction. The TEM
images are taken close to the [100] zone axis. }
\label{fig:06}
\end{figure*}
Since MA at \SI{840}{\celsius} is critical for the formation of a
well-defined mosaic pattern that produces maximum $H_\text{cj}$, experiments
were designed to study the effect of the MA time at \SI{840}{\celsius}
on the magnetic properties. Samples were treated at \SI{840}{\celsius}
for time ranging from \SI{30}{\second} to \SI{60}{\minute}, followed
by furnace cooling. The measured $H_\text{cj}$ values are shown in
\rfig{fig:05}. The blue line denotes samples with the MA treatment
only, while the green line denotes samples treated with the FHT
process. The $H_\text{cj}$ first increases to $\sim$\SI{1400}{Oe} as MA time
increases from $0$ to $\sim$\SI{3}{\minute}. Between $3$ and
\SI{20}{\minute}, the $H_\text{cj}$ stays relatively constant. After a
\SI{60}{\minute} MA, the $H_\text{cj}$ decreases to
$\sim$\SI{1100}{Oe}. Drawing improves the $H_\text{cj}$ of all samples and
the increase of $H_\text{cj}$ with low temperature annealing for the drawn
samples follows the same trend as that for samples without
drawing. Samples labeled as $\smpl{G}$ and $\smpl{H}$ in \rfig{fig:05}
were selected for further TEM analysis.
\rFig{fig:06} shows the SD morphologies viewed along the transverse
direction of samples $\smpl{G}$, $\smpl{B}$ (same as in Fig.3 but used
again here for comparison), and $\smpl{H}$ with MA time of
\SIlist{1.5;10;60}{\minute}, respectively. The diameter of the $\alpha_1$
phase increases from $\sim$\SI{15}{\nm} to $\sim$\SI{35}{\nm} with
increasing MA time from \SI{1.5}{\minute} to \SI{10}{\minute},
respectively. Most of the area imaged in sample $\smpl{G}$
(\rfig{fig:06}a) formed a mosaic pattern, but some regions exhibited a
blurry contrast, which appears to originate from incomplete phase
separation. Increasing the MA time to \SI{60}{\minute} drastically
modifies the SD morphology, as shown in \rfig{fig:06}c (sample
$\smpl{H}$). The size of the $\alpha_1$ phase transformed into a bimodal
distribution during the extended MA. The morphology consists of a
larger $\alpha_1$ phase with a size of $\sim$\SIrange{60}{80}{\nm} and a
smaller round (non-faceted) $\alpha_1$ phase with a size of
$\sim$\SI{10}{\nm}. The larger $\alpha_1$ phase has alternating $\{100\}$
and $\{110\}$ facets. Some of the smaller $\alpha_1$ phases are located
at the regions between two $\{100\}$ facets of the bigger $\alpha_1$
phases and some of them form clusters. Cu-enriched phases were also
observed at the corners of two larger $\alpha_1$ facets, but these
Cu-enriched phases are slightly smaller compared to those in sample
$\smpl{B}$.
\subsection{Effect of magnetic field}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.42\textwidth,clip]{fig07.pdf}%
\caption{Comparison on effect of magnetic field on (a) $H_\text{cj}$ and (b)
$B_\text{r}$ of samples treated at different temperatures. Red line are
from samples treated with a magnetic field, while the green line are
from samples treated without a magnetic field.}
\label{fig:07}
\end{figure}
Finally, the magnetic field effect was analyzed by comparing the
magnetic properties ($H_\text{cj}$ and $B_\text{r}$) of samples treated with/without
external magnetic fields at different temperatures, as shown in
\rfig{fig:07}. Samples treated in the absence of a magnetic field at
\SIlist{800;840;860}{\celsius} are labeled as $\smpl{I}$, $\smpl{J}$,
and $\smpl{K}$, respectively, and are compared to samples $\smpl{A}$,
$\smpl{B}$, and $\smpl{C}$ in \rfig{fig:07}. For both sets of samples,
there is a maximum $H_\text{cj}$ at \SI{840}{\celsius}, but the $H_\text{cj}$ of
sample $\smpl{J}$ is $\sim$\SI{840}{Oe}, which is smaller than that of
sample $\smpl{B}$. The presence of an external magnetic field barely
affects the $H_\text{cj}$ of samples treated at 800 and
\SI{860}{\celsius}. On the other hand, all samples treated using a
magnetic field have an obviously higher $B_\text{r}$
($\sim$\SI{45}{\percent}) compared with those treated without a
magnetic field, as shown in \rfig{fig:07}b.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.70\textwidth,clip]{fig08.pdf}%
\caption{ Transverse HAADF STEM images of samples $\smpl{I}$ (a),
$\smpl{J}$ (b), $\smpl{K}$ (c), and longitudinal HAADF STEM images
of samples $\smpl{I}$ (d), $\smpl{J}$ (e), $\smpl{K}$ (f).}
\label{fig:08}
\end{figure*}
TEM analysis along the transverse direction shows that the morphology
of the SD phases in samples $\smpl{I}$ and $\smpl{K}$ (\rfig{fig:08}a
and ~\ref{fig:08}c, respectively) is similar to that of samples
$\smpl{A}$ and $\smpl{C}$ (\rfig{fig:03}a and ~\ref{fig:03}c),
indicating no biasing of the microstructure along the transverse
direction if MA is performed away from the optimum temperature
(\SI{840}{\celsius}). However, a SD morphology difference
(\rfig{fig:03}b and \rfig{fig:08}b) was observed between sample
$\smpl{B}$ and sample $\smpl{J}$. Only some areas of sample $\smpl{J}$
exhibit a mosaic pattern. The faceting of $\alpha_1$ rods is not as well
defined as those in sample $\smpl{B}$. Moreover, between regions with
a mosaic pattern, there are clusters of the $\alpha_1$ phase with smaller
diameters, as indicated by the white arrow in \rfig{fig:08}b. Viewing
samples along the longitudinal direction showed that the SD phase
morphology is similar to that observed along the transverse direction,
which obviously suggests that the $\alpha_1$ phases in samples
($\smpl{I}$, $\smpl{J}$, and $\smpl{K}$) are oriented isotropically,
as expected without an external field biasing effect.
\begin{figure}[bht]
\centering
\includegraphics[width=0.48\textwidth,clip]{fig09.pdf}%
\caption{APT reconstruction of the alnico 8 sample annealed at
\SI{840}{\celsius} for \SI{90}{\second} with (a) an external
magnetic field and without (b) an external magnetic field. The
initial nanostructures evolution is viewed in perpendicular
$\langle001\rangle$ directions, together with 30 at.\si{\percent} Fe
isoconcentration surfaces and 10 at.\si{\percent} Cu
iso-surfaces. The dark arrow in (a) indicates two orthogonally
aligned $\alpha_1$ rods. }
\label{fig:09}
\end{figure}
The effects of external magnetic field on the morphology of SD phases
at the early stage of SD are shown in the 3D APT reconstructed
iso-surfaces in \rfig{fig:09}. The non-field annealed (Non-MA) sample
has $\alpha_1$ rods with their longer axis aligned along all three
orthogonal $\langle100\rangle$ directions
(\rfig{fig:09}a)~\cite{guo.mm2016}. In contrast, in the MA sample, the
$\alpha_1$ phase is much longer and preferentially elongates along the
$\langle100\rangle$ direction closest to the external magnetic field
direction (\rfig{fig:09}b). The APT bulk composition from the center
of the $\alpha_1$ and $\alpha_2$ phases of both samples
(\rtbl{tbl:composition}) shows similar concentrations in the two
$\alpha_1$ and $\alpha_2$ phases, indicating that the field mainly biases the
morphology of the SD phases and has no obvious effects on their
chemistry.
\begin{table}[hbtp]
\footnotesize
\centering
\caption{Phase compositions of the $\alpha_1$ and $\alpha_2$ phases in MA and
{Non-MA} samples.}
\label{tbl:composition}
\bgroup
\def\arraystretch{1.2
\begin{tabular}[0.9\linewidth]{clcccccc}
\hline
\\[-1.1em]
Phase & Sample & Al & Ti & Fe & Co & Ni & Cu\\
\\[-1.1em]
\hline
\\[-1.2em]
\multirow{2}{*}{$\alpha_1$} & MA & 4.28 & 1.01 & 52.47 & 36.48 & 4.55 & 1.22\\
& Non-MA & 4.35 & 1.04 & 51.92 & 36.65 & 4.79 & 1.22\\
\multirow{2}{*}{$\alpha_2$} & MA & 22.45 & 14.98 & 13.08 & 33.36 & 14.29 & 1.86\\
& Non-MA & 22.06 & 14.52 & 13.72 & 33.08 & 14.69 & 1.89\\
\\[-1.2em]
\hline
\end{tabular}
\egroup
\end{table}
\section{Discussion}
The results presented clearly show that the magnetic properties of
alnico are closely related to the morphology of the SD phases, which
in turn are sensitive to both thermal MA and post-MA processing. The
best magnetic properties are promoted by well-aligned rods of the
$\alpha_1$ phase in a uniformly spaced mosaic pattern. This pattern is
observed only if a magnetic field is applied in a very narrow
temperature range for a limited time (e.g. $\sim$\SI{840}{\celsius}
for \SI{10}{\minute} for this alloy). Subsequent low temperature
anneals are required to further increase the coercivity, which can be
increased as much as \SI{30}{\percent}. There are no apparent
morphological changes corresponding to the increase in the coercivity,
except for the elongation of the Cu-enriched phase.
These results suggest that there is likely several mechanisms at play
in addition to SD. For alloys with a miscibility gap, a two-phase
mixture can be formed by either classical nucleation and growth (NG)
or by an SD path~\cite{laughlin.asmhandbook1985}. The NG path arises
with small undercooling (low supersaturation) and requires relatively
large localized composition
fluctuations~\cite{laughlin.asmhandbook1985}. The size of the
nucleated second phase grows with phase separation time. The SD path
occurs at large undercooling and the transformation occurs
homogeneously throughout the alloy by continuous growth of initially
small compositional
fluctuations~\cite{laughlin.asmhandbook1985}. Theoretical studies on
magnetic aging of the SD alloy showed that early in the SD process,
the composition wave length is determined by the interplay of chemical
free energy, elastic energy, and magnetic
energy~\cite{cahn.acta1961,cahn.jap1963}. The elastic energy
suppresses the system SD, while the magnetic energy suppresses the
development of composition waves along the external field
direction~\cite{cahn.jap1963}. The magnetic energy can have a serious
effect on the SD process only when the alloy is annealed just below
the Curie temperature, where the change of magnetic energy is the
largest~\cite{cahn.jap1963}. Our experiments showed that the MA
process is very effective in forming uniformly spaced rods only when
MA is conducted at a specific temperature range
($\sim$\SI{840}{\celsius} for this alloy, samples $\smpl{B}$ and
$\smpl{E}$ in \rtbl{tbl:ht-condition}), which tends to support this
prediction.
\begin{figure}[htb]
\centering
\includegraphics[width=0.44\textwidth,clip]{fig10.pdf}%
\caption{Schematic of phase diagram of alnico 8 by considering it as a
quasi-binary system with $\alpha_1$ and $\alpha_2$ phases.}
\label{fig:10}
\end{figure}
Considering alnico as a pseudo-binary alloy that consists of $\alpha_1$
and $\alpha_2$ phases with a miscibility gap, its phase diagram may be
illustrated schematically in \rfig{fig:10}. The chemical spinodal line
is defined by the inflexion point of the isothermal free energy ($G$)
composition curve \newline ($\partial G^2/\partial C^2$=$0$). The
coherent SD line is defined by the displacement of the SD curve due to
strain energy. Since the SD process is homogeneous and the SD phase is
typically uniform within a grain~\cite{laughlin.asmhandbook1985}, it
is reasonable to assume that a homogeneous structure with a periodic
pattern in the alloy implies that the alloy decomposes along the path
of SD. Transverse observations from TEM indicate that samples
$\smpl{A}$, $\smpl{B}$, and $\smpl{I}$ (MA @ \SI{840}{\celsius} or
lower) follow a SD path, while samples $\smpl{C}$, $\smpl{J}$, and
$\smpl{K}$ (MA@ $>$ \SI{840}{\celsius}) follow a NG path. As a result,
the SD coherent spinodal line shifts upward due to application of an
external field, as outlined in \rfig{fig:09}. This may be because the
application of an external magnetic field promotes elongation of the
$\alpha_1$ phase along the field direction substantially at
\SI{840}{\celsius}, which can slightly modify the strain energy of the
system and change the phase separation mechanism perpendicular to the
field direction from NG to SD.
The similarity in the SD morphologies of samples treated with MA only
(samples $\smpl{A}$, $\smpl{B}$, and $\smpl{C}$ in
\rtbl{tbl:ht-condition}) and FHT processes (samples $\smpl{D}$,
$\smpl{E}$, and $\smpl{F}$ in \rtbl{tbl:ht-condition}) reveals that
the shape of the $\alpha_1$ phase is established during the MA stage by
SD. Drawing may further ``tune'' the chemistry differences between the
$\alpha_1$ and $\alpha_2$ phases, which can cause larger magnetization
differences between $\alpha_1$ and $\alpha_2$ phases. This is because at a
much lower temperature (draw annealing), the long-range kinetic
effects during MA are suppressed, but the chemical driving forces
still allow for small, but important chemical separation to
continue. As a result, there is a large increase in $H_\text{cj}$ with
FHT. Details on the morphology and chemistry changes, as well as the
evolution of Cu-enriched phase during MA at \SI{840}{\celsius} and the
drawing process, will be reported elsewhere.
The changes of the magnetic properties and microstructures with
different MA time at \SI{840}{\celsius} (samples $\smpl{G}$,
$\smpl{B}$, and $\smpl{H}$) suggest that the mosaic structure set by
the MA process after $\sim$\SI{10}{\minute}, which gives the best
magnetic properties, is a metastable structure. The $\alpha_1$ phase will
grow into a different morphology if the sample is subjected to MA for
a long enough time. The mosaic structure may be temporarily stabilized
by the Cu-enriched phase at the corners of $\alpha_1$ rods, which
introduces a large amount of strain and constrains the growth of the
$\alpha_1$ phase. Observation of the Cu-enriched phase with a smaller size
in sample $\smpl{H}$ tends to support the hypothesis that the $\alpha_1$
phase can grow larger only after the Cu-enriched phase is broken into
smaller particles by extended magnetic annealing.
\begin{figure}[htb]
\centering
\includegraphics[width=0.42\textwidth,clip]{fig11.pdf}%
\caption{Arrangement of $\alpha_1$ phase with (a) and without (b) an
external magnetic field. Red arrows indicate the direction of
magnetic moment inside the $\alpha_1$ phase.}
\label{fig:11}
\end{figure}
Particle elongation during MA seems to be driven by magnetostatic
interaction, which is the same provider of magnetic shape anisotropy
and coercivity in alnico. For shape anisotropy, the shape of the
elongated magnetic particles is fixed and the magnetization tends to
align along the longer axis of the particles to lower magnetostatic
energy. Similarly, for particle elongation during MA, the
magnetization vectors of the magnetic phase are fixed along the field
direction (For simplicity, we assume that a saturated field is applied
along the [001] direction.) and the magnetic phases tend to evolve
into elongated shapes with their long axes along or close to the field
direction.
The two major phases in alnico, $\alpha_1$ and $\alpha_2$, have coherent
lattices and the corresponding interfacial energy is relatively small
~\cite{cahn.jap1963,zijlstra.zap1962,neel.cr1947}. This results in a
high sensitivity of the microstructure to the external magnetic field
during MA, especially at the onset of SD and just below the Curie
temperature, when the composition difference between two phases is
still small, but the magnetization difference may be large. The
composition waves along the field direction are thus effectively
suppressed at the initial stage of SD~\cite{cahn.jap1963}. Thus, a
well-defined mosaic pattern is developed within a short time during MA
(\rfig{fig:11}a) perpendicular to the field direction. After the SD
phases become fully developed, the composition difference between the
two phases and their interfacial energy both increase. For an extended
MA, $\alpha_1$ rods may become larger and more round (with both $\{100\}$
and $\{110\}$ faceting), as shown in \rfig{fig:06}c, to minimize the
interfacial energy. In the absence of an external magnetic field, the
$\alpha_1$ phase may still elongate along a $\langle100\rangle$ direction
and form closed loops to minimize the magnetostatic energy. Elongated
$\alpha_1$ phases embedded in $\alpha_2$ phases still form in three
dimensions, but the overall microstructure is isotropic, as
illustrated in \rfig{fig:11}b.
\section{Conclusions}
The correlations between magnetic properties, microstructure and HT
conditions have been comprehensively studied for alnico. The
morphology of SD phases is very sensitive to MA temperature and time
for the alnico composition in this study. A structure formed by
uniformly spaced, $\sim$\SI{40}{\nm} (diameter), $\{110\}$ faceted
$\alpha_1$ rods in a $\alpha_2$ matrix can only be achieved for a narrow MA
temperature range ($\sim$\SI{840}{\celsius}), and within limited MA
time ($\sim$\SIrange{3}{10}{\minute}), resulting in suitable magnetic
properties, similar to commercial magnets of this type. The external
magnetic field aids the formation of a well-faceted mosaic structure
only at the optimal temperature ($\sim$\SI{840}{\celsius}). MA at
different temperatures causes elongation of the $\alpha_1$ phase as well,
which improves $B_\text{r}$ of alnico, but has no obvious effect on
$H_\text{cj}$. Low temperature drawing can further increase the alloy’s
coercivity by modifying local chemistry in the $\alpha_1$ and $\alpha_2$
phases and their interfaces, without substantially changing the $\alpha_1$
rod diameter and spacing.
\section*{Acknowledgments}
Research was supported by U.S. DOE, Office of Energy Efficiency and
Renewable Energy (EERE), under its Vehicle Technologies Office,
Electric Drive Technology Program, through the Ames Laboratory, Iowa
State University under contract DE-AC02-07CH11358. APT research was
conducted as part of a user project at ORNL’s Center for Nanophase
Materials Sciences (CNMS), which is a DOE Office of Science User
Facility.
\section*{References}
|
\section{Introduction}
Surface nanobubbles are gaseous nanoscopic entities on an immersed surface that are less than 1 $\mu$m in height \cite{lohse2015rmp}. The presence of surface nanobubbles has significant implications to many physical and chemical processes at solid-liquid interfaces, for example, in fabrication of bubble-templated nanostructures \cite{fan2005am,hui2009,darwich2011}, propelling microscopic swimmers fuelled by catalytic reactions \cite{paxton2004,wilson2012}, light conversion by plasmonic effects\cite{fang2013}, heterogeneous cavitation under ultrasound\cite{belova2011,belova2013}, onset of boiling on microscopic scale \cite{zhang2014prl}, and flotation in mineral processing \cite{hampton2009me,hampton2010}. Hence the stability and the dynamics of surface nanobubbles is interesting from both a fundamental and from an applied point of view \cite{lohse2015rmp}.
Experimental studies have shown that surface nanobubbles have long lifetime, even for only a mild oversaturation level supplied to the system\cite{parker1994,stevens2005,lou2000,zhang2007prl,zhang2013JSR}. When a bubble dissolves or grows, the morphological features of the bubble suggest a stick-slip motion of the contact line \cite{zhang2013langmuir,german2014}. The origin of pinning is attributed to intrinsic features of the solid surface. Different theories were proposed to explain the stability of surface nanobubbles \cite{ducker2009,brenner2008,weijs2013prl,petsev2013}. Recently
\citeauthor{lohse2015}\cite{lohse2015} provided the exact calculation for the stability of a single surface nanobubble\cite{lohse2015}. Their derivation assumes pinning and builds on, not more than only the diffusion equation,
Henry's law, and the Laplace equation. Their calculation reveals that contact line pinning and the gas oversaturation $\zeta>0$ in the bulk liquid are crucial for the equilibrium of the single surface nanobubble to be stable. Here $\zeta= \frac{C_\infty}{C_s} -1$ is the gas oversaturation, $C_\infty$ the gas concentration, and $C_s$ the gas solubility.
For a given gas oversaturation $\zeta>0$, there exists an equilibrium in which the outflux of gas molecules from the nanobubble due to the large Laplace pressure
is compensated by the influx into the bubble due to gas oversaturation, i.e., there is no net flux.
Though the flux of gas particles out or into the nanobubble is not spatially uniform along the interface as shown by
\citeauthor{weijs2012prl}\cite{weijs2012prl} and \citeauthor{yasui2015}\cite{yasui2015},
the {\it net} flux integrated over the whole interface is zero.
The contact angle of the surface nanobubble is not given by Young's equation, but determined by the equilibrium,
\begin{equation}
\sin\theta_e = \zeta\frac{L}{L_c}, \label{eq:eqangle}
\end{equation}
\nobreak where L is the pinned length of the footprint and $L_c = 4\gamma/P_0$ a capillary length scale,
with $\gamma$ the surface tension and $P_0$ the ambient pressure as shown in the sketch in Figure \ref{fig:simbox}.
Meanwhile, molecular dynamics (MD) simulations have provided important insight into the dynamics of surface nanobubbles, in particular under the conditions that are difficult to achieve in experiments. \citeauthor{weijs2012prl}\cite{weijs2012prl} performed molecular dynamics (MD) simulations of surface nanobubbles without any heterogeneities and found that nanobubbles are not stable.
\citeauthor{liu2013jcp}\cite{liu2013jcp,liu2014jcp2} showed with the help of kinetic lattice density functional theory and MD simulations
that contact line pinning on geometrical heterogeneities also leads to stable surface nanobubble. However, it remains unknown how exactly the oversaturation in the liquid and chemical patterns with a nanoscale dimension on a surface can mediate the stability of nanobubbles.
Moreover, given that it is extremely difficult to confirm eq \ref{eq:eqangle} experimentally, in this work we want to confirm
it with the help of MD simulations.
We will present our simulations of the dynamics (stability, growth, and dissolution) of a single surface nanobubble on a chemically patterned surface, in response to different oversaturation levels. The chemical patterns act as pinning sites enabling us to study the growth or dissolution of surface nanobubbles on a more realistic solid substrate,
exhibiting pinning and de-pinning of the contact line.
\section{Approach and methodology}
Molecular Dynamics (MD) simulations were performed to simulate the nanobubble on a solid substrate
for which we used the open source code GROMACS\cite{gromacs}.
We have used four types of particles or molecules in our simulations: two types of solid particles ($\rm{S}$ and $\rm{S}_P$), which
remain fixed in a fcc lattice during the whole simulation,
and two types of moving particles, Liquid (L) and Gas (G). The $\rm{S}_P$ particles form the pinning sites, and have
different interaction strength towards the two types of moving particles. The L particles form a bulk liquid phase (and hence we refer to these
as "liquid particles") as the system temperature and pressure are below the critical point of L particles whereas the G particles form a bulk gaseous phase (to which we refer as "gas particles")
because the critical point for G particles is much below the thermodynamic conditions at which we are performing our simulations.
A typical simulation box is shown in Figure \ref{fig:simbox}.
\begin{figure}
\includegraphics[width=0.95\textwidth]{Fig1.eps}
\caption{
Left: Schematic of a single surface nanobubble on a hydrophobic surface and immersed in an oversaturated solution. L is the footprint diameter
of the bubble, $\theta$ the contact angle, R the radius of the curvature of the bubble and $C_\infty>C_s$ the concentration of the gas far away from the interface. Right:
A typical simulation box which consists of four kind of particles. Red particles ($\rm{S}$) form the solid surface, yellow particles
($\rm{S}_P$) form pinning sites, blue particles (L), which are predominantly in the liquid phase, and green particles (G), which are
predominantly in the gas phase. The nanobubble shown in the figure is cylindrical in shape
with the x-axis along the length of the cylinder.}
\label{fig:simbox}
\end{figure}
The interaction between the particles is described by a Lennard-Jones potential:
\begin{equation}
\phi^{\rm{LJ}}_{ij} (r) = 4\epsilon_{ij}\Big[\Big({\frac{\sigma_{ij}}{r}}\Big)^{12} - \Big({\frac{\sigma_{ij}}{r}}\Big)^6\Big], \label{eq:lj}
\end{equation}
\nolinebreak in which $\epsilon_{ij}$ is the interaction strength between particles $i$ and $j$, and $\sigma_{ij}$ is the characteristic
size of particles.
The potential is truncated at a relatively large cut-off radius ($r_{c}$) of $5\sigma_{LL}$. The time step for updating the particle
velocities and positions was set at $dt = \sigma_{LL}\sqrt(m/\epsilon_{LL})/400$, where $m$ is mass of the liquid
particles and $\epsilon_{LL} $ is the Lennard-Jones interaction parameter for the liquid particles.
The time step has been chosen such that its value is sufficiently less than the shortest time scale available in the system\cite{frenkel2002}.
Periodic boundary conditions have been employed in all three directions, which suggest that same solid substrate is also present above the liquid layer.
Simulations have been performed in a $NPT$ ensemble where the temperature is fixed at $300 K$,
which is below the critical point for the Lennard-Jones parameters ($\sigma_{LL}$, $\epsilon_{LL}$) that we have set for the liquid particles.
Semi-isotropic pressure coupling is used for maintaining constant pressure which means that the simulation box can expand or contract
only in the z-direction to keep the pressure constant. This has been done to avoid the creation of gaps along the
solid surface boundaries in the x and y directions.
Simulations were performed in a quasi-2D manner in which the length of the simulation
box along the x-axis is considerably smaller than the lengths in the other two directions, which means the shape of the nanobubble is cylindrical
instead of spherical. Simulations of cylindrical nanobubbles save computation time; of course
the effect of the modified shape on the
Laplace pressure is taken into consideration.
The complete set of Lennard-Jones parameters that we have used in our
simulations are given in table \ref{tab:ljparam}. The typical system size is $5.6\times40\times24~\mbox{nm}^3$ in x, y and z direction respectively, where
we note that the length of the z dimension changes during the simulation
to keep the pressure constant.
\begin{table}
\centering
\begin{tabular} { l | c c }
$\mathbf{i-j}$ & $\mathbf{\sigma_{ij}}$, \rm{nm} & $\mathbf{\epsilon_{ij}}$, \rm{kJ/mol} \\ \hline
$\rm{S}$-L & 0.34 & 1.8 \\
$\rm{S}_P$-L & 0.34 & 1.5 \\
$\rm{S}$-G & 0.40 & 2.0 \\
$\rm{S}_P$-G & 0.40 & 5.0 \\
L-G & 0.40 & 1.55 \\
G-G & 0.46 & 0.8 \\
L-L & 0.34 & 3.0 \\ \hline
\end{tabular}
\caption{Value of various LJ parameters used in the MD simulations} \label{tab:ljparam}
\end{table}
In the initial configuration, gas and liquid particles are arranged in a fcc lattice above the solid substrate.
Initially, the liquid near the surface is highly oversaturated with gas particles in order to aid the bubble nucleation on the surface which decreases the equilibration time,
which is around $5\times10^{7}$ time steps (around $\sim$9 ns).
For simulations at different pressures, we have used the final configuration of the previous simulation as an initial configuration to save computation time.
In Figure \ref{fig:simbox} we show a typical equilibrium profile of a nanobubble on a chemical heterogenous surface.
After the equilibrium has been reached, the time-average density field of liquid particles is calculated,
correcting for the center of mass motion in the lateral direction.
Quantities like the radius of curvature of the bubble and the contact angle are obtained by
fitting a circle to the iso-density contour of $0.5$ of the normalised density field,
$\rho^{*}(r)$, defined as $\rho^{*}(r) = \frac{\rho(r)-\rho_{V}}{\rho_{L}-\rho_{V}}$, where $\rho_{V}$ and $\rho_{L}$ are the bulk vapour and liquid density,
respectively. Since the liquid very near to the solid
substrate is subject to layering, we have excluded
the density field in the range of $2\sigma_{LL}$ from the substrate for the circular cap fitting.
From the intersection of the circular fit with the substrate, the contact angle and the radius of curvature are evaluated (see figure \ref{fig:circlefit}).
In order to study the time evolution for these quantities, we have calculated these time-averages for subsequent subsets of 25000 time steps.
\begin{figure}
\includegraphics[width=0.99\textwidth]{circle_fit.eps}
\caption{Example of a spherical cap fit (black line) to the iso-density contour of 0.5 (points) for a surface nanobubble}
\label{fig:circlefit}
\end{figure}
The contact line is pinned by the pinning sites, which are formed by the solid particles with higher hydrophobicity ($\rm{S}_P$) as
shown by different colours in Figure \ref{fig:simbox}, which means that the interaction strength with gas particles is much higher than for the liquid particles.
Exact values for the interaction strengths are tabulated in table \ref{tab:ljparam}.
One can also notice from Figure \ref{fig:simbox} that the gas particles
accumulate near the solid surface. This is due to the
hydrophobic nature of the solid surface which strongly attracts the gas particles, as discussed
in \cite{dammer2006}.
The oversaturation of gas particles in the bulk liquid is given by $\zeta= \frac{C_\infty}{C_s}-1$, which involves the calculation of $C_\infty$ and $C_s$.
We have defined the concentration in this work as the ratio of gas particles to liquid particles in a certain amount of fixed volume, which is equivalent to the mole fraction of gas particles in the liquid.
The oversaturation $\zeta$ is controlled by changing the system pressure, since Henry's law dictates that the
solubility of gas in the liquid increases linearly with the system pressure.
In our simulations, first a nanobubble is formed on the chemically heterogenous surface by a NVT simulation,
and then the pressure coupling is switched on, to keep the solubility of gas particles in the bulk liquid fixed. To change the solubility, and hence the oversaturation level,
the system pressure is slowly increased or decreased which respectively leads to the dissolution or growth of the bubble. For dissolution, the system pressure is increased which results in
the increase of solubility. Because of the finite system size, an increase in solubility leads to the migration of gas particles from the nanobubble to the bulk liquid, which leads to the dissolution
of the nanobubble. In order to determine $C_\infty$, we have divided the liquid domain into concentric shells of thickness $0.5\sigma_{LL}$, concentric around the
nanobubble, and then calculated $C(r)$, the ratio of
gas to liquid particles in each shell (see Figure \ref{fig:conc}). $C_\infty$ is then calculated by averaging
$C(r)$ over the range of $7.5\sigma_{LL}$ to $25\sigma_{LL}$, where it is almost constant. In Figure \ref{fig:henry} we show $C_\infty$ as a
function of the pressure of the gas phase, which at equilibrium is equal to the sum of the bulk pressure and the Laplace pressure.
Our data are found to obey Henry's law, where we estimate Henry constant as $k_H = 1.43\pm0.027\times10^{-5}~\rm{bars}^{-1}$.
Note that the concentration of gas particles around the nanobubble is assumed to vary with radial distance only. In general, it varies in both radial and azimuthal direction
but it is a very weak function of the azimuthal angle when contact angle is around $\sim 90^{\circ}$\cite{popov2005}. In our simulations the contact angle is varying between $70^{\circ}$
and $90^{\circ}$, so the azimuthal angle dependence of the concentration field can be neglected.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{concentration.eps}
\caption{(a) Radial variation of concentration of gas particles around a nanobubble. The inset shows the concentration of gas particles near the interface.
(b) Variation of concentration of gas against the distance from the interface at few pressures. The schematic in the inset shows the range of radial distance over which the concentration is plotted.
}
\label{fig:conc}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.99\textwidth]{henry.eps}
\caption{Concentration of gas particles in the bulk phase as measures in our simulations, for different gas pressures. The straight line suggests that the concentration of the gas particles indeed follows Henry's law
$C_\infty = k_{H}P$.}
\label{fig:henry}
\end{figure}
\section{Results and Discussion}
\subsection{Pinning provides stability to surface nanobubbles}
We start with testing the prerequisite conditions for the stability of nanobubbles given by \citeauthor{lohse2015}\cite{lohse2015}.
They theoretically showed that pinning of three-phase contact line by hydrophobic heterogeneities in a gas oversaturated liquid leads
to stability of nanobubbles. We now show systematically by performing MD simulations that these pinning conditions are indeed sufficient to
get a stable surface nanobubble.
To this end we have performed four MD simulations at the same temperature and pressure, the time evolution of which is shown in Figures \ref{fig:stability}
(snapshots) and \ref{fig:timeevolution} ($\theta$, L , R, and H as function of time).
Figure \ref{fig:stability} (a) shows the time evolution of a nanobubble on a homogenous solid substrate, i.e. without any pinning sites, at a particular gas oversaturation, $\zeta>0$.
We find that the nanobubble in this simulation grows quickly due to large oversaturation of gas particles in the bulk liquid, so that after the few nanoseconds
it comes into contact with its periodic image, and ultimately forms a vapour liquid equilibrium with a planar interface.
It clearly shows the importance of heterogeneities on the solid surface without which it is impossible to form a stable nanobubble.
This behaviour is also shown in terms of the variation of $\theta$, L, R and H of the nanobubble with time in Figure \ref{fig:timeevolution}.
It is apparent from the plot that $\theta$, L, R and H increase rapidly with time and couldn't form a stable nanobubble.
Figure \ref{fig:stability} (b) shows the time evolution of a nanobubble on a chemical heterogeneous surface with hydrophobic pinning sites
at a particular gas undersaturation, $\zeta<0$. One can observe that the nanobubble slowly shrinks and dissolves with time.
The nanobubble is not stable because the surrounding liquid is undersaturated with gas particles and it shows
the "stick-jump" mechanism during the dissolution, which means that the contact line remains pinned to the hydrophobic pinning sites.
This "stick-jump" dissolution mode is evident from Figure \ref{fig:timeevolution}, especially from the variation of L and R with time.
This kind of motion of the contact line in the dissolution of nanobubbles is very similar to the dissolution mode exhibited by microdrops dissolving in an
another liquid\cite{dietrich2015}.
This simulation clearly demonstrates the necessity of having gas oversaturation ($\zeta>0$) for the stability of nanobubbles.
Figure \ref{fig:stability} (c) shows the time evolution of a nanobubble on a chemical heterogeneous surface with hydrophilic pinning sites
at a particular gas oversaturation, $\zeta>0$. Hydrophilicity is introduced by interchanging the value of $\epsilon_{S_{P}G}$ and $\epsilon_{S_{P}L}$.
In this case we have not observed pinning of the three-phase
contact line and the nanobubble eventually dissolves, although the dissolution time is quite high.
Though it looks like from Figure \ref{fig:timeevolution} that a stable nanobubble is formed but in reality it slowly (compared to
the undersaturation case) dissolves.
Time scale for dissolution of a bubble for unpinned contact line case is given by the classical theory of \citeauthor{epstein1950}\cite{epstein1950,lohse2015,ljunggren1997}, namely
\begin{equation}
\tau_{life} \approx \frac{R_{0}^2\rho_g}{3DC_s}, \label{lifetime}
\end{equation}
\nobreak where $R_0$ is the radius of curvature of the bubble, $\rho_g$ the gas density, and $D$ the diffusion coefficient of gas particles in the liquid, which are extracted
from MD simulation of the nanobubble with hydrophilic pinning sites.
Accoding to this eq \ref{lifetime}, the lifetime of the nanobubble in our case is calculated as $\sim 2.9\mu s$. Simulating as long times as $\mu s$ is still very challenging for MD simulations; therefore we could not show
the whole dissolution dynamics in this case.
However, one can infer from this simulation that only hydrophobic pinning sites lead to the pinning of the contact line and hence stable nanobubble.
Figure \ref{fig:stability} (d) shows the time evolution of a nanobubble on a chemical heterogeneous surface with hydrophobic pinning sites
at a particular gas oversaturation, $\zeta>0$. In this case, the nanobubble on the solid surface remains stable
throughout the simulation as chemical heterogeneities fixate the contact line, and the outflux due to Laplace pressure is balanced by the influx due to the
oversaturation of the gas particles in the bulk liquid.
From Figure \ref{fig:timeevolution}, it is clear that in tis case $\theta$, L, R and H remain constant with time and form a stable equilibrium.
These four simulations confirm that hydrophobic chemical heterogeneities and gas oversaturation ($\zeta>0$) are necessary and sufficient conditions for forming a
stable nanobubble. Note that also geometrical pinning sites lead to the stability of nanobubbles as already shown by \citeauthor{liu2014jcp2}\cite{liu2014jcp2}.
\begin{figure*}
\center
\includegraphics[width=\textwidth]{stability.eps}
\caption{Time evolution of a surface nanobubble in a $NPT$ ensemble at a temperature of $300 K$ and pressure of 26 bars (a) without chemical heterogeneities and
gas oversaturated liquid ($\zeta>0$), (b) with hydrophobic chemical heterogeneities and gas undersaturated liquid ($\zeta<0$), (c) with hydrophilic chemical heterogeneities and gas oversaturated liquid ($\zeta>0$), and
(d) with hydrophobic chemical heterogeneities and gas oversaturated liquid ($\zeta>0$) surrounding the nanobubble.}
\label{fig:stability}
\end{figure*}
\begin{figure*}
\center
\includegraphics[width=0.8\textwidth]{cangle_time.eps}
\caption{Time evolution of $\theta$, L, R and H of a surface nanobubble in the four cases shown in Figure \ref{fig:stability}.}
\label{fig:timeevolution}
\end{figure*}
\subsection{Stable surface nanobubbles at equilibrium}
In this section we study the stable configurations of surface nanobubbles at equilibrium at various gas oversaturation levels.
The primary aim of this section is to validate the expression for the equilibrium contact angle of a surface nanobubble derived by \citeauthor{lohse2015}\cite{lohse2015}.
In the previous section we have qualitatively shown the necessary conditions for forming a stable nanobubble. Here, we systematically and quantitavely study the
stable nanobubble at equilibrium with the help of MD simulations.
To this end, a nanobubble is nucleated on a chemical heterogenous surface with the help of local oversaturation
at a particular temperature and pressure. After reaching equilibrium, the system pressure is slightly increased which results in an increase of the solubility of gas particles in the bulk liquid.
After increasing the pressure, the simulation was run for another 21 ns to ensure equilibrium of the system.
As explained before, the increased solubility (and the fixed number of the gas particles in the system) will lead to the dissolution of the surface nanobubble.
This mechanism results in a sudden jump of the three phase contact line, at some particular pressure, towards a new pinning site, decreasing the
lateral length and suddenly increasing the contact angle of the nanobubble.
Figure \ref{fig:jump} shows the contact angle $\theta$, the footprint diameter L, the radius of curvature R of nanobubble, and its height H, as function of the bulk pressure $P_0$.
One can clearly observe the discontinuities in the variation which corresponds to
the jump of the three-phase contact line over the pinning sites.
The magnitude of change in the footprint diameter at discontinuities is equal to the distance between the pinning sites which confirms the "jump" of the
contact line on the pinning sites.
We would like to emphasise that for each data point shown in Figure \ref{fig:jump}, the system is at equilibrium for that particular pressure.
\begin{figure}
\includegraphics[width=0.8\textwidth]{jump1.eps}
\caption{Variation of $\sin\theta$, footprint diameter (L) , radius of curvature of the nanobubble (R), and height of the nanobubble (H) with the bulk pressure $P_0$.
Dashed lines show the sticking of the contact line between the two particular pinning sites
and solid lines show the jump of the contact line from one pinning site to another. }
\label{fig:jump}
\end{figure}
Note that $P_0$ is the bulk liquid pressure which is calculated in a liquid slab away from the interface. Although all the simulations are performed in a NPT ensemble,
where the pressure is controlled by semi-isotropic coupling, there is a large difference between the pressure set in GROMACS, and the bulk pressure $P_0$,
on which we elaborate in the appendix.
Theoretically, the equilibrium contact angle of a single surface nanobubble is given by eq \ref{eq:eqangle} $\sin\theta_e = \zeta L/L_c$.
Note that here we had to modify the definition of $L_c$ from $4\gamma/P_0$ to $2\gamma/P_0$, owing to the fact that
for cylindrical bubbles the Laplace pressure is equal to $\gamma/R$, instead of the usual value $2\gamma/R$ that holds for spherical bubbles.
Since Henry's law is obeyed in our simulations, we can write $C_\infty = k_{H}P$ and $C_s = k_{H}P_0$, where
$P$ is the pressure inside the nanobubble and $P_0$ is the ambient bulk pressure. Note that now $P_0$ is calculated from the actual MD simulations and
not evaluated from the expression as given in the appendix.
We have plotted $\sin\theta_{e}/L$ against the oversaturation $\zeta{P_{0}}$ in Figure \ref{fig:final_av},
which shows that the equilibrium contact angle obeys eq \ref{eq:eqangle}. Note that there are some data points
which are slightly away from the theoretical line, which is mainly due to the large errors in the pressure calculation.
Because of the relatively small system size, the instantaneous value for the pressure can deviate significantly from the mean which leads to large error bars\cite{gromacs}.
Another reason for this deviation is the assumption of the surface tension being independent of pressure.
The pressure inside the bubble directly affects the concentration of gas particles at the interface, which can have an influence on the magnitude of the surface tension.
The three points in the Figure \ref{fig:final_av} corresponds to the three radii of curvature or
footprint diameter levels as shown in Figure \ref{fig:jump}. Each of the three points is the average of all the small data points for a nanobubble pinned on two particular pinning sites, which is also
shown in Figure \ref{fig:final_av}.
Note that $\sin\theta_{e}/L$ is plotted against $\zeta{P_{0}}$ instead of $\zeta/L_{c}$, as we do not a priori know the value of the surface tension $\gamma$.
In fact we can use eq \ref{eq:eqangle}, to obtain an estimate for the surface tension $\gamma$. since the slope of the straight line in Figure \ref{fig:final_av} is $1/(2\gamma)$, which yields a surface tension $\gamma = 0.072\pm0.0048$ N/m.
\begin{figure}
\includegraphics[width=\textwidth]{final_average.eps}
\caption{Variation of $\sin\theta/L$ against oversaturation the $\zeta{P_{0}}$. Small black data points represent the complete dataset shown in Figure \ref{fig:jump} while the three red points
corresponds to the average of all data points for a nanobubble pinned on two particular pinning sites. The three data points refer to the three levels of L seen in Figure \ref{fig:jump}.}
\label{fig:final_av}
\end{figure}
\section{Conclusions}
Molecular dynamics simulations were performed for single surface nanobubbles on a chemical heterogenous surface and in a gas oversaturated liquid. Heterogeneities act as pinning sites
which fix the three-phase contact line and stabilise the equilibrium, in which the gas influx due to the gas oversaturation is balanced by the outflux due to the Laplace pressure.
The theoretically predicted conditions\cite{lohse2015} for stable single surface nanobubble were tested with MD simulations, finding full confirmation.
We have found that only hydrophobic heterogeneities lead to the pinning of the contact line and not the hydrophilic ones.
In addition, we have also studied the dissolution of surface nanobubbles on a chemical heterogenous surface in a gas undersaturated liquid.
The dissolution of a surface nanobubble shows "stick-jump" behaviour which is very similar to the behaviour shown by the dissolution of microdrops in another liquid\cite{dietrich2015,dietrich2016,liu2013pre}.
We have also simulated the stable nanobubble at various levels of oversaturation and showed the variation of the contact angle, the footprint diameter, the radius of curvature and the height of the nanobubble
with system pressure.
The equilibrium contact angle is following the analytical expression calculated in ref. \cite{lohse2015},
which herewith we thus confirm.
The next step will be to simulate interacting surface nanobubbles.
We thank Shantanu Choudhary for assistance with parallelizing some of the codes, Michiel van Limbeek for fruitful discussions and Sander Wildeman, Varghese Mathai, Poorvi Shukla for constructive comments on the manuscript,
SURFsara (funded by NWO) for providing the computational facilities for the simulations, and FOM
for financial support.
\section{Appendix}
In this section we describe the calculation of $P_0$, that we have used in Figures \ref{fig:jump}. We have performed NPT
simulations in which the pressure is prescribed. Yet, there is a clear difference between the input pressure and the actual bulk liquid pressure, $P_0$ that we have used.
The pressure in any MD simulation with pair interaction $\phi_{ij}(r)$ can be calculated from:
\begin{equation}
P = \rho{k_{B}}T + \frac{1}{3V}\left\langle\sum_{i<j}\frac{d\phi_{ij}}{dr_{ij}}.{r_{ij}} \right\rangle
\label{eq:press}
\end{equation}
As one can notice from the eq \ref{eq:press}, that calculation of pressure involves the interaction between all the possible pairs of particles which include
the interaction between solid and moving particles also, which is thus included in the pressure that is set for GROMACS.
However, due to the finite size of the system, this interaction between fixed solid and moving particles can have a huge effect on the overall pressure.
In order to evaluate the "true" bulk pressure $P_0$, we have excluded this interaction from solid
particles, that is, $P_0$ is calculated by considering only the interactions in a rectangular slab
above the bubble. In the calculation of $P_0$, we have also considered the interaction between the pairs of $r_{ij}$ which are
intersecting the boundary planes. Relation between the input pressure in GROMACS and $P_0$ is shown in Figure \ref{fig:p0}.
\begin{figure}
\includegraphics[width=\textwidth]{pressure_slab1.eps}
\caption{Variation of $P_0$ with input pressure in GROMACS}
\label{fig:p0}
\end{figure}
It can be noticed that $P_0$ is directly proportional to the input pressure in GROMACS and a straight line is fitted to get the relation between the two,
which is given as $P_0 = 2.73565{P_{gro}} + 54.9801$. $P_0$ in Figure \ref{fig:jump} is calculated from the input pressure using this relation, whereas
in Figure \ref{fig:final_av}, $P_0$ is the actual calculated value from the MD simulation. Now the question arises,
why we chose to use this relation instead of using $P_0$ directly calculated from the simulations. From Figure \ref{fig:p0}, it is clear that on an average $P_0$ is increasing
linearly with input pressure but there are still some local fluctuations in the pressure which overcasts the jump in $\theta$, L, R, and H of surface nanobubble shown in Figure \ref{fig:jump}.
$P$, used in Figure \ref{fig:final_av} for calculation of $\zeta$, is the pressure inside the nanobubble which is calculated by considering
the virial function of all pairs of particles within the bubble and half the virial function of the pairs whose
$r_{ij}$ vector is intersecting the bubble boundary.
\bibliographystyle{unsrtnat}
|
\section{Introduction}
The relative homogeneity of Type Ia supernovae (SN~Ia) has allowed them to serve as precise
extragalactic distance indicators after the application of empirically derived relations for light curve shape \citep{phillips93} and color \citep{tripp98}. The similarity among the vast majority of SNe~Ia suggests a commonality among their progenitor stars and their explosion physics.
SNe~Ia are generally thought to be the complete disruption of Chandrasekhar-mass white dwarfs (WDs) which undergo a thermonuclear runaway process of explosive carbon and oxygen burning.
Despite a large population of `normal' SNe~Ia, increasing numbers of apparently similar events, but with spectroscopic and light curve peculiarities have been discovered. It is not clear whether such under- \citep[e.g.][]{filippenko92b,leibundgut93, turatto96} and over-luminous \citep{ruizlapuente92,phillips92,filippenko92a, howell06} events represent extreme extensions of the SN Ia progenitor model or are distinct in their progenitor stars and explosion mechanisms.
One particular event, SN~2002cx, also appeared to fall into the category of SN~Ia, following the traditional SN classification scheme \citep{filippenko97}, but garnered the title of the `most peculiar known' SN~Ia \citep{li03}.
With early spectral similarity to the over-luminous SN~1991T-like events -- consisting of conspicuous high-ionization spectral features associated with \ion{Fe}{ii} and \ion{Fe}{iii} -- but being relatively faint at peak ($M_R \sim -17.7$~mag), SN~2002cx prompted the classification of other `SN~2002cx-like events' \citep[e.g.,][]{jha06, narayan11}. A comprehensive study of all objects designated as SN~2002cx-like events (hereafter SNe~Iax) was presented by \citet{foley13}. The sample of
has since expanded and a recent review by \citet{jha17} discusses the current status of SN~Iax observations and theory.
The primary distinctions of SNe~Iax from normal SNe~Ia are fainter peak magnitudes, lower ejecta velocities, no secondary near-infrared maxima (with light curves peaking in optical before near-infrared, unlike normal SNe~Ia), and late time spectra that do not become fully nebular \citep[see][for more in depth discussion]{jha06, foley13, foley16}.
However even within the events classified as members of this group there appears a large degree of diversity.
The low ejecta velocities of SNe~Iax ($2000-8000$ cf. $10000-15000$~km~s$^{-1}${} for normal SNe~Ia around peak) implies that their explosions are of significantly lower energies, and their faint peak magnitudes indicate production of smaller amounts of radioactive $^{56}$Ni{} ($0.003-0.27$ cf. $\sim 0.4-1.0$~M$_\odot${} for normal SN Ia, e.g. \citealt{stritzinger06, childress15}). One of the leading scenarios to explain SNe~Iax is that they are deflagrations of Chandrasekhar mass WDs \citep[e.g.][]{branch04, phillips07, jordan12, kromer13} that have accreted helium-rich material from a companion star \citep{foley09, liuzw15a}. Due to the extension of SNe~Iax to low explosion energies and inferred ejecta masses, there is indirect evidence that at least some fraction of these SNe could leave bound remnants.
Support for the helium-accreting WD scenario has come from direct progenitor constraints and environmental constraints. \citet{mccully14} present the detection of a blue source at the location of SN~2012Z that is similar to the Galactic helium nova V445 Pup and limits for SN~2014dt provided by \citet{foley15b} are also consistent with a similar progenitor system. \citet{liuzw15b} suggest a hybrid CONe WD + He-star as the most favourable progenitor system for SN~2012Z. The delay time to explosion for WD+He-star systems has been investigated by \citet{wang13, wang14} and \citet{liuzw15a} who found a timescale of $\sim$100~Myr, in good agreement with observational constraints on their ages. \citet{lyman13} found that SN~Iax host galaxy types closely match one drawn from the star-formation (SF) rate density of the local universe and the environments of SNe~Iax trace the SF of their hosts at a similar level to core-collapse SNe~II with expected delay times of tens-hundred Myr. This statistical estimate of their environment ages was backed up through analyses of the local stellar populations (SP) of SNe 2008ha \citep{foley14} and 2012Z \citep{mccully14}, which found environments harbouring young SP with ages of 10--80~Myr.
Consistent with their young environments, it has also been suggested that some SNe~Iax may be due to the collapse of massive stars. One particular example that has drawn attention is SN~2008ha. This was an extremely weak SN, even in comparison to other SN~Iax, producing 0.003~M$_\odot${} of $^{56}$Ni{} and possibly ejecting only a few tenths~M$_\odot${} of material \citep{valenti09,foley09}, although it is not alone in this regard \citep[e.g., SN~2010ae;][]{stritzinger14}. Massive star models explored to explain these very low-energy and -luminosity events include the collapse of very massive Wolf-Rayet (WR) stars with significant fall-back onto a nascent black hole \citep{valenti09, moriya10} and stripped-envelope electron-capture (EC) SNe \citep{pumo09, moriya16}. The scenario of very low-luminosity SNe accompanying massive stellar collapse may be appealing to explain apparently SN-less gamma ray bursts \citep[GRBs, e.g.][]{fynbo06, dellavalle06, michalowski16}, if indeed they are related to other long-duration GRBs (LGRBs), and thus a result of the collapse of massive stars \citep[see, e.g.,][]{gehrels06,galyam06}.
Adding further diversity to the sample we note that for one event, SN~2008ge, no evidence for a young SP at the explosion site, or within the S0 host galaxy, was found by \citet{foley10b}. Furthermore, for two events, SNe~2004cs and 2007J, helium (and potentially hydrogen) was detected in their spectra. Doubt as to their validity as SNe~Iax members has been raised by \citet{white15} who classify them instead as core-collapse Type IIb SNe, although this was refuted by \citet{foley16}.
It is apparent that the group of SNe that have been labelled SN~Iax are a heterogeneous group, and multiple progenitor systems and explosion mechanisms could be responsible \citep{foley13, jha17}.
The sample has now grown to a size where statistical studies can elucidate further information on the progenitor systems, beyond characterising individual examples. Environmental diagnostics have played a central role in the understanding of other well-known transient types \citep[see][and references therein]{anderson15} and, in particular, the Multi-Unit Spectroscopic Explorer \citep[MUSE;][]{bacon10} instrument mounted at the Very Large Telescope (VLT) is revolutionising the way such studies are performed. MUSE is an integral field spectrograph (IFS) providing spatial and spectral information at the transient explosion site and across the host galaxy system in a single pointing \citep[e.g.,][]{galbany16, prieto16, kruehler17, izzo17}.
In this paper we perform spectroscopic environmental measurements for a large number of SNe Iax using both IFS and long-slit data from VLT/MUSE and Nordic Optical Telescope/Andalucia Faint Object Spectrograph and Camera (NOT/ALFOSC), thereby providing additional constraints on the nature of the progenitor systems and investigating whether their wide-ranging explosion characteristics are matched by large spreads in environmental measures. In \cref{sec:obs} we present our observations. Our methods for the MUSE-observed sample are given in \cref{sec:musemethods} and in \cref{sec:notmethods} we highlight similarities and differences for our NOT-observed sample's analysis. Results are presented and discussed in \cref{sec:results,sec:discussion}, respectively. Throughout the paper we use $\log(\textrm{O}/\textrm{H})+12 = 8.69$ dex as the solar abundance \citep{asplund09} and adopt the cosmological parameters $H_0 = 73.24$~km~s$^{-1}${}~Mpc$^{-1}$ \citep{riess16} and $\Omega_\text{m} = 0.3$.
\section{Observations and Data reduction}
\label{sec:obs}
The current number of SN Iax discovered is around 50 events
\citep{jha17}.
Our sample consists of 37 SNe Iax host galaxies observed with VLT/MUSE or NOT/ALFOSC.
The hosts span a redshift range of 0.004 to 0.08. Membership of our sample as SN Iax up to SN~2012Z is presented in \citet{foley13}. Additionally we add SN~2013dh \citep{jha13}, SN~2013en \citep{liuzw15c}, SN~2013gr \citep{hsiao13}, SN~2014ey\footnote{\url{https://wis-tns.weizmann.ac.il/object/2014ey}}
({\em Carnegie Supernova Project}, in prep), SN~2014ck \citep{tomasella16}, SN~2014dt \citep{ochner14}, SN~2015H \citep{magee16}, PS~15aic \citep{pan15}, PS~15csd \citep{harmanen15}, SN~2015ce \citep{balam15} and SN~2016ado \citep{tonry16}.
Here we detail the observations taken for our sample on the two instruments.
\subsection{VLT/MUSE}
\label{sec:museobs}
Observations of the host galaxies were carried out in a MUSE programme running between September 2015 and March 2016 apart from the host of SN~2002bp, which was observed as part of an earlier programme (as this galaxy has also hosted other SNe) in April 2015.
The time lag between the SN and our observations is more than two years for all but two events, SNe 2014dt and 2015H. In both these cases it is possible to identify broad forbidden line emission arising from the SN ejecta (see \citealp{foley16} for an overview of SN~Iax late-time spectra). These are analysed further in \cref{app:14dt15H} where the late time SN spectra and any impact on our environmental analyses are presented. In essence, we do not consider the presence of faint underlying signal from these hydrogen-poor, non-interacting events to significantly impact our emission-line based environmental analyses.
The strategies for sky-subtraction with MUSE depended on the angular size of each host galaxy. For galaxies that did not cover the field of view of MUSE, blank sky spaxels within the on-source exposures could be used to create the sky-spectrum to subtract. Otherwise, we included two off-source shorter exposures of nearby blank sky interspersed among the on-source exposures. Four on-source exposures, rotated 90~degrees from each other, were used in each case to account for detector artefacts. The details of the exposures taken are given in \cref{tab:musesample}. The seeing values were determined by the FWHM of a point source in the (spectrally) flattened data cubes, in the absence of a suitable source we used the observatory's estimate of the conditions provided in the headers. All data were reduced and combined using version 1.6.2 of the ESO MUSE pipeline via {\sc reflex} \citep{freudling13} with default parameters. The ESO MUSE pipeline employs an empirical method for the removal of the sky signal, however this can leave significant residuals, particularly in the redder part of the spectrum where sky lines are prominent. To combat this, Zurich Atmospheric Purge (ZAP\footnote{\url{http://muse-vlt.eu/science/tools}}; \citealt{soto16}) has been developed. We additionally applied this method to our already reduced data cubes to correct these residuals. Where we took off-source sky exposures, we applied ZAP to the reduced off-source blank-sky exposures, before applying the results to the on-source combined data cubes. For those hosts that did not fill the field of view of MUSE, the correction to be applied by ZAP was calculated using blank regions of the on-source data cubes. Although some residuals were still present, they were significantly reduced compared to the standard sky-subtraction method.
\begin{table*}
\begin{threeparttable}
\caption{Observations of SN Iax host galaxies with VLT/MUSE.}
\begin{tabular}{llcrlcc}
\hline
SN name & Host galaxy & $z$ & $R_{25}$ & Date Obs. & Exp. time & Seeing \\
& & & (arcsec) & & (s) & (arcsec) \\
\hline
1991bj & IC~344 & 0.018 & 14.4 & Sept 2015 & $4 \times 700$\tnote{b} & 1.7 \\
2002bp & UGC~6332 & 0.021 & 36.9 & Apr 2015 & $4 \times 555$ & 1.2 \\
2002cx & CGCG~044-035 & 0.024 & 20.8 & Jan 2016 & $4 \times 700$\tnote{b} & 1.7 \\
2004cs & UGC~11001 & 0.014 & 38.7 & Mar 2016 & $4 \times 555$ & 1.1 \\
2005P & NGC~5468 & 0.009 & 78.9 & Jan 2016 & $4 \times 555$ & 1.0 \\
2005hk & UGC~272 & 0.013 & 43.4 & Oct 2015 & $4 \times 555$ & 1.4 \\
2008ae & IC~577 & 0.030 & 15.8 & Nov 2015 & $4 \times 700$\tnote{b} & 0.9 \\
2008ge & NGC~1527 & 0.004 & 111.5 & Sept 2015 & $4 \times 555$ & 0.6 \\
2008ha & UGC~12682 & 0.005 & 40.5 & Nov 2015 & $4 \times 555$ & 1.4 \\
2009J & IC~2160 & 0.016 & 61.3 & Sept 2015 & $4 \times 555$ & 2.0 \\ 2010ae & ESO~162-017 & 0.004 & 61.3 & Sept 2015 & $4 \times 700$\tnote{b} & 1.8 \\
2010el & NGC~1566 & 0.005 & 249.6 & Sept 2015 & $4 \times 520$ & 1.5 \\
2011ce & NGC~6708 & 0.009 & 34.5 & Sept 2015 & $4 \times 555$ & 0.9 \\
2012Z & NGC~1309 & 0.007 & 65.7 & Oct 2015 & $4 \times 555$ & 0.5 \\
2013gr & ESO~114-007 & 0.007 & 54.6 & Sept 2015 & $4 \times 555$ & 1.8 \\
2014dt & NGC~4303 & 0.005 & 193.7 & Jan 2016 & $4 \times 520$ & 1.9 \\
2014ey & CGCG~048-099 & 0.032 & 19.8& Feb/Mar 2016\tnote{a} & $5 \times 700$\tnote{b} & 0.7 \\
2015H & NGC~3464 & 0.012 & 77.1 & Dec 2015 & $4 \times 555$ & 1.1 \\
\hline
\end{tabular}
\label{tab:musesample}
\begin{tablenotes}
\item [a] {The best five exposures from two incomplete observing attempts were combined.}
\item [b] {Sky subtraction was done using blank regions of on-source exposures.}
\end{tablenotes}
\end{threeparttable}
\end{table*}
\subsection{NOT/ALFOSC}
\label{sec:notobs}
We observed 21 host galaxies of SNe~Iax at the NOT using ALFOSC over two campaigns during March, September 2016 and April 2017. The data were reduced largely following the same procedures described in \citet{taddia13, taddia15b}, which we additionally describe here.
We obtained long-exposure ($\gtrsim$~1800~s), long-slit spectra of the \ion{H}{ii}{} regions within the host galaxies, by placing the slit at the SN position and choosing a position angle such that the galaxy centre or other bright \ion{H}{ii}{} regions near the SN location were captured. In most cases the slit contained a few \ion{H}{ii}{} regions bright enough for our analysis. For all but SN~2008A the slit was positioned to go through the host galaxy nucleus. The instrumental setup chosen was the same as adopted for the study presented in \citet{taddia13,taddia15}, i.e. ALFOSC with grism \#4 (wide wavelength range $\sim$3500$-$9000~\AA) and a 1 arcsec-wide slit, resulting in a FWHM (full width half maximum) spectral resolution of $\sim$16--17~\AA. Details of the observations and exposure times adopted for each host galaxy observation are listed in \cref{tab:notsample}.
First, the 2-D (dimensional) spectra were bias subtracted and flat-field corrected in a standard way. When available, multiple exposures were then median-combined to remove any spikes produced by cosmic rays. Otherwise, we removed them using the L.A.Cosmic removal algorithm \citep{vandokkum01}.
We extracted the trace of the brightest object in the 2-D spectrum (either the galaxy nucleus, a bright star, or a \ion{H}{ii}{} region with a bright continuum) and fitted
with a low-order polynomial. The precision of this trace was checked by plotting it over the 2-D spectrum. We then shifted the same trace in the spatial direction to match the position of each \ion{H}{ii}{} region visible in the 2-D spectrum, and then extracted a 1-D spectrum for each \ion{H}{ii}{} region. The extraction regions were chosen by looking at the \text{H}\ensuremath{\alpha}{} flux profile. The extracted spectra were wavelength and flux calibrated using an arc-lamp spectrum and a spectrophotometric standard star, observed the same night or (in March 2016) during the same week, respectively. Following \citet{stanishev07}, we removed the second order
contamination, which characterises the spectra obtained with grism~\#4, from each spectrum. In this study, we included all the
spectra showing at least \text{H}\ensuremath{\alpha}{} and [\text{N {\sc ii}}]{}~$\lambda$6583 emission lines.
\begin{table*}
\begin{threeparttable}
\caption{Observations of SN Iax host galaxies with NOT/ALFOSC.}
\begin{tabular}{llcrlcc}
\hline
SN name & Host galaxy & $z$ & $R_{25}$ & Date Obs. & Exp. time & Seeing \\
& & & (arcsec) & & (s) & (arcsec) \\
\hline
1999ax & SDSS J140358.27+155101.2 & 0.023\tnote{a} & \ldots & Mar 2016 & $1200 \times 3$ & 0.8 \\
2002cx & CGCG~044-035 & 0.024 & 20.8 & Mar 2016 & $1200 \times 3$ & 1.1 \\
2003gq & NGC 7407 & 0.021 & 59.9 & Sep 2016 & $1800 \times 3$ & 0.8 \\
2004gw & CGCG 283-003 & 0.017 & 38.8 & Mar 2016 & $1200 \times 3$ & 1.1 \\
2005cc & NGC 5383 & 0.008 & 94.9 & Mar 2016 & $1800 \times 1$ & 1.2 \\
2006hn & UGC 6154 & 0.017 & 29.3 & Mar 2016 & $1800 \times 2$ & 1.7 \\
2007J & UGC 1778 & 0.017 & 36.1 & Sep 2016 & $1800 \times 3$ & 1.3 \\
2007qd & SDSS J020932.73-005959.8 & 0.043 & 25.1 & Sep 2016 & $1800 \times 3$ & 1.0 \\
2008A & NGC 634 & 0.016 & 62.7 & Sep 2016 & $1800 \times 3$ & 0.7 \\
2009ku & APMUKS(BJ) B032747.73-281526.1 & 0.079 & \ldots & Sep 2016 & $1800 \times 3$ & 1.7 \\
2011ay & NGC 2315 & 0.021 & 40.5 & Mar 2016 & $1200 \times 3$ & 1.0 \\
PS1-12bwh & CGCG 205-021 & 0.023 & 23.8 & Apr 2017 & $1800 \times 3$ & 1.1 \\
2013dh & NGC 5936 & 0.013 & 43.4 & Mar 2016 & $1200 \times 3$ & 1.6 \\
2013en & UGC 11369 & 0.012 & 34.5 & Mar 2016 & $1200 \times 2$ & 1.5 \\
2014ck & UGC 12182 & 0.005 & 38.8 & Mar 2016 & $1800 \times 1$ & 1.7 \\
2014ek & UGC 12850 & 0.023 & 28.7 & Sep 2016 & $1800 \times 3$ & 1.0 \\
2015H & NGC 3464 & 0.012 & 77.1 & Mar 2016 & $1200 \times 3$ & 1.3 \\
2015ce & UGC 12156 & 0.017 & 28.7 & Sep 2016 & $1800 \times 3$ & 1.0 \\
PS~15aic & SDSS J133047.95+380645.0 & 0.056 & 22.3 & Mar 2016 & $1800 \times 2$ & 1.2 \\
PS~15csd & SDSS J020455.52+184815.0 & 0.044\tnote{b} & \ldots & Sep 2016 & $1800 \times 3$ & 0.7 \\
2016ado & SDSS J020305.81-035024.5 & 0.043 & \ldots & Sep 2016 & $1800 \times 3$ & 1.3 \\
\hline
\end{tabular}
\label{tab:notsample}
\begin{tablenotes}
\item [a] {An adjacent galaxy is present at $z = 0.059$ (as measured from \text{H}\ensuremath{\alpha}{} in the NOT spectrum).}
\item [b] {Redshift from \text{H}\ensuremath{\alpha}{} emission of host galaxy in PESSTO classification spectrum.}
\end{tablenotes}
\end{threeparttable}
\end{table*}
\section{MUSE data analysis and methods}
\label{sec:musemethods}
To analyse the data cubes, we used {\sc ifuanal}\footnote{\url{https://github.com/Lyalpha/ifuanal}}, a package developed in Python to perform spaxel binning, stellar continuum and emission line fitting of IFU data following work done by \citet{stanishev12} and \citet{galbany14}. Further documentation is available for the package but we also detail the relevant analysis procedures in the following subsections.
The effects of Galactic reddening were removed for each cube using the extinction maps of \citet{schlafly11} and adopting a $R= 3.1$ \citet{cardelli89} reddening law before correcting to rest-frame using redshifts for each host given by the NASA/IPAC Extragalactic Database (NED)\footnote{\url{http://ned.ipac.caltech.edu/}}.
\begin{figure*}
\centering
\subfloat{\includegraphics[width=1\linewidth]{spaxelbins}}\\
\subfloat{\includegraphics[width=0.333\linewidth]{continuumfit}}
\subfloat{\includegraphics[width=0.666\linewidth]{emissionfit}}
\caption{An example of the binning and fitting routines performed on each data cube described in \cref{sec:binning,sec:contfitting,sec:emisfitting}, as shown for NGC~6708 (the host of SN 2011ce). {\em Top:} The spectral axis has been convolved with the transmission profile of the appropriate filter to produce an $R$-band {\em (left)} and \text{H}\ensuremath{\alpha}{} {\em (middle)} image of the MUSE cube; North is up, East is left. The bins that were produced by the binning algorithm (\cref{sec:hiibinning}) are shown (arbitrarily coloured based on their \text{H}\ensuremath{\alpha}{} brightness) overlaid by the location of the nucleus and the SN explosion site, given by black $\times$ and `star' symbols, respectively {\em (right)}. {\em Bottom:} The continuum fitting of an example spectrum {\em (left)}. Black shows the observed spectrum and orange the {\sc starlight} fit. Regions around strong emission lines were masked and these are shown as vertical shaded regions. Red elements in the residual plot are those flagged as bad pixels or clipped by {\sc starlight}. The continuum fits were subtracted from the spectra before emission line fitting {\em (right)}. Orange dashed line shows the multiple gaussian fit over the observed residual emission spectrum -- line identifications for the main lines used in our analysis are shown.}
\label{fig:museanalysis}
\end{figure*}
\subsection{Spaxel binning}
\label{sec:binning}
As the spaxels in MUSE data are typically smaller than the seeing and typical sizes of objects of interest (i.e. \ion{H}{ii}{} regions), we utilised binning methods in order to combine the signal of adjacent physically-related spaxels, as well as improve the SNR for faint regions. For every galaxy except NGC~1527 (the host of SN~2008ge) we adopted a binning method aimed at segmenting \ion{H}{ii}{} regions. For NGC~1527 we used the Voronoi binning method based on achieving a target SNR of the bins in the continuum. Additional to these algorithmically created bins, we manually added two custom bins, a fibre-like 2~arcsec bin on the host nucleus and a bin at the explosion site of the SN using the same radius limit as for our \ion{H}{ii}{} region binning (\cref{sec:hiibinning}). For any cases where it was evident that foreground stars were affecting the binning algorithm, these were removed using a circular mask before repeating the binning. A single weighted-average spectrum was used to represent each bin, which was forwarded for stellar continuum and emission line fitting. An overview of the binning performed for each data cube is given in \cref{tab:bins}.
\subsubsection{\ion{H}{ii}{} region binning}
\label{sec:hiibinning}
In order to separate \ion{H}{ii}{} regions into bins we use a method based on the {\sc Hiiexplorer}\footnote{\url{http://www.caha.es/sanchez/HII_explorer/}} algorithm \citep{sanchez12b, galbany16}. A narrow 30~\AA{} filter was simulated in the data cubes, centred on the \text{H}\ensuremath{\alpha}{} emission line, and from this we subtracted a continuum level determined by interpolation of the flux in two simulated neighbouring continuum filters. From the \text{H}\ensuremath{\alpha}{} map a region of blank sky was used to determine the background median and standard deviation. Seeds of potential bins were found as all peaks in the flux distribution that were $>8\sigma$ above the background median. The \text{H}\ensuremath{\alpha}{} map was smoothed with a narrow gaussian before the peak detection algorithm was run to avoid finding spurious peaks due to noise fluctuations. Seeds were ordered in descending \text{H}\ensuremath{\alpha}{} flux and, starting with the brightest seed, every pixel in the \text{H}\ensuremath{\alpha}{} map that satisfied the following conditions was added to the candidate bin:
\begin{enumerate}
\item within 250~pc (or 1/2 $\times$ FWHM of the MUSE cube if this was larger) of the seed pixel;
\item a flux at least 10~per~cent of the seed pixel flux;
\item $>5\sigma$ above the background median.
\end{enumerate}
If the pixels satisfying these conditions were a contiguous group of 8 or more then the bin was stored, otherwise the bin was rejected. The next seed (that had not already been assigned a bin) was then used. This process was repeated until all seeds were used.
The nature of the {\sc Hiiexplorer} algorithm means that bins in bright areas of the galaxy are grown to a similar size (i.e. the maximum radius specified). Our limit is conservative in comparison to observed sizes of \ion{H}{ii}{} regions, however our choice was driven by the approximate mode of \ion{H}{ii}{} region sizes \citep[e.g.][]{oey03}. Thus, although we will crop any giant regions that can have diameters up to $\sim$1~kpc, our limit means the majority of bins are representative of \ion{H}{ii}{} region sizes. As shown in the top panels of \cref{fig:museanalysis}, the algorithm captures each of the individual regions above our flux limit, although some of the diffuse emission and the outer extents of very large complexes may be missed.
The typical minimum flux in the \text{H}\ensuremath{\alpha}{} narrow band image we consider as a bin seed was \mbox{a few~$\times 10^{-18}$~erg~sec$^{-1}$~cm$^{-2}$~\AA{}$^{-1}$}. The number of bins found by the algorithm for each galaxy is shown in \cref{tab:bins}.
\subsubsection{Continuum binning}
\label{sec:contbinning}
NGC~1527 is a smooth-profiled, early-type S0 galaxy \citep{foley10b} and as such we chose another method of binning the spaxels. Specifically, we adopted the Voronoi binning method of \citet{cappellari03}, designed to attain a minimum signal-to-noise in the continuum ($5590-5680$~\AA{}) for each bin whilst maximising spatial resolution. Individual spaxels were required to have a signal-to-noise ratio (SNR) of 3 in the continuum window and these were binned by the algorithm to a target of SNR = 40 per bin. In practice, individual spaxels in the central $\sim 2$~kpc of the galaxy had SNR $\gtrsim$~40 and so these were not binned. Even with the reasonably high SNR target, bin sizes only reached radii of $\sim 100-150$~pc in the outskirts of the field of view.
\subsubsection{Host nucleus bins}
\label{sec:surveybin}
For each MUSE cube where the central nucleus of the host galaxy was in the field of view we simulated a representative survey fibre by creating a 2~arcsec diameter bin centred on the nucleus. The results pertaining to these bins will be referred to as ``nucleus'' values.
\subsubsection{SN explosion site bins}
\label{sec:expsites}
We manually added a bin at the location of the SN explosion site in each host. The explosion site bins were circular with a diameter equal to that used for the \ion{H}{ii}{} region binning (\cref{sec:hiibinning}). The location was found via astrometric alignment of a SN discovery image or, if no image could be obtained, we used offsets from the host centre given in discovery IAU circulars (the method used for each is given in \cref{tab:bins}). Where a suitable image was found, the data cube was convolved with a filter transmission matching the SN image filter to create a MUSE image. The MUSE field of view is comparatively small at $\sim 1$~arcmin and so, generally, only a small number of common sources with which to tie the astrometric alignment were present. We therefore first registered the SN and MUSE images with their WCS information before fitting a transformation that allowed a shift only.
Two of our MUSE images detected their respective SN at late times - SN~2014dt and SN~2015H, and we used these as a test of our alignment procedures. With 3 sources to tie the fit, our transformed location of SN~2015H was 0.6 pixels (0.12 arcsec) offset from the centroid location of the SN in the MUSE image itself. For SN 2014dt the MUSE image is quite poor seeing and we are limited to using diffuse \ion{H}{ii}{} regions as common sources; nevertheless we find our transformed location to be only 1 pixel offset from the centroid of the SN. When compared to the typical sizes of the explosion site bins, uncertainties in the alignment of $\sim 1$ pixel do not significantly impact our results.
\begin{table*}
\begin{threeparttable}
\caption{Details of spatial binning for the data cubes.}
\begin{tabular}{lccll}
\hline
SN name & Binning & $N_\text{bins}$ & Survey fibre & Explosion site \\
& algorithm\tnote{a} & & bin? & astrometry\tnote{b} \\
\hline & & & \\
1991bj & \ion{H}{ii}{} region & 31 & Y & NTT/EMMI (R) \\
2002bp & \ion{H}{ii}{} region & 67 & Y & IAUC offsets \\
2002cx & \ion{H}{ii}{} region & 17 & Y & IAUC offsets \\
2004cs & \ion{H}{ii}{} region & 90 & Y & IAUC offsets \\
2005P & \ion{H}{ii}{} region & 127 & N & Swope/SITe3 ($r$) \\
2005hk & \ion{H}{ii}{} region & 87 & Y & VLT/FORS1 (R) \\
2008ae & \ion{H}{ii}{} region & 58 & Y & Swope/SITe3 ($r$) \\
2008ge & continuum & 10237 & Y & Gemini/GMOS-S ($r$) \\
2008ha & \ion{H}{ii}{} region & 50 & Y & Swope/SITe3 ($r$) \\
2009J & \ion{H}{ii}{} region & 61 & Y & Swope/SITe3 ($r$) \\
2010ae & \ion{H}{ii}{} region & 39 & Y & NTT/EFOSC (R) \\
2010el & \ion{H}{ii}{} region & 88 & Y & NTT/EFOSC (R) \\
2011ce & \ion{H}{ii}{} region & 182 & Y & NTT/EFOSC (R) \\
2012Z & \ion{H}{ii}{} region & 311 & Y & Swope/SITe3 ($r$) \\
2013gr & \ion{H}{ii}{} region & 48 & Y & NTT/EFOSC (V) \\
2014dt & \ion{H}{ii}{} region & 109 & N & SOAR/Goodman (clear) \\
2014ey & \ion{H}{ii}{} region & 95 & Y & Swope/SITe3 ($r$) \\
2015H & \ion{H}{ii}{} region & 148 & Y & NTT/EFOSC (V) \\
\hline
\end{tabular}
\label{tab:bins}
\begin{tablenotes}
\item [a] {The \ion{H}{ii}{} region and continuum binning are described in \cref{sec:hiibinning,sec:contbinning}, respectively.}
\item [b] {If an image of the SN was used to astrometrically align the MUSE cube, the Telescope/instrument (filter) is given. Otherwise offsets from the host given in the discovery IAU Circular were used.}
\end{tablenotes}
\end{threeparttable}
\end{table*}
\subsection{Stellar continuum fitting}
\label{sec:contfitting}
Although our analysis of the underlying stellar continuum is restricted as our regions of interest are emission dominated, and generally without sufficient signal in the continuum to permit a robust analysis from the fitting of absorption indices, we nevertheless must account for the continuum in order to determine accurate emission line results. In order to fit the stellar continuum we used the spectral synthesis package {\sc starlight} \citep{cidfernandes05}. A minimisation of the (model $-$ data) residuals is performed by summing contributions of single stellar populations (SSPs) after masking regions dominated by emission lines and assuming a dust screen following the same reddening law we assume for Galactic extinction. SSP base models were taken from \citet{bc03}\footnote{Using the 2016 update at \url{http://www.bruzual.org/bc03/}.} for the MILES spectral library \citep{sanchezblazquez06} using a \citet{chabrier03} stellar initial mass function (IMF) over the range 0.1--100~M$_\odot${}.
The components for the base models comprised 16 ages from 1 Myr to 13 Gyr for each of 4 metallicities (Z~$= 0.004, 0.008, 0.02, 0.05$). We note that tests with somewhat different base model sets and choice of IMF did not significantly affect continuum subtraction and so do not affect the emission line results on which our analyses rely. Our procedure largely mirrors that performed on MUSE data of supernova hosts elsewhere \citep[e.g.,][]{galbany16}.
\subsection{Emission line fitting}
\label{sec:emisfitting}
Emission line fitting was performed on the spectra after subtracting the best-fit continuum, as determined in \cref{sec:contfitting}. Prior to fitting, the continuum-subtracted spectra were median filtered with a width of 120~\AA{}\ -- this width was chosen in order to remove any broad residuals left by the imperfect continuum fitting, whilst leaving the narrow emission lines unaffected. The emission line model was a composite model of single gaussians initially centred at the wavelengths of \text{H}\ensuremath{\beta}{}, [\text{O {\sc iii}}]{} $\lambda\lambda$4959,5007, \text{H}\ensuremath{\alpha}{}, [\text{N {\sc ii}}]{}~$\lambda\lambda$6548,6583 and [\text{S {\sc ii}}]{} $\lambda\lambda$6716,6731 (the two exceptions to this set up, SNe 2002bp and 2008ge, are detailed later). The fitting was performed via the weighted Levenberg-Marquardt least-squares algorithm. Such fitting is however susceptible to finding local minima solutions that are dependent on the initial guesses for the gaussian parameters. In order to circumvent the dependency on initial guesses, and to restrict to physically plausible solutions we added the following caveats and steps to our fitting routine:
\begin{enumerate}
\item The velocity offset of the gaussian means were restricted to within 500~km~s$^{-1}${} relative to the galaxy's rest-frame (i.e., after accounting for the redshift of the galaxy nucleus). This comfortably encloses the limits of rotational velocities seen in late-type galaxies \citep{sofue01} whilst preventing misidentification of lines.
\item The velocity offset of the Balmer and forbidden lines were each tied, i.e a single velocity offset for \text{H}\ensuremath{\alpha}{} and \text{H}\ensuremath{\beta}{} was fit and a separate offset was fit for the other lines.
\item The standard deviations of the gaussians were restricted to $\leq 130$~km~s$^{-1}${} and pairs of lines for [\text{O {\sc iii}}]{}, [\text{N {\sc ii}}]{} and [\text{S {\sc ii}}]{} had tied values for the standard deviation (in km~s$^{-1}${}).
\item In order to semi-brute force locate the global minimum, a list of several initial guesses for each of the means, standard deviations and amplitudes to be fit were made. A grid was then formed of all possible combinations of these parameters to use as initial guesses for the Levenberg-Marquardt fitter, which found the local minimum for each initial guess. The residual between the data and model was used as an estimator of the goodness of fit for each grid position and used to identify the best overall fit.
\end{enumerate}
Where the fitting process failed, we manually inspected the emission line spectrum. In most cases, low signal-to-noise or non-detection of the emission features was the cause and we did not consider these bins in our further analysis. Another source of failure was the nuclear bins of some hosts that harbour an active galactic nucleus (AGN). AGN can have a variety of spectral morphologies including broad emission line components. Their presence prevents much of our emission line analyses of the host nuclei bins since metallicity and age indicators are calibrated based on ionising radiation from recently-formed young stars, these cases were identified as described in \cref{sec:ionising}.
Observations of UGC~6332, the host of SN~2002bp, were taken under relatively poor sky conditions, and its redshift means that the [\text{S {\sc ii}}]{} $\lambda 6731$ lies in a region of telluric absorption. We attempted to correct this with {\sc molecfit} \citep{smette15} using the galaxy nucleus (as there are no bright stars in the MUSE field of view for this host). However we were not able to clean the spectrum sufficiently to recover the line satisfactorily. Instead we opted to set the flux of [\text{S {\sc ii}}]{} $\lambda 6731$ equal to $0.7 \times$~[\text{S {\sc ii}}]{} $\lambda 6716$ for each bin. This was chosen as the relation provided a good fit to the vast majority of all other bins in our MUSE sample. We assigned a factor of two uncertainty in the estimated flux, which encompasses the range of intensity ratios for this doublet for typical \ion{H}{ii}{} regions \citep{osterbrock06}.
As mentioned, the host of SN~2008ge is an S0 galaxy with the consequential expectation of weak or no emission lines from SF. We found after a initial run with the full emission line list detailed above, that the fitting routine was failing as it relies on finding a fit simultaneously for each line. Thus, although inspection of certain bin spectra showed conspicuous emission in stronger lines (e.g.\ \text{H}\ensuremath{\alpha}{}, [\text{N {\sc ii}}]{}), the fits were failing due to non-detections of the weaker lines. We opted to fit this observation for \text{H}\ensuremath{\beta}{}, \text{H}\ensuremath{\alpha}{} and N{\sc ii} $\lambda\lambda$6548,6583 only, which allows us to still estimate \ebv{gas} and determine N2 metallicities for the small number of bins where these were detected. This host galaxy is discussed further in \cref{sec:08ge}.
The fitted gaussian parameters were used to determine the signal-to-noise ratio, flux, equivalent width (EW), FWHM and velocity offset of the lines. Uncertainties on these quantities were found by propagating the statistical uncertainties of the fitted parameters and accounting for photon noise in the manner of \citet[][and references therein]{gomes16b}. The \text{H}\ensuremath{\alpha}{} and \text{H}\ensuremath{\beta}{} fluxes were used to determine the Balmer decrement compared to the expected ratio of 2.86 (assuming Case B recombination, $T_e = 10^4$~K and $n_e = 10^2$~cm$^{-3}$, \citealt{osterbrock06}) and give the reddening of the gas component, \ebv{gas}, again adopting a $R_V= 3.1$ \citet{cardelli89} reddening law. This was used to correct the flux and EW values for the effects of dust extinction. Although assumptions for the physical parameters of the gas are inherent to this correction, they are representative of observed \ion{H}{ii}{} regions. Furthermore, our metallicity measurements, for example, rely on emission lines nearby in wavelength and as such are largely insensitive to reddening.
Luminosity measurements have their statistical uncertainties quoted, however the systematic uncertainty due to the distance of the hosts is a dominating source in most cases. For example, a 500~km~s$^{-1}${} peculiar galaxy velocity at the median redshift of our sample ($z$ = 0.0165) equates to an uncertainty of $\sim 0.084$ dex in $L$(H$\alpha$).
\subsubsection{Ionising source}
\label{sec:ionising}
We created a BPT diagram \citep{baldwin81} for each host, in order to discard bins where the emission lines are not driven by the radiation from young, hot stars, such as regions around AGN, and thus where measures of line strengths and ratios are not appropriate tracers of metallicity, SFR etc. We used the redshift-dependent classification criterion for regular \ion{H}{ii}{} regions being powered by SF of \citet{kewley13}. The theoretically-determined limit on the ionisation ratios from pure SF is also used \citep{kewley01}. We analyse any region below the theoretical SF limit as the region is still consistent with being driven by SF, although there may be smaller contributions from other sources of ionisation. Above this limit we consider the ionising source of the bin to no longer be dominated by SF and do not include these in our results.
An example for the host of SN~2009J, which harbours an AGN and strongly star-forming regions, is shown in \cref{fig:bptmap}. Those bins that were found to be above the theoretical SF line (shown in this plot) were discarded from further analysis as they are powered primarily via other sources of ionisation.
\begin{figure*}
\centering
\includegraphics[width=1\linewidth]{BPTmap}
\caption{An example BPT map and diagram shown for IC~2160 (the host of SN 2009J). {\em Left:} Heatmap classification of the ionising source (SF: star-formation, CP: composite, AGN: active galactic nucleus) of each bin determined using the relations of \citet[][see text]{kewley01, kewley13}, overlaid on a \text{H}\ensuremath{\alpha}{} map of the galaxy; North is up, East is left. The location of the host nucleus and the SN explosion site are indicated by the black $\times$ and `star' symbols, respectively. {\em Right:} The classification limits for each ionising source shown for [\text{N {\sc ii}}]{}/\text{H}\ensuremath{\alpha}{} and [\text{O {\sc iii}}]{}/\text{H}\ensuremath{\beta}{} line ratios with the location of each bin in this parameter space.}
\label{fig:bptmap}
\end{figure*}
\subsubsection{Metallicity}
\label{sec:methmetallicity}
Many emission line metallicity abundance indicators are present in the literature, with significant systematic differences in results between methods, in particular based on whether the method is empirical or theoretically motivated \citep[e.g.][]{kewley08}. The wavelength range of MUSE for low-redshift galaxies covers many emission lines used in various strong emission line methods for determining the gas-phase metallicity, although it does not extend sufficiently into the blue to cover the most commonly used temperature-sensitive lines -- a thorough discussion of metallicity determinations with a view to MUSE data of low redshift galaxies is given in \citet{kruehler17}.
We concentrate here on determining metallicities on the scale of relative oxygen abundance and use the theoretically-motivated calibration based on photo-ionization models presented by \citet[][hereafter \citetalias{dopita16}]{dopita16}. This indicator uses the [\text{S {\sc ii}}]{} $\lambda\lambda6717, 6731$ lines as well as \text{H}\ensuremath{\alpha}{} and [\text{N {\sc ii}}]{} $\lambda6583$ and is thus well-suited to MUSE observations of \ion{H}{ii}{} regions in the local Universe. The relation is given as:
\begin{equation}
12 + \log_{10}(\text{O/H}) = 8.77 + y,
\label{eq:D16}
\end{equation}
where
\begin{equation}
y = \log_{10}\frac{[\text{N {\sc ii}}]{} \lambda6583}{[\text{S {\sc ii}}]{} \lambda\lambda6717, 6731} + 0.264 \cdot \log_{10}\frac{[\text{N {\sc ii}}]{} \lambda6583}{\text{H}\ensuremath{\alpha}}.
\end{equation}
\noindent This is a good description of the theoretical results over a wide range of metallicities from very sub- to super-solar. Although this indicator is robust to changing ionization parameter \citepalias{dopita16}, it was noted by \citet{kruehler17} that there exists an apparent $\sim 0.15$~dex systematic offset to lower abundances with this indicator compared to $T_e$-based values.
We also present values based on the O3N2 \mbox{($ = \log_{10}\left [\frac{[\text{O {\sc iii}}]{} \lambda5007}{\text{H}\ensuremath{\beta}} \times \frac{\text{H}\ensuremath{\alpha}}{[\text{N {\sc ii}}]{} \lambda6583}\right ]$)} calibration of \citet[][hereafter \citetalias{marino13}]{marino13}\footnote{Where we did not reliably detect [\text{O {\sc iii}}]{} $\lambda$5007 in our cubes we use the N2 relation of \citetalias{marino13}. These cases are highlighted in tables of results.}, primarily to facilitate comparison with literature values for the environments of other transient types. There is a 1$\sigma$ uncertainty of 0.18~dex associated to measurements with this indicator due to the observed spread of $T_e$-based abundances about this relation. The indicator is given over the $Z$ range $\sim8.15-8.75$ dex. Where literature values were presented in the O3N2 calibration of \citet{pettini04}, we convert these to \citetalias{marino13}. We adopt a binning regime that means our effective spatial scale is similar to the size of \ion{H}{ii}{} regions, and we are therefore not as sensitive to significant variations of the ionisation parameter that can occur within individual regions on small spatial scales. For ionisation parameter-dependent abundance indicators, such as O3N2, this creates strong gradients within individual regions \citep{kruehler17}.
Although we correct our emission line fluxes for reddening (\ebv{gas} estimated by the Balmer decrement), we note that since the line ratios used in the above metallicity indicators are relatively nearby in wavelength, we are not strongly affected by uncertain reddening values.
We determined metallicity gradients for each host with a linear fit to the bin metallicities and deprojected galactocentric distances.
Deprojected distances were found following the method of \citet{andersen13} to determine the position angle and inclination from the velocity map of the \text{H}\ensuremath{\alpha}{} line and then normalised by the $R_{25}$ (25th $B$-band mag~arcsec$^{-2}$ surface brightness radius) value of the host. $R_{25}$ values were taken from NED. For SN~2008ge the velocity map used was that of the stellar continuum as determined by {\sc starlight}. For SNe where we do not directly detect emission lines at the location of the SN we use these metallicity gradients to estimate the metallicity of the region local to the SN. The uncertainty on this estimate is taken as the rms of the observed metallicity values about the linear fit.
An example of a metallicity map and gradient is shown in \cref{fig:d16map} and stamps for all hosts (except SN~2008ge) are given in \cref{app:zstamps}.
\begin{figure*}
\centering
\includegraphics[width=1\linewidth]{D16_Zmap}
\caption{A gas-phase metallicity map for NGC~1309 (the host of SN~2012Z). Metallicities are given on the oxygen abundance scale using the relations of \citetalias{dopita16}. {\em Left:} A heatmap of the metallicity of each bin overlaid on a \text{H}\ensuremath{\alpha}{} map of the galaxy; North is up, East is left.
The adopted locations of the nucleus and the SN explosion site are indicated by the black $\times$ and `star' symbols, respectively. {\em Right:} A cumulative distribution of the metallicity in the galaxy weighted by the SFR -- i.e. $L(\text{H}\ensuremath{\alpha}{})$ -- of each bin {\em (top)} and the deprojected metallicity gradient of the galaxy {\em (bottom)}. The SN explosion site metallicity is highlighted in these panels also.}
\label{fig:d16map}
\end{figure*}
\subsubsection{SF rates}
\label{sec:methsfr}
We use \citet{kennicutt98} to convert our extinction-corrected $L$(\text{H}\ensuremath{\alpha}{}) values (determined as in \cref{sec:emisfitting}) into SFRs with the relation:
\begin{equation}
SFR~[M_\odot \text{yr}^{-1}] = 7.9 \times 10^{-42} L(\text{H}\alpha)~[\text{ergs s}^{-1}].
\end{equation}
\subsubsection{Ages}
\label{sec:methages}
Ages for the youngest SP in galactic regions and at SN explosion sites are often made based on the EW of (primarily Balmer) emission lines in comparison to theoretical predictions. Such mappings between EW values and ages are subject to significant uncertainty due to potentially unaccounted-for effects (stellar multiplicity, SN feedback) and the large spread in physical characteristics of nebular gas in star forming regions (electron density, physical size).
With these caveats in mind, it is nevertheless widely accepted that the EW of nebular emission lines should have some inverse relation with the age of the youngest SP for an instantaneous or rapidly-declining SF history for the region. We used results from the Binary Population and Spectral Synthesis code, {\sc bpass} \citep[e.g.][]{eldridge09,eldridge12}, and processed the stellar continua with {\sc cloudy} \citep{ferland98}. This was done with {\sc bpass} version 2.1 \citep[][Eldridge \& Stanway, in prep]{stanway16} in the manner described in \citet{stanway14b} for a fiducial nebula gas model of 10$^2$~cm$^{-3}$ hydrogen gas density in a spherical distribution with inner radius of 10~pc. From this we obtained the evolution of EW measurements for \text{H}\ensuremath{\alpha}{} with age for instantaneous SF episodes, which is shown in \cref{fig:bpassew}. These were compared with our measured EW values at the explosion sites to provide estimates of the age of the youngest SP at these locations. We note here the strong effect of including binary stellar evolution for these calculations, which act to strengthen the emission relative to the continuum, beginning after several~Myr. A binary fraction $< 1.0$ would result in evolution somewhere between the two cases shown in \cref{fig:bpassew}.
A pertinent study of the effect of including binary stellar evolution when determining age estimates of CCSN-hosting regions has found that regions are generally older than previously thought, thus revising progenitor initial mass estimates down \citep[][Xiao et al. in prep]{xiao17}
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{BPASS_EW}}
\caption{The evolution of EW(\text{H}\ensuremath{\alpha}{}) for an instantaneous SF episode calculated with {\sc bpass} and {\sc cloudy} (see text). The IMF used has a power-law slope of $\alpha = 2.35$ over the range 0.5 to 100 M$_\odot${}. We stress here that this is for a single fiducial gas model of assumed size and density and that the physical properties of the gas distribution can affect results significantly \citep[see][Eldridge \& Stanway, in prep]{stanway14b}. Different metallicities are represented by colours with binaries and single populations as solid and dashed lines, respectively.}
\label{fig:bpassew}
\end{figure}
\subsubsection{Pixel statistics}
\label{sec:pixelstat}
The location of transients within the light distribution of their hosts has been used to infer further constraints on the nature of the progenitor systems \citep[see][for a review]{anderson15}. We use the Normalised Cumulative Rank (NCR) method presented in \citet{james06}: briefly, pixel values for an image of a transient host are sorted, cumulatively summed and then normalised by the total sum of the values. The location of the transient's explosion site in this cumulative sum provides the fraction of light in the host at a level lower than the intensity at the explosion site. When using an appropriate SF tracer, such as \text{H}\ensuremath{\alpha}{}, the fraction of SF in the host below the level of the explosion site can be inferred; a distribution of \text{H}\ensuremath{\alpha}{} NCR values describes the association to SF for a sample of transients.
The NCR analysis has been presented for an initial sample of SNe~Iax in \citet{lyman13}. We used the same \text{H}\ensuremath{\alpha}{} maps constructed from our MUSE data cubes that were used to create our spaxel bins (\cref{sec:hiibinning}) in order to calculate NCR values for a larger sample of SN~Iax. The maps were binned 3$\times$3 to circumvent astrometric uncertainties before applying the NCR method. Where we had an existing NCR value from \citet{lyman13} and our MUSE observation did not cover the extent of the host, we use the pre-existing value.
We further introduce a metallicity cumulative rank in \cref{sec:reszrank}, which we use to assess the presence of metallicity bias in the production of SNe~Iax within their host galaxies.
\section{NOT data analysis and methods}
\label{sec:notmethods}
\begin{figure*}
\centering
\includegraphics[height=9cm,angle=90]{NOT_slit}
\includegraphics[width=7.0cm,angle=0]{NOT_HaZslit}\\
\includegraphics[width=17cm,angle=0]{NOT_Zgrad}
\caption{An example of the analysis performed on our NOT observations for UGC~1778 (the host of SN~2007J). {\em Top left:} The open filter acquisition image. Overlaid are the 25th $B$-band magnitude elliptic contour (black solid line), the position of SN~2007J (red star) and the centre of the galaxy (red circle). The slit position is shown, and a color code is used to show the \citetalias{dopita16} metallicity measurements at the position of each \ion{H}{ii}{} region that we inspected. {\em Top right:} The flux at \text{H}\ensuremath{\alpha}{} along the slit, with the location of the SN and host nucleus marked. The \citetalias{dopita16} metallicity measurements are shown at the corresponding positions in the top sub-panel. {\em Bottom:} The metallicity gradient of UGC~1778. The linear fit on our D16 measurements is shown with a solid red line. The interpolated metallicity at the SN distance is marked with a empty red square and its uncertainty corresponds to the fit error. The positions of SN and nucleus are marked with vertical dotted lines. The solar metallicity and the LMC metallicity are indicated with horizontal dotted lines.}
\label{fig:notanalysis}.
\end{figure*}
We here detail the analysis of our NOT observations, highlighting differences from the MUSE analysis. Further details of the analysis methods used can be found in \citet{taddia13,taddia15}. Prior to analysis, NOT spectra were corrected for Galactic extinction and dereddened in the same manner as the MUSE data cubes. The main differences from the MUSE analysis were the identification of \ion{H}{ii}{} regions versus spaxel binning routines and the fitting of emission lines due to a lower spectral resolution. Stellar continuum fitting, identification of the ionising source and our metallicity, SFR and age measurements followed those described in \cref{sec:musemethods}.
Intrinsic (deprojected) galactocentric offsets of SN explosion sites and other \ion{H}{ii}{} regions were calculated and normalised to $R_{25}$ following the method of \citet{hakobyan09,hakobyan12}. Host galaxy position angles and inclinations were obtained from NED, supplemented by data from HYPERLEDA\footnote{\url{http://atlas.obs-hp.fr/hyperleda/}}, since we were not able to determine a velocity map as was done for the MUSE data (\cref{sec:methmetallicity}). Again, these were normalised by the $R_{25}$ of the hosts. These deprojected distances were used to determine metallicity gradients for the host galaxies where more than two \ion{H}{ii}{} regions were extracted. For SNe where we could not extract a \ion{H}{ii}{} region underlying the explosion site we used these gradients to estimate the metallicity of the SN. Statistical uncertainties on these interpolated or extrapolated estimates were determined based on the uncertainties of our linear gradient fit (which generally dominate over the rms of \ion{H}{ii}{} regions about the relation, whereas the opposite is true for MUSE, \cref{sec:methmetallicity}).
Since we only have long-slit spectra for these targets, we do not perform our pixel statistics, as done for the MUSE sample (\cref{sec:pixelstat}).
For two observations, those of UGC~12182 (SN~2014ck) and SDSS J020305.81-035024.5 (SN~2016ado) we were not able to perform our main emission line-based analysis at any location within the slit due to poor observing conditions. We remove these two events from our sample when discussing our spectroscopic results as they provide no meaningful measurements or limits (however we include the host-normalised offset of SN~2014ck, $r/R_{25} = 0.18$).
For the hosts of SN~1999ax and PS~15aic we were only able to extract emission at the host nuclei.\footnote{For SN~1999ax we additionally detect another nearby source of \ion{H}{ii}{} at an apparently higher redshift, this is discussed in \cref{sec:hetero}.} As we have no additional spatial information to be gained from the NOT long-slit spectra, we opted to use existing SDSS spectra which offer improved spectral resolution and depth on which to perform our analysis.
For SN~2009ku and PS~15csd we take the host nucleus measurements as those of the SN explosion site owing to the small apparent size of the hosts and the very small offset of the SNe.
Within our NOT-observed sample we have two duplicates of MUSE-observed events: SNe~2002cx and 2015H. We found very good agreement (within 1$\sigma$) between our results obtained for the explosion site of SN~2002cx and host nuclei of each (the explosion site of SN~2015H was not detected in emission lines in the NOT data) for the two sets of data. We present our results for these SNe based on the MUSE data results since these are subject to smaller statistical uncertainties owing to greater signal-to-noise of the emission lines and increased spectral resolution, and also allowed for more robust determinations of the metallicity gradients as more \ion{H}{ii}{} regions could be analysed.
\subsection{H{\sc ii} region identification}
We performed an inspection of each acquisition image as well the \text{H}\ensuremath{\alpha}{} line flux trace in order to identify \ion{H}{ii}{} regions from which to extract spectra (i.e. peaks in the flux trace). An example of this procedure is shown in the top of \cref{fig:notanalysis}, extraction widths were ad hoc in order to maximise the signal to noise ratio.
For our observation of SN 2008A the slit did not cover the host nucleus and thus we do not have a measurement of the nucleus for this host.
\subsection{Emission line fitting}
Emission line fitting followed broadly the same methodology as for the MUSE spectra (\cref{sec:emisfitting}), however, the lower resolution of the NOT spectra meant we had to use a different fitting routine as nearby lines are blended. Firstly, the \text{H}\ensuremath{\alpha}{} and [\text{N {\sc ii}}]{} $\lambda\lambda$6548,6583 lines were simultaneously fit with the gaussians of fixed width (determined by the spectral resolution), and fixed centroid offsets using the known wavelengths of the lines. Furthermore, the flux of
[\text{N {\sc ii}}]{} $\lambda$6548 was fixed as one third that of [\text{N {\sc ii}}]{} $\lambda$6583 \citep{osterbrock06} -- this assumption was required to allow for a proper fit of this faint line, although it is not used in our metallicity determinations, to remove its contamination from \text{H}\ensuremath{\alpha}{}. Similarly, the [\text{S {\sc ii}}]{} $\lambda\lambda$6716,6731 doublet was fitted with two gaussians of fixed width and separation. We also attempted fits to \text{H}\ensuremath{\beta}{}, [\text{O {\sc iii}}]{} $\lambda$5007 for each extracted spectrum, however, due to poorer sensitivity in the blue part of the spectrum we were were not able to measure these in several cases.
\begin{table*}
\begin{threeparttable}
\caption{Results for SN explosion sites and host nuclei.}
\begin{tabular}{L{3cm}ccccc@{\hskip 0.4cm}clcl}
\hline
Galaxy & SN & Offset\tnote{a} & $L$(H$\alpha$) & EW(H$\alpha$) & NCR & \multicolumn{2}{c}{$Z$ (SN site)\tnote{b}} & \multicolumn{2}{c}{$Z$ (Host nucleus)\tnote{c}} \\
& & [${R_{25}}$] & [$\log_{10}$ erg s$^{-1}$] & [\AA{}] & & D16 & \multicolumn{1}{c}{M13} & D16 & \multicolumn{1}{c}{M13} \\
\hline
\multicolumn{10}{c}{MUSE sample}\\
\hline
IC~344 & 1991bj & 1.383 & 39.40$\pm$0.01 & 98.9$\pm$3.4 & 0.29 & 8.26$\pm$0.01 & 8.37$\pm$0.01 & 8.55$\pm$0.01 & 8.51$\pm$0.01 \\
UGC~6332 & 2002bp & 0.920 & 37.13$\pm$0.36 & 1.8$\pm$1.5 & 0.00 & \ldots & 8.77$\pm$0.25\tnote{$\dagger$} & \multicolumn{2}{c}{AGN} \\
CGCG~044-035 & 2002cx & 0.765 & 39.19$\pm$0.01 & 45.8$\pm$1.9 & 0.33 & 8.25$\pm$0.02 & 8.36$\pm$0.01 & 8.52$\pm$0.01 & 8.50$\pm$0.01 \\
UGC~11001 & 2004cs\tnote{\ddag} & 0.519 & 39.23$\pm$0.01 & 78.6$\pm$2.2 & 0.72 & 8.42$\pm$0.02 & 8.43$\pm$0.01 & 8.68$\pm$0.01 & 8.56$\pm$0.02 \\
NGC~5468 & 2005P & 0.443 & 37.50$\pm$0.02 & 9.2$\pm$0.9 & 0.06\tnote{d} & 8.07$\pm$0.13 & 8.36$\pm$0.06 & \multicolumn{2}{c}{not in FOV} \\
UGC~272 & 2005hk & 1.089 & 37.25$\pm$0.04 & 20.9$\pm$5.1 & 0.00 & 7.85$\pm$0.28 & 8.34$\pm$0.09 & 8.39$\pm$0.02 & 8.45$\pm$0.01 \\
IC~577 & 2008ae & 0.881 & 39.06$\pm$0.01 & 58.6$\pm$5.4 & 0.07 & 8.66$\pm$0.03 & 8.50$\pm$0.02 & 8.99$\pm$0.04 & 8.53$\pm$0.01 \\
NGC~1527 & 2008ge & 0.055 & \ldots & \ldots & 0.00 & \ldots & \ldots & \ldots & \ldots \\
UGC~12682 & 2008ha & 0.263 & 37.91$\pm$0.01 & 24.9$\pm$1.0 & 0.41\tnote{d} & 7.91$\pm$0.04 & 8.22$\pm$0.01 & 7.99$\pm$0.02 & 8.19$\pm$0.00 \\
IC~2160 & 2009J & 0.551 & 38.81$\pm$0.01 & 34.3$\pm$1.7 & 0.00\tnote{d} & 8.51$\pm$0.03 & 8.41$\pm$0.02 & \multicolumn{2}{c}{AGN} \\
ESO~162-017 & 2010ae & 0.070 & 39.24$\pm$0.01 & 39.7$\pm$0.8 & 0.61 & 8.26$\pm$0.01 & 8.32$\pm$0.00 & 8.28$\pm$0.01 & 8.31$\pm$0.00 \\
NGC~1566 & 2010el & 0.106 & 38.66$\pm$0.01 & 20.4$\pm$0.6 & 0.03 & 8.85$\pm$0.02 & 8.51$\pm$0.02 & \multicolumn{2}{c}{AGN} \\
NGC~6708 & 2011ce & 0.136 & 39.94$\pm$0.01 & 86.4$\pm$1.9 & 0.59 & 8.76$\pm$0.01 & 8.59$\pm$0.01 & 8.88$\pm$0.01 & 8.60$\pm$0.02 \\
NGC~1309 & 2012Z & 0.766 & 37.80$\pm$0.01 & 36.1$\pm$1.9 & 0.00\tnote{d} & 8.26$\pm$0.03 & 8.41$\pm$0.01 & \multicolumn{2}{c}{AGN} \\
ESO~114-007 & 2013gr & 0.266 & 37.78$\pm$0.01 & 31.0$\pm$1.5 & 0.06 & 7.84$\pm$0.05 & 8.27$\pm$0.01 & 7.96$\pm$0.03 & 8.27$\pm$0.01 \\
NGC~4303 & 2014dt & 0.218 & 38.78$\pm$0.01 & 33.0$\pm$0.6 & 0.04 & 8.85$\pm$0.02 & 8.52$\pm$0.02 & \multicolumn{2}{c}{not in FOV} \\
CGCG~048-099 & 2014ey & 0.616 & 38.87$\pm$0.01 & 47.8$\pm$3.6 & 0.22 & 8.36$\pm$0.04 & 8.43$\pm$0.02 & 8.67$\pm$0.01 & 8.57$\pm$0.02 \\
NGC~3464 & 2015H & 0.426 & 37.97$\pm$0.02 & 7.6$\pm$0.4 & 0.01 & 8.66$\pm$0.07 & 8.50$\pm$0.10 & 8.97$\pm$0.05 & 8.58$\pm$0.01\tnote{$\dagger$} \\
\hline
\multicolumn{10}{c}{NOT sample}\\
\hline
SDSS J140358.27+155101.2 & 1999ax & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots & 7.78$\pm$0.10 & 8.23$\pm$0.02 \\
NGC 7407 & 2003gq & 0.133 & 39.57$\pm$0.01 & 23.5$\pm$0.9 & \ldots & 8.51$\pm$0.07 & 8.33$\pm$0.02\tnote{$\dagger$} & \multicolumn{2}{c}{AGN} \\
CGCG 283-003 & 2004gw & 0.568 & \ldots & \ldots & \ldots & \ldots & \ldots & 9.23$\pm$0.24 & 8.54$\pm$0.02 \\
NGC 5383 & 2005cc & 0.061 & 38.81$\pm$0.01 & 28.7$\pm$0.4 & \ldots & 8.80$\pm$0.07 & 8.50$\pm$0.02\tnote{$\dagger$} & 8.86$\pm$0.08 & 8.51$\pm$0.02\tnote{$\dagger$} \\
UGC 6154 & 2006hn & 0.498 & 38.98$\pm$0.01 & 18.4$\pm$0.6 & \ldots & 8.81$\pm$0.17 & 8.50$\pm$0.02 & 8.95$\pm$0.05 & 8.48$\pm$0.01 \\
UGC 1778 & 2007J\tnote{\ddag} & 0.514 & 39.35$\pm$0.01 & 116.9$\pm$9.4 & \ldots & 8.41$\pm$0.04 & 8.38$\pm$0.02 & 8.84$\pm$0.04 & 8.45$\pm$0.01 \\
SDSS J020932.73-005959.8 & 2007qd & 0.531 & \ldots & \ldots & \ldots & \ldots & \ldots & 8.54$\pm$0.03 & 8.46$\pm$0.01 \\
NGC 634 & 2008A & 0.789 & \ldots & \ldots & \ldots & \ldots & \ldots & \multicolumn{2}{c}{not in FOV} \\
APMUKS(BJ) B032747.73-281526.1 & 2009ku\tnote{e}& \ldots & 40.41$\pm$0.01 & 47.86$\pm$0.82 & \ldots & 8.43$\pm$0.06 & 8.37$\pm$0.01 & 8.43$\pm$0.06 & 8.37$\pm$0.01 \\
NGC 2315 & 2011ay & 0.468 & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots & 8.69$\pm$0.01 \\
CGCG 205-021 & PS1-12bwh & 0.318 & \ldots & \ldots & \ldots & \ldots & \ldots & 8.98$\pm$0.07 & 8.56$\pm$0.01 \\
NGC 5936 & 2013dh & 0.219 & 39.14$\pm$0.01 & 34.3$\pm$3.7 & \ldots & 8.77$\pm$0.05 & 8.58$\pm$0.01 & 9.18$\pm$0.03 & 8.63$\pm$0.01 \\
UGC 11369 & 2013en & 0.480 & 39.32$\pm$0.01 & 20.7$\pm$0.7 & \ldots & 8.48$\pm$0.09 & 8.46$\pm$0.03\tnote{$\dagger$} & 9.05$\pm$0.02 & 8.63$\pm$0.01\tnote{$\dagger$} \\
UGC 12850 & 2014ek & 0.248 & 39.83$\pm$0.01 & 34.8$\pm$0.4 & \ldots & \ldots & 8.50$\pm$0.02\tnote{$\dagger$} & \multicolumn{2}{c}{AGN} \\
UGC 12156 & 2015ce & 0.394 & \ldots & \ldots & \ldots & \ldots & \ldots & 8.92$\pm$0.09 & 8.63$\pm$0.01 \\
SDSS J133047.95+380645.0 & PS~15aic & 0.666 & 39.50$\pm$0.02 & 2.3$\pm$0.1 & \ldots & \ldots & \ldots & \ldots & 8.72$\pm$0.02\tnote{$\dagger$} \\
SDSS J020455.52+184815.0 & PS~15csd\tnote{e} & \ldots & 39.55$\pm$0.02 & 25.1$\pm$1.0 & \ldots & 8.11$\pm$0.17 & 8.29$\pm$0.03 & 8.11$\pm$0.17 & 8.29$\pm$0.03 \\
\hline
\end{tabular}
\label{tab:results}
\begin{tablenotes}
\item[$\dagger$] The N2 relation of \citetalias{marino13} was used instead of O3N2.
\item[\ddag] Helium detected in SN spectra.
\item[a] The deprojected offset of the explosion site normalised by the $R_{25}$ value of the host.
\item[b] The $12 + \log_{10}(\textrm{O/H})$ abundance measured in the scales of \citetalias{dopita16} and \citetalias{marino13} for the SN explosion site. Uncertainties are statistical only: \citetalias{marino13} have a systematic uncertainty of 0.18 dex; \citetalias{dopita16} do not have well quantified systematics but appear lower than $T_e$-based estimates \citep{kruehler17}.
\item[c] As for $^\text{b}$ but for the host nucleus. `AGN' means the host was deemed to host an active galactic nucleus, `not in FOV' means the host nucleus was not captured in the IFU or slit.
\item[d] NCR value taken from \citet{lyman13}.
\item[e] Due to the SN being overlaid on the nucleus of an unresolved host at very small offset, we use the host nucleus results as those of the SN site also.
\end{tablenotes}
\end{threeparttable}
\end{table*}
\begin{landscape}
\begin{table}
\begin{center}
\begin{threeparttable}
\caption{Metallicity gradients for SN Iax host galaxies and SN explosion site metallicities. The metallicity gradients for SN Iax host galaxies are parameterised as $Z = m/R_{25} \times D + c$, where $D$ is the deprojected offset normalised by the host $R_{25}$, in the scales of \citetalias{dopita16} and \citetalias{marino13}. $Z_\textrm{SN, grad}$ are estimates of $Z$ based on these metallicity gradients at the offset of the SN, and $Z_\textrm{SN, meas}$ are our directly measured values for the \ion{H}{ii}{} underlying the SN position.}
\begin{tabular}{ll@{\hskip 0.75cm}rccc@{\hskip 0.75cm}rccc}
\hline
Galaxy & SN name & \multicolumn{4}{c}{D16} & \multicolumn{4}{c}{M13} \\
& & $m/R_{25}$ & $c$ & $Z_\textrm{SN, grad}$ & $Z_\textrm{SN, meas}$ & $m/R_{25}$ & $c$ & $Z_\textrm{SN, grad}$ & $Z_\textrm{SN, meas}$ \\
\hline
\multicolumn{10}{c}{MUSE sample}\\
\hline
IC~344 & 1991bj & $-0.27\pm0.08$ & $ 8.59\pm0.08$ & $ 8.22\pm0.05 $ & $ 8.26\pm0.01 $ & $-0.13\pm0.07$ & $ 8.51\pm0.06$ & $ 8.33\pm0.04 $ & $ 8.37\pm0.01 $ \\
UGC~6332\tnote{$\dagger$} &2002bp & $ 0.07\pm0.19$ & $ 8.82\pm0.12$ & $ 8.88\pm0.08 $ & \ldots & $ 0.01\pm0.06$ & $ 8.56\pm0.04$ & $ 8.57\pm0.02 $ & $ 8.77\pm0.25 $ \\
CGCG~044-035 & 2002cx & $-0.43\pm0.11$ & $ 8.54\pm0.05$ & $ 8.21\pm0.03 $ & $ 8.25\pm0.02 $ & $-0.18\pm0.08$ & $ 8.49\pm0.03$ & $ 8.35\pm0.02 $ & $ 8.36\pm0.01 $ \\
UGC~11001 & 2004cs & $-0.28\pm0.13$ & $ 8.62\pm0.07$ & $ 8.47\pm0.05 $ & $ 8.42\pm0.02 $ & $-0.22\pm0.07$ & $ 8.56\pm0.04$ & $ 8.45\pm0.02 $ & $ 8.43\pm0.01 $ \\
NGC~5468 & 2005P & $-0.68\pm0.15$ & $ 8.70\pm0.08$ & $ 8.40\pm0.05 $ & $ 8.07\pm0.13 $ & $-0.38\pm0.10$ & $ 8.57\pm0.05$ & $ 8.40\pm0.04 $ & $ 8.36\pm0.06 $ \\
UGC~272 & 2005hk & $-0.38\pm0.10$ & $ 8.30\pm0.06$ & $ 7.88\pm0.05 $ & $ 7.85\pm0.28 $ & $-0.18\pm0.08$ & $ 8.41\pm0.05$ & $ 8.22\pm0.04 $ & $ 8.34\pm0.09 $ \\
IC~577 & 2008ae & $-0.23\pm0.11$ & $ 8.88\pm0.07$ & $ 8.68\pm0.04 $ & $ 8.66\pm0.03 $ & $-0.17\pm0.05$ & $ 8.64\pm0.04$ & $ 8.49\pm0.03 $ & $ 8.50\pm0.02 $ \\
UGC~12682 & 2008ha & $-0.14\pm0.13$ & $ 7.97\pm0.07$ & $ 7.93\pm0.07 $ & $ 7.91\pm0.04 $ & $-0.06\pm0.12$ & $ 8.25\pm0.06$ & $ 8.24\pm0.05 $ & $ 8.22\pm0.01 $ \\
IC~2160 & 2009J & $-0.79\pm0.23$ & $ 8.97\pm0.10$ & $ 8.53\pm0.06 $ & $ 8.51\pm0.03 $ & $-0.50\pm0.20$ & $ 8.66\pm0.10$ & $ 8.38\pm0.05 $ & $ 8.41\pm0.02 $ \\
ESO~162-017 & 2010ae & $-0.24\pm0.17$ & $ 8.26\pm0.07$ & $ 8.24\pm0.06 $ & $ 8.26\pm0.01 $ & $-0.08\pm0.16$ & $ 8.35\pm0.06$ & $ 8.35\pm0.05 $ & $ 8.32\pm0.01 $ \\
NGC~1566 & 2010el & $-0.64\pm0.75$ & $ 8.91\pm0.10$ & $ 8.84\pm0.05 $ & $ 8.85\pm0.02 $ & $ 0.13\pm0.44$ & $ 8.55\pm0.07$ & $ 8.57\pm0.03 $ & $ 8.51\pm0.02 $ \\
NGC~6708 & 2011ce & $-0.54\pm0.10$ & $ 8.85\pm0.05$ & $ 8.78\pm0.04 $ & $ 8.76\pm0.01 $ & $-0.30\pm0.09$ & $ 8.62\pm0.04$ & $ 8.58\pm0.04 $ & $ 8.59\pm0.01 $ \\
NGC~1309 & 2012Z & $-0.41\pm0.14$ & $ 8.60\pm0.05$ & $ 8.29\pm0.05 $ & $ 8.26\pm0.03 $ & $-0.24\pm0.10$ & $ 8.55\pm0.04$ & $ 8.37\pm0.03 $ & $ 8.41\pm0.01 $ \\
ESO~114-007 & 2013gr & $-0.23\pm0.14$ & $ 7.99\pm0.06$ & $ 7.93\pm0.06 $ & $ 7.84\pm0.05 $ & $-0.16\pm0.18$ & $ 8.29\pm0.08$ & $ 8.24\pm0.07 $ & $ 8.27\pm0.01 $ \\
NGC~4303 & 2014dt & $-0.27\pm0.30$ & $ 8.88\pm0.07$ & $ 8.82\pm0.05 $ & $ 8.85\pm0.02 $ & $ 0.07\pm0.28$ & $ 8.55\pm0.07$ & $ 8.57\pm0.03 $ & $ 8.52\pm0.02 $ \\
CGCG~048-099 & 2014ey & $-0.36\pm0.09$ & $ 8.65\pm0.06$ & $ 8.42\pm0.06 $ & $ 8.36\pm0.04 $ & $-0.17\pm0.07$ & $ 8.51\pm0.06$ & $ 8.40\pm0.05 $ & $ 8.43\pm0.02 $ \\
NGC~3464 & 2015H & $-0.51\pm0.20$ & $ 8.92\pm0.08$ & $ 8.70\pm0.05 $ & $ 8.66\pm0.07 $ & $-0.52\pm0.11$ & $ 8.77\pm0.05$ & $ 8.55\pm0.03 $ & $8.50\pm0.10$ \\
\hline
\multicolumn{10}{c}{NOT sample}\\
\hline
NGC 7407\tnote{$\dagger$} & SN~2003gq & $-0.04\pm0.43$ & $8.61\pm0.13$ & $8.60\pm0.14$ & $8.51\pm0.07$ & $ 0.30\pm0.32$ & $8.38\pm0.10$ & $8.42\pm0.11$ & $8.33\pm0.02$ \\
CGCG 283-003\tnote{$\dagger$} & SN~2004gw & $-1.21\pm0.21$ & $9.18\pm0.07$ & $8.03\pm0.21$ & \ldots & $-0.12\pm0.08$ & $8.51\pm0.02$ & $8.39\pm0.08$ & \ldots \\
NGC 5383\tnote{$\dagger$} & SN~2005cc & $-0.97\pm0.21$ & $8.86\pm0.01$ & $8.80\pm0.02$ & $8.80\pm0.07$ & $-0.30\pm0.30$ & $8.51\pm0.01$ & $8.49\pm0.02$ & $8.50\pm0.02$ \\
UGC 6154 & SN~2006hn & $-0.14\pm0.03$ & $8.94\pm0.02$ & $8.82\pm0.03$ & $8.81\pm0.17$ & $ 0.06\pm0.10$ & $8.49\pm0.06$ & $8.55\pm0.11$ & $8.50\pm0.02$ \\
UGC 1778 & SN~2007J & $-0.86\pm0.16$ & $8.85\pm0.06$ & $8.40\pm0.10$ & $8.41\pm0.04$ & $-0.13\pm0.02$ & $8.45\pm0.01$ & $8.38\pm0.01$ & $8.38\pm0.02$ \\
SDSS J020932.73-005959.8\tnote{$\dagger$} & SN~2007qd & $0.07$ & $8.54$ & $8.60\pm0.15$\tnote{a} & \ldots & $0.04$ & $8.46$ & $8.50\pm0.15$\tnote{a} & \ldots \\
NGC 2315\tnote{$\dagger$} & SN~2011ay & \ldots & \ldots & \ldots & \ldots & $-0.42 $ & $8.69$ & $8.50\pm0.15$\tnote{a} & \ldots \\
CGCG 205-021\tnote{$\dagger$} & PS1-12bwh & $-0.42$ & $8.98$ & $8.84\pm0.15$\tnote{a} & \ldots & $ 0.00\pm0.05$ & $8.51\pm0.04$ & $8.51\pm0.04$ & \ldots \\
NGC 5936 & SN~2013dh & $-1.92\pm0.55$ & $9.17\pm0.10$ & $8.75\pm0.16$ & $8.77\pm0.05$ & $-0.21\pm0.02$ & $8.63\pm0.00$ & $8.59\pm0.00$ & $8.58\pm0.01$ \\
UGC 11369\tnote{$\dagger$} & SN~2013en & $-0.96\pm0.37$ & $9.07\pm0.14$ & $8.61\pm0.24$ & $8.48\pm0.09$ & $-0.28\pm0.11$ & $8.63\pm0.04$ & $8.50\pm0.07$ & $8.46\pm0.03$ \\
UGC 12850\tnote{$\dagger$} & SN~2014ek & \ldots & \ldots & \ldots & \ldots & $-0.00\pm0.03$ & $8.48\pm0.02$ & $8.48\pm0.02$ & $8.50\pm0.02$ \\
UGC 12156\tnote{$\dagger$} & SN~2015ce & $-1.49\pm0.25$ & $8.88\pm0.08$ & $8.29\pm0.13$ & \ldots & $-0.36\pm0.16$ & $8.61\pm0.05$ & $8.46\pm0.08$ & \ldots \\
\hline
\end{tabular}
\label{tab:zgrad}
\begin{tablenotes}
\item[$\dagger$] The N2 relation of \citetalias{marino13} was used instead of O3N2.
\item[a] Only two \ion{H}{ii}{} regions were used to determine the gradient. We adopt an uncertainty of $\pm$0.15 dex on $Z_\textrm{SN, grad}$.
\end{tablenotes}
\end{threeparttable}
\end{center}
\end{table}
\end{landscape}
\section{Results}
\label{sec:results}
We present our main results of the SNe explosion sites and host nuclei in \cref{tab:results}. We indicate those observations where the host nucleus was not captured in the FOV or where the ionising source was deemed to be powered by an AGN (\cref{sec:ionising,fig:bptmap}). Individual line fluxes for the explosion sites (where they could be measured) are given in \cref{tab:fluxes}.
\subsection{Metallicity}
\label{sec:resz}
\subsubsection{Gradients and estimating explosion site metallicities}
\label{sec:zgrads}
In some cases we were not able to directly measure emission lines at the explosion site of the SN. As discussed in \cref{sec:notmethods}, in these cases we resort to estimating the metallicity based on the observed metallicity gradient of the galaxy as has been done previously for other studies of SN environments.
Here we utilise the power of the MUSE data, in combination with the results of NOT to assess the reliability of this method in the extreme regimes of many and few \ion{H}{ii}{} regions from which to determine the gradient. In \cref{tab:zgrad} we present our determined metallicity gradients alongside the estimated (based on the gradient) and measured metallicities of the SN explosion sites, which are plotted in \cref{fig:Zgrad}.\footnote{We exclude NGC~1527, the host of SN~2008ge, as even if the ionised gas we detect is driven by SF (see \cref{sec:08ge}), it is very irregular in morphology, and perhaps not intrinsic to the luminous S0 host. Thus a metallicity gradient is not likely to be an appropriate description.} The estimates agree very well with the measured values -- the one outlier in the figure is SN~2005P, although this is still only at the $\sim2.3\sigma$ level and interestingly only in the \citetalias{dopita16} indicator. We crucially also see that even with a low number of regions from which to fit a metallicity gradient (in the case of NOT data), the measured and estimated values agree well, offering further support of results derived using this method. The associated uncertainties (determined from the residuals of the measured \ion{H}{ii}{} region metallicities about the gradient or based on the uncertainties of the linear fit; \cref{sec:musemethods,sec:notmethods}), are generally larger than the statistical uncertainties on the measured values, but are still mainly dominated by the systematic uncertainties on the calibrations. We note that these findings corroborate those of \citet{galbany16b} who investigated a number of local metallicity estimators in IFS data, finding that an interpolation of the metallicity gradient is a robust estimate.
For some observations we were not able to either measure the metallicity at the SN explosion site or determine a metallicity gradient. In the cases of SNe~1999ax, 2009ku, PS~15aic and PS~15csd we were only able to measure the host nucleus metallicities, and for SN~2008A we could only measure a single bright \ion{H}{ii}{} region slightly offset from the host nucleus\footnote{In agreement with our findings based on the NOT spectroscopy here, \citet{lyman13} show a \text{H}\ensuremath{\alpha}{} image of the SN location with no underlying detected emission and mainly faint, diffuse emission throughout the host. \citet{mccully14b} also present {\em HST} broadband observations of the environment, showing it to be in the outskirts of its host, in a fairly low density environment.}. SN~2009ku and PS~15csd lie on top of their respective hosts which are of small apparent size and so we adopt the host nucleus metallicities as those of the SN (\cref{tab:results}). Our determined metallicity gradients cover a large range, making any estimates for SNe~1999ax, 2008A or PS~15csd based on an average gradient of limited use as the associated error would be very large (\citealt{galbany16b} also caution against this methodology). We are limited to supposing their metallicities are less than or equal to their respective host nucleus metallicities (for SN~2008A our sole \ion{H}{ii}{} region metallicity is at an offset of $r$/$R_{25}$ = 0.168 with a metallicity of $Z_\textrm{D16} = 9.10\pm0.07_{stat}$ dex, $Z_\textrm{M13, N2} = 8.60\pm0.01_{stat}$ dex). This does not introduce the requirement for especially low metallicities (compared to the rest of the sample) for these events.
Given the very good agreement between our direct explosion site metallicities and those estimated based on the hosts' gradients, we supplement our explosion site metallicity values with gradient estimates where appropriate, and present and discuss this enlarged sample in our results.
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{Zgrad_D16M13}}
\caption{A comparison between the metallicity determined directly at the location of SN explosion sites ($Z_\textrm{SN, meas}$) and that which would be estimated based on the metallicity gradient alone ($Z_\textrm{SN, grad}$). Metallicities are determined using the calibrations of \citetalias{dopita16} and \citetalias{marino13} in the top and bottom panels, respectively. Empty markers in the lower panel indicate the N2 relation was used (versus O3N2 for filled markers). Uncertainties on the measured metallicities are statistical only. The gradient uncertainties are either the uncertainty on the gradient linear fit, or the root mean square of \ion{H}{ii}{} region metallicity residuals about the gradient (see text). Orange dashed lines indicate the one to one relation for each.}
\label{fig:Zgrad}
\end{figure}
\subsubsection{Distribution}
\label{sec:reszdist}
In \cref{fig:Zall} the metallicity distribution of SNe Iax explosion sites is shown alongside that of the host nuclei and all \ion{H}{ii}{} regions extracted in our MUSE sample\footnote{We do not include the emission from the host of SN~2008ge, see \cref{sec:08ge}.}. The cumulative weight of each \ion{H}{ii}{} region in this plot is given by its \text{H}\ensuremath{\alpha}{} luminosity, i.e. SFR, in order to show the cumulative distribution of SFR with metallicity in SN~Iax hosts. Although we only have global metallicity determinations for the MUSE half of our sample, the main discriminating factor between this and the NOT sample is only declination of the source, and so we expect no bias in terms of host properties. Additionally, a few of our MUSE observations do not cover the entire hosts, but they do cover areas around the same location and offset as the SNe, and cover reasonable fractions of their respective hosts. As a check, we also plot in \cref{fig:Zall} the metallicity distribution of SF in our MUSE sample only including observations that cover approximately all the host galaxy (i.e., we exclude observations of SNe 2005P, 2010el, 2012Z, 2013gr and 2014dt; see \cref{fig:zstamps}). The distribution for this sub-sample of our MUSE galaxies is almost identical to the full MUSE sample (as was found in our other environmental measure also) and so the partially covered hosts do not introduce significant biases into the distributions.
We find that the metallicity of SNe~Iax sites are at lower metallicities than where typical SF is occurring in the hosts (in each metallicity indicator shown the KS test gives $p \sim 0.01$ that the two are drawn from the same distribution). As expected, given the SN explosion sites are offset from their host nuclei and we find mostly negative radial metallicity gradients in our hosts, there is an offset between the host nuclei and explosion site metallicities.
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{Zall_D16}}\\
\subfloat{\includegraphics[width=\columnwidth]{Zall_M13}}
\caption{The metallicity distributions of SNe~Iax explosion sites, their host nuclei, and all detected H{\sc ii} regions in the MUSE sample. Metallicities are determined using the calibrations of \citetalias{dopita16} and \citetalias{marino13} (O3N2) in the top and bottom panels, respectively. When appropriate, SN explosion site metallicity estimates based on the gradients of the hosts are included (see \cref{sec:zgrads}). Dotted lines are the distributions after adding/subtracting $1\sigma$ statistical uncertainties to/from all values and are somewhat extreme limits for the true distribution. The cumulative sum of SN explosion sites and host nuclei are unweighted (i.e. a step in these histograms equals one SN or host nucleus, as appropriate). The H{\sc ii} regions have been weighted by their \text{H}\ensuremath{\alpha}{} luminosity such that this histogram shows the cumulative fraction of all detected SF with metallicity across the MUSE sample. The thin orange dashed line shows the sub sample of MUSE \ion{H}{ii}{} regions for which the observations covered the whole host and appears almost indistinguishable from the full \ion{H}{ii}{} region sample (see text).}
\label{fig:Zall}
\end{figure}
\subsubsection{Metallicity ranking of explosion sites}
\label{sec:reszrank}
In order to further assess the indications of \cref{sec:reszdist}, we have calculated SFR-weighted metallicity NCR ($Z$NCR) values for the explosion sites of out MUSE sample of SNe~Iax. The $Z$NCR is calculated in a similar manner to the traditional NCR value of SNe \citep[\cref{sec:pixelstat}, see also][]{james06, anderson12}, but the cumulative distribution of the SFR-weighted metallicity in the galaxy is used to determine the rank of each SN. The $Z$NCR is thus the fraction of stars being formed at or below the metallicity of the explosion site in that host. This may be seen visually in the top right panel of \cref{fig:d16map}, where the SN in question has $Z$NCR = 0.003. In \cref{fig:ZNCR} we show the distributions of $Z$NCR for \citetalias{dopita16} and \citetalias{marino13}. In the case of metallicity-unbiased formation of the progenitors, a uniform distribution (indicated by dashed lines) should be recovered, i.e. the SNe have no dependence on where in the metallicity distribution of SF they form, and so sample it uniformly. We see for both distributions the SN explosion sites are systematically shifted to lower values than the uniform distribution, indicating the SNe are preferentially exploding in metal poor regions, and not unbiasedly tracing the SF in their hosts. For \citetalias{marino13}, although the distribution is systematically below the uniform distribution, the difference is more marginal than for \citetalias{dopita16}.
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{ZNCR}}
\caption{The SFR-weighted metallicity NCR ($Z$NCR) values of SNe~Iax explosion sites in two metallicity indicators, as observed with MUSE. In the case of metallicity-unbiased production of the progenitors, the SNe should create a uniform distribution in $Z$NCR (thick dashed lines). Uncertainties on the distributions were found by creating many realisations of the cumulative sums (based on the uncertainties of the values) for each SN host and recalculating each explosion site's rank. Dotted lines represent the lower and upper distributions of the 95~per~cent confidence interval of the ranks.}
\label{fig:ZNCR}
\end{figure}
\subsection{Ages and SFR}
\label{sec:resultsages}
We show the EW(\text{H}\ensuremath{\alpha}{}) and L(\text{H}\ensuremath{\alpha}{}) measurements for our explosion site bins in \cref{fig:HaEWall,fig:Halumall}, respectively. As was done for the metallicities, we also show the cumulative distribution of SF in these measurements for our MUSE-observed hosts. The regions of SF at the locations of SNe~Iax are occurring at significantly lower EW and luminosities than the overall SF in the hosts (KS test $>3\sigma$ for each). Adopting a \text{H}\ensuremath{\alpha}{} luminosity--SFR relation \citep{kennicutt98}, the \text{H}\ensuremath{\alpha}{} luminosities imply that the explosions sites are typically of lower SFR than that of the population of SF regions in the hosts, ranging from $10^{-4}-10^{-2}$ M$_\odot${}~yr$^{-1}$.
As discussed in \cref{sec:methages}, ages for the young stellar component of star forming regions can be estimated from EW measurements of emission lines but are subject to sources of uncertainty. The bulk of our explosion site EW(\text{H}\ensuremath{\alpha}{}) measurements are $\sim$tens, up to $\sim 100$~\AA{}. We show in \cref{fig:bpassew} (see \cref{sec:methages}) the evolution of EW(\text{H}\ensuremath{\alpha}{}) with age for {\sc bpass} models for an instantaneous SF episode at a range of metallicities. As the consensus is that the vast majority of (particularly massive) stars are in some form of binaries \citep[e.g.][]{sana12} we compare results with the binary population models.
Our observed range of EW(\text{H}\ensuremath{\alpha}{}) suggests young SP components of several~$\times 10^7$ to $10^8$~yr at the locations of SNe Iax. We again note the inherent uncertainties in selecting a single fiducial model for the nebular gas properties, however these values indicate that there are moderately young SP components at the location of the majority of SNe~Iax explosion sites.\footnote{Although we were not able to measure emission lines at the locations of some NOT-observed examples, these data were comparatively shallower than the MUSE data, where signatures of ongoing SF were found for all but one explosion site (SN~2008ge). This is demonstrated by our duplicate observations of SN~2015H: in the NOT data we were not able to extract an emission line spectrum at the explosion site but we could in the MUSE data.} Notably, although our EW values are actually somewhat lower limits of the true young SP EW (as there will be an existing underlying, older SP that contributes to the continuum but not emission lines), they are not exceptionally high. This would seem to disfavour very young, and therefore very massive, SPs at their locations with ages of several Myr, since initially very high EW values tend to drop off quickly and largely independent of model differences and reasonable gas parameters.
As an additional check for the presence of very young SPs at the location of SNe~Iax we also attempted fitting for He {\sc i} $\lambda 4922$ in our emission fitting routine (\cref{sec:emisfitting}) for the MUSE sample.\footnote{Our comparatively shallower NOT data with lower spectral resolution is not conducive to providing meaningful detections of this faint line, which is close to other, much stronger features.} He {\sc i} $\lambda 4922$ in emission is present only for ages up to a few Myr \citep[e.g.][]{gonzalezdelgado99}. We did not detect this line at any SN~Iax explosion site. Within our MUSE data we found detections of He {\sc i} $\lambda 4922$ within 3~kpc of the SN for SNe~2008ha, 2009J, 2010ae, 2012Z and 2013gr. Assuming a 5~Myr age for the young SP (also consistent with EW(\text{H}\ensuremath{\alpha}{}) $\gtrsim 100$ found in these regions), the average velocities of the SN~Iax progenitors would have to have been in excess of 234, 128, 269, 580, 159 km~s$^{-1}${}, respectively, in order to have originated from these regions.
Two events with higher EW(\text{H}\ensuremath{\alpha}{}) measurements at their explosion sites are debated members of the SN~Iax sample as they displayed helium and perhaps hydrogen in their spectra. The removal of these events (SNe~2004cs and 2007J) would lean the typical ages of the young SPs at the locations of the rest of the sample slightly higher (\cref{sec:hetero}).
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{HaEWall}}
\caption{As for \cref{fig:Zall} but here showing EW(\text{H}\ensuremath{\alpha}{}) of SN~Iax explosion sites.}
\label{fig:HaEWall}
\end{figure}
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{Halumall}}
\caption{As for \cref{fig:Zall} but here showing L(\text{H}\ensuremath{\alpha}{}) of SN~Iax explosion sites. The top axis indicates the corresponding SFR \citep{kennicutt98}.}
\label{fig:Halumall}
\end{figure}
\subsection{Offsets}
\label{sec:resultsoffsets}
Offsets of SNe~Iax and \ion{H}{ii}{} regions are shown for kpc and host-normalised values in \cref{fig:Offsetall}. SNe~Iax appear to trace a similar offset distribution as the overall SF of their hosts when accounting for the varying sizes of the host galaxies, although shifted systematically to slightly larger offsets. Since the outer regions of late type galaxies are likely to be more metal poor and less intensely star forming, this may be a contributing factor to the difference we observe in our other measurements.
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{Offsetall}}\\
\subfloat{\includegraphics[width=\columnwidth]{Offsetr25all}}
\caption{As for \cref{fig:Zall} but here showing the deprojected galactocentric offsets of SN~Iax explosion sites. The offsets are given in kpc and normalised to the host galaxies' $R_{25}$ values in the top and bottom panels, respectively.}
\label{fig:Offsetall}
\end{figure}
\section{Discussion}
\label{sec:discussion}
\subsection{Correlation between environment and SN properties}
Tying the diversity of progenitor systems to the observed properties of SNe is an open issue amongst all SN types. It may be expected that some imprint of the nature of the progenitor (which can be determined via direct detection or, as here, inferred through environmental analysis) is evident in the SN light curve and spectra.
To investigate this for SNe~Iax we plot SN properties, where they have been measured in the literature, against our environmental measures in \cref{fig:snprops}. The light curve peak ($M_\textrm{peak}$) and decline rate ($\Delta m_{15}$) in $R$- or $r$-band, as well as an estimate of the photospheric velocity around peak light ($v_\textrm{ph}$) are shown versus the \citetalias{dopita16} metallicity, EW(\text{H}\ensuremath{\alpha}{}) and the host-normalised offsets. Values and references for SN properties are provided in \cref{tab:litsn}. In order to expand our comparison sample of SN properties we include preliminary analyses for SNe~2010el, 2013gr and 2014ey (Stritzinger et al., in preparation)\footnote{For the three SNe we find $M_\textrm{peak,r} \simeq [-15.4, -15.2, -18.1]$~mag, $\Delta m_{15,r} \simeq [1.08, 0.99, 0.62]$~mag and $v_\textrm{ph} \simeq [3000, 5300, 5000]$~km~s$^{-1}${}, respectively. For SN~2010el we assign an uncertainty of 0.5~mag on the peak owing to the quite uncertain distance to NGC~1566.} -- these objects are to be the subject of more detailed studies in preparation, but our values here are representative.
\begin{table*}
\centering
\begin{threeparttable}
\caption{Values for literature SN properties.}
\begin{tabular}{lccccl}
\hline
SN name & Filter & $\Delta m_{15}$ & $M_\textrm{peak}$ & $v_\textrm{ph}$ & Reference \\
\hline
2002cx & $R$ & 0.54$\pm$0.06 & -17.64$\pm$0.15 & 5600 & \citet{foley13} \\
2003gq & $R$ & 0.71$\pm$0.10 & -17.37$\pm$0.15 & 5200 & \citet{foley13} \\
2004cs & $R$ & 1.11$\pm$0.07 & -16.55$\pm$0.45 & \ldots & \citet{magee16} \\
2005cc & $R$ & 0.65$\pm$0.01 & -17.13$\pm$0.15 & 5000 & \citet{foley13} \\
2005hk & $r$ & 0.70$\pm$0.02 & -18.07$\pm$0.25 & 6000 & \citet{stritzinger15} \\
2007qd & \ldots & \ldots & \ldots & 2800 & \citet{foley13} \\
2008A & $R$ & 0.51$\pm$0.01 & -18.23$\pm$0.15 & 6400 & \citet{foley13} \\
2008ae & $r$ & 0.71$\pm$0.13 & -17.76$\pm$0.16 & 6100 & \citet{foley13} \\
2008ha & $R$ & 0.97$\pm$0.02 & -14.41$\pm$0.15 & 3700 & \citet{foley10a} \\
2009J & $R$ & 0.79$\pm$0.05 & -15.47$\pm$0.22 & 2200 & \citet{foley13} \\
2009ku & $r$ & 0.25$\pm$0.03 & -18.70$\pm$0.15 & 3300 & \citet{foley13} \\
2010ae & $r$ & 1.01$\pm$0.03 & -14.59$\pm$0.81 & 5500 & \citet{stritzinger14} \\
2010el & $R$ & 1.08$\pm$0.03 & -15.43$\pm$0.50 & 3000 & This work \\
2011ay & $r$ & 0.44$\pm$0.14 & -18.43$\pm$0.19 & 5600 & \citet{foley13} \\
2012Z & $r$ & 0.66$\pm$0.02 & -18.60$\pm$0.09 & 8000 & \citet{stritzinger15} \\
PS1-12bwh & $r$ & 0.60$\pm$0.05 & -17.69$\pm$0.24 & 5700 & \citet{magee17} \\
2013gr & $r$ & 0.99$\pm$0.20 & -15.17$\pm$0.50 & 5300 & This work \\
2014ck & $r$ & 0.58$\pm$0.05 & -17.29$\pm$0.15 & 3000 & \citet{tomasella16} \\
2014dt & \ldots & \ldots & \ldots & 4100\tnote{\dag} & \citet{foley15} \\
2014ey & $r$ & 0.62$\pm$0.02 & -18.13$\pm$0.10 & 5000 & This work \\
2015H & $r$ & 0.69$\pm$0.04 & -17.27$\pm$0.07 & 5500 & \citet{magee16} \\
PS~15csd & $r$ & 1.06$\pm$0.06 & -17.75$\pm$0.06 & \ldots & \citet{magee16} \\
\hline
\end{tabular}
\label{tab:litsn}
\begin{tablenotes}
\item[\dag] Based on a similarity to SN~2002cx post-peak.
\end{tablenotes}
\end{threeparttable}
\end{table*}
From these plots we see that brighter ($M_R \lesssim -16$~mag) members appear to cover almost the full range of the metallicity, EW and offset distributions as found for the full sample, with large spreads in each parameter. Although we have fewer fainter members with peak absolute peak magnitude determinations, we find none at large galactocentric offsets. This trend with galactocentric offset cannot be attributed to observational biases as these work in the opposite direction (very faint transients are more difficult to detect in the brighter central regions of galaxies) and so may actually be more pronounced than is shown.
The faint members appear to cluster at a very small range in EW(\text{H}\ensuremath{\alpha}{}) of $\sim$ 25--40 \AA{}, perhaps indicating more strict age constraints for these members, assuming they arise from the local young stellar population. Their metallicities appear diverse although most appear quite metal-poor.\footnote{For another member, SN~2007qd, \citet{mcclelland10} find a faint peak magnitude ($M_R \simeq -15.8$~mag), although this is uncertain since the light curve does not have full coverage \citep{foley13}. We note for this event we only have a poorly constrained (2 regions) metallicity gradient from which to estimate the \citetalias{dopita16} abundance as 8.60.}
Evidence has been presented for a relation between $M_R$ and $\Delta m_{15}$ for SNe~Iax, but any relation is certainly less tight than that seen for normal SNe~Ia and with notable outliers \citep[e.g.][]{mcclelland10,narayan11,foley13}. The slowly fading ($\Delta m_{15} \lesssim 0.75$~mag) members of our sample occupy a wide range of environments, with the faster declining members generally occupying more restricted ranges, analogous to the trends seen for the bright and faint members. In particular we observe distinct clustering in metallicity with $\Delta m_{15}$ -- slowly declining members are systematically more metal-rich than the faster decliners (barring the fast declining, metal-rich SN~2010el).
A composite figure showing the $M_R$ and $\Delta m_{15}$ parameter space for SNe~Iax now coded by the explosion site metallicity is shown in \cref{fig:mm15}.
SN~2004cs (based on unfiltered imaging) was a fast declining member although somewhat brighter than the other fast-decliners in this plot, and shows a significantly larger EW(\text{H}\ensuremath{\alpha}{}) than the other fast decliners (or any of the sample for which we have light-curve information). This was one of two proposed SN~Iax members that showed He features, and is further discussed in \cref{sec:hetero}.
We do not observe any strong clustering or correlated behaviour between our environmental measures and the estimates of photospheric velocity near peak. We caution here, however, that these values have been determined using a variety of methods (spectral synthesis, line measurements based on differing elements) at slightly varying epochs around maximum light, and there thus may be some systematics within the sample. We assign 1000~km~s$^{-1}${} uncertainties when plotting, which are likely to be overestimates for individual measurements but account better for potential systematics arising from differing methods.
\begin{figure*}
\centering
\subfloat{\includegraphics[width=\linewidth]{propsvsenvs}}
\caption{The properties of SNe~Iax and their environments. $M_\textrm{peak}$ and $\Delta m_{15}$ values are given in the $R$- (red squares), $r$-band (red circles) or unfiltered (for SN~2004cs, empty squares). $v_\textrm{ph}$ values are estimates of the photospheric velocity.}
\label{fig:snprops}
\end{figure*}
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{mm15}}
\caption{$R$- and $r$-band $M_\textrm{peak}$ versus $\Delta m_{15}$ for SNe~Iax \citep[e.g.][]{narayan11, foley13, white15, tomasella16} in our sample, now colour coded by their environmentally-derived metallicities in the indicator of \citetalias{dopita16} where possible. SN~2004cs is highlighted as the classification of this event is debated (\cref{sec:hetero}).}
\label{fig:mm15}
\end{figure}
\subsection{SNe~Iax environments in the context of other transients}
\label{sec:sncompare}
Despite sharing some similarities to SNe~Ia in the general spectral classification sense, the host galaxies of SN~Iax are almost exclusively (barring SN~2008ge) late type and star-forming \citep[e.g.][]{perets10,foley13} and the transients appear to be associated with regions of ongoing SF \citep[this work;][]{lyman13}. This is the case for CCSNe and thus a comparison between our results here and those of other SN types in the literature can inform on similarities or differences in the progenitor environments.
The construction of our comparison SN samples do not constitute unbiased, blindly-targeted events by any means. However, the same holds for the SN~Iax sample, which were discovered over a variety of surveys. Our comparison is thus limited to an initial, indicative comparison until such time as a reasonable sample of homogeneously discovered SNe~Iax exists.
\subsubsection{Metallicity}
As we are working with literature samples, there are almost no abundance measurements in the \citetalias{dopita16} scale owing to its recent inclusion in the literature. We therefore collect available values based on the O3N2 indicator (these were generally presented in the calibrations of \citealt{pettini04} and have been translated to those of \citetalias{marino13}). We used only values based at or nearby the SN explosion site, or based on determined gradient values (i.e. we exclude those where the metallicity was determined based on the host nucleus but the SN was at a significant offset and not covered by the slit/fibre).
We take values from the analyses of \citet{anderson10,modjaz11,leloudas11,sanders12,kuncarayakti13a,kuncarayakti13b,stoll13} and \citet{galbany16}.
In the case of SNe~II (including both II-L and II-P) we take values from \citet[][which includes the results of \citealt{anderson10}]{anderson16} and \citet{galbany16} directly in \citetalias{marino13}, and values from \citet{stoll13}. On the values of \citet{anderson16} we impose a distance cut of 3~kpc on the \ion{H}{ii}{} region---SN distance.
We present SNe Ib and Ic separately, however uncertain classifications mean there is likely to be some cross-contamination. Where appropriate we use the updated classifications of \citet{modjaz14, shivvers17} and count still uncertain `Ib/c' designations with half weight in each cumulative distribution.
We plot the explosion-site metallicity distributions for SNe~Iax and our comparison samples in \cref{fig:Zcompare}. There is no significant difference between SNe~Iax explosion-site metallicities and those of SNe Ib, Ic, II or IIb and the median value for SNe~Iax (8.41 dex) is close to that of the others (Ib: 8.42 dex, Ic: 8.49 dex, II: 8.47 dex, IIb: 8.39 dex). The distribution for SNe~Iax covers broadly the same range as these other SN types, perhaps not extending as metal-rich or -poor as SNe~Ib, Ic or II, however there are only a very small number of events in these samples on the extreme edges of their distributions. Compared to the SNe~Ia distribution \citep[local metallicity measurements explosion sites from IFS data taken from][]{galbany16}, median = 8.53, we see SNe~Iax are metal-poor (KS test Ia vs Iax metallicities $p = 5\times10^{-4}$).
Although the overall SN~Iax distribution appears to follow broadly the distributions of other well-known SN types, barring SNe~Ia, it is overall quite different from the distribution of low redshift LGRB explosion sites, which are in the range $Z_\textrm{O3N2} \simeq 8.0-8.4$ dex \citep[e.g.][]{modjaz08, sanders12}. Similar very low metallicities have also been found for the the vast majority of super-luminous supernova hosts (e.g. \citealt{lunnan14,leloudas15}).
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{Zcompare_M13}}
\caption{The explosion site metallicity distributions of SNe~Iax compared to literature samples of other principal SN types. Metallicities are based on the O3N2 indicator of \citetalias{marino13}.}
\label{fig:Zcompare}
\end{figure}
\subsubsection{Offsets}
$R_{25}$-normalised offset distributions for SNe Ib, Ic, II and IIb were taken from \citet{taddia13}. SNe~Ia offsets were taken from the catalogue of \citet{hakobyan16} and calculated following the prescription described therein. The SNe~Ia host-normalisation are based on $g$-band $R_{25}$ measurements as opposed to $B$-band measurements used for the other samples. This may introduce a systematic offset in the values, however, given the considerable filter overlap and the modest variation of apparent galaxy sizes around these wavelength ranges \citep{vulcani14}, the effect is likely to be small. SNe~Ia in the catalogue are those that exploded in disk galaxies (types S0-Sm).
Similarly to the metallicities, we see in \cref{fig:offsetcompare} that the offsets of SNe~Iax closely match other SNe types, in particular SNe~Ia and Ib. Offsets do not appear to be a strong factor in distinguishing SNe~Iax, whereas their host galaxy type distribution proves more discriminatory against that seen for SNe~Ia \citep{perets10,lyman13}.
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{Offsetcompare}}
\caption{The explosion site galactocentric offset distributions of SNe~Iax compared to literature samples of other well-known SN types. Offsets have been normalised by their hosts' respective $R_{25}$ values.}
\label{fig:offsetcompare}
\end{figure}
\subsubsection{Pixel Statistics}
\label{sec:pixelstatres}
An analysis of the NCR (\cref{sec:pixelstat}) distribution of SNe~Iax was presented by \citet{lyman13}, who found the association of SNe~Iax to SF in their hosts was at a similar level to that of SNe~IIP and inferred similar progenitor ages (tens of Myr). As we are able to create \text{H}\ensuremath{\alpha}{} maps from our MUSE data cubes we present an updated comparison figure with this extended combined sample of SNe~Iax in \cref{fig:ncr}. Other SN sample data are taken from \citet{anderson08, anderson12, anderson15} and, following \citet{lyman13}, we have corrected the SN~Ia sample to account for the fact that only late-type galaxy hosts were used but $\sim$27~per~cent of SNe~Ia explode in early-type hosts \citep{li11}, with no appreciable ongoing SF. For SNe~2005P, 2008ha, 2009J and 2012Z we use values from \citet{lyman13} as \text{H}\ensuremath{\alpha}{} imaging presented there covers the full spatial extent of the hosts. The NCR values for these determined in \citet{lyman13} and here are [0.06, 0.41, 0.00, 0.00] and [0.01, 0.33, 0.06, 0.03], respectively.
Following \citet{lyman13}, we confirm for a larger sample that SNe~Iax display a level of association to the SF of their hosts most similar to that of SNe II. Their distribution is formally discrepant, based on the KS test, with the SNe~Ia and Ic distributions at $3\sigma$ (considering only SNe~Ia in late-type hosts and excluding SN~2008ge from the SN~Iax sample, the SN~Ia vs Iax discrepancy is $p = 0.05$).\footnote{Anderson-Darling tests provided very similar levels of significance, with slightly higher significance for the uncorrected SNe~Ia vs Iax ($p \simeq 0.006$).} Although there are few \text{H}\ensuremath{\alpha}{} NCR values for LGRBs (due to their typically much larger distances), studies have shown these to be very strongly associated to the brightest SF regions of their hosts, at a level exceeding SNe~Ic \citep{fruchter06, svensson10,lyman17}. Assuming the same would hold for SF as traced by \text{H}\ensuremath{\alpha}{}, this infers drastically different environments of SNe~Iax compared to LGRBs.
We note that although we detected \text{H}\ensuremath{\alpha}{} at the locations of all our SNe~Iax in the MUSE sample, we have some where NCR = 0.00. In order to facilitate comparison to literature samples, which were performed using narrow band imaging, we construct \text{H}\ensuremath{\alpha}{} narrow band images (filter width 30\AA{}) and subtract the neighbouring continuum within our data cubes (\cref{sec:hiibinning}). As such, very faint sources of emission may become dominated by shot noise or inaccuracies in continuum subtraction during the \text{H}\ensuremath{\alpha}{} map construction, leaving the pixel within the noise floor of the image \citep[defined as NCR=0, see][for more thorough discussion]{james06,anderson08}. Additionally, the spatial extend of our spaxel bins is larger than the $3 \times 3$ binning employed for the NCR calculations in order to replicate the method used in the literature.
\begin{figure}
\centering
\subfloat{\includegraphics[width=\columnwidth]{ncr}}
\caption{The explosion site Normalised Cumulative Rank (NCR) distributions of SNe~Iax compared to literature samples of other well-known SN types. NCR values were calculated using \text{H}\ensuremath{\alpha}{} maps from our MUSE data cubes and supplemented with values presented in \citet{lyman13}.}
\label{fig:ncr}
\end{figure}
\subsection{Ionised gas in the early-type host of SN~2008ge}
\label{sec:08ge}
Although there are likely to be contributions from different progenitor channels to such a diverse class of objects (see \cref{sec:hetero}), one SN in the Iax sample has been marked out by its unique environment and host. SN~2008ge was hosted by NGC~1527, an early, weakly-barred S0 galaxy \citep{deVauc91}, in contrast to the strongly star-forming, late-type hosts of other known Iax \citep{perets10, lyman13}. Motivated by this, \citet{foley10b} investigated the host galaxy and explosion site for signs of a young SP. They concluded from a smooth galaxy profile, non-detection of far-infrared or H{\sc I} 21cm emission in the host, and the lack of narrow nebular emission lines in their host galaxy spectrum that there is no significant ongoing SF (and thus young SP) in NGC~1527.
Our MUSE observations provide deep optical spectra, spatially resolved, across a reasonable fraction of the central regions of the host galaxy. After an initial pass with our emission line fitting as presented in \cref{sec:musemethods}, we reran the emission line fitting this time fitting only for \text{H}\ensuremath{\alpha}{} and [\text{N {\sc ii}}]{}~$\lambda$6548,6583 in order to determine a subset of the bins with evidence for ionised emission lines. We then attempted to fit \text{H}\ensuremath{\beta}{} and [\text{O {\sc iii}}]~$\lambda$5007 for these bins. In accordance with \citet{foley10b}, we found no significant emission lines at the location or in the close vicinity of the SN explosion site (either within the Voronoi bin directly underlying the explosion site or by extracting circular apertures of varying sizes). However, within our MUSE data we detected an ionised gas stream arcing from the NE of the nucleus with evidence for a further region of emission to the SE of the host nucleus. As NGC~1527 is particularly nearby ($\sim$ 17 Mpc) the MUSE observations only cover the central $\sim$2--3 kpc around the nucleus. We have summed the \text{H}\ensuremath{\alpha}{} flux captured by the MUSE FOV and determine $\log_{10}$(L(\text{H}\ensuremath{\alpha}{}) = 37.6~erg~s$^{-1}$ and note that this is not corrected for the effects of reddening as we were unable to detect \text{H}\ensuremath{\beta}{} in many of the bins. For the case that this emission is driven by ongoing SF, the relation of \citet{kennicutt98} suggests the detected emission equates to $\log_{10}$(SFR)~=~-3.48 M$_\odot${}~yr$^{-1}$. This is consistent with the limit that \citet{foley10b} provide for NGC~1527 of $\log_{10}$(SFR)$ < -2.14$~M$_\odot${}~yr$^{-1}$. The metallicity (for SF-driven ionisation) appears to be around solar to slightly sub-solar based on the N2 indicator \citepalias{marino13}. The EW(\text{H}\ensuremath{\alpha}{}) of the bins are $\sim 0.8$\AA{}.
Morphologically the main stream appears somewhat coherent, in a thin parabolic stream about the nucleus. The emission looks different to the more extended, and generally more symmetric, spiral arms seen in some early-type galaxies \citep[ETG; see][and references therein]{gomes16a}. The nature of the ionising sources in ETGs is debated and there appears to be contributions from a variety of phenomena, with different mechanisms dominating in different galaxies \citep[e.g.]{goudfrooij99,sarzi06,sarzi10}. Unfortunately, we are limited in our analysis for the ionising source based on emission line flux ratios as we are close to our detection limits and thus flux ratios are not well constrained. Within \cref{fig:08ge} we plot the BPT diagram for bins where all lines were detected at SNR > 1. Although the position of each bin is rather uncertain within this diagram, there appears to be a general clustering of the bins around or above the maximal star-formation relation of \citet{kewley01}, indicative that ongoing SF may not be the dominant process driving this emission. (Those bins that are more consistent with being ionised solely due young, massive stars are those SE of the nucleus, not in the main arc). With a comparison to the study of \citet{sarzi10}, the stronger line ratios and morphology may be more reminiscent of emission due to shocks in the galaxy, although these are not expected to be dominating sources of ionisation in ETGs. Alternatively it may be due to diffuse ionising SPs, such as post-asymptotic giant branch (pAGB) stars.
The ionising flux from pAGB stars can lead to line ratios similar to those seen for low-ionization emission-line regions \citep[LIERs][]{binette94,sarzi06}, and the bulk of the bins in \cref{fig:08ge} are located close to LIER areas of the parameter space \citep[e.g. $\log_{10}\frac{[\text{N {\sc ii}}]{} \lambda6583}{\text{H}\ensuremath{\alpha}} \gtrsim -0.4$ and EW(\text{H}\ensuremath{\alpha}{})~$< 6$\AA{},][]{cidfernandes11}. One would, however, expect a pAGB population to be more or less pervasive across the galaxy and follow the distribution of stellar mass. This results in the correlation between ionised line flux and stellar continuum flux seen in ETGs with pAGB ionising sources \citep{sarzi10}. A more structured ionised region (as seen here) could therefore be due to variations in gas column density within the galaxy.
The ionising output and subsequent EW(\text{H}\ensuremath{\alpha}{}) predictions from various models for pAGB star contributions predict roughly constant low values from $\sim 10^8$ to $10^{9.5}$~years \citep{cidfernandes11}, meaning such values do not offer strong constraints on the relative age of the underlying population.
\begin{figure*}
\centering
\subfloat{\includegraphics[width=\linewidth]{SN2008ge_plot}}
\caption{The central ionised gas component of NGC~1527 (the host of SN~2008ge). {\em Left:} A heatmap of the metallicity (using the N2 relation of \citetalias{marino13}) for each bin where both \text{H}\ensuremath{\alpha}{} and [\text{N {\sc ii}}]{}~$\lambda$6583 has SNR~$> 2$, overlaid on an inverted $R$-band image of the galaxy; North is up, East is left. The adopted location of the nucleus and the SN explosion site are indicated by the black $\times$ and `star' symbols, respectively. {\em Left, inset:} An inverted, continuum-subtracted [\text{O {\sc iii}}]{} image of the central region of NGC~1527. Each Voronoi bin has had its best fitting stellar continuum model (\cref{sec:contfitting}) subtracted. The residual ionised gas component is evident {\em Top right:} A BPT diagram for those bins with detected emission. The format is the same as shown in \cref{fig:bptmap}. {\em Bottom right:} The observed spectrum (black) with the {\sc starlight} fit to the continuum (orange, dashed) and the residual emission (purple) around \text{H}\ensuremath{\alpha}{} and [\text{N {\sc ii}}]{}. This is shown for the SN explosion site, where we do not detect any significant emission lines, and in one of the bins NE of the galaxy nucleus, from which we were able to measure a metallicity. Wavelengths are in the rest-frame of the nucleus of NGC~1527.}
\label{fig:08ge}
\end{figure*}
\subsection{Heterogeneity of the class and contaminants}
\label{sec:hetero}
As is the case for many astrophysical transients, in particular for peculiar and relatively rare objects, membership to a particular class (or even designating a single class) can be contentious. Often the underlying continua of the events properties are at odds with defining distinct regions of parameter space to assign one class or another. This is exemplified in the case of SNe Iax where multiple progenitor channels and explosion mechanisms may be contributing to empirically similar transients.
Two examples in the sample of \citet{foley13} showed evidence for helium in their spectra: SNe 2004cs and 2007J. Helium can remain hidden in SN spectra as it is difficult to ionise and so it is possible that helium is present in the ejecta of other members, however at the time of writing these two members remain the only posited SNe~Iax to show detections. One potential confusion, as highlighted by \citet{foley13}, is with SN~2005E-like events \citep[][also known as Ca-rich transients/SNe]{perets10} -- these similarly faint-and-fast events display helium features in their spectra before quickly evolving to unusually Ca-dominated nebular spectra \citep{kasliwal12}. Instead \citet{white15} favour the original classifications of SNe~2004cs and 2007J as CCSNe. In addition to helium they also find evidence for hydrogen in their spectra, prompting classifications of SNe~IIb. However, a reanalysis by \citet{foley16} argued against the presence of hydrogen in SN~2007J, citing an identification as [Fe{\sc ii}] of the same feature in SN~2002cx. Furthermore, they find inconsistencies between the light curve of SN~2004cs and other known SNe~IIb.
As we are discussing only 2 objects, statistical inferences from environmental measures are weak. This is further exacerbated by the fact that our measures overlap significantly for SNe~Iax and SNe~IIb (\cref{sec:sncompare}). For completeness we note EW values at the helium-SNe~Iax explosion sites are the among the highest (1st and 4th) of our sample, pointing to the presence of younger SPs at their explosion sites than typical SNe~Iax. Their offsets ($\simeq 0.5 \times R_{25}$) and metallicities (O3N2~$\simeq 8.4$) are typical of the samples of both SNe~Iax and IIb. Their exclusion would revise our estimates of the typical ages of the young SP at SN~Iax explosion sites slightly higher, although the quantitative effect is likely to be dwarfed by the inherent uncertainties present in determining ages from nebular gas spectra \citep[][\cref{sec:methages}]{stanway14b}, and would not significantly affect our discussion elsewhere.
SN~1999ax was classified as a ``somewhat peculiar'' SN~Ia by \citet{galyam08} at $z \simeq 0.05$. On the basis of SDSS spectra of a nearby potential host showing it to be at $z = 0.023$, \citet{foley13} re-analysed its spectrum and classified it as a SN~Iax based on similarity to SN~2002cx post-peak. We note our NOT spectroscopy of the explosion site also covered another nearby object in SDSS that is at a similar (projected) offset from SN~1999ax as the galaxy at $z = 0.023$ (with the SN being located between the two). We find this galaxy is at $z = 0.059$ based on the \text{H}\ensuremath{\alpha}{} line in the NOT spectrum. As the SN is located directly between the two potential hosts it is difficult to distinguish the likely host from environments alone. The velocities of the features in the spectrum appear low compared to normal SN~Ia, and a consideration of the SN at $z \simeq 0.05$ still poses peculiarities to a typical SN~Ia classification, as such we retain the SN~Iax classification making the lower redshift galaxy the probable host.
\subsection{Implications for progenitor models}
\label{sec:progmodels}
Pure deflagrations \citep[e.g.][]{branch04,jha06,phillips07,jordan12,kromer13,fink14,magee16}, pulsational-delayed detonations \citep{stritzinger15} and helium-ignition \citep{wang14} of WDs have been presented as progenitor models to explain SNe~Iax. Although too weak to account for normal SNe~Ia, they broadly agree with the low $^{56}$Ni masses and kinetic energies displayed by SNe~Iax. There are currently limited predictions as to the progenitor environments for these based on population studies,
with the main constraints in the form of delay-time distributions and age constraints.
Any double-degenerate channel (either merger or accretion between two WDs), would require circumstances to explain the young environments since such progenitor systems would be expected to be prevalent in old SPs and early-type galaxies, inconsistent with the locations of SNe~Iax as a whole. As such, single-degenerate formation channels are generally favoured for SN~Iax.
\citet{meng14} found that delay times as low as 30~Myr were possible for massive hybrid CONe WDs accreting material until the Chandrasekhar mass, and would likely produce lower-luminosity events (cf. normal SNe~Ia) making them an attractive possibility for SNe~Iax.
\citet{wang13} explore detonation of CO WDs via ignition in a helium envelope that is accreted from a He companion. For the case of a non-degenerate companion their delay times to explosion are $\sim 10^8$~yrs, consistent with the ages of young SPs at the location of most SNe~Iax. If the companion is a He WD the delay time can be significantly extended (up to the Hubble time), potentially providing an explanation for SNe~Iax in old populations (e.g., SN~2008ge) or those with no detected signs of SF at the explosion site and in the outskirts of their hosts (e.g., SN~2008A). We note that these are for solar metallicity populations, which is metal-rich compared to the SN~Iax population (\cref{fig:Zall}). Similarly, \citet{liuzw15a} find that the WD + He star channel best reproduces the young ages of SN~Iax explosion sites, but note an extended delay time exists for main sequence or red giant companions.
Our age constraints on SNe~Iax seem to be in good agreement with predictions of young explosion scenarios of single-degenerate thermonuclear progenitor models.
\subsubsection{Faint and fast `SN~2008ha-like events'}
\label{sec:08halike}
The faintest, lowest-energy members of SNe~Iax could represent a distinct population of events, with the most famous example being SN~2008ha \citep[see][for further discussion of models in light of the physical properties of the SNe]{foley09}. These very faint examples are difficult to produce with models proposed for other SNe~Iax \citep{fink14}, perhaps indicating a single model is not able to explain the full diversity of all events labelled as such. For the purposes of this section we simply refer to faint ($M_R > -16$~mag) and (/or) fast evolving ($\Delta m_{15} \gtrsim 1$~mag) members of the sample, ignoring a detailed spectroscopic distinction.
\citet{valenti09} suggested SN~2008ha and, by extension, other SNe~Iax were of core-collapse origin. Our environmental analyses indicate that the fainter members arise from low metallicity regions, quite centrally located on their host galaxies (\cref{fig:snprops}). From an environmental viewpoint, their distributions are largely consistent with those seen for well-known CCSN types. The question of whether they are due to lower mass, $\sim 8$~M$_\odot${}, stars stripped of their hydrogen envelopes or very massive WR stars, probable progenitors of LGRBs, experiencing significant fall-back, was posed by \citet{valenti09}. Our analysis of the ages of the youngest SPs here suggests ages of order tens of Myr for their explosion sites (\cref{sec:resultsages}). This is in agreement with resolved SP studies of explosion sites of nearby SNe~Iax \citep{foley14,mccully14} and the comparative association of SNe~Iax to \text{H}\ensuremath{\alpha}{} emission in their host galaxies at the level of SNe~II \citep[\cref{sec:pixelstatres}, ][]{lyman13}. For WR stars, however, we may expect much younger ages, $\lesssim 10$~Myr (for stars with initial masses of $\gtrsim 25$~M$_\odot${}). Consequently we would expect larger EW values at the explosion sites \citep[e.g. see the low redshift LGRB environment studies of][]{thoene08,christensen08,kruehler17,izzo17}, and an association to ongoing star formation of their hosts at a level much higher than is seen for SNe~Iax (\cref{sec:pixelstatres}, see also \citealt{kangas17}).
A search for He {\sc i} $\lambda 4922$ in emission (indicative of stellar populations of only a few Myr) in the hosts of our MUSE sample found 5 SNe~Iax with such young regions within 3~kpc. This include 4 of the 5 events with peak luminosities known to be $\geq -16$~mag plus SN~2012Z (the missing low-luminosity event is SN~2010el). The host galaxies of these fainter members are typically more irregular and exhibit signatures of extremely young stellar populations within them, with EW(\text{H}\ensuremath{\alpha}{}) reaching hundreds of \AA{}. In order to establish a causal link between the SNe and their nearest very young regions, progenitor velocities of hundreds of km~s$^{-1}${} must be invoked (\cref{sec:resultsages}). To produce such high mass and velocity runaways requires very rare dynamical ejections or unfeasibly large ejection velocities after a binary companion goes SN \citep[][see also discussion in \citealt{kruehler17}]{eldridge11}.
Our results would therefore disfavour the very massive progenitor scenario (for the SN~Iax sample as a whole and the SN~2008ha subset). \citet{moriya10} concur, finding that their 25 and 40~M$_\odot${} fall-back models cannot reproduce the observations of SN~2008ha, but find better agreement with a 13~M$_\odot${} model.
Deflagrations of hybrid CONe have been proposed to explain the weak SN~2008ha by \citet{kromer15}. Although occurring in Chandrasekhar-mass WDs, the deflagration is not propagated into the outer ONe layer, which results instead in lower energy release and consequently low ejecta masses (cf. CO WD explosions), leaving behind a bound remnant. The ejecta mass in the model is below estimates for SN~2008ha, resulting in too quickly rising and fading light curves, and also affecting the spectral comparison. The delay time to explosion for CONe WDs with helium donor stars has been estimated at $\sim$few~$\times 10^7$ to $10^8$~yrs by \citet{wang14, kromer15}, albeit at solar metallicity which may not be representative of the typical environments of SN~2008ha-like events. This is in agreement with this study and \citet{foley14}.
\citet{moriya16} present an initial study of the light curves and rates for binary evolution channels of stripped-envelope electron-capture (SE-EC) SNe, which may have relation to the low-luminosity CCSN scenario of \citet{valenti09}. They find broad agreement with SN~2008ha-like events in terms of peak magnitude and their rapidly-evolving light curves, and that the rates of SE-ECSNe are significantly enhanced at low-metallicities. For our 5 events peaking at $M_R > -16$~mag we determined a wide spread of metallicities \citep[see also][]{foley09, stritzinger14}, including very low metallicity events and the quite metal-rich SN~2010el.
In the case of the rapidly declining events (note that these are not all overlapping with the faint events, e.g., PS~15csd; \citealt{magee16}), we similarly find low metallicities for all but SN~2010el. SE-ECSNe progenitors are expected to come from the lower end of the mass range of massive SNe progenitors $\sim 7-8$~M$_\odot${} and we therefore expect to see evidence for young, but not exceptionally young SPs at their explosion sites -- this is borne out in the relatively modest EW(\text{H}\ensuremath{\alpha}{}) values we find of 25--40~\AA{}. We therefore confirm here the preference for faint and/or fast subsets of SNe~Iax (SN~2008ha-like events) to be preferentially in lower-metallicity environments with signatures of relatively young SPs at their explosion sites, consistent with current predictions for SE-ECSNe by \citet[][see also \citealt{pumo09}]{moriya16}. In this scenario the proposed donor or remnant of SN~2008ha \citep{foley14} would be the companion star that may have stripped the envelope of the SN progenitor. The spectral predictions for SE-ECSNe that are needed to provide a full comparison to SN~2008ha-like events are currently lacking.
\section{Conclusions}
We have presented a spectroscopic survey of the explosion sites and host galaxies of SNe~Iax and fitted emission line regions in our VLT/MUSE IFS and NOT/ALFOSC long-slit observations.
We determined deprojected metallicity gradients for all hosts where possible and found that estimates of the explosion sites' metallicities based on these gradients agreed well with our direct measurements. With this our explosion site metallicity sample was supplemented with gradient estimates where we could not measure the location directly. The majority of SN~Iax explosion sites were found to be sub-solar ($12 + \log_{10} (\text{O/H}) \simeq 7.75-8.85$ dex in the scale of \citetalias{dopita16}), and generally metal poor compared to the distribution of SF in their hosts (\cref{sec:resz}).
SNe~Iax explosion sites appeared to be less intensely star-forming and somewhat older than the typical SF region of the hosts ({\cref{sec:resultsages}), although they do follow a similar host-normalised offset distribution (\cref{sec:resultsoffsets}). Through comparison with a fiducial gas model and the results of the SP synthesis code {\sc bpass} we estimate the ages of the young SPs at the explosion sites of SNe~Iax to be several $\times10^7$ to $10^8$ years old (although note the caveats on assigning quantitative values in \cref{sec:methages}).
The relatively young ages at the explosion sites are confirmed through a similarity in their association to the ongoing SF of their host galaxies to that seen for SNe~II (\cref{sec:pixelstatres}), which are expected to have typical ages of tens of Myr, extending up to $\simeq 10^8$ years \citep{zapartas17}.
Assessing our environmental measures in terms of SN properties, we find that the brighter, more slowly fading objects define our range of metallicities, EW(\text{H}\ensuremath{\alpha}{}) and offsets by covering the entire ranges (\cref{sec:sncompare}). Faint and faster events appear to occupy more restricted ranges of EW values and host offsets although we are limited to only a few objects for which we can investigate. Their metallicities extend across a wide range including some of the most metal-poor and -rich in the whole sample. We find no correlations between the velocity of the SNe and our environmental parameters. The only two designated SNe~Iax to display helium (SNe~2004cs, 2007J) have among the highest EW(\text{H}\ensuremath{\alpha}{}) values of the sample, indicate the presence of comparatively younger SPs. Their membership as SNe~Iax has been debated \citep{white15, foley16}. With just two events we are limited in searching for statistical difference between these and the rest of the SNe~Iax and their environments appear typical of both SNe~Iax and IIb (the suggested typing of \citealt{white15}).
When compared to other SN types, SNe~Iax as a whole display a similar metallicity distribution to that seen for SNe Ib/c and II, with similar median values and covering a broadly similar range in metallicity. The host-normalised offset distribution of SNe~Iax follows closely that seen for SNe~Ia in disk galaxies and SNe~Ib. Using pixel statistics, SNe~Iax trace the ongoing SF of their hosts at a level most similar to SN~II and significantly more so than SNe~Ia, but less than SNe~Ic (and LGRBs).
For the S0 host of SN~2008ge, NGC~1527, we discover a stream and clumps of ionised gas not seen in previous studies. A limited analysis of emission line ratios would disfavour a SF-driven ionising flux, with an evolved SP or shocks more likely to be powering it. We find no evidence of ionised gas at, or nearby, the location of SN~2008ge and confirm its environment is old compared to the rest of SNe~Iax. The ionised component suggests there may have been SF in NGC~1527 relatively recently and thus $\sim$Gyrs old progenitors are not required.
The young explosion sites we confirm here remain in good agreement with the expectations for progenitors consisting of CO/CONe WD + He stars and moderately massive CCSNe with fall-back or SE-ECSNe. The lack of features associated with very young environments disfavour the presence of very massive stars at the explosion site. We would therefore consider the model of massive WR stars suffering fall-back on to a black hole upon collapse as unlikely progenitors (without having to invoke implausibly large runaway velocities), ruling out faint SNe~Iax as a way to explain local, apparently SN-less LGRBs.
\section*{Acknowledgements}
We thank Ryan Foley for providing the late time spectra of SN~2014dt. David Balam is thanked for providing the spectrum of PSN~J22412689+3917220 (now SN~2015ce) for its registration on the Transient Name Server.
Eric Hsiao and the iPTF collaboration are thanked for their assistance registering PTF~14ans (now SN~2014ey) onto the Transient Name Server.
Lise Christensen is thanked for comments on a draft of the manuscript.
JDL gratefully acknowledges support from the UK Science and Technology Facilities Council (grant IDs ST/L000733/1 and ST/P000495/1).
MDS gratefully acknowledges support provided by the Danish Agency for Science and Technology and Innovation realized through a Sapere Aude Level 2 grant, the VILLUM FONDEN (research grant 3261), and the Instrument Centre for Danish Astrophysics (IDA).
Based on observations collected at the European Organisation for Astronomical Research in the Southern Hemisphere under ESO programmes 095.D0091, 096.D-0263 and 099.D-0022.
Based on observations made with the Nordic Optical Telescope (proposal numbers 52-004 and P3-005; PI: Stritzinger), operated by the Nordic Optical Telescope Scientific Association at the Observatorio del Roque de los Muchachos, La Palma, Spain, of the Instituto de Astrofisica de Canarias.
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. We acknowledge the usage of the HyperLeda database (http://leda.univ-lyon1.fr)
This research made use of Astropy, a community-developed core Python package for Astronomy \citep{astropy13}.
\bibliographystyle{mnras}
|
\section{Introduction}
The traditional way of describing an inverse problem is to define a forward map relating the unknown to an observed quantity, and to look for an estimate of the unknown when the data is corrupted by noise. In this description, it is often tacitly assumed that an underlying "truth" exists, and the noiseless data arises from applying the forward map on this true value. On the other hand, it is commonly acknowledged that a mathematical model does not coincide with the reality, and therefore part of the noise must be attributed to the model discrepancy, or the mismatch between the model and the reality. Modeling this discrepancy is an active research topic in statistics -- see
\cite{kennedy2001bayesian,brynjarsdottir,briol2015probabilistic} and the references
therein; it is also a closely related to the concept of the ``inverse crime'', a procedure of testing a computational method with data that has been generated by the same model that is used to solve the inverse problem \cite{Colton,KSart}.
Common sources of modeling errors in inverse problems include:
\begin{enumerate}
\item model reduction -- a complex, computationally intensive model is replaced by a simpler, less demanding model;
\item parametric reduction -- in a model depending on poorly known parameters, some of them are frozen to fixed values, assuming that the solution is not sensitive to them;
\item unknown geometry -- a computational domain of unknown shape is approximated by a standard geometry.
\end{enumerate}
Including the modeling error into the computations in the traditional deterministic
setting may not be straightforward. Recasting the inverse problem via Bayesian
inference provides tools to carry this out in a natural statistical fashion. The present article introduces and analyzes
a Bayesian methodology for model error estimation which demonstrably leads to improved estimates of the true unknown function
generating the data.
\subsection{Background}
\label{ssec:background}
In this article, we consider the problem of estimating an unknown quantity $u$ based on indirect observations. In the Bayesian framework, the prior belief about the quantity $u$ is encoded in the prior probability distribution $\mathbb{P}(u)$, and given the observed data $b$, the posterior
distribution $\mathbb{P}(u\mid b)$ follows from Bayes' formula,
\[
\mathbb{P}(u\mid b) \propto \mathbb{P}(b\mid u)\mathbb{P}(u),
\]
where $\propto$ denotes proportionality up to a scaling constant depending on
$b$ but not on $u$, and the distribution $\mathbb{P}(b\mid u)$ of $b$ is the likelihood. To construct the likelihood, a forward model from $u$ to $b$ needs to be specified. A commonly used model, assuming additive observation noise that is independent of the unknown $u$, is
\[
b = F(u) + \varepsilon
\]
where $\varepsilon \sim \pi_{\mathrm{noise}}(\cdot)$ is a realization of random noise,
and $F$ is a mapping defined typically on the parameter space for $u$,
which is often infinite-dimensional ($u$ is a function) or high dimensional
(a function has been discretized to obtain $u$). Under these assumptions, the likelihood is equal to the distribution of $\varepsilon$ with mean shifted by $F(u)$, i.e.,
\[
\mathbb{P}(b\mid u) = \pi_{\mathrm{noise}}(b-F(u)).
\]
To estimate $u$, or derived quantities based on it, numerical approximations of integrals with respect to the posterior distribution are required, and a common approach is to use sampling methods such as Markov chain Monte Carlo (MCMC). This requires a large number of repeated evaluations of the forward map $F$, which is often expensive to evaluate numerically. A particular instance that we have in mind is the situation where
evaluation of $F$ requires numerical solution of a partial differential equation. If computational resources or time are an issue, an attractive approach is to trade off the accuracy of evaluations with the computational cost by adjusting the resolution of the mesh that the PDE is solved upon. Denoting by $f$ an approximation to $F$ on a coarse mesh, a model for the data can be written as
\[
b = f(u) + m + \varepsilon
\]
where $m = F(u) - f(u)$ denotes the model error induced by moving from the accurate model $F$ to the approximate one.
If we ignore the fact that $m$ depends on $u$, and instead model it as additive independent noise, the conditional likelihood $\mathbb{P}(b\mid u,m)$ is then given by
\[
\mathbb{P}(b\mid u,m) = \pi_{\mathrm{noise}}(b-f(u)-m);
\]
evaluations of this map then only require evaluation of the approximate map $f$. Furthermore, the likelihood $\mathbb{P}(b\mid u)$ can be found by marginalizing out the model error $m$. However, the marginalization requires the distribution of $m$ which is not known. As suggested in \cite{KSbook}, the Bayesian approach provides a natural approximate solution to this problem: By using the prior distribution of $u$ and the model error mapping $M(u) = F(u)-f(u)$, one can generate a sample of model errors to estimate the model error distribution. This approach, referred to as the \emph{enhanced error model}, has been shown to produce more accurate point estimates than those that come from neglecting the model error (the \emph{conventional error model}), see, e.g., \cite{arridge2006approximation, Heino, KSart, banasiak2012improving} for static inverse problems, and \cite{huttunen2007approximation, huttunen2, huttunen2010importance} for extensions to dynamic inverse problems.
In \cite{Oliver}, the enhanced error model was developed further using the observation that the the posterior distribution based on the error model contains refined information about the unknown $u$ beyond the point estimate determined by the enhanced error model; as a consequence the model error distribution can be updated by pushing forward the distribution under the model error mapping $M$. When the data are particularly informative, posterior samples may differ significantly from prior samples, and this should produce a much better approximation to the distribution of the model error, potentially yielding a better approximation of the posterior distribution of $u$, and at the
very least providing point estimates of higher accuracy. The procedure can be iterated to produce a sequence of approximate posterior distributions that have the potential to yield a sequence of point estimates of improved accuracy;
they may also approximate the true posterior distribution with increasing accuracy. In this article, we address this approach in a systematic way, with particular focus on convergence of the iterative updating.
The effect of model error and model discrepancy in Bayesian inference is a widely studied topic. Early works focus primarily on code uncertainty -- the uncertainty that arises due to expense of forward model evaluations meaning that it is only practical
to compute outputs of the model for a limited finite number of inputs. A review of work in this direction is given in \cite{sacks1989design}, including the problems of optimal choice of inputs at which to evaluate the forward model, and how to predict the output for inputs away from the computed ones. In \cite{kennedy2001bayesian} the numerous sources of uncertainty within computational inverse problems are discussed, including those arising from model inadequacy and code uncertainty. The authors model this error as a function independent of the approximate model, which can be justified in certain cases. Hierarchical Gaussian priors are placed upon the model and the model error, and the model and error are then linked by imposing correlations between the hyperparameters. The technique has subsequently been developed further in \cite{brynjarsdottir}, and used in, for example, the context of model validation \cite{bayarri2007framework} and uncertainty quantification \cite{higdon2004combining}.
More recent work in probabilistic numerical methods \cite{briol2015probabilistic}
provides a unifying perspective on this
body of work, linking it to earlier research connecting numerical algorithms to
Bayesian statistics \cite{diaconis1988bayesian}.
\subsection{Our Contribution}
\label{ssec:contribution}
\begin{itemize}
\item We develop further the iterative updating of the posterior probability densities based on repeated updating of the model error distribution, leading to an approximation of the posterior probability density. While the approximation error is defined
through the computationally expensive accurate model, the posterior approximation
we introduce relies primarily on the computationally inexpensive approximate model,
and a limited number of evaluations of the accurate model.
\item In the case where the models are linear and the noise and prior distributions are Gaussian, we show that the means and covariances of the resulting sequence of posterior distributions converge to a limit geometrically fast.
\item For more general models and prior/noise distributions we introduce particle approximations to allow the algorithm to be implemented numerically, and show convergence of these approximations in the large particle limit.
\item We illustrate numerically the effectiveness of the algorithms in multiple different settings, showing the advantage over the conventional and enhanced error models.
\end{itemize}
\subsection{Outline}
\label{ssec:outline}
The iterative approach of updating the posterior distributions is introduced in Section \ref{sec:formulation}.
In Section \ref{sec:gaussian} we focus on the particular case where the forward model is linear, and the noise and prior distributions are Gaussian. The assumptions imply that the approximate posterior distributions are also Gaussian, and can therefore be characterized by their means and covariances. We identify conditions guaranteeing the convergence of the sequence of approximate posteriors to a non-degenerate limit as the number of iterations tends to infinity. In Section \ref{sec:sampling} we discuss different sampling methods which may be used to implement the algorithm in practice. In particular we focus on particle methods that require a finite number of full model evaluations in order to estimate the modeling error and posterior distribution, and show convergence
to the correct sequence of approximate distributions in the large particle limit. Finally, in Section \ref{sec:numerics}, we provide numerical illustrations of the behavior and performance of the algorithm for three different forward models.
Section \ref{sec:conc} contains the conclusions and discussion.
\section{Problem Formulation}
\label{sec:formulation}
We start by introducing the main ingredients of the iterative algorithm: the accurate and approximate models are defined in Subsection \ref{ssec:models}, along with some examples whose details will be discussed later on. The enhanced error model \cite{KSbook} is reviewed in Subsection \ref{ssec:enhanced}, prompting the question of how to update the density of the modeling error. In subsection \ref{ssec:algorithm} we provide an iterative algorithm for doing this, in the case where all measures involved have Lebesgue densities; a more general algorithm is provided in the appendix for cases such as those
arising when the measures are defined on infinite-dimensional function spaces.
\subsection{Accurate vs. Approximate Model}
\label{ssec:models}
Let $X, Y$ be two Banach spaces representing the parameter and data spaces, respectively. Let
\[
F:X\to Y, \quad u\mapsto b
\]
denote a reference forward model, referred to as the {\em accurate model}, and let the {\em approximate model} be denoted by
\[
f:X\to Y, \quad u\mapsto b.
\]
We write the observation model using the accurate model,
\begin{equation}\label{obs model}
b = F(u) + \varepsilon,
\end{equation}
and equivalently, using the approximate model, as
\begin{eqnarray}\label{appr model}
b &=& f(u) + \big(F(u) - f(u)\big) + \varepsilon \nonumber \\
&=& f(u) + m + \varepsilon,
\end{eqnarray}
where $m$ represents the modeling error,
\[
m = F(u) - f(u) = M(u).
\]
In light of the above observation, we may view the data as coming from the approximate model,
with an error term which reflects both observational noise
$\varepsilon$ and modeling error $m$. The main problem
addressed in this paper is how
to model the probability distribution of the model error. We study this
question with the goal of providing computations to estimate the unknown $u$
from $b$ using only the less expensive model $f$, and not the true model $F$,
without creating modeling artifacts.
We conclude this subsection by giving two examples of approximate models, both
of which may be relevant in applications.
\begin{example}[Linearization] In electrical impedance tomography (EIT), the goal is to estimate the conductivity distribution inside a body from a finite set of current/voltage measurements at the boundary, as discussed in more detail in Subsection \ref{ssec:eit}.
We denote by $F$ the differentiable non-linear forward model, mapping the appropriately parametrized conductivity distribution to the voltage measurements. We define the approximate model through the linearization,
\[
f(u) = F(u_0) + {\mathsf D}F(u_0)(u - u_0),
\]
where ${\mathsf D} F(u_0)$ is the Jacobian of the forward map, and $u_0$ is a fixed parameter value, representing, e.g., a constant conductivity background.
\end{example}
\begin{example}[Coarse Mesh] In the EIT model, the accurate model represents the forward model computed with a FEM grid fine enough to guarantee that the numerical solution approximates the solution of the underlying PDE within a required precision. To speed up computations, we introduce the reduced model $f$ as the forward model based on FEM built on a coarse grid. We assume that both computational grids are built on an underlying independent discretization of the conductivity $\sigma$, appropriately parametrized, and the FEM stiffness matrices that require integration over elements are computed by evaluating $\sigma$ in the Gauss points of the elements, respectively.
\end{example}
\subsection{The Enhanced Error Model}
\label{ssec:enhanced}
We start by reviewing the basic ideas of the enhanced error model and, for simplicity, assume here that $X$ and $Y$ are Euclidean
spaces and that all probability distributions are expressible in terms of Lebesgue densities. We assume that $u$ is an $X$-valued random variable with a given {\em a priori} density,
\[
u \sim \pi_{\rm prior}(u).
\]
Furthermore, we assume that the additive noise $\varepsilon$ is a
$Y$-valued random variable, independent of $u$, with the density
\[
\varepsilon\sim \pi_{\rm noise}(\varepsilon).
\]
In view of the model (\ref{obs model}), we may write the likelihood as
\[
\pi(b\mid u) = \pi_{\rm noise}(b - F(u)),
\]
and the posterior density is, according to Bayes' formula, given by
\[
\pi(u\mid b) \propto \pi_{\rm prior}(u)\pi_{\rm noise}(b - F(u)).
\]
If, instead, we want to use the approximate model $f(u)$ for the data, the modeling error $m$ needs to be taken into account. The idea of the {\em enhanced error model} in \cite{KSbook} is the following: Given the prior distribution $\mu_0({d} u) = \pi_{\rm prior}(u){d} u$, with no other information about $u$, the probability distribution of $m=M(u)$ is obtained as a push-forward of the prior distribution:
\[
m\sim M^\#\mu_0.
\]
To obtain a computationally efficient formalism, the distribution of $m$ is approximated by a Gaussian distribution sharing the mean and
covariance of the push-forward measure,
\[
m\sim{\mathcal N}(\overline m,{\mathsf \Sigma}),
\]
where in practice the mean and covariance may be estimated numerically by sampling the modeling error.
The Gaussian approximation is particularly convenient if the additive noise is Gaussian,
\[
\varepsilon\sim{\mathcal N}(0,{\mathsf \Gamma}),
\]
as it leads to the approximate likelihood model
\[
\pi(b\mid u) \propto {\rm exp}\left(-\frac 12 (b - f(u) - \overline m)^{\mathsf T}({\mathsf \Gamma} + {\mathsf \Sigma})^{-1}(b - f(u) - \overline m)\right),
\]
and, consequently, to the posterior model
\[
\pi(u\mid b) \propto \pi_{\rm prior}(u)\, {\rm exp}\left(-\frac 12 (b - f(u) - \overline m)^{\mathsf T}(
{\mathsf \Gamma} + {\mathsf \Sigma})^{-1}(b - f(u) - \overline m)\right).
\]
Note that the enhanced error model can be interpreted as a form of variance inflation: if the model error is neglected as in the conventional error model, then the covariance matrix in the likelihood would be smaller in the sense of positive-definite quadratic forms, since ${\mathsf \Sigma}$ is non-negative definite. This is to be expected as the lack of accuracy in the model contributes additional uncertainty to the problem; we do not wish to be over-confident in an inaccurate model. In the next section we explain how this idea may be built upon to
produce the algorithms studied in this paper.
\subsection{The Iterative Algorithm}
\label{ssec:algorithm}
In this subsection we generalize the preceding enhanced error model
in two ways: (i) we iterate the construction of the model error, updating its probability distribution by pushing
forward by $M$ the measure $\mu_{\ell}$ , the posterior distribution of $u$ when the model
error distribution is computed as the pushforward under
$M$ of the measure $\mu_{\ell-1};$ in this iterative method
we choose $\mu_0$ to be the prior and so the first step is analogous to what is described
in the previous subsection;
(ii) we do not invoke a Gaussian approximation of
the model error, leaving open other possibilities for practical
implementation.
We describe the resulting algorithm here in the case where Lebesgue
densities exist, and refer to the appendix for its formulation in a more abstract setting.
{\bf Algorithm (Lebesgue densities).} Let $\mu_\ell$ denote the posterior distribution at stage $\ell$, with density $\pi_\ell$, so that $\mu_\ell({d} u) = \pi_\ell(u)\,{d} u$. Denote $\pi_{\ell}(b\mid u)$ the likelihood at stage $\ell$.
\begin{enumerate}[\hspace{0.2cm}1.]
\item Set $\pi_0(u) = \pi_{\rm prior}(u)$ and $\ell = 0$.
\item Given $\mu_\ell$, assume $m \sim M^\#\mu_\ell$. Assuming that $u$ and $m$ are mutually independent, we have
\[
\pi(b\mid u,m) = \pi_{\rm noise}(b - f(u) - m),
\]
and by marginalization,
\begin{eqnarray*}
\pi_{\ell+1}(b\mid u) &= \int_Y \pi_{\rm noise}(b - f(u) - m)(M^\#\mu_\ell)({d} m)\\
&= \int_X \pi_{\rm noise}(b - f(u) - M(z))\pi_\ell(z){d} z.
\end{eqnarray*}
Hence using Bayes' theorem, update the posterior distribution:
\begin{eqnarray}
\label{eq:update_leb}
\pi_{\ell+1}(u) \propto \pi_{\rm prior}(u)\int_X \pi_{\rm noise}(b - f(u) - M(z))\pi_\ell(z){d} z.
\end{eqnarray}
\item Set $\ell \mapsto \ell+1$ and go to 2.\qed
\end{enumerate}
We can give an explicit expression for the above density $\pi_{\ell}(u)$:
\[
\fl \pi_\ell(u) \propto \pi_{\rm prior}(u)\int_X\cdots\int_X \left(\prod_{i=1}^\ell\pi_{\rm noise}(b - f(z_{i+1}) - M(z_i))\pi_{\rm prior}(z_i)\right)\,{d} z_1\ldots{d} z_\ell
\]
where we define $z_{\ell+1} = u$.
The above algorithm can be generalized to the case when no Lebesgue densities exist, such as will be the case on infinite dimensional Banach spaces, see the Appendix.
\section{The Linear Gaussian Case}
\label{sec:gaussian}
We analyze the convergence of the general algorithm in the case where both the accurate model $F$ and the approximate model $f$ are linear, and the noise and prior distributions are Gaussian. With these assumptions the posterior distributions forming the sequence remain Gaussian, and are hence characterized by the sequences of means and covariances. Convergence properties of the iteration can be established by using the updating formulas for these sequences. Though the iteration is not immediately implementable, since the full matrix for the accurate model $F$ is required to calculate each covariance matrix, the explicit convergence results give insight into the implementable variants of the algorithm introduced in Section \ref{sec:sampling}, as well as for non-linear forward maps.
In subsection \ref{ssec:linear_evolution}, we first describe how the posterior density evolves, establishing the iterations that the means and covariances satisfy. In subsection \ref{ssec:conv_infinite} we show that if the model error is sufficiently small, the sequences of posterior means and covariances converge geometrically fast. Moreover, despite the repeated incorporation of the data into the posterior, the limiting distribution does not become singular. Additionally, in subsection \ref{ssec:conv_finite} we show that in finite dimensions the assumption of small model error is not needed in order to establish convergence of the covariance.
\subsection{Evolution of the Posterior Distribution}
\label{ssec:linear_evolution}
When the noise distribution $\pi_{\rm noise}$ and prior distribution $\pi_{\rm prior}$ are Gaussian, the measure $\mu_\ell$ is Gaussian at each stage $\ell$, and we can write down expressions for the evolution of its mean and covariance.
Let $X,Y$ be separable Hilbert spaces and ${\mathsf A}_\star:X\rightarrow Y$ a linear operator. Assume that the data $b \in Y$ arise from ${\mathsf A}_\star$ via
\[
b = {\mathsf A}_\star u + \varepsilon,\;\;\;\varepsilon \sim \mathcal{N}(0,{\mathsf \Gamma})
\]
where ${\mathsf \Gamma}:Y\rightarrow Y$ is a symmetric positive definite covariance operator. Let ${\mathsf A}:X\rightarrow Y$ denote an approximation to ${\mathsf A}_\star$ so that the expression for $b$ may be rewritten
\begin{eqnarray*}
b &= {\mathsf A} u + ({\mathsf A}_\star - {\mathsf A})u + \varepsilon.
\end{eqnarray*}
We define the model error operator ${\mathsf M}:X\rightarrow Y$ by ${\mathsf M} = {\mathsf A}_\star - {\mathsf A}$, so that ${\mathsf M} u$ represents the (unknown) model error. We assume that the model error is Gaussian, with unknown mean $w$ and covariance ${\mathsf \Sigma}$. Additionally we assume that it is independent of the observation noise. The data is now given by
\begin{eqnarray*}
b &= {\mathsf A} u + \hat{\varepsilon},\;\;\;\hat{\varepsilon} \sim \mathcal{N}(w,{\mathsf \Sigma} + {\mathsf \Gamma}).
\end{eqnarray*}
Let $\mu_0 = N(m_0,{\mathsf C}_0)$ denote the prior distribution on $u$. We first estimate $w$ and ${\mathsf \Sigma}$ by pushing forward $\mu_0$ by the model error operator ${\mathsf M}$:
\[
{\mathsf M}^{\#}\mu_0 = \mathcal{N}({\mathsf M} m_0, {\mathsf M} {\mathsf C}_0 {\mathsf M}^*) \equiv \mathcal{N}(\hat{w}_1,\hat{{\mathsf \Sigma}}_1).
\]
Then, assuming for now that the model error has this distribution, the resulting posterior distribution on $u$ can be calculated as $\mu_1 = N(m_1,{\mathsf C}_1)$, where
\begin{eqnarray*}
{\mathsf C}_1 &= \big({\mathsf A}^*({\mathsf \Gamma} + {\mathsf M} {\mathsf C}_0 {\mathsf M}^*)^{-1}{\mathsf A} + {\mathsf C}_0^{-1}\big)^{-1},\\
m_1 &= {\mathsf C}_1\big({\mathsf A}^*({\mathsf \Gamma} + {\mathsf M}{\mathsf C}_0 {\mathsf M}^*)^{-1}(b - {\mathsf M} m_0) + {\mathsf C}_0^{-1}m_0\big).
\end{eqnarray*}
As described previously, in order to obtain a better approximation of the model error, the above step can be repeated so that the measure pushed forward to approximate the model error is closer to the posterior distribution.
Therefore, in the next step, we approximate the distribution of the model error by
\[
{\mathsf M}^{\#}\mu_1 = \mathcal{N}({\mathsf M} m_1,{\mathsf M}{\mathsf C}_1{\mathsf M}^*) \equiv \mathcal{N}(\hat{w}_2,\hat{{\mathsf \Sigma}}_2),
\]
and make a further posterior approximation. Iterating this process, we see that the posterior mean and covariance evolve via
\begin{eqnarray}
\label{eq:ck_standard}
{\mathsf C}_{\ell+1} &= \big({\mathsf A}^*({\mathsf \Gamma} + {\mathsf M} {\mathsf C}_\ell {\mathsf M}^*)^{-1}{\mathsf A} + {\mathsf C}_0^{-1}\big)^{-1},\\
m_{\ell+1} &= {\mathsf C}_{\ell+1}\big({\mathsf A}^*({\mathsf \Gamma} + {\mathsf M}{\mathsf C}_\ell {\mathsf M}^*)^{-1}(b - {\mathsf M} m_\ell) + {\mathsf C}_0^{-1}m_0\big)\nonumber.
\end{eqnarray}
We wish to show that these sequences are convergent.
We may write the above iteration in an equivalent form as
\begin{eqnarray*}
{\mathsf C}_{\ell+1} &= {\mathsf C}_0 - {\mathsf C}_0{\mathsf A}^*({\mathsf \Gamma} + {\mathsf M}{\mathsf C}_\ell {\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}{\mathsf A}{\mathsf C}_0,\\
m_{\ell+1} &= m_0 + {\mathsf C}_0{\mathsf A}^*({\mathsf \Gamma} + {\mathsf M}{\mathsf C}_\ell {\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}(b - {\mathsf A} m_0 - {\mathsf M} m_\ell),
\end{eqnarray*}
using results from \cite{lps,mandelbaum}, assuming that ${\mathsf C}_0$ is trace-class and ${\mathsf \Gamma}$ is trace-class or white. This form has the advantage that the unbounded operator ${\mathsf C}_0^{-1}$ does not appear, and so we need not worry about its domain of definition. In what follows we simply assume that this equivalent expression
for the evolution of the means and covariances is valid.
In the following subsections, we consider two different cases. In the first one, we limit the data into a finite dimensional space $Y$ but let the space $X$ to be a Hilbert space. The convergence of the algorithm is demonstrated under certain restrictive conditions: the modeling error needs to be small enough. In the second case, we also limit the unknown $u$ to a finite dimensional space, and show that in this case the convergence proof can be obtained without the restrictions needed in the former case.
We emphasize that, although we establish convergence of the iteration in
various settings, the limiting distribution does not coincide with the
true posterior distribution found in the absence of model error. Nonetheless
our numerical experiments will show that the limiting distribution can
significantly improve point estimates of the underlying value used to generate the data.
\subsection{Convergence in Infinite Dimensions}
\label{ssec:conv_infinite}
We introduce a scalar parameter $\delta$ controlling the accuracy of the approximation ${\mathsf A}$ of ${\mathsf A}_\star$, writing $\delta {\mathsf M}$ in place of ${\mathsf M}$. By writing explicitly the dependence of the mean and covariance on $\delta$, we have
\begin{eqnarray*}
{\mathsf C}_{\ell+1}(\delta) &= {\mathsf C}_0 - {\mathsf C}_0 {\mathsf A}^*\big({\mathsf \Gamma} + \delta^2 {\mathsf M}{\mathsf C}_\ell(\delta) {\mathsf M}^* + {\mathsf A}{\mathsf C}_0 {\mathsf A}^*\big)^{-1}{\mathsf A}{\mathsf C}_0,\\
m_{\ell+1}(\delta) &= m_0 + {\mathsf C}_0{\mathsf A}^*\big({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf C}_\ell(\delta){\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*\big)^{-1}\big(b - {\mathsf A} m_0 - \delta {\mathsf M} m_\ell(\delta)\big).
\end{eqnarray*}
Let $\mathcal{L}(X)$ denote the space of bounded linear operators on $X$, equipped with the operator norm. Let $S_+(X)\subseteq\mathcal{L}(X)$ denote the set of positive bounded linear operators on $X$, and $\overline{S_+(X)}$ the set of non-negative bounded linear operators on $X$. We may write the iteration for ${\mathsf C}_\ell(\delta)$ as
\begin{eqnarray}
\label{eq:ck}
{\mathsf C}_{\ell+1}(\delta) = {\mathscr F}\big({\mathsf C}_\ell(\delta),\delta\big)
\end{eqnarray}
where ${\mathscr F}:\overline{S_+(X)}\times\mathbb{R}\rightarrow \overline{S_+(X)}$ is given by
\[
{\mathscr F}({\mathsf B},\delta) = {\mathsf C}_0 - {\mathsf C}_0 {\mathsf A}^*\big({\mathsf \Gamma} + \delta^2 {\mathsf M}{\mathsf B}{\mathsf M}^* + {\mathsf A}{\mathsf C}_0 {\mathsf A}^*\big)^{-1}{\mathsf A}{\mathsf C}_0.
\]
We show that under certain assumptions, for all $\delta$ sufficiently small, ${\mathscr F}(\,\cdot\,,\delta)$ has a unique stable fixed point ${\mathsf C}(\delta)$. The assumptions we make are as follows.
\begin{assumptions}
\label{assump}
\begin{enumerate}[(i)]
\item ${\mathsf C}_0 \in S_+(X)$ and is trace-class.
\item $Y$ is finite dimensional, and ${\mathsf \Gamma} \in S_+(Y)$.
\item ${\mathsf A}_\star,\,{\mathsf A}:X\rightarrow Y$ are bounded.
\end{enumerate}
\end{assumptions}
We first establish the following result concerning convergence of the sequence of covariance operators:
\begin{proposition}
\label{prop:ckconv}
Let Assumptions \ref{assump} hold. Then there is a $\beta > 0$ such that for all $\delta < 1/\beta$, a unique ${\mathsf C}(\delta) \in S_+(X)$ exists with
\[
{\mathsf C}(\delta) = {\mathscr F}\big({\mathsf C}(\delta),\delta\big).
\]
Moreover, ${\mathsf C}(\delta)$ is a stable fixed point of ${\mathscr F}(\,\cdot\,,\delta)$, and there is a constant $\alpha_1$ such that
\[
\|{\mathsf C}_\ell(\delta) - {\mathsf C}(\delta)\|_{\mathcal{L}(X)} \leq \alpha_1(\beta\delta)^{2\ell}\;\;\;\mbox{for all }\ell \geq 1.
\]
In particular, for $\delta < 1/\beta$, the sequence $\{{\mathsf C}_\ell(\delta)\}_{\ell\geq 1}$ converges geometrically fast.
\end{proposition}
From this geometric convergence of $\{{\mathsf C}_\ell(\delta)\}_{\ell\geq 1}$ we can deduce that also the means $m_\ell(\delta)$ converge: Define the maps $G_\ell:X\times\mathbb{R}\rightarrow X$, $\ell \geq 0$, by
\[
G_\ell(m,\delta) = m_0 + {\mathsf C}_0{\mathsf A}^*\big({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf C}_\ell(\delta){\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*\big)^{-1}\big(b - {\mathsf A} m_0 - \delta {\mathsf M} m\big)
\]
so that the update for $m_\ell(\delta)$ is given by
\[
m_{\ell+1}(\delta) = G_\ell\big(m_\ell(\delta),\delta\big).
\]
Define also the limiting map $G:X\times\mathbb{R}\rightarrow X$ by
\[
G(m,\delta) = m_0 + {\mathsf C}_0{\mathsf A}^*\big({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf C}(\delta){\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*\big)^{-1}\big(b - {\mathsf A} m_0 - \delta {\mathsf M} m\big).
\]
Then we have the following result:
\begin{proposition}
\label{prop:mkconv}
Let Assumptions \ref{assump} hold, and let $\beta$ be as in Proposition \ref{prop:ckconv}. Then for all $\delta < 1/\beta$, there exists a unique $m(\delta) \in X$ with
\[
m(\delta) = G\big(m(\delta),\delta\big).
\]
Moreover, there is an $\alpha_2 > 0$ such that
\[
\|m_\ell(\delta) - m(\delta)\|_X \leq \alpha_2(\beta\delta)^\ell\;\;\;\mbox{ for all }\ell \geq 1.
\]
Hence, for $\delta < 1/\beta$, the sequence $\{m_\ell(\delta)\}_{\ell\geq 1}$ converges geometrically fast.
\end{proposition}
To prove the above propositions we first prove the following lemma. In the proof, the following notation is used: Given symmetric non-negative linear operators ${\mathsf B}_1,{\mathsf B}_2 \in \mathcal{L}(X)$, we write ${\mathsf B}_1 \leq {\mathsf B}_2$ to mean that ${\mathsf B}_2 - {\mathsf B}_1$ is non-negative.
\begin{lemma}
\label{lem:ckbound}
Let Assumptions \ref{assump} hold. Then the family of operators ${\mathsf K}({\mathsf B},\delta):Y\rightarrow X$ given by
\[
{\mathsf K}({\mathsf B},\delta) = {\mathsf C}_0{\mathsf A}^*({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf B}{\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}
\]
is bounded uniformly over ${\mathsf B} \in \overline{S_+(X)}$ and $\delta \in \mathbb{R}$.
\end{lemma}
\begin{proof}
We have that
\[
{\mathsf \Gamma} + {\mathsf A}{\mathsf C}_0{\mathsf A}^* \leq {\mathsf \Gamma} + {\mathsf Q} + {\mathsf A}{\mathsf C}_0 {\mathsf A}^*,
\]
for any ${\mathsf Q} \in \overline{S_+(Y)}$, which implies that
\[
({\mathsf \Gamma} + {\mathsf Q} + {\mathsf A}{\mathsf C}_0 {\mathsf A}^*)^{-1} \leq ({\mathsf \Gamma} + {\mathsf A}{\mathsf C}_0A^*)^{-1},
\]
and consequently
\[
\|({\mathsf \Gamma} + {\mathsf Q} + {\mathsf A}{\mathsf C}_0 {\mathsf A}^*)^{-1}\|_{\mathcal{L}(Y)} \leq \|({\mathsf \Gamma} + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}\|_{\mathcal{L}(Y)}.
\]
By choosing ${\mathsf Q} = \delta^2{\mathsf M}{\mathsf B}{\mathsf M}^*$ the claim follows.
\end{proof}
In what follows we will denote $K_{\max} = \sup\big\{\|{\mathsf K}({\mathsf B},\delta)\|_{\mathcal{L}(Y,X)}\;|\;{\mathsf B} \in \overline{S_+(X)}, \delta \in \mathbb{R}\big\}$. Furthermore, we define the parameter $\beta = K_{\max}\| {\mathsf M}\|_{{\mathcal L}(Y)}$.
\begin{proof}[Proof of Proposition \ref{prop:ckconv}]
We first show that for $\delta$ sufficiently small, the map ${\mathscr F}(\,\cdot\,,\delta)$ is a contraction on $\overline{S_+(X)}$. To do this, we look at the Fr\'{e}chet derivative of the map, which may be calculated explicitly. For $B \in \overline{S_+(X)}$ and ${\mathsf V} \in \mathcal{L}(X)$, we have
\begin{eqnarray*}
\fl {\mathsf D}_{\mathsf B}{\mathscr F}&({\mathsf B},\delta){\mathsf V}\\
\fl &= \delta^2{\mathsf C}_0{\mathsf A}^*({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf B}{\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}{\mathsf M}{\mathsf V}{\mathsf M}^*({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf B}{\mathsf M}^*+{\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}{\mathsf A}{\mathsf C}_0\\
\fl &= \delta^2{\mathsf K}({\mathsf B},\delta){\mathsf M}{\mathsf V}{\mathsf M}^*{\mathsf K}({\mathsf B},\delta)^*
\end{eqnarray*}
where ${\mathsf K}({\mathsf B},\delta)$ is as defined in Lemma \ref{lem:ckbound}. The norm of the derivative can be estimated as
\begin{eqnarray*}
\big\|{\mathsf D}_{\mathsf B}{\mathscr F}\big({\mathsf B},\delta\big)\big\|_{\mathcal{L}(X)\rightarrow\mathcal{L}(X)} &= \sup_{\|{\mathsf V}\|_{\mathcal{L}(X)} = 1} \big\|{\mathsf D}_{\mathsf B}{\mathscr F}\big({\mathsf B},\delta\big){\mathsf V}\big\|_{\mathcal{L}(X)}\\
&\leq \sup_{\|{\mathsf V}\|_{\mathcal{L}(X)} = 1} \delta^2\|{\mathsf K}({\mathsf B},\delta)\|_{{\mathcal L}(Y,X)}^2\|{\mathsf M}\|_{\mathcal{L}(Y)}^2 \|{\mathsf V}\|_{\mathcal{L}(X)}\\
&\leq (\beta\delta)^2
\end{eqnarray*}
by the estimate of Lemma \ref{lem:ckbound}. Since the above bound is uniform in ${\mathsf B}$, we may use the mean value theorem to deduce that for all ${\mathsf B}_1,{\mathsf B}_2 \in \overline{S_+(X)}$,
\[
\|{\mathscr F}({\mathsf B}_1,\delta) - {\mathscr F}({\mathsf B}_2,\delta)\|_{\mathcal{L}(X)} \leq (\beta\delta)^2\|{\mathsf B}_1-{\mathsf B}_2\|_{\mathcal{L}(X)}
\]
and so ${\mathscr F}(\,\cdot\,,\delta)$ is a contraction for $\delta < 1/\beta$. The set $\overline{S_+(X)}$ is a complete subset of the space $\mathcal{L}(X)$, and so by the Banach fixed-point theorem there exists a unique ${\mathsf C}(\delta) \in \overline{S_+(X)}$ such that
\[
{\mathsf C}(\delta) = {\mathscr F}({\mathsf C}(\delta),\delta),
\]
and we have ${\mathsf C}_\ell(\delta)\rightarrow {\mathsf C}(\delta)$. Moreover,
\begin{eqnarray*}
\|{\mathsf C}_\ell(\delta) - {\mathsf C}(\delta)\|_{\mathcal{L}(X)} &=& \|{\mathscr F}({\mathsf C}_{\ell-1}(\delta),\delta) - {\mathscr F}({\mathsf C}(\delta),\delta)\|_{\mathcal{L}(X)} \\
&\leq& (\beta\delta)^2\|{\mathsf C}_{\ell-1}(\delta) - {\mathsf C}(\delta)\|_{\mathcal{L}(X)},
\end{eqnarray*}
and recursively,
\[
\|{\mathsf C}_\ell(\delta) - {\mathsf C}(\delta)\|_{\mathcal{L}(X)} \leq (\beta\delta)^{2\ell} \|{\mathsf C}_1(\delta) - {\mathsf C}_0\|_{{\mathcal L}(X)} = \alpha_1(\beta\delta)^{2\ell}.
\]
We finally show that we actually have ${\mathsf C}(\delta) \in S_+(X)$ and so the covariance does not become degenerate in the limit. We denote ${\mathsf C}_{\mathrm{post}} = {\mathscr F}({\mathsf C}_{\mathrm{post}},0)$ the exact posterior covariance in the absence of model error, noting that ${\mathsf C}_{\mathrm{post}} \in S_+(X)$ as we assume ${\mathsf C}_0 \in S_+(X)$ and ${\mathsf \Gamma} \in S_+(Y)$. From a similar argument as in the proof of Lemma \ref{lem:ckbound}, we have
\begin{eqnarray*}
0 < {\mathsf C}_{\mathrm{post}} &= {\mathsf C}_0 - {\mathsf C}_0{\mathsf A}^*({\mathsf \Gamma} + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}{\mathsf A}{\mathsf C}_0\\
&\leq {\mathsf C}_0 - {\mathsf C}_0{\mathsf A}^*({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf C}(\delta){\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*)^{-1}{\mathsf A}{\mathsf C}_0\\
&= {\mathsf C}(\delta),
\end{eqnarray*}
which gives the result.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:mkconv}]
We may express $G_\ell$ in the form of an affine mapping,
\[
G_\ell(m,\delta) = {\mathsf H}_\ell(\delta)m + g_\ell(\delta)
\]
where ${\mathsf H}_\ell(\delta)$ and $g_\ell(\delta)$ are given by
\begin{eqnarray*}
{\mathsf H}_\ell(\delta) &= -\delta {\mathsf C}_0{\mathsf A}^*\big({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf C}_\ell(\delta){\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*\big)^{-1}{\mathsf M} \\
&= -\delta{\mathsf K}({\mathsf C}_\ell(\delta),\delta){\mathsf M},\\
g_\ell(\delta) &= m_0 + {\mathsf C}_0{\mathsf A}^*\big({\mathsf \Gamma} + \delta^2{\mathsf M}{\mathsf C}_\ell(\delta){\mathsf M}^* + {\mathsf A}{\mathsf C}_0{\mathsf A}^*\big)^{-1}(b - {\mathsf A} m_0) \\
&= m_0 + {\mathsf K}({\mathsf C}_\ell(\delta),\delta)(b - {\mathsf A} m_0),
\end{eqnarray*}
respectively. From the estimates of Lemma \ref{lem:ckbound}, we obtain the uniform bounds
\begin{eqnarray*}
\|{\mathsf H}_\ell(\delta)\|_{\mathcal{L}(X)} &\leq \beta\delta<1,\\
\|g_\ell(\delta)\|_X &\leq \|m_0\|_X + \big\|{\mathsf K}({\mathsf C}_\ell(\delta),\delta)\big\|_{\mathcal{L}(Y)}\|b - {\mathsf A} m_0\|_X\\
&= \|m_0\|_X +\beta \|b - {\mathsf A} m_0\|_X = L.
\end{eqnarray*}
From the convergence of the sequence $\{{\mathsf C}_\ell(\delta)\}_{\ell\geq 1}$ in the previous proposition, we see that $\{{\mathsf H}_\ell(\delta)\}_{\ell\geq 1}$ and $\{g_\ell(\delta)\}_{\ell\geq 1}$ also converge, the limits being denoted by ${\mathsf H}(\delta)$ and $g(\delta)$, respectively. Explicitly,
\begin{eqnarray*}
{\mathsf H}(\delta) &= -\delta {\mathsf K}({\mathsf C}(\delta),\delta){\mathsf M},\\
g(\delta) &= m_0 + {\mathsf K}({\mathsf C}(\delta),\delta)(b - {\mathsf A} m_0).
\end{eqnarray*}
Moreover, since ${\mathsf B}\mapsto {\mathsf K}({\mathsf B},\delta)$ is Fr\'{e}chet differentiable, this convergence occurs at the same rate as the convergence of $\{{\mathsf C}_\ell(\delta)\}_{\ell\geq 1}$.
Next we show that $\{m_\ell(\delta)\}_{\ell \geq 1}$ remains bounded for sufficiently small $\delta$. From the bounds above, we have
\begin{eqnarray*}
\|m_\ell(\delta)\|_X &\leq \|{\mathsf H}_{\ell-1}(\delta)\|_{\mathcal{L}(X)}\|m_{\ell-1}(\delta)\|_X + \|g_{\ell-1}(\delta)\|_X\\
&\leq \beta\delta\|m_{\ell-1}\|_X + L,
\end{eqnarray*}
and therefore, by repeatedly applying the estimate, we obtain
\begin{eqnarray*}
\|m_\ell(\delta)\|_X &\leq (\beta\delta)^\ell\|m_0\|_X + L\sum_{j=0}^{\ell-1}(\beta\delta)^j\\
&\leq (\beta\delta)^\ell\|m_0\|_X + \frac{L}{1-\beta\delta},
\end{eqnarray*}
which provides a uniform bound for $\delta < 1/\beta$.
To prove the convergence, we write first for $i \geq 1$ the estimate
\begin{eqnarray*}
\|m_{i+1}(\delta) - m_i(\delta)\|_X &= \|{\mathsf H}_i(\delta)m_i(\delta) + g_i(\delta) - {\mathsf H}_{i-1}(\delta)m_{i-1}(\delta) - g_{i-1}(\delta)\|_X\\
&\leq \|{\mathsf H}_i(\delta)m_i(\delta) - {\mathsf H}_i(\delta)m_{i-1}(\delta)\|_X\\
&\hspace{1cm} + \|{\mathsf H}_i(\delta)m_{i-1}(\delta) - {\mathsf H}_{i-1}(\delta)m_{i-1}(\delta)\|_X\\
&\hspace{1cm} + \|g_i(\delta) - g_{i-1}(\delta)\|_X\\
&\leq \|{\mathsf H}_i(\delta)\|_{\mathcal{L}(X)}\|m_i(\delta) - m_{i-1}(\delta)\|_X\\
&\hspace{1cm} + \|{\mathsf H}_i(\delta) - {\mathsf H}_{i-1}(\delta)\|_{\mathcal{L}(X)}\|m_{i-1}(\delta)\|_X\\
&\hspace{1cm} + \|g_i(\delta) - g_{i-1}(\delta)\|_X,
\end{eqnarray*}
and further, by the geometric convergence of the sequences $\{{\mathsf H}_i(\delta)\}$ and $\{g_{i}(\delta)\}$, and the uniform boundedness, for some $\gamma>0$,
\[
\|m_{i+1}(\delta) - m_i(\delta)\|_X \leq \beta\delta\|m_i(\delta) - m_{i-1}(\delta)\|_X + \gamma(\beta\delta)^{2i}.
\]
By by repeatedly applying the estimate, we arrive at
\begin{eqnarray*}
\|m_{i+1}(\delta) - m_i(\delta)\|_X&\leq (\beta\delta)^{i+1}\|m_0\|_X + \gamma\sum_{j=0}^i(\beta\delta)^j\big((\beta\delta)^2\big)^{i-j}\\
&= (\beta\delta)^{i+1}\|m_0\|_X + \gamma\frac{\big((\beta\delta)^2\big)^{i+1} - (\beta\delta)^{i+1}}{(\beta\delta)^2 - \beta\delta},
\end{eqnarray*}
From this bound, it follows that $\{m_\ell(\delta)\}_{\ell \geq 1}$ is a Cauchy sequence: For $k > \ell$, we have
\begin{eqnarray*}
\|m_k(\delta) - m_\ell(\delta)\|_X &\leq \sum_{i=\ell}^{k-1} \|m_{i+1}(\delta) - m_i(\delta)\|_X\\
&\leq \|m_0\|_X\sum_{i=\ell}^{k-1}(\beta\delta)^{i+1} + \gamma\sum_{i=\ell}^{k-1}\frac{\big((\beta\delta)^2\big)^{i+1} - (\beta\delta)^{i+1}}{(\beta\delta)^2 - \beta\delta}\\
&= \|m_0\|_X\frac{(\beta\delta)^{\ell+1} - (\beta\delta)^{k+1}}{1-\beta\delta} + \gamma\frac{(\beta\delta)^{\ell+1}-(\beta\delta)^{k+1}}{(\beta\delta - 1)((\beta\delta)^2 - \beta\delta)}\\
&\hspace{1cm} + \gamma\frac{\big((\beta\delta)^2\big)^{\ell+1} - \big((\beta\delta)^2\big)^{k+1}}{(\delta^2 - 1)(\delta^2 - \beta\delta)}
\end{eqnarray*}
which tends to zero as $k,\ell\rightarrow\infty$, provided $\delta$ is small enough. Thus the sequence $\{m_\ell(\delta)\}$ converges, and we denote the limit by $m(\delta)$. Taking the limit as $k\rightarrow\infty$ in the above inequality, we have
\begin{eqnarray*}
\fl\|m_\ell(\delta) - m(\delta)\|_X &\leq \frac{(\beta\delta)^{\ell+1}}{\beta\delta - 1} + \gamma\frac{(\beta\delta)^{\ell+1}}{(\beta\delta - 1)((\beta\delta)^2 - \beta\delta)} + \gamma\frac{\big((\beta\delta)^2\big)^{\ell+1}}{(\delta^2 - 1)((\beta\delta)^2 - \beta\delta)}\\
&= \mathcal{O}\big((\beta\delta)^\ell\big)
\end{eqnarray*}
for all $\ell \geq 1$, and it follows that $\{m_\ell(\delta)\}_{\ell\geq 1}$ converges geometrically with rate $\beta\delta$.
To show that the limit $m(\delta)$ is indeed a fixed point of $G(\cdot,\delta)$, we first we note that
\begin{eqnarray*}
\fl\|G_\ell(m_\ell(\delta),\delta) - G(m(\delta),\delta)\|_X &= \|{\mathsf H}_\ell(\delta)m_\ell(\delta) + g_\ell(\delta) - {\mathsf H}(\delta)m(\delta) - g(\delta)\|_X\\
&\leq \|{\mathsf H}_\ell(\delta) - {\mathsf H}(\delta)\|_{\mathcal{L}(X)}\|m_\ell(\delta)\|_X \\
&\hspace{1cm} + \|{\mathsf H}(\delta)\|_{\mathcal{L}(X)}\|m_\ell(\delta) - m(\delta)\|_X + \|g_\ell(\delta) - g(\delta)\|_X\\
&\rightarrow 0,
\end{eqnarray*}
as $\ell\rightarrow \infty$, and so it follows that
\[
m(\delta) = \lim_{\ell\rightarrow\infty} m_{\ell+1}(\delta)
= \lim_{\ell\rightarrow\infty} G_\ell(m_\ell(\delta),\delta)
= G(m(\delta),\delta).
\]
All that remains is to check that the fixed point is unique. Supposing that $h(\delta)$ is another fixed point, it follows that
\begin{eqnarray*}
\|m(\delta) - h(\delta)\|_X &= \|G(m(\delta),\delta) - G(h(\delta),\delta)\|_X\\
&\leq \|{\mathsf H}(\delta)\|_{\mathcal{L}(X)}\|m(\delta) - h(\delta)\|_X\\
&\leq \beta\delta\|m(\delta) - h(\delta)\|_X.
\end{eqnarray*}
Hence, for $\delta < 1/\beta$ we must have that $m(\delta) = h(\delta)$, and the result follows.
\end{proof}
\subsection{Convergence in Finite Dimensions}
\label{ssec:conv_finite}
Above we had to assume that the prior distribution was sufficiently close to the posterior in order to guarantee convergence; in finite dimensions we may drop this assumption and still get convergence of the covariances. Here we assume that $X$ and $Y$ are Euclidian spaces, $X = \mathbb{R}^N$ and $Y = \mathbb{R}^J$.
In the following, we use the following convention: For symmetric matrices
${\mathsf B}_1,{\mathsf B}_2$, we write ${\mathsf B}_1 \geq {\mathsf B}_2$, or ${\mathsf B}_1 > {\mathsf B}_2$ to indicate that ${\mathsf B}_1 - {\mathsf B}_2$ is non-negative definite, or positive definite, respectively. We denote by $S_+^N \subseteq \mathbb{R}^{N\times N}$ the set of positive definite $N\times N$ matrices. In this section, the adjoint operators are denoted as transposes.
We start by showing that the iterative updating formula (\ref{eq:ck_standard}) gives a convergent sequence. However, instead of the covariance matrices, it is more convenient to work with the precision matrices, defined as ${\mathsf B}_\ell := {\mathsf C}_\ell^{-1}$. Observe that formula (\ref{eq:ck_standard}) can be written as
\[
{\mathsf C}_{\ell+1}^{-1} = {\mathsf A}^{\mathsf T}({\mathsf \Gamma} + {\mathsf M} {\mathsf C}_\ell {\mathsf M}^{\mathsf T})^{-1}{\mathsf A} + {\mathsf C}_0^{-1},
\]
which motivates the following result.
\begin{proposition}
Let ${\mathsf B}_0 = {\mathsf C}_0^{-1} \in S^N_+$ be a positive definite precision matrix, and let the sequence $\{{\mathsf B}_{\ell}\}_{\ell \geq 0}$ be generated iteratively by the formula ${\mathsf B}_{\ell+1} = R({\mathsf B}_\ell)$, where $R:S^N_+\rightarrow S^N_+$ is given by
\[
R({\mathsf B}) = {\mathsf A}^{\mathsf T}({\mathsf \Gamma} + {\mathsf M} {\mathsf B}^{-1} {\mathsf M}^{\mathsf T})^{-1}{\mathsf A} + {\mathsf B}_0.
\]
Then the sequence $\{{\mathsf B}_\ell\}_{\ell\geq 0}$ is increasing in the sense of quadratic forms, and there exists a positive definite ${\mathsf B} \in S^N_{+}$ such that ${\mathsf B}_\ell\uparrow {\mathsf B}$.
Consequently, the sequence of covariances $\{{\mathsf C}_\ell\}_{\ell\geq 0}$ defined by (\ref{eq:ck_standard}) satisfies ${\mathsf C}_\ell\downarrow {\mathsf C} := {\mathsf B}^{-1}$.
\end{proposition}
\begin{proof}
We first show that ${\mathsf B}_{\ell+1} \geq {\mathsf B}_\ell$ for all $\ell$ using induction. Write
\[
r({\mathsf B}) = {\mathsf A}^{\mathsf T}({\mathsf \Gamma} + {\mathsf M} {\mathsf B}^{-1} {\mathsf M}^{\mathsf T})^{-1}{\mathsf A},
\]
so that $R({\mathsf B}) = r({\mathsf B}) + {\mathsf B}_0$. If ${\mathsf B} > 0$, then $r({\mathsf B}) \geq 0$, proving that ${\mathsf B}_1 = r({\mathsf B}_0) + {\mathsf B}_0 \geq {\mathsf B}_0$.
Now assume that ${\mathsf B}_{\ell} \geq {\mathsf B}_{\ell-1}$. Then we have
\begin{eqnarray*}
{\mathsf B}_{\ell+1} - {\mathsf B}_\ell &= R({\mathsf B}_\ell) - R({\mathsf B}_{\ell-1})\\
&= r({\mathsf B}_\ell) - r({\mathsf B}_{\ell-1})\\
&= {\mathsf A}^{\mathsf T}\Big(\big({\mathsf \Gamma} + {\mathsf M} {\mathsf B}_\ell^{-1} {\mathsf M}^{\mathsf T}\big)^{-1} - \big({\mathsf \Gamma} + {\mathsf M} {\mathsf B}_{\ell-1}^{-1}{\mathsf M}^{\mathsf T}\big)^{-1}\Big){\mathsf A}.
\end{eqnarray*}
To prove the claim, it suffices to show that the bracketed difference is non-negative definite. Consider the difference
\[
\big({\mathsf \Gamma} + {\mathsf M}{\mathsf B}_{\ell-1}^{-1}{\mathsf M}^{\mathsf T}\big) - \big({\mathsf \Gamma} + {\mathsf M}{\mathsf B}_\ell^{-1}{\mathsf M}^{\mathsf T}\big)
= {\mathsf M}\big({\mathsf B}_{\ell-1}^{-1} - {\mathsf B}_\ell^{-1}\big){\mathsf M}^{\mathsf T} \geq 0
\]
by the induction assumption. Therefore,
\[
\big({\mathsf \Gamma} + {\mathsf M}{\mathsf B}_{\ell-1}^{-1}{\mathsf M}^{\mathsf T}\big)^{-1} \leq \big({\mathsf \Gamma} + {\mathsf M}{\mathsf B}_\ell^{-1}{\mathsf M}^{\mathsf T}\big)^{-1},
\]
which implies the desired non-negative definiteness.
To prove that the sequence $\{{\mathsf B}_\ell\}_{\ell\geq 0}$ is bounded as quadratic forms, denote the quadratic form as
\[
{\mathsf Q}_\ell(u,v) = u^{\mathsf T} {\mathsf B}_\ell v,\quad u,v\in{\mathbb R}^N.
\]
Since
\[
\big({\mathsf \Gamma} + {\mathsf M}{\mathsf B}_{\ell-1}^{-1}{\mathsf M}^{\mathsf T}\big)^{-1} \leq {\mathsf \Gamma}^{-1},
\]
we have
\[
{\mathsf Q}_\ell(u,u) \leq ({\mathsf A} u)^{\mathsf T} {\mathsf \Gamma}^{-1} {\mathsf A} u + u^{\mathsf T} {\mathsf B}_0 u,
\]
proving the boundedness of the sequence.
In particular, it follows that for each $u\in{\mathbb R}^N$,
\[
{\mathsf Q}_\ell(u,u)\rightarrow u^{\mathsf T} {\mathsf B} u, \quad\mbox{as $\ell\to\infty$}
\]
for some symmetric positive definite matrix ${\mathsf B}\in {\mathbb R}^{N\times N}$. The fact that the matrix entries of ${\mathsf B}_\ell$ converge to the corresponding entries of ${\mathsf B}$ follows from the polar identity,
\[
u^{\mathsf T}{\mathsf B}_\ell v = \frac 14\left({\mathsf Q}_\ell(u+v) - {\mathsf Q}_\ell(u-v)\right),
\]
with $u,v$ being the canonical basis vectors. This completes the proof.
\end{proof}
\section{The General Case}
\label{sec:sampling}
In general, the sequence of distributions $\pi_\ell$ are not Gaussian, and so it
is considerably harder to analyze the convergence as we did in the previous section. In this section we consider how the algorithm may be implemented in practice, and in particular, how to produce an approximation to the posterior distribution using a finite number of full model evaluations. This approximate distribution can be used for generating samples using only approximate model evaluations,
leading to a significantly lower computational cost over sampling using the true posterior based on the full model.
In subsection \ref{ssec:update_densities} we outline the general framework for sampling from the approximate posterior sequence and updating the density, making use of particle approximations. In subsection \ref{ssec:iteration_op} we reformulate the iteration (\ref{eq:update_leb}) in terms of operators on the set of probability measures, and provide results on properties of these operators. Convergence in the large particle limit is shown, using the new formulation of the update. In subsection \ref{ssec:gauss_sampling} a particular rejection sampling method, based on a Gaussian mixture proposal,
is studied. Importance sampling is then considered in subsection \ref{ssec:importance} and similar convergence is shown.
\subsection{Updating the Densities via Sampling}
\label{ssec:update_densities}
We consider the algorithm in Section~\ref{ssec:algorithm}, and in particular, address the question of how to generate a sequence of approximate samples from the
iteratively defined densities $\mu_\ell$ given by (\ref{eq:update_leb}). We shall use particle approximations to do this. Assume that
\[
{\mathscr S}_\ell = \big\{(u_\ell^1,w_\ell^1),(u_\ell^2,w_\ell^2),\ldots,(u_\ell^N,w_\ell^N)\big\},\quad \ell = 0,1,2,\ldots
\]
is the current approximate sample of the unknowns with relative weights $w_\ell^j$. For $\ell=0$, the sample is obtained by independent sampling from the prior, and $w_0^j = 1/N$. We then compute the modeling error sample,
\[
{\mathscr M}_\ell = \big\{m_\ell^1,m_\ell^2,\ldots,m_\ell^N\big\},\quad \ell = 0,1,2,\ldots
\]
by defining
\[
m_\ell^j = M(u_\ell^j).
\]
Consider now the model (\ref{appr model}). Assuming that the modeling error is independent of the unknown $u$, we may write a conditional likelihood model,
\[
\pi(b\mid u,m) \propto \pi_{\rm noise}(b - f(u) - m).
\]
Let $\nu_\ell(m)$ denote the probability density of the modeling error based on our current information. Then, the updated likelihood model based on the approximate model is
\[
\pi_{\ell+1}(b\mid u) = \int \pi(b\mid u,m)\nu_\ell(m) dm,
\]
and, using a Monte Carlo integral approximation, postulating that the realizations $m_\ell^j$ inherit the weights of the sample points $u_\ell^j$, we obtain
\[
\pi_{\ell+1}(b\mid u) \approx \sum_{j=1}^Nw_\ell^j \pi(b\mid u,m_\ell^j).
\]
The current approximation for the posterior density is
\[
\pi_{\ell+1}(u\mid b) \propto \pi_{\rm prior}(u)\sum_{j=1}^Nw_\ell^j \pi(b\mid u,m_\ell^j),
\]
suggesting an updating scheme for ${\mathscr S}_\ell\rightarrow {\mathscr S}_{\ell+1}$:
\begin{itemize}
\item[(a)] Draw an index $k_j$ by replacement from $\{1,2,\ldots,N\}$, using the probabilities $w_\ell^j$;
\item[(b)] Draw the sample point
\begin{eqnarray}
\label{eq:update_draw}
u_{\ell+1}^j\sim \pi_{\rm prior}(u)\pi(b\mid u,m_\ell^{k_j}).
\end{eqnarray}
\end{itemize}
Part (b) above is straightforward, in particular, if the model is Gaussian and $f$ is linear, such as in the linearized model for EIT, since the measure (\ref{eq:update_draw}) is then a Gaussian. We will demonstrate the effectiveness of this approach in Section~\ref{ssec:eit}. Otherwise we may consider other sampling methods such as importance sampling; this is what is done in the following subsections.
\subsection{A convergence result for particle approximations}
\label{ssec:iteration_op}
In this section, we rewrite the updating formula in terms of mappings of measures, and analyze the convergence of the particle approximation under certain limited conditions.
Let $\mu_\ell$ denote the current approximation of the posterior density for $u$. The updated likelihood based on the modeling error is
\[
\pi_{\ell+1}( b\mid u) \propto \int_X \pi_{\rm noise}(b - f(u) - M(z))\mu_{\ell}(dz),
\]
and therefore, the updating, by Bayes' formula, is given by
\begin{equation}\label{def of P}
\mu_{\ell+1}(du) \propto \mu_{\rm prior}(du) \int_X \pi_{\rm noise}(b - f(u) - M(z))\mu_{\ell}(dz) = P\mu_{\ell}(du).
\end{equation}
Furthermore, we write the normalization formally as an operator,
\[
L\mu = \frac{\mu}{\mu(1)}, \quad \mu(1) = \int_X\mu(du).
\]
The model updating algorithm can therefore be written concisely as
\[
\mu_{\ell+1} = LP\mu_\ell,\quad \mu_0 = \mu_{\rm prior}.
\]
Let $\mathsf{M}(X)$ denote the set of finite measures on $X$. Denote by $\mathsf{P}(X)$ the set of probability measures on $X$, and for $p \in (0,1)$ denote by
$\mathsf{M}_p(X)$ the set of finite measures with total mass lying in the interval $[p,p^{-1}]$.
Let $\mu$ and $\nu$ denote two random $\mathsf{M}(X)$-valued measures, i.e., $\mu^\omega,\nu^\omega\in \mathsf{M}(X)$ for $\omega\in\Omega$, where $\Omega$ is a probability space. Denoting by ${\mathbb E}$ the expectation, we define
the distance between random measures through
\[
d(\mu,\nu)^2 = \sup_{\|\varphi\|_\infty=1} {\mathbb E} |\mu(\varphi)-\nu(\varphi)|^2,
\]
where the functions $\varphi$ are continuous over $X$. For non-random measures, the definition coincides with the total variation distance.
In the following two lemmas, which we need for the large particle convergence
result that follows them,
we make a restrictive assumption about the noise distribution.
\begin{assumptions}
\label{ass:particle}
There exists $\kappa \in (0,1)$ such that for all $\varepsilon \in Y$, $\kappa \leq \pi_{\mathrm{noise}}(\varepsilon) \leq \kappa^{-1}$.
\end{assumptions}
Under this assumption, we show the following results concerning the mappings $P$ and $L$ defined before.
\begin{lemma}
Let Assumptions \ref{ass:particle} hold. Then $P:\mathsf{P}(X)\rightarrow\mathsf{M}_\kappa(X)$, and
\[
d(P\mu,P\nu) \leq \kappa^{-1} d(\mu,\nu).
\]
\end{lemma}
\begin{proof}
First note that $\kappa \leq \pi_{\mathrm{noise}} <\kappa^{-1}$ implies that $\kappa \leq (P\mu)(1) \leq \kappa^{-1}$, and so $P$ does indeed map ${\mathsf P}(X)$ into $\mathsf{M}_\kappa(X)$. Exchanging the order of integration, we see for any bounded measurable $\varphi$,
\[
(P\mu)(\varphi) = \int_X\underbrace{\left(\int_X \pi_{\mathrm{prior}}(u)\pi_{\mathrm{noise}}(b-f(u) - M(z))\varphi(u){d} u\right)}_{=:\psi(z)}\mu({d} z)
\]
and so
\[
|(P\mu)(\varphi)-(P\nu)(\varphi)|^2 = |\mu(\psi) - \nu(\psi)|^2.
\]
Using that $\pi_{\mathrm{noise}} <\kappa^{-1}$, we see that $\|\varphi\|_\infty \leq 1$ implies that $\|\psi\|_\infty \leq \kappa^{-1}$, and so
\begin{eqnarray*}
d(P\mu,P\nu)^2 &\leq \sup_{\|\psi\|_\infty \leq \kappa^{-1}} {\mathbb E}|\mu(\psi) - \nu(\psi)|^2\\
&\leq \sup_{\|\psi\|_\infty \leq 1} \kappa^{-2}{\mathbb E} |\mu(\psi) - \nu(\psi)|^2\\
&= \kappa^{-2} d(\mu,\nu)^2,
\end{eqnarray*}
implying the claim.
\end{proof}
A similar result for the mapping $L$ can be obtained.
\begin{lemma}
Let Assumptions \ref{ass:particle} hold. Then it follows that $L:\mathsf{M}_\kappa(X)\rightarrow\mathsf{P}(X)$, and furthermore,
for $\mu,\nu \in \mathsf{M}_\kappa(X)$, we have
\[
d(L\mu,L\nu) \leq 2\kappa^{-2}d(\mu,\nu).
\]
\end{lemma}
\begin{proof}
The proof is essentially identical to that of Lemma 5.17 in \cite{lecturenotes}, with $1$ in place of $g$. We skip the details here.
\end{proof}
We use the above results to analyze the convergence of particle approximations of the measures. We introduce the sampling operator $S^N:\mathsf{P}(X)\rightarrow\mathsf{P}(X)$,
\[
S^N\mu = \frac{1}{N}\sum_{j=1}^N \delta_{u^j},\quad u^1,\ldots,u^N \sim \mu\;\mathrm{i.i.d.}
\]
and we have
\[
(S^N\mu)(\varphi) = \frac{1}{N}\sum_{j=1}^N \varphi(u^j),\quad u^1,\ldots,u^N \sim \mu\;\mathrm{i.i.d.}
\]
Observe that $S^N\mu$ is a random measure, as it depends on the sample.
It is shown in \cite{lecturenotes}, Lemma 5.15, that the operator $S^N$ satisfies
\[
\sup_{\mu \in \mathsf{P}(X)} d(S^N\mu,\mu) \leq \frac{1}{\sqrt{N}}.
\]
Define the sequence of particle approximations $\{\mu_\ell^N\}_{\ell\geq 0}$ to $\{\mu_\ell\}_{\ell\geq 0}$ by
\begin{eqnarray}
\label{eq:iteration_exact}
\nonumber\mu_0^N &= S^N \mu_0,\\
\mu_{\ell+1}^N &= S^NLP\mu_\ell^N.
\end{eqnarray}
in light of the previous lemmas, now we prove the following result regarding convergence of this approximation as $N\rightarrow\infty$:
\begin{proposition}
\label{prop:exact_sampling}
Let Assumptions \ref{ass:particle} hold. Define $\{\mu_\ell\}_{\ell\geq 0}$, $\{\mu_\ell^N\}_{\ell\geq 0}$ as above. Then, for each $\ell$,
\[
d(\mu_\ell^N,\mu_\ell) \leq \frac{1}{\sqrt{N}}\sum_{k=0}^\ell (2\kappa^{-3})^k.
\]
In particular, $d(\mu_\ell^N,\mu_\ell)\rightarrow 0$ as $N\rightarrow\infty$.
\end{proposition}
\begin{proof}
The triangle inequality for $d(\cdot,\cdot)$ yields
\begin{eqnarray*}
e_\ell := d(\mu_\ell^N,\mu_\ell) \leq d(S^NLP\mu_{\ell-1}^N,LP\mu_{\ell-1}^N) + d(LP\mu_{\ell-1}^N,LP\mu_{\ell-1}),
\end{eqnarray*}
and applying the bounds given by the previous lemmas, we obtain
\begin{eqnarray*}
e_\ell &\leq \frac{1}{\sqrt{N}} + 2\kappa^{-2}d(P\mu_{\ell-1}^N,P\mu_{\ell-1})\\
&\leq \frac{1}{\sqrt{N}} + 2\kappa^{-3} d(\mu_{\ell-1}^N,\mu_{\ell-1})\\
&= \frac{1}{\sqrt{N}} + 2\kappa^{-3} e_{\ell-1}.
\end{eqnarray*}
The result follows since $e_0 = d(S^N\mu_0,\mu_0) \leq 1/\sqrt{N}$.
\end{proof}
\subsection{Particle approximation with Gaussian densities}
\label{ssec:gauss_sampling}
In this section, we consider the particle approximation when the approximate model is linear, while the accurate model need not be. This is the situation in the computed examples that will be discussed later.
Suppose that the approximate model $f$ is linear, $f(u) = {\mathsf A} u$, and the noise and prior distributions are Gaussian,
\[
\pi_{\mathrm{noise}} = \mathcal{N}(0,{\mathsf \Gamma}),\quad \pi_{\mathrm{prior}} = \mathcal{N}(m_0,{\mathsf C}_0).
\]
Then the measure (\ref{eq:update_draw}) is a Gaussian mixture:
\[
\pi_{\mathrm{prior}}(u)\pi(b\mid u,m_\ell^j) = \frac{1}{N}\sum_{j=1}^N \mathcal{N}(u\mid p_\ell^j,{\mathsf C})
\]
where the means and covariance are given by
\begin{eqnarray*}
{\mathsf C} &= ({\mathsf A}^{\mathsf T}{\mathsf \Gamma}^{-1}{\mathsf A} + {\mathsf C}_0)^{-1},\\
p_\ell^j &= {\mathsf C}({\mathsf A}^{\mathsf T} {\mathsf \Gamma}^{-1}(b-m_\ell^j) + {\mathsf C}_0^{-1}m_0).
\end{eqnarray*}
The collection of samples $\mathscr{S}_\ell$ can then be evolved via the following algorithm.
{\bf Algorithm (Linear Approximate Model):}
\begin{enumerate}[\hspace{0.2cm}1.]
\item Set $\ell = 0$. Define the covariance operator ${\mathsf C} = ({\mathsf A}^{\mathsf T} {\mathsf \Gamma}^{-1}{\mathsf A} + {\mathsf C}_0)^{-1}$. Draw an initial ensemble of particles $\{u_\ell^j\}_{j=1}^N$ from the prior measure $\mu_0({d} u) = \pi_{\mathrm{prior}}(u){d} u$, and define the collection $\mathscr{S}_\ell = \{u_\ell^1,u_\ell^2,\ldots,u_\ell^N\}$.
\item Define the means $p_\ell^j = {\mathsf C}({\mathsf A}^{\mathsf T} {\mathsf \Gamma}^{-1}(b-M(u_\ell^j)) + {\mathsf C}_0^{-1}m_0)$, $j=1,\ldots,N$.
\item For each $j=1,\ldots,N$
\begin{enumerate}[(i)]
\item Sample $k_j$ uniformly from the set $\{1,\ldots,N\}$
\item Sample $u_{\ell+1}^j \sim N(p_\ell^{k_j},{\mathsf C})$
\end{enumerate}
\item Set $\mathscr{S}_{\ell+1} = \{u_{\ell+1}^1,u_{\ell+1}^2,\ldots,u_{\ell+1}^N\}$.
\item Set $\ell \mapsto \ell+1$ and go to 2.
\end{enumerate}
\begin{remark}
For more general models, one could use a method such as rejection sampling in order to produce exact samples from the measure (\ref{eq:update_draw}). A suitable proposal distribution for this rejection sampling could be, for example, a Gaussian mixture with appropriately chosen means and covariances \cite{CSbook}.
Two natural candidates for non-Gaussian priors, that retain some of the simplicity of the Gaussian models without being as limited, are:
\begin{enumerate}
\item Hierarchical, conditionally Gaussian prior models,
\[
\pi_{\rm prior}(u\mid\theta) \sim {\mathcal N}(\mu_\theta,{\mathsf C}_\theta),
\]
where the mean and covariance depend on a hyperparameter vector $\theta$ that follows a hyperprior distribution,
\[
\theta\sim\pi_{\rm hyper}.
\]
The hypermodels allow the introduction of sparsity promoting priors, similar to total variation; \cite{hypermodels,MEG}.
\item Gaussian mixtures, which allow a fast sampling from non-Gaussian distributions through a local approximation by Gaussian or other simple distributions \cite{west}.
\end{enumerate}
\end{remark}
\begin{remark}
We point out here that the approximation result of Proposition~\ref{prop:exact_sampling} does not apply to the Gaussian likelihood, as the noise density is not bounded from below.
\end{remark}
\subsection{Importance Sampling and Convergence}
\label{ssec:importance}
In this section we consider an approximate sampling based updating scheme of the probability densities using importance sampling.
This method effectively turns a collection of prior samples into samples from the posterior by weighting them appropriately, using the fact that the posterior is absolutely continuous with respect to the prior.
Assume that at stage $\ell$ of the approximation scheme, we have a collection of $N$ particles and the corresponding weights,
${\mathscr S}_\ell = \{(u_\ell^j,w_\ell^j)\}_{j=1}^N$. The associated particle approximation $\mu_\ell^N$ of the probability distribution acting on a test function $\varphi$ is
\[
\mu_\ell^N(\varphi) = \sum_{j=1}^N w_\ell^j\varphi(u_\ell^j).
\]
We evolve this distribution by acting on it with $P$ and $L$. By the definition (\ref{def of P}) of $P$, we first get an approximation
\begin{eqnarray*}
P\mu_\ell^N(du) & = \left(\sum_{j=1}^N w_\ell^j\pi_{\mathrm{noise}}(b-f(u)-M(u_\ell^j))\right)\mu_{\rm prior}({d} u) \\
&=: g_\ell(u)\mu_{\rm prior}({d} u).
\end{eqnarray*}
To generate an updated sample based on this approximation, we
use independent sampling to draw $u_{\ell+1}^j \sim \mu_0$, $j=1,\ldots,N$, and define the particle approximation by
\[
\mu_{\ell+1}^N(\varphi) = \sum_{j=1}^N w_{\ell+1}^j\varphi(u_{\ell+1}^j),\;\;\;w_{\ell+1}^j = \frac{g_\ell(u_{\ell+1}^j)}{\sum_{j=1}^N g_\ell(u_{\ell+1}^j)}.
\]
Denoting by $T^N:\mathsf{P}(X)\rightarrow\mathsf{P}(X)$ the importance sampling step, consisting of independent sampling and weighting, we may define an iterative algorithm symbolically as
\begin{eqnarray}
\label{eq:iteration_importance}
\nonumber\mu_0^N &= T^N \mu_0,\\
\mu_{\ell+1}^N &= T^NLP\mu_\ell^N.
\end{eqnarray}
Explicitly, the algorithm can be described as follows.
{\bf Algorithm (Importance sampling)}
\begin{enumerate}[\hspace{0.2cm}1.]
\item Set $\ell = 0$. Draw an initial ensemble of particles $\{u_\ell^j\}_{j=1}^N$ from the prior measure $\mu_0({d} u) = \pi_{\mathrm{prior}}(u){d} u$, and initialize the weights $w_\ell^j = 1/N$ for each $j=1,\ldots,N$. Define the collection $\mathscr{S}_\ell = \{(u_\ell^1,w_\ell^1),(u_\ell^2,w_\ell^2),\ldots,(u_\ell^N,w_\ell^N)\}$.
\item Define
\[
g_\ell(u) = \sum_{j=1}^N w_\ell^j\pi_{\mathrm{noise}}(b-f(u)-M(u_\ell^j)).
\]
\item Sample $u_{\ell+1}^j \sim \mu_0$, $j=1,\ldots,N$ i.i.d. and define the weights
\[
w_{\ell+1}^j = \frac{g_\ell(u_{\ell+1}^j)}{\sum_{j=1}^N g_\ell(u_{\ell+1}^j)},\quad j=1,\ldots,N.
\]
\item Set $\mathscr{S}_{\ell+1} = \{(u_{\ell+1}^1,w_{\ell+1}^1),(u_{\ell+1}^2,w_{\ell+1}^2),\ldots,(u_{\ell+1}^N,w_{\ell+1}^N)\}$.
\item Set $\ell \mapsto \ell+1$ and go to 2.
\end{enumerate}
As in the previous section, we establish a convergence result for $N\to\infty$ only under the restrictive condition of Assumption~\ref{ass:particle}.
We recall the following result from \cite{intrinsicdim}:
\begin{lemma}
Let $\mu \in \mathsf{P}(X)$ be absolutely continuous with respect to the prior measure $\mu_0$,
\[
\mu({d} u) \propto g(u)\mu_0({d} u),
\]
where $\mu_0(g^2) < \infty$. Define the quantity $\rho \geq 1$ by $\rho = \mu_0(g^2)/\mu_0(g)^2$. Then
\[
d(T^N\mu,\mu) \leq 2\sqrt{\frac{\rho}{N}}.
\]
\end{lemma}
By Assumption \ref{ass:particle}, there exists $\kappa \in (0,1)$ such that $\kappa \leq g_\ell(\varepsilon) \leq \kappa^{-1}$ for all $\varepsilon \in Y$, implying that
\[
\frac{\mu_0(g_\ell^2)}{\mu_0(g_\ell)^2} \leq \kappa^{-4}.
\]
In particular, by applying the above lemma to the measure $\hat{\mu}_{\ell+1} =LP\mu_{\ell}$, we see that
\[
d(T^N LP \mu_{\ell},LP{\mu}_{\ell}) \leq \frac{2\kappa^{-2}}{\sqrt{N}}.
\]
We are ready to prove the following proposition establishing the convergence of the particle approximations as $N\rightarrow\infty$:
\begin{proposition}
Let Assumptions \ref{ass:particle} hold for the noise distribution, and let $\{\mu_\ell\}_{\ell\geq 0}$ be the sequence of the model error approximations of the posterior, and $\{\mu_\ell^N\}_{\ell\geq 0}$ a sequence of importance sampling approximations obtained as above. Then, for each $\ell$,
\[
d(\mu_\ell^N,\mu_\ell) \leq \frac{2\kappa^{-2}}{\sqrt{N}}\sum_{k=0}^{\ell-1} (2\kappa^{-3})^k + \frac{(2\kappa^{-3})^\ell}{\sqrt{N}}.
\]
In particular, $d(\mu_\ell^N,\mu_\ell)\rightarrow 0$ as $N\rightarrow\infty$.
\end{proposition}
\begin{proof}
From the triangle inequality for $d(\,\cdot\,,\,\cdot\,)$, we have
\begin{eqnarray*}
e_\ell := d(\mu_\ell^N,\mu_\ell) \leq d(T^NLP\mu_{\ell-1}^N,LP\mu_{\ell-1}^N) + d(LP\mu_{\ell-1}^N,LP\mu_{\ell-1}).
\end{eqnarray*}
The bounds derived above yield
\begin{eqnarray*}
e_\ell &\leq \frac{2\kappa^{-2}}{\sqrt{N}} + 2\kappa^{-2}d(P\mu_{\ell-1}^N,P\mu_{\ell-1})\\
&\leq \frac{2\kappa^{-2}}{\sqrt{N}} + 2\kappa^{-3} d(\mu_{\ell-1}^N,\mu_{\ell-1})\\
&= \frac{2\kappa^{-2}}{\sqrt{N}} + 2\kappa^{-3}e_{\ell-1}.
\end{eqnarray*}
Since we have $e_0 = d(T^N\mu_0,\mu_0) = d(S^N\mu_0,\mu_0) \leq 1/\sqrt{N}$, the result follows.
\end{proof}
\begin{remark}
In theory, the importance sampling method described above can be used with very weak assumptions on the forward maps and prior/noise distributions. However in practice it may be ineffective if the posterior is significantly far from the prior, such as when the size of the observational noise is small. To overcome this issue, one could instead consider Sequential Monte Carlo or Sequential Importance Sampling methods to evolve prior samples into posterior samples by introducing a sequence of intermediate measures \cite{BJMS15,maceachern1999sequential}.
\end{remark}
\begin{comment}
Let $\nu(du)$ be a probability measure in ${\mathbb R}^n$ with finite mean $\overline u$ and covariance ${\mathsf C}$. The Gaussian mixture approximation of $\nu$
is defined as
\begin{equation}\label{Gaussian mixture}
\nu(du) \approx \frac 1N\sum_{j=1}^N {\mathcal N}(du\mid T_a y_j,h^2{\mathsf C}),\quad 0<h<1,
\end{equation}
where $T_a$ is a shrinkage,
\[
T_a u = a u + (1-a)\overline u,\quad a = \sqrt{1-h^2},
\]
and the points $y_j$ are sampled independently from $\nu$.
We define a continuous version of the Gaussian mixture by
\[
\nu^h(du) = \int {\mathcal N}(du \mid T_ay,h^2{\mathsf C})\nu(dy).
\]
Then, the discrete approximation can be viewed as an approximation by Monte Carlo integration. It is straightforward to verify that $\nu^h$ has the same mean and variance as $\nu$.
In the following, consider first a modification of the above approximation, in which the shrinkage is omitted, that is
\begin{equation}\label{no shrinkage}
\widetilde\nu^h(du) = \int {\mathcal N}(du \mid y,h^2{\mathsf C})\nu(dy).
\end{equation}
This approximation has variance $(1+h^2){\mathsf C}$.
{\bf Lemma:} {\em As $h\to 0+$, the approximation $\nu^h$ converges to $\nu$ in the Wasserstein metric.}
{\em Proof.} Let $\varphi\in {\rm Lip}_\alpha$, $\|\varphi\|_{{\rm Lip}_\alpha} = 1$. We have
\[
\widetilde\nu^h(\varphi) = \int\varphi(x) \widetilde\nu^h(dx) = \int\int \varphi(x){\mathcal N}(dx\mid y,h^2{\mathsf C}) \nu(dy) = \int \varphi^h(y) \nu(dy),
\]
where
\[
\varphi^h(y) = \int\varphi(x){\mathcal N}(dx\mid y,h^2{\mathsf C}) = \int \varphi(x) G_h(y-x) dx,
\]
the kernel $G_h$ being the scaled Gaussian,
\[
G_h(z) = \frac 1{h^n\sqrt{(2\pi)^n|{\mathsf C}|}} {\rm exp}\left(-\frac 1{2h^2}\|z\|_{\mathsf C}^2\right).
\]
Then,
\[
\widetilde \nu^h(\varphi) - \nu(\varphi) = \int(\varphi^h(y) - \varphi(y))\nu(dy),
\]
and
\[
\varphi^h(y) - \varphi(y) = \int (\varphi(x) - \varphi(y))G_h(y-x)dx = \int(\varphi(y-hz) -\varphi(y))G(z) dz,
\]
where $G(z) = G_1(z)$.
Therefore,
\[
|\varphi^h(y) - \varphi(y)| \leq h^\alpha \int |z|^\alpha G(z) dz,
\]
implying that
\[
| \widetilde \nu^h(\varphi) - \nu(\varphi)| \leq C h^\alpha.
\]
\hfill $\Box$.
To measure the distance from the Gaussian mixture approximation with the shrinkage, we write as above,
\[
\nu^h(\varphi) - \nu(\varphi) = \int (\varphi^h(T_ay) -\varphi(y))\nu(dy) = \int (\varphi^h(T_ay) -\varphi(T_a y))\nu(dy) +
\int (\varphi(T_ay) -\varphi(y))\nu(dy).
\]
The first term goes to zero pointwise in $L^\infty$, and uniformly to zero over functions of bounded Lipshitz norm. For the latter term, define the shrinkage $T_a^*$ of the measure,
\[
\int \varphi(y) T_a^*\nu(dy) = \int\varphi (T_a y) \nu(dy).
\]
In general we cannot argue that $d(T_a\nu,\nu)\to 0$ as $a\to 1$: Take $\nu = \delta_1 + \delta_{-1}$ in ${\mathbb R}$ to see this. However,
if $\varphi$ is Lipshitz, we have
\[
|\varphi(T_ay) - \varphi(y)| = |\varphi(y + (1-a)(\overline x - y))-\varphi(y)|\leq (1-a)^\alpha |\overline x -y|^\alpha,
\]
and therefore,
\[
|T_a^*\nu(\varphi) -\nu(\varphi)| \leq (1-a)^\alpha \int |\overline x - y|^\alpha \nu(dy) \rightarrow 0,
\]
as $h\to 0$, showing that
the Wasserstein distance does go to zero.
{\em Sampling:} Denote the sampled version of $\nu^h$ by
\[
S^N\nu^h(dx) = \frac 1N\sum_{j=1}^N {\mathcal N}(dx\mid T_ay_j,h^2{\mathsf C}).
\]
Then,
\[
S^N\nu^h(\varphi) = \frac 1N\sum_{j=1}^N \varphi(x) {\mathcal N}(dx\mid T_a y_j,h^2{\mathsf C}) = \frac 1N \sum_{j=1}^N \varphi^h(y_j),
\]
where we defined
\[
\varphi^h(y) = \int\varphi(x) {\mathcal N}(dx\mid T_a y,h^2{\mathsf C}).
\]
Observe that
\[
\|\varphi^h\|_\infty \leq \|\varphi\|_\infty.
\]
On the other hand,
\[
\nu^h(\varphi) = \int \int\varphi(x){\mathcal N}(dx\mid y,h^2{\mathsf C}) \nu(dy) = \int \varphi^h(y) \nu(dy).
\]
Therefore,
\[
S^N\nu^h(\varphi) = S^N\nu(\varphi^h),
\]
and therefore, for $ \|\varphi\|_\infty=1$,
\[
{\mathsf E}\{ |S^N\nu^h(\varphi) - \nu^h(\varphi)|\} = {\mathsf E} \{ |S^N\nu(\varphi^h) - \nu(\varphi^h)|\} \leq \frac 1{\sqrt{N}},
\]
where the expectation is over the realizations of the sampled values $y_j$. (Dashti\&Stuart)
\end{comment}
\section{Numerical Illustrations}
\label{sec:numerics}
In this section, we demonstrate the convergence properties established in the
preceding sections by means of computed examples. Furthermore, we demonstrate
the enhanced reconstructions obtained by modelling error as advocated
in this paper. The first example is a linear inverse source problem, elucidating the geometric convergence in the linear Gaussian case. The second example is the EIT problem with linearized approximate model with a coarse FEM mesh, allowing for straightforward particle updates. In the last example we consider the problem of recovering the permeability field in the steady state Darcy flow model, again with a linearized approximate model.
\subsection{Inverse Source Problem}
\label{ssec:num_inv}
As a proof of concept, we start by considering a simple one-dimensional inverse source problem.
Let $\Omega = (0,1)$ and define $X = L^2(\Omega)$. Given $u \in X$, let $p = P(u) \in H^1_0(\Omega)$ be the solution to the Laplace equation,
\[
\cases{\pushright{- p'' = u}&$x\in\Omega$\\
\pushright{p=0}&$x\in\partial\Omega$
}
\]
The inverse problem is to estimate the source $u$ from pointwise observations of $p$. Therefore, define the observation operator $\mathcal{O}:H^1_0(\Omega)\rightarrow\mathbb{R}^J$ by
\[
\mathcal{O}(u) = \big(u(q_1),\ldots,u(q_J)\big)
\]
for some set of points $\{q_1,\ldots,q_J\} \subseteq \Omega$. We define the exact forward operator ${\mathsf A}_\star^{\rm exact} = \mathcal{O}\circ P$. For numerical simulations, the exact forward model ${\mathsf A}_\star^{\rm exact}$ is approximated by a high fidelity proxy, ${\mathsf A}_\star$, obtained by approximating the solution $p$ through a finite difference solution on a fine mesh. The coarse mesh approximation of ${\mathsf A}_\star^{\rm exact}$, used in the inverse model, is denoted by ${\mathsf A} = {\mathsf A}_n$. In our computed example, we use $2^{10} - 1=1\,023$ equally spaced interior points for ${\mathsf A}_\star$, while the coarse mesh model ${\mathsf A}_n$ is computed with of $2^n - 1$ equally spaced interior points, $n < 10$.
We let $q_j = j/16$, $j=1,\ldots,15=J$ be equally spaced observation points, and to generate the simulated data, we corrupt the high fidelity data with a small amount of white noise, $\varepsilon \sim \mathcal{N}(0,{\mathsf \Gamma})$, where we set ${\mathsf \Gamma} = 10^{-8}{\mathsf I}_J$. The prior is chosen to be a standard Brownian motion, specifically we take $\mu_0 = \mathcal{N}(0,{\mathsf C}_0)$ with
\begin{eqnarray*}
{\mathsf C}_0 &= (-\Delta)^{-1},\quad \Delta = \frac{{d}^2}{{d} x^2},\\
\mathcal{D}(-\Delta) &= \bigg\{u \in H^2(\Omega)\,\bigg|\,u(0) = 0,\quad\frac{{d} u}{{d} x}(1) = 0\bigg\},
\end{eqnarray*}
and the true source used for data generation is drawn from the prior. Numerically the precision operator ${\mathsf C}_0^{-1}$ is implemented as the finite difference Laplacian matrix. We perform $L = 30$ iterations in each simulation.
The posterior mean and covariance, $m_{\mathrm{post}}, {\mathsf C}_{\mathrm{post}}$, corresponding to the high fidelity model, and the corresponding mean and covariance, $m_{\mathrm{post}}^n, {\mathsf C}_{\mathrm{post}}^n$, based on the approximate model are given by
\begin{eqnarray*}
m_{\mathrm{post}} &= m_0 + {\mathsf C}_0{\mathsf A}_\star^{\mathsf T}({\mathsf \Gamma} + {\mathsf A}_\star{\mathsf C}_0{\mathsf A}_\star^{\mathsf T})^{-1}(b-{\mathsf A}_\star m_0),\\
{\mathsf C}_{\mathrm{post}} &= {\mathsf C}_0 - {\mathsf C}_0{\mathsf A}_\star^{\mathsf T}({\mathsf \Gamma} + {\mathsf A}_\star{\mathsf C}_0{\mathsf A}_\star^{\mathsf T})^{-1}{\mathsf A}_\star{\mathsf C}_0,\\
m_{\mathrm{post}}^n &= m_0 + {\mathsf C}_0{\mathsf A}_n^{\mathsf T}({\mathsf \Gamma} + {\mathsf A}_n{\mathsf C}_0{\mathsf A}_n^{\mathsf T})^{-1}(b-{\mathsf A}_n m_0),\\
{\mathsf C}_{\mathrm{post}}^n &= {\mathsf C}_0 - {\mathsf C}_0{\mathsf A}_n^{\mathsf T}({\mathsf \Gamma} + {\mathsf A}_n{\mathsf C}_0{\mathsf A}_n^{\mathsf T})^{-1}{\mathsf A}_n{\mathsf C}_0,
\end{eqnarray*}
respectively. The approximate posterior mean and covariances (\ref{eq:ck_standard}) obtained by the modeling error approach, after $\ell$ iterations, are denoted by $m_\ell$ and ${\mathsf C}_\ell$, respectively.
Table \ref{tab:table_pde} shows the approximation errors arising from both approximations of ${\mathsf C}_\mathrm{post}$ and $m_\mathrm{post}$ with different discretization levels. The table shows that the modeling error approach produces a better approximation of the posterior mean than the model ignoring the modeling error, while the approximate covariances are
slightly less accurate as approximations of the posterior covariance than those
found without allowing for the modeling error correction. These experiments
confirm our assertion at the start of the paper, namely that allowing for
model error can result in improved point estimates (here the posterior mean)
but that the iteration introduced does not converge to the true posterior
distribution (as evidenced by the error in the covariance at fixed $n$ and large $L$.)
To demonstrate the convergence rate, Figure \ref{fig:errors_pde} shows the
mean and covariance errors for various approximation levels as functions of the number of iterations. The plots, as well as the tabulated values, Table \ref{tab:table_pde_slope}, of the logarithmic slopes of the approximation errors verify the geometric convergence rates, with their dependence on the approximation level. Observe that the logarithm of the convergence rate for the covariance, a quadratic quantity, is twice that of the mean.
\begin{table}
\begin{center}
\caption{The approximation errors of the approximate posterior means and covariances for various approximation levels with and without the inclusion of the modeling error correction. The matrix norms are the Frobenius norms.}
\label{tab:table_pde}
\begin{tabular}{l|c|c|c|c}
$n$ & $\|m_L - m_\mathrm{post}\|$ & $\|m_\mathrm{post}^n - m_\mathrm{post}\|$ & $\|{\mathsf C}_L - {\mathsf C}_\mathrm{post}\|$ & $\|{\mathsf C}_\mathrm{post}^n - {\mathsf C}_\mathrm{post}\|$\\
\hline
4 & 0.1906 & 0.2986 & 0.0676 & 0.0381\\
5 & 0.0455 & 0.0739 & 0.0170 & 0.0095\\
6 & 0.0111 & 0.0182 & 0.0043 & 0.0024 \\
7 & 0.0028 & 0.0045 & 0.0011 & 0.0006 \\
8 & 0.0007 & 0.0011 & 0.00026 & 0.00015 \\
9 & 0.0002 & 0.0003 & 0.00006 & 0.00004 \\
\end{tabular}
\end{center}
\end{table}
\begin{table}
\begin{center}
\caption{The convergence rates quantified in terms of the slope of the logarithmic plot of mean and covariance deviation from the limit values $m_L$ and ${\mathsf C}_L$ before the deviations plateau, indicating that the algorithm has converged.}
\label{tab:table_pde_slope}
\begin{tabular}{l|c|c|c}
$n$ & Slope of $\log\|m_\ell - m_L\|$ & Slope of $\log\|{\mathsf C}_\ell - {\mathsf C}_L\|$ & $\|{\mathsf A}_n-{\mathsf A}_\star\|$\\
\hline
4 & -1.80 & -3.60 & 0.101\\
5 & -3.13 & -6.26 & 0.0706\\
6 & -4.57 & -9.00 & 0.0491\\
7 & -6.30 & -11.9 & 0.0335\\
8 & -7.73 & -14.8 & 0.0219\\
9 & -9.34 & -14.9 & 0.0127
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\centerline{
\includegraphics[width=5.2cm]{images_new/pde_err_4.eps}
\includegraphics[width=5.2cm]{images_new/pde_err_5.eps}
\includegraphics[width=5.2cm]{images_new/pde_err_6.eps}
}
\centerline{
\includegraphics[width=5.2cm]{images_new/pde_err_7.eps}
\includegraphics[width=5.2cm]{images_new/pde_err_8.eps}
\includegraphics[width=5.2cm]{images_new/pde_err_9.eps}
}
\caption{The trace of the errors $\|m_\ell - m_L\|$ and $\|{\mathsf C}_\ell - {\mathsf C}_L\|$, illustrating their convergence rates. From left to right, top to bottom, the approximation level $n$ is increased from 4 to 9.}
\label{fig:errors_pde}
\end{figure}
\subsection{Electrical Impedance Tomography (EIT)}
\label{ssec:eit}
In this section, we revisit the modeling error due to coarse discretization of a PDE model in the context of Electrical Impedance Tomography (EIT). Let $\Omega\subset {\mathbb R}^d$, $d=2,3$, denote a bounded connected set with boundary $\partial\Omega$, and let $\sigma:\Omega\to {\mathbb R}$ be a function modeling the electric conductivity in $\Omega$, $0<\sigma_m\leq\sigma\leq\sigma_M<\infty$. We assume that $S$ electrodes are attached to the boundary $\partial\Omega$, and we model them as open disjoint surface patches $e_s\subset\partial\Omega$, $1\leq s\leq S$.
Assuming that an electric current $I_s$ is injected through $e_s$ into the body modeled by $\Omega$, the electric voltage potential $v$ in $\Omega$, and the electrode voltages $V_s$ on the electrodes can be found as a solution of the Complete Electrode Model (CEM) boundary value problem \cite{SIC},
\begin{eqnarray*}
\cases{\pushright{\nabla\cdot\big(\sigma\nabla v\big) = 0} & $x \in \Omega$\\
\pushright{\sigma\frac{\partial v}{\partial n} = 0} & $x \in \partial\Omega\setminus\bigcup_{s=1}^S e_s$\\
\pushright{v+z_s \sigma\frac{\partial v}{\partial n} = V_s} & $x \in e_s$, $1\leq s\leq S$\\
\pushright{\int_{e_s} \sigma\frac{\partial v}{\partial n} dS = I_s} & $1\leq s\leq S$.}
\end{eqnarray*}
Here, the parameters $z_s>0$ are the presumably known contact impedances, and the currents satisfy the Kirchhoff's law, or conservation of charge condition,
\[
\sum_{s=1}^S I_s \in{\mathbb R}^S_0 =\bigg\{V\in{\mathbb R}^L\,\bigg|\, \sum_{s=1}^S V_s = 0\bigg\}.
\]
The solution of the boundary value problem is the unique solution $(v,V)\in H^1(\Omega)\times {\mathbb R}^S_0$ of the weak form variational problem
\[
{\mathscr B}((w,W),(v,V)) = \sum_{s=1}^S I_s W_s = \langle (w,W),b_I\rangle,
\quad\mbox{for all $(w,W)\in H^1(\Omega)\times{\mathbb R}^S_0$,}
\]
where $b_I = (0,I)\in H^1(\Omega)\times{\mathbb R}^S_0$, and
\[
{\mathscr B}((w,W),(v,V)) = \int_\Omega \sigma\nabla w \cdot \nabla v dx + \sum_{s=1}^S \frac 1{z_s}\int_{e_s}(w - W_s)(v-V_s) dS.
\]
To discretize the problem, assume that $\Omega$ is approximated by the union of triangular or tetrahedral elements, the mesh containing $n_{\rm f}$ nodes (`f' for fine), and let $\{\psi_j\}_{j=1}^{n_{\rm f}}$ denote a nodal-based piecewise polynomial Lagrange basis. Further, let $\{\phi_s\}_{s=1}^{S-1}$ denote a basis of ${\mathbb R}^S_0$. We define the basis functions $\overline \psi_j\in H^1(\Omega)\times {\mathbb R}^S_0$ as
\[
\overline \psi_j = (\psi_j,0),\quad 1\leq j\leq n_{\rm f},\quad \overline\psi_{n_{\rm f}+s} = (0,\phi_s),\quad 1\leq s\leq S-1.
\]
We approximate the potential-voltage pair $(v,V)$ as
\[
(v,V) = \sum_{j=1}^{n_{\rm f}+S-1} \alpha_j\overline \psi_j,
\]
and discretize the forward problem by choosing $(w,W) = \overline\psi_k$, to arrive at the Galerkin approximation,
\begin{equation}\label{eq for alpha}
\sum_{j=1}^{n_{\rm f}+S-1} {\mathscr B}(\overline\psi_k,\overline \psi_j) \alpha_j = \langle \overline\psi_k, b_I\rangle, \quad 1\leq k\leq n_{\rm f}+S-1.
\end{equation}
Further, to parametrize the conductivity, we define a discretization of $\Omega$ by triangular or tetrahedral elements, independent of the discretization above, with $K$ nodes, and denote by $\{\eta_j\}_{j=1}^K$ the nodal-based piecewise polynomial Lagrange basis functions. We then parametrize the conductivity by writing
\[
\sigma(x) = \sigma_0{\rm exp}\left(\sum_{j=1}^K u_j \eta_j(x)\right),\quad x\in\Omega,
\]
where $\sigma_0>0$ is a fixed background conductivity. The matrix $[{\mathscr B}(\overline\psi_k,\overline \psi_j)]$ defining the system (\ref{eq for alpha}) is parametrized by the vector $u\in{\mathbb R}^K$, and we write the equation in matrix form concisely as
\[
{\mathsf A}^{n_{\rm f}}_u \alpha = b(I),
\]
where we have indicated explicitly the dependency on the discretization by the number $n_{\rm f}$ of nodes. Solving this system for $\alpha$, extracting the last $S-1$ components $\alpha_{n_{\rm f} + s}$, $1\leq s\leq S-1$, and representing the voltage in terms of the basis functions $\phi_s$ defines the forward map
\[
u \mapsto V = \sum_{s=1}^{S-1}\alpha_{n_{\rm f}+s}\phi_s = {\mathsf R}_u^{n_{\rm f}} I, \quad\mbox{where $\alpha =({\mathsf A}_u^{n_{\rm f}})^{-1}b(I)$,}
\]
where ${\mathsf R}^{n_{\rm f}}_u \in{\mathbb R}^{S\times S}$ is the resistance matrix. We repeat the calculation for a full frame of $S-1$ linearly independent current patterns, $I^1,\ldots, I^{S-1}\in{\mathbb R}^S_0$, obtaining the full frame of voltage patterns $V^1,\ldots,V^{S-1}$. Finally, the voltage patterns are stacked together in a vector, constituting the forward model for the observation,
\[
{\mathsf V} = \left[\begin{array}{c} V^1 \\ \vdots \\ V^{S-1}\end{array}\right] = F^{n_{\rm f}}(u), \quad F^{n_{\rm f}}:{\mathbb R}^K\to{\mathbb R}^{S(S-1)}.
\]
To guarantee satisfactory accuracy of the forward model, the discretization needs to be fine enough, in particular to capture the singularities of the voltage potential $v$ at the electrode edges. To demonstrate the modeling error effect, we construct a forward map defined over a coarser FEM mesh with $n_{\rm c}$ nodes ('c' for coarse), $n_{\rm c}<n_{\rm f}$, and denote the corresponding forward map by
\[
{\mathsf V} = F^{n_{\rm c}}(u), \quad F^{n_{\rm c}}:{\mathbb R}^K\to{\mathbb R}^{S(S-1)}, \quad n_{\rm c}<n_{\rm f}.
\]
Observe that the discretization of $u$ is independent of the FEM mesh, and is not changed when passing to a coarser computational mesh. In our computed examples, we use a piecewise linear Lagrange basis to represent both $u$ and $v$ over the different meshes. The three meshes that we base our simulations, generated with the mesh generator described in \cite{perssonStrang}, on are shown in Figure~\ref{fig:meshes}. The number of electrodes is $S=16$.
\begin{figure}[h!]
\centerline{
\includegraphics[width=4cm]{images/ConductivityMesh.eps}
\includegraphics[width=4cm]{images/CoarseMesh.eps}
\includegraphics[width=4cm]{images/FineMesh.eps}
}
\caption{\label{fig:meshes} Triangular meshes used in the numerical simulations. The number of electrodes is $L=16$, and they are indicated by red nodal points in the plot. The mesh for representing the conductivity distribution (left) has $K=733$ vertices and 1\,364 elements. The coarse mesh for the forward solver (middle) has $n_{\rm c}=877$ vertices and 1\,592 elements, and the fine scale mesh (right) consist of $n_{\rm f} =2\,418$ vertices and 4\,562 elements.
}
\end{figure}
We assign the Whittle-Mat\'{e}rn prior \cite{roininen2011, roininen2014} for the vector $u$ defining the conductivity so that
\begin{eqnarray}
\label{eq:wmprior}
\zeta \lambda^{-1} \left(-\lambda^2 {\mathsf L}_g + {\mathsf I}_K \right) u \sim {\mathcal N}(0, {\mathsf I}_K),
\end{eqnarray}
where ${\mathsf L}_g\in {\mathbb R}^{K \times K}$ is the graph Laplacian defined on the conductivity mesh, $\lambda>0$ is a correlation length parameter, $\zeta >0$ is amplitude scaling, and ${\mathsf I}_K$ is the identity matrix. In Figure~\ref{fig:prior draws}, three independently drawn realizations of the conductivity distributions are shown. The values of the model parameters are indicated in the figure caption.
\begin{figure}[h!]
\centerline{
\includegraphics[width=4cm]{images/ConductivityDraw_1.eps}
\includegraphics[width=4cm]{images/ConductivityDraw_2.eps}
\includegraphics[width=4cm]{images/ConductivityDraw_3.eps}
}
\caption{\label{fig:prior draws} Conductivities $\sigma = \sigma_0{\rm exp} (u)$ corresponding to three independent draws of $u$ from the prior density. The parameter values used here are $\lambda = 0.2$ and $\zeta = 1/15$.
The background conductivity is $\sigma_0 = 1.5$. The radius of the disc is unity, and the units are arbitrary.
}
\end{figure}
We generate the data using the fine scale model $F = F^{n_{\rm f}}$, and using the Conventional Error Model, i.e., ignoring the modeling error, compute a MAP estimate $u_{\rm MAP}$ using the forward map $f = F^{n_{\rm c}}$ in the inverse solver. The estimate is based on a simple Gauss-Newton iteration. The additive noise covariance in this simulation is ${\mathsf \Gamma} = \gamma^2{\mathsf I}_{S(S-1)}$ with $\gamma= 10^{-3} V_{\rm max}$, where $V_{\rm max}$ is the maximum of all noiseless electrode voltages over the full frame of $S-1$ voltage patterns. The noise level is assumed to be low so that the modeling error is the predominant part in the uncertainty. In Figure~\ref{fig:true etc}, we show the conductivity distribution that was used to generate the synthetic data with the model $F$, the Conventional Error Model MAP estimate based on the coarse mesh model $f$, as well as the Enhanced Error Model estimate. In the latter, the modeling error mean and covariance are estimated from a sample of 1\,500 random draws from the prior of $u$.
\begin{figure}[h!]
\centerline{
\includegraphics[width=4cm]{images/TrueConductivity.eps}
\includegraphics[width=4cm]{images/GN_ME_off.eps}
\includegraphics[width=4cm]{images/GN_ME_on.eps}
}
\caption{\label{fig:true etc} Left: The true conductivity used to generate the test data using the finely discetized FEM forward model. Center: A Gauss-Newton-based MAP estimate based on the coarsely discretized FEM forward model, using the Conventional Error Model that ignores the modeling error. Right: The MAP estimate computed by using the Enhanced Error Model, in which the modeling error mean and covariance are estimated from 1\,500 random draws from the prior.
}
\end{figure}
Observe that in the reconstruction based on the Conventional Error Model, the true inclusions are completely overshadowed by the boundary artifacts that are concentrated around the edges of the electrodes. This is to be expected, since the basis functions in the coarse FEM mesh do not capture the voltage singularities at the electrode edges, and the inverse solution compensates the modeling error with elevated conductivity at the edges to mitigate the singularity. In agreement with previously published results, the Enhanced Error Model produces a solution without modeling error artifacts.
The computation of the MAP estimate, regardless of the error model, requires repeated linearization of the forward map. The re-evaluation of the Jacobian may be time consuming, and therefore it is tempting to replace the coarse mesh FEM model with a linearized approximation around the background conductivity $\sigma_0$ corresponding to $u=0$,
\[
f(u) = F^{n_{\rm c}}(0) + {\mathsf D} F^{n_{\rm c}}(0) u.
\]
The solution of the inverse problem with the linearized model and Gaussian prior is particularly straightforward, requiring a solution of a linear system. We iterate the posterior updating algorithm, generating samples ${\mathscr S}_\ell = \{u_\ell^1,\ldots,u_\ell^N\}$, $\ell=0,1,2,\cdots$ using the modeling error updating scheme. In Figure~\ref{fig:estimates, linear}, we plot the conductivities corresponding to the posterior means,
\[
\overline\sigma_\ell = \sigma_0{\rm exp}(\overline u_\ell),\quad \overline u_\ell = \frac 1N\sum_{j=1}^N u_\ell^j,
\]
as well as the marginal variances of the parameters $u_\ell$, that is,
\[
{\rm var}_\ell = {\rm diag}\left(\frac 1N\sum_{j=1}^N(u^j_\ell - \overline u_\ell)(u^j_\ell - \overline u_\ell)^{\mathsf T}\right).
\]
The sample size here was $N = 5\,000$.
\begin{figure}
\centerline{\includegraphics[width=3.2cm]{images/MeanCoarseLinearNew_1.pdf}
\includegraphics[width=3.2cm]{images/MeanCoarseLinearNew_2.pdf}
\includegraphics[width=3.2cm]{images/MeanCoarseLinearNew_3.pdf}
\includegraphics[width=3.2cm]{images/MeanCoarseLinearNew_4.pdf}
\includegraphics[width=3.2cm]{images/MeanCoarseLinearNew_5.pdf}
}
\centerline{\includegraphics[width=3.2cm]{images/VarCoarseLinearNew_1.pdf}
\includegraphics[width=3.2cm]{images/VarCoarseLinearNew_2.pdf}
\includegraphics[width=3.2cm]{images/VarCoarseLinearNew_3.pdf}
\includegraphics[width=3.2cm]{images/VarCoarseLinearNew_4.pdf}
\includegraphics[width=3.2cm]{images/VarCoarseLinearNew_5.pdf}
}
\caption{\label{fig:estimates, linear} Sample means of the conductivity $\overline u_\ell$ for $\ell=1,2,3,4$ (upper row), and sample marginal variances of the components of the vectors $u$ at the same iterations.
}
\end{figure}
Finally, we consider the convergence of the iterated densities $\pi_\ell$ towards the posterior density by means of the Kullback-Leibler divergence, which we approximate using the particles drawn from $\pi_\ell$,
\begin{eqnarray*}
D_{\rm KL}(\pi_\ell \| \pi_{\rm post}) &=& \int\pi_\ell(u) \log\left(\frac{\pi_\ell(u)}{\pi_{\rm post}(u)}\right)du \approx \frac 1N \sum_{j=1}^N
\log\left(\frac{\pi_\ell(u_\ell^j)}{\pi_{\rm post}(u_\ell^j)}\right) \\
&=&\frac 1N \sum_{j=1}^N
\log\left(\frac{\pi_\ell(b\mid u_\ell^j)}{\pi(b\mid u_\ell^j)}\right) - \log\left(\frac{\pi(b)}{\pi_\ell(b)}\right),
\end{eqnarray*}
the second term corresponding to the normalization factors of the true and approximate posteriors.
Observe that to evaluate the posterior density, the fine mesh model needs not to be evaluated anew, since the fine mesh evaluations are already computed for the modeling error sample. The sample-based approximation of the KL divergence is straightforward to compute up to the normalizing constants.
Figure~\ref{fig:mean error,linear} shows the sample-based estimates of the Kullback-Leibler divergence for $\ell = 1,2,\ldots,\ell_{\rm max} = 5$. To subtract the unknown normalization offset, we plot the differences
\begin{equation}\label{DeltaKL}
\Delta D_{\rm KL}(\pi_\ell \| \pi_{\rm post}) = D_{\rm KL}(\pi_\ell \| \pi_{\rm post}) - D_{\rm KL}(\pi_{\ell_{\rm max}} \| \pi_{\rm post}).
\end{equation}
The figure shows also the relative error of the sample mean approximating the true conductivity,
\[
e_r(\overline u_\ell) = \frac{\| \sigma -\overline\sigma_\ell\|}{\|\sigma\|}.
\]
As in the previous subsection, the numerical results demonstrate that the
approach to model error advocated in this paper leads to improved estimates
of the true value used to generate the data.
\begin{figure}
\centerline{\includegraphics[height=5.3cm]{images/KLdivergence.eps}
\includegraphics[height=5cm]{images/DistFromTrue.eps}}
\caption{\label{fig:mean error,linear} Left: The Kullback-Leibler divergence (\ref{DeltaKL}) difference estimated by using a sample of $5\,000$ realizations drawn from the approximate posterior densities $\pi_\ell$. Right: The relative distance of the sample mean from the true conductivity over five iterations.}
\end{figure}
\subsection{Steady State Darcy Flow}
\label{ssec:gwf}
In the last computed example, we consider the inverse problem of estimating the permeability distribution in porous medium from a discrete set of pressure measurements. More precisely, let the computational domain be $\Omega = (0,1)^2$, and define $X = L^\infty(\Omega)$. For a given $u \in X$, called {\em log-permeability} define the pressure filed $p = P(u) \in H^1_0(\Omega)$ to be the solution to the steady-state Darcy equation with Dirichlet boundary conditions,
\begin{eqnarray}
\label{eq:gwf_pde}
\cases{\pushright{-\nabla\cdot(e^u\nabla p) = g} &$x\in\Omega$\\
\pushright{\hfill p=0} &$x\in\partial\Omega$
}
\end{eqnarray}
for some fixed and presumably known source term $g \in H^{-1}(\Omega)$.
We define now the observation operator $\mathcal{O}:H^1_0(\Omega)\rightarrow\mathbb{R}^J$ by
\[
\mathcal{O}_j(p) = \frac{1}{2\pi\varepsilon}\int_\Omega p(x)e^{-\frac{1}{2\varepsilon^2}(x-q_j)^2}\,{d} x,\quad j=1,\ldots,J,
\]
for some set of points $\{q_1,\ldots,q_J\} \subseteq \Omega$. The observations are smoothed observations at the points $\{q_1,\ldots,q_J\}$, converging to point observations as $\varepsilon\to 0$. Note that each $\mathcal{O}_j$ is a bounded linear functional on $H^1_0(\Omega)$. In what follows, we choose $\varepsilon = 0.02$, $g(x_1,x_2) = 100\sin(\pi x_1)\sin(\pi x_2)$, and let $\{q_1,\ldots,q_{25}\}$ be a uniformly spaced grid of $25$ points in $\Omega$. The accurate model is then defined by the composition $F = \mathcal{O}\circ P$. As in the EIT example, the approximate model is defined through linearization,
\[
f(u) = F(u_0) + \mathsf{D}F(u_0)u.
\]
for some fixed $u_0 \in X$; in this example we choose $u_0 = 0$. To construct the linear model, the derivative may be computed inexpensively using an adjoint method. The computation of the full Jacobian $\mathsf{D}F(u_0)$ requires $J+1$ numerical solutions of a PDE of the form (\ref{eq:gwf_pde}), and needs to be performed only once.
In this example, we generate three different data sets
corresponding to different noise levels: The noiseless data generated by using the non-linear model is
perturbed by additive observational noise drawn from normal distribution
$\mathcal{N}(0,{\mathsf \Gamma}_i)$, where ${\mathsf \Gamma}_i = 10^{-i-1}{\mathsf I}$, and $i=1,2,3$.
The true log-permeability $u^\dagger$, defined as the sum of two unnormalized Gaussian densities, is shown in Figure \ref{fig:gwf_truth}. In the same figure, the computed pressure field is shown, with the observation points indicated by black dots. Each data set is generated using a uniform mesh of $128\times 128$ points, while in the inverse computations, we use a reduced model with a uniform mesh of $64\times 64$ points. We perform $10$ iterations of the posterior updating algorithm, using as few as $100$ particles in the particle approximation model.
As in the EIT simulations, we choose a Whittle-Mat\'ern prior distribution for the vector $u$ defining the permeability as given in (\ref{eq:wmprior}). We make the choices $\lambda=0.1$ and $\zeta=1$, and note that here ${\mathsf L}_g$ corresponds to the the finite-difference Laplacian on the reduced mesh using the standard 5-point stencil.
In Figure \ref{fig:gwf_means} the conditional means arising from the different error models and data sets are shown. In this example the conventional and enhanced error models have very similar performance in terms of inferring the conditional mean. They are both able to infer the geometry of the log-permeability field, particularly when the observational noise is small, however they fail to obtain the magnitude. The iterative algorithm proposed in this article is able to obtain both the geometry and magnitude with good accuracy in a small number of iterations. Figure \ref{fig:gwf_conv} shows the evolution of the size of the error between the conditional mean at iteration $\ell$ and the true log-permeability field. In all cases the error has converged in 4 or 5 iterations, similarly to what was observed in the EIT experiments.
As in the previous two subsections, the numerical results demonstrate that the approach to model error advocated in this paper leads to improved accuracy of the point estimates of the true parameter underlying the data.
\begin{figure}
\includegraphics[width=\linewidth, trim=1cm 1cm 1cm 1cm, clip]{images/gwf_truth.eps}
\caption{(Left) The true log-permeability used to generate the data. (Right) The true pressure and the observation points $\{q_j\}_{j=1}^{25}$.}
\label{fig:gwf_truth}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[width = 15cm]{images_new/gwf_1.eps}
}
\centerline{
\includegraphics[width = 15cm]{images_new/gwf_2.eps}
}
\centerline{
\includegraphics[width = 15cm]{images_new/gwf_3.eps}
}
\caption{(Left column) Conditional mean arising from conventional error model. (Middle column) Conditional mean arising from enhanced error model. (Right column) Conditional mean arising from iterative error model, iteration 10. From top to bottom, observational noise standard deviation is $10^{-2}$, $10^{-3}$, $10^{-4}$ respectively.}
\label{fig:gwf_means}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[width=5.5cm]{images_new/gwf_conv_1.eps}
\includegraphics[width=5.5cm]{images_new/gwf_conv_2.eps}
\includegraphics[width=5.5cm]{images_new/gwf_conv_3.eps}
}
\caption{Convergence of the error $\|u_\ell - u^\dagger\|$ between the conditional mean and the truth as the number of iterations increases. From left to to right observational noise standard deviation is $10^{-2}$, $10^{-3}$, $10^{-4}$ respectively.}
\label{fig:gwf_conv}
\end{figure}
\section{Conclusions}
\label{sec:conc}
Ill-posedness is a characteristic feature of inverse problems, and therefore, special attention needs to be paid to model uncertainties and model discrepancies that manifest themselves as highly correlated noise, deviating the measured data from the value predicted by the forward model. The modeling error is particularly detrimental when the quality of the data is good, and the exogenous noise does not mask the modeling errors that may become the predominant component of the noise. Quantification of the uncertainty due to the modeling errors is therefore an important part of successfully solving
the inverse problem. Modeling error depends on the unknown that is the target of the inverse problem, and therefore, the Bayesian framework provides a natural basis for attacking the problem: the unknown of interest, modeled as a random variable, can be used in a natural way to define the modeling error as a random variable, thus allowing a statistical interpretation of the modeling error. In this article we introduce,
and study the properties of, an iterative method of refining the statistical description of the modeling error as our information about the unknown increases.
From the implementational point of view, two cases in which the refinement of the modeling error distribution can be computed are identified. When the model is linear and the distributions are Gaussian, a fairly straightforward updating strategy of the posterior estimate is found, and convergence of this iteration can be shown. For non-linear inverse problems, a linearized approximate model leads to a tractable iterative algorithm based on particle approximations of the posterior, and as demonstrated in the numerical experiments, the computed point estimates can be very good, significantly improving on estimates which ignore model error. However, as pointed out in the article, the limiting approximate probability density obtained by the iterative algorithm is not identical to the Bayesian posterior density, although it may be close to it. Regarding both the point estimate and the posterior it is important to
recognize that while the approximation error approach does requires a number of
evaluations of the expensive forward model, unlike traditional MCMC algorithms
no rejections occur. Thus the methodology has potential to compute point estimates
more economically than conventional non-Bayesian approaches such as Tikhonov
regularization; and it also holds the potential to produce reasonable posterior
distributions at considerably lower cost than MCMC using the fully accurate
Bayesian posterior. One of the future directions of research is to see how the approximation process proposed in this article can be effectively used to produce an estimate of the true posterior density.
\ack{The work of D Calvetti is partially supported by NSF grant DMS-1522334.
E Somersalo's work is partly supported by the NSF grant DMS-1312424.
The research of AM Stuart was partially supported by the EPSRC programme grant EQUIP, by AFOSR Grant FA9550-17-1-0185 and ONR Grant N00014-17-1-2079. M Dunlop was partially supported by the EPSRC MASDOC Graduate Training Program. Both M Dunlop and AM Stuart
are supported by DARPA funded program Enabling Quantification of Uncertainty in Physical Systems (EQUiPS), contract W911NF-15-2-0121.}
|
\section{Introduction}
Cloud Computing allows customers to build online applications that can cost effectively scale as necessary \cite{cc}. It provides a massive pool of resources for online applications that can be flexibly obtained when needed and then returned back to the pool at a later time. Such resources can be provided to different applications on top of the infrastructure built and maintained by a cloud company. Examples of hosted online applications include video streaming, data storage and sharing, and big data processing. Most companies that provide cloud computing services own multiple datacenters placed in different cities and countries in order to improve availability, reduce end-to-end delays for end users, and provide customized regional services. At the time of writing this paper, Amazon has more than two dozen availability zones each consisting of one or more discrete datacenters \cite{aws}, Microsoft has 22 regions with plans to build 8 more \cite{azure}, and Google relies on more than a dozen datacenters \cite{google}.
Many applications hosted on these datacenters need to transfer data to their peers in other datacenters for the purpose of data replication and synchronization \cite{sync, amoeba}. The aim is to improve fault-tolerance and user quality of service by making multiple copies of data and getting data closer to end users. Most of these transfers have to be completed before a deadline in order to meet customer service level agreements (SLAs) and they can take hours to complete \cite{tempus}. For example, search engines may have to exchange data among datacenters in order to synchronize their search databases and storage applications may need to back up user data over certain periods.
As mentioned in \cite{swan}, inter-datacenter traffic can be categorized into three groups: \textit{interactive} traffic that has to be transmitted as soon as possible since it is in the critical path of user experience, \textit{elastic} traffic that requires timely delivery which can be modeled in the form of a deadline, and \textit{background} traffic which is bandwidth hungry and has less priority than the other two categories of traffic. In this paper, we focus on elastic traffic with user specified deadlines.
Previous work in the context of inter-datacenter traffic scheduling either fails to consider the negative effects of packet reordering caused by multiplexing packets over different paths (AMOEBA \cite{amoeba}) or cannot guarantee that admitted requests will actually complete transmission before the deadlines specified by customers (B4 \cite{b4}, SWAN \cite{swan}, TEMPUS \cite{tempus}). In addition, AMOEBA and TEMPUS which are the state-of-the-art techniques in this area, model the allocation problem as large linear programs (LP), with possibly hundreds of thousands of variables, solving which incurs large memory and CPU overhead and can take a long time.
Avoiding packet reordering allows data to be instantly delivered to applications upon arrival of packets. In addition, inter-datacenter networks have characteristics similar to WAN networks (including asymmetric link delays and large delays for links that connect distant locations) for which multiplexing packets over different paths has been shown to considerably degrade TCP performance \cite{wan-reordering}. Putting out of order packets and segments back in order can be expensive in terms of memory and CPU usage, especially when transmitting at high rates.
As explained in \cite{real-mptcp}, TCP needs to buffer as much as the bandwidth-delay product ($BDP$) of the network in lossless and $2 \times BDP$ in lossy networks to put out of order packets back in order. For high speed inter-datacenter networks with tens of gigabits of speed and tens of milliseconds of latency considerable amount of buffering may be needed. In \cite{juggler}, authors show how Vanilla kernel uses $50\%$ more CPU in presence of severe packet reordering and present Juggler, a reordering resilient network stack designed for low latency datacenter networks which can reduce the extra CPU utilization to $10\%$: still a considerable amount. For higher latency networks, due to large variation of RTT over multiple paths, reordering may further increase CPU utilization.
In \cite{rcd}, we proposed RCD, a technique that speeds up the traffic allocation problem by scheduling transfers close to their deadlines. Through simulations, we showed that RCD speeds up the allocation process by allowing new transfers to be scheduled only considering the residual bandwidth which would result in creation of much smaller LP models. However, we did not discuss the reordering problem and only evaluated our technique for a single link scenario.
In this paper, we propose DCRoute, a fast and efficient routing algorithm which eliminates the need for LP modeling and:
\vspace{0.5em}
\begin{itemize}
\item \textbf{Guarantees that admitted transfers complete prior to specified deadlines.}
\item \textbf{Schedules all packets of each transfer over the same path to avoid packet reordering.}
\item \textbf{Works much faster than other techniques while admitting almost equal traffic to the system.}
\end{itemize}
\vspace{0.5em}
The input to DCRoute is the network topology as well as a list of transfers including their volumes and deadlines which are submitted to DCRoute in the order of arrival. DCRoute assigns a single path to every transfer and generates a transmission schedule which specifies the rate at which every transfer should be sent over their assigned path during current timeslot. In this paper, we verify the performance of DCRoute through long running simulations. In practice, label switching (such as VLAN tagging) can be used to enforce paths and rate-limiting at the end hosts can be used to enforce transmission rates as in SWAN \cite{swan}.
\section{Problem Description}
Figure \ref{problem-setup} shows our problem setup which is comprised of multiple datacenters in different locations managed by a central controller. The datacenters are connected using high speed WAN links. Applications hosted on these datacenters will make transfers that may take hours to complete and have to be finished before specified deadlines. Due to large transfer time, the time it takes for the source datacenter to request a path from the central controller and the time it takes to setup such a path is considered negligible. However, the time it takes for the scheduling algorithm to prepare a traffic schedule depends on the scheduling algorithm which we aim at minimizing. In addition, in order to avoid packet reordering, we would prefer to send all packets of a transfer on a single path.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{datacenter_setup}
\caption{Problem Setup}
\label{problem-setup}
\end{figure}
As in \cite{amoeba, swan}, we assume the timeline is divided into properly sized timeslots over which the transmission rate is constant. Using a slotted timeline allows for schedules with variable transmission rates over time. In our model, an inter-datacenter request $R$ is represented with four parameters $src(R)$, $dst(R)$, $vol(R)$ and $dl(R)$ which stand for source, destination, volume of data to be transferred and the deadline prior to which the transfer has to be completed.
Similar to previous work \cite{amoeba}, we assume arriving requests are put into a queue and processed in the order of arrival and that there is no preemption: once a request is allocated, it cannot be unallocated from the system. At each moment, we have two parameters $t_{now}$ and $t_{end}$ which represent current timeslot and the latest deadline among all active requests, respectively. A request arriving sometime in timeslot $t$ can be allocated starting timeslot $t+1$ since the schedule and transmission rate for current timeslot is already decided upon and broadcast into all datacenters. Also, at any moment $t$, $t_{now}$ is the timeslot that includes $t$ (current timeslot), and $t_{now}+1$ is the next available timeslot for allocation (next timeslot).
Upon arrival of a request, a central controller decides whether it is possible to allocate it considering some criteria that includes the total available bandwidth over future timeslots. If there is not enough room to allocate a request, the request is rejected and can be submitted to the system again later with a new deadline.
A request is considered \textbf{active} if it is accepted into the system and its deadline has not passed yet. Some active requests may take many timeslots to complete transmission. The total unsatisfied demand of an active request is called the residual demand of that request.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{ALAP}
\caption{An example of ``As Late As Possible" allocation}
\label{ALAP}
\end{figure}
\textbf{Allocation Problem:} \textit{Given active requests $R_1$ through $R_n$ with residual demands $D_1$ to $D_n$ ($0 \le D_i, ~ 1 \le i \le n$), is it possible to allocate a new request $R_{n+1}$? If yes, what is a possible schedule?}
An important characteristic of network traffic is that the size of smallest traffic unit (which is a packet) is significantly smaller than link capacities (which are in the range of gigabits nowadays). This allows us to solve the allocation problem by forming a linear program (LP) considering capacity constraints of the network edges as well as demand constraints of requests. The answer will give us a possible allocation if the constructed LP is feasible. Although this solution maybe straightforward, considering the number of active requests, number of links in network graph, and how far we are planning ahead into the future ($t_{end}$), the resulting LP could be large and may take a long time to solve.
One of the ways to speed up this process is to limit the number of possible paths between every pair of nodes \cite{tempus}, for example, using k-shortest paths \cite{amoeba}. While solving the LP, another speedup method is to limit the number of considered active requests based on some criterion \cite{amoeba} such as having a common link with the new request. It is also possible to use customized iterative methods to solve the resulting LP models faster based on the solutions of previous LP models in a way similar to the water filling process \cite{tempus}.
\emph{Our proposition is to avoid building an LP model in the first place by trying to allocate new requests only knowing the residual bandwidth on the edges for different timeslots.}
\section{DCRoute}
DCRoute relies on the following three techniques:
\vspace{0.5em}
\begin{itemize}
\item \textbf{Requests are initially allocated as late as possible (ALAP) \cite{alap}}
\item \textbf{Utilization is maximized by pulling traffic from closest future timeslots into the upcoming timeslot}
\item \textbf{A variant of BFS search is used for path selection}
\end{itemize}
\vspace{0.5em}
Figure \ref{ALAP} provides an example of the ALAP allocation technique. As can be seen, when the first transfer is received the timeline is empty and therefore it is allocated adjacent to its deadline. The second transfer is allocated as close as possible to its deadline. The benefit of this type of scheduling is that requests do not use resources until it is absolutely necessary. This means resources will be available to other requests that currently demand them. Now when the third transfer is submitted, the resources are free and it just grabs as much bandwidth as needed. If we had allocated the first two requests closer to current time we may have had to either reject the third transfer or move the first two transfers ahead freeing resources for the third transfer.
Assume requests $R_1$ through $R_n$ are current active requests and we would like to allocate $R_{n+1}$. For every request $R_i,~ 1 \le i \le n$, we either have $dl(R_i) \le dl(R_{n+1})$ or $dl(R_i) > dl(R_{n+1})$. In the former case, since there is no preemption, there is no way to increase the chance of new request being accepted by shifting the traffic allocation of request $i$ away as it has to be completed before $dl(R_{n+1})$. For the latter case, since we allocated all requests ALAP, the traffic is already shifted out of $R_{n+1}$'s window as much as possible. As a result, it is possible to decide on admission of new request by just looking at the residual bandwidth on the links. For a single link, since all requests use a single shared resource (link capacity) this technique allows us to optimally decide whether a new request can be allocated. For a network, each request is routed on multiple links and there are many ways to schedule requests ALAP. If some link is used by multiple requests that are routed on different edges, how traffic is allocated on the common link can affect multiple other links which will affect the requests that use those links later on. Despite this uncertainty, we will show that using this feature, we can greatly speed up the allocation process.
Using ALAP alone can result in poor utilization as we always push traffic towards future timeslots and leave the current timeslot underutilized. In order to maximize utilization, upon beginning of a new timeslot, the scheduler looks at future timeslots and pulls as much traffic as possible from the closest timeslots in the future to the upcoming timeslot. Pulling from closest timeslot allows the ALAP characteristic of allocation to hold true afterwards: all residual demands will still be allocated as close to their deadlines as possible.
While pulling traffic from future timeslots, a request can span over multiple links and if there is some traffic fully occupying the next timeslot on one of those links, it is impossible to pull traffic back for that request from the future timeslots to the next timeslot. In such cases, we may need to pull traffic from requests that may not be the closest to current timeslot. This can result in an allocation in which some requests are not scheduled ALAP because they can be pushed further into the future. To address this problem, after pulling traffic from future timeslots, we sweep requests allocated on future timeslots and push them forward as much as possible.
Figure \ref{ex} shows an example of this process. There are three different requests all of which having the same deadline. It is not possible to pull back the green request as the link $E1$ is already occupied. Therefore, we have to pull the orange request (PullBack phase). Afterwards, the allocation is not ALAP anymore, so we push the green request towards its deadline (PushForward phase). The final allocation is ALAP and the utilization of upcoming timeslot is maximum.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{DCRoute_ex.pdf}
\caption{An example of improving utilization while keeping the final allocation ALAP}
\label{ex}
\end{figure*}
\subsection{DCRoute Algorithms}
We assume a graph $G(V,E)$ connecting datacenters with $M$ bidirectional links. For simplicity, we also assume all links/edges have equal capacity of $1.0$. Every edge $e$ has a boolean \textit{use} property which identifies whether that edge can be used in the course of routing. Also each node $v$ has $4$ parameters \textit{v}, \textit{r}, \textit{rv} and \textit{rs}. If $p$ is the path on BFS tree from src($R$) ending at $v$, these parameters represent whether $v$ has been visited before, the number of hops to $v$ from source, the load of bottleneck link, and the sum of loads on all edges from source to $v$, respectively. Moreover, we have variables $S_{t,m}$ that represent the total sum of traffic over link $m$ from time $t_{now}+1$ to $t$. Every time a new request is submitted to the system, $t_{end}$ is updated so that it covers all active requests. Finally, we define the \textbf{active window} as the set of all timeslots over all edges from time $t_{now}+1$ to $t_{end}$. Our algorithms only operate on the active window.
\textbf{Allocate($R$):} Algorithm \ref{algo1} is executed upon arrival of a new request $R$ and performs multiple BFS searches. In every search, we calculate multiple costs, remove edges with highest costs, and compare the result with previous steps. While examining different paths, our heuristic finds the path with most preferred characteristics: the total sum of traffic before dl($R$) over the chosen path will be minimal compared to other paths when $R$ gets allocated on that path.
In each round, the path assignment algorithm starts by choosing the path with the least number of hops and calculates the total cost of sending the request on that path. Also, the bottleneck link on that path is identified. If the total cost is less than the best path found in previous steps, this path replaces the best path. Next, all edges with an equal or higher cost than the bottleneck edge are removed from the network. This process continues until there is no path from source to destination. Now, if it is possible to allocate the request on the selected path, we apply the allocation, otherwise, the request is rejected since admitting it might result in many more future requests to be rejected.
\begin{algorithm}
\scriptsize
\caption{}
\label{algo1}
\begin{algorithmic}[1]
\Procedure{Allocate($R$)}{}
\For {$t=t_{end} + 1$ \textbf{to} $R.dl$}
\For {$e \in E$}
\State $S_{t,e} \gets S_{t-1,e}$
\EndFor
\EndFor
\State $t_{end} \gets$ \textbf{max}($\emph{R.dl}, t_{end}$)
\For {$e \in E$}
\State $e.use \gets True$
\EndFor
\For {$v \in V$}
\State \{$v.v$, $v.r$, $v.rv$, $v.rs$\} $\gets$ \{$False$, $0$, $0$, $0$\}
\EndFor
\State $SE \gets$ array of edges sorted by $S_{R.dl,e}$ descending
\State $i \gets 0$, $j \gets 1$, $flag \gets True$, $r_{b} \gets$ BFS\_test($R$)
\While{$r_{b}.r$ \textbf{is not} $-1$}
\If {$(r_{b}.rv > 0)$ \textbf{and} $flag$}
\State $flag \gets False$
\While {$S_{R.dl,SE[i]} > r_{b}.rv$}
\State $SE[i].use \gets False$
\State $i \gets i+1$
\EndWhile
\EndIf
\State $SE[i].use \gets False$, $i \gets i+1$
\State $r_{n} \gets$ BFS\_test($R$)
\If {$r_{n}.r == -1$}
\textbf{break}
\EndIf
\State $\alpha \gets r_{b}.r \times R.vol + r_{b}.rs$
\State $\beta \gets r_{n}.r \times R.vol + r_{n}.rs$
\If {$(\alpha > \beta)$ \textbf{or} $((\alpha == \beta)$ \textbf{and} $(r_{b}.rv > r_{n}.rv))$}
\State $j \gets i+1$, $r_{b} \gets r_{n}$, $flag \gets True$
\EndIf
\EndWhile
\While {$i \ge j$}
\State $i \gets i-1$, $SE[i].use \gets True$
\EndWhile
\State $path \gets$ \textit{edges of the path ending at $R.dst$ on the BFS}
\State \hspace{3.5em} \textit{tree starting at $R.src$}
\If{PathAllocate($path,R,False$)}
\State \textbf{return} PathAllocate($path,R,True$)
\Else
\State \textbf{return} $False$
\EndIf
\EndProcedure
\vspace{1.5em}
\Procedure{BFS\_test($R$)}{}
\For{$v \in V$}
\State \{$v.v$, $v.r$, $v.rv$, $v.rs$\} $\gets$ \{$False$, $0$, $0$, $0$\}
\EndFor
\State $Q \gets$ a queue having $R.src$ as first element
\State $R.src.v \gets True$
\While{$Q$ is not empty}
\State $node \gets$ head of the $Q$ removed
\If {$node == R.dst$}
\State \textbf{return} $\{node.r, node.rv, node.rs\}$
\EndIf
\For{$next \in$ all neighbors of $node$}
\State $e \gets$ edge connecting $node$ to $next$
\If{$e.use == True$ \textbf{and} $next.v == False$}
\State add $next$ to $Q$
\State $next.v \gets True$
\State $next.r \gets next.r$ $+$ $1$
\State $next.rv \gets$ \textbf{max}$(node.rv, S_{R.dl, e})$
\State $next.rs \gets node.rv+S_{R.dl, e}$
\EndIf
\EndFor
\EndWhile
\State \textbf{return} $\{-1, -1, -1\}$
\EndProcedure
\vspace{1.5em}
\Procedure{PathAllocate($path,R,apply$)}{}
\State $vol \gets R.vol$
\For {$t=R.dl$ \textbf{to} $t_{now}+2$ \textbf{step} $-1$}
\State $space \gets vol$
\For {$e \in path$}
\State $space \gets$ \textbf{min}$(space,1-S_{t_{now}+1,e})$
\EndFor
\If {$space > 0$}
\State $vol \gets vol-space$
\If{$apply$}
\For {$e \in path$}
\State add $space$ of $R$ to edge $e$ at time $t$
\For{$t'=t$ \textbf{to} $t_{end}$}
\State $S_{t',e} \gets S_{t',e}+space$
\EndFor
\EndFor
\EndIf
\EndIf
\EndFor
\State \textbf{return} $vol == 0$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\textbf{PullBack():} Algorithm \ref{algo2:1} looks at the timeslots starting $t_{now}+2$ to $t_{end}$ and pulls back traffic to $t_{now}+1$ (next timeslot to be scheduled). When pulling back traffic, all edges on a request's path have to be checked for unused capacity and updated together as we pull traffic back.
\begin{algorithm}
\scriptsize
\caption{}
\label{algo2:1}
\begin{algorithmic}[1]
\Procedure{PullBack()}{}
\For {$t=t_{now}+1$ \textbf{to} $t=t_{end}$}
\For {$e \in E$}
\State $reqs \gets$ all requests allocated on $e$ at $t$
\For{$R \in reqs$}
\State $vol \gets$ how much of $R$ allocated on $e$ at $t$
\State $path \gets$ path assigned to $R$ upon allocation
\For{$e' \in path$}
\State $vol \gets$ \textbf{min}$(vol,1-S_{t_{now}+1,e'})$
\EndFor
\If{$vol > 0$}
\For{$e' \in path$}
\State move $vol$ of $R$ from edge $e'$ at $t$
\State to edge $e'$ at $t_{now}+1$
\For{$t' = t_{now}+1$ \textbf{to} $t' = t$}
\State $S_{t',e'} \gets S_{t',e'}+vol$
\EndFor
\EndFor
\EndIf
\EndFor
\EndFor
\EndFor
\EndProcedure
\end{algorithmic}
\end{algorithm}
\textbf{PushForward():} After pulling some traffic back, it may be possible for some other traffic to be pushed ahead even further. Algorithm \ref{algo3:1} scans all future timeslots starting $t_{now}+2$ and makes sure that all demands are allocated ALAP. If not, it moves as much traffic as possible to the future timeslots until all residual demands are ALAP.
\begin{algorithm}
\scriptsize
\caption{}
\label{algo3:1}
\begin{algorithmic}[1]
\Procedure{PushForward()}{}
\For {$t=t_{now}+2$ \textbf{to} $t=t_{end}$}
\For {$e \in E$}
\State $reqs \gets$ all requests allocated on $e$ at $t$
\For{$R \in reqs$}
\State $vol \gets$ how much of $R$ allocated on $e$ at $t$
\State $path \gets$ path assigned to $R$ upon allocation
\For {$t_2=R.dl$ \textbf{to} $t_2=t+1$ \textbf{step} $-1$}
\State $free \gets$ \textbf{min}$(vol,$ free space on $e$ at $t_2)$
\For{$e' \in path$}
\State $free \gets$ \textbf{min}$(free,$ free space on $e'$ at $t_2)$
\EndFor
\If{$free > 0$}
\For{$e' \in path$}
\State remove $free$ of $R$ from edge $e'$ at $t$
\State add $free$ of $R$ to edge $e'$ at $t_2$
\For{$t' = t$ \textbf{to} $t' = t_2$}
\State $S_{t',e'} \gets S_{t',e'}-free$
\EndFor
\EndFor
\EndIf
\EndFor
\EndFor
\EndFor
\EndFor
\EndProcedure
\end{algorithmic}
\end{algorithm}
\textbf{Walk():} Algorithm \ref{algo4:1} is executed when the allocation for next timeslot is final. This algorithm tells each datacenter of the decided allocation and adjusts request demands accordingly by deducting what is scheduled to be sent from the total demand.
\begin{algorithm}
\scriptsize
\caption{}
\label{algo4:1}
\begin{algorithmic}[1]
\Procedure{Walk()}{}
\State \textbf{Broadcast the schedule for $t_{now}+1$ to all datacenters}
\For{$e \in E$}
\State $D \gets S_{t_{now}+1}$
\For{$t = t_{now}+1$ \textbf{to} $t = t_{end}$}
\State $S_{t, e} \gets S_{t, e} - D$
\EndFor
\EndFor
\State $t_{now} \gets t_{now}+1$
\State $t_{end} \gets$ \textbf{max}$(t_{end}, t_{now}+1)$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\subsection{DCRoute and Multi-Path Routing}
As mentioned earlier, to avoid packet reordering, DCRoute maps every transfer to exactly one path and if there is no single path that can allocate a transfer, it will be rejected. Although this sounds too restrictive, we show in the next section that limiting every transfer to a single path results in $2\%$ less utilization in the worst case.
To use the aggregate bandwidth on multiple paths, one can use Multi-Path TCP (MPTCP) \cite{mptcp} which allows sending traffic from multiple interfaces of a single host. MPTCP is based on multiple sub-flows that behave similar to single TCP flows. While using MPTCP, the overall CPU and memory footprint can increase significantly compared to TCP which is the cost paid for increased bandwidth. However, by carefully choosing which parts of the data goes over what path, it is possible to improve the CPU footprint \cite{real-mptcp}.
To increase network utilization, it is possible to use MPTCP and apply DCRoute to all sub-flows created. To do that, one can divide the total demand of a request over multiple sub-flows and assign the same deadline as the deadline of the original request to all of them. If all sub-flows can be allocated with DCRoute, then the request is accepted.
Deciding on the number of sub-flows and the portion of traffic that goes over each one of them is not a trivial problem. It may be possible to extend DCRoute by ranking the paths found in \textbf{Allocate} procedure and choosing a subset of the best paths. Further discussion of this subject is out of the scope of this paper.
\section{Simulation Results}
In this section, we perform simulations to evaluate the performance of DCRoute. We generate synthetic traffic requests with Poisson arrival and input the traffic to both DCRoute and a few other techniques that can be used for traffic allocation. Two metrics are being measured and compared: \textbf{allocation time} and \textbf{portion of rejected traffic} both of which are desired to be small.
\textbf{Simulation Parameters:} We used the same traffic distributions as described in \cite{amoeba}. Requests arrive with Poisson distribution of rate $\lambda$. Also, total demand of each request $R$ is distributed exponentially with mean $\frac{1}{8}$ proportional to the maximum transmission volume possible prior to $dl(R)$. In addition, the length of requests is exponentially distributed for which we assumed a mean of $10$ timeslots. We performed the simulations over $500$ timeslots.
All simulations were performed on a machine with an Intel Core i$7$-$6700$T CPU and $24$ GBs of memory and the algorithms are coded in Java. To solve linear programs faster, we used Gurobi Optimizer \cite{gurobi} with a free academic license. Gurobi Optimizer can speedup the LP solving process using several techniques including parallel processing. All simulations presented here are performed $3$ times and the average is reported. We compare DCRoute with the following allocation schemes for all of which we used the same objective function as \cite{amoeba}:
\textbf{Global LP:} This technique is the most general and flexible way of allocation which routes traffic over all possible edges. All active requests are considered for all timeslots on all edges creating a potentially large linear program. The solution here gives us a lower bound on traffic rejection rate.
\textbf{K-Shortest Paths:} Same as Global LP, however, only the K-Shortest Paths between each pair of nodes are considered in routing. The traffic is allocated using a linear program over such paths. We simulated four cases of $K \in \{1, 3, 5, 7\}$. It is obvious that as $K$ increases, the overall rejection rate will decrease as we have higher flexibility for choosing paths and multiplexing traffic.
\textbf{Pseudo-Integer Programming (PIP):} In terms of traffic rejection rate, comparing DCRoute with the previous two techniques is not fair as they allow multiplexing packets on multiple paths. \textbf{The aim of this technique is to find a lower bound on traffic rejection rate when all packets of each request are sent over a single path.} To do so, the general way is to create an integer program involving a list of possible paths (maybe all paths) for the new request and fixed paths for requests already allocated. The resulting model would be a non-linear integer program which cannot be solved using standard optimization libraries available. We instead created a number of linear programs each assigning one of the possible K-Shortest Paths for the newly arriving request. We then compare the objective values manually and choose the best possible path. In our implementation, we chose $K = 20$. This $K$ seems to be more than necessary as we saw negligible improvement in traffic rejection rate even when increasing $K$ from $5$ to $7$. Using PIP, the path over which a request is transferred is decided upon admission and does not change afterwards. We implemented two versions of this scheme:
\begin{itemize}
\item \textbf{Pure Minimum Cost (PMC):} We choose the path that results in smallest objective value.
\item \textbf{Shortest Path, Minimum Cost (SPMC):} Amongst all shortest paths that result in a feasible solution and have the least number of hops, we choose the one with smallest objective value.
\end{itemize}
\subsection{Google's GScale Network}
GScale network \cite{b4} comprised of $12$ nodes and $19$ links (at $2013$, they have $15$ datacenters as of $2016$ \cite{google}) is a private network that connects Google data centers. We used the same topology to evaluate DCRoute as well as other allocation schemes. Figure \ref{115} shows the rejection rate of different techniques for different arrival rates from low load ($\lambda=1$) to high load ($\lambda=15$). We have included the schemes that potentially multiplex traffic over multiple paths just to provide a lower bound. Comparing with PMC and SPMC schemes over all arrival rates, DCRoute performs $< 2\%$ worse than the one with minimum rejection rate. Also, compared to all schemes, DCRoute rejects at most $4\%$ more traffic.
Figure \ref{115} shows the relative time to process a request using different schemes. This time is calculated dividing the total time to allocate/adjust all requests over all timeslots by the total number of requests. DCRoute is about $3$ orders of magnitude faster than either PMC or SPMC. It should be noted that the rate at which time complexity grows drops as we move towards higher arrival rates since there is less capacity available for new requests and many arriving requests get rejected by failing simple capacity constraint checks.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{115.pdf}
\caption{Total \% of rejected traffic and relative request processing time for GScale network with 12 nodes and 19 links}
\label{115}
\end{figure*}
\subsection{Variable Network Size}
We simulated different methods against four networks from $5$ to $20$ nodes:
$(N, M) \in \{(5, 7), (10, 17), (15, 27), (20, 37)\}$
In our topology, each node was connected to $3$ or $4$ other nodes at most $2$ hops away. The arrival rate was kept constant at $\lambda=6.0$ for all cases.
Figure \ref{520} shows the rejection rate of different schemes for different network sizes. As network size increases, since $\lambda$ is kept constant, the total cpacity of network increases compared to the total demand of requests. As a result, for a scheme that multiplexes request traffic over different paths, we expect to see decrease in rejection rate. For the K-Shortest Paths case with $K \in \{1,3\}$ we see increase in rejection rate which we think is because these schemes cannot multiplex packets that much. Increasing the network size for these cases can cause more requests to have common links as the network is sparsely connected and create more bottlenecks resulting in a higher rejection rate.
PMC has a high rejection rate for small networks since choosing the minimum cost path might result in selecting longer (more hops) paths that create larger number of bottlenecks due to collision with other requests. Increasing network size, there are more paths to choose from and that results in less bottlenecks and therefore less rejection rate. In contrast, SPMC enforces the selection of paths with smaller number of hops resulting in lower rejection rates for small networks (due to request paths colliding less) and more rejections as network grows due to less diversity of chosen paths.
Compared to these two approaches, DCRoute balances the choice between smaller and longer paths. The assigned path has the least sum of load on the entire path and the least bottleneck load among all such paths. Paths with heavily loaded links and unnecessarily larger number of hops are avoided. As a result, rejection rate compared to \textit{min(PMC, SPMC)} is relatively small ($<3\%$) for all network sizes. Also, as Figure \ref{520} shows, similar to previous simulation, DCRoute is almost three orders of magnitude faster than PIP schemes and more than $200$ times faster than all considered schemes.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{520.pdf}
\caption{Total \% of rejected traffic and relative request processing time for different network sizes}
\label{520}
\end{figure*}
\subsection{Effect of Timeslot Length}
As mentioned earlier, the timeline is divided into timeslots. In this paper, we do not discuss the exact duration of timeslots, however we would like to see how smaller timeslots affect the amount of rejected traffic as well as the speed of different methods. The transmission rate of transfers only changes when moving from one timeslot to the next. Having timeslots that last longer than most requests would essentially result in a fixed transmission rate for such transfers. This causes less critical transfers (with a later deadline) to have to share the bandwidth with more critical transfers (with a closer deadline), reducing the flexibility of the system and increasing the number of rejected transfers. Having very short timeslots on the other hand can result in unnecessary scheduling overhead and practical complications. For example, rate-limiters at the end hosts may not be able to converge to specified rate if they have to change rate too often. Therefore, a proper timeslot length could be a few times smaller than most requests but large enough to avoid unnecessary processing overhead.
As shown in Figure \ref{51}, we divide every timeslot into $2$ to $5$ smaller timeslots. We again use GScale topology and consider $(\lambda = 6)$. We only considered the K-Shortest Paths techniques with $K \in \{3, 7\}$ since during previous tests they provided the best utilization at the least processing cost. It appears that if the timeslots are short enough, making them shorter does not improve the load accepted into the system: DCRoute rejects $2\%$ more traffic compared to K-Shortest Paths schemes for all timeslot resolutions. Although DCRoute is always more than $2$ orders of magnitude faster than K-Shortest Paths schemes, it does have a slightly higher growth rate in processing time as the timeslot resolution increases.
\begin{figure*}
\centering
\includegraphics[width=0.99\textwidth]{timeslot_scale_factor.pdf}
\caption{Total \% of rejected traffic and relative request processing time as timeslot resolution increases}
\label{51}
\end{figure*}
\subsection{Discussion}
We do not directly compare our scheme with any of the previous schemes, such as AMOEBA or TEMPUS, since we pursue the following two objectives together which is not the case for previous work:
\begin{itemize}
\item Avoiding any packet reordering
\item Guaranteeing deadlines for admitted transfers
\end{itemize}
However, since AMOEBA is the most similar to DCRoute, we would like to provide an approximate comparison. We argue that AMOEBA is essentially an extension to the K-Shortest Paths LP technique $(K = 10)$ introduced here. By intelligently avoiding LP modeling for some special cases, AMOEBA provides a speedup of $30$ times compared to 10-Shortest Paths LP for $(\lambda = 8)$ \cite{amoeba}. For the same arrival rate and request volume/deadline distribution, DCRoute performs more than $1000$ times faster than 10-Shortest Paths LP technique. Also in a previous work called RCD \cite{rcd}, we showed how the close to deadline scheduling technique can speed up the traffic allocation by up to $15$ times compared to AMOEBA while resulting in same utilization over a single link. DCRoute is based on RCD coupled with a path selection heuristic that eliminates the need for LP modeling.
\section{Conclusions and Future Work}
In this paper, we proposed DCRoute, a routing algorithm for Inter-Datacenter networks which guarantees that transfers complete before their deadlines and to avoid reordering, schedules all packets of a request on the same path. Inspired by the fact that allocating ALAP allows for new requests to be scheduled considering only residual bandwidth, DCRoute performs much faster than schemes based on LP modeling. It is more than $2$ orders of magnitude faster than all simulated schemes and $3$ orders of magnitude faster than schemes that do not multiplex transfer packets on multiple paths.
We showed that DCRoute admits at most $4\%$ less traffic compared to schemes that multiplex packets on multiple paths (which provide an estimate of lower bound on rejected traffic) and $2\%$ less traffic compared to schemes that schedule all packets of each transfer on the same path.
Finally, we studied the effect of timeslot resolution and found out that making timeslots smaller than necessary does not increase the admitted load and only incurs extra processing costs. In this paper, we evaluated DCRoute with synthetic traffic. Evaluation of DCRoute using real inter-datacenter traffic is suggested as a future work. In addition, studying the effects of link failures and developing methods to properly handle them can be a subject for future work.
\section*{Acknowledgement}
We would like to thank the anonymous reviewers of HiPC whose suggestions and comments helped us improve the quality of this paper.
{\footnotesize \bibliographystyle{unsrt}
|
\section{Introduction}
For a based loop space $\Omega X$, let $\Omega X\{k\}$ denote the homotopy fibre of the $k^{th}$ power map $k:\Omega X\to \Omega X$. In \cite{S Odd} and \cite{S Decomposition}, Selick showed that after localizing at an odd prime $p$, there is a homotopy decomposition $\Omega^2S^{2p+1}\{p\} \simeq \Omega^2S^3\langle 3\rangle \times W_p$, where $S^3\langle 3\rangle$ is the $3$-connected cover of $S^3$ and $W_n$ is the homotopy fibre of the double suspension $E^2:S^{2n-1}\to \Omega^2S^{2n+1}$. Since $\Omega^2S^{2p+1}\{p\}$ is homotopy equivalent to the pointed mapping space $\mathrm{Map}_\ast(P^3(p), S^{2p+1})$ and the degree $p$ map on the Moore space $P^3(p)$ is nullhomotopic, an immediate consequence is that $p$ annihilates the $p$-torsion in $\pi_\ast(S^3)$ when $p$ is odd. In \cite{S Reformulation}, Ravenel's solution to the odd primary Arf--Kervaire invariant problem \cite{Ra} was used to show that, at least for $p \ge 5$, similar decompositions of $\Omega^2S^{2n+1}\{p\}$ are not possible if $n \ne 1$ or $p$.
The $2$-primary analogue of Selick's decomposition, namely that there is a $2$-local homotopy equivalence $\Omega^2S^5\{2\} \simeq \Omega^2S^3\langle 3\rangle \times W_2$, was later proved by Cohen \cite{C 2-primary}. Similarly, since $\Omega^2S^5\{2\}$ is homotopy equivalent to $\mathrm{Map}_\ast(P^3(2), S^5)$ and the degree $4$ map on $P^3(2) \simeq \Sigma\mathbb{R}P^2$ is nullhomotopic, this product decomposition gives a ``geometric" proof of James' classical result that $4$ annihilates the $2$-torsion in $\pi_\ast(S^3)$. Unlike the odd primary case however, for reasons related to the divisibility of the Whitehead square $[\iota_{2n-1},\iota_{2n-1}] \in \pi_{4n-3}(S^{2n-1})$, the fibre of the squaring map on $\Omega^2S^{2n+1}$ admits nontrivial product decompositions for some other values of $n$.
First, in their investigation of the homology of spaces of maps from mod-$2$ Moore spaces to spheres, Campbell, Cohen, Peterson and Selick \cite{CCPS} found that if $2n+1 \ne 3, 5, 9$ or $17$, then $\Omega^2S^{2n+1}\{2\}$ is atomic and hence indecomposable. Following this, it was shown in \cite{CS} that after localization at the prime $2$, there is a homotopy decomposition $\Omega^2S^9\{2\} \simeq BW_2 \times W_4$ and $W_4$ is a retract of $\Omega^3S^{17}\{2\}$. Here $BW_n$ denotes the classifying space of $W_n$ first constructed by Gray \cite{G}. Since $BW_1$ is known to be homotopy equivalent to $\Omega^2S^3\langle 3\rangle$, the pattern suggested by the decompositions of $\Omega^2S^{5}\{2\}$ and $\Omega^2S^{9}\{2\}$ led Cohen and Selick to conjecture that $\Omega^2S^{17}\{2\} \simeq BW_4 \times W_8$. In this note we prove this is true after looping once. (This weaker statement was also conjectured in \cite{C Fibration}.)
\begin{theorem} \label{mainthm}
There is a $2$-local homotopy equivalence $\Omega^3S^{17}\{2\} \simeq W_4\times \Omega W_8$.
\end{theorem}
In addition to the exponent results mentioned above, decompositions of $\Omega^mS^{2n+1}\{p\}$ also give decompositions of homotopy groups of spheres with $\mathbb{Z}/p\mathbb{Z}$ coefficients. Recall that the mod-$p$ homotopy groups of $X$ are defined by $\pi_k(X; \mathbb{Z}/p\mathbb{Z}) = [P^k(p), X]$.
\begin{corollary}
$\pi_k(S^{17}; \mathbb{Z}/2\mathbb{Z}) \cong \pi_{k-4}(W_4) \oplus \pi_{k-3}(W_8)$ for all $k\ge 4$.
\end{corollary}
In section 3 we relate the problem of decomposing $\Omega^2S^{2n+1}\{2\}$ to a problem considered by Mukai and Skopenkov in \cite{MSk} of computing a certain summand in a homotopy group of the mod-$2$ Moore space $P^{2n+1}(2)$---more specifically, the problem of determining when the Whitehead square $[i_{2n}, i_{2n}] \in \pi_{4n-1}(P^{2n+1}(2))$ of the inclusion of the bottom cell $i_{2n}:S^{2n} \to P^{2n+1}(2)$ is divisible by $2$. The indecomposability result for $\Omega^2S^{2n+1}\{2\}$ in \cite{CCPS} (see also \cite{C Course}) was proved by showing that for $n>1$ the existence of a spherical homology class in $H_{4n-3}(\Omega^2S^{2n+1}\{2\})$ imposed by a nontrivial product decomposition implies the existence of an element $\theta \in \pi_{2n-2}^S$ of Kervaire invariant one such that $\theta\eta$ is divisible by $2$, where $\eta$ is the generator of the stable $1$-stem $\pi_1^S$. Such elements are known to exist only for $2n=4, 8$ or $16$. We show that the divisibility of the Whitehead square $[i_{2n},i_{2n}]$ similarly implies the existence of such Kervaire invariant elements to obtain the following.
\begin{theorem}
The Whitehead square $[i_{2n}, i_{2n}] \in \pi_{4n-1}(P^{2n+1}(2))$ is divisible by $2$ if and only if $2n=2,4,8$ or $16$.
\end{theorem}
This will follow from a preliminary result (Proposition \ref{Whitehead products}) equating the divisibility of $[i_{2n}, i_{2n}]$ with the vanishing of a Whitehead product in the mod-$2$ homotopy of the Stiefel manifold $V_{2n+1,2}$, i.e., the unit tangent bundle over $S^{2n}$. It is shown in \cite{S Indecomposability} that there do not exist maps $S^{2n-1}\times P^{2n}(2) \to V_{2n+1,2}$ extending the inclusions of the bottom cell $S^{2n-1}$ and bottom Moore space $P^{2n}(2)$ if $2n \ne 2,4,8$ or $16$. When $2n=2,4$ or $8$, the Whitehead product obstructing an extension is known to vanish for reasons related to Hopf invariant one, leaving only the boundary case $2n=16$ unresolved. We find that the Whitehead product is also trivial in this case.
\section{The decomposition of $\Omega^3S^{17}\{2\}$}
The proof of Theorem \ref{mainthm} will make use of the $2$-primary version of Richter's proof of Gray's conjecture, so we begin by reviewing this conjecture and spelling out some of its consequences. In his construction of a classifying space of the fibre $W_n$ of the double suspension, Gray \cite{G} introduced two $p$-local homotopy fibrations
\[
S^{2n-1} \xrightarrow{\mathmakebox[0.5cm]{E^2}} \Omega^2S^{2n+1} \xrightarrow{\mathmakebox[0.5cm]{\nu}} BW_n
\]
\[
BW_n \xrightarrow{\mathmakebox[0.5cm]{j}} \Omega^2S^{2np+1} \xrightarrow{\mathmakebox[0.5cm]{\phi}} S^{2np-1}
\]
with the property that $j \circ \nu \simeq \Omega H$, where $H: \Omega S^{2n+1} \to \Omega S^{2np+1}$ is the $p^{th}$ James--Hopf invariant. In addition, Gray showed that the composite $BW_n \xrightarrow{j} \Omega^2S^{2np+1} \xrightarrow{p} \Omega^2S^{2np+1}$ is nullhomotopic and conjectured that the composite $\Omega^2S^{2np+1} \xrightarrow{\phi} S^{2np-1} \xrightarrow{E^2} \Omega^2S^{2np+1}$ is homotopic to the $p^{th}$ power map on $\Omega^2S^{2np+1}$. This was recently proved by Richter in \cite{Ri}.
\begin{theorem}[\cite{Ri}] \label{Richter}
For any prime $p$, there is a homotopy fibration \[BW_n \xrightarrow{\mathmakebox[0.5cm]{j}} \Omega^2 S^{2np+1} \xrightarrow{\mathmakebox[0.5cm]{\phi_n}} S^{2np-1}\] such that $E^2 \circ \phi_n \simeq p$.
\end{theorem}
For odd primes, it was shown in \cite{T Anick's} that there is a homotopy fibration $\Omega W_{np} \to BW_n \to \Omega^2S^{2np+1}\{p\}$ based on the fact that a lift $\overline{S}:BW_n \to \Omega^2S^{2np+1}\{p\}$ of $j$ can be chosen to be an $H$-map when $p$ is odd. One consequence of Theorem \ref{Richter} is that this homotopy fibration exists for all primes and can be extended one step to the right by a map $\Omega^2S^{2np+1}\{p\} \to W_{np}$.
\begin{lemma} \label{fibration}
For any prime $p$, there is a homotopy fibration \[BW_n \xrightarrow{\mathmakebox[0.5cm]{}} \Omega^2S^{2np+1}\{p\} \xrightarrow{\mathmakebox[0.5cm]{}} W_{np}.\]
\end{lemma}
\begin{proof}
The homotopy pullback of $\phi_n$ and the fibre inclusion $W_{np}\to S^{2np-1}$ of the double suspension defines a map $\Omega^2S^{2np+1}\{p\} \to W_{np}$ with homotopy fibre $BW_n$, which can be seen by comparing fibres in the homotopy pullback diagram
\begin{equation} \label{diagram2}
\begin{gathered}
\xymatrix{
BW_n \ar[r] \ar@{=}[d] & \Omega^2S^{2np+1}\{p\} \ar[r] \ar[d] & W_{np} \ar[d] \\
BW_n \ar[r]^-j & \Omega^2S^{2np+1} \ar[r]^-{\phi_n} \ar[d]^-p & S^{2np-1} \ar[d]^-{E^2} \\
& \Omega^2S^{2np+1} \ar@{=}[r] & \Omega^2S^{2np+1}.
}
\end{gathered}
\end{equation}
\end{proof}
Looping once, we obtain a homotopy fibration $$W_n \xrightarrow{\mathmakebox[0.5cm]{}} \Omega^3S^{2np+1}\{p\} \xrightarrow{\mathmakebox[0.5cm]{}} \Omega W_{np}$$ which we will show is split when $p=2$ and $n=4$.
We now fix $p=2$ and localize all spaces and maps at the prime $2$. Homology will be taken with mod-$2$ coefficients unless otherwise stated.
The next lemma describes a factorization of the looped second James--Hopf invariant, an odd primary version of which appears in \cite{T Anick's}. By a well-known result due to Barratt, $\Omega H: \Omega^2S^{2n+1} \to \Omega^2S^{4n+1}$ has order $2$ in the group $[\Omega^2S^{2n+1}, \Omega^2S^{4n+1}]$ and hence lifts to a map $\Omega^2S^{2n+1}\to \Omega^2S^{4n+1}\{2\}$. Improving on this, a feature of Richter's construction of the map $\phi_n$ is that the composite $\Omega^2S^{2n+1} \xrightarrow{\Omega H} \Omega^2S^{4n+1} \xrightarrow{\phi_n} S^{4n-1}$ is nullhomotopic \cite[Lemma 4.2]{Ri}. This recovers Gray's fibration $S^{2n-1} \xrightarrow{E^2} \Omega^2S^{2n+1} \xrightarrow{\nu} BW_n$ and the relation $j \circ \nu \simeq \Omega H$ since there then exists a lift $\nu:\Omega^2S^{2n+1} \to BW_n$ making the diagram
\[
\xymatrix{
& BW_n \ar[d]^j \\
\Omega^2S^{2n+1} \ar[r]^-{\Omega H} \ar[ur]^\nu & \Omega^2S^{4n+1}
}
\]
commute up to homotopy. Since $j$ factors through $\Omega^2S^{4n+1}\{2\}$, by composing the lift $\nu$ with the map $BW_n \to \Omega^2S^{4n+1}\{2\}$ we obtain a choice of lift $S:\Omega^2S^{2n+1} \to \Omega^2S^{4n+1}\{2\}$ of the looped James--Hopf invariant. Hence we have the following consequence of Richter's theorem.
\begin{lemma} \label{S}
There is a homotopy commutative diagram
\[
\xymatrix{
\Omega^2S^{2n+1} \ar[d]^\nu \ar[r]^-S & \Omega^2S^{4n+1}\{2\} \ar@{=}[d] \\
BW_n \ar[r] & \Omega^2S^{4n+1}\{2\}
}
\]
where $S$ is a lift of the looped second James--Hopf invariant $\Omega H: \Omega^2S^{2n+1} \to \Omega^2S^{4n+1}$ and the map $BW_n \to \Omega^2S^{4n+1}\{2\}$ has homotopy fibre $\Omega W_{2n}$.
\end{lemma}
The following homological result was proved in \cite{CCPS} and used to obtain the homotopy decompositions of \cite{C 2-primary} and \cite{CS}.
\begin{lemma}[\cite{CCPS}] \label{CCPS}
Let $n\ge 2$ and let $f:X\to \Omega^2S^{2n+1}\{2\}$ be a map which induces an isomorphism on the module of primitives in degrees $2n-2$ and $4n-3$. If the mod-$2$ homology of $X$ is isomorphic to that of $\Omega^2S^{2n+1}\{2\}$ as a coalgebra over the Steenrod algebra, then $f$ is a homology isomorphism.
\end{lemma}
\begin{theorem}
There is a homotopy equivalence $\Omega^3S^{17}\{2\} \simeq W_4\times \Omega W_8$.
\end{theorem}
\begin{proof}
Let $\tau_n$ denote the map $BW_n\to \Omega^2S^{4n+1}\{2\}$ appearing in Lemma \ref{fibration}. By \eqref{diagram2}, $\tau_n$ is a lift of $j$, implying that $\tau_n$ is nonzero in $H_{4n-2}(\;)$ by naturality of the Bockstein since $j$ is nonzero in $H_{4n-1}(\;)$. We can therefore use the maps $\tau_n$ in place of the (potentially different) maps $\sigma_n$ used in \cite{CS} to obtain product decompositions of $\Omega^2S^{4n+1}\{2\}$ for $n=1$ and $2$, the advantage being that $\tau_n$ has fibre $\Omega W_{2n}$. Explicitly, for $n=2$ this is done as follows. By \cite[Corollary 2.1]{CS}, there exists a map $g:\Omega^3S^{17}\{2\} \to \Omega^2S^9\{2\}$ which is nonzero in $H_{13}(\;)$. Letting $\mu$ denote the loop multiplication on $\Omega^2S^9\{2\}$, it follows that the composite
\[
\psi:BW_2\times W_4 \xrightarrow{\mathmakebox[1.5cm]{\tau_2\times (g\circ \Omega \tau_4)}} \Omega^2S^9\{2\}\times\Omega^2S^9\{2\} \xrightarrow{\mathmakebox[0.5cm]{\mu}} \Omega^2S^9\{2\}
\]
induces an isomorphism on the module of primitives in degrees $6$ and $13$. Since $H_\ast(BW_2\times W_4)$ and $H_\ast(\Omega^2S^9\{2\})$ are isomorphic as coalgebras over the Steenrod algebra, the map above is a homology isomorphism by Lemma \ref{CCPS} and hence a homotopy equivalence.
Now the map $\Omega \tau_4$ fits in the homotopy fibration
\[ W_4 \xrightarrow{\mathmakebox[0.5cm]{\Omega \tau_4}} \Omega^3S^{17}\{2\} \xrightarrow{\mathmakebox[0.5cm]{}} \Omega W_8 \]
and has a left homotopy inverse given by $\pi_2 \circ \psi^{-1}\circ g$ where $\psi^{-1}$ is a homotopy inverse of $\psi$ and $\pi_2:BW_2\times W_4 \to W_4$ is the projection onto the second factor. (Alternatively, composing $g:\Omega^3S^{17}\{2\} \to \Omega^2S^{9}\{2\}$ with the map $\Omega^2S^{9}\{2\} \to W_4$ of Lemma \ref{fibration} yields a left homotopy inverse of $\Omega \tau_4$.) It follows that the homotopy fibration above is fibre homotopy equivalent to the trivial fibration $W_4 \times \Omega W_8 \to \Omega W_8$.
\end{proof}
\begin{corollary}
$\pi_k(S^{17}; \mathbb{Z}/2\mathbb{Z}) \cong \pi_{k-4}(W_4) \oplus \pi_{k-3}(W_8)$ for all $k\ge 4$.
\end{corollary}
One consequence of the splitting of the fibration $W_n\to \Omega^3S^{4n+1}\{p\} \to \Omega W_{2n}$ when $n\in \{1,2,4\}$ is a corresponding homotopy decomposition of the fibre of the map $S$ appearing in Lemma \ref{S}. As in \cite{T 2-primary}, we define the space $Y$ and the map $t$ by the homotopy fibration
\[ Y\xrightarrow{\mathmakebox[0.5cm]{t}} \Omega^2S^{2n+1}\xrightarrow{\mathmakebox[0.5cm]{S}} \Omega^2S^{4n+1}\{2\}. \]
This space and its odd primary analogue play a central role in the construction of Anick's fibration in \cite{T 2-primary}, \cite{T Anick's} and the alternative proof given in \cite{T A new proof} of Cohen, Moore and Neisendorfer's determination of the odd primary homotopy exponent of spheres. Unlike at odd primes, the lift $S$ of $\Omega H$ cannot be chosen to be an $H$-map. Nevertheless, the corollary below shows that its fibre has the structure of an $H$-space in cases of Hopf invariant one.
\begin{corollary}\label{Y}
There is a homotopy fibration $S^{2n-1} \xrightarrow{f} Y \xrightarrow{g} \Omega W_{2n}$ with the property that the composite $S^{2n-1} \xrightarrow{f} Y \xrightarrow{t} \Omega^2S^{2n+1}$ is homotopic to the double suspension $E^2$. Moreover, if $n=1,2$ or $4$ then the fibration splits, giving a homotopy equivalence
\[ Y \simeq S^{2n-1} \times \Omega W_{2n}. \]
\end{corollary}
\begin{proof}
By Lemma \ref{S}, the homotopy fibration defining $Y$ fits in a homotopy pullback diagram
\[
\xymatrix{
S^{2n-1} \ar[d]^f \ar@{=}[r] & S^{2n-1} \ar[d]^{E^2} \\
Y \ar[d]^g \ar[r]^-t & \Omega^2S^{2n+1} \ar[d]^\nu \ar[r]^-S & \Omega^2S^{4n+1}\{2\} \ar@{=}[d] \\
\Omega W_{2n} \ar[r] & BW_n \ar[r] & \Omega^2S^{4n+1}\{2\},
}
\]
which proves the first statement. Note that when $n=1,2$ or $4$, the map $\Omega W_{2n}\to BW_n$ is nullhomotopic by Theorem \ref{mainthm}, hence $t$ lifts through the double suspension. Since any choice of a lift $Y\to S^{2n-1}$ is degree one in $H_{2n-1}(\;)$, it also serves as a left homotopy inverse of $f$, which implies the asserted splitting.
\end{proof}
\begin{remark}
The first part of Corollary \ref{Y} and an odd primary version are proved by different means in \cite{T 2-primary} and \cite{T A new proof}, respectively (see Remark 6.2 of \cite{T 2-primary}). At odd primes, there is an analogous splitting for $n=1$:
$$Y \simeq S^1 \times \Omega W_p \simeq S^1 \times \Omega^3 T^{2p^2+1}(p)$$
where $T^{2p^2+1}(p)$ is Anick's space (see \cite{T A case}).
\end{remark}
\section{Relations to Whitehead products in Moore spaces and Stiefel manifolds}
The special homotopy decompositions of $\Omega^3S^{2n+1}\{2\}$ discussed in the previous section are made possible by the existence of special elements in the stable homotopy groups of spheres, namely elements of Arf--Kervaire invariant one $\theta \in \pi_{2n-2}^S$ such that $\theta\eta$ is divisible by $2$. In this section, we give several reformulations of the existence of such elements in terms of mod-$2$ Moore spaces and Stiefel manifolds.
Let $i_{n-1}:S^{n-1} \to P^n(2)$ be the inclusion of the bottom cell and let $j_n:P^n(2) \to P^n(2)$ be the identity map. Similarly, let $i'_{2n-1}:S^{2n-1} \to V_{2n+1,2}$ and $j'_{2n}:P^{2n}(2) \to V_{2n+1,2}$ denote the inclusions of the bottom cell and bottom Moore space, respectively.\footnote{Note that we index these maps by the dimension of their source rather than their target, so the element of $\pi_{4n-1}(P^{2n+1}(2))$ we call $[i_{2n}, i_{2n}]$ is called $[i_{2n+1}, i_{2n+1}]$ in \cite{MSk}.}
\begin{proposition} \label{Whitehead products}
The Whitehead product $[i'_{2n-1}, j'_{2n}] \in \pi_{4n-2}(V_{2n+1,2}; \mathbb{Z}/2\mathbb{Z})$ is trivial if and only if the Whitehead square $[i_{2n},i_{2n}] \in \pi_{4n-1}(P^{2n+1}(2))$ is divisible by $2$.
\end{proposition}
\begin{proof}
Let $\lambda:S^{4n-2}\to P^{2n}(2)$ denote the attaching map of the top cell in $V_{2n+1,2} \simeq P^{2n}(2) \cup_\lambda e^{4n-1}$ and note that $[i'_{2n-1}, j'_{2n}] = j'_{2n}\circ [i_{2n-1}, j_{2n}]$ by naturality of the Whitehead product. The map $[i_{2n-1}, j_{2n}]: P^{4n-2}(2) \to P^{2n}(2)$ is essential since its adjoint is a Samelson product with nontrivial Hurewicz image $[u,v] \in H_{4n-3}(\Omega P^{2n}(2))$, where $H_\ast(\Omega P^{2n}(2))$ is isomorphic as an algebra to the tensor algebra $T(u,v)$ with $|u|=2n-2$ and $|v|=2n-1$ by the Bott--Samelson theorem. Since the homotopy fibre of the inclusion $j'_{2n}:P^{2n}(2)\to V_{2n+1,2}$ has $(4n-2)$-skeleton $S^{4n-2}$ which maps into $P^{2n}(2)$ by the attaching map $\lambda$, it follows that $[i'_{2n-1}, j'_{2n}]$ is trivial if and only if $[i_{2n-1}, j_{2n}]$ is homotopic to the composite
\[ P^{4n-2}(2) \xrightarrow{\mathmakebox[0.5cm]{q}} S^{4n-2} \xrightarrow{\mathmakebox[0.5cm]{\lambda}} P^{2n}(2) \]
where $q$ is the pinch map.
To ease notation let $P^n$ denote the mod-$2$ Moore space $P^n(2)$ and consider the morphism of $EHP$ sequences
\[
\xymatrix{
[S^{4n}, P^{2n+1}]\ar[r]^-H \ar[d]^{q^\ast} & [S^{4n}, \Sigma P^{2n}\wedge P^{2n}]\ar[r]^-P\ar[d]^{q^\ast} & [S^{4n-2}, P^{2n}]\ar[r]^-E\ar[d]^{q^\ast} & [S^{4n-1}, P^{2n+1}] \ar[d]^{q^\ast} \\
[P^{4n}, P^{2n+1}]\ar[r]^-H & [P^{4n}, \Sigma P^{2n}\wedge P^{2n}] \ar[r]^-P & [P^{4n-2}, P^{2n}] \ar[r]^-E & [P^{4n-1}, P^{2n+1}]
}
\]
induced by the pinch map. A homology calculation shows that the $(4n)$-skeleton of $\Sigma P^{2n}\wedge P^{2n}$ is homotopy equivalent to $P^{4n} \vee S^{4n}$. Let $k_1:P^{4n} \to \Sigma P^{2n}\wedge P^{2n}$ and $k_2:S^{4n} \to \Sigma P^{2n}\wedge P^{2n}$ be the composites
\[ P^{4n} \hookrightarrow P^{4n} \vee S^{4n} \simeq \mathrm{sk}_{4n}(\Sigma P^{2n}\wedge P^{2n}) \hookrightarrow \Sigma P^{2n}\wedge P^{2n} \]
and
\[ S^{4n} \hookrightarrow P^{4n} \vee S^{4n} \simeq \mathrm{sk}_{4n}(\Sigma P^{2n}\wedge P^{2n}) \hookrightarrow \Sigma P^{2n}\wedge P^{2n} \]
defined by the left and right wedge summand inclusions, respectively. Then we have that $\pi_{4n}(\Sigma P^{2n}\wedge P^{2n}) = \mathbb{Z}/4\mathbb{Z}\{k_2\}$ and $P(k_2)=\pm 2\lambda$ by \cite[Lemma 12]{M attaching}.
It follows from the universal coefficient exact sequence
\[ 0 \to \pi_{4n}(\Sigma P^{2n}\wedge P^{2n}) \otimes \mathbb{Z}/2\mathbb{Z} \to \pi_{4n}(\Sigma P^{2n}\wedge P^{2n};\mathbb{Z}/2\mathbb{Z}) \to \mathrm{Tor}(\pi_{4n-1}(\Sigma P^{2n}\wedge P^{2n}), \mathbb{Z}/2\mathbb{Z}) \to 0 \]
that
\begin{align*}
\pi_{4n}(\Sigma P^{2n}\wedge P^{2n}; \mathbb{Z}/2\mathbb{Z}) &= [P^{4n}, \Sigma P^{2n}\wedge P^{2n}] \\
&= \mathbb{Z}/2\mathbb{Z}\{k_1\} \oplus \mathbb{Z}/2\mathbb{Z}\{k_2 \circ q\}
\end{align*}
and that the generator $k_2 \circ q$ is in the kernel of $P$ since $P(k_2)=\pm 2\lambda$ implies $$P(k_2 \circ q)=P(q^\ast (k_2))=q^\ast(P(k_2))= \pm \lambda \circ 2 \circ q=0$$ by the commutativity of the above diagram and the fact that $q:P^{4n-2}\to S^{4n-2}$ and $2:S^{4n-2}\to S^{4n-2}$ are consecutive maps in a cofibration sequence. Therefore $[i_{2n-1}, j_{2n}] = P(k_1)$ since the suspension of a Whitehead product is trivial. On the other hand, $\Sigma \lambda$ is homotopic to the composite $S^{4n-1} \xrightarrow{[\iota_{2n}, \iota_{2n}]} S^{2n} \xrightarrow{i_{2n}} P^{2n+1}$ by \cite{M attaching}, which implies $E(\lambda \circ q) = i_{2n} \circ [\iota_{2n}, \iota_{2n}] \circ q = [i_{2n}, i_{2n}] \circ q$ is trivial in $[P^{4n-1}, P^{2n+1}]$ precisely when $[i_{2n}, i_{2n}]$ is divisible by $2$. Hence $[i_{2n}, i_{2n}]$ is divisible by $2$ if and only if $\lambda \circ q = P(k_1) = [i_{2n-1}, j_{2n}] \in [P^{4n-2}, P^{2n}]$, and the proposition follows.
\end{proof}
We use Proposition \ref{Whitehead products} in two ways. First, since the calculation of $\pi_{31}(P^{17}(2))$ in \cite{MSh} shows that $[i_{16}, i_{16}] = 2\widetilde{\sigma}_{16}^2$ for a suitable choice of representative $\widetilde{\sigma}_{16}^2$ of the Toda bracket $\{\sigma_{16}^2, 2\iota_{16}, i_{16}\}$, it follows that the Whitehead product $[i'_{15}, j'_{16}]: P^{30}(2) \to V_{17,2}$ is nullhomotopic and hence there exists a map $S^{15} \times P^{16}(2) \to V_{17,2}$ extending the wedge of skeletal inclusions $S^{15} \vee P^{16}(2) \to V_{17,2}$. This resolves the only case left unsettled by Theorem 3.2 of \cite{S Indecomposability}.
In the other direction, note that such maps $S^{2n-1}\times P^{2n}(2) \to V_{2n+1,2}$ restrict to maps $S^{2n-1}\times S^{2n-1}\to V_{2n+1,2}$ which exist only in cases of Kervaire invariant one by \cite[Proposition 2.27]{W}, so Proposition \ref{Whitehead products} shows that when $2n \ne 2^k$ for some $k \ge1$ the Whitehead square $[i_{2n}, i_{2n}]$ cannot be divisible by $2$ for the same reasons that the Whitehead square $[\iota_{2n-1}, \iota_{2n-1}] \in \pi_{4n-3}(S^{2n-1})$ cannot be divisible by $2$. Moreover, since maps $S^{2n-1}\times P^{2n}(2) \to V_{2n+1,2}$ extending the inclusions of $S^{2n-1}$ and $P^{2n}(2)$ are shown not to exist for $2n>16$ in \cite{S Indecomposability}, Proposition \ref{Whitehead products} implies that the Whitehead square $[i_{2n}, i_{2n}]$ is divisible by $2$ if and only if $2n=2,4,8$ or $16$. In all other cases it generates a $\mathbb{Z}/2\mathbb{Z}$ summand in $\pi_{4n-1}(P^{2n+1}(2))$. This improves on the main theorem of \cite{MSk} which shows by other means that $[i_{2n}, i_{2n}]$ is not divisible by $2$ when $2n$ is not a power of $2$.
These results are summarized in Theorem \ref{TFAE} below. First we recall the following well-known equivalent formulations of the Kervaire invariant problem.
\begin{theorem} [\cite{C Course}, \cite{W}]
The following are equivalent:
\begin{enumerate}[label=(\alph*),ref=(\alph*)]
\item The Whitehead square $[\iota_{2n-1}, \iota_{2n-1}] \in \pi_{4n-3}(S^{2n-1})$ is divisible by $2$;
\item There is a map $P^{4n-2}(2) \to \Omega S^{2n}$ which is nonzero in homology;
\item There exists a space $X$ with mod-$2$ cohomology $\widetilde{H}^i(X)\cong \mathbb{Z}/2\mathbb{Z}$ for $i=2n$, $4n-1$, $4n$ and zero otherwise with $Sq^{2n}:H^{2n}(X) \to H^{4n}(X)$ and $Sq^1:H^{4n-1}(X) \to H^{4n}(X)$ isomorphisms;
\item There exists a map $f:S^{2n-1}\times S^{2n-1} \to V_{2n+1,2}$ such that $f|_{S^{2n-1}\times\ast}=f|_{\ast\times S^{2n-1}}$ is the inclusion of the bottom cell;
\item $n =1$ or there exists an element $\theta \in \pi_{2n-2}^S$ of Kervaire invariant one.
\end{enumerate}
\end{theorem}
The above conditions hold for $2n=2, 4, 8, 16, 32$ and $64$, and the recent solution to the Kervaire invariant problem by Hill, Hopkins and Ravenel \cite{HHR} implies that, with the possible exception of $2n=128$, these are the only values for which the conditions hold. Mimicking the reformulations above we obtain the following.
\begin{theorem} \label{TFAE}
The following are equivalent:
\begin{enumerate}[label=(\alph*),ref=(\alph*)]
\item The Whitehead square $[i_{2n}, i_{2n}] \in \pi_{4n-1}(P^{2n+1}(2))$ is divisible by $2$; \label{statement1}
\item There is a map $P^{4n}(2) \to \Omega P^{2n+2}(2)$ which is nonzero in homology; \label{statement2}
\item There exists a space $X$ with mod-$2$ cohomology $\widetilde{H}^i(X)\cong\mathbb{Z}/2\mathbb{Z}$ for $i=2n+1$, $2n+2$, $4n+1$, $4n+2$ and zero otherwise with $Sq^{2n}: H^{2n+1}(X)\to H^{4n+1}(X)$, $Sq^1:H^{2n+1}(X)\to H^{2n+2}(X)$ and $Sq^1:H^{4n+1}(X)\to H^{4n+2}(X)$ isomorphisms; \label{statement3}
\item There exists a map $f: S^{2n-1}\times P^{2n}(2) \to V_{2n+1,2}$ such that $f|_{S^{2n-1}\times\ast}$ and $f|_{\ast\times P^{2n}(2)}$ are the skeletal inclusions of $S^{2n-1}$ and $P^{2n}(2)$, respectively; \label{statement4}
\item $n =1$ or there exists an element $\theta \in \pi_{2n-2}^S$ of Kervaire invariant one such that $\theta\eta$ is divisible by $2$; \label{statement5}
\item $2n=2, 4, 8$ or $16$. \label{statement6}
\end{enumerate}
\end{theorem}
\begin{proof}
\ref{statement1} is equivalent to \ref{statement2}: In the $n=1$ case, $[\iota_2, \iota_2]=2\eta_2$ implies $[i_2, i_2]=0$, and since $\eta_3 \in \pi_4(S^3)$ has order $2$ its adjoint $\widetilde{\eta}_3:S^3 \to \Omega S^3$ extends to a map $P^4(2) \to \Omega S^3$. If this map desuspended, then $\widetilde{\eta}_3$ would be homotopic to a composite $S^3\to P^4(2)\to S^2 \xrightarrow{E} \Omega S^3$, a contradiction since $\pi_3(S^2)\cong \mathbb{Z}$ implies that any map $S^3 \to S^2$ that factors through $P^4(2)$ is nullhomotopic. Hence the map $P^4(2) \to \Omega S^3$ has nontrivial Hopf invariant in $[P^4(2), \Omega S^5]$ from which it follows that $P^4(2) \to \Omega S^3$ is nonzero in $H_4(\;)$. Composing with the inclusion $\Omega S^3\to \Omega P^4(2)$ gives a map $P^4(2) \to \Omega P^4(2)$ which is nonzero in $H_4(\;)$.
Now suppose $n>1$ and $[i_{2n}, i_{2n}] = 2\alpha$ for some $\alpha \in \pi_{4n-1}(P^{2n+1}(2))$. Then $\Sigma\alpha$ has order $2$ so there is an extension $P^{4n+1}(2) \to P^{2n+2}(2)$ whose adjoint $f: P^{4n}(2)\to \Omega P^{2n+2}(2)$ satisfies $f|_{S^{4n-1}} = E \circ \alpha$. Since $\Omega\Sigma(P^{2n+1}(2) \wedge P^{2n+1}(2))$ has $4n$-skeleton $S^{4n}$, to show that $f_\ast$ is nonzero on $H_{4n}(P^{4n}(2))$ it suffices to show that $H_2 \circ f$ is nontrivial in $[P^{4n}(2), \Omega\Sigma(P^{2n+1}(2) \wedge P^{2n+1}(2))]$ where $H_2: \Omega P^{2n+2}(2) \to \Omega\Sigma(P^{2n+1}(2) \wedge P^{2n+1}(2))$ is the second James--Hopf invariant. If $H_2 \circ f$ is nullhomotopic, then there is a map $g:P^{4n}(2) \to P^{2n+1}(2)$ making the diagram
\[
\xymatrix{
P^{2n+1}(2) \ar[r]^-E & \Omega P^{2n+2}(2) \ar[r]^-{H_2} & \Omega\Sigma(P^{2n+1}(2) \wedge P^{2n+1}(2)) \\
P^{4n}(2) \ar[u]^g \ar[ur]_f
}
\]
commute. But then $\alpha - g|_{S^{4n-1}}$ is in the kernel of $E_\ast: \pi_{4n-1}(P^{2n+1}(2)) \to \pi_{4n}(P^{2n+2}(2))$ which is generated by $[i_{2n}, i_{2n}]$, so $\alpha - g|_{S^{4n-1}}$ is a multiple of $[i_{2n}, i_{2n}]$. Since $[i_{2n}, i_{2n}]$ has order $2$ and clearly $2g|_{S^{4n-1}}=0$, it follows that $[i_{2n}, i_{2n}]=2\alpha=0$, a contradiction. Therefore $f_\ast$ is nonzero on $H_{4n}(P^{4n}(2))$.
Conversely, assume $n>1$ and $f: P^{4n}(2)\to \Omega P^{2n+2}(2)$ is nonzero in $H_{4n}(\;)$. Since the restriction $f|_{S^{4n-1}}$ lifts through the $(4n-1)$-skeleton of $\Omega P^{2n+2}(2)$, there is a homotopy commutative diagram
\[
\xymatrix{
S^{4n-1} \ar[r] \ar[d]^g & P^{4n}(2) \ar[d]^f \\
P^{2n+1}(2) \ar[r]^-E & \Omega P^{2n+2}(2)
}
\]
for some map $g:S^{4n-1} \to P^{2n+1}(2)$. Since $E\circ 2g$ is nullhomotopic, $2g$ is a multiple of $[i_{2n}, i_{2n}]$. But if $2g=0$, then $g$ admits an extension $e:P^{4n}(2) \to P^{2n+1}(2)$ and it follows that $f-E\circ e$ factors through the pinch map $q:P^{4n}(2)\to S^{4n}$. This makes the Pontrjagin square $u^2 \in H_{4n}(\Omega P^{2n+2}(2))$ a spherical homology class, and this is a contradiction which can be seen as follows. If $u^2$ is spherical, then the $4n$-skeleton of $\Omega P^{2n+2}(2)$ is homotopy equivalent to $P^{2n+1}(2) \vee S^{4n}$. On the other hand, it is easy to see that the attaching map of the $4n$-cell in $\Omega P^{2n+2}(2)$ is given by the Whitehead square $[i_{2n}, i_{2n}]$ which is nontrivial as $n>1$, whence $P^{2n+1} \cup_{[i_{2n}, i_{2n}]} e^{4n} \not\simeq P^{2n+1}(2) \vee S^{4n}$.
\ref{statement1} is equivalent to \ref{statement4}: Since the Whitehead product $[i'_{2n-1}, j'_{2n}] \in \pi_{4n-2}(V_{2n+1,2}; \mathbb{Z}/2\mathbb{Z})$ is the obstruction to extending $i'_{2n-1} \vee j'_{2n}:S^{2n-1} \vee P^{2n}(2) \to V_{2n+1,2}$ to $S^{2n-1} \times P^{2n}(2)$, this follows immediately from Proposition \ref{Whitehead products}.
As described in \cite{S Indecomposability}, applying the Hopf construction to a map $f: S^{2n-1}\times P^{2n}(2) \to V_{2n+1,2}$ as in \ref{statement4} yields a map $H(f):P^{4n}(2)\to \Sigma V_{2n+1,2}$ with $Sq^{2n}$ acting nontrivially on $H^{2n}(C_{H(f)})$. Since $\Sigma^2 V_{2n+1,2} \simeq P^{2n+2}(2)\vee S^{4n+1}$, composing the suspension of the Hopf construction $H(f)$ with a retract $\Sigma^2 V_{2n+1,2} \to P^{2n+2}(2)$ defines a map $g:P^{4n+1}(2) \to P^{2n+2}(2)$ with $Sq^{2n}$ acting nontrivially on $H^{2n+1}(C_g)$, so \ref{statement4} implies \ref{statement3}.
By the proof of \cite[Theorem 3.1]{S Indecomposability}, \ref{statement3} implies \ref{statement5}, and \ref{statement5} implies \ref{statement6}.
The triviality of the Whitehead product $[i'_{2n-1}, j'_{2n}] \in \pi_{4n-2}(V_{2n+1,2}; \mathbb{Z}/2\mathbb{Z})$ when $n=1, 2$ or $4$ is implied by \cite[Theorem 2.1]{S Indecomposability}, for example, and Proposition \ref{Whitehead products} implies $[i'_{15}, j'_{16}] \in \pi_{30}(V_{17,2}; \mathbb{Z}/2\mathbb{Z})$ is trivial as well since $[i_{16}, i_{16}] \in \pi_{31}(P^{17}(2))$ is divisible by $2$ by \cite[Lemma 3.10]{MSh}. Thus \ref{statement6} implies \ref{statement4}.
\end{proof}
\section{A loop space decomposition of $J_3(S^2)$}
In this section, we consider some relations between the fibre bundle $S^{4n-1} \to V_{4n+1,2} \to S^{4n}$ defined by projection onto the first vector of an orthonormal $2$-frame in $\mathbb{R}^{4n+1}$ (equivalently, the unit tangent bundle over $S^{4n}$) and the fibration $BW_n \to \Omega^2S^{4n+1}\{2\} \to W_{2n}$ of Lemma \ref{fibration}. Letting $\partial: \Omega S^{4n} \to S^{4n-1}$ denote the connecting map of the first fibration, we will show that there is a morphism of homotopy fibrations
\begin{equation} \label{diagram3}
\begin{gathered}
\xymatrix{
\Omega^2S^{4n} \ar[r]^-{\Omega\partial} \ar[d] & \Omega S^{4n-1} \ar[r] \ar[d] & \Omega V_{4n+1,2} \ar[d] \\
\Omega W_{2n} \ar[r] & BW_n \ar[r] & \Omega^2S^{4n+1}\{2\}
}
\end{gathered}
\end{equation}
from which it will follow that for $n=1,2$ or $4$, $\Omega\partial$ lifts through $\Omega\phi_n:\Omega^3S^{4n+1} \to \Omega S^{4n-1}$. If this lift can be chosen to be $\Omega^2 E$, then it follows that there is a homotopy pullback diagram
\begin{equation} \label{diagram4}
\begin{gathered}
\xymatrix{
\Omega^2V_{4n+1,2} \ar[r] \ar[d] & \Omega^2S^{4n} \ar[r]^{\Omega\partial} \ar[d]^{\Omega^2E} & \Omega S^{4n-1} \ar@{=}[d] \\
W_n \ar[r]^-{\Omega j} \ar[d] & \Omega^3S^{4n+1} \ar[r]^{\Omega\phi_n} \ar[d]^{\Omega^2H} & \Omega S^{4n-1} \\
\Omega^3S^{8n+1} \ar@{=}[r] & \Omega^3S^{8n+1}
}
\end{gathered}
\end{equation}
which identifies $\Omega^2V_{4n+1,2}$ with $\Omega M_3(n)$ where $\{M_k(n)\}_{k\ge 1}$ is the filtration of $BW_n$ studied in \cite{GTZ} beginning with the familiar spaces $M_1(n) \simeq\Omega S^{4n-1}$ and $M_2(n) \simeq S^{4n-1}\{\underline{2}\}$. (Spaces are localized at an odd prime throughout \cite{GTZ} but the construction of the filtration works in the same way for $p=2$.) We verify this (and deloop it) for $n=1$ since it leads to an interesting loop space decomposition which gives isomorphisms $\pi_k(V_{5,2}) \cong \pi_k(J_3(S^2))$ for all $k\ge 3$.
In his factorization of the $4^{th}$-power map on $\Omega^2S^{2n+1}$ through the double suspension, Theriault constructs in \cite{T 2-primary} a space $A$ and a map $\overline{E}: A\to \Omega S^{2n+1}\{2\}$ with the following properties:
\begin{enumerate}[label=(\alph*),ref=(\alph*)]
\item $H_\ast(A) \cong \Lambda(x_{2n-1},x_{2n})$ with Bockstein $\beta x_{2n}=x_{2n-1}$;
\item $\overline{E}$ induces a monomorphism in homology;
\item There is a homotopy fibration $S^{2n-1} \to A \to S^{2n}$ and a homotopy fibration diagram
\[
\xymatrix{
S^{2n-1} \ar[r] \ar[d]^{E^2} & A \ar[r] \ar[d]^{\overline{E}} & S^{2n} \ar[d]^E \\
\Omega^2S^{2n+1} \ar[r] & \Omega S^{2n+1}\{2\} \ar[r] & \Omega S^{2n+1}.
}
\]
\end{enumerate}
Noting that the homology of $A$ is isomorphic to the homology of the unit tangent bundle $\tau(S^{2n})$ as a coalgebra over the Steenrod algebra, Theriault raises the question of whether $A$ is homotopy equivalent to $\tau(S^{2n}) = V_{2n+1,2}$. Our next proposition shows this is true for any space $A$ with the properties above.
\begin{proposition}
There is a homotopy equivalence $A\simeq V_{2n+1,2}$.
\end{proposition}
\begin{proof}
First we show that $A$ splits stably as $P^{2n}\vee S^{4n-1}$. As in \cite{T 2-primary}, let $Y$ denote the $(4n-1)$-skeleton of $\Omega S^{2n+1}\{2\}$. Consider the homotopy fibration
\[ \Omega S^{2n+1}\{2\} \xrightarrow{\mathmakebox[0.5cm]{}} \Omega S^{2n+1} \xrightarrow{\mathmakebox[0.5cm]{2}} \Omega S^{2n+1} \]
and recall that $H_\ast(\Omega S^{2n+1}\{2\}) \cong H_\ast(\Omega S^{2n+1}) \otimes H_\ast(\Omega^2 S^{2n+1})$.
Restricting the fibre inclusion to $Y$ and suspending once we obtain a homotopy commutative diagram
\[
\xymatrix{
& & S^{2n+1} \ar[d] \ar[r]^{\underline{2}} & S^{2n+1} \ar[d] \\
\Sigma Y \ar[urr]^\ell \ar[r] & \Sigma\Omega S^{2n+1}\{2\} \ar[r] & \Sigma\Omega S^{2n+1} \ar[r]^{\Sigma 2} & \Sigma\Omega S^{2n+1}
}
\]
where $\underline{2}$ is the degree $2$ map, the vertical maps are inclusions of the bottom cell of $\Sigma\Omega S^{2n+1}$ and a lift $\ell$ inducing an isomorphism in $H_{2n+1}(\;)$ exists since $\Sigma Y$ is a $4n$-dimensional complex and $\mathrm{sk}_{4n}(\Sigma\Omega S^{2n+1}) = S^{2n+1}$. It follows from the James splitting $\Sigma\Omega S^{2n+1} \simeq \bigvee_{i=1}^\infty S^{2ni+1}$ and the commutativity of the diagram that $\underline{2} \circ \ell$ is nullhomotopic, so in particular $\Sigma\ell$ lifts to the fibre $S^{2n+2}\{\underline{2}\}$ of the degree $2$ map on $S^{2n+2}$. Since $H_\ast(S^{2n+2}\{\underline{2}\}) \cong \mathbb{Z}/2\mathbb{Z}[u_{2n+1}] \otimes \Lambda(v_{2n+2})$ with $\beta v_{2n+2}=u_{2n+1}$, this implies $\Sigma\ell$ factors through a map $r:\Sigma^2Y \to P^{2n+2}(2)$ which is an epimorphism in homology by naturality of the Bockstein, and hence $P^{2n+2}(2)$ is a retract of $\Sigma^2Y$. (Alternatively, $r$ can be obtained by suspending a lift $\Sigma Y \to S^{2n+1}\{\underline{2}\}$ of $\ell$ and using the well-known fact that $\Sigma S^{2n+1}\{\underline{2}\}$ splits as a wedge of Moore spaces.) Now since $\overline{E}:A\to \Omega S^{2n+1}\{2\}$ factors through $Y$ and induces a monomorphism in homology, composing $\Sigma^2A \to \Sigma^2Y$ with the retraction $r$ shows that $\Sigma^2A \simeq \Sigma^2(P^{2n}(2)\vee S^{4n-1})$.
Next, let $E^\infty: A \to QA$ denote the stabilization map and let $F$ denote the homotopy fibre of a map $g: QP^{2n}(2) \to K(\mathbb{Z}/2\mathbb{Z}, 4n-2)$ representing the mod-$2$ cohomology class $u_{2n-1}^2 \in H^{4n-2}(QP^{2n}(2))$. A homology calculation shows that the $(4n-1)$-skeleton of $F$ is a three-cell complex with homology isomorphic to $\Lambda(x_{2n-1},x_{2n})$ as a coalgebra. The splitting $\Sigma^2A \simeq \Sigma^2(P^{2n}(2)\vee S^{4n-1})$ gives rise to a map $\pi_1:QA \simeq QP^{2n}(2) \times QS^{4n-1} \to QP^{2n}(2)$ inducing isomorphisms in $H_{2n-1}(\;)$ and $H_{2n}(\;)$, and since the composite $g \circ \pi_1 \circ E^\infty: A \to K(\mathbb{Z}/2\mathbb{Z}, 4n-2)$ is nullhomotopic, there is a lift $A \to F$ inducing isomorphisms in $H_{2n-1}(\;)$ and $H_{2n}(\;)$. The coalgebra structure of $H_\ast(A)$ then implies this lift is a $(4n-1)$-equivalence and the result follows as $V_{2n+1,2}$ can similarly be seen to be homotopy equivalent to the $(4n-1)$-skeleton of $F$.
\end{proof}
The homotopy commutative diagram \eqref{diagram3} is now obtained by noting that the composite $\Omega S^{4n-1} \xrightarrow{} \Omega V_{4n+1,2} \xrightarrow{\Omega\overline{E}} \Omega^2S^{4n+1}\{2\}$ is homotopic to $\Omega S^{4n-1} \xrightarrow{\Omega E^2} \Omega^3S^{4n+1} \to \Omega^2S^{4n+1}\{2\}$, which in turn is homotopic to a composite $\Omega S^{4n-1} \to BW_n \to \Omega^2S^{4n+1}\{2\}$ since by Theorem \ref{Richter} there is a homotopy fibration diagram
\[
\xymatrix{
\Omega S^{4n-1} \ar[d]^{\Omega E^2} \ar[r] & BW_n \ar[d] \ar[r]^j & \Omega^2S^{4n+1} \ar@{=}[d] \ar[r]^{\phi_n} & S^{4n-1} \ar[d]^{E^2} \\
\Omega^3S^{4n+1} \ar[r] & \Omega^2S^{4n+1}\{2\} \ar[r] & \Omega^2S^{4n+1} \ar[r]^2 & \Omega^2S^{4n+1}.
}
\]
Specializing to the case $n=1$, the proof of Proposition \ref{J_3(S^2)} will show that $\Omega V_{5,2}$ fits in a delooping of diagram \eqref{diagram4} and hence that $\Omega V_{5,2} \simeq M_3(1).$ We will need the following cohomological characterization of $V_{5,2}$.
\begin{lemma} \label{Stiefel}
Let $E$ be the total space of a fibration $S^3 \to E \to S^4$. If $E$ has integral cohomology group $H^4(E; \mathbb{Z})=\mathbb{Z}/2\mathbb{Z}$ and mod-$2$ cohomology ring $H^\ast(E)$ an exterior algebra $\Lambda(u,v)$ with $|u|=3$ and $|v|=4$, then $E$ is homotopy equivalent to the Stiefel manifold $V_{5,2}$.
\end{lemma}
\begin{proof}
As shown in \cite[Theorem 5.8]{W}, the top row of the homotopy pullback diagram
\[
\xymatrix{
X^4 \ar[r] \ar[d] & P^4(2) \ar[r] \ar[d]^q & BS^3 \ar@{=}[d] \\
S^7 \ar[r]^{\nu} & S^4 \ar[r] & BS^3
}
\]
induces a split short exact sequence
\[ 0 \xrightarrow{\mathmakebox[0.5cm]{}} \mathbb{Z}/4\mathbb{Z} \xrightarrow{\mathmakebox[0.5cm]{}} \pi_6(P^4(2)) \xrightarrow{\mathmakebox[0.5cm]{}} \pi_5(S^3) \xrightarrow{\mathmakebox[0.5cm]{}} 0 \]
from which it follows that $\pi_6(P^4(2)) = \mathbb{Z}/4\mathbb{Z}\{\lambda\} \oplus \mathbb{Z}/2\mathbb{Z}\{\widetilde{\eta}_3^2\}$ where $\lambda$ is the attaching map of the top cell of $V_{5,2}$ and $\widetilde{\eta}_3^2$ maps to the generator $\eta_3^2$ of $\pi_5(S^3)$. It follows from the cohomological assumptions that $E \simeq P^4(2)\cup_f e^7$, where $f=a\lambda + b\widetilde{\eta}_3^2$ for some $a \in \mathbb{Z}/4\mathbb{Z}$, $b \in \mathbb{Z}/2\mathbb{Z}$, and that $H_\ast(\Omega E)$ is isomorphic to a polynomial algebra $\mathbb{Z}/2\mathbb{Z}[u_2,v_3]$. Since the looped inclusion $\Omega P^4(2) \to \Omega E$ induces the abelianization map $T(u_2,v_3) \to \mathbb{Z}/2\mathbb{Z}[u_2,v_3]$ in homology, it is easy to see that the adjoint $f':S^5 \to \Omega P^4(2)$ of $f$ has Hurewicz image $[u_2, v_3]=u_2\otimes v_3 + v_3\otimes u_2$ and hence $f$ is not divisible by $2$. Moreover, since $E$ is an $S^3$-fibration over $S^4$, the pinch map $q:P^4(2)\to S^4$ must extend over $E$. This implies the composite $S^6 \xrightarrow{f} P^4(2) \xrightarrow{q} S^4$ is nullhomotopic and therefore $b=0$ by the commutativity of the diagram above. It now follows that $f=\pm \lambda$ which implies $E \simeq V_{5,2}$.
\end{proof}
\begin{proposition} \label{J_3(S^2)}
There is a homotopy fibration
$$V_{5,2} \xrightarrow{\mathmakebox[0.5cm]{}} J_3(S^2) \xrightarrow{\mathmakebox[0.5cm]{}} K(\mathbb{Z}, 2)$$
which is split after looping.
\end{proposition}
\begin{proof}
Let $h$ denote the composite $\Omega S^3\langle 3\rangle \xrightarrow{} \Omega S^3 \xrightarrow{H} \Omega S^5$ and consider the pullback
\[
\xymatrix{
P \ar[r] \ar[d] & S^4 \ar[d]^E \\
\Omega S^3\langle 3\rangle \ar[r]^-h & \Omega S^5.
}
\]
Since $h$ has homotopy fibre $S^3$, so does the map $P\to S^4$. Next, observe that $P$ is the homotopy fibre of the composite $\Omega S^3\langle 3\rangle \xrightarrow{h} \Omega S^5 \xrightarrow{H} \Omega S^9$ and since $\Omega S^9$ is $7$-connected, the inclusion of the $7$-skeleton of $\Omega S^3\langle 3\rangle$ lifts to a map $\mathrm{sk}_7(\Omega S^3\langle 3\rangle) \to P$. Recalling that $H^4(\Omega S^3\langle 3\rangle; \mathbb{Z}) \cong \mathbb{Z}/2\mathbb{Z}$ and $H_\ast(\Omega S^3\langle 3\rangle) \cong \Lambda(u_3) \otimes \mathbb{Z}/2\mathbb{Z}[v_4]$ with generators in degrees $|u_3|=3$ and $|v_4|=4$, it follows that this lift must be a homology isomorphism and hence a homotopy equivalence. So $P$ is homotopy equivalent to the total space of a fibration satisfying the hypotheses of Lemma \ref{Stiefel} and there is a homotopy equivalence $P\simeq V_{5,2}$.
It is well known that the iterated composite of the $p^{th}$ James--Hopf invariant $H^{\circ k}: \Omega S^{2n+1} \to \Omega S^{2np^k+1}$ has homotopy fibre $J_{p^k-1}(S^{2n})$, the $(p^k-1)^{st}$ stage of the James construction on $S^{2n}$. The argument above identifies $V_{5,2}$ with the homotopy fibre of the composite
\[ \Omega S^3\langle 3\rangle \xrightarrow{\mathmakebox[0.5cm]{}} \Omega S^3 \xrightarrow{\mathmakebox[0.5cm]{H}} \Omega S^5 \xrightarrow{\mathmakebox[0.5cm]{H}} \Omega S^9, \]
so there is a homotopy pullback diagram
\[
\xymatrix{
V_{5,2} \ar[r] \ar[d] & J_3(S^2) \ar[r] \ar[d] & K(\mathbb{Z}, 2) \ar@{=}[d] \\
\Omega S^3\langle 3\rangle \ar[r] \ar[d]^{H\circ h} & \Omega S^3 \ar[r] \ar[d]^{H^{\circ 2}} & K(\mathbb{Z}, 2) \\
\Omega S^9 \ar@{=}[r] & \Omega S^9
}
\]
where the maps into $K(\mathbb{Z}, 2)$ represent generators of $H^2(J_3(S^2); \mathbb{Z}) \cong \mathbb{Z}$ and $H^2(\Omega S^3; \mathbb{Z}) \cong \mathbb{Z}$. To see that the homotopy fibration along the top row splits after looping, note that the connecting map $\Omega K(\mathbb{Z}, 2)=S^1 \to V_{5,2}$ is nullhomotopic since $V_{5,2}$ is simply-connected. Therefore the looped projection map $\Omega J_3(S^2) \to S^1$ has a right homotopy inverse producing a splitting $\Omega J_3(S^2) \simeq S^1 \times \Omega V_{5,2}$.
\end{proof}
\begin{corollary}
$\pi_k(J_3(S^2)) \cong \pi_k(V_{5,2})$ for all $k\ge 3$.
\end{corollary}
|
\section{Introduction}
Nonholonomic constraints have been an important part of classical mechanics for few centuries, see, e.g., \cite{Bloch2003} and\cite{ArKoNe1997}. These types of constraints, incorporating the velocity in an essential way, usually arise as an appropriate mathematical idealization of some contact conditions between parts of a mechanical system. For example, the no-slip rolling conditions are in general nonholonomic except for the simplest cases. The nonholonomic rolling condition states that the infinitesimally small point of contact experiences enough total force not to move with respect to the substrate. In other words, the force per area for the rolling constraint must be infinite for the nonholonomic condition to be valid.
Since nonholonomic constraints are mathematical idealizations, there has been considerable interest in deriving nonholonomic systems as a limit of physically realizable physical forces. As an example, we shall point to the monograph \cite{ArKoNe1997} where the nonholonomic equations are discussed from the point of view of limits of viscous forces and parameter limits.
Alternatively, realising the nonholonomic constraint as an appropriate limit of potential forces can be also useful for example for the quantization of nonholonomic systems, as discussed in \cite{BlRo2008}.
In these notes, we consider the effects of introducing noise in the nonholonomic constraints directly, i.e., study the case of \emph{stochastic constraints}. Physically our study explores what would happen if there is a stochastic error in the nonholonomic constraint. For example, a ball or a disk rolling on a surface may have random slippage introducing a mismatch in velocity. Alternatively, that surface may have a small roughness or exhibits microscopic random motions forcing the constraint to have a stochastic part. These cases are analysed and discussed in details in the recent paper by the authors \cite{FGBPu2016}. The main focus of that paper was the consideration of energy conservation, as well as other integrals of motion, under stochastic constraints. The reader is encouraged to consult that paper for the in-depth consideration of the rolling ball case, as well as other particular applications. The present notes should be viewed as supplementary to that paper as we develop in details two types of stochastic deformations of the Suslov problem in Section \ref{sec:Suslov} and study the possibility of extending to the stochastic case the preservation of some integrals of motion such as the Kharlamova or Clebsch-Tisserand integrals, see \cite{Ko1985}.
Stochastic perturbations of Hamiltonian systems have been considered before in
\cite{Ne1967,Bi1981,LaOr2008,BROw2009,Ho2015,GBHo2017}. Closer to our interests here,
the stochastic extensions of nonholonomically constrained systems, obeying \emph{exact} nonholonomic constraints, was studied in \cite{Ho2010,HoRa2015}. In that work, the focus was on physical systems that are sensitive to noise, while preserving the nature of the nonholonomic constraints exactly. Our paper considers, in some sense, the opposite problem of intrinsically (and somewhat loosely speaking) non-noisy system with noisy constraints. Of course, the stochasticity of constraints leads to stochasticity of the system itself, so the full system is stochastic. However, the stochasticity is quite different from the one discussed before, as it allows for the preservation of integrals of motion. For example, the energy can be conserved under quite general perturbations of the constraints, as we derive in the general form below in Section \ref{sec:stochastic_constraints}.
\section{The Lagrange-d'Alembert equations in nonholonomic mechanics}
\label{sec:LdA}
In this section, we quickly recall the Lagrange-d'Alembert principle in nonholonomic mechanics and two important classes of nonholonomic systems on Lie groups with symmetry.
\subsection{Lagrange-d'Alembert principle}
Consider a mechanical system with configuration manifold $Q$ and Lagrangian $L= L( q , \dot q): TQ \rightarrow \mathbb{R} $ defined on the tangent bundle $TQ$ of the manifold $Q$. Let us assume that the mechanical system is subject to a linear constraint on velocity, encoded in a nonintegrable distribution $ \Delta \subset TQ$ on $Q$. The equations of motion for the nonholonomic system can be derived from the \textit{Lagrange-d'Alembert principle}
\begin{equation}\label{LdA}
\left.\frac{d}{d\varepsilon}\right|_{\varepsilon=0} \int_{0 }^{ T}L(q _\varepsilon (t) , \dot q_ \varepsilon (t) )dt=0,
\end{equation}
where $\dot q(t) \in\Delta $ and for variations $\delta q (t) = \left.\frac{d}{d\varepsilon}\right|_{\varepsilon=0} q _\varepsilon(t)$
of the curve $q(t)$ such that $ \delta q (t) \in \Delta $ and $ \delta q(0)= \delta q(T)=0$. Note that the curves $q_\varepsilon (t)$ do not need to satisfy the constraint when $ \varepsilon \neq 0$.
From \eqref{LdA} one derives the Lagrange-d'Alembert equations
\begin{equation}\label{LdA_equations}
\frac{d}{dt} \frac{\partial L}{\partial \dot q}- \frac{\partial L}{\partial q}\in \Delta(q) ^\circ, \quad \dot q\in \Delta (q),
\end{equation}
where $ \Delta (q) \subset T_qQ$ is the vector fiber of $ \Delta$ at $q \in Q$ and $ \Delta (q) ^\circ \subset T^*_qQ$ is the annihilator of $ \Delta (q)$ in $T^*_q Q$. When there are no constraints, \eqref{LdA_equations} reduces to the Euler-Lagrange equations for the unconstrained system.
\medskip
It is also useful to write \eqref{LdA_equations} explicitly in local coordinates as follows.
Let $ \omega \in \Omega ^1 (Q, \mathbb{R} ^m )$ be a $ \mathbb{R} ^m $-valued one-form such that $ \Delta (q)=\ker ( \omega (q))$. We assume that the $m$ components of $\omega $ are independent. One can choose, in a neighborhood of each point, a local coordinate chart such that the one-form $ \omega $ reads
\begin{equation}\label{local_expression}
\omega ^a (q)= ds ^a + A_ \alpha ^a (r,s)dr ^\alpha , \quad a=1,...,m,
\end{equation}
where $q=(r,s) \in U \subset \mathbb{R} ^{n-m} \times \mathbb{R} ^m $. In these coordinates, equations \eqref{LdA_equations} read
\begin{equation}
\label{LdA_local}
\frac{d}{dt} \frac{\partial L}{\partial \dot r ^\alpha }- \frac{\partial L}{\partial r ^\alpha }= A _\alpha ^a \left( \frac{d}{dt} \frac{\partial L}{\partial \dot s ^a }- \frac{\partial L}{\partial s ^a }\right) , \;\; \alpha =1,...,n-m,\quad \dot s ^a =- A _\alpha ^a \dot r ^\alpha ,\;\; a=1,...,m.
\end{equation}
For later comparisons with the case of stochastic nonholonomic constraints, it is instructive to rewrite the Lagrange-d'Alembert equations in terms of the \textit{constrained Lagrangian}
\[
L_C( r ^\alpha , s ^a , \dot r ^\alpha ):= L( r ^\alpha , s ^a , \dot r ^\alpha , -A_ \alpha ^a (r,s)\dot r ^\alpha ),
\]
which is obtained by substituting the expression for $\dot s^a$ from the constraint in \eqref{LdA_local} into the Lagrangian $L(r,\dot r, s, \dot s)$.
We get
\begin{equation}\label{constrained_L}
\frac{d}{dt}\frac{\partial L_C}{\partial \dot r ^\alpha }- \frac{\partial L_C}{\partial r ^\alpha }+ A _\alpha ^a \frac{\partial L_C}{\partial s ^a }=- \frac{\partial L}{\partial \dot s ^b} B ^b _{ \alpha \beta } \dot r^\beta,
\end{equation}
where we have defined
\begin{equation}\label{def_B}
B ^b _{\alpha \beta } =\frac{\partial A ^b _ \alpha }{\partial r ^\beta }-\frac{\partial A ^b _ \beta }{\partial r ^ \alpha }+ A ^a _\alpha \frac{\partial A ^b _ \beta }{\partial s ^a }- A_\beta ^\alpha \frac{\partial A ^b _ \alpha }{\partial s ^ \alpha }.
\end{equation}
We refer to \cite{Bloch2003} and \cite{BlKrMaMy1996} for a detailed treatment of the Lagrange-d'Alembert equations.
\subsection{Invariant nonholonomic systems on Lie groups}
Let us assume that the configuration manifold is a Lie group $Q=G$ and that the Lagrangian $L:TG \rightarrow \mathbb{R} $ is (left) $G$-invariant. We can thus define the associated reduced Lagrangian $\ell: \mathfrak{g} \rightarrow \mathbb{R} $ on the Lie algebra $ \mathfrak{g}$ of $G$, such that $L(g,v)=\ell( g ^{-1} v)$, for all $g \in G$ and $v \in T_gG$.
We also assume that the distribution $ \Delta \subset TG$ is (left) $G$-invariant. In terms of the one-form $\omega \in \Omega ^1 (G, \mathbb{R} ^k )$, we thus have $\omega (g) \cdot v= \omega (e)\cdot g^{-1} v$, for all $(g,v) \in TG$, and we define the subspace
\[
\mathfrak{g} ^\Delta := \Delta (e)=\{ \xi \in \mathfrak{g} \mid \omega (e) \cdot \xi =0\} \subset \mathfrak{g}.
\]
The Lagrange-d'Alembert equations for $L$ and $ \Delta $ can be equivalently written in terms of $ \ell$ and $ \mathfrak{g} ^ \Delta $ as
\begin{equation}\label{EPS}
\frac{d}{dt} \frac{\delta \ell}{\delta \xi }- \operatorname{ad}^*_ \xi \frac{\delta \ell}{\delta \xi } \in (\mathfrak{g} ^\Delta ) ^\circ, \qquad \xi \in\mathfrak{g} ^\Delta.
\end{equation}
These are the \textit{Euler-Poincar\'e-Suslov equations}, see \cite{Bloch2003,Kozlov1988}. In \eqref{EPS}, $ (\mathfrak{g} ^\Delta ) ^\circ :=\{ \mu \in \mathfrak{g} ^\ast \mid \left\langle \mu , \xi \right\rangle =0, \;\text{for all} \; \xi \in \mathfrak{g} ^\Delta\} \subset \mathfrak{g} ^\ast $ denotes the annihilator of $\mathfrak{g} ^\Delta $ in $\mathfrak{g} ^\ast $, $\frac{\delta \ell}{\delta \xi } \in\mathfrak{g} ^\ast $ defined by
\[
\left\langle \frac{\delta \ell}{\delta \xi }, \eta \right\rangle = \left.\frac{d}{d\varepsilon}\right|_{\varepsilon=0} \ell( \xi + \varepsilon\eta ),
\]
is the functional derivative of $ \ell$, and $ \operatorname{ad}^*_ \xi :\mathfrak{g} ^\ast \rightarrow\mathfrak{g} ^\ast $, defined by $ \left\langle\operatorname{ad}^*_ \xi \mu ,\eta \right\rangle =\left\langle\mu , [\xi , \eta ]\right\rangle $, is the coadjoint operator.
Let us illustrate these concepts for the case of the Lie group $G=SO(3)$ of $3 \times 3$ rotation matrices. Physically, such a group defines the attitude of a rigid body about a fixed point in space, usually assumed to be the center of mass. The Lie algebra of this group, denoted $\mathfrak{so}(3)$, is the algebra of all skew-symmetric matrices, identified with $\mathbb{R}^3$ through the hat map $\;\widehat{\;}:\mathbb{R} ^3\rightarrow \mathfrak{so}(3)$:
\begin{equation}
\widehat{a}_{i j} = - \epsilon_{ijk} a_k , \quad \mathbf{a}=\left(a_1, a_2, a_3 \right)^T \in \mathbb{R}^3 \, .
\label{so3def}
\end{equation}
For this definition, the Lie algebra bracket is related to the cross product as
\begin{equation}
\left[ \widehat{ \mathbf{a} } , \widehat{ \mathbf{b} } \right] = \widehat{ \mathbf{a} \times \mathbf{b} } \, , \quad \forall \, \,
\widehat{ \mathbf{a} } , \widehat{ \mathbf{b} } \in \mathfrak{so}(3) .
\label{widehatproperties}
\end{equation}
Let us now consider the equations of motion for a left-invariant Lagrangian $L:TSO(3) \rightarrow \mathbb{R} $, with associated reduced Lagrangian $\ell: \mathfrak{so}(3) \rightarrow \mathbb{R} $. The reduced velocity (the variable $ \xi $ above) has the physical meaning of the body angular velocity $\widehat{\boldsymbol{\Omega} }=\Lambda^T \dot \Lambda \in \mathfrak{so}(3)$, where $ \Lambda (t) \in SO(3)$ is the attitude of the body. The constraint reads $\boldsymbol{\Omega} \cdot \mathbf{a} =0$ for some constant vector $ \mathbf{a} \in \mathbb{R}^3$. This is the formulation of the classical Suslov problem \cite{Suslov1951}, which is perhaps one of the simplest nonholonomic problems having a non-trivial behavior. The equations of motion \eqref{EPS} are written explicitly as
\begin{equation}
\frac{d }{d t} \frac{ \delta \ell}{ \delta \boldsymbol{\Omega}}+ \boldsymbol{\Omega} \times \frac{ \delta \ell}{ \delta \boldsymbol{\Omega}}= \lambda \mathbf{a} \, , \quad \boldsymbol{\Omega} \cdot \mathbf{a} =0 ,
\label{Suslov_simple}
\end{equation}
where $\lambda$ is a scalar enforcing the constraint. We will study Suslov's problem in considerable more details, and in a more general formulation, in Section \ref{sec:Sus_prob} below.
\color{black}
\subsection{Rolling ball type systems on semidirect products}\label{geom_setting}
Let us consider a slightly more involved situation, namely, the case of a (possibly nonhomegeneous) rolling ball on a flat surface.
In this case the configuration Lie group is given by the semidirect product
\[
SE(3)= SO(3) \,\circledS\, \mathbb{R} ^3 \ni ( \Lambda , \mathbf{x} ),
\]
with group multiplication $( \Lambda_1 , \mathbf{x} _1 )(\Lambda _2 , \mathbf{x} _2 )= ( \Lambda _1\Lambda _2 , \mathbf{x} _1 + \Lambda _1 \mathbf{x} _2 )$.
Here $ \Lambda $ describes the attitude of the ball and $ \mathbf{x} $ the position of its center of mass. As opposed to the previous case, the Lagrangian $L:TSE(3) \rightarrow \mathbb{R} $ and the distribution $ \Delta \subset TSE(3)$ are only invariant under a subgroup of $SE(3)$, namely, the group $S ^1 \,\circledS\, \mathbb{R} ^3$, where $S^1 $ is the group of rotations around the $ \mathbf{z} $-axis. The quotiented velocity space $TSE(3)/ (S ^1 \,\circledS\, \mathbb{R} ^3)$ is identified as follows
\[
TSE(3)/ (S ^1 \,\circledS\, \mathbb{R} ^3)\simeq \mathfrak{se}(3) \times S ^2\ni ( \widehat{\boldsymbol{\Omega} }, \mathbf{Y} ,\boldsymbol{\Gamma} ),
\]
where $\widehat{\boldsymbol{\Omega} }= \Lambda ^T \dot \Lambda $ is the body angular velocity, $ \mathbf{Y} = \Lambda ^{-1} \dot{ \mathbf{x} }$ is the body velocity of the center of mass, and $ \boldsymbol{\Gamma}= \Lambda ^{-1} \mathbf{z} $ is (minus) the direction of gravity measured from the body frame.
The reduced Lagrangian $\ell$ associated to $L$ is now defined as $\ell: \mathfrak{se}(3) \times S ^2 \rightarrow \mathbb{R} $. The reduced constraint associated to $ \Delta \subset TSE(3)$ is a $ \boldsymbol{\Gamma} $-dependent subspace $ \mathfrak{se}(3)^ \Delta ( \boldsymbol{\Gamma} ) \subset \mathfrak{se}(3)$, $\boldsymbol{\Gamma} \in S ^2$. We shall assume the form
\begin{equation}\label{constraint}
\mathfrak{se}(3)^ \Delta ( \boldsymbol{\Gamma} )=\{( \boldsymbol{\Omega} , \mathbf{Y} ) \in \mathfrak{se}(3) \mid \mathbf{Y} = \alpha ( \boldsymbol{\Gamma} ) \cdot\boldsymbol{\Omega} \},
\end{equation}
where $ \alpha (\boldsymbol{\Gamma} ): \mathbb{R} ^3 \rightarrow \mathbb{R}^3 $ is a $\boldsymbol{\Gamma} $-dependent linear map. For an explicit expression of the map $\alpha$ in terms of $\boldsymbol{\Gamma}$ and physical parameters of the rolling ball, and appropriate physical discussion of slipping without rolling, we refer the reader to \cite{Bloch2003,Ho2008} and in the stochastic formulation to \cite{FGBPu2016}.
In terms of $L:T SE(3) \rightarrow \mathbb{R} $ and $ \Delta \subset TSE(3)$, the equations of motion are given by the Lagrange-d'Alembert equations. In terms of the reduced Lagrangian $\ell$ and reduced distribution $ \mathfrak{se}(3) ^\Delta (\boldsymbol{\Gamma} )$, these equations take the form
\begin{equation}\label{reduced_equations}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \frac{d}{dt} \frac{\delta \ell}{\delta \boldsymbol{\Omega} } + \boldsymbol{\Omega} \times \frac{\delta \ell}{\delta\boldsymbol{\Omega} }+ \mathbf{Y}\times \frac{\delta \ell}{\delta \mathbf{Y} } + \alpha( \boldsymbol{\Gamma} )^{T} \cdot \left( \frac{d}{dt} \frac{\delta \ell}{\delta\mathbf{Y} } +\boldsymbol{\Omega} \times \frac{\delta \ell}{\delta \mathbf{Y} }\right) =\frac{\delta\ell}{\delta\boldsymbol{\Gamma} }\times\boldsymbol{\Gamma} \\
\vspace{0.2cm} \displaystyle\frac{d}{dt}\boldsymbol{\Gamma} + \boldsymbol{\Omega} \times\boldsymbol{\Gamma} =0, \quad \mathbf{Y} = \alpha(\boldsymbol{\Gamma} ) \cdot \boldsymbol{\Omega},
\end{array}
\right.
\end{equation}
where $ \frac{\delta \ell}{\delta \boldsymbol{\Omega} }, \frac{\delta \ell}{\delta \mathbf{Y} }, \frac{\delta \ell}{\delta\boldsymbol{\Gamma} } \in \mathbb{R} ^3 $ are the partial derivatives of $\ell$.
We refer to \cite{GBYo2015} for the general theory of rolling ball type systems on semidirect products and for a detailed derivation of the reduced Lagrange-d'Alembert equations \eqref{reduced_equations}, both from the point of view of variational principles and from the point of view of Dirac structures.
\section{Stochastic extensions of nonholonomic constraints}\label{Stoch_LdA}
\label{sec:stochastic_constraints}
Let us fix a configuration manifold $Q$, a Lagrangian function $L: TQ \rightarrow \mathbb{R} $, and a constraint distribution $ \Delta \subset TQ$. Let us assume that the constraint distribution is written as $ \Delta (q)=\{v \in T_qQ\mid \omega (q) \cdot v=0\}$ for a $\mathbb{R} ^m$-valued one-form $ \omega \in\Omega^1(Q, \mathbb{R} ^m )$. We assume that the $m$ components of $\omega $ are independent.
\subsection{Stochastic Lagrange-d'Alembert equations}\label{I_and_II}
We present two ways to introduce stochasticity in the nonholonomic constraint.
\paragraph{I. Affine stochastic constraint.}
The first way of introducing stochasticity in the constraints consists in replacing the linear constraint $\omega (q) \cdot v = 0 $ by an affine constraint
\begin{equation}\label{affine}
\omega(q) \cdot v=N,
\end{equation}
for $N \in \mathbb{R} ^m $.
The stochasticity is inserted in the dynamics through the constraint by assuming that $N(t)$ is a stochastic process verifying the Stratonovich stochastic differential equation
\begin{equation}\label{dN}
\mbox{d}N=F(q,v,N) \mbox{d}t+ \Sigma (q,v,N) \circ \mbox{d}W,
\end{equation}
for given functions $F, \Sigma : TQ \times \mathbb{R} ^m \rightarrow \mathbb{R} ^m $.
We associate to this constraint the stochastic Lagrange-d'Alembert equations
\begin{equation}\label{LdA1}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}\frac{\partial L}{\partial v}- \frac{\partial L}{\partial q} \mbox{d}t \in \Delta (q) ^\circ\\
\vspace{0.2cm}\displaystyle \mbox{d}q=v\mbox{d}t\\
\vspace{0.2cm}\displaystyle \omega (q) \cdot v=N\\
\mbox{d}N= F(q,v, N)\mbox{d}t+ \Sigma (q,v, N) \circ \mbox{d}W,
\end{array}
\right.
\end{equation}
understood in the Stratonovich sense and whose precise meaning is explained below. In general, the energy of the system, defined by
\begin{equation}\label{energy_L}
E(q,v):= \left\langle \frac{\partial L}{\partial v},v \right\rangle -L(q,v),
\end{equation}
is not preserved along the solutions of \eqref{LdA1}.
\paragraph{Stratonovich form.} Taking local coordinates $(r ^\alpha , s ^a )$ such that $ \omega $ takes the expression \eqref{local_expression}, we can rewrite system \eqref{LdA1} as
\begin{equation}\label{LdA1_rs}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}\frac{\partial L}{\partial u ^\alpha }- \frac{\partial L}{\partial r ^\alpha }\mbox{d}t= A _\alpha ^a (r,s) \left( \mbox{d}\frac{\partial L}{\partial w ^a }- \frac{\partial L}{\partial s ^a } \mbox{d}t\right) \\
\vspace{0.2cm}\displaystyle \mbox{d}r ^\alpha = u ^\alpha \mbox{d}t, \;\; \mbox{d}s ^a =w ^a \mbox{d}t\\
\vspace{0.2cm}\displaystyle w ^a + A _\alpha ^a (r,s) u^\alpha = N ^a \\
\mbox{d}N ^a = F ^a (r,s,u,w, N)\mbox{d}t+ \Sigma ^a (r,s,u,w, N) \circ \mbox{d}W.
\end{array}
\right.
\end{equation}
Defining the constrained Lagrangian as
\begin{equation}\label{def_LC1}
\mathcal{L} _C(r ^\alpha ,s ^a ,u ^\alpha , N ^a )= L( r ^\alpha ,s ^a ,u ^\alpha ,N ^a-A _\alpha ^a (r,s) u ^\alpha ),
\end{equation}
similar computations as in the deterministic case, yield the first equation of \eqref{LdA1_rs} in the form
\begin{equation}\label{LC_form1}
\mbox{d}\frac{\partial \mathcal{L} _C}{\partial u ^\alpha }- \frac{\partial \mathcal{L} _C}{\partial r ^\alpha }\mbox{d}t+ A _\alpha ^a \frac{\partial\mathcal{L} _C}{\partial s ^a }\mbox{d}t=- \frac{\partial L}{\partial w^b} \left( B ^b _{ \alpha \beta } w^\beta +\frac{\partial A _\alpha ^b }{\partial s ^a }N ^a \right) \mbox{d}t,
\end{equation}
where $B$ is defined as in \eqref{def_B}.
The difference with the deterministic counterpart \eqref{constrained_L}, lies in two facts. Firstly, the constrained Lagrangian $ \mathcal{L} _C$ (and hence its partial derivatives) depends on the stochastic process $N$ through its definition \eqref{def_LC1}. Secondly, there is the additional last term in \eqref{LC_form1} that explicitly involves $N$.
By developing the term $ \mbox{d}\frac{\partial \mathcal{L} _C}{\partial u ^\alpha }$ using the Stratonovich chain rule, and assuming that $ \frac{\partial ^2 \mathcal{L} _C}{\partial u ^\alpha u ^\beta }$ is invertible, we can rewrite the system \eqref{LdA1_rs} in the standard Stratonovich form
for the stochastic processes $(r,s,u,N)$ as
\begin{equation}\label{LdA1_rs_Stratonovich}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}r ^\alpha = u ^\alpha \mbox{d}t, \quad \mbox{d}s ^a =(N ^a- A _\alpha ^a (r,s) u^\alpha ) \mbox{d}t\\
\vspace{0.2cm}\displaystyle \mbox{d} u ^\alpha = P(r,s,u, N) \mbox{d}t - \left( \frac{\partial ^2 \mathcal{L} _C}{\partial u ^\alpha \partial u ^\beta } \right) ^{-1} \frac{\partial ^2\mathcal{L} _C}{\partial u ^\beta \partial N ^a } \Sigma _C ^a \circ \mbox{d}W\\
\vspace{0.2cm}\displaystyle \mbox{d}N ^a = F ^a _C \mbox{d}t+ \Sigma ^a_C \circ \mbox{d}W,
\end{array}
\right.
\end{equation}
where $F_C$ is defined by $F _C(r^\alpha , s^a ,u ^\alpha , N ^a ):= F ( r ^\alpha ,s ^a ,u ^\alpha ,N ^a-A _\alpha ^a (r,s) u ^\alpha , N ^a )$, similarly for $ \Sigma _C $, and $P$ is a function depending on the variables through the Lagrangian and its derivatives, as well as the functions $A^a_\alpha$. The function $P(r,s,u, N)$ has an explicit expression involving the partial derivatives of $ \mathcal{L} _C$, the partial derivative $ \frac{\partial L}{\partial w ^b }$, and the quantities $A_ \alpha ^b $, $ \frac{\partial A _\alpha ^b }{\partial s ^a }$, $B_{ \alpha \beta } ^b $.
\rem{
\begin{remark}\label{functions_of_noise}{\rm In some examples, like the ones in Section~\ref{sec:shake}, the variable $N$ does not have the meaning of the physical noise experienced by the system. Rather, $N$ is a function of the noise $n$ as well as another variable $q$, i.e., $N= \gamma (q,n)$. Given a Stratonovich SDE $\mbox{d}n= a(q,v,n)\mbox{d}t+ b(q,v,n) \circ \mbox{d}W$ for $n$, one can compute $\mbox{d}N$ by applying the Stratonovich chain rule. To accomplish that, we will need to use the inverted relationship $n=n(N,q)$, assuming this inverse function of $N$ and $q$ exists for $(N,q)$ belonging to the region of interest. For example, for the case of the rolling ball, $\mathbf{N}$ and $\mathbf{n}$ are related by $\mathbf{N}=\Lambda^{-1} \mathbf{n}$, which is invertible for any $\Lambda \in SO(3)$. This will allow us to deduce the corresponding generalizations of \eqref{LdA1}, \eqref{LdA1_rs}, and \eqref{LdA1_rs_Stratonovich}. This exercise is left to the reader.
}
\end{remark}
\paragraph{II. Ideal stochastic constraints.} The second way of altering constraints modifies the one-form $ \omega $ so that it includes a dependency on an additional stochastic variable $\mathsf{N} \in \mathbb{R} ^p$. Let us denote by $\widetilde{\omega} $ this $\mathsf{N}$-dependent one-form on $ Q$ with values in $ \mathbb{R} ^m $. The explicit construction of $\widetilde{\omega}$ from $ \omega $ depends on the problem at hand. We associate the following $\mathsf{N}$-dependent linear constraint to the form $ \widetilde{\omega}$:
\begin{equation}\label{linear_Stoch}
\Delta (q,\mathsf{N}):=\{ v \in T_qQ \mid \widetilde{\omega}(q,\mathsf{N}) \cdot v=0\} \subset T_qQ.
\end{equation}
The stochasticity is inserted in the dynamics through the constraint as before by assuming that $\mathsf{N}(t)$ verifies the Stratonovich stochastic differential equation
\begin{equation}\label{dN_2}
\mbox{d}\mathsf{N}=F(q,v,\mathsf{N}) \mbox{d}t+ \Sigma (q,v,\mathsf{N}) \circ \mbox{d}W,
\end{equation}
for given functions $F, \Sigma : TQ \times \mathbb{R} ^p \rightarrow \mathbb{R} ^p$.
The associated stochastic Lagrange-d'Alembert equations read
\begin{equation}\label{LdA2}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}\frac{\partial L}{\partial v}- \frac{\partial L}{\partial q} \mbox{d}t \in \Delta (q,\mathsf{N}) ^\circ\\
\vspace{0.2cm}\displaystyle \mbox{d}q=v\mbox{d}t\\
\vspace{0.2cm}\displaystyle v \in \Delta (q,\mathsf{N})\\
\mbox{d}\mathsf{N}= F(q,v, \mathsf{N})\mbox{d}t+ \Sigma (q,v, \mathsf{N}) \circ \mbox{d}W,
\end{array}
\right.
\end{equation}
understood in the Stratonovich sense. In this case, one verifies that $\mbox{d} \left( E(q,v) \right) =0$, where $E(q,v)$ is defined in \eqref{energy_L}, meaning that the energy is preserved along the solutions of \eqref{LdA2}. Note that this setting can be easily generalized to the case when the stochastic process $ \mathsf{N}$ takes values in a manifold. See \cite{FGBPu2016} for details and applications to the rolling ball case.
\paragraph{Stratonovich form.} Let us assume the existence of local coordinates $(r ^\alpha , s ^a )$ such that $ \widetilde{\omega}$ takes the expression
\begin{equation}\label{local_expression_2}
\widetilde{\omega}^a (q,\mathsf{N})= ds ^a + \widetilde{A}_ \alpha ^a (r,s,\mathsf{N})dr ^\alpha , \quad a=1,...,m.
\end{equation}
In these coordinates, we can rewrite system \eqref{LdA2} as
\begin{equation}\label{LdA2_rs}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}\frac{\partial L}{\partial u ^\alpha }- \frac{\partial L}{\partial r ^\alpha }\mbox{d}t= \widetilde{A} _\alpha ^a (r,s,\mathsf{N}) \left( \mbox{d}\frac{\partial L}{\partial w ^a }- \frac{\partial L}{\partial s ^a } \mbox{d}t\right) \\
\vspace{0.2cm}\displaystyle \mbox{d}r ^\alpha = u ^\alpha \mbox{d}t, \;\; \mbox{d}s ^a =w ^a \mbox{d}t\\
\vspace{0.2cm}\displaystyle w ^a + \widetilde{A} _\alpha ^a (r,s,\mathsf{N}) u^\alpha =0\\
\mbox{d}\mathsf{N}^j = F ^j (r,s,u,w, \mathsf{N})\mbox{d}t+ \Sigma ^j(r,s,u,w,\mathsf{N}) \circ \mbox{d}W.
\end{array}
\right.
\end{equation}
Defining the constrained Lagrangian as
\begin{equation}\label{def_LC2}
\tilde{\mathcal{L}}_C(r ^\alpha ,s ^a ,u ^\alpha , \mathsf{N} ^j )= L( r ^\alpha ,s ^a ,u ^\alpha ,-\widetilde{A} _\alpha ^a (r,s,\mathsf{N}) u ^\alpha )\,,
\end{equation}
similar computations as in the deterministic case, yield the first equation of \eqref{LdA2_rs} in the form
\begin{equation}\label{LC_form2}
\mbox{d}\frac{\partial \tilde{\mathcal{L}}_C}{\partial u ^\alpha }- \frac{\partial \tilde{\mathcal{L}}_C}{\partial r ^\alpha }\mbox{d}t+ \widetilde{A} _\alpha ^a \frac{\partial \tilde{\mathcal{L}}_C}{\partial s ^a }\mbox{d}t=- \frac{\partial L}{\partial w^b} \left( \widetilde{B} ^b _{ \alpha \beta } w^\beta \mbox{d}t+\frac{\partial \widetilde{A} _\alpha ^b }{\partial \mathsf{N} ^j }\mbox{d} \mathsf{N} ^j \right),
\end{equation}
where $\widetilde{B}(r,s,\mathsf{N})$ is defined in terms of $\widetilde{A}(r,s,\mathsf{N})$ exactly as in \eqref{def_B}.
The difference with the deterministic counterpart \eqref{constrained_L}, lies in two facts. Firstly, the quantities $A _\alpha ^a $ and $B_{ \alpha \beta } ^a $ of \eqref{constrained_L} are replaced by $\widetilde{A}_\alpha ^a$ and $\widetilde{B}_{ \alpha \beta } ^a$ which now depend on $\mathsf{N}$. As a consequence, the constrained Lagrangian $\tilde{\mathcal{L}}_C$ also depends on $\mathsf{N}$ through its dependence on $\widetilde{A}_ \alpha ^a $ in \eqref{def_LC2}. Secondly, there is the additional last term that involves $\mbox{d}\mathsf{N}$.
As above, by developing the term $ \mbox{d}\frac{\partial \tilde{\mathcal{L}}_C}{\partial u ^\alpha }$ using the Stratonovich chain rule, and assuming that $ \frac{\partial ^2 \tilde{\mathcal{L}}_C}{\partial u ^\alpha \partial u ^\beta }$ is invertible, we can rewrite the system \eqref{LdA2_rs} in the standard Stratonovich form
for the stochastic processes $(r,s,u,\mathsf{N})$ as
\begin{equation}\label{LdA2_rs_Stratonovich}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}r ^\alpha = u ^\alpha \mbox{d}t, \quad \mbox{d}s ^a =- \widetilde{A} _\alpha ^a (r,s, \mathsf{N}) u^\alpha \mbox{d}t\\
\vspace{0.2cm}\displaystyle \mbox{d} u ^\alpha =P(r,s,u, \mathsf{N}) \mbox{d}t - \left( \frac{\partial ^2 \tilde{\mathcal{L}}_C}{\partial u ^\alpha \partial u ^\beta } \right) ^{-1} \left( \frac{\partial ^2 \tilde{\mathcal{L}}_C}{\partial u ^\beta \partial \mathsf{N} ^j }+ \frac{\partial L}{\partial w ^b } \frac{\partial \widetilde{A}_ \beta ^b }{\partial \mathsf{N} ^j } \right) \Sigma _C ^ j \circ \mbox{d}W\\
\vspace{0.2cm}\displaystyle \mbox{d}\mathsf{N} ^j = F ^j _C \mbox{d}t+ \Sigma ^j_C \circ \mbox{d}W,
\end{array}
\right.
\end{equation}
where $F_C$ is defined by $F_C(r^\alpha , s^a ,u ^\alpha , \mathsf{N} ^j ):= F ( r ^\alpha ,s ^a ,u ^\alpha ,-\widetilde{A} _\alpha ^a (r,s, \mathsf{N}) u ^\alpha , \mathsf{N} ^j )$, similarly for $\Sigma_C $. The function $P(r,s,u, \mathsf{N})$ has an explicit expression involving the partial derivatives of $\tilde{\mathcal{L}}_C$, the partial derivative $ \frac{\partial L}{\partial w ^b }$, and the quantities $\widetilde{A}_ \alpha ^b $, $\frac{\partial \widetilde{A}_ \alpha ^b}{\partial \mathsf{N} ^j } $, $\widetilde{B}_{ \alpha \beta } ^b $.
\medskip
Expressing the two variants of the stochastic Lagrange-d'Alembert equation in their Stratonovich forms \eqref{LdA1_rs_Stratonovich} and \eqref{LdA2_rs_Stratonovich} helps to identify the role played by the stochastic processes $N$ and $\mathsf{N}$. One observes that in absence of these noisy perturbations of the constraints, both \eqref{LdA1_rs_Stratonovich} and \eqref{LdA2_rs_Stratonovich} coincide and recover the deterministic Lagrange-d'Alembert equations. It would be interesting to explore how the diffusion following from equations \eqref{LdA2_rs_Stratonovich} relate to the nonholonomic diffusion derived in \cite{HoRa2015}, especially in the presence of symmetry. We will postpone the study of this interesting question for future work.
\rem{
We shall show below in \S\ref{subsec_Rolling}, that the equations for the rolling ball presented earlier can be obtained as symmetry reduced versions of either \eqref{LdA1} or \eqref{LdA2}. The reduction process being quite involved, we shall first present here a simpler case of reduction, namely, the case of a $G$-invariant system on the Lie group $G$.
}
\subsection{Stochastic Euler-Poincar\'e-Suslov systems}\label{EPSuslov}
Let us now consider the corresponding two stochastic extensions of the Euler-Poincar\'e-Suslov equations \eqref{EPS}. In the first case, we assume that the functions $F$ and $ \Sigma $ in \eqref{LdA1} are $G$-invariant, so we have $F(g,v,N)=f( g ^{-1} v,N)$ and $ \Sigma (g,v,N)= \sigma (g^{-1} v,N)$, for functions $f, \sigma :\mathfrak{g} \times \mathbb{R} ^m \rightarrow \mathbb{R} ^m $. Using the (left) $G$-invariance of the Lagrangian and the constraints, the stochastic Lagrange-d'Alembert equations \eqref{LdA1} can be rewritten in terms of the reduced quantities as
\begin{equation}\label{Stoch_EPS1}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d} \frac{ \delta \ell}{ \delta \xi } - \operatorname{ad}^*_\xi \frac{ \delta \ell}{ \delta \xi } \mbox{d}t \in (\mathfrak{g} ^\Delta ) ^\circ, \qquad \omega (e) \cdot \xi =N,\\
\mbox{d}N= f( \xi , N)\mbox{d}t+ \sigma ( \xi , N) \circ \mbox{d}W.
\end{array}
\right.
\end{equation}
For the second case, we assume that the $\mathsf{N}$-dependent one-form $\widetilde{ \omega }$ is $G$-invariant, i.e., $\widetilde{ \omega }(g,\mathsf{N}) \cdot v= \widetilde{ \omega }(e,\mathsf{N}) \cdot g^{-1} v$, so we obtain the $\mathsf{N}$-dependent subspace
\[
\mathfrak{g} ^\Delta (\mathsf{N}):=\{ \xi \in \mathfrak{g} \mid \widetilde{ \omega }(e,\mathsf{N}) \cdot\xi =0\} \subset \mathfrak{g}.
\]
The stochastic Lagrange-d'Alembert equations \eqref{LdA2} can be rewritten in terms of the reduced quantities as
\begin{equation}\label{Stoch_EPS2}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d} \frac{ \delta \ell}{ \delta \xi } - \operatorname{ad}^*_\xi \frac{ \delta \ell}{ \delta \xi } \mbox{d}t \in (\mathfrak{g} ^\Delta (\mathsf{N}) ) ^\circ, \qquad \xi \in \mathfrak{g} ^\Delta (\mathsf{N}) ,\\
\mbox{d}\mathsf{N}= \mathsf{f}( \xi , \mathsf{N})\mbox{d}t+ \mathsf{\sigma }( \xi , \mathsf{N}) \circ \mbox{d}W.
\end{array}
\right.
\end{equation}
We shall consider below a more general reduction process that allows us to derive the stochastic equations for the semidirect products, with applications to systems such as the rolling ball. This reduction process, applied to \eqref{LdA1} or \eqref{LdA2}, is a stochastic extension of the nonholonomic reduction process presented in \cite[\S4.3]{GBYo2015} following \cite{Schneider2002}. As in \S\ref{geom_setting}, we will focus on the group $SE(3)$.
\rem{
\begin{remark}[Symmetry breaking due to noise]{\rm As in Remark \ref{functions_of_noise}, one can also consider the more general situation in which the noise $N$ is given as a function $N= \gamma (g,n)$. In this case, the $G$-symmetry of the system may be broken by the presence of the noise and one needs to add to the system the reconstruction equation $\mbox{d}g= \xi g\mbox{d}t$.}
\end{remark}
}
\subsection{Rolling ball type systems with noisy nonholonomic constraints}\label{subsec_Rolling}
We shall now present the reduced version of the stochastic Lagrange-d'Alembert equations \eqref{LdA1} and \eqref{LdA2}, in the context of the geometric setting recalled in \S\ref{geom_setting} and for the ``rolling ball type" constraints considered in \eqref{constraint}. We focus here on the case of the motion on the special Euclidean group $SE(3)$, relevant for physical application of the rolling ball. A reader interested in the results written fully in abstract geometric form should refer to \cite{FGBPu2016}.
We thus fix a Lagrangian $L:TSE(3) \rightarrow \mathbb{R} $ and a constraint distribution $ \Delta \subset TSE(3)$ that are both $S ^1 \,\circledS\, \mathbb{R} ^3 $-invariant, and we consider the associated reduced Lagrangian $\ell: \mathfrak{se}(3) \times S ^2 \rightarrow\mathbb{R} $ and the associated constraint subspaces $ \mathfrak{se}(3) ^ \Delta ( \boldsymbol{\Gamma} ) \subset\mathfrak{se}(3)$.
Following our developments in \S\ref{Stoch_LdA}, we consider
two stochastic deformations of the nonholonomic constraint.
\paragraph{I. Affine stochastic constraint.} We assume that the functions $F, \Sigma :TSE(3) \times\mathbb{R} ^3 \rightarrow\mathbb{R} ^3 $ are $S ^1 \,\circledS\,\mathbb{R} ^3 $-invariant, so that they induce the functions $f, \sigma : \mathfrak{se}(3) \times S ^2 \times\mathbb{R} ^3 \rightarrow \mathbb{R} ^3 $. In the context of constraints of ``rolling ball type" \eqref{constraint}, the reduced version of the stochastic deformation considered in \eqref{affine} and \eqref{dN} thus reads
\begin{equation}\label{noisy_affine}
\mathbf{Y} = \alpha(\boldsymbol{\Gamma} ) \cdot \boldsymbol{\Omega} +\mathbf{N} ,
\end{equation}
where the stochastic process $ \mathbf{N} (t) \in\mathbb{R} ^3 $ verifies the Stratonovich stochastic differential equation
\[
\mbox{d} \mathbf{N} = f ( \boldsymbol{\Omega} ,\mathbf{Y} ,\boldsymbol{\Gamma} , \mathbf{N} ) \mbox{d}t+ \sigma (\boldsymbol{\Omega} ,\mathbf{Y} ,\boldsymbol{\Gamma} , \mathbf{N} ) \circ \mbox{d}W.
\]
Using the (left) $G_{a _0}$-invariance of the Lagrangian and the constraints, the stochastic Lagrange-d'Alembert equations \eqref{LdA1} can be rewritten in terms of the reduced quantities as
\begin{equation}\label{system1}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d} \frac{\delta \ell}{\delta \boldsymbol{\Omega} } + \boldsymbol{\Omega} \times \frac{\delta \ell}{\delta\boldsymbol{\Omega} } \mbox{d}t+ \mathbf{Y}\times \frac{\delta \ell}{\delta \mathbf{Y} } \mbox{d}t+ \alpha( \boldsymbol{\Gamma} )^{T} \cdot \left( \mbox{d} \frac{\delta \ell}{\delta\mathbf{Y} } +\boldsymbol{\Omega} \times \frac{\delta \ell}{\delta \mathbf{Y} } \mbox{d}t\right) =\frac{\delta\ell}{\delta\boldsymbol{\Gamma} }\times\boldsymbol{\Gamma} \mbox{d}t\\
\vspace{0.2cm} \displaystyle \mbox{d}\boldsymbol{\Gamma} + \boldsymbol{\Omega} \times\boldsymbol{\Gamma} \mbox{d}t =0, \quad \mathbf{Y} = \alpha(\boldsymbol{\Gamma} ) \cdot \boldsymbol{\Omega}+ \mathbf{N},
\end{array}
\right.
\end{equation}
understood in the Stratonovich sense. The energy, given in terms of the reduced variables as \begin{equation}\label{reduced_energy}
E( \boldsymbol{\Omega} , \mathbf{Y} , \boldsymbol{\Gamma} )= \frac{\delta \ell}{\delta \boldsymbol{\Omega} } \cdot \boldsymbol{\Omega} +\frac{\delta \ell}{\delta \mathbf{Y} } \cdot \mathbf{Y} - \ell( \boldsymbol{\Omega} , \mathbf{Y} , \boldsymbol{\Gamma} ),
\end{equation}
verifies
\begin{equation}
\label{energy_noncons}
\begin{aligned}
\mbox{d} \left( E( \boldsymbol{\Omega} , \mathbf{Y} , \boldsymbol{\Gamma} ) \right) &=\left( \mbox{d} \frac{\delta \ell}{\delta\mathbf{Y} }+ \boldsymbol{\Omega} \times \frac{\delta \ell}{\delta\mathbf{Y} }\mbox{d}t\right) \cdot \left( \mathbf{Y} - \alpha ( \boldsymbol{\Gamma} ) \cdot \boldsymbol{\Omega}\right) \\
& = \left( \mbox{d} \frac{\delta \ell}{\delta\mathbf{Y} }+ \boldsymbol{\Omega} \times \frac{\delta \ell}{\delta\mathbf{Y} }\mbox{d}t\right)\cdot \mathbf{N}
\end{aligned}
\end{equation}
and is therefore not conserved in general.
\paragraph{II. Ideal stochastic constraints.} Following the second way of introducing the stochasticity in \S\ref{Stoch_LdA}, we assume that the one-form defining $ \Delta $ includes a dependency on an additional variable $\mathsf{N} \in \mathbb{R} ^p$. We also assume that this $\mathsf{N}$-dependent one-form is $S ^1\,\circledS\, \mathbb{R} ^3 $-invariant. Similarly, we assume that the functions $F, \Sigma :TSE(3) \times \mathbb{R} ^p \rightarrow \mathbb{R} ^p $ are $S ^1\,\circledS\, \mathbb{R} ^3 $-invariant, so that they induce the functions $f, \sigma : \mathfrak{se}(3) \times S ^2 \times \mathbb{R} ^p \rightarrow \mathbb{R} ^p $. In the context of constraints of ``rolling ball type" \eqref{constraint}, the reduced version of the stochastic deformation considered in \eqref{linear_Stoch} and \eqref{dN_2} thus reads
\begin{equation}\label{noisy_affine2}
\mathbf{Y} = \widetilde{\alpha}( \boldsymbol{\Gamma} , \mathsf{N}) \cdot\boldsymbol{\Omega} ,
\end{equation}
where the stochastic process $\mathsf{N} (t) \in \mathbb{R} ^p $ verifies the Stratonovich stochastic differential equation
\[
\mbox{d} \mathsf{N}= f ( \boldsymbol{\Omega} ,\mathbf{Y} ,\boldsymbol{\Gamma} , \mathsf{N}) \mbox{d}t+ \sigma ( \boldsymbol{\Omega} , \mathbf{Y} , \boldsymbol{\Gamma} ,\mathsf{N}) \circ \mbox{d}W.
\]
Using the (left) $S ^1 \,\circledS\, \mathbb{R} ^3 $-invariance of the Lagrangian and of the constraints, the stochastic Lagrange-d'Alembert equations \eqref{LdA2} can be rewritten in terms of the reduced quantities as
\begin{equation}\label{system2}
\left\{
\begin{array}{l}
\vspace{0.2cm}\displaystyle \mbox{d}\frac{\delta \ell}{\delta \boldsymbol{\Omega} } + \boldsymbol{\Omega} \times \frac{\delta \ell}{\delta\boldsymbol{\Omega} } \mbox{d}t+ \mathbf{Y}\times \frac{\delta \ell}{\delta \mathbf{Y} } \mbox{d}t+ \widetilde{\alpha}( \boldsymbol{\Gamma} , \mathsf{N})^{T} \cdot \left( \mbox{d} \frac{\delta \ell}{\delta\mathbf{Y} } +\boldsymbol{\Omega} \times \frac{\delta \ell}{\delta \mathbf{Y} } \mbox{d}t\right) =\frac{\delta\ell}{\delta\boldsymbol{\Gamma} }\times\boldsymbol{\Gamma} \mbox{d}t \\
\vspace{0.2cm} \displaystyle \mbox{d} \boldsymbol{\Gamma} + \boldsymbol{\Omega} \times\boldsymbol{\Gamma} \mbox{d}t =0, \quad \mathbf{Y} = \widetilde{\alpha}(\boldsymbol{\Gamma}, \mathsf{N}) \cdot \boldsymbol{\Omega}+ \mathbf{N},
\end{array}
\right.
\end{equation}
understood in the Stratonovich sense.
In this case, the energy $E( \boldsymbol{\Omega} , \mathbf{Y} , \boldsymbol{\Gamma} ) $ defined in \eqref{reduced_energy} is conserved since we can compute
\[
\mbox{d} \left( E( \boldsymbol{\Omega} , \mathbf{Y} , \boldsymbol{\Gamma} ) \right) =\left( \mbox{d} \frac{\delta \ell}{\delta\mathbf{Y} }+ \boldsymbol{\Omega} \times \frac{\delta \ell}{\delta\mathbf{Y} }\mbox{d}t\right) \cdot \left( \mathbf{Y} - \widetilde{\alpha} ( \boldsymbol{\Gamma}, \mathsf{N} ) \cdot \boldsymbol{\Omega}\right) = 0.
\]
\section{Illustration of general theory for the case of Suslov problem}
\label{sec:Suslov}
\rem{
\subsection{Rolling disk}
\color{magenta}
The final example we consider here is that of the vertical rolling disk. This system is characterized by coordinates in the plane $(x,y) \in \mathbb{R}^2$, and two angles of rotation, one along the axis of symmetry of the disk $\theta$, and another orientation angle $\varphi$ defining the axis of the disk's plane with respect to $x$-axis.
The Lagrangian $L: T( \mathbb{R} ^2 \times S ^1 \times S^1 ) \rightarrow \mathbb{R} $, is then defined as
\begin{equation}
L(x,y, \theta , \varphi , \dot x, \dot y , \dot\theta , \dot \varphi )= \frac{1}{2} m( \dot x^2 +\dot y ^2 )+ \frac{1}{2} I \dot \theta^2 + \frac{1}{2} J \dot \varphi^2
\label{disk_lagrangian}
\end{equation}
and the nonholonomic rolling constraints are
\begin{equation}\label{RC}
\dot x= R\dot \theta \cos\varphi, \qquad \dot y = R \dot\theta \sin\varphi \,.
\end{equation}
The Lagrange-d'Alembert principle is formulated as
\[
\delta \int_0^T L(x,y, \theta , \varphi , \dot x, \dot y , \dot\theta , \dot \varphi )dt=0, \quad \delta x= R \delta \theta \cos\varphi \quad \delta y = R \delta \theta \sin\varphi \, ,
\]
giving equations of motion
\[
mR( \ddot x \cos\varphi +\ddot y \sin \varphi )+I \ddot\theta =0, \quad J \ddot \varphi =0
\]
or, equivalently
\[
(m R^2 +I) \ddot \theta =0, \quad J \ddot \varphi =0\,.
\]
These equations can be solved explicitly with $\phi$ and $\theta$ being linear functions of time.
To derive the noisy analogue of \eqref{RC},
for the affine case we choose
\begin{equation}\label{RC_noisy}
\mbox{d} x= R\mbox{d} \theta \cos\varphi + N_1 \quad \mbox{d} y = R \mbox{d} \theta \sin\varphi +N_2
\end{equation}
with
\begin{equation}
\label{dNi_def}
\mbox{d} N_i= f_i(x,y, \theta , \varphi ,N,t) \mbox{d} t+ \sigma_i (x,y, \theta , \varphi,N,t) \circ \mbox{d} W \, .
\end{equation}
The equations of motion are obtained from \eqref{LdA1} as
\begin{equation}
\label{rolling_disk_noisy}
(m R ^2 +I) \mbox{d} \omega + mR ( \mbox{d}N_1\cos \varphi + \mbox{d}N_2 \sin\varphi )=0, \quad \omega \mbox{d} t= \mbox{d}\theta, \quad J \mbox{d} \nu =0, \quad \nu \mbox{d} t= \mbox{d} \varphi.
\end{equation}
The solution is $ \varphi = \nu _0 t+\varphi_0$, so the noise does not affect the dynamics of the orientation angle $ \varphi $ of the rolling disk.
The dynamics of the rotation angle $ \theta $ is, however, affected by the noise. The solution for $\omega(t)$ is obtained from \eqref{rolling_disk_noisy} as
\begin{align}
\omega (t)=& \omega (0) - \frac{mR}{m R ^2 +I} \left[ \int_0^t \cos ( \nu _0 s+\varphi_0) f_1(s) + \sin ( \nu _0 s+\varphi_0) f_2 (s) ) \mbox{d} s \right.
\nonumber
\\
& \qquad + \left.
\int_0^t \left( \cos ( \nu _0 s+\varphi_0) \sigma_1(s) + \sin ( \nu _0 s+\varphi_0) \sigma_2(s) \right)\circ \mbox{d}W(s) \right] \,.
\label{omega_sol_rotation}
\end{align}
The energy $E$ is equal to the Lagrangian defined by \eqref{disk_lagrangian}, $E=L$. As is expected, $E$ is not conserved:
\[
\mbox{d}E= m N_1 {\rm d} \dot x+ m N_2 {\rm d} \dot y = m N _1 \mbox{d} ( R\omega \cos \varphi +N _1)+ m N _2 \mbox{d} ( R\omega \sin \varphi +N _2) \neq 0 \, .
\]
From \eqref{rolling_disk_noisy}, we can also compute the Fokker-Planck equation for the probability density $p(\omega,t)$ as
\begin{equation}
\label{FP_rolling_disk}
\pp{p}{t}=-\pp{}{\omega} \left( F p \right) + \frac{1}{2} \pp{}{\omega} G \pp{ }{\omega} G p \, ,
\end{equation}
where we have defined for shortness
\begin{equation}
\label{FG_def}
F:= f_1 \cos (\nu t +\varphi_0 ) + f_2 \sin (\nu t +\varphi_0 ) \, ,
G:= \sigma_1 \cos (\nu t +\varphi_0 ) + \sigma_2 \sin (\nu t +\varphi_0 ) \, .
\end{equation}
\color{black}
\subsection{Suslov problem}
}
\label{sec:Sus_prob}
The Suslov problem, \cite{Suslov1951} (see also \cite{Kozlov1988,Bloch2003}) considers the motion of a rigid body with a given point (for example, center of mass) fixed in space, subject to a constraint on the angular velocity. We have introduced the classical Suslov problem earlier in \eqref{Suslov_simple} in the general context of motion on $SO(3)$ group with constraints. We remind the reader that if $\Lambda(t) \in SO(3)$ is the attitude of the body in space, and $\widehat{\boldsymbol{\Omega}} (t) =\Lambda (t) ^T \dot \Lambda(t) \in \mathfrak{so}(3)$ is the angular velocity, then the nonholonomic constraint is written in terms of the vector $\mathbf{a}$ that is {\em fixed in the body frame}, as
\[
\mathbf{a} \cdot \boldsymbol{\Omega} =0.
\]
We extend \eqref{Suslov_simple} by incorporating possible effects of gravity or other external potential and take the Lagrangian of the form
\[
\ell( \boldsymbol{\Omega}, \boldsymbol{\Gamma} )= \frac{1}{2} \mathbb{I}\boldsymbol{\Omega} \cdot \boldsymbol{\Omega}-U(\boldsymbol{\Gamma} ),
\]
where $ \boldsymbol{\Gamma} = \Lambda ^{-1} \mathbf{z}$. Then, the Euler-Poincar\'e-Suslov equations \eqref{Suslov_simple} generalize as
\begin{equation}
\label{Suslov_inhomog}
\frac{d}{dt} \mathbb{I} \boldsymbol{\Omega} +\boldsymbol{\Omega}\times \mathbb{I} \boldsymbol{\Omega} =\boldsymbol{\Gamma} \times \frac{\partial U}{\partial\boldsymbol{\Gamma}} + \lambda \mathbf{a}, \quad \frac{d}{dt} \boldsymbol{\Gamma} = \boldsymbol{\Gamma} \times \boldsymbol{\Omega} ,\quad \mathbf{a} \cdot \boldsymbol{\Omega} =0,
\end{equation}
see \cite{Suslov1951}.
The Lagrange multiplier can be solved by taking the scalar product with $ \mathbb{I} ^{-1} \mathbf{a} $, which yields:
\begin{equation}\label{lambda}
\lambda = \frac{ \left( \boldsymbol{\Omega}\times \mathbb{I} \boldsymbol{\Omega} +\displaystyle \frac{\partial U}{\partial\boldsymbol{\Gamma}} \times \boldsymbol{\Gamma} \right) \cdot \mathbb{I} ^{-1} \mathbf{a} }{ \mathbf{a} \cdot\mathbb{I} ^{-1} \mathbf{a} } .
\end{equation}
Integrals of motion are considered in \cite{Ko1985}. Evidently, the equation always has the three independent integrals:
\[
E= \frac{1}{2} \mathbb{I} \boldsymbol{\Omega} \cdot \boldsymbol{\Omega} +U( \boldsymbol{\Gamma} ),\quad \boldsymbol{\Gamma} \cdot \boldsymbol{\Gamma} ,\quad \mathbf{a}\cdot \boldsymbol{\Omega}
\]
(a) If $U=0$, and $ \mathbf{a} $ is an eigenvector of $ \mathbb{I} $, then
\[
\frac{1}{2} \mathbb{I} \boldsymbol{\Omega} \cdot \mathbb{I} \boldsymbol{\Omega}
\]
is an additional integral.
From now on, we suppose that $ \mathbf{a} $ is an eigenvector of $ \mathbb{I} $. Without loss of generality let us take $ \mathbf{a} = \mathbf{e} _3 $.
(b) If $U( \boldsymbol{\Gamma} )= \boldsymbol{\chi} \cdot\boldsymbol{\Gamma} $ with $ \mathbf{a}\cdot \boldsymbol{\chi} =0$, then
\[
\mathbb{I} \boldsymbol{\Omega}\cdot \boldsymbol{\chi}
\]
(Kharlamova) is an additional integral. Indeed, we have $\frac{d}{dt} \mathbb{I} \boldsymbol{\Omega}\cdot \boldsymbol{\chi}= - (\mathbb{I} \boldsymbol{\Omega} \times \boldsymbol{\Omega} ) \cdot \boldsymbol{\chi} =0$
since $\boldsymbol{\Omega}\cdot \mathbf{a} =0$ and $ \boldsymbol{\chi} \cdot \mathbf{a} =0$, see p.158 in \cite{Ko1985}.
(c) If $U( \boldsymbol{\Gamma} )= \boldsymbol{\chi} \cdot\boldsymbol{\Gamma} $ with $\boldsymbol{\chi} = \varepsilon \mathbf{a} $ and if $I_1=I_2$ (Lagrange top), then
\[
\mathbb{I} \boldsymbol{\Omega}\cdot \boldsymbol{\Gamma}
\]
is an additional integral. Indeed, we have $\frac{d}{dt} \mathbb{I} \boldsymbol{\Omega}\cdot \boldsymbol{\Gamma} = \lambda \mathbf{a}\cdot\boldsymbol{\Gamma}$.
However, from \eqref{lambda} we see that $ \lambda =0$ since $I_1=I_2$. This corresponds to the fact that the constraint $\boldsymbol{\Omega} \cdot \mathbf{a} $ is always preserved in a Lagrange top (recall $ \mathbf{a} = \mathbf{e} _3 $).
(d) If $U( \boldsymbol{\Gamma} )= \frac{\varepsilon }{2}\mathbb{I} \boldsymbol{\Gamma} \cdot \boldsymbol{\Gamma} $, then
\[
\frac{1}{2} \mathbb{I} \boldsymbol{\Omega}\cdot \mathbb{I} \boldsymbol{\Omega} - \frac{1}{2} A\boldsymbol{\Gamma}\cdot \boldsymbol{\Gamma} , \quad A: =\varepsilon \mathbb{I} ^{-1}\det\mathbb{I}
\]
(Clebsch-Tisserand integral) is an additional integral (Theorem 3 in \cite{Ko1985}). This is a generalization of situation in (a).
\paragraph{Noisy constraint I.} Let us consider the first stochastic extension introduced in \S\ref{I_and_II}. It corresponds to the stochastic constraint
\begin{equation}
\label{Noisy_constr_1}
\mathbf{a} \cdot \boldsymbol{\Omega} =N, \quad \mbox{d}N= f( \boldsymbol{\Omega} ,\boldsymbol{\Gamma} , N)\mbox{d}t+ \sigma (\boldsymbol{\Omega} ,\boldsymbol{\Gamma} , N) \circ \mbox{d}W,
\end{equation}
for two functions $f, \sigma : \mathbb{R} ^3 \times\mathbb{R} ^3 \times \mathbb{R} \rightarrow \mathbb{R} $.
From the stochastic Euler-Poincar\'e-Suslov equations \eqref{Stoch_EPS1}, we get the system
\begin{equation}\label{case_1}
\left\{
\begin{array}{l}
\displaystyle\vspace{0.2cm}\mbox{d} \mathbb{I} \boldsymbol{\Omega} +\boldsymbol{\Omega}\times \mathbb{I} \boldsymbol{\Omega} \mbox{d}t=\boldsymbol{\Gamma} \times \frac{\partial U}{\partial\boldsymbol{\Gamma}} \mbox{d}t + \lambda \mathbf{a} \mbox{d}t, \quad \mathbf{a} \cdot \boldsymbol{\Omega} =N,\\
\displaystyle\mbox{d}\boldsymbol{\Gamma} = \boldsymbol{\Gamma} \times \boldsymbol{\Omega}\mbox{d}t ,\quad \mbox{d}N=f\mbox{d}t+ \sigma \circ \mbox{d}W,
\end{array}
\right.
\end{equation}
whose Lagrange multiplier is computed as
\[
\lambda \mbox{d}t= \frac{ \left( \boldsymbol{\Omega}\times \mathbb{I} \boldsymbol{\Omega} +\displaystyle \frac{\partial U}{\partial\boldsymbol{\Gamma}} \times \boldsymbol{\Gamma} \right) \cdot \mathbb{I} ^{-1} \mathbf{a}\, \mbox{d}t+ \mbox{d}N}{ \mathbf{a} \cdot\mathbb{I} ^{-1} \mathbf{a} } .
\]
If $N$ is a constant number $c$, then the constraint $ \mathbf{a} \cdot \boldsymbol{\Omega} =c$ corresponds to the inhomogeneous Suslov problem, see \cite{Suslov1951}.
By choosing the eigenvector $\mathbf{a}=\mathbf{e}_3$, equations \eqref{case_1} can be written as
\begin{equation}
\left\{
\begin{array}{l}
\vspace{0.1cm}I _1 \mbox{d} \Omega _1 + \left[ (I_3-I_2) \Omega _2 N -(\boldsymbol{\Gamma}\times \frac{\partial U}{\partial \boldsymbol{\Gamma} } )\cdot \mathbf{e}_1\right] \mbox{d} t=0 \\
\vspace{0.1cm}I _2 \mbox{d} \Omega _2 + \left[ (I_1-I_3) \Omega _1 N -(\boldsymbol{\Gamma}\times \frac{\partial U}{\partial\boldsymbol{\Gamma} } )\cdot \mathbf{e}_2\right] \mbox{d}t=0\\
\vspace{0.1cm}\mbox{d} \boldsymbol{\Gamma} + \boldsymbol{\Omega} \times \boldsymbol{\Gamma} \mbox{d}t =\mathbf{0} \\
\vspace{0.1cm}\mbox{d} N = f( \Omega _1 ,\Omega _2 ,\boldsymbol{\Gamma} ,N) \mbox{d}t+ \sigma ( \Omega _1 ,\Omega _2 \boldsymbol{\Gamma} ,N) \circ \mbox{d}W
\end{array}
\right.
\label{Suslov_PE}
\end{equation}
and the Lagrange multiplier $\lambda$ can be computed independently as
\begin{equation}
\lambda \mbox{d}t= \left( \boldsymbol{\Omega} \times \mathbb{I} \boldsymbol{\Omega} + \frac{\partial U}{\partial \boldsymbol{\Gamma} } \times \boldsymbol{\Gamma} \right) \cdot \mathbf{e} _3\, \mbox{d}t+ I _3 \mbox{d}N \, .
\label{lambda_eq}
\end{equation}
By assuming $U=0$ for simplicity, so that the variable $ \boldsymbol{\Gamma} $ can be discarded, the associated Fokker-Planck equations for the probability density $p( \Omega _1 , \Omega _2, N,t)$ is
\begin{equation}
\partial _t p=- \partial _{\Omega_1} ( J_1 \Omega _2 Np)-\partial _{\Omega_2} ( J_1 \Omega _1 Np) - \partial _N(fp)+ \frac{1}{2} \partial_N ( \sigma\partial _N( \sigma p)),
\label{FPSuslov_simple}
\end{equation}
where
\[
J _1 = \frac{I_3-I_2}{I_1}, \quad J _2 = \frac{I_1-I_3}{I_2}.
\]
The Fokker-Planck equations for the general case can be derived similarly.
\medskip
The stochastic extension \eqref{Suslov_PE} does not preserve the integrals mentioned in (a)--(d) if no further assumptions are made. For example for (a), $U=0$ and we have
\[
\frac{1}{2} \mbox{d}(\mathbb{I} \boldsymbol{\Omega} \cdot \mathbb{I} \boldsymbol{\Omega} )= \lambda I_3 N\mbox{d}t.
\]
For the Kharlamova integral (b), $U(\boldsymbol{\Gamma} )=\bchi\cdot \boldsymbol{\Gamma}$ and we have
\[
\mbox{d}( \mathbb{I} \boldsymbol{\Omega} \cdot \boldsymbol{\chi} )= -(\mathbb{I} \boldsymbol{\Omega} \times \boldsymbol{\Omega} ) \cdot \boldsymbol{\chi}\mbox{d}t
\]
which does not vanish in general. Indeed, the conservation of Kharlamova integral hinges on the fact that the result of $\boldsymbol{\Omega} \times \mathbb{I} \boldsymbol{\Omega} $ points in $\mathbf{e}_3=\mathbf{a}$ direction, and $\bchi \cdot \mathbf{a}=0$ affords vanishing of the appropriate terms. This condition is broken in the case of noisy constraints \eqref{Noisy_constr_1}. We shall see below under which hypothesis on the body this integral of motion is preserved.
For the integrals (c) and (d) we have, respectively,
\[
\mbox{d} (\mathbb{I} \boldsymbol{\Omega} \cdot \boldsymbol{\Gamma} )=I_3\Gamma _3\mbox{d}N \quad\text{and}\quad \frac{1}{2}\mbox{d} \left( \mathbb{I} \boldsymbol{\Omega}\cdot \mathbb{I} \boldsymbol{\Omega} - A\boldsymbol{\Gamma}\cdot \boldsymbol{\Gamma}\right) = \lambda I_3N \mbox{d}t.
\]
\paragraph{Noisy constraint II.} The energy preserving stochastic constraint introduced in \S\ref{I_and_II} corresponds to
\begin{equation}
\mathbf{N} \cdot\boldsymbol{\Omega} =0,\quad \mbox{d} \mathbf{N} = \mathbf{f} ( \boldsymbol{\Omega} ,\boldsymbol{\Gamma} , \mathbf{N} )\mbox{d}t+ \boldsymbol{\sigma} ( \boldsymbol{\Omega} ,\boldsymbol{\Gamma} , \mathbf{N} ) \circ \mbox{d}W,
\label{eqN}
\end{equation}
for two functions $ \mathbf{f} , \boldsymbol{\sigma} : \mathbb{R} ^3 \times\mathbb{R} ^3 \times \mathbb{R} ^3 \rightarrow \mathbb{R} $.
From the stochastic Euler-Poincar\'e-Suslov equations \eqref{Stoch_EPS2} we get the stochastic system
\begin{equation}
\label{Suslov_homogeneous}
\left\{
\begin{array}{l}
\displaystyle\vspace{0.2cm}
\mbox{d} \mathbb{I} \boldsymbol{\Omega} +\boldsymbol{\Omega}\times \mathbb{I} \boldsymbol{\Omega} \mbox{d}t=\boldsymbol{\Gamma} \times \frac{\partial U}{\partial\boldsymbol{\Gamma}} \mbox{d}t + \lambda \mathbf{N} \mbox{d}t, \quad \mathbf{N} \cdot \boldsymbol{\Omega} =0,\\
\mbox{d}\boldsymbol{\Gamma} = \boldsymbol{\Gamma} \times \boldsymbol{\Omega}\mbox{d}t,\quad \mbox{d} \mathbf{N} =\mathbf{f} \mbox{d}t+ \boldsymbol{\sigma} \circ \mbox{d}W,
\end{array}
\right.
\end{equation}
so the Lagrange multiplier is
\begin{equation}
\label{lambda_homogeneous}
\lambda \mbox{d} t = \frac{ \left( \boldsymbol{\Omega}\times \mathbb{I} \boldsymbol{\Omega} +\displaystyle \frac{\partial U}{\partial\boldsymbol{\Gamma}} \times \boldsymbol{\Gamma} \right) \cdot \mathbb{I} ^{-1} \mathbf{N} \mbox{d} t \, -\boldsymbol{\Omega} \cdot \mbox{d} \mathbf{N} }{ \mathbf{N} \cdot\mathbb{I} ^{-1} \mathbf{N} } .
\end{equation}
As opposed to \eqref{case_1}, the stochastic equation \eqref{lambda_homogeneous} preserves the energy $E= \frac{1}{2} \boldsymbol{\Omega}\cdot \mathbb{I} \boldsymbol{\Omega} -U( \boldsymbol{\Gamma} )$.
The stochastic extension \eqref{Suslov_homogeneous} does not preserve the integrals mentioned in (a)--(d) if no further assumptions are made. Indeed, for the integrals (a)--(d), we have, under the same assumptions as before, except when we consider the fixed vector $ \mathbf{a}=\mathbf{e}_3 $, since this vector is replaced by the process $ \mathbf{N} $:
\begin{equation}\label{integrals}
\begin{aligned}
{\rm (a)} &\quad \frac{1}{2} \mbox{d}\left( \mathbb{I} \boldsymbol{\Omega} \cdot \mathbb{I} \boldsymbol{\Omega} \right) = \lambda \mathbb{I} \boldsymbol{\Omega} \cdot \mathbf{N} \mbox{d}t\\
{\rm (b)} &\quad \mbox{d} (\mathbb{I} \boldsymbol{\Omega} \cdot \boldsymbol{\chi} )= -(\mathbb{I} \boldsymbol{\Omega} \times \boldsymbol{\Omega} ) \cdot \boldsymbol{\chi}\mbox{d}t+ \lambda \mathbf{N}\cdot\boldsymbol{\chi} \mbox{d}t \textcolor{white}{ \frac{1}{2} } \\
{\rm (c)}&\quad \mbox{d}( \mathbb{I} \boldsymbol{\Omega} \cdot \boldsymbol{\Gamma} )= \lambda \boldsymbol{\Gamma}\cdot \mathbf{N} \mbox{d}t \textcolor{white}{ \frac{1}{2} } \\
{\rm (d)}&\quad \frac{1}{2} \mbox{d}\left( \mathbb{I} \boldsymbol{\Omega} \cdot \mathbb{I} \boldsymbol{\Omega} - A\boldsymbol{\Gamma} \cdot \boldsymbol{\Gamma} \right) = \lambda \mathbb{I} \boldsymbol{\Omega} \cdot \mathbf{N} \mbox{d}t.
\end{aligned}
\end{equation}
Note that we can longer assume that the vector $\mathbf{N}$ is an eigenvector of $\mathbb{I}$, since it is not stationary.
\paragraph{Evolution of the Kharlamova integral with stochastic constraints.}
Let us now analyze the conservation of the Kharlamova integral $\mathbb{I} \boldsymbol{\Omega} \cdot \bchi$.
In general, this integral is not conserved in both settings of the stochastic nonholonomic constraints.
For the first case, i.e., equations \eqref{case_1}, we compute
\[
\mbox{d}( \mathbb{I} \boldsymbol{\Omega} \cdot \boldsymbol{\chi} )= \chi _1 (I_2-I_3) \Omega _2 N\mbox{d}t+ \chi _2 (I_3-I_1) \Omega _1 N\mbox{d}t,
\]
where we assumed, as in (b), $U ( \boldsymbol{\Gamma} )= \bchi \cdot \boldsymbol{\Gamma}$ with $\bchi \cdot \mathbf{a}=0$ for constant vectors $\bchi$ and $\mathbf{a}= \mathbf{e} _3 $.
It is easy to see that $\mathbb{I} \boldsymbol{\Omega} \cdot \bchi$ is an integral of motion if either one of the following conditions is satisfied:
\begin{enumerate}
\item $ \boldsymbol{\chi} = \mathbf{e}_1$ and $I_1=I_3$,
\item $\boldsymbol{\chi} = \mathbf{e}_2$ and $I_2=I_3$, or
\item $ \mathbb{I} = I_0 {\rm Id}_{3 \times 3}$.
\end{enumerate}
On the other hand, for the second case, i.e., equations \eqref{Suslov_homogeneous}, if one of the three cases 1., 2., or 3. above is realised, we get the evolution equation for the quantity $\mathbb{I} \boldsymbol{\Omega} \cdot \bchi$ as
\begin{equation}
\label{Suslov_homogeneous2}
\mathbb{I} \mbox{d} \boldsymbol{\Omega} \cdot \bchi =\lambda \mathbf{N} \cdot \bchi \mbox{d}t \, ,
\end{equation}
so $ \mathbb{I}\boldsymbol{\Omega} \cdot \bchi$ is only conserved in time if $\mathbf{N} \cdot \bchi=0$ for all times. This can happen, for example, if $\mathbf{N}(0) \cdot \bchi=0$ and the evolution of $\mathbf{N}$ satisfies
\begin{equation}
\label{eq_N2}
d \mathbf{N} = (\mathbf{g} \times \bchi )\mbox{d} t + (\boldsymbol{\eta} \times \bchi)\circ \mbox{d}W,
\end{equation}
for arbitrary functions $ \mathbf{g} , \boldsymbol{\eta}$ of $( \boldsymbol{\Omega} , \boldsymbol{\Gamma} , \mathbf{N} )$.
\paragraph{Preservation of the other integrals.} From the evolutions equations \eqref{integrals}, we see that for the second type of stochastic constraints, it is always possible to choose the noise such that one of the integral is preserved.
For the integral (c), it suffices to choose $\mathbf{N}(0) \cdot \boldsymbol{\Gamma} (0)=0$ and the evolution of $\mathbf{N}$ given by
\begin{equation}
\label{eq_N2_c}
\mbox{d} \mathbf{N} = (\mathbf{g} \times\boldsymbol{\Gamma} )\mbox{d} t + (\boldsymbol{\eta} \times\boldsymbol{\Gamma} )\circ \mbox{d}W,
\end{equation}
for arbitrary functions $ \mathbf{g} , \boldsymbol{\eta}$ of $( \boldsymbol{\Omega} , \boldsymbol{\Gamma} , \mathbf{N} )$.
For the Clebsch-Tisserand integral (d), which contains (a) as a particular case, it suffices to choose $\mathbf{N}(0) \cdot \mathbb{I} \boldsymbol{\Omega} (0)=0$ and the evolution of $\mathbf{N}$ given by
\begin{equation}
\label{eq_N2_c}
\mbox{d} \mathbf{N} = (\mathbf{g} \times\mathbb{I} \boldsymbol{\Omega} )\mbox{d} t + (\boldsymbol{\eta} \times\mathbb{I} \boldsymbol{\Omega} )\circ \mbox{d}W,
\end{equation}
for arbitrary functions $ \mathbf{g} , \boldsymbol{\eta}$ of $( \boldsymbol{\Omega} , \boldsymbol{\Gamma} , \mathbf{N} )$.
\paragraph{Analytical solutions and Fokker-Planck equations.}
Let us simplify the system further and assume that $ \mathbb{I} =I_0 {\rm Id}_{3 \times 3}$ and $U=0$ in the first case \eqref{case_1}. Then, the equations of motion become simply
\begin{equation}
I_0 \mbox{d} \boldsymbol{\Omega} =I_0 \frac{ \mathbf{a} }{ |\mathbf{a}|^2} \mbox{d} N \quad \Rightarrow \quad \boldsymbol{\Omega} (t) = \boldsymbol{\Omega}_0 + \frac{ \mathbf{a} }{ |\mathbf{a}|^2} \left( N (t) - N_0 \right) \, .
\label{solN}
\end{equation}
It is easy to verify that indeed, $\boldsymbol{\Omega}(t) \cdot \mathbf{a}=N(t)$ for all $t$. Using this solution, we can reduce the evolution equation to that of $N$.
To contrast this with the second case \eqref{Suslov_homogeneous}, let us also assume $\mathbb{I}=I_0 {\rm Id}_{3 \times 3}$ and $U=0$. Then, we get
\begin{equation}
I_0 \mbox{d} \boldsymbol{\Omega} = \lambda \mathbf{N} \mbox{d} t \, , \quad
\lambda = - I_0 \frac{ \boldsymbol{\Omega} \cdot \mbox{d} \mathbf{N}}{|\mathbf{N}|^2} \, ,
\label{dOmega0}
\end{equation}
and, by using $\mbox{d} \mathbf{N} =\mathbf{f} \mbox{d}t+ \boldsymbol{\sigma} \circ \mbox{d}W$,
\begin{equation}
\mbox{d} \boldsymbol{\Omega} = - \mathbf{N} \frac{ \boldsymbol{\Omega} \cdot \mathbf{f}}{|\mathbf{N}|^2} \mbox{d} t
-\mathbf{N} \frac{ \boldsymbol{\Omega} \cdot \boldsymbol{\sigma}}{|\mathbf{N}|^2} \circ \mbox{d} W :=
\alpha \mathbf{N} \mbox{d} t + \beta \mathbf{N} \circ \mbox{d} W \, .
\label{dOmega}
\end{equation}
No further progress is possible without additional assumptions on $\mathbf{f}$ and $\boldsymbol{\sigma}$ in \eqref{Suslov_homogeneous}. This shows the fundamental difference between these two cases.
It is also possible to write the equation of evolution for the probability density.
Assuming that in the simplest case $\mathbf{f}$ and $\boldsymbol{\sigma}$ do not depend on $\boldsymbol{\Gamma}$, we obtain the Fokker-Planck equation for the probability density $p(\boldsymbol{\Omega},\mathbf{N},t)$ as
\begin{equation}
\partial _t p = - {\rm div}_{\boldsymbol{\Omega}} (\alpha \mathbf{N} p ) - {\rm div}_{\mathbf{N}} (\mathbf{f} p )
+
\frac{1}{2}
\left(
{\rm div}_{\boldsymbol{\Omega}} \beta \mathbf{N} + {\rm div}_{\mathbf{N}} \boldsymbol{\sigma}
\right)
\left(
{\rm div}_{\boldsymbol{\Omega}} \beta \mathbf{N} p + {\rm div}_{\mathbf{N}} \boldsymbol{\sigma} p
\right) \, ,
\label{FP_linear}
\end{equation}
where it is understood that the diffusion terms in the left parentheses of the last term on the right-hand-side apply as an operator to the the right parentheses of the same term.
\section{Conclusion}
In these notes, we have outlined the procedure for introducing stochasticity into nonholonomic constraints. We have shown that such introduction of noise allows to preserve various integrals of motion. We have also shown that for general nonholonomic systems, the energy is conserved for arbitrary noisy perturbation connecting the velocities, as long as the perturbation preserves the linear structure of the constraints between the velocity components. On the other hand, the energy is in general not conserved in the case of affine, or inhomogeneous, noisy constraints.
Furthermore, in \cite{FGBPu2016}, we have proved that the expectation value for the energy can either increase indefinitely, or remain finite, depending on the type of noise used in the equations. These results are derived using analytical studies stemming from the cases of reduced dynamics, such as rolling of a ball along a one-dimensional line. The choice of noise in constraints preventing infinite energy growth is an interesting question which deserves further studies.
Another interesting problem is the study of Fokker-Planck equations such as \eqref{FPSuslov_simple} and \eqref{FP_linear}
resulting from the Stratonovich SDEs \eqref{LdA1_rs} or \eqref{LdA2_rs}. It remains to be seen what conditions on the system and noise lead to the resulting Fokker-Planck equation being hypoelliptic, so that the traditional methods of probability theory \cite{Pa2014} apply here. A direct attempt to use \eqref{LdA1_rs} and \eqref{LdA2_rs} to derive Fokker-Planck equations with standard methods gives highly non-intuitive and cumbersome formulas. It is possible that a more elegant and geometric approach developed by the method of Hamel applied to nonholonomic constraints \cite{ZeLeBl2012,ShBeZeBl2015}, can be useful in approaching this problem. In this method, the velocity coordinates are chosen in such a way that the nonholonomic constraints take a particularly simple form, which may possibly lead to tractable expressions for the diffusion operator in Fokker-Planck equations. We also defer the discussion of this interesting problem for the future.
\section*{Acknowledgements}
We acknowledge fruitful and enlightening discussions with Profs. M. Barlow, L. Bates, A. M. Bloch, D. D. Holm, G. Pavliotis, T. S. Ratiu, J. Sniatycki, and D. V. Zenkov. FGB is partially supported by the ANR project GEOMFLUID 14-CE23-0002-01. VP acknowledges support from NSERC Discovery Grant and the University of Alberta Centennial Fund.
\addcontentsline{toc}{section}{References}
|
\section{Background}
\label{sec:background}
In this note we shall work with $\Gray$-categories, i.e., categories enriched in the category $\V = \Gray$ (see~\cite{kelly:book} for the theory of enriched categories), where $\Gray$ is a symmetric monoidal closed category of 2-categories with the $\Gray$ tensor product (as described in~\cite{gordon+power+street} or~\cite{gurski}).
\subsection*{$\Gray$-categories}
Recall that a $\Gray$-category $\gkat{K}$ with objects $A$, $B$, $C$, has a hom-2-category $\gkat{K}(A,B)$ for each pair $A$, $B$ of objects, the unit 2-functor $u_{A}: \I \to \gkat{K}(A,A)$ sends the unique $i$-cell ($i = 0,1,2$) to the identity $(i+1)$-cell of $A$ (e.g., the unique object $*$ of $\I$ gets sent to $1_{A}: A \to A$), and the \emph{composition}
\[
\gkat{K}(B,C) \tensor \gkat{K}(A,B) \to \gkat{K}(A,C)
\]
is essentially the composition \emph{cubical} functor
\[
\gkat{K}(B,C) \times \gkat{K}(A,B) \to \gkat{K}(A,C)
\]
yielding, for any $F$ in $\gkat{K}(A,B)$ and any $G$
in $\gkat{K}(B,C)$, two 2-functors
\begin{align*}
(\blank) F: \gkat{K}(B,C) & \to
\gkat{K}(A,C)
\\
G (\blank): \gkat{K}(A,B) & \to
\gkat{K}(A,C)
\end{align*}
(``precomposition'' and ``postcomposition''), with $(\blank) F$ acting on the data
\[
\begin{tikzcd}
B
\arrow[rr, bend left, "H"]
\arrow[rr, bend right, "K", swap]
\arrow[phantom, rr, bend left=15, "\tau"{middle}]
\arrow[phantom, rr, "\alpha \Downarrow \, \Rrightarrow \, \Downarrow \beta"{middle}]
&
&
C
\end{tikzcd}
\]
to give
\[
\begin{tikzcd}
B
\arrow[rr, bend left, "HF"]
\arrow[rr, bend right, "KF", swap]
\arrow[phantom, rr, bend left=15, "\tau F"{middle}]
\arrow[phantom, rr, "\alpha F \Downarrow \, \Rrightarrow \, \Downarrow \beta F"{middle}]
&
&
C
\end{tikzcd}
\]
and $G (\blank)$
acting on
\[
\begin{tikzcd}
A
\arrow[rr, bend left, "H"]
\arrow[rr, bend right, "K", swap]
\arrow[phantom, rr, bend left=15, "\tau"{middle}]
\arrow[phantom, rr, "\alpha \Downarrow \, \Rrightarrow \, \Downarrow \beta"{middle}]
&
&
B
\end{tikzcd}
\]
to give
\[
\begin{tikzcd}
A
\arrow[rr, bend left, "GH"]
\arrow[rr, bend right, "GK", swap]
\arrow[phantom, rr, bend left=15, "G \tau"{middle}]
\arrow[phantom, rr, "G \alpha \Downarrow \, \Rrightarrow \, \Downarrow G \beta"{middle}]
&
&
B.
\end{tikzcd}
\]
The 2-functors $(\blank) F$ and $G (\blank)$ are subject to the equality
\[
(G) F = G (F) = GF
\]
and for every pair
\[
\begin{tikzcd}
A
\arrow[r, bend left, "F"]
\arrow[r, bend right, "F'", swap]
\arrow[phantom, r, "\alpha \dDar"{middle}]
&
B
\end{tikzcd}
\qquad
\begin{tikzcd}
B
\arrow[r, shift left, bend left, "G"]
\arrow[r, bend right, "G'", swap]
\arrow[phantom, r, "\beta \dDar"{middle}]
&
C
\end{tikzcd}
\]
there is an isomorphism
\[
\begin{tikzcd}
G F
\arrow[Rightarrow, r, "\beta F"]
\arrow[Rightarrow, d, "G \alpha", swap]
&
G' F
\arrow[phantom, dl, "\Rrightarrow"{middle,
rotate=225},
"\beta_\alpha"{above left}]
\arrow[Rightarrow, d, "G' \alpha"]
\\
G F'
\arrow[Rightarrow, r, "\beta F'", swap]
&
G' F'
\end{tikzcd}
\]
subject to the cubical functor axioms:
\begin{enumerate}
\item Composition axioms
\[
\begin{tikzcd}[arrows=Rightarrow]
GF
\arrow[r, "\beta F"]
\arrow[d, "G \alpha", swap]
&
G'F
\arrow[d, "G' \alpha"]
\arrow[phantom, dl, "\Ldar"{middle}, "\beta_\alpha"{below right}]
\\
GF'
\arrow[r, "\beta F'", swap]
\arrow[d, "G \gamma", swap]
&
G'F'
\arrow[d, "G' \gamma"]
\arrow[phantom, dl, "\Ldar"{middle}, "\beta_\gamma"{below right}]
\\
GF''
\arrow[r, "\beta F''", swap]
&
G'F''
\end{tikzcd}
\quad
=
\quad
\begin{tikzcd}[arrows=Rightarrow]
GF
\arrow[r, "\beta F"]
\arrow[d, "G \alpha", swap]
&
G'F
\arrow[d, "G' \alpha"]
\arrow[phantom, ddl, "\Ldar"{middle}, "\beta_{\gamma \cdot \alpha}"{below right}]
\\
GF'
\arrow[d, "G \gamma", swap]
&
G'F'
\arrow[d, "G' \gamma"]
\\
GF''
\arrow[r, "\beta F''", swap]
&
G'F''
\end{tikzcd}
\]
and
\[
\begin{tikzcd}[arrows=Rightarrow]
GF
\arrow[r, "\beta F"]
\arrow[d, "G \alpha", swap]
&
G'F
\arrow[r, "\delta F"]
\arrow[d, "G' \alpha"]
\arrow[phantom, dl, "\Ldar"{middle},
"\beta_\alpha"{below right}]
&
G''F
\arrow[d, "G'' \alpha"]
\arrow[phantom, dl, "\Ldar"{middle}, "\delta_\alpha"{below right}]
\\
GF'
\arrow[r, "\beta F'", swap]
&
G'F'
\arrow[r, "\delta F'", swap]
&
G''F'
\\
&
{=}
&
\end{tikzcd}
\]
\[
\begin{tikzcd}[arrows=Rightarrow]
GF
\arrow[r, "\beta F"]
\arrow[d, "G \alpha", swap]
&
G'F
\arrow[r, "\delta F"]
&
G''F
\arrow[d, "G'' \alpha"]
\arrow[phantom, dll, "\Ldar"{middle}, "{(\delta \cdot \beta)}_\alpha"{below right}]
\\
GF'
\arrow[r, "\beta F'", swap]
&
G'F'
\arrow[r, "\delta F'", swap]
&
G''F'.
\end{tikzcd}
\]
\item ``Modification'' axioms
\[
\begin{tikzcd}
G F
\arrow[Rightarrow, r, "\beta F"]
\arrow[Rightarrow, d, "G \alpha", swap]
&
G' F
\arrow[phantom, dl, "\Rrightarrow"{middle, rotate=225}, "\beta_\alpha"{above left}]
\arrow[Rightarrow, d, "G' \alpha", swap, ""{name=A, middle}]
\arrow[Rightarrow, d, shiftarr={xshift=10ex}, "G' \alpha'", ""{name=B, middle}]
\arrow[from=A, to=B, phantom, "\Llar"{below}, "G' s"{above}]
\\
G F'
\arrow[Rightarrow, r, "\beta F'", swap]
&
G' F'
\end{tikzcd}
\qquad
=
\qquad
\begin{tikzcd}
G F
\arrow[Rightarrow, r, "\beta F"]
\arrow[Rightarrow, d, "G \alpha'", ""{name=A, middle}]
\arrow[Rightarrow, d, shiftarr={xshift=-10ex}, "G \alpha", swap, ""{name=B, middle}]
\arrow[from=A, to=B, phantom, "\Llar"{below}, "G s"{above}]
&
G' F
\arrow[phantom, dl, "\Rrightarrow"{middle, rotate=225}, "\beta_{\alpha'}"{above left}]
\arrow[Rightarrow, d, "G' \alpha'"]
\\
G F'
\arrow[Rightarrow, r, "\beta F'", swap]
&
G' F'
\end{tikzcd}
\]
for any 3-cell $s: \alpha' \Rrightarrow \alpha$ and
\[
\begin{tikzcd}[arrows=Rightarrow]
G F
\arrow[Rightarrow, r, "\beta F", swap]
\arrow[Rightarrow, r, shiftarr={yshift=8ex}, "\beta' F", ""{name=B, middle}]
\arrow[r, bend left, phantom, "t F \Ddar"{middle}]
\arrow[Rightarrow, d, "G \alpha", swap]
&
G' F
\arrow[phantom, dl, "\Rrightarrow"{middle, rotate=225}, "\beta_\alpha"{above left}]
\arrow[Rightarrow, d, "G' \alpha", ""{name=A, middle}]
\\
G F'
\arrow[Rightarrow, r, "\beta F'", swap]
&
G' F'
\end{tikzcd}
\qquad
=
\qquad
\begin{tikzcd}[arrows=Rightarrow]
G F
\arrow[Rightarrow, r, "\beta' F"]
\arrow[Rightarrow, d, "G \alpha", swap]
&
G' F
\arrow[phantom, dl, "\Rrightarrow"{middle, rotate=225}, "{\beta'}_\alpha"{above left}]
\arrow[Rightarrow, d, "G' \alpha", ""{name=A, middle}]
\\
G F'
\arrow[Rightarrow, r, "\beta' F'"]
\arrow[Rightarrow, r, shiftarr={yshift=-8ex}, "\beta F", swap, ""{name=B, middle}]
\arrow[r, bend right, phantom, "t F' \Ddar"{middle}]
&
G' F'
\end{tikzcd}
\]
for any 3-cell $t: \beta' \Rrightarrow \beta$.
\end{enumerate}
Given any triple
\[
\begin{tikzcd}
A
\arrow[r, bend left, "F"]
\arrow[r, bend right, "F'", swap]
\arrow[phantom, r, "\alpha \dDar"{middle}]
&
B
\end{tikzcd}
\qquad
\begin{tikzcd}
B
\arrow[r, shift left, bend left, "G"]
\arrow[r, bend right, "G'", swap]
\arrow[phantom, r, "\beta \dDar"{middle}]
&
C
\end{tikzcd}
\qquad
\begin{tikzcd}
C
\arrow[r, shift left, bend left, "H"]
\arrow[r, bend right, "H'", swap]
\arrow[phantom, r, "\gamma \dDar"{middle}]
&
D
\end{tikzcd}
\]
the associativity equalities
\[
\gamma_{(\beta F)} = (\gamma_\beta) F, \quad
\gamma_{(G
\alpha)} = {(\gamma G)}_\alpha, \quad H(\beta_\alpha)
=
{(H\beta)}_\alpha
\]
hold, allowing us to relax the notation when working
with the invertible 3-cells $\beta_\alpha$.
Finally, the unit equalities
\[
1 F = F, \quad G 1 = G
\]
hold, where $1$ can stand for the identity $1$-cell,
$2$-cell or $3$-cell on $B$.
\subsection*{Duality}
B\'{e}nabou's theorem admits variations based on duality, both in the case of ordinary categories and in the setting of $\Gray$-categories. We introduce two dual constructions on a $\Gray$-category $\gkat{K}$.
\begin{itemize}
\item The \emph{horizontal dual} $\gkat{K}^\op$ of $\gkat{K}$ is defined by reversing the 1-cells of $\gkat{K}$. That is,
\[
\gkat{K}^\op(A,B) = \gkat{K}(B,A).
\]
Composition in $\gkat{K}^\op$ is defined by the symmetry of the $\Gray$-tensor product, as is usual in the context of enriched categories.
\item The \emph{vertical dual} $\gkat{K}^\co$ of $\gkat{K}$ is defined by reversing the 2-cells of $\gkat{K}$. That is, we put
\[
\gkat{K}^\co(A,B) = (\gkat{K}(A,B))^\op;
\]
observe that the 1-cells of $\gkat{K}$ are not reversed. In this definition we use that $(\gkat{K}(A,B))$ is a \emph{2-category} and that we can therefore form \emph{its} opposite.
\end{itemize}
\section{Pseudoadjunctions and pseudoextensions}
\label{sec:pseudo-notions-in-gray-cats}
We will recall the notion of a pseudoadjunction and a pseudolifting \emph{in} a general $\Gray$-category.
\begin{definition}[Pseudoadjunctions in
$\Gray$-categories]
\label{def:psadj-in-gray}
Let $\gkat{K}$ be a $\Gray$-category.
We say that 1-cells $U: A \to X$, $F: X \to A$ together with the data
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "1", swap, ""{name=beg, middle}]
\arrow[dr, bend left, phantom, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "U", ""{name=fin, middle}]
\\
&
X
\end{tikzcd}
\quad
\begin{tikzcd}
A
\arrow[d, "U", swap]
\arrow[dr, "1_A", ""{name=fin, middle}]
\arrow[dr, bend right, phantom, "\eps \uRar"{middle}]
&
\\
|[alias=X]|X
\arrow[r, "F", swap]
&
A
\end{tikzcd}
\quad
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s \Urar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F"]
\\
&
F
\end{tikzcd}
\quad
\begin{tikzcd}
U
\arrow[Rightarrow, r, "\eta U"]
\arrow[Rightarrow, dr, "1_U", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "t \Ldar"{middle}]
&
|[alias=UFU]|UFU
\arrow[Rightarrow, d, "U \eps"]
\\
&
U
\end{tikzcd}
\]
(with $s$ and $t$ being isomorphisms) constitute a
\emph{pseudoadjunction} in $\gkat{K}$ with
\emph{unit}
$\eta$
and \emph{counit} $\eps$ if these data satisfy two
coherence
identities: the 3-cell
\[
\begin{tikzcd}
&
UF
\arrow[Rightarrow, dr, "UF \eta"]
\arrow[Rightarrow, drr, bend left, "1_{UF}",
""{name=upp,
middle}]
\arrow[phantom, dd, "\Rrightarrow"{middle,
rotate=-90}, "{\eta_\eta}"{above left}]
& & \\
1_X
\arrow[Rightarrow, ur, "\eta"]
\arrow[Rightarrow, dr, "\eta", swap]
& &
UFUF
\arrow[Rightarrow, r, "U \eps F"]
\arrow[phantom, "\Rrightarrow"{middle,
sloped, rotate=180}, "U s"{above left},
from=upp]
& UF \\
&
UF
\arrow[Rightarrow, ur, "\eta UF"]
\arrow[Rightarrow, urr, bend right, "1_{UF}", swap,
""{name=downn,
middle}]
\arrow[phantom, urr, "\Rrightarrow"{middle,
sloped, rotate=-90}, "t F"{above left}]
& &
\end{tikzcd}
\]
has to be equal to the identity 3-cell on $\eta$, and
the 3-cell
\[
\begin{tikzcd}
& &
FU
\arrow[Rightarrow, rd, "\eps"]
\arrow[
phantom,
dd,
"\Rrightarrow"{sloped, middle},
"{\eps_\eps}"{above left}
]
& \\
FU
\arrow[Rightarrow, urr, bend left, "1_{FU}",
""{name=upp,
middle}]
\arrow[Rightarrow, r, "F \eta U"]
\arrow[Rightarrow, drr, bend right, "1_{FU}", swap,
""{name=downn,
middle}]
&
FUFU
\arrow[Rightarrow, ur, "\eps FU"]
\arrow[Rightarrow, dr, "FU \eps", swap]
\arrow[
phantom,
"\Rrightarrow"{sloped, middle},
"s U"{above left},
from=upp
]
\arrow[
phantom,
"\Rrightarrow"{sloped, middle, rotate=180},
"F t"{above left},
to=downn
]
& & 1_A \\
& &
FU
\arrow[Rightarrow, ur, "\eps", swap]
&
\end{tikzcd}
\]
has to be equal to the identity 3-cell on $\eps$. We write $F \adj U: A \to X$ for this pseudoadjunction.
\end{definition}
Duality operations with the $\Gray$-category $\gkat{K}$ transform pseudoadjunctions into pseudoadjunctions. The roles of the defining data have to be swapped accordingly.
\begin{remark}
Suppose we are given the category $\gkat{K}$ as in
Definition~\ref{def:psadj-in-gray} and the data for
the pseudoadjunction $F \adj U: A \to X$.
\begin{enumerate}
\item In $\gkat{K}^\op$, the same data transform into
a pseudoadjunction $U \adj F: X \to A$ due to the
reversal of 1-cells. The unit $\eta$ and counit
$\eps$
stay the same, as well as the coherence 3-cells $s$
and $t$.
\item In $\gkat{K}^\co$, the same data transform into
a pseudoadjunction $U \adj F: X \to A$, but with unit
$\eps$ and counit $\eta$; the coherence 3-cells $s$
and $t$ stay the same, although their role as
witnesses for the triangle isomorphisms is swapped.
\end{enumerate}
\end{remark}
The notion of a (left) pseudoextension is the appropriate weakening of the usual notion of a (left) Kan extension.
\begin{definition}[Left pseudoextension~\cite{fiore:pseudo-limits,nunes-pseudo-kan}]
\label{def:left-pseudoextension}
In a $\Gray$-category $\gkat{K}$, we say that
\[
\begin{tikzcd}
X
\arrow[r, "J"]
\arrow[dr, "H", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "L", ""{name=fin, middle}]
\\
&
B
\end{tikzcd}
\]
exhibits $L$ as a left pseudoextension of $H$ along
$J$ if for each
\[
\begin{tikzcd}
X
\arrow[r, "J"]
\arrow[dr, "H", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "f \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "K", ""{name=fin, middle}]
\\
&
B
\end{tikzcd}
\]
(i.e., $f: H \Rightarrow KJ$) there is a 2-cell
$f^\sharp: L
\Rightarrow K$ and an isomorphism 3-cell
\[
\begin{tikzcd}
H
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin, middle}]
\arrow[dr, phantom, bend left, "\Ldar \mu(f)"{middle}]
&
|[alias=LJ]|LJ
\arrow[Rightarrow, d, "f^\sharp J"]
\\
&
KJ
\end{tikzcd}
\]
such that for each $k: L \Rightarrow K$ and a 3-cell
\[
\begin{tikzcd}
H
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar \omega"{middle}]
&
|[alias=LJ]|LJ
\arrow[Rightarrow, d, "kJ"]
\\
&
KJ
\end{tikzcd}
\]
there is a unique 3-cell $\lift{\omega}: k
\Rrightarrow f^\sharp$ such that
\[
\begin{tikzcd}
H
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin1,
middle}]
&
|[alias=LJ]|LJ
\arrow[Rightarrow, d, "f^\sharp J", swap, ""{name=fin2, middle}]
\\
&
|[alias=KJ]|KJ
\arrow[phantom, from=LJ, to=fin1, "\Rrightarrow"{near
end, rotate=225},
"\mu(f)"{near
start}]
\arrow[Rightarrow, from=LJ, to=KJ,
shiftarr={xshift=10ex},
"kJ", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\lift{\omega} J"{below}, phantom]
\end{tikzcd}
=
\begin{tikzcd}
H
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin, middle}]
\arrow[dr, phantom, bend left, "\Ldar \omega"{middle}]
&
|[alias=LJ]|LJ
\arrow[Rightarrow, d, "kJ"]
\\
&
KJ
\end{tikzcd}
\]
We say that the pseudoextension $\eta: H \to LJ$ is
\emph{preserved by $G: B \to C$} if the 2-cell
\[
\begin{tikzcd}
X
\arrow[r, "J"]
\arrow[dr, "GH", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "G \eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "GL", ""{name=fin, middle}]
\\
&
C
\end{tikzcd}
\]
exhibits $GL$ as a left pseudoextension of $GH$ along
$J$.
The pseudoextension $\eta: H \to LJ$ is said to be
\emph{absolute} if it is preserved by \emph{any}
1-cell $G: B \to C$.
\end{definition}
The various notions of duality for $\Gray$-categories allow us to express compactly the definition of (left/right) pseudoextensions and pseudoliftings via the definition of a left pseudoextension.
\begin{definition}
Given a left pseudoextension
\[
\begin{tikzcd}
X
\arrow[r, "J"]
\arrow[dr, "H", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "L", ""{name=fin, middle}]
\\
&
B
\end{tikzcd}
\]
in a $\Gray$-category $\gkat{K}$, we call it
\begin{enumerate}
\item a \emph{right} pseudoextension of $H$ along $J$
in $\gkat{K}^\co$.
\item a \emph{left} pseudolifting of $H$ through $J$
in $\gkat{K}^\op$.
\item a \emph{right} pseudolifting of $H$ through $J$
in $\gkat{K}^{\co \op}$.
\end{enumerate}
\end{definition}
With these definitions at hand, we can move to the statement and proof of B\'{e}nabou's theorem.
\section{B\'{e}nabou's theorem for pseudoextensions}
\label{sec:benabou}
Let us first recall ordinary B\'{e}nabou's theorem~\cite{benabou}.
\begin{theorem}
For functors $U: \A \to \X$ and $F: \X \to \A$ the following are equivalent:
\begin{enumerate}
\item $F \adj U$ holds with unit $\eta$.
\item $\eta$ exhibits $U$ as an absolute left extension of $1_\X$ along $F$.
\item $\eta$ exhibits $U$ as a left extension of $1_\X$ along $F$, and this extension is preserved by $F$.
\item $\eta$ exhibits $F$ as an absolute left lifting of $1_\X$ through $U$.
\item $\eta$ exhibits $F$ as a left lifting of $1_\X$ through $U$, and this lifting is preserved by $U$.
\end{enumerate}
\end{theorem}
The pseudo-version of B\'{e}nabou's theorem can be stated in the same way, changing the notions of adjunction and extension/lifting to the notions of pseudoadjunction and pseudoextension/pseudolifting.
\begin{theorem}
\label{thm:pseudo-benabou}
Given two 1-cells $U: A \to X$ and $F: X \to A$ in a $\Gray$-category $\gkat{K}$, the following are equivalent:
\begin{enumerate}
\item $F \adj U$ is a pseudoadjunction with unit $\eta$.
\item $\eta$ exhibits $U$ as an absolute left pseudoextension of $1_X$ along $F$.
\item $\eta$ exhibits $U$ as a left pseudoextension of $1_X$ along $F$, and this extension is preserved by $F$.
\item $\eta$ exhibits $F$ as an absolute left pseudolifting of $1_X$ through $U$.
\item $\eta$ exhibits $F$ as a left pseudolifting of $1_X$ through $U$, and this lifting is preserved by $U$.
\end{enumerate}
\end{theorem}
The proof strategy in the ordinary case and in the pseudo-case is the same: it is enough to prove the implications (1) $\implies$ (2) and (3) $\implies$ (1). This is because (2) $\implies$ (3) is trivial, and because the equivalence of (1), (4) and (5) follows by duality. Moreover, the ordinary proofs can serve as a guidance for the proofs of the pseudo-case.
\begin{lemma}[The implication (3) $\implies$ (1)]
Suppose that
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "1", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "U", ""{name=fin, middle}]
\\
&
X
\end{tikzcd}
\]
is a left (Kan) pseudoextension preserved by $F$.
Then
$\eta$ can be made a unit of a pseudoadjunction $F
\adj U$.
\end{lemma}
\begin{proof}
Recall that
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "1", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "U", ""{name=fin, middle}]
\\
&
X
\end{tikzcd}
\]
is a left pseudoextension if for each
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "1", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "f \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "K", ""{name=fin, middle}]
\\
&
X
\end{tikzcd}
\]
(i.e., $f: 1_X \Rightarrow KF$) there is a 2-cell
$f^\sharp: U
\Rightarrow K$ and an isomorphism 3-cell
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar \mu(f)"{middle}]
&
|[alias=UF]|UF
\arrow[Rightarrow, d, "f^\sharp F"]
\\
&
KF
\end{tikzcd}
\]
such that for each $k: U \Rightarrow K$ and a 3-cell
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar \omega"{middle}]
&
|[alias=UF]|UF
\arrow[Rightarrow, d, "kF"]
\\
&
KF
\end{tikzcd}
\]
there is a unique 3-cell $\lift{\omega}: k
\Rrightarrow f^\sharp$ such that
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin1,
middle}]
&
|[alias=UF]|UF
\arrow[Rightarrow, d, "f^\sharp F", swap,
""{name=fin2, middle}]
\\
&
|[alias=KF]|KF
\arrow[phantom, from=UF, to=fin1, "\Rrightarrow"{near
end, rotate=225},
"\mu(f)"{near
start}]
\arrow[Rightarrow, from=UF, to=KF,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\lift{\omega} F"{below}, phantom]
\end{tikzcd}
=
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "f", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar \omega"{middle}]
&
|[alias=UF]|UF
\arrow[Rightarrow, d, "kF"]
\\
&
KF
\end{tikzcd}
\]
For the purpose of establishing notation, we describe
the data concerning the left pseudoextension
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "F", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "F \eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "FU", ""{name=fin, middle}]
\\
&
A
\end{tikzcd}
\]
Given, e.g., the identity
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "F", swap, ""{name=beg, middle}]
\arrow[dr, phantom, bend left, "1_F \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "1_A", ""{name=fin, middle}]
\\
&
A
\end{tikzcd}
\]
(the 2-cell $1_F: F \Rightarrow F$), we have a 2-cell
$(1_A)^\sharp: FU \Rightarrow 1_A$ that we will
denote
by $\eps$ and which will be the counit of the
pseudoadjunction we construct. With this counit comes
an isomorphism
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar s^{-1}"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F"]
\\
&
F
\end{tikzcd}
\]
such that for each $h: FU \Rightarrow 1_A$ and a
3-cell
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar \nu"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "h F"]
\\
&
F
\end{tikzcd}
\]
there is a unique $\lift{\nu}: h \Rrightarrow \eps$
satisfying
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin1,
middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F", swap,
""{name=fin2, middle}]
\\
&
|[alias=F]|F
\arrow[phantom, from=FUF, to=fin1,
"\Rrightarrow"{near
end, rotate=225},
"s^{-1}"{near
start}]
\arrow[Rightarrow, from=FUF, to=F,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\lift{\nu} F"{below}, phantom]
\end{tikzcd}
=
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin, middle}]
\arrow[dr, phantom, bend left, "\Ldar \nu"{middle}]
&
|[alias=FUF]|UF
\arrow[Rightarrow, d, "h F"]
\\
&
F
\end{tikzcd}
\]
Let us first observe that $s^{-1}$ (or, equivalently,
its inverse $s$) witnesses the first
triangle axiom of a pseudomonad. To obtain the second
triangle isomorphism, consider that $\eta: 1_X
\Rightarrow UF$ lifts to the identity $1_U =
\eta^\sharp: U \Rightarrow U$ with the identity
3-cell
($\mu(\eta) = 1_\eta$)
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, dr, "\eta", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar \mu(\eta)"{middle}]
&
|[alias=UF]|UF
\arrow[Rightarrow, d, "1_{UF}"]
\\
&
UF
\end{tikzcd}
\]
By the universal property of the left pseudoextension
given by $\eta$ we get that for the 2-cell $$U
\xRightarrow{\eta U} UFU \xRightarrow{U \eps} U$$ and
the 3-cell
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, d, "\eta", swap]
&
UF
\arrow[Rightarrow, d, "\eta UF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"{\eta_\eta}^{-1}"{near start}]
\\
UF
\arrow[Rightarrow, r, "UF \eta"]
\arrow[Rightarrow, dr, "1_{UF}", swap, ""{name=fin,
middle}]
\arrow[dr, phantom, bend left, "\Ldar Us^{-1}"{middle}]
&
|[alias=UFUF]|UFUF
\arrow[Rightarrow, d, "U \eps F"]
\\
&
UF
\end{tikzcd}
\]
there is a unique 3-cell
\[
\begin{tikzcd}
U
\arrow[Rightarrow, r, "\eta U"]
\arrow[Rightarrow, dr, "1_U", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "t \Ldar"{middle}]
&
|[alias=UFU]|UFU
\arrow[Rightarrow, d, "U \eps"]
\\
&
U
\end{tikzcd}
\]
such that the 3-cell
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
&
UF
\arrow[Rightarrow, r, "\eta UF"]
\arrow[Rightarrow, dr, "1_{UF}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "tF \Ldar"{middle}]
&
|[alias=UFUF]|UFUF
\arrow[Rightarrow, d, "U \eps F"]
\\
&&
UF
\end{tikzcd}
\]
equals
\[
\begin{tikzcd}
1_X
\arrow[Rightarrow, r, "\eta"]
\arrow[Rightarrow, d, "\eta", swap]
&
UF
\arrow[Rightarrow, d, "\eta UF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"{\eta_\eta}^{-1}"{near start}]
\\
UF
\arrow[Rightarrow, r, "UF \eta"]
\arrow[Rightarrow, dr, "1_{UF}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "Us^{-1} \Ldar"{middle}]
&
|[alias=UFUF]|UFUF
\arrow[Rightarrow, d, "U \eps F"]
\\
&
UF
\end{tikzcd}
\]
or, written differently, that the 3-cell
\[
\begin{tikzcd}
&
UF
\arrow[Rightarrow, dr, "UF \eta"]
\arrow[Rightarrow, drr, bend left, "1_{UF}",
""{name=upp,
middle}]
\arrow[phantom, dd, "\Rrightarrow"{middle,
rotate=-90}, "{\eta_\eta}"{above left}]
& & \\
1_X
\arrow[Rightarrow, ur, "\eta"]
\arrow[Rightarrow, dr, "\eta", swap]
& &
UFUF
\arrow[Rightarrow, r, "U \eps F"]
\arrow[phantom, "\Rrightarrow"{middle,
sloped, rotate=180}, "U s"{above left},
from=upp]
& UF \\
&
UF
\arrow[Rightarrow, ur, "\eta UF"]
\arrow[Rightarrow, urr, bend right, "1_{UF}", swap,
""{name=downn,
middle}]
\arrow[phantom, urr, "\Rrightarrow"{middle,
sloped, rotate=-90}, "t F"{above left}]
& &
\end{tikzcd}
\]
is equal to identity. This is precisely the first
coherence axiom for pseudoadjunctions.
For the other coherence axiom, we need the 3-cell
\[
\begin{tikzcd}
& &
FU
\arrow[Rightarrow, rd, "\eps"]
\arrow[
phantom,
dd,
"\Rrightarrow"{sloped, middle},
"{\eps_\eps}"{above left}
]
& \\
FU
\arrow[Rightarrow, urr, bend left, "1_{FU}",
""{name=upp,
middle}]
\arrow[Rightarrow, r, "F \eta U"]
\arrow[Rightarrow, drr, bend right, "1_{FU}", swap,
""{name=downn,
middle}]
&
FUFU
\arrow[Rightarrow, ur, "\eps FU"]
\arrow[Rightarrow, dr, "FU \eps", swap]
\arrow[
phantom,
"\Rrightarrow"{sloped, middle},
"s U"{above left},
from=upp
]
\arrow[
phantom,
"\Rrightarrow"{sloped, middle, rotate=180},
"F t"{above left},
to=downn
]
& & 1_A \\
& &
FU
\arrow[Rightarrow, ur, "\eps", swap]
&
\end{tikzcd}
\]
to be equal to identity as well.
We shall use that for each $h: FU \Rightarrow 1_A$
and
a 3-cell
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "\nu \Ldar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "h F"]
\\
&
F
\end{tikzcd}
\]
there is a \emph{unique} $\lift{\nu}: h \Rrightarrow
\eps$
satisfying
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin1,
middle}]
\arrow[dr, bend left, phantom, "s^{-1} \Ldar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F", swap,
""{name=fin2, middle}]
\\
&
|[alias=F]|F
\arrow[Rightarrow, from=FUF, to=F,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\lift{\nu} F"{below}, phantom]
\end{tikzcd}
=
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "\nu \Ldar"{middle}]
&
|[alias=FUF]|UF
\arrow[Rightarrow, d, "h F"]
\\
&
F
\end{tikzcd}
\]
Thus if we find \emph{two} 3-cells $\alpha, \beta: h
\Rrightarrow \eps$ with
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin1,
middle}]
\arrow[dr, bend left, phantom, "s^{-1} \Ldar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F", swap,
""{name=fin2, middle}]
\\
&
|[alias=F]|F
\arrow[Rightarrow, from=FUF, to=F,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\alpha F"{below}, phantom]
\end{tikzcd}
=
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin1,
middle}]
\arrow[dr, bend left, phantom, "s^{-1} \Ldar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F", swap,
""{name=fin2, middle}]
\\
&
|[alias=F]|F
\arrow[Rightarrow, from=FUF, to=F,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\beta F"{below}, phantom]
\end{tikzcd}
\]
it means that $\alpha = \beta$. Take now the 3-cell
\[
\begin{tikzcd}
FU
\arrow[Rightarrow, r, "F\eta U"]
\arrow[Rightarrow, dr, "1_{FU}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "Ft \Ldar"{middle}]
&
|[alias=FUFU]|FUFU
\arrow[Rightarrow, d, "FU \eps"]
&
\\
&
FU
\arrow[Rightarrow, r, "\eps"]
&
1_A
\end{tikzcd}
\]
for $\alpha$ and the 3-cell
\[
\begin{tikzcd}
FU
\arrow[Rightarrow, r, "F \eta U"]
\arrow[Rightarrow, dr, "1_{FU}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s^{-1}U \Ldar"{middle}]
&
|[alias=FUFU]|FUFU
\arrow[Rightarrow, d, "\eps FU"]
\arrow[Rightarrow, r, "FU \eps"]
&
FU
\arrow[Rightarrow, d, "\eps"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"{\eps_\eps}^{-1}"{near start}]
\\
&
FU
\arrow[Rightarrow, r, "\eps", swap]
&
1_A
\end{tikzcd}
\]
for $\beta$.
We ask whether the 3-cells
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, rrdd, "1_F", swap, ""{name=fin2,
middle}]
&
FUF
\arrow[Rightarrow, r, "F \eta UF"]
\arrow[Rightarrow, dr, "1_{FUF}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "FtF \Ldar"{middle}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, d, "FU \eps F"]
\\
&
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F"]
\\
&
&
F
\arrow[from=FUF, to=fin2, phantom,
"\Rrightarrow"{rotate=225},
"s^{-1}"{near start}]
\end{tikzcd}
\]
and
\[
\begin{tikzcd}
&
FUF
\arrow[Rightarrow, r, "F \eta U F"]
\arrow[Rightarrow, dr, "1_{FUF}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s^{-1}UF \Ldar"{middle}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, d, "\eps FUF"]
\arrow[Rightarrow, r, "FU \eps F"]
&
FUF
\arrow[Rightarrow, d, "\eps F"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"{\eps_{\eps F}}^{-1}"{near start}]
\\
F
\arrow[Rightarrow, ur, "F \eta"]
\arrow[Rightarrow, rrr, bend right, "1_F", swap]
\arrow[rr, phantom, "\Rrightarrow"{rotate=225,
above},
"s^{-1}"{above}]
&
&
FUF
\arrow[Rightarrow, r, "\eps F", swap]
&
F
\end{tikzcd}
\]
are equal. Pasting $s$ and $F\eta_\eta$, we can
equivalently ask whether the 3-cells
\begin{equation}
\label{diag:tmp1}
\begin{tikzcd}
F
\arrow[Rightarrow, d, "F \eta", swap]
\arrow[Rightarrow, r, "F \eta"]
&
FUF
\arrow[Rightarrow, d, "FUF \eta"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225}, "F
\eta_\eta"{near
start}]
&
\\
FUF
\arrow[Rightarrow, r, "F \eta UF"]
\arrow[Rightarrow, dr, "1_{FUF}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "FtF \Ldar"{middle}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, d, "FU \eps F"]
&
\\
&
|[alias=FUF]|FUF
\arrow[Rightarrow, r, "\eps F"]
&
F
\\
&
&
\end{tikzcd}
\end{equation}
and
\begin{equation}
\label{diag:tmp2}
\begin{tikzcd}
F
\arrow[Rightarrow, d, "F \eta", swap]
\arrow[Rightarrow, r, "F \eta"]
&
FUF
\arrow[Rightarrow, d, "FUF \eta"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"F{\eta_{\eta}}"{near start}]
&
\\
FUF
\arrow[Rightarrow, r, "F \eta U F"]
\arrow[Rightarrow, dr, "1_{FUF}", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s^{-1}UF \Ldar"{middle}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, d, "\eps FUF"]
\arrow[Rightarrow, r, "FU \eps F"]
&
FUF
\arrow[Rightarrow, d, "\eps F"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"{\eps_{\eps F}}^{-1}"{near start}]
\\
&
FUF
\arrow[Rightarrow, r, "\eps F", swap]
&
F
\end{tikzcd}
\end{equation}
are equal. Using the first coherence axiom, the
diagram~\eqref{diag:tmp1} is equal to
\begin{equation}
\label{diag:tmp3}
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
&
FUF
\arrow[Rightarrow, r, "FUF \eta"]
\arrow[Rightarrow, dr, "1_{FUF}", swap, ""{name=fin,
middle}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, d, "FU \eps F"]
&
\\
&
&
|[alias=FUF]|FUF
\arrow[Rightarrow, r, "\eps F"]
&
F
\\
&
&
&
\arrow[phantom, from=FUFUF, to=fin,
"\Rrightarrow"{near
end, rotate=225},
"FU s^{-1}"{near
start}]
\end{tikzcd}
\end{equation}
Let us take diagrams~\eqref{diag:tmp2} and
\eqref{diag:tmp3} and paste $\eps_{\eps F}$ and
$\eps_{F\eta}$. The resulting diagrams
\begin{equation}
\label{diag:tmp4}
\begin{tikzcd}
F
\arrow[Rightarrow, d, "F \eta", swap]
\arrow[Rightarrow, r, "F \eta"]
&
FUF
\arrow[Rightarrow, d, "FUF \eta", swap]
\arrow[Rightarrow, dr, "\eps F"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225},
"F{\eta_{\eta}}"{near start}]
&
\\
FUF
\arrow[Rightarrow, r, "F \eta U F"]
\arrow[Rightarrow, dr, "1_{FUF}", swap, ""{name=fin,
middle}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, d, "\eps FUF", swap]
&
F
\arrow[Rightarrow, dl, "F \eta"]
\arrow[phantom, l, "\Rrightarrow"{rotate=180, above},
"\eps_{F \eta}"{above}]
\\
&
FUF
\arrow[Rightarrow, r, "\eps F", swap]
&
F
\arrow[phantom, from=FUFUF, to=fin,
"\Rrightarrow"{near
end, rotate=225},
"s^{-1} UF"{near
start}]
\end{tikzcd}
\end{equation}
and
\begin{equation}
\begin{tikzcd}
& &
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[phantom, d, "\Rrightarrow"{rotate=-90},
"\eps_{F \eta}"{near start}]
&
FUF
\arrow[Rightarrow, dr, "\eps F"]
\arrow[d, phantom, "\Rrightarrow"{rotate=-90},
"\eps_{\eps F}"{near start}]
&
\\
F
\arrow[Rightarrow, r, "F \eta"]
&
FUF
\arrow[Rightarrow, r, "F \eta UF"]
\arrow[Rightarrow, ur, "\eps F",
""{name=beg,
middle}]
\arrow[Rightarrow, rr, "1_{FUF}", swap, ""{name=fin,
middle}, shiftarr={yshift=-10ex}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, r, "FU \eps F"]
\arrow[Rightarrow, ur, "\eps FUF"]
&
FUF
\arrow[Rightarrow, r, "\eps F"]
&
F
\arrow[phantom, from=FUFUF, to=fin,
"\Rrightarrow"{near
end, rotate=-90},
"FU s^{-1}"{near
start}]
\end{tikzcd}
\end{equation}
are equal by using the identities
\[
\begin{tikzcd}
FUF
\arrow[Rightarrow, r]
\arrow[Rightarrow, d]
\arrow[Rightarrow, dd, shiftarr={xshift=-15ex},
""{middle, name=fin}]
&
F
\arrow[Rightarrow, d]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225,
above},
"\eps_{F \eta}"{above}]
\\
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, r]
\arrow[Rightarrow, d]
&
FUF
\arrow[Rightarrow, d]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225,
above},
"\eps_{\eps F}"{above}]
\\
FUF
\arrow[Rightarrow, r]
&
F
\arrow[phantom, from=FUFUF, to=fin,
"\Rrightarrow"{above, rotate=225}, "FUs^{-1}"{above}]
\end{tikzcd}
=
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s^{-1} \Ldar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F"]
\\
&
F
\end{tikzcd}
\]
and
\[
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, d, "F \eta", swap]
&
FUF
\arrow[Rightarrow, r, "\eps F"]
\arrow[Rightarrow, d, "FUF \eta"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225,
above},
"F \eta_{\eta}"{above}]
&
F
\arrow[Rightarrow, d, "F \eta"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225,
above},
"\eps_{F \eta}"{above}]
\\
FUF
\arrow[Rightarrow, r, "F \eta UF", swap]
\arrow[Rightarrow, rr, shiftarr={yshift=-10ex},
"1_{FUF}", swap, ""{middle, name=fin}]
&
|[alias=FUFUF]|FUFUF
\arrow[Rightarrow, r, "\eps FUF", swap]
&
FUF
\arrow[phantom, from=FUFUF, to=fin,
"\Rrightarrow"{rotate=-90, near end}, "s^{-1}
UF"{near
start}]
\end{tikzcd}
=
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s^{-1} \Ldar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F"]
\\
&
F
\end{tikzcd}
\]
The proof is therefore finished.
\end{proof}
\begin{lemma}[The implication (1) $\implies$ (2)]
Suppose that $F \adj U: A \to X$ is a
pseudoadjunction
in a $\Gray$-category $\gkat{K}$ with
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "1", swap, ""{name=beg, middle}]
\arrow[dr, bend left, phantom, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "U", ""{name=fin, middle}]
\\
&
X
\end{tikzcd}
\quad
\begin{tikzcd}
A
\arrow[d, "U", swap]
\arrow[dr, "1_A", ""{name=fin, middle}]
\arrow[dr, bend right, phantom, "\eps \uRar"{middle}]
&
\\
|[alias=X]|X
\arrow[r, "F", swap]
&
A
\end{tikzcd}
\quad
\begin{tikzcd}
F
\arrow[Rightarrow, r, "F \eta"]
\arrow[Rightarrow, dr, "1_F", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "s \Urar"{middle}]
&
|[alias=FUF]|FUF
\arrow[Rightarrow, d, "\eps F"]
\\
&
F
\end{tikzcd}
\quad
\begin{tikzcd}
U
\arrow[Rightarrow, r, "\eta U"]
\arrow[Rightarrow, dr, "1_U", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "t \Ldar"{middle}]
&
|[alias=UFU]|UFU
\arrow[Rightarrow, d, "U \eps"]
\\
&
U
\end{tikzcd}
\]
Then
\[
\begin{tikzcd}
X
\arrow[r, "F"]
\arrow[dr, "1", swap, ""{name=beg, middle}]
\arrow[dr, bend left, phantom, "\eta \uRar"{middle}]
&
|[alias=A]|A
\arrow[d, "U", ""{name=fin, middle}]
\\
&
X
\end{tikzcd}
\]
is an absolute left pseudolifting.
\end{lemma}
\begin{proof}
We need to show that for each $G: X \to Y$ and $g: G
\Rightarrow KF$ there is a 2-cell $g^\sharp: GU
\Rightarrow K$ and an isomorphism
\[
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G\eta"]
\arrow[Rightarrow, dr, "G", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "\mu(g) \Ldar"{middle}]
&
|[alias=GUF]|GUF
\arrow[Rightarrow, d, "g^\sharp F"]
\\
&
KF
\end{tikzcd}
\]
satisfying that for each $k: GU \Rightarrow K$ and
\[
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, dr, "g", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "\omega \Ldar"{middle}]
&
|[alias=GUF]|GUF
\arrow[Rightarrow, d, "kF"]
\\
&
KF
\end{tikzcd}
\]
there is a unique 3-cell $\lift{\omega}: k
\Rrightarrow g^\sharp$ such that
\[
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, dr, "g", swap, ""{name=fin1,
middle}]
&
|[alias=GUF]|GUF
\arrow[Rightarrow, d, "g^\sharp F", swap,
""{name=fin2, middle}]
\\
&
|[alias=KF]|KF
\arrow[phantom, from=GUF, to=fin1,
"\Rrightarrow"{near
end, rotate=225},
"\mu(g)"{near
start}]
\arrow[Rightarrow, from=GUF, to=KF,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\lift{\omega} F"{below}, phantom]
\end{tikzcd}
=
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, dr, "g", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "\omega \Ldar"{middle}]
&
|[alias=GUF]|GUF
\arrow[Rightarrow, d, "kF"]
\\
&
KF
\end{tikzcd}
\]
We shall define $g^\sharp: GU \Rightarrow K$ as the
2-cell
\[
\begin{tikzcd}
GU
\arrow[Rightarrow, r, "gU"]
&
KFU
\arrow[Rightarrow, r, "K \eps"]
&
K
\end{tikzcd}
\]
Then, for the 2-cell $g: G \Rightarrow KF$ we define
$\mu(g)$ to be the 3-cell
\[
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, d, "g", swap]
&
GUF
\arrow[Rightarrow, d, "g UF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225,
above},
"{g_{\eta}}^{-1}"{above}]
\\
KF
\arrow[Rightarrow, r, "KF \eta"]
\arrow[Rightarrow, dr, "1_{KF}", swap, ""{name=fin,
middle}]
&
|[alias=KFUF]|KFUF
\arrow[Rightarrow, d, "K \eps F"]
\\
&
KF
\arrow[phantom, from=KFUF, to=fin,
"\Rrightarrow"{near
end, rotate=225},
"Ks^{-1}"{near
start}]
\end{tikzcd}
\]
Now given a 3-cell
\[
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, dr, "g", swap, ""{name=fin,
middle}]
\arrow[dr, bend left, phantom, "\omega \Ldar"{middle}]
&
|[alias=GUF]|GUF
\arrow[Rightarrow, d, "kF"]
\\
&
KF
\end{tikzcd}
\]
we will show that the ``lifted'' 3-cell
$\lift{\omega}: k \Rrightarrow g^\sharp$
is the 3-cell
\[
\begin{tikzcd}
GU
\arrow[Rightarrow, d, "gU", swap]
\arrow[Rightarrow, r, "1_{GU}"]
&
GU
\arrow[Rightarrow, d, "G \eta U"]
\arrow[Rightarrow, dd, shiftarr={xshift=15ex},
""{middle, name=beg}, "1_{GU}"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"\omega U"{above}]
\\
KFU
\arrow[Rightarrow, d, "K \eps", swap]
&
|[alias=GUFU]|GUFU
\arrow[Rightarrow, l, "kFU"]
\arrow[Rightarrow, d, "GU \eps"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_\eps}^{-1}"{above}]
\\
K
&
GU
\arrow[Rightarrow, l, "k"]
\arrow[phantom, from=beg, to=GUFU,
"\Rrightarrow"{rotate=180, above}, "Gt^{-1}"{above}]
\end{tikzcd}
\]
Indeed, observe that the 3-cell
\[
\begin{tikzcd}
G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, dr, "g", swap, ""{name=fin1,
middle}]
&
|[alias=GUF]|GUF
\arrow[Rightarrow, d, "g^\sharp F", swap,
""{name=fin2, middle}]
\\
&
|[alias=KF]|KF
\arrow[phantom, from=GUF, to=fin1,
"\Rrightarrow"{near
end, rotate=225},
"\mu(g)"{near
start}]
\arrow[Rightarrow, from=GUF, to=KF,
shiftarr={xshift=10ex},
"kF", ""{name=beg2, middle}]
\arrow[from=beg2, to=fin2, "\Rrightarrow"{below,
rotate=180},
"\lift{\omega} F"{below}, phantom]
\end{tikzcd}
\]
is the composite
\[
\begin{tikzcd}[
execute at end picture={
\begin{pgfonlayer}{background}
\foreach \Nombre in {n1,n2,GUFUF,n3}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[greeo!20]
(n1) -- (n2) -- (GUFUF) -- (n3) -- cycle;
\end{pgfonlayer}
}
]
|[alias=n1]|G
\arrow[Rightarrow, r, "G \eta"]
\arrow[Rightarrow, d, "g", swap]
&
GUF
\arrow[Rightarrow, d, "g UF"]
\arrow[Rightarrow, r, "1_{GUF}"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=225,
above},
"{g_{\eta}}^{-1}"{above}]
&
|[alias=n2]|GUF
\arrow[Rightarrow, d, "G \eta UF"]
\arrow[Rightarrow, dd, shiftarr={xshift=15ex},
""{middle, name=beg}, "1_{GUF}"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=180,
above},
"\omega UF"{above}]
\\
|[alias=n3]|KF
\arrow[Rightarrow, r, "KF \eta"]
\arrow[Rightarrow, dr, "1_{KF}", swap, ""{name=fin,
middle}]
&
|[alias=KFUF]|KFUF
\arrow[Rightarrow, d, "K \eps F"]
&
|[alias=GUFUF]|GUFUF
\arrow[Rightarrow, l, "kFUF"]
\arrow[Rightarrow, d, "GU \eps F"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_{\eps F}}^{-1}"{above}]
\\
&
KF
\arrow[phantom, from=KFUF, to=fin,
"\Rrightarrow"{near
end, rotate=225},
"Ks^{-1}"{near
start}]
&
GUF
\arrow[Rightarrow, l, "kF"]
\arrow[phantom, from=beg, to=GUFUF,
"\Rrightarrow"{rotate=180, above}, "GtF^{-1}"{above}]
\end{tikzcd}
\]
which is equal (transforming the green subdiagram) to
the 3-cell
\[
\begin{tikzcd}[
execute at end picture={
\begin{pgfonlayer}{background}
\foreach \Nombre in {n1,n2,n3,n4}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[greeo!20]
(n1) -- (n2) -- (n3) -- (n4) -- cycle;
\end{pgfonlayer}
}
]
&
&
&
|[alias=n1]|G
\arrow[Rightarrow, dd, "G \eta"]
\arrow[Rightarrow, dl, "G \eta"]
\arrow[Rightarrow, dll, "g", swap, ""{middle,
name=fin1}]
\arrow[ddl, phantom, "\Rrightarrow"{rotate=135,
above},
"{G \eta_{\eta}}^{-1}"{above}]
\\
|[alias=m1]|{}
&
|[alias=n2]|KF
\arrow[Rightarrow, d, "KF \eta", swap]
\arrow[Rightarrow, dd, shiftarr={xshift=-16ex},
""{middle, name=fin}, "1_{KF}", swap]
&
|[alias=m2]|GUF
\arrow[Rightarrow, d, "GUF \eta", swap]
\arrow[Rightarrow, l, "kF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_{F \eta}}^{-1}"{above}]
\arrow[phantom, to=fin1, "\Rrightarrow"{rotate=135,
above}, "\omega"{above}]
&
\\
&
|[alias=n3]|KFUF
\arrow[Rightarrow, d, "K \eps F", swap]
\arrow[Rightarrow, phantom, to=fin,
"\Rrightarrow"{rotate=180, above}, "{Ks}^{-1}"{above}]
&
|[alias=GUFUF]|GUFUF
\arrow[Rightarrow, d, "GU \eps F", swap]
\arrow[Rightarrow, l, "k FUF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_{\eps F}}^{-1}"{above}]
&
|[alias=n4]|GUF
\arrow[Rightarrow, l, "G \eta UF", swap]
\arrow[Rightarrow, dl, "1_{GUF}", ""{middle,
name=beg1}]
\arrow[Rightarrow, phantom, from=beg1, to=GUFUF,
"\Rrightarrow"{rotate=135, above},
"{GtF}^{-1}"{above}]
\\
|[alias=m4]|{}
&
KF
&
|[alias=m3]|GUF
\arrow[Rightarrow, l, "kF"]
&
\end{tikzcd}
\]
and, transforming ${Ks}^{-1}$ to ${GUs}^{-1}$, the
3-cell
\[
\begin{tikzcd}[
execute at end picture={
\begin{pgfonlayer}{background}
\foreach \Nombre in {m1,m2,m3,m4}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[cof!20]
(m1) -- (m2) -- (m3) -- (m4) -- cycle;
\end{pgfonlayer}
}
]
&
&
&
|[alias=n1]|G
\arrow[Rightarrow, dd, "G \eta"]
\arrow[Rightarrow, dl, "G \eta"]
\arrow[Rightarrow, dll, "g", swap, ""{middle,
name=fin1}]
\arrow[ddl, phantom, "\Rrightarrow"{rotate=135,
above},
"{G \eta_{\eta}}^{-1}"{above}]
\\
|[alias=m1]|{}
&
|[alias=n2]|KF
\arrow[Rightarrow, d, "KF \eta", swap]
\arrow[Rightarrow, dd, shiftarr={xshift=-16ex},
""{middle, name=fin}, "1_{KF}", swap]
&
|[alias=m2]|GUF
\arrow[Rightarrow, d, "GUF \eta", swap]
\arrow[Rightarrow, l, "kF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_{F \eta}}^{-1}"{above}]
\arrow[phantom, to=fin1, "\Rrightarrow"{rotate=135,
above}, "\omega"{above}]
&
\\
&
|[alias=n3]|KFUF
\arrow[Rightarrow, d, "K \eps F", swap]
\arrow[Rightarrow, phantom, to=fin,
"\Rrightarrow"{rotate=180, above}, "{Ks}^{-1}"{above}]
&
|[alias=GUFUF]|GUFUF
\arrow[Rightarrow, d, "GU \eps F", swap]
\arrow[Rightarrow, l, "k FUF"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_{\eps F}}^{-1}"{above}]
&
|[alias=n4]|GUF
\arrow[Rightarrow, l, "G \eta UF", swap]
\arrow[Rightarrow, dl, "1_{GUF}", ""{middle,
name=beg1}]
\arrow[Rightarrow, phantom, from=beg1, to=GUFUF,
"\Rrightarrow"{rotate=135, above},
"{GtF}^{-1}"{above}]
\\
|[alias=m4]|{}
&
KF
&
|[alias=m3]|GUF
\arrow[Rightarrow, l, "kF"]
&
\end{tikzcd}
\]
is equal to
\[
\begin{tikzcd}[arrows=Rightarrow,
execute at end picture={
\begin{pgfonlayer}{background}
\foreach \Nombre in {m1,m2,m3}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[pur!20]
(m1) -- (m2) -- (m3) -- cycle;
\foreach \Nombre in {n1,n2,GUFUF}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[cof!20]
(n1) -- (n2) -- (GUFUF) -- cycle;
\end{pgfonlayer}
}
]
|[alias=m1]|G
\arrow[rrr, "G \eta"]
\arrow[dr, "G \eta"]
\arrow[dddrrr, shiftarr={xshift=-8ex,yshift=-8ex},
"g", swap, ""{middle, name=fin2}]
&
&
&
|[alias=m2]|GUF
\arrow[dl, "G \eta UF"]
\arrow[ddd, "1_{GUF}", ""{middle, name=beg1}]
\arrow[dll, phantom, "\Rrightarrow"{rotate=225,
above}, "{G\eta_{\eta}}^{-1}"{above}]
\arrow[from=beg1, to=GUFUF, phantom,
"\Rrightarrow"{rotate=225, above},
"{GtF}^{-1}"{above}]
\\
&
|[alias=n1]|GUF
\arrow[r, "GUF \eta"]
\arrow[dr, "1_{GUF}", swap, ""{middle, name=fin1}]
&
|[alias=GUFUF]|GUFUF
\arrow[d, "GU \eps F"]
\arrow[to=fin1, phantom, "\Rrightarrow"{rotate=225,
above},
"{GUs}^{-1}"{above}]
&
\\
&
&
|[alias=n2]|GUF
\arrow[rd, "kF"]
\arrow[to=fin2, phantom, "\Rrightarrow"{rotate=225,
above},
"\omega"{above}]
&
\\
&
&
&
|[alias=m3]|KF
\end{tikzcd}
\]
And since the whole coloured subdiagram equals
identity by
the pseudomonad coherence axiom, the diagram
simplifies to
$\omega$, showing that our choice of $\lift{\omega}$
was
correct. Indeed, the choice of $\lift{\omega}$ is
even the
only possible one: each diagram in the following
series is
equal to $\lift{\omega}$.
\begin{equation}
\begin{tikzcd}[arrows=Rightarrow]
GU
\arrow[r, bend left, "k", ""{middle, name=beg}]
\arrow[r, bend right, "g^\sharp", swap, ""{middle,
name=fin}]
\arrow[r, phantom, "\lift{\omega} \Ddar"{middle}]
&
K
\end{tikzcd}
\end{equation}
\begin{equation}
\begin{tikzcd}[arrows=Rightarrow]
&
|[alias=GUFU]|GUFU
\arrow[dr, "GU \eps"]
\arrow[to=fin2, phantom, "\Rrightarrow"{rotate=-90},
"Gt"{near start}]
&
&
\\
GU
\arrow[ur, "G \eta U"]
\arrow[rr, "1_{GU}", swap, ""{middle, name=fin2}]
\arrow[d, "gU", swap]
\arrow[rr, shiftarr={yshift=24ex}, "1_{GU}",
""{middle, name=beg1}]
&
&
|[alias=GU]|GU
\arrow[r, "k"]
\arrow[d, "gU", swap]
\arrow[dll, phantom, "\Rrightarrow"{rotate=225,
above}, "1_{gU}"{above}]
&
K
\\
KFU
\arrow[rr, "1_{KFU}", swap]
&
&
KFU
\arrow[ur, "K \eps", swap, ""{middle, name=fin1}]
&
\arrow[from=GU, to=fin1, phantom,
"\Rrightarrow"{rotate=-45, below},
"\lift{\omega}"{above}]
\arrow[from=beg1, to=GUFU, phantom,
"\Rrightarrow"{rotate=-90, near end},
"{Gt}^{-1}"{near
start}]
\end{tikzcd}
\end{equation}
\begin{equation}
\begin{tikzcd}[arrows=Rightarrow]
&
|[alias=GUFU]|GUFU
\arrow[dr, "GU \eps"]
\arrow[d, "gUFU"]
&
&
\\
GU
\arrow[ur, "G \eta U"]
\arrow[d, "gU", swap]
\arrow[rr, shiftarr={yshift=24ex}, "1_{GU}",
""{middle, name=beg1}]
\arrow[r, phantom, "\Rrightarrow"{near start,
rotate=225},
"{g_{\eta U}}^{-1}"{near end}]
&
|[alias=KFUFU]|KFUFU
\arrow[rd, "KFU \eps"]
\arrow[r, phantom, "\Rrightarrow"{near start,
rotate=225},
"{g_{U \eps}}^{-1}"{near end}]
&
|[alias=GU]|GU
\arrow[r, "k"]
\arrow[d, "gU", swap]
&
K
\\
KFU
\arrow[rr, "1_{KFU}", swap, ""{middle, name=fin2}]
\arrow[ur, "KF \eta U"]
&
&
KFU
\arrow[ur, "K \eps", swap, ""{middle, name=fin1}]
&
\arrow[from=GU, to=fin1, phantom,
"\Rrightarrow"{rotate=-45, below},
"\lift{\omega}"{above}]
\arrow[from=beg1, to=GUFU, phantom,
"\Rrightarrow"{rotate=-90, near end},
"{Gt}^{-1}"{near
start}]
\arrow[from=KFUFU, to=fin2, phantom,
"\Rrightarrow"{rotate=-90},
"KFt"{near
start}]
\end{tikzcd}
\end{equation}
\begin{equation}
\begin{tikzcd}[arrows=Rightarrow]
&
|[alias=GUFU]|GUFU
\arrow[dr, "kFU", ""{middle, name=beg2}]
\arrow[d, "gUFU", swap]
\arrow[r, "GU \eps"]
&
GU
\arrow[dr, "k"]
\arrow[d, phantom, "\Rrightarrow"{rotate=-90},
"{k_\eps}^{-1}"{near start}]
&
\\
GU
\arrow[ur, "G \eta U"]
\arrow[d, "gU", swap]
\arrow[rru, shiftarr={yshift=16ex}, "1_{GU}",
""{middle, name=beg1}]
\arrow[r, phantom, "\Rrightarrow"{near start,
rotate=225}, "{g_{\eta U}}^{-1}"{near end}]
&
|[alias=KFUFU]|KFUFU
\arrow[rd, "KFU \eps", swap]
\arrow[r, "K \eps FU", swap]
\arrow[from=beg2, phantom, "\Rrightarrow"{rotate=225,
above},
"\lift{\omega} FU"{above}]
&
|[alias=KFU]|KFU
\arrow[r, "K \eps", swap]
\arrow[d, phantom, "\Rrightarrow"{rotate=-90},
"K\eps_\eps"{near start}]
&
K
\\
KFU
\arrow[rr, "1_{KFU}", swap, ""{middle, name=fin2}]
\arrow[ur, "KF \eta U"]
&
&
KFU
\arrow[ur, "K \eps", swap, ""{middle, name=fin1}]
&
\arrow[from=beg1, to=GUFU, phantom,
"\Rrightarrow"{rotate=-90, near end},
"{Gt}^{-1}"{near
start}]
\arrow[from=KFUFU, to=fin2, phantom,
"\Rrightarrow"{rotate=-90},
"KFt"{near
start}]
\end{tikzcd}
\end{equation}
\begin{equation}
\begin{tikzcd}[arrows=Rightarrow]
&
|[alias=GUFU]|GUFU
\arrow[dr, "kFU", ""{middle, name=beg2}]
\arrow[d, "gUFU", swap]
\arrow[r, "GU \eps"]
&
GU
\arrow[dr, "k"]
\arrow[d, phantom, "\Rrightarrow"{rotate=-90},
"{k_\eps}^{-1}"{near start}]
&
\\
GU
\arrow[ur, "G \eta U"]
\arrow[d, "gU", swap]
\arrow[rru, shiftarr={yshift=16ex}, "1_{GU}",
""{middle, name=beg1}]
\arrow[r, phantom, "\Rrightarrow"{near start,
rotate=225}, "{g_{\eta U}}^{-1}"{near end}]
&
|[alias=KFUFU]|KFUFU
\arrow[r, "K \eps FU", swap]
\arrow[from=beg2, phantom, "\Rrightarrow"{rotate=225,
above},
"\lift{\omega} FU"{above}]
&
|[alias=KFU]|KFU
\arrow[r, "K \eps", swap]
&
K
\\
KFU
\arrow[ur, "KF \eta U"]
\arrow[rrru, bend right, "1_{KFU}", swap, ""{middle,
name=fin1}]
&
&
&
\arrow[from=beg1, to=GUFU, phantom,
"\Rrightarrow"{rotate=-90, near end},
"{Gt}^{-1}"{near
start}]
\arrow[from=KFUFU, to=fin1, phantom,
"\Rrightarrow"{rotate=-90, near end},
"{KsU}^{-1}"]
\end{tikzcd}
\end{equation}
\begin{equation}
\begin{tikzcd}
GU
\arrow[Rightarrow, d, "gU", swap]
\arrow[Rightarrow, r, "1_{GU}"]
&
GU
\arrow[Rightarrow, d, "G \eta U"]
\arrow[Rightarrow, dd, shiftarr={xshift=15ex},
""{middle, name=beg}, "1_{GU}"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=180,
above},
"\omega U"{above}]
\\
KFU
\arrow[Rightarrow, d, "K \eps", swap]
&
|[alias=GUFU]|GUFU
\arrow[Rightarrow, l, "kFU"]
\arrow[Rightarrow, d, "GU \eps"]
\arrow[dl, phantom, "\Rrightarrow"{rotate=135,
above},
"{k_\eps}^{-1}"{above}]
\\
K
&
GU
\arrow[Rightarrow, l, "k"]
\arrow[phantom, from=beg, to=GUFU,
"\Rrightarrow"{rotate=180, above}, "Gt^{-1}"{above}]
\end{tikzcd}
\end{equation}
The proof is therefore complete.
\end{proof}
\begin{remark}
Having proved the implications (1) $\implies$ (2) and (3) $\implies$ (1), the proof of Theorem~\ref{thm:pseudo-benabou} is complete.
\end{remark}
|
\section{Introduction}
\medskip
Let us begin with an overview of values for the strong coupling,
$\alpha_s(m_\tau^2)$ obtained in the most recent analyses of hadronic
$\tau$ decays. Taking the averages of ``FOPT'' and ``CIPT'' values,\footnote{FOPT stands for fixed-order perturbation theory, and CIPT for
contour-improved perturbation theory; for a review, see for instance Ref.~\cite{jamin2005}.
For comparison we use these averaged values in this section; however, we do not recommend using such averages in general, as this averaging is without theoretical basis.}
the most recent values obtained from ALEPH data \cite{aleph,davier2014} are
\begin{eqnarray}
\label{alphasvalues}
\alpha_s(m_\tau^2)&=&0.332(12)\ ,\qquad\cite{davier2014}\ ,\\
&=&0.328(12)\ ,\qquad\cite{pich2016}\ ,\nonumber\\
&=&0.301(10)\ ,\qquad\cite{boito2015}\ ,\nonumber\\
&=&0.309(9)\ ,\qquad\ \,\cite{boito2015}\ \mbox{including OPAL data \cite{opal,boito2012b}}\ .\nonumber
\end{eqnarray}
While the values of Refs.~\cite{davier2014} and \cite{pich2016} are in
good agreement, there is a significant discrepancy with the values from Ref.~\cite{boito2015}. It is interesting to also quote some values from the PDG \cite{pdg}, adjusted to the $\tau$ mass:
\begin{eqnarray}
\label{pdgvalues}
\alpha_s(m_\tau^2)&=&0.325(15)\ ,\qquad\mbox{from $\tau$ decays}\ ,\\
&=&0.315(9)\ ,\qquad\ \,\mbox{from lattice QCD}\ ,\nonumber\\
&=&0.314(9)\ ,\qquad\ \,\mbox{world average without $\tau$}\ .\nonumber\end{eqnarray}
While these values are nominally in agreement with each other, the value from
$\tau$ decays is on the high side, and this is clearly related to the higher values in Eq.~(\ref{alphasvalues}), which push up the $\tau$-based PDG average.
Two natural questions arise regarding the average from $\tau$ decays: (i) is the average actually reliable, and (ii) can we do better? We address these questions in what follows.
As we will see, the differences originate in the way the non-perturbative
``contamination'' afflicting the determination of $\alpha_s$ from $\tau$ decays
is treated in the analysis. While the value of the $\tau$ mass is large enough
that one may attempt to apply QCD perturbation theory, it is sufficiently low that
such a determination is not free from non-perturbative effects.
Here, we will compare the analysis methods used in Refs.~\cite{davier2014,pich2016,boito2015}, and argue that the different assumptions
underlying these methods lead to the different values reported in Eq.~(\ref{alphasvalues}). We will argue that the method used in Refs.~\cite{davier2014,pich2016} is unreliable, while the method used in Ref.~\cite{boito2015} appears to be consistent with the data. However, also Ref.~\cite{boito2015} employs an assumption
in order to deal with the non-perturbative contamination. While tests based on
ALEPH and OPAL data support the validity of this assumption, more precise
data, as for instance obtained from Belle and Belle-II, would make it possible to subject also the method of Ref.~\cite{boito2015} to more stringent
consistency tests.
\section{Background}
\medskip
The ratio of the decay width for non-strange hadronic $\tau$ decays to the
width for $\tau\to\nu_\tau e\overline\nu_e$ can be calculated in QCD
perturbation theory, and takes the form\footnote{In FOPT.}
\begin{equation}
\label{Rtau}
R^{ud}_\tau\equiv\frac{\Gamma(\tau\to\mbox{hadrons}_{ud})}{\Gamma(\tau\to\nu_\tau e\overline\nu_e)}=3S_{\rm EW}|V_{ud}|^2\left(1+\frac{\alpha_s}{\pi}+
5.2\left(\frac{\alpha_s}{\pi}\right)^2+\dots\right)\ ,
\end{equation}
where $S_{\rm EW}$ is a short-distance electroweak correction. The
perturbative series has been calculated to order $\alpha_s^4$ \cite{baikov2008}.
In principle, this result can be used to determine $\alpha_s$ from the
vector ($V$) and axial-vector ($A$) non-strange decays, for which
ALEPH and OPAL obtained relatively precise data. However, at $\alpha_s(m_\tau^2)/\pi\approx 0.1$, the series is rather slowly converging, indicating that
QCD perturbation theory may not be all there is at the $\tau$ mass. On the
experimental side, what one sees is not $\tau\to\nu_\tau\mbox{jets}$, but
rather (in the $V$ channel) $\tau\to\nu_\tau\rho^*\to\nu_\tau\mbox{pions}$
with $\rho^*=\rho(770)$, $\rho(1450)$, {\it etc.}, calling into question the
relation between observed hadronic $\tau$ decays and the perturbative
regime.
\begin{figure}[t]
\includegraphics[width=14pc]{cauchy.pdf}\hspace{2pc}%
\begin{minipage}[b]{14pc}\caption{\label{contour}{\it Contour used in Eq.~(\ref{cauchy}).}}
\end{minipage}
\end{figure}
The finite-energy sum-rule (FESR) approach allows us to look into this
more systematically \cite{braaten1992}. First, using the optical theorem,
$R^{ud}_\tau$ can be expressed in terms of the $I=1$ vacuum polarization
$\Pi_{V+A}(q^2)$ as \cite{tsai1971}
\begin{equation}
\label{Rtauvacpol}
R^{ud}_\tau=12\pi^2 S_{\rm EW}|V_{ud}|^2\int_0^{m_\tau^2}\frac{ds}{m_\tau^2}
\left(1-\frac{s}{m_\tau^2}\right)^2\left(1+\frac{2s}{m_\tau^2}\right)\rho_{V+A}(s)\ ,
\end{equation}
where $\rho_{V+A}(s)=\frac{1}{\pi}\mbox{Im}\ \Pi_{V+A}(s)$ is the non-strange
inclusive $V+A$ spectral function measured in hadronic $\tau$ decays. Since
in this talk we will always consider the case $V+A$ we drop the
subscript $V+A$ from now on. Second, Cauchy's theorem, applied to
the contour shown in Fig.~\ref{contour}, implies that\footnote{The same equation holds for $V$ and $A$ channels separately.}
\begin{equation}
\label{cauchy}
\int_0^{s_0}ds\,w(s)\,\rho(s)=-\frac{1}{2\pi i}\oint_{|z|=s_0}dz\,w(z)\,\Pi(z)\ ,
\end{equation}
where $w(s)$ can be any analytic function. The idea is now to split $\Pi(z=q^2)$ into a perturbative part $\Pi^{\rm pert}$ and non-perturbative parts,
\begin{equation}
\label{Piz}
\Pi(z)=\Pi^{\rm pert}(z)+\Pi^{\rm OPE}(z)+\Pi^{\rm DV}(z)\ ,
\end{equation}
and substitute this into the contour integral on the right-hand side of
Eq.~(\ref{cauchy}).
Here $\Pi^{\rm OPE}$ is given by the operator product expansion (OPE),\footnote{The coefficients $C_{2k}$ are logarithmically dependent on $q^2$,
but this dependence can be numerically neglected in the application to $\tau$ decays \cite{boito2012}. For the non-strange channel, $C_2$ is negligibly
small, being proportional to the squares of the light quark masses.}
\begin{equation}
\label{OPE}
\Pi^{\rm OPE}(q^2)=\frac{C_4}{q^4}-\frac{C_6}{q^6}+\frac{C_8}{q^8}+\dots\ ,
\end{equation}
and $\Pi^{\rm DV}$ represents, by definition, the non-perturbative part not
encapsulated by the OPE.\footnote{In the literature, the perturbative part is often defined to be the leading-order term of the OPE. Here we will use ``OPE'' to refer to power-correction terms (\ref{OPE}) only.} Neglecting this latter part,
and truncating the OPE, for $w(x)$ a polynomial of degree $N$
the right-hand side then depends on a finite number of parameters, $\alpha_s$
and $C_{2k}$ for $k=2,\ 3,\ \dots,\ k_{max}=2N+2$. These parameters can then be fit to the integral
on the left-hand side, computed from the data for the spectral function, as a function of $s_0$.
The physics not represented by the OPE is the physics of resonances, which
are visible in the data along the positive real $z=q^2$ axis in the spectral
function, {\it cf.} Sec.~3 below. In other words, the OPE does not provide
a reliable representation of the non-perturbative part of $\Pi(z)$ in the region
where the circle with radius $s_0$ crosses the positive real axis in Fig.~\ref{contour}. As we will see below, this leads to the need to introduce a
non-trivial {\it ansatz} for $\Pi^{\rm DV}(z)$, in order to take into account the
resonance contributions to $\Pi(q^2)$. Under mild conditions on
$\rho^{\rm DV}(s)=\frac{1}{\pi}\mbox{Im}\ \Pi^{\rm DV}(s)$, Eq.~(\ref{cauchy})
can be rewritten as \cite{cata2005,cata2008}
\begin{equation}
\label{fesr}
\int_0^{s_0}ds\,w(s)\,\rho(s)
=-\frac{1}{2\pi i}\oint_{|z|=s_0}dz\,w(z)\,\left(\Pi^{\rm pert}(z)+\Pi^{\rm OPE}(z)\right)
-\int_{s_0}^\infty ds\,w(s)\,\rho^{\rm DV}(s)\ .
\end{equation}
If we supply a parametrization for $\rho^{\rm DV}(s)$ in terms of a finite
number of parameters, this constitutes the FESR we will apply to determine
$\alpha_s(m_\tau^2)$ (one of the parameters of the ``theory''
part on the right-hand side) from the experimental data, which allow us to
compute the integrals on the left-hand side.\footnote{The parameters in
$\Pi^{\rm OPE}$ and $\Pi^{\rm DV}$ can be viewed as ``nuisance'' parameters
in these fits, although they are of physical interest in their own right.}
The representation on the right-hand side is valid for $s_0$ sufficiently large,
and the hope is that it already applies for $s_0$ (not too far) below $m_\tau^2$.
\section{Data}
\medskip
Plots showing the $V$ and $A$ non-strange spectral functions can be
found in Fig.~5 of Ref.~\cite{davier2014} and Fig.~5 of Ref.~\cite{opal}.
First, these figures suggest that, short of a complete determination of
the inclusive spectral functions from Belle and Belle-II data, a
determination of the $4\pi$ decay modes might already significantly reduce
the error on the inclusive spectral functions if combined with the ALEPH
data, given the vastly increased statistics of the Belle experiments. Of
course, a complete determination of the inclusive spectral function from
Belle data is highly desirable as well, given the much higher statistics, and would avoid
potential pitfalls with combining data from different experiments.
\begin{figure}[t]
\includegraphics[width=20pc]{CIPT_V+A-1.pdf}\hspace{2pc}%
\begin{minipage}[b]{16pc}\caption{\label{blowup}{\it Blow-up of the ALEPH data (red data points) in the
large-$s$ region of the $V+A$ non-strange spectral function
($s$ in {\rm GeV}$^2$).
Black dashed line: perturbation theory (CIPT) contributions to the spectral function. The $\alpha_s$-independent parton-model contribution to $\rho(s)$ has been subtracted from both.}}
\end{minipage}
\end{figure}
Second, it is interesting to consider the data from the point of view of
what they indicate about non-perturbative effects, in particular, about
resonance effects. In Fig.~\ref{blowup}, we show the high-$s$ part of the $V+A$ non-strange spectral function measured by ALEPH.
What is shown in this figure is $2\pi^2\rho_{V+A}-1$, {\it i.e.},
the dynamical QCD contribution to the spectral distribution, subtracting the
contribution from QCD without gluons (the parton model). The
black, dashed curve shows the theoretical representation of the spectral
function without ``duality violations,''
obtained from the right-hand side of Eq.~(\ref{fesr}) setting $\rho^{\rm DV}=0$.
The figure shows that the oscillations of the data, due to the presence of
resonances, are not small compared to the contribution from QCD perturbation
theory, and thus need to be taken into account when fitting the data using
Eq.~(\ref{fesr}).
We will turn to the analysis of the data using the two approaches, the
``truncated-OPE'' approach and the ``DV-model'' approach, next.
\section{Truncated-OPE approach}
\medskip
In the truncated-OPE approach \cite{diberder1992}, employed in Refs.~\cite{davier2014,pich2016}, one always takes $s_0=m_\tau^2$, and
employs the weights ($x=s/s_0$)
\begin{eqnarray}
\label{ALEPHweights}
w_{00}(x)&=&(1-x)^2(1+2x)\ ,\\
w_{10}(x)&=&(1-x)^3(1+2x)\ ,\nonumber\\
w_{11}(x)&=&(1-x)^3(1+2x)x\ ,\nonumber\\
w_{12}(x)&=&(1-x)^3(1+2x)x^2\ ,\nonumber\\
w_{13}(x)&=&(1-x)^3(1+2x)x^3\ .\nonumber
\end{eqnarray}
One assumes that only $C_{4,6,8}$ are non-negligible, and that duality violations, represented
by $\rho^{\rm DV}$ in Eq.~(\ref{fesr}) can be neglected. We note that the
weights in Eq.~(\ref{ALEPHweights}) all suppress the region near $s=s_0$
by a double or triple zero, which supports (but does not necessarily validate) the assumption that duality violations can be
neglected. However, since, from the residue theorem,
\begin{equation}
\label{residue}
\frac{1}{2\pi i}\oint dz\,z^n\,\frac{C_{2k}}{z^k}=C_{2n+2}\,\delta_{k,n+1}\ ,
\end{equation}
we see that, using the weights (\ref{ALEPHweights}), which have values of
$n$ up to $7$, we would need to keep terms in the OPE up to $C_{16}$,
since $2k=2(n+1)=16$ for $n=7$. The truncated-OPE approach thus
neglects, without theoretical foundation, the OPE coefficients
$C_{10,12,14,16}$. Of course, the practical reason for doing this is that
at $s_0=m_\tau^2$, only 5 data integrals are available, so that a maximum
of 4 parameters, $\alpha_s$, $C_4$, $C_6$ and $C_8$, can be fit.
The truncated-OPE approach thus has two shortcomings: the truncation of
the OPE, and the neglect of resonance oscillations around perturbation
theory plus the OPE clearly visible in the $V+A$ spectral function, {\it cf.}
Fig.~\ref{blowup}.
\begin{figure}[t]
\begin{minipage}{18pc}
\includegraphics[width=18pc]{CIPT-VpA-SFmatch.pdf}
\end{minipage}\hspace{2pc}%
\begin{minipage}{18pc}
\includegraphics[width=18pc]{Fake-V+A_CIPT.pdf}
\end{minipage}
\caption{\label{fake}{\it Left figure: true data, with the model curve for $s\ge 1.55$~{\rm GeV}$^2$. Right figure: fake data for $s\ge 1.55$~{\rm GeV}$^2$,
true data for $s<1.55$~{\it GeV}$^2$.}}
\end{figure}
This approach can be tested, using a set of fake data based on a model with
known values for the parameters \cite{boito2016}. Using a fit from Ref.~\cite{boito2015}, which describes the $V$, $A$ and thus $V+A$ spectral
functions very well, one can use this fit as a model to generate fake data
from a gaussian distribution with the model as central values, and the
real-world, ALEPH covariance matrix. Since the model is only valid for
$s\ge 1.55$~GeV$^2$, we generate fake data only in this region, and use
the true data for $s<1.55$~GeV$^2$. The model has $\alpha_s(m_\tau^2)=0.312$ (CIPT), and known OPE coefficients (for details, we refer to
Ref.~\cite{boito2016}).
\begin{table}[h!]
\hspace{0cm}\begin{tabular}{|c|c|c|c|c|c|}
\hline
ALEPH &
$\alpha_s(m_\tau^2)$ & $C_{4}$ (GeV$^4$) & $C_{6}$ (GeV$^6$)& $C_{8}$ (GeV$^8$)& $\chi^2/$dof \\
\hline
true values & 0.312 & 0.0027 & -0.013 & 0.035 & \\
\hline
fake data fit & 0.334(4) & -0.0024(4) & 0.0007(3) & -0.0008(4) & 0.95/1 \\
\hline
\end{tabular}
\caption{\label{table1}{\it CIPT fits employing the truncated-OPE approach
on the fake data, based on the weights of Eq.~(\ref{ALEPHweights}).
By assumption, $C_{10}=C_{12}=C_{14}=C_{16}=0$.
Errors are statistical only.}}
\end{table}%
In Table~\ref{table1} we show the results of fitting the fake data with the
truncated-OPE approach. We see that the truncated-OPE approach gets
it wrong, and leads to a systematic overestimate of the value of $\alpha_s(m_\tau^2)$, despite an excellent value for the $\chi^2$ per degree of freedom.
In addition, we also note that there is a large difference between the fitted
and true values of the OPE coefficients: clearly, the fit is trying to compensate
for the ``missing'' contributions from the higher-dimension OPE coefficients
and/or duality violations that have, by fiat, been set equal to zero.
Reference~\cite{pich2016} considered many other choices of sets of
weights, in addition to the ``ALEPH'' weights shown in Eq.~(\ref{ALEPHweights}). In particular,
another set of ``optimal'' weights was considered,
given by
\begin{eqnarray}
\label{optimal}
w^{(2,1)}(x)&=&1-3x^2+2x^3\ ,\\
w^{(2,2)}(x)&=&1-4x^3+3x^4\ ,\nonumber\\
w^{(2,3)}(x)&=&1-5x^4+4x^5\ ,\nonumber\\
w^{(2,4)}(x)&=&1-6x^5+5x^6\ ,\nonumber\\
w^{(2,5)}(x)&=&1-7x^6+6x^7\ .\nonumber
\end{eqnarray}
These weights differ from the ALEPH weights of Eq.~(\ref{ALEPHweights})
in that no linear term in $x$ appears in these weights. Therefore, they do
not involve $C_4$, allowing one to fit also $C_{10}$ using the same strategy.
This reduces the reliance on truncation of the OPE by one order, although
the OPE terms proportional to $C_{12}$, $C_{14}$ and $C_{16}$ still need
to be truncated in this approach.
\begin{table}[h!]
\hspace{0cm}\begin{tabular}{|c|c|c|c|c|c|}
\hline
optimal &
$\alpha_s(m_\tau^2)$ & $C_{6}$ (GeV$^6$) & $C_{8}$ (GeV$^8$)& $C_{10}$ (GeV$^{10}$)& $\chi^2/$dof \\
\hline
true values & 0.312 & 0.0027 & -0.013 & 0.035 & \\
\hline
fake data fit & 0.334(4) & 0.0008(4) & -0.0008(5) & 0.0001(3) & 0.92/1 \\
\hline
\end{tabular}
\caption{\label{table2}{\it CIPT fits employing the truncated-OPE approach
on the fake data, based on the weights of Eq.~(\ref{optimal}).
By assumption, $C_{12}=C_{14}=C_{16}=0$.
Errors are statistical only.}}
\end{table}%
Table~\ref{table2} shows the result of the fake-data test. The conclusion is the
same as before: while the fit results are completely consistent between Tables~\ref{table1} and \ref{table2}, they again disagree with the true model values.
A much more detailed analysis of this problem was given in Ref.~\cite{boito2016}, with the conclusion that the truncated-OPE approach is
inherently unreliable, and thus should no longer be used.
\section{DV-model approach}
\medskip
Our conclusion is that the truncation of the OPE, employed in the approach
described in the previous section, leads to a systematic bias of the value for
$\alpha_s$ (as well as the OPE coefficients) obtained in that approach.
Already based on earlier investigations raising doubts about the truncated-OPE
approach \cite{maltman2008,cata2009}, we
employed a simpler choice of weights,\footnote{Note that the sets of weights considered here share $w_{00}(x)=w^{(2,1)}(x)=w_3(x)$.}
\begin{eqnarray}
\label{simple}
w_0(x)&=&1\ ,\\
w_2(x)&=&1-x^2\ ,\nonumber\\
w_3(x)&=&(1-x)^2(1+2x)\ .\nonumber
\end{eqnarray}
It follows from Eq.~(\ref{residue}) that only the coefficients $C_6$ and $C_8$
need to be kept in the OPE.\footnote{This assumes no logarithmic corrections
to the OPE coefficients. These are suppressed by powers of $\alpha_s$,
and found to be numerically negligible \cite{boito2012}.} However, the first
weight does not suppress the region $s\approx s_0$ at all, and the second
weight does this only linearly. It is thus even less likely than in the truncated-OPE
approach that duality violations can be ignored. We introduce an {\it ansatz} for
$\rho^{\rm DV}(s)$ in Eq.~(\ref{fesr}); in each of the channels, $V$ and $A$,
we assume that, for large enough $s$,
\begin{equation}
\label{ansatz}
\rho^{\rm DV}(s)=e^{-\delta-\gamma s}\sin{(\alpha+\beta s)}\ ,
\end{equation}
thus introducing 8 new parameters, $\delta_{V,A}$, $\gamma_{V,A}$,
$\alpha_{V,A}$ and $\beta_{V,A}$ to the fits. Clearly, we cannot restrict
ourselves to $s_0=m_\tau^2$; instead, we let $s_0$ vary with a minimum
value $s_{min}$, to be determined by the quality of the fits. It turns out that
the optimal value is $s_{min}\approx 1.55$~GeV$^2$, with good stability
for small variations of this value.
Of course, the use of Eq.~(\ref{ansatz}) implies that this approach is model
dependent. This is unavoidable, since the physics of resonances is not
captured by QCD perturbation theory or the OPE, while it is clearly visible
in the data ({\it cf.} Fig.~\ref{blowup}). The ingredients going into the
construction of the model
are generally thought of as approximate properties of QCD. One starts
with assuming
Regge behavior of the spectrum for $s>s_{min}$,
which, in the large-$N_c$ limit implies that the spectrum in each channel
looks like $M^2(n)=M^2(0)+\sigma n$, with $n=0,\ 1,\ 2,\ \dots$ and
$\sigma\sim 1$~GeV$^2$. Relaxing the large-$N_c$ limit, these poles
pick up an imaginary part leading to a decay width $\Gamma(n)\propto
M(n)/N_c$, providing a more realistic picture of the resonance spectrum
in each channel. A model which incorporates all these features and
satisfies the analyticity constraints of QCD everywhere in the complex
plane exists \cite{cata2005,cata2008,blok1998,bigi1999,golterman2002}, and leads to the {\it ansatz}~(\ref{ansatz}).
While this approach, the ``DV-model'' approach, is model dependent, it should be stressed that also the truncated-OPE approach is model dependent in the same sense: it assumes an even simpler model for duality violations, by
simply setting $\rho^{\rm DV}(s)=0$. Since the data clearly show the presence of duality violations, this choice of model is almost certainly worse than the
choice~(\ref{ansatz}). Pinching, {\it i.e.} the suppression of the region
$s\approx s_0$ by the choice of weights~(\ref{ALEPHweights}) or (\ref{optimal}) may help, but without an explicit
model for duality violations this cannot be checked. As we have shown,
internal consistency between different sets of weights in the truncated-OPE
approach is not sufficient.
Within this framework, many different fits can be performed. One can choose
any subset of the weights~(\ref{simple}), fit only the $V$ channel or the combined $V$ and $A$ channels, or just the sum $V+A$. A large suite of
fits has been considered in Refs.~\cite{boito2015,boito2012b,boito2012},
applying this DV-model approach to both the OPAL and ALEPH data.
The results of these fits show excellent internal consistency.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=7cm]{CIFO_w0fit_w0_V.pdf}
\hspace{0.5cm}
\includegraphics*[width=7cm]{CIFO_w0fit_spectrum_V.pdf}
\end{center}
\begin{quotation}
\caption{\label{CIFOw0fit} {{\it Left panel:
comparison of $V$-channel, $w_0$ fits to the data. Right panel: comparison of the theoretical
spectral function resulting from this fit with the experimental results.
CIPT fits are shown in red (dashed) and FOPT in blue (solid).
The (much flatter) black curves in the left panel represent the OPE parts of the
fits, {\it i.e.}, the fit results with the DV parts
removed.
The vertical dashed line indicates the location of $s_{\rm min}$.
From Ref.~\cite{boito2015}.
}}}
\end{quotation}
\vspace*{-4ex}
\end{figure}
Figure~\ref{CIFOw0fit} shows an example of the simplest possible fit based on the
DV-model approach, by considering only the $V$ channel, and only the weight
$w_0(x)=1$. There is excellent agreement between the data and the fits,
shown in the left panel. The spectral function following from this fit is shown
in the right panel, along with the ALEPH data. Another example of a fit
following this approach, now involving all weights in Eq.~(\ref{simple}), again
considering only the $V$ channel, is shown in Fig.~\ref{CIFOw023fit}.
\begin{figure}[t!]
\begin{center}
\includegraphics*[width=7cm]{CIFO_w023fit_w0_V.pdf}
\hspace{0.5cm}
\includegraphics*[width=7cm]{CIFO_w023fit_w2_V.pdf}
\vspace{1cm}
\includegraphics*[width=7cm]{CIFO_w023fit_w3_V.pdf}
\hspace{0.5cm}
\includegraphics*[width=7cm]{CIFO_w023fit_spectrum_V.pdf}
\end{center}
\vspace*{-6ex}
\begin{quotation}
\caption{\label{CIFOw023fit} {{\it
Comparison of $V$ channel fits using the weights $w_0$ (upper left panel),
$w_2$ (upper right panel) and $w_3$ (lower left panel). CIPT fits are shown in red (dashed) and
FOPT in blue (solid).
Lower right panel: comparison
of the theoretical spectral function resulting
from this fit with the experimental results.
The black curves (which are much
flatter for the $w_0$ case)
represent the OPE parts of the fits. The vertical dashed line
indicates the location of $s_{\rm min}$. From Ref.~\cite{boito2015}.}}}
\end{quotation}
\vspace*{-4ex}
\end{figure}
For many other fits, also involving the $A$ channel, and many consistency checks, we refer to Ref.~\cite{boito2015}. The results we obtain, from the
ALEPH data, for $\alpha_s$ are
\begin{eqnarray}
\label{results}
\alpha_s(m_\tau^2)&=&0.296(10)\ ,\qquad\mbox{FOPT}\ ,\\
&=&0.310(14)\ ,\qquad\mbox{CIPT}\ .
\end{eqnarray}
The average of these leads to the third value reported in Eq.~(\ref{alphasvalues});
averaging this with the OPAL-based value obtained in Ref.~\cite{boito2012b}
yields the fourth value reported in this equation.
\begin{figure}[t!]
\includegraphics*[width=7.5cm]{udvpa_w00_our_cipt_opepdv_fit_match_oct13_16.pdf}
\hspace{0.0cm}
\includegraphics*[width=7.5cm]{udvpa_w10_our_cipt_opepdv_fit_match_oct13_16.pdf}
\vspace{1cm}
\includegraphics*[width=7.5cm]{udvpa_w11_our_cipt_opepdv_fit_match_oct13_16.pdf}
\hspace{0.0cm}
\includegraphics*[width=7.5cm]{udvpa_w12_our_cipt_opepdv_fit_match_oct13_16.pdf}
\begin{minipage}{14pc}
\vspace{-2cm}
\includegraphics*[width=7.5cm]{udvpa_w13_our_cipt_opepdv_fit_match_oct13_16.pdf}
\end{minipage}\hspace{6pc}
\begin{minipage}[b]{14pc}\caption{\label{ALEPHDVmodel}
{\it Comparison between data and fitted theory for the moments of
Eq.~(\ref{ALEPHweights}). Fits using the DV-model approach.
Shown are differences between the values at $s_0=m_\tau^2$
minus the values at varying $s_0$.}}
\end{minipage}
\vspace*{-4ex}
\end{figure}
A test that is of particular interest is to compare the moments with weights~(\ref{ALEPHweights}) between the fits, which allow us to determine
all OPE coefficients \cite{boito2015}, and the data. These comparisons are
shown in Fig.~\ref{ALEPHDVmodel}. The red points show the data, {\it i.e.},
the left-hand side of Eq.~(\ref{fesr}) (``spectral integrals''), while the black
curves show the fitted theory, {\it i.e.},
the right-hand side of Eq.~(\ref{fesr}) (``OPE$+$DV integrals'').
What is actually shown is the value
of these moments evaluated at $s_0=m_\tau^2$ minus the moment at varying
$s_0$, in order to account for the strong correlations between both the theory integrals at different $s_0$ and the data integrals at different $s_0$.
We note that there is good agreement for all $s_0$
shown in the figures.
\begin{figure}[t!]
\includegraphics*[width=7.5cm]{udvpa_w00_santi_davierfit_ope_vs_specintdiff_mtausqminuss0_oct13_16.pdf}
\hspace{0.0cm}
\includegraphics*[width=7.5cm]{udvpa_w10_santi_davierfit_ope_vs_specintdiff_mtausqminuss0_oct13_16.pdf}
\vspace{1cm}
\includegraphics*[width=7.5cm]{udvpa_w11_santi_davierfit_ope_vs_specintdiff_mtausqminuss0_oct13_16.pdf}
\hspace{0.0cm}
\includegraphics*[width=7.5cm]{udvpa_w12_santi_davierfit_ope_vs_specintdiff_mtausqminuss0_oct13_16.pdf}
\begin{minipage}{14pc}
\vspace{-2cm}
\includegraphics*[width=7.5cm]{udvpa_w13_santi_davierfit_ope_vs_specintdiff_mtausqminuss0_oct13_16.pdf}
\end{minipage}\hspace{6pc}
\begin{minipage}[b]{14pc}\caption{\label{ALEPHtruncatedOPE}
{\it Comparison between data and fitted theory for the moments of
Eq.~(\ref{ALEPHweights}). Fits using the truncated-OPE approach.
Shown are differences between the values at $s_0=m_\tau^2$
minus the values at varying $s_0$.}}
\end{minipage}
\vspace*{-4ex}
\end{figure}
In Fig.~\ref{ALEPHtruncatedOPE} we show the same comparison, but now
using the fit results obtained with the truncated-OPE approach, where all
moments shown were employed in the fit. The black curves now only show
the perturbative plus OPE contribution (``OPE integrals''), because duality
violations, represented by the last term on the right-hand side of Eq.~(\ref{fesr}),
are assumed negligible in the truncated-OPE approach. The agreement between data and
fitted theory is much less good than for the DV-model approach, for which
the corresponding comparisons are shown in Fig.~\ref{ALEPHDVmodel}.
Of course, near $s_0=m_\tau^2$ data and fitted theory agree, because of the
differences which are shown.
Differences between theory and experimental integrals are, of course, by definition, zero at $s_0=m_\tau^2$;
but deviation of the theory curves from the data at lower $s_0$ signals shortcomings in the theoretical representation.
For the moments with weights $w_{12}$
and $w_{13}$, the agreement already deteriorates well above $s_0=2$~GeV$^2$. For more detail, we refer to Ref.~\cite{boito2016}.
\section{Discussion and outlook}
\medskip
The determination of the strong coupling, $\alpha_s$, from hadronic $\tau$
decays is interesting for at least two reasons. First, it provides a
value of $\alpha_s$ at the $Z$ mass which is quite competitive in comparison
with other methods. The reason for this is that because of the running, relatively modest errors get reduced when the obtained values for $\alpha_s$ are converted to values at the $Z$ mass. Second, because it is at present the
only determination with rather high precision at a rather low energy, it provides a nice test of the running of the strong coupling predicted by QCD. With the presently
available precision, the QCD prediction is nicely confirmed, as indicated, for
instance, by the values shown in Eq.~(\ref{pdgvalues}).
However, as we have discussed in detail, the determination from $\tau$ decays
is afflicted by contamination from non-perturbative effects, because of the rather
low value of the $\tau$ mass. Some of these non-perturbative effects are captured by the OPE, but not all: resonance effects are clearly important, and
appear as significant deviations from the quark-gluon picture underlying
perturbation theory and the OPE. Another, probably related, issue is the
discrepancy between values obtained using CIPT or FOPT, which correspond to two different resummation schemes.\footnote{Other resummation schemes are possible, and have been considered.}
Two methods for extracting $\alpha_s$ from $\tau$ decays have been pursued.
One of these, the truncated-OPE approach, has now been shown to fail; it cannot, in general, account for all non-perturbative effects that afflict the analysis. The
other approach is based on a model for duality violations, and has thus far been shown to be internally consistent, and consistent with the data. However, it is
important to test this approach more stringently, and much more precise data
for the $V$ channel would provide the possibility to carry out such tests, and,
if successful, lead to a possible reduction in the systematic error.\footnote{For other ideas on possible tests, see Ref.~\cite{caprini2014}.} With thousands times more $\tau$ pairs, inclusive spectral functions obtained from
Belle and Belle-II data would almost certainly have a significant impact on our understanding of QCD from hadronic $\tau$ decays. Even a precise determination of the $\tau\to\nu_\tau 4\pi$ channels might already have an
impact, as the large errors in the large-$s$ region on the currently available $V$ spectral functions are dominated by these channels. In this respect, we believe that it will suffice to focus on the non-strange $V$ channel, since, contrary to popular belief, there is no indication that a higher
precision can be obtained from $V+A$. It is not unlikely that the
asymptotic regime where the right-hand side of Eq.~(\ref{fesr}) is valid sets in
at lower values of $s$ in the $V$ than in the $A$ channel.
It is more difficult to make progress on the optimal resummation scheme.
The typical difference between values of $\alpha_s(m_\tau^2)$ obtained
from FOPT or CIPT is always about 5\%, with CIPT providing the higher
value, irrespective of the analysis approach employed. Theoretical
considerations based on a renormalon analysis somewhat favor FOPT
\cite{beneke2008,beneke2013}, but other resummation schemes cannot
be ruled out \cite{caprini2009}. A direct comparison between the value of $\alpha_s$
from $\tau$ decays and from the lattice, for instance, tends to favor
the CIPT-based value. Again, higher precision determinations based on
higher precision data may contribute to a deeper understanding.
\ack
\medskip
MG would like to thank the organizers of the Mini Workshop on Tau Physics
for the opportunity to present this work, and for warm hospitality.
The work of DB is supported by the S{\~a}o Paulo Research Foundation
(FAPESP) Grant No. 2015/20689-9 and by CNPq Grant No. 305431/2015-3.
This material is based
upon work supported by the U.S. Department of Energy, Office of Science,
Office of High Energy Physics, under Award Number DE-FG03-92ER40711
(MG).
KM is supported by a grant from the Natural Sciences and Engineering Research
Council of Canada. SP is supported by CICYTFEDER-FPA2014-55613-P, 2014-SGR-1450.
\section*{References}
|
\section{Introduction}
The aim of this paper is to give a systematic construction of a distinguished class of Iwanaga-Gorenstein algebras, called minimal Auslander-Gorenstein algebras, which are defined by the property
\[\operatorname{idim} A=d+1=\operatorname{domdim} A\]
for some positive integer $d$, where $\operatorname{idim} A$ and $\operatorname{domdim} A$ are the self-injective dimension and dominant dimension of $A$ respectively.
This class includes higher Auslander algebras which are defined by replacing $\operatorname{idim} A$ with the global dimension $\operatorname{gldim} A$.
These algebras appear naturally in the higher Auslander-Reiten theory on $d$-precluster-tilting subcategories \cite{Iya2,IS}.
Our starting point is a famous article of Kunio Yamagata \cite{Yam} on Frobenius algebras, where the following construction was introduced:
For a given finite dimensional algebra $A=A^{[0]}$, we define a series of finite dimensional algebras by
\[A^{[m+1]}:=\operatorname{End}_{A^{[m]}}(A^{[m]}\oplus DA^{[m]}).\]
We call $A^{[m]}$ the \emph{$m$-th SGC extension} of $A$.
At first glance, one would expect that the algebras $A^{[m]}$ become complicated very quickly as $m$ increases.
Nevertheless, Yamagata observed that in the case when $A$ is hereditary, every $A^{[m]}$ has finite global dimension and large dominant dimension - although he did not calculate their explicit values.
One of our main results gives explicit formulae for these dimensions in the case of $A$ being hereditary.
\begin{theorem}\label{first main}
Suppose $A$ is a hereditary algebra.
\begin{enumerate}
\item If $A$ is representation-infinite, then
$\operatorname{gldim} A^{[m]}=2m+1$ and $\operatorname{domdim} A^{[m]}=2m$.
\item If $A$ is representation-finite, then $\operatorname{gldim} A^{[m]}=\operatorname{domdim} A^{[m]}$ holds for infinitely many $m$.
\end{enumerate}
\end{theorem}
The majority of Theorem \ref{first main} is proved by considering the same problem on the \emph{replicated algebras} of $A$.
The $m$-th replicated algebra of $A$ is defined as a $(m+1)\times (m+1)$ matrix algebra:
\[A^{(m)}=\left(\begin{array}{ccccc}
A & 0 & 0 & \cdots & 0 \\
DA & A & 0 & & 0 \\
\vdots & \ddots & \ddots & \ddots & \vdots \\
0 & & \ddots & \ddots &\vdots \\
0 & 0 & \cdots & DA & A
\end{array}\right).
\]
Such construction of algebras was studied in, for example, \cite{AssIwa,ABST,LvZh}; but it seems to us that its connection with SGC extensions (Theorem \ref{thm-SGC-replicate}) has not been exploited before; in fact, these two constructions of new algebras from $A$ coincide under a mild assumption.
\begin{theorem}\label{comparison}
Let $A$ be a finite dimensional algebra such that $\operatorname{Hom}_A(DA,A)=0$.
Then the $m$-th SGC extension $A^{[m]}$ is Morita equivalent to the $m$-th replicated algebra $A^{(m)}$.
\end{theorem}
It is usually easier to understand the behaviour of $A^{(m)}$ compared to that of $A^{[m]}$.
As such, the majority of the calculations carried out in this paper are on the replicated algebras.
In fact, we can find the formulae for the dominant dimension and the injective dimension of each indecomposable projective $A^{(m)}$-module in terms of certain combinatorial data (including the Coxeter numbers; see Corollaries \ref{cor-RF} and \ref{cor-RI}).
Theorem \ref{comparison} enables us to deduce Theorem \ref{first main} from these calculations for $A^{(m)}$, apart from one particular case (namely, when $A$ is a Nakayama algebra).
In fact, our main result is much more general, and is stated for a much wider class of algebras - we call them \emph{Serre-formal algebras}.
Recall that the perfect derived category of an Iwanaga-Gorenstein algebra $A$ has a Serre functor $\nu$.
We call an Iwanaga-Gorenstein algebra $A$ \emph{Serre-formal} if $\nu^i(A)$ is isomorphic to a direct sum of stalk complexes for any $i\in\mathbb{Z}$ in the bounded derived category $\mathrm{D}^{\mathrm{b}}(\mod A)$ of finitely generated $A$-modules.
This class of algebras is closed under tensor products (Proposition \ref{prop-tensor}), and contains several important subclasses which are central in representation theory.
For example, self-injective algebras and higher hereditary algebras - such as path algebras of finite acyclic quivers and Beilinson algebras - are all Serre-formal.
Moreover, algebras arising from symmetrisable Cartan matrices \cite{GLS}, and Ringel's canonical algebras \cite{Rin} as well as higher canonical algebras of \cite{HIMO} (Theorem \ref{thm-canonical}), are also Serre-formal.
From the perspective of this paper, the main advantage of considering Serre-formal algebras is that we can determine the explicit formulae for the dominant and injective dimensions of each indecomposable projective $A^{(m)}$-module (Theorem \ref{thm-dims2}).
Hence, we can obtain a replicated algebra version of Theorem \ref{first main} for Serre-formal algebras.
We call an algebra $A$ \emph{periodically Serre-formal} if for each indecomposable projective $A$-module $P$, there exists a positive integer $\ell$ so that $\nu^{-\ell}(P)$ is also projective.
It is easy to see that in the case of $A$ being hereditary, this is equivalent to $A$ being representation-finite.
Moreover, recall from \cite{HerIya} that an algebra $A$ is \emph{twisted fractionally Calabi-Yau} if there is a pair of integers $(m\neq 0,n)$ so that $\nu^m(A)\cong A[n]$ in $\mathrm{D}^{\mathrm{b}}(\mod A)$.
In such a case, we say that the twisted Calabi-Yau dimension of $A$ is $(m,n)$ if $m$ is the smallest positive integer for which such an isomorphism holds.
We will show the following equivalences (see Proposition \ref{prop-twistedCY}, Corollary \ref{cor-minAG2}).
\begin{theorem}\label{second main}
Let $A$ be a finite dimensional algebra which is Serre-formal.
Then the following conditions are equivalent.
\begin{itemize}
\item $A$ is periodically Serre-formal.
\item $A$ is twisted fractionally Calabi-Yau.
\item $A^{(m)}$ is minimal Auslander-Gorenstein for some (or infinitely many) positive integer $m$.
\end{itemize}
In such a case, let $(h,s)$ be the twisted Calabi-Yau dimension of $A$, then $A^{(m)}$ being minimal Auslander-Gorenstein is equivalent to $m=ht-1$ for some integer $t>0$.
Moreover, we have
\[
\operatorname{idim} A^{(m)} = (h+s)t-1 =\operatorname{domdim} A^{(m)}.
\]
\end{theorem}
In the special case of higher hereditary algebras, we have the following result (see Corollary \ref{cor-dim-dhered}, Proposition \ref{prop-RF-d}).
We refer the reader to Subsection \ref{subsec-NF-Bkgd} and Subsection \ref{subsec-NF-d-hered} for details on the terminologies.
\begin{theorem}\label{first main-d}
Let $A$ be a $d$-hereditary algebra with $d\ge1$.
\begin{enumerate}
\item If $A$ is $d$-representation-infinite, then
$\operatorname{gldim} A^{(m)}=(d+1)m+d$ and $\operatorname{domdim} A^{(m)}=(d+1)m$.
\item If $A$ is $d$-representation-finite, then $\operatorname{gldim} A^{(m)}=\operatorname{domdim} A^{(m)}$ holds for infinitely many $m$.
\end{enumerate}
\end{theorem}
The structure of the paper is as follows.
In Section \ref{sec-Repl-Alg}, we present some elementary properties of $m$-th replicated algebras $A^{(m)}$ of arbitrary finite dimensional algebra $A$.
In Section \ref{sec-Hered-base}, we give more in-depth studies - in particular, the homological dimensions - of $A^{(m)}$ in the case when $A$ is hereditary.
In Section \ref{sec-Nu-formal}, we introduce Serre-formal algebras, study their basic properties (Subsection \ref{subsec-NF-Bkgd}), and give basic examples (Subsections \ref{subsec-NF-naka}, \ref{subsec-NF-canonical}).
Moreover, we investigate $A^{(m)}$ for the case when $A$ is Serre-formal in a similar manner as Section \ref{sec-Hered-base}, which yields Theorem \ref{second main} (Subsection \ref{subsec-NF-dims}) and Theorem \ref{first main-d} (Subsection \ref{subsec-NF-d-hered}).
In Section \ref{sec-SGC}, we explain the connection between the $m$-th replicated algebra $A^{(m)}$ and the $m$-th SGC extension $A^{[m]}$ (Subsection \ref{subsec-SGC-replicate}); hence, showing Theorem \ref{comparison}.
We then complete the proof of Theorem \ref{first main} by calculating the dominant and global dimension of certain Nakayama algebras explicitly (Subsection \ref{subsec-SGC-typeA}).
\section*{Acknowledgement}
This project started with numerous fruitful experiments thanks to \cite{QPA}.
AC is supported by JSPS International Research Fellowship, and partly by IAR Research Project, Institute for Advanced Research, Nagoya University.
\section{Preliminaries}\label{sec-Prelim}
Fix throughout a field $K$.
Algebras will always mean finite dimensional $K$-algebras.
We follow the notation and conventions of \cite{SkoYam}.
In particular, all modules are finite dimensional right modules and all maps are module maps unless otherwise stated.
The $K$-linear duality is denoted by $D(-)$.
Unless otherwise specified, all algebras are ring-indecomposable, non-simple, finite dimensional over $K$, and basic.
Let $A$ be an algebra.
The Nakayama functor $D\operatorname{Hom}_A(-,A_A)$ and its inverse are denoted by $\nu_A$ and $\nu_A^{-}$ respectively.
The Auslander-Reiten translate and its inverse are denoted by $\tau_A$ and $\tau_A^{-}$ respectively.
The full subcategory of projective (resp. injective) $A$-modules is denoted by $\operatorname{proj} A$ (resp. $\operatorname{inj} A$).
For a module $M$, we denote by $|M|$ the number of indecomposable direct summands of $M$ up to isomorphism.
We will usually denote an idempotent by $e$, or by $e_i$ for a label $i$ of (an isomorphism class of) a simple module.
\medskip
Let $M,N$ be $A$-modules.
We say that $M$ has an ($\operatorname{\mathrm{add}} N$)-resolution (resp. ($\operatorname{\mathrm{add}} N$)-coresolution) if there is an infinite exact sequence
\begin{align}
\begin{array}{rlllll}
\cdots & \to N_2 & \to N_1 & \to N_0 & \to M & \to 0 \\
(\text{resp. } \quad 0 &\to M & \to N^0& \to N^1& \to N^2 &\to \cdots)
\end{array}\notag
\end{align}
so that $N_i\in \operatorname{\mathrm{add}} N$ (resp. $N^i\in\operatorname{\mathrm{add}} N$).
Let $M$ be an $A$-module with minimal injective coresolution $I^\bullet=(I^n)_{n\geq 0}$ and minimal projective resolution $P_\bullet=(P_n)_{n\geq 0}$.
The \emph{dominant dimension} and \emph{codominant dimension} of $M$ are
\begin{align}
\operatorname{domdim} M& :=\begin{cases}
1+\sup \{ n \geq 0 \mid I^j \in \operatorname{proj} A \text{ for } j=0,1,...,n \} & \mbox{ if $I^0$ is projective,}\\
0 & \mbox{ if $I^0$ is not projective;}
\end{cases}\notag \\
\operatorname{codomdim} M & :=\begin{cases}
1+\sup \{ n \geq 0 \mid P_j \in \operatorname{inj} A \text{ for } j=0,1,...,n \} & \mbox{ if $P_0$ is injective,}\\
0 & \mbox{ if $P_0$ is not injective.}
\end{cases}\notag
\end{align}
respectively.
The \emph{dominant dimension} of $A$, denoted by $\operatorname{domdim} A$, is the dominant dimension of the regular $A$-module $A_A$.
Note that we always have $\operatorname{domdim} A = \operatorname{codomdim} DA$, see \cite{Ta} for example.
\medskip
The \emph{left} (resp. \emph{right}) \emph{self-injective dimension of $A$} is the injective dimension of $A$ as a left (resp. right) $A$-module.
An algebra $A$ is called \emph{Iwanaga-Gorenstein} if both its left and right self-injective dimensions are finite.
When these dimensions are finite, then they are equal.
If this number is equal to $d<\infty$, then we simply call $A$ a $d$-Iwanaga-Gorenstein algebra \cite{EncJen}.
Note that this is \emph{not} the same as the $k$-Gorenstein algebras which appear in \cite{AR}.
For convenience, we recall the following.
\begin{lemma}\label{lem-IGfacts}
Let $A$ be a $d$-Iwanaga-Gorenstein algebra, and $M$ be an $A$-module.
\begin{enumerate}[(1)]
\item {\rm (Iwanaga, 1980)} The following are equivalent.
\begin{itemize}
\item $\operatorname{pdim}_AM\leq d$.
\item $\operatorname{idim}_AM\leq d$.
\item $\operatorname{pdim}_AM< \infty$.
\item $\operatorname{idim}_AM< \infty$.
\end{itemize}
\item The following are equivalent for non-projective $M$.
\begin{itemize}
\item $M$ is the zero-th homology of a totally acyclic complex in $\mod A$ (i.e. it is Gorenstein projective).
\item $M\cong \Omega_A^d(X)$ for some $X\in\mod A$.
\item $\operatorname{Ext}_A^i(M,A)=0$ for all $i\in\{1,2,\ldots, d\}$ (i.e. $M$ is maximal Cohen-Macaulay).
\end{itemize}
\end{enumerate}
\end{lemma}
Every algebra of finite global dimension $d$ is $d$-Iwanaga-Gorenstein.
On the other hand, a $d$-Iwanaga-Gorenstein algebra has global dimension either $d$ or $\infty$.
\begin{definition}
A non-self-injective algebra $A$ is \emph{higher Auslander} (resp. \emph{minimal Auslander-Gorenstein}), or more specifically \emph{$d$-Auslander} (resp. \emph{$d$-Auslander-Gorenstein}), if $\operatorname{gldim} A=d+1=\operatorname{domdim} A$ (resp. $\operatorname{idim} {}_AA=\operatorname{idim} A_A=d+1=\operatorname{domdim} A$) for some positive $d$.
\end{definition}
\section{Replicated Algberas}\label{sec-Repl-Alg}
Let us review the construction of replicated algebras and the description of their modules.
Recall that for a finite dimensional algebra $A$, there is an infinite dimensional locally bounded algebra called \emph{repetitive algebra} $\hat{A}$.
The underlying vector space is given by $(\bigoplus_{i\in\mathbb{Z}}A)\oplus(\bigoplus_{i\in\mathbb{Z}}DA)$, with elements of the form $(a_i,f_i)_{i\in\mathbb{Z}}$ with finitely many non-zero $a_i$ and $f_i$'s.
The multiplication is defined by $(a_i,f_i)_i(b_i,g_i)_i=(a_ib_i,a_{i+1}g_i+f_ib_i)_i$.
$\hat{A}$ can be regarded as a $\mathbb{Z}\times \mathbb{Z}$ matrix (non-unital but locally unital) algebra with diagonal entries $\hat{A}_{i,i}=A$, sub-diagonal entries $\hat{A}_{i,i-1}=DA$, and everywhere else zero.
\begin{definition}
Let $A$ be a finite dimensional algebra.
Denote by $e^{(m)}$ the idempotent of $\hat{A}$ given by the $\mathbb{Z}\times\mathbb{Z}$ matrix with $(i,i)$-th entry $1$ for all $i\in\{0,1,\ldots,m\}$ and zero everywhere else.
For $m\geq 0$, the \emph{$m$-replicated algebra} $A^{(m)}$ of $A$ is the idempotent truncation $e^{(m)}\hat{A}e^{(m)}$.
In other words, we have
\begin{align}
A^{(m)}:=e^{(m)} \hat{A}e^{(m)} =
\left(\begin{array}{cccccc}
A & 0 & 0 & \cdots & 0 & 0 \\
DA & A & 0 & & & 0 \\
0 & DA & A & & & 0 \\
\vdots & \ddots & \ddots & \ddots & &\vdots \\
0 & & \ddots & \ddots & \ddots &\vdots \\
0 & 0 & 0 & \cdots & DA & A
\end{array}\right), \notag
\end{align}
where the matrix is of size $(m+1)\times (m+1)$.
We call $A$ the \emph{base algebra}.
\end{definition}
\begin{example}\label{eg-typeA-Kronecker}
(1) Let $A$ be the path algebra of type $\mathbb{A}_n$ with linear orientation (i.e. $A$ is a Nakayama algebra with $n$ simple modules).
Then $A^{(m)}=KQ/I$ with $Q$ of type $\mathbb{A}_{n(m+1)}$ and $I$ generated by all paths of length $n+1$.
(2) Let $A$ be the Kronecker algebra $K(1\rightrightarrows 2)$.
Then $A^{(m)} = KQ/I$ with
\[
Q: \xymatrix@C=50pt{ 1 \ar@/^/[r]^{a_1} \ar@/_/[r]_{b_1} & 2 \ar@/^/[r]^{a_2} \ar@/_/[r]_{b_2} & 3 \ar@{.}[rr] & & 2m+1\ar@/^/[r]^{a_{2m+1}} \ar@/_/[r]_{b_{2m+1}} & 2m+2 }
\]
and $I$ is generated by $a_ja_{j+1}$, $b_jb_{j+1}$, $a_jb_{j+1}-b_ja_{j+1}$, and all paths of length 3.
\end{example}
\begin{proposition}\label{prop-repli-tensor-SI}
Suppose $A, B$ are finite dimensional algebras with $B$ basic and self-injective.
Then there is an algebra isomorphism $(A\otimes_KB)^{(m)}\cong A^{(m)}\otimes_KB$.
In particular, $(A\otimes_KB)^{(m)}$ is minimal Auslander-Gorenstein if, and only if, so is $A^{(m)}$.
\end{proposition}
\begin{proof}
For simplicity, we take $\otimes:=\otimes_K$.
Suppose ${}_\Gamma M_\Lambda$ is a $\Gamma$-$\Lambda$-bimodule and $\phi$ is an algebra automorphism of $\Lambda$.
Then there is an algebra isomorphism
\begin{align*}
\left(\begin{array}{cc}\phi^{-1} & 0\\ \mathrm{id} & \mathrm{id}\end{array}\right):
\left(\begin{array}{cc}\Lambda & 0\\ M & \Gamma\end{array}\right)
\to \left(\begin{array}{cc}\Lambda & 0 \\ {}_1M_\phi & \Gamma\end{array}\right),
\end{align*}
where ${}_1M_\phi$ denotes the $\Gamma$-$\Lambda$-bimodule whose right $\Lambda$-action is twisted by $\phi$.
Since $B$ is basic and self-injective, there is an algebra automorphism $\phi$ of $B$ so that $DB\cong {}_1B_\phi$ as $B$-bimodule.
In particular, we have $A\otimes B$-bimodules isomorphisms $D(A\otimes B)\cong DA\otimes DB \cong DA\otimes {}_1B_\phi \cong {}_1(DA\otimes B)_{\mathrm{id}\otimes\phi}$.
Now the claim for $m=1$ follows from the discussion of the first paragraph by taking $\Lambda=\Gamma=A\otimes B$.
We omit the similar proof for $m>1$.
\end{proof}
Any $A^{(m)}$-module can be described by a sequence
\[
(M_0, f_1, M_1, f_2, \ldots, f_m, M_m) ,
\]
where $M_i\in \mod A$ and $f_i\in \operatorname{Hom}_A(M_i\otimes_ADA,M_{i-1})$ satisfying $f_{i-1}\circ (f_i\otimes 1)=0$ for all $i$.
We denote such a module simply by $(M_i,f_i)_i$ whenever $m$ is understood.
Let $(a_i,\phi_i)_i$ be an element of $A^{(m)}$, and $(x_i)_i$ be an element of the module given by the above sequence.
Then the $A^{(m)}$-action is given by $(x_i)_i\cdot (a_i,\phi_i)_i = (x_ia_i+ f_{i+1}(x_{i+1}\otimes \phi_{i+1}))_i$.
Using this description, an $A^{(m)}$-module homomorphism can be described by a sequence $(\theta_i)_i: (M_i,f_i)_i \to (N_i,g_i)_i$ with $\theta_i\in \operatorname{Hom}_A(M_i,N_i)$ satisfying the natural commutation relation (c.f. \cite[III.2]{ARS}).
\begin{observation}\label{eg-modules}
Let $M$ be an $A$-module
\begin{enumerate}[(1)]
\item For $0\leq k\leq m$, $[M]_k$ is the ``stalk module" given by a row vector with $m+1$ entries, where the $k$-th entry being $M$ and the remaining zero:
\[
[M]_k := (\underbrace{\stackrel{0}{0}, \ldots, \stackrel{k-1}{0}, \stackrel{k}{M}, \stackrel{k+1}{0}, \ldots, \stackrel{m}{0}}_{m+1\text{ entries}});
\]
the action of $A^{(m)}$ is obvious.
Note that, for every $0\leq k\leq m$, $[?]_k:\mod A\to \mod A^{(m)}$ is a full exact embedding of categories which preserves indecomposable modules, almost split sequences, and irreducible morphisms (c.f. \cite[Lemma 5]{ABST}).
\item If $M$ is projective (resp. injective), for $0\leq k < m$, $[\nu_A(M),M]_k$ (resp. $[M,\nu_A^{-1}(M)]_k$) is the $A^{(m)}$-module given by $M_{k+1}=\nu_A(M)$ (resp. $M$), $f_{k+1}=\mathrm{id}$, $M_{k}=M$ (resp. $\nu_A^-(M)$), i.e.
\begin{align}
[\nu_A(M),M]_k &:= (\underbrace{\stackrel{0}{0}, \ldots, \stackrel{k-1}{0}, \stackrel{k}{\nu_A(M)}, \stackrel{k+1}{M}, \stackrel{k+2}{0},\ldots, \stackrel{m}{0}}_{m+1\text{ entries}}) \quad \text{for projective $M$;}\notag \\
[M,\nu_A^-(M)]_k &:= (\underbrace{\stackrel{0}{0}, \ldots, \stackrel{k-1}{0}, \stackrel{k}{M}, \stackrel{k+1}{\nu_A^-(M)}, \stackrel{k+2}{0}, \ldots, \stackrel{m}{0}}_{m+1\text{ entries}}) \quad \text{for injective $M$.}\notag
\end{align}
This also shows that $[\nu_A(M),M]_k$ is projective-injective as it is a direct summand of the $(k-1)$-st row of $A^{(m)}$, as well as the dual of a direct summand of the $k$-th column of $A^{(m)}$.
\item The injective envelope $I_M$ and cosyzygy $\Omega_A^-(M)$ form a short exact sequence
\[
0\to M\xrightarrow{\iota} I_M \xrightarrow{\pi} \Omega_A^-(M)\to 0.
\]
For $0\leq k< m$, the injective envelop of $M$ is $[I_M,\nu_A^-(I_M)]_k$.
Moreover, we have a short exact sequence
\begin{align}\label{eq-cosyzygy}
0\to [M]_k\xrightarrow{[\iota,0]_k} [I_M,\nu_A^-(I_M)]_k \xrightarrow{[\pi,\mathrm{id}]_k} \Omega_{A^{(m)}}^-([M]_k)\to 0,
\end{align}
where the cosyzygy $\Omega_{A^{(m)}}^-([M]_k)$ of $[M]_k$ is given by $(M_i,f_i)_i$ with $M_{k+1}=\Omega_A^-(M)$, $f_{k+1}=\pi$, $M_{k}=\nu_A^-(I_M)$, and everything else zero.
In particular, if $M$ is injective, then $\Omega_{A^{(m)}}^-([I]_k)=[\nu_A^-(I)]_{k+1}$.
Note that $[\pi,\mathrm{id}]_k$ is the map given by the sequence $(\theta_i)_{i=0,1,\ldots, m}$ with $\theta_k=\pi, \theta_{k+1}=\mathrm{id}, \theta_i=0$ for all $i\neq k$; similarly for $[\iota,0]_k$.
We leave it to the reader to write down the analogous description of $\Omega_{A^{(m)}}([M]_k)$.
\end{enumerate}
\end{observation}
In addition to (2), we remark that
\begin{align}
\operatorname{proj} A^{(m)} &= \operatorname{\mathrm{add}}\{ [A]_0, [DA,A]_i \mid i=0,1,\ldots,m-1 \}; \notag \\
\operatorname{inj} A^{(m)} &= \operatorname{\mathrm{add}}\{ [DA]_m, [DA,A]_i \mid i=0,1,\ldots,m-1 \}. \notag
\end{align}
The following is well-known.
\begin{lemma}{\rm \cite{AssIwa}}\label{lem-gldim}
There is an inequality $m+\operatorname{gldim} A \leq \operatorname{gldim} A^{(m)} \leq (m+1)\operatorname{gldim} A+m.$
\end{lemma}
\section{Hereditary base algebra}\label{sec-Hered-base}
In subsection \ref{subsec-hered-dim}, we calculate the homological dimensions of the replicated algebras of hereditary algebras, and use them to determine when higher Auslander algebras appear.
Then we will give some other simple observations on these replicated algebras in subsection \ref{subsec-hered-further}.
Throughout this section, we assume $A$ is a basic, ring-indecomposable, hereditary, non-simple algebra.
This means that there is a finite acyclic valued quiver $Q$ associated to $A$ which is not an isolated vertex.
For detailed reference, including the classification of representation-finite hereditary algebras, we refer the reader to \cite[Chapter VIII]{ARS}.
Before we begin our exposition, we specifically remind the reader that $\nu_A^-(M)=\operatorname{Hom}_A(DA,M)=0$ for all indecomposable non-injective $A$-modules $M$.
Moreover, there is an automorphism $\nu_1:=\operatorname{\mathsf{R}Hom}_A(DA,?[1])$ on the bounded derived category $\mathrm{D}^{\mathrm{b}}(\mod A)$.
It is well-known that
\begin{align}\label{vanish}
\nu_1^{-k}(M)\in (\mod A)[s_M(k)]
\end{align}
for some $s_M(k)\in \mathbb{Z}_{\geq 0}$, for each indecomposable $A$-module $M$.
We denote by $M^{\succ k}$ the $A$-module $\nu_1^{-k}(M)[-s_M(k)]$.
Note that $\nu_1^{-1}(M)=\tau_A^-(M)$ for any non-injective $A$-module $M$.
\subsection{Homological dimensions}\label{subsec-hered-dim}
By Lemma \ref{lem-gldim} and the assumption that $\operatorname{gldim} A\leq 1$, we have
\[
\operatorname{gldim} A^{(m)} = \operatorname{idim} A^{(m)} < \infty \;\;\text{ for all }m\geq 1.
\]
\begin{lemma}\label{lem-injcopres}
Let $M$ be a non-injective $A$-module with minimal injective coresolution $0\to I^0 \xrightarrow{f} I^1 \to 0$.
For $0\leq i< m$, we have an exact sequence of $A^{(m)}$-modules:
\begin{align}\label{eq-inj-pres}
0 \to [M]_i \to [I^0,\nu_A^-(I^0)]_i \xrightarrow{[f,\nu_A^-(f)]_i} [I^1,\nu_A^-(I^1)]_i\to [\tau_A^-(M)]_{i+1} \to 0,
\end{align}
where the middle two terms are the minimal injective copresentation of $[M]_i$ in $\mod A^{(m)}$.
\end{lemma}
Note that the middle two terms are in $\operatorname{\mathrm{add}}([DA,A]_i)$; hence, they are projective-injective.
\begin{proof}
Using Observation \ref{eg-modules}, we obtain the first two terms of (\ref{eq-inj-pres}).
To obtain the third term, observe that $[f,\nu_A^-(f)]_i$ factors through $\Omega_{A^{(m)}}^-([M]_i)$ as follows:
\[
[I^0,\nu_A^-(I^0)]_i \xrightarrow{[\pi,\mathrm{id}]_i} \Omega_{A^{(m)}}^-([M]_i) \xrightarrow{[\mathrm{id},\nu_A^-(f)]_i} [I^1, \nu_A^-(I^1)]_i,
\]
where $\pi$ is the canonical projection from $I^0$ to $\Omega_A^-(M)$.
As $\ker(\nu_A^-(f))=\nu_A^-(M)=0$, the sequence is exact at $[I^0,\nu_A^-(I)]_i$.
Since $\mathrm{Coker}(f)=0$, we have $\mathrm{Coker}[f,\nu_A^-(f)]_i=[\mathrm{Coker}(\nu_A^-(f))]_{i+1}=[\tau_A^-(M)]_{i+1}$.
\end{proof}
Recall that for any indecomposable $A$-module $M$, we have
\begin{align}
\nu_1^-(M) \cong \begin{cases}
\tau_A^-(M), & \text{if $M$ is non-injective;}\\
\nu_A^-(M)[1], & \text{if $M$ is injective.}
\end{cases}\label{eq-nuD}
\end{align}
Using this and the fact that $\Omega_{A^{(m)}}^-([I]_i)=[\nu_A^-(I)]_{i+1}$ for any injective $A$-module $I$ yields the following list of exact sequences.
Let $M$ be an indecomposable projective $A$-module.
\begin{enumerate}[(i)]
\item If $M^{\succ k}$ is injective and $0\leq i< m$, then we have
\begin{align}
\xi(M,k,i):\qquad 0 \to [M^{\succ k}]_i \to I \to [M^{\succ k}]_{i+1} \to 0, \label{eq-seq1}
\end{align}
where $I$ is a projective-injective $A^{(m)}$-module.
\item If $M^{\succ k}$ is non-injective and $0\leq i< m$, then we have
\begin{align}
\xi(M,k,i):\qquad 0 \to [M^{\succ k}]_i \to I \to I' \to [M^{\succ k+1}]_{i+1} \to 0,\label{eq-seq2}
\end{align}
where $I,I'$ are projective-injective $A^{(m)}$-modules.
\item If $i=m$, then we have
\begin{align}
\xi(M,k,m):\qquad 0 \to [M^{\succ k}]_m \to [I]_m \to [I']_m \to 0,\label{eq-seq3}
\end{align}
where $I$ is a non-zero injective $A$-module, and so is $I'$ when $M^{\succ k}$ is non-injective; otherwise, $I'=0$.
\end{enumerate}
For any indecomposable $A$-module $M$ and any $0\leq i\leq m$, the minimal injective coresolution of the $A^{(m)}$-module $[M]_i$ can then be obtained by iteratively taking the Yoneda products of
\[
\xi(M,0,i), \xi(M,1,i+1), \cdots, \xi(M,m-1-i,m-1), \xi(M,m-i,m).
\]
\begin{proposition}\label{prop-dim}
Let $M$ be an indecomposable $A$-module and $i$ be an integer with $0\leq i\leq m$.
Then the following hold.
\begin{enumerate}
\item $\operatorname{idim}_{A^{(m)}}[M]_i-\operatorname{domdim}_{A^{(m)}}[M]_i = \operatorname{idim}_A M^{\succ m-i}$.
\item $\operatorname{domdim}_{A^{(m)}}[M]_i= 2(m-i)-s_M(m-i)$.
\item $\operatorname{idim}_{A^{(m)}}[M]_i = 2(m-i)-s_M(m-i+1)$
\end{enumerate}
\end{proposition}
\begin{proof}
From the preceding discussion, we see that every term in the minimal injective coresolution of $[M]_i$ is projective-injective apart from those given by $\xi(M,m-i,m)$.
Therefore, these values follow immediately from the description (\ref{eq-seq3}) of $\xi(M,m-i,m)$ and the fact that $[I]_m$ is a non-projective $A^{(m)}$-module for any injective non-projective $A$-module $I$.
\end{proof}
From (1) of the proposition, it makes sense to define
\[
\epsilon_M(k):=\begin{cases}
0, & \text{if $M^{\succ k}$ is injective;}\\
1, & \text{otherwise.}
\end{cases}
\]
Then we can write down the dominant and global dimension of $A$ in two simple formulae.
\begin{theorem}\label{thm-dims}
Let $A$ be a non-simple hereditary algebra, then we have
\begin{align}
\operatorname{domdim} A^{(m)} & = 2m-\max\{s_P(m)\mid P\mbox{ an indecomposable projective $A$-module}\} \notag \\
\text{ and }\quad \operatorname{gldim} A^{(m)} &= 2m-\min\{s_P(m)-\epsilon_P(m)\mid P\mbox{ an indecomposable projective $A$-module}\}\notag
\end{align}
for all $m\geq 1$.
\end{theorem}
\begin{proof}
Every indecomposable projective non-injective $A^{(m)}$-module is of the form $[P]_0$, where $P$ is an indecomposable projective $A$-module.
By Proposition \ref{prop-dim}, we have
\begin{align*}
\operatorname{domdim}_{A^{(m)}}[P]_0 = 2m-s_M(m) \quad \text{ and }\quad \operatorname{idim}_{A^{(m)}}[P]_0 = 2m-(s_M(m)-\epsilon_P(m)).
\end{align*}
By definition, the dominant and injective dimensions of $A^{(m)}$ are the minimum and maximum of these respectively.
Since $\operatorname{gldim} A^{(m)}$ is finite by Lemma \ref{lem-gldim}, it is equal to $\operatorname{idim}_{A^{(m)}} A^{(m)}$.
\end{proof}
Theorem \ref{thm-dims} implies for $m\geq 1$ that $\operatorname{domdim} A^{(m)} \geq m+1$ when $A$ is non-uniserial or $m\geq 2$.
This is because we have $s_P(m)<m$ in these situations.
On the other hand, the dominant dimension is usually much larger than this bound, which we will explain in the following.
\begin{corollary}\label{cor-RI}
If $A$ is representation-infinite hereditary algebra, then we have, for all $m\geq 1$,
\[
\operatorname{domdim} A^{(m)} = 2m \quad \text{ and }\quad \operatorname{gldim} A^{(m)}=\operatorname{idim} A^{(m)}= 2m+1.
\]
In particular, $A^{(m)}$ is neither minimal Auslander-Gorenstein nor higher Auslander for all $m\geq 1$.
\end{corollary}
\begin{proof}
When $A$ is representation-infinite, then we have $\nu_1^{-k}(P)\cong \tau_A^{-k}(P)$ (i.e. $s_M(k)=0$) for all $k\geq 0$.
The claim follows from Theorem \ref{thm-dims}.
\end{proof}
Part of this result, namely $\operatorname{gldim} (A^{(m)})=2m+1$, was also shown in \cite[Lem 4.3]{LvZh}.
In the following, we assume $A$ is representation-finite hereditary, and determine the dominant and global dimension of $A^{(m)}$.
We fix the enumeration of vertices of Dynkin graphs which have non-trivial automorphisms as follows:
\begin{align}
\mathbb{A}_n: & \xymatrix{ 1 \ar@{-}[r] & 2 \ar@{-}[r] & \cdots \ar@{-}[r]& n-1 \ar@{-}[r] & n} \notag \\
\mathbb{D}_n: & \xymatrix@R=10pt{ 1 \ar@{-}[r] & 2 \ar@{-}[r] & \cdots \ar@{-}[r]& n-2\ar@{-}[r]\ar@{-}[d] & n-1 \\ & & &n& } \notag \\
\mathbb{E}_6: & \xymatrix@R=10pt{ 1 \ar@{-}[r] & 2 \ar@{-}[r] & 3 \ar@{-}[d]\ar@{-}[r] & 4\ar@{-}[r] & 5 & \\ & & 6&&& } \notag
\end{align}
For a Dynkin graph $\Delta \in \{\mathbb{A}_n, \mathbb{B}_n, \mathbb{C}_n,\mathbb{D}_n, \mathbb{E}_6, \mathbb{E}_7, \mathbb{E}_8, \mathbb{F}_4, \mathbb{G}_2\}$, we denote by $h_\Delta$ the Coxeter number of type $\Delta$, which are given in the following table.
Denote by $\nu_\Delta$ a certain involutive graph automorphism of $\Delta$; we describe the non-identity ones by swapping of vertices as follows.
\[
\begin{array}{c||c|c|c|c|c|c|c|c}
\Delta & \mathbb{A}_n & \mathbb{B}_n,\mathbb{C}_n & \mathbb{D}_n & \mathbb{E}_6 & \mathbb{E}_7 & \mathbb{E}_8 & \mathbb{F}_4 & \mathbb{G}_2 \\ \hline
h_\Delta & n+1 & 2n & 2n-2 & 12 & 18 & 30 & 12 & 6 \\
\nu_\Delta & i \leftrightarrow n+1-i & \text{identity}& n-1 \leftrightarrow n \text{, if $n$ odd} & 1\leftrightarrow 5 & \text{identity} & \text{identity}& \text{identity}& \text{identity}\\
& & & \text{identity, if $n$ even} & 2\leftrightarrow 4 & & & &
\end{array}
\]
Since $A$ is representation-finite, its associated valued quiver $Q$ is of Dynkin type $\Delta$.
For a vertex $i\in \Delta_0$, we fix $e_i$ to be the primitive idempotent of the path algebra of type $\Delta$.
Denote by $P_i:=e_iA$ (resp. $I_i:=D(Ae_i)$) the corresponding indecomposable projective (resp. injective) module.
We define $\ell_i$ the cardinality of the $\tau_A^-$-orbit of $P_i$.
Equivalently, we have
\[
\ell_i = 1+\max\{j\geq 1\mid s_{P_i}(j)=0\} = \min\{j\geq 1\mid s_{P_i}(j)=1\}.
\]
Note that $P_i^{\succ \ell_i-1}=\tau_A^{-(\ell_i-1)}(P_i)\cong I_{\nu_\Delta(i)}$.
The following fact is well-known; see, for example, \cite{MiyYek}.
\begin{lemma}\label{lem-h-Delta}
There is a natural isomorphism of auto-equivalences $\nu_{\mathcal{D}}^{-h_\Delta}\cong [2]$ on $\mathcal{D}$.
\end{lemma}
In particular, by combining with (\ref{eq-nuD}), we get that
\begin{align}\label{eq-li-h}
\ell_{i}+\ell_{\nu_\Delta(i)}=h_\Delta.
\end{align}
\begin{lemma}\label{lem-RF0}
The following are equivalent.
\begin{enumerate}[(1)]
\item $\operatorname{idim}_{A^{(m)}}[P_i]_0=2m-s_{P_i}(m)$;
\item $\operatorname{idim}_{A^{(m)}}[P_i]_0=\operatorname{domdim}_{A^{(m)}}([P_i]_0)$;
\item $P_i^{\succ m}$ is injective;
\item $P_i^{\succ m} \cong I_i$ with $s_{P_i}(m)$ odd, or $P_i^{\succ m}\cong I_{\nu_\Delta(i)}$ with $s_{P_i}(m)$ even;
\item $m\in\{\ell_i-1, h_\Delta-1\}+h_\Delta\mathbb{Z}_{\geq 0}$
\end{enumerate}
Moreover, in such a case, we have
\begin{align}\label{eq-spi(m)-hered}
s_{P_i}(m) = \begin{cases}
2t, & \text{if $m=\ell_i-1+th_\Delta$ with $t\geq 0$;}\\
2t+1, & \text{if $m=h_\Delta-1+th_\Delta$ with $t\geq 0$,}
\end{cases}
\end{align}
\end{lemma}
\begin{proof}
\underline{(1)$\Leftrightarrow$(2)$\Leftrightarrow$(3)}: Follows from Proposition \ref{prop-dim}.
\underline{(3)$\Leftrightarrow$(4)}: Follows from the facts that $\nu_\Delta^2=\mathrm{id}$, $P_i^{\ell-1}\cong I_{\nu(i)}$, and $\nu_1^-(I_i)\cong P_i[1]$.
\underline{(4)$\Leftrightarrow$(5)}:
This follows from Lemma \ref{lem-h-Delta}; namely, we have the following equivalent conditions.
\begin{align}
m=\ell_i-1+th_\Delta\text{ for some }t\geq 0 & \Leftrightarrow \nu_1^{-m}(P_i)\cong \nu_1^{-(\ell_i-1)}(P_i)[2t] \cong I_{\nu_\Delta(i)}[2t] \in \operatorname{\mathrm{add}}(DA[2t]), \notag \\
m=-1+th_\Delta\text{ for some }t\geq 1 & \Leftrightarrow \nu_1^{-m}(P_i)\cong \nu_1(P_i)[2t] \cong I_{i}[2t-1] \in \operatorname{\mathrm{add}}(DA[tk-1]). \notag
\end{align}
The last statement is clear from these equivalences.
\end{proof}
\begin{corollary}\label{cor-RF}
Suppose $A$ is a representation-finite hereditary algebra.
\begin{enumerate}[(1)]
\item If $\ell_i\neq \ell_j$ for some $i,j$, then $A^{(m)}$ being a higher Auslander algebra is equivalent to $m= th_\Delta-1$ for some $t>0$.
In such a case, we have
\[
\operatorname{domdim} A^{(m)}=2t(h_\Delta-1)-1=\operatorname{gldim} A^{(m)}.
\]
\item If all $\ell_i$'s are equal, then $A^{(m)}$ being a higher Auslander algebra is equivalent to $m=th_\Delta/2-1$ for some $t>0$.
In such a case, we have
\[
\operatorname{domdim} A^{(m)}=t(h_\Delta-1)-1=\operatorname{gldim} A^{(m)}.
\]
\end{enumerate}
\end{corollary}
\begin{proof}
Note that $\operatorname{domdim} A^{(m)}=\operatorname{idim} A^{(m)}$ (the latter being equal to $\operatorname{gldim} A^{(m)}$) if, and only if, $\operatorname{domdim}_{A^{(m)}}[P_i]_0=\operatorname{idim}_{A^{(m)}}[P_i]_0$ for all $i\in Q_0$.
Therefore, in the case when equality holds, by (2)$\Rightarrow$(5) of Lemma \ref{lem-RF0}, $m$ has to be $th_\Delta-1$, or $\ell+th_\Delta-1$ when $\ell_i=\ell$ for all $i$.
Note that $\ell=h_\Delta/2$ in the latter case.
This proves one direction for both (1) and (2).
For the converse, if $m$ is as stated, then by (5)$\Rightarrow$(2) of Lemma \ref{lem-RF0}, $\operatorname{idim}_{A^{(m)}} [P_i]_0=\operatorname{domdim}_{A^{(m)}} [P_i]_0$ for all $i$, and this is equal to $2m-s_{P_i}(m)$.
By the formula (\ref{eq-spi(m)-hered}) in Lemma \ref{lem-RF0}, this value does not depend on $i$; hence, it is precisely the dominant dimension as well as the injective dimension of $A^{(m)}$.
\end{proof}
\begin{remark}
Note that the condition $\ell_i=\ell_j(=h_\Delta/2)$ for all $i,j$ is satisfied if, and only if, the orientation of $Q$ is stable under the involution $\nu$ (c.f. the description of the AR-quiver of $KQ$ from \cite{Gab1}).
A path algebra $KQ$ which satisfies this condition is called \emph{$(h_\Delta/2)$-homogeneous} in \cite{HerIya}.
\end{remark}
Combining Corollary \ref{cor-RI} and Corollary \ref{cor-RF}, we obtain a new characterisation of representation-(in)finiteness of a hereditary algebra.
\begin{corollary}\label{cor-summary-hered}
The following are equivalent for a hereditary algebra $A$.
\begin{itemize}
\item $A$ is representation-finite (resp. representation-infinite);
\item there exist infinitely many (resp. there does not exists any) $m\in \mathbb{Z}_+$ such that $A^{(m)}$ is a higher Auslander algebra.
\end{itemize}
Moreover, in the representation-finite case, the global and dominant dimension of the higher Auslander replicated algebra is equal to $t(h_\Delta-1)-1$ for some $t>0$.
\end{corollary}
\subsection{Further Properties}\label{subsec-hered-further}
The fact that the base algebra $A$ is hereditary allows us to obtain a few more properties about $A^{(m)}$ compared to the more general case of base algebras which will be studied in the next section.
The key property of $A$ being hereditary allows us list the indecomposable $A^{(m)}$-modules explicitly - they are all in the forms listed in Observation \ref{eg-modules}.
\begin{proposition}\label{prop-reptype}
Every indecomposable $A^{(m)}$-module is in one of the following forms.
\begin{itemize}
\item $[M]_i$ with $M$ an indecomposable $A$-module and $0\leq i\leq m$.
\item $[\nu_A(P),P]_i$ with $P$ an indecomposable projective $A$-module and $0\leq i<m$.
\item $\Omega_{A^{(m)}}^{-}([M]_i)$ with $M$ an indecomposable non-injective $A$-module and $0\leq i<m$.
\end{itemize}
In particular, $A^{(m)}$ preserves the representation-type of $A$ for all $m\geq 1$.
Moreover, if $A$ is representation-finite with $r$ indecomposable modules, then there are $(2m+1)r$ indecomposable $A^{(m)}$-modules.
\end{proposition}
\begin{proof}
According to \cite{ABST}, this is somewhat well known - as it is just a consequence of $A^{(m)}$ being a ``nice" idempotent truncation of the repetitive algebra $\hat{A}$.
The stated list can then be obtained using the theory of repetitive algebras.
\end{proof}
This result hinted that there should be a close connection between indecomposable $A^{(m)}$-modules and dimension vectors.
We closely follow the article \cite{PS} in the following.
Consider a basic algebra $A=KQ/I$ over an algebraically closed field $K$.
It is called \emph{triangular} if the quiver $Q$ contains no oriented cycles.
Given a triangular algebra $A$, the \emph{Tits form} $q_A(x)$ of $A$ is defined as
\begin{align}
q_A(x) := \sum\limits_{i \in Q_0}^{}{x_i^2}- \sum\limits_{\alpha \in Q_1}{x_{s(\alpha)} x_{t(\alpha)}}+ \sum\limits_{i,j \in Q_0}{\dim_K \operatorname{Ext}_A^2(S_i,S_j) x_i x_j},
\end{align}
where $x=(x_i)_{i\in Q_0}$, and $s(-), t(-)$ denote the start and target of an arrow respectively.
The \emph{roots} of $q_A$ are the vectors $v \in \mathbb{Z}^{|Q_0|}$ such that $q_A(v)=1$.
A root $v$ of $q_A$ is \emph{positive} if all of its entries are non-negative.
Gabriel's famous theorem \cite{Gab2} says that for a representation-finite hereditary algebra, there is a one-to-one correspondence between the dimension vectors of the indecomposable modules and the positive roots of the Tits form.
It is not difficult to see that the same holds for $A^{(m)}$ when the base algebra $A$ is representation-finite hereditary, and we suspect this is known to people who have worked with replicated algebras.
Since we cannot find this precise statement in the literature, a brief explanation is provided below.
\begin{proposition}\label{prop-tits}
If $A$ is representation-finite hereditary algebra over an algebraically closed field $K$, then
$X\mapsto \mathbf{dim}(X)$ induces a bijection from the set of isoclasses of indecomposable $A^{(m)}$-modules to the set of of positive roots of the Tits form of $A^{(m)}$
\end{proposition}
\begin{proof}
It follows from the definition of $A^{(m)}$ and heredity of $A$ that $A^{(m)}$ is triangular; so it makes sense to talk about Tit forms of $A^{(m)}$.
Moreover, $A^{(m)}$ is a so-called convex subalgebra of the repetitive algebra $\hat{A}$, which is simply-connected.
This implies that $A^{(m)}$ is also simply-connected.
Recall from Proposition \ref{prop-reptype} that $A^{(m)}$ is representation-finite.
Hence, $A^{(m)}$ satisfies all the assumption required to apply a theorem of Bongartz (see \cite[Theorem 7.2]{PS}), which gives the claim.
\end{proof}
Recall the following generalisations of quasi-Frobenius (i.e. self-injective) algebras from \cite{Thr}.
\begin{definition}\label{def-QF123}
Let $A$ be a finite dimensional algebra.
\begin{itemize}
\item $A$ is \emph{QF-1} if every faithful left $A$-module $M$ has double centraliser property, i.e. there is a canonical ring epimorphism $A\to \operatorname{End}_{\Gamma}(M_{\Gamma})$, where $\Gamma:=\operatorname{End}_A(M)^{\mathrm{op}}$.
\item $A$ is \emph{QF-2} if every indecomposable projective $A$-module has a simple socle.
\item $A$ is \emph{QF-3} if $\operatorname{domdim} (A)\geq 1$.
\end{itemize}
\end{definition}
Note that any self-injective algebra satisfies all three conditions.
An algebra is \emph{QF-13} if it is QF-1 and QF-3.
\begin{proposition}\label{prop-yamagata-QF13}{\rm \cite[Corollary 3.5.1]{Yam}}
Suppose $A$ is a QF-3 algebra with $\operatorname{domdim} A\geq 2$.
Then $A$ is also QF-1 if, and only if, every indecomposable module $M$ has $\operatorname{codomdim}(M)\geq 1$ or $\operatorname{domdim} (M)\geq 1$
\end{proposition}
\begin{proposition}\label{prop-QF}
Suppose $A$ is a hereditary algebra and $m$ is a positive integer.
For any indecomposable $A^{(m)}$-module $X$ which is not projective-injective, we have
\begin{itemize}
\item $\operatorname{domdim}_{A^{(m)}}(X)\geq 1$ if, and only if, $X=[M]_i$ with $i<m$ for some indecomposable $A$-module, or $X=\Omega_{A^{(m)}}^-([M]_i)$ with $i<m$ for some indecomposable non-injective $A$-module $M$.
\item $\operatorname{codomdim}_{A^{(m)}}(X)\geq 1$ if, and only if, $X=[M]_i$ with $i>1$ for some indecomposable $A$-module, or $X=\Omega_{A^{(m)}}^-([M]_i)$ with $i\leq m$ for some indecomposable non-injective $A$-module $M$.
\end{itemize}
In particular, $A^{(m)}$ is a QF-13 algebra for all $m\geq 2$, or for all $m\geq 1$ when $A$ is not uniserial.
\end{proposition}
\begin{proof}
It is clear from Lemma \ref{lem-injcopres} that $\Omega_{A^{(m)}}^-([M]_i)$ has projective cover being injective, and injective envelope being projective, as long as $i<m$.
For $X=[M]_i$, Lemma \ref{lem-injcopres} shows that the injective envelop of $[M]_i$ is the projective-injective module $[I_M,\nu_A^-(I_M)]_{i}$ for all $i<m$.
Hence, it has dominant dimension at least 1.
On the other hand, $[M]_{m}$ has injective envelope $[I_M]_{m}$, which is not projective.
Likewise, the projective cover of $[M]_i$ with $i>0$ is the projective-injective module $[\nu_A(P_M),P_M]_{i-1}$, where $P_M$ is the projective cover of $M$ in $\mod A$; whereas $[M]_0$ has a non-injective projective cover $[P_M]_0$.
It follows from the discussion after Theorem \ref{thm-dims} that $A^{(m)}$ is QF-3 under the stated situations; hence, the final claim follows from Proposition \ref{prop-yamagata-QF13}.
\end{proof}
\begin{remark}
In general, $A^{(m)}$ is not a QF-2 algebra.
For example, if $A$ is the Kronecker algebra, then one can see from Example \ref{eg-typeA-Kronecker} (2) that $e_{2m+1}A^{(1)}$ has non-simple socle, namely $S_{2m+2}\oplus S_{2m+2}$.
\end{remark}
\section{Serre-formal algebras and their replicated algebras}\label{sec-Nu-formal}
The aim of this section is to introduce a new class of algebras which we call \emph{Serre-formal}, and give some basic properties.
We can naturally extend the methods and results from the previous section on hereditary algebras to this new class of algebras.
\subsection{Definition and basic properties}\label{subsec-NF-Bkgd}
We assume throughout this section that $A$ is a ring-indecomposable finite dimensional algebra.
Suppose $A$ is Iwanaga-Gorenstein.
Consider the bounded derived category $\mathrm{D}^{\mathrm{b}}(\mod A)$ of $A$-modules.
Denote by $\mathcal{D}$ the subcategory $\mathrm{perf}(A)$ given by the perfect complexes.
In this setting, we have endofunctors on $\mathcal{D}$ given by
\[
\nu:= DA\otimes^\mathsf{L}_A- = D\circ \operatorname{\mathsf{R}Hom}_A(-,A) \quad\text{ and }\quad \nu^{-1}:= \operatorname{\mathsf{R}Hom}_A(DA,-)
\]
which are quasi-inverses of each other; see \cite{Hap1}.
Note that $\nu$ is the Serre functor of $\mathcal{D}$.
\begin{definition}
An algebra $A$ is \emph{Serre-formal}, or \emph{$\nu$-formal}, if it is Iwanaga-Gorenstein and $\nu^{k}(A)$ is isomorphic to a direct sum of stalk complexes of $A$-module, i.e. $\nu^k(A)\cong \bigoplus_{n\in \mathbb{Z}} H^n(\nu^{k}(A))[-n]$, for all integers $k$, in $\mathrm{D}^{\mathrm{b}}(\mod A)$.
\end{definition}
The typical examples of $\nu$-formal algebras are self-injective algebras and hereditary algebras.
The Auslander-Reiten theoretic generalisation of hereditary algebras - $d$-hereditary algebras - are also examples of $\nu$-formal algebra; this can be seen immediately from the following definition.
\begin{definition}\label{def-dhered}
Let $d$ be a non-negative integer.
An algebra $A$ is \emph{$d$-hereditary} \cite{HIO} if $\operatorname{gldim} A=d$ and, for all $i\in \mathbb{Z}$, $\nu^i(A)$ belongs to $\operatorname{\mathrm{add}}\bigcup_{k\in \mathbb{Z}}(\mod A)[dk]$.
\end{definition}
\begin{example}\label{eg-alg}
We give two easy examples, as well as a non-example.
\begin{enumerate}[(1)]
\item If $A$ is $\nu$-formal, then so is $A^{\mathrm{op}}$. This follows from the following commutative diagram \[
\xymatrix@C=60pt{\mathrm{perf}(A) \ar[r]^{\nu_{\mathrm{perf}(A)}} \ar@{<->}[d]_{D}^{\wr}& \mathrm{perf}(A)\ar@{<->}[d]_{D}^{\wr}\\
\mathrm{perf}(A^{\mathrm{op}}) \ar[r]^{\nu^{-1}_{\mathrm{perf}(A^{\mathrm{op}})}} & \mathrm{perf}(A^{\mathrm{op}}).
}
\]
\item The family of 1-Iwanaga-Gorenstein algebras arising from symmetrizable Cartan matrix introduced in \cite{GLS} are $\nu$-formal.
This follows from combining \cite[Prop 11.6]{GLS} and \cite[Theorem 1.1]{GLS}.
Namely, for such an algebra $A$, any preprojective and preinjective $A$-module (i.e. direct summands of $\tau^{k}(A)$ for some $k\in\mathbb{Z}$) has injective and projective dimension at most 1, which implies that $A$ is $\nu$-formal.
Note that this family contains all the path algebras associated to finite acyclic quivers as well as many other algebras of infinite global dimension.
\item (Non-example) A generalisation, called ``pro-species algebras", of the class of algebras in (2) was introduced \cite{Kue}; the following example shows that such an algebra is not necessarily $\nu$-formal.
Let $A=KQ/I$ where
\[
Q:= \xymatrix@R=20pt{ 1 \ar[r]^{\alpha} & 2\ar@/^1pc/[r]^{\beta} & 3\ar@/^1pc/[l]^{\gamma} }
\]
and $I$ is generated by $\beta\gamma, \gamma\beta$.
One can check that this is a 1-Iwanaga-Gorenstein algebra; in fact, this is a ``locally self-injective" pro-species algebra (c.f. \cite{Kue}).
The minimal injective coresolution of $e_2A$ is $0\to D(Ae_3) \to D(Ae_1)\to 0$.
So we get that $\nu^{-1}(e_2A) = (e_3A\to e_1A)$, which has non-zero cohomology at both degree 0 and 1; hence, $A$ is not $\nu$-formal.
\end{enumerate}
\end{example}
Note that the class of $\nu$-formal algebras with finite global dimension is still strictly larger than that of $d$-hereditary algebras.
For example, the tensor product algebra of two hereditary algebras is not necessarily $d$-hereditary, but it is $\nu$-formal.
More generally, we have the following.
\begin{proposition}\label{prop-tensor}
If $A,A'$ are $\nu$-formal algebras, then $A\otimes_K A'$ is also $\nu$-formal.
\end{proposition}
\begin{proof}
Since $A,A'$ are both Iwanaga-Gorenstein, $A\otimes_K A'$ is also.
We will omit the subscript $K$ on the tensor product for clarity in the following.
For $X\in\mod A$ and $Y\in \mod A'$, we have isomorphisms
\begin{align}
\nu(X\otimes Y) & \cong D(A\otimes A')\otimes^\mathsf{L}_{A\otimes A'} (X\otimes Y) \cong (DA\otimes^\mathsf{L}_A X)\otimes DA'\otimes^\mathsf{L}_{A'}Y \cong \nu(X) \otimes \nu(Y),\notag
\end{align}
where the last two $\nu$ are the auto-equivalences on $\mathrm{perf}(A)$ and $\mathrm{perf}(A')$ respectively.
This means that, for all $i\geq 0$, if $\nu^{i}(X)$ and $\nu^{i}(Y)$ are both direct sum of stalk complexes, then so is $\nu^{i}(X\otimes Y)$.
Hence, the required condition on $\nu^{i}(A\otimes A')$ is satisfied by induction on $i\geq 0$.
Similarly, the isomorphisms
\begin{align}
\nu^{-1}(X\otimes Y) & \cong \operatorname{\mathsf{R}Hom}_{A\otimes A'}(D(A\otimes A'), X\otimes Y) \cong \operatorname{\mathsf{R}Hom}_{A\otimes A'}(DA\otimes DA', X\otimes Y) \notag \\
& \cong \operatorname{\mathsf{R}Hom}_A(DA,X)\otimes \operatorname{\mathsf{R}Hom}_{A'}(DA',Y) \cong \nu^{-1}(X) \otimes \nu^{-1}(Y)\notag
\end{align}
yields $\nu^{-i}(A\otimes A')\cong \nu^{-i}(A)\otimes \nu^{-i}(A')$ for all $i\geq 0$, which shows that the required condition on $\nu^{-i}(A\otimes A')$ is satisfied.
\end{proof}
We have particular interest in the modules $\bigoplus_{k\in\mathbb{Z}}\nu^{k}(A)$ for a $\nu$-formal algebra $A$.
If one bears in mind that the toy model for $\nu$-formal algebras are the hereditary algebras, then one can think of them as analogue of preprojective and preinjective modules.
For ease of further exposition, let us fix some notations.
We denote by $\mathcal{S}$ the indexing set of (isomorphism classes of) simple $A$-modules.
The indecomposable projective (resp. injective) $A$-module corresponding to $x\in \mathcal{S}$ is denoted by $P_x$ (resp. $I_x=\nu_A(P_x)$).
For $x\in \mathcal{S}$ and $k\in\mathbb{Z}_{\geq 0}$, we have
\[
\nu^{-k}(P_x)\in (\mod A)[s_x^-(k)]\quad \text{ and }\quad \nu^{k}(I_x)\in (\mod A)[s_x^+(k)]
\]
for some integers $s_x^{\pm}(k)$.
Note that $s_x^-$ and $s_x^+$ are decreasing and increasing functions on $\mathbb{Z}_{\geq 0}$ respectively with $s_x^-(0)=s_x^+(0)=0$.
We define $s_x^-$ and $s_x^+$ separately as it is more convenient to state our forthcoming results in this way.
We will denote the indecomposable $A$-modules $\nu^{-k}(P_x)[-s_x^-(k)]$ and $\nu^{k}(I_x)[-s_x^+(k)]$ by $P_x^{\succ k}$ and $I_x^{\prec k}$ respectively.
Note that as $\nu$ is an auto-equivalence on $\mathcal{D}=\mathrm{perf}(A)$, these modules have finite projective dimensions.
By definitions, we have
\begin{align}
& P_x^{\succ k+1} = \nu^{-(k+1)}P_x[-s_x^-(k+1)] \cong \nu^{-1}P_x^{\succ k}[s_x^-(k)-s_x^-(k+1)],\notag \\
\text{and }\quad & I_x^{\prec k+1} = \nu^{k+1}I_i[-s_x^+(k+1)] \cong \nu I_x^{\prec k}[s_x^+(k)-s_x^+(k+1)].\notag
\end{align}
Our notations are chosen so that for a hereditary algebra $A$, $P_x^{\succ k+1}$ lies on the right hand side of $P_x^{\succ k}$ on the Auslander-Reiten quiver of $\mathcal{D}$.
\begin{lemma}\label{lem-vanish}
Let $A$ be a $\nu$-formal algebra and fix $x\in \mathcal{S}$.
Then the following holds for all $k\in \mathbb{Z}_{\geq 0}$.
\begin{itemize}
\item[(1)] $\operatorname{Ext}_A^p(DA,P_x^{\succ k})\neq 0$ is equivalent to $p=s_x^-(k)-s_x^-(k+1)$;
\item[(1')] $\operatorname{Ext}_A^p(I_x^{\prec k},A)\neq 0$ is equivalent to $\mathrm{Tor}_p^A(DA,I_x^{\prec k})\neq 0$, and also to $p=s_x^+(k+1)-s_x^+(k)$;
\item[(2)] $\operatorname{idim}_{A} P_x^{\succ k} = s_x^-(k)-s_x^-(k+1) = \operatorname{pdim}_{A} P_x^{\succ k+1}$;
\item[(2')] $\operatorname{pdim}_{A} I_x^{\prec k} = s_x^+(k+1)-s_x^+(k) = \operatorname{idim}_{A} I_x^{\prec k+1}$;
\end{itemize}
In particular, $P_x^{\succ k}$ is injective (resp. projective) if, and only if, $s_x^-(k)=s_x^-(k+1)$ (resp. $s_x^-(k)=s_x^-(k-1)$).
Dually, $I_x^{\prec k}$ is projective (resp. injective) if, and only if, $s_x^+(k)=s_x^+(k+1)$ (resp. $s_x^+(k)=s_x^+(k-1)$).
\end{lemma}
\begin{proof}
Since $\operatorname{Ext}_A^p(DA,P_x^{\succ k})=H^p(\operatorname{\mathsf{R}Hom}_A(DA,P_x^{\succ k}))=H^p(\nu^{-1}(P_x^{\succ k}))$ for all $p\in\mathbb{Z}$, $\nu$-formality of $A$ implies that $\operatorname{Ext}_A^p(DA,P_x^{\succ k})$ vanishes for all indices but $p=s_x^-(k)-s_x^-(k+1)$.
This gives us (1).
(1') can be proved similarly by considering the cohomology of $DA\otimes^\mathsf{L}_A-$, and the functorial isomorphism $D\operatorname{Ext}_A^p(-,A)\cong \mathrm{Tor}_p^A(DA,-)$.
We prove (2) by induction on $k$.
The case of $k=0$ is trivial.
Suppose the claim holds for $P_x^{\succ k}$ for some $k\geq 0$.
If $P_x^{\succ k}$ is injective, then $\nu^{-1}(P_x^{\succ k})=\nu_A^-(P_x^{\succ k})\in \operatorname{proj} A$ and we are done.
Thus, we can assume $P_x^{\succ k}$ is non-injective.
Since $A$ is Iwanaga-Gorenstein, it follows from Lemma \ref{lem-IGfacts} that finiteness of $\operatorname{pdim}_AP_x^{\succ k}$ implies that of $\operatorname{idim}_A P_x^{\succ k}$.
It now suffices to prove that $\operatorname{pdim}_A P_x^{\succ k}<\infty$, as the vanishing of the Ext-spaces forces $\operatorname{idim}_A P_x^{\succ k}=s_x^-(k)-s_x^-(k+1)$.
By induction hypothesis $P_x^{\succ k}$ has a minimal injective coresolution
\[
0\to P_x^{\succ k} \to I^0 \to I^1 \to \cdots \to I^p\to 0,
\]
where $p=s_x^-(k)-s_x^-(k+1)$.
Applying $\nu_A^{-}=\operatorname{Hom}_A(DA,-)$ yields a sequence
\[
0 \to \nu_A^-(P_x^{\succ k}) \to \nu_A^-(I^0) \to \nu_A^-(I^1) \to \cdots \to \nu_A^-(I^p) \to M \to 0,
\]
where $M=\mathrm{Coker}(\nu_A^-(f))$ and $f$ is the differential $I^{p-1}\to I^p$.
By construction, we have $M=H^p(\nu^{-1}(P_x^{\succ k}))$, which means that $M\cong P_x^{\succ k+1}$ and $p=s_x^-(k)-s_x^-(k+1)$.
It now follows from (1) that the sequence above is exact.
In particular, it is the minimal projective resolution of $P_x^{\succ k+1}$.
Hence, we have $\operatorname{pdim}_A P_x^{\succ k+1} = s_x^-(k)-s_x^-(k+1)$ as required.
(2') can be proved analogously, which we leave the reader to check.
\end{proof}
\begin{example}
When $A$ is $d$-hereditary, an indecomposable $A$-module $M$ is called $d$-preprojective (resp. $d$-preinjective) if $M=\nu^{k}(P_x)[-dk]$ for some integer $k\leq 0$ (resp. $k\geq 1$).
Moreover, we have
\begin{align}
\operatorname{idim}_A P_x^{\succ k} &= s_x^-(k)-s_x^-(k+1)=\begin{cases} d, & \text{if $P_x^{\succ k}$ is non-injective;} \\ 0, & \text{otherwise.}\end{cases}\notag \\
\operatorname{pdim}_A I_x^{\prec k} &= s_x^+(k+1)-s_x^+(k)=\begin{cases} d, & \text{if $I_x^{\prec k}$ is non-injective;} \\ 0, & \text{otherwise.}\end{cases}\notag
\end{align}
\end{example}
Having Theorem \ref{thm-dims} in mind, it is natural to ask what subclass of the $\nu$-formal algebras play the analogous role of representation-finite hereditary algebras.
\begin{definition}
Let $A$ be a $\nu$-formal algebra and $x\in \mathcal{S}$.
Define
\[
\ell_x := \inf \{ k\geq 0 \mid P_x^{\succ k}\in \operatorname{\mathrm{add}} A \} = \inf\{ k\geq 1 \mid P_x^{\succ k-1}\in \operatorname{\mathrm{add}} DA\}.
\]
We say that $A$ is \emph{periodically $\nu$-formal} if $\ell_x<\infty$ for all $x\in \mathcal{S}$; otherwise, \emph{aperiodically $\nu$-formal}.
In the former case, we define $\sigma$ to be the permutation on $\mathcal{S}$ so that $P_x^{\succ \ell_x}\cong P_{\sigma(x)}$, and $k_A$ to be its order.
\end{definition}
By Lemma \ref{lem-vanish}, $A$ being periodically $\nu$-formal is the same as the existence of $k_x\in \mathbb{Z}_{\geq 0}$ so that $s_x^-(k_x)=s_x^-(k_x+1)$ (or $s_x^+(k_x)=s_x^+(k_x+1)$) for all $x\in \mathcal{S}$.
Note also that $\ell_x$ can also be defined in terms of $I_x^{\prec k}$ instead, namely
\[
\ell_x = \inf \{ k\geq 0\mid I_{\sigma(x)}^{\prec k}\in \operatorname{\mathrm{add}} DA \} = \inf\{ k\geq 1\mid I_{\sigma(x)}^{\prec k-1}\in \operatorname{\mathrm{add}} A\}.
\]
Recall that if $\phi$ is an automorphism on $A$, then there is a natural auto-equivalence $\phi^*$ on $\mod A$ (hence, on $\mathrm{perf}(A)$) given by twisting the $A$-action on the modules by $\phi$.
\begin{definition}
Following \cite{HerIya}, an algebra $A$ is said to be \emph{twisted fractionally Calabi-Yau} if there exist integers $n$ and $m\neq 0$, as well as an automorphism $\phi$ of $A$, so that there is a natural isomorphism of auto-equivalences $\nu^m\simeq [n]\circ \phi^*$ on $\mathrm{perf}(A)$.
It follows from \cite[Prop 4.3]{HerIya} that this is equivalent to $\nu^m(A)\cong A[n]$ in $\mathrm{D}^{\mathrm{b}}(\mod A)$.
We say that the \emph{twisted Calabi-Yau dimension} of $A$ is $(n,m)$ if
$m$ is the minimal positive integer so that $\nu^m(A)\cong A[n]$ in $\mathrm{perf}(A)$.
\end{definition}
The following result generalises \cite[Theorem 1.1]{HerIya}.
\begin{proposition}\label{prop-twistedCY}
Suppose $A$ is a $\nu$-formal algebra.
Then $A$ is periodically $\nu$-formal if, and only if, $A$ is twisted fractionally Calabi-Yau.
\end{proposition}
\begin{proof}
The if-part follows from the definition of twisted fractionally Calabi-Yau.
We concentrate on showing the converse direction in the rest.
Let us first define $h_x:= \ell_x + \ell_{\sigma(x)} + \cdots +\ell_{\sigma^{k_A-1}(x)}$ and $p_x:=s_x^-(h_x)$.
We claim that the rational number $p_x/h_x$ is independent of $x\in \mathcal{S}$.
Indeed, this follows from a simple modification of the argument in \cite[subsection 4.1]{HerIya}, which we explain below for completeness.
Suppose $x,y\in \mathcal{S}$ are chosen so that $\operatorname{Hom}_A(P_x,P_y)\neq 0$.
Since
\[
\nu^{-h_x}(P_x) \cong P_{\sigma^{k_A}(x)}[s_x^-(h_x)] = P_x[p_x]
\]
by definition, we get that
\[
0 \neq \operatorname{Hom}_{\mathcal{D}}\big(\nu^{-h_yh_x}(P_x),\nu^{-h_xh_y}(P_y)\big) \cong \operatorname{Hom}_{\mathcal{D}}(P_x[h_yp_x],P_y[h_xp_y]).
\]
This implies that $h_yp_x = h_xp_y$.
It follows from the assumption of $A$ being ring-indecomposable that $p_x/h_x=p_y/h_y$ for all $x,y\in\mathcal{S}$.
Take $m$ to be the least common multiple of $h_x$'s over all $x\in\mathcal{S}$, then it suffices to show that $s_x^-(m)$ is independent of $x\in\mathcal{S}$ to finish the proof.
Indeed, since the isomorphism $\nu^{-m}(P_x)=\nu^{-(m/h_x)h_x}(P_x)\cong P_x[(m/h_x)p_x]$ holds for all $x\in \mathcal{S}$, we have $s_x^-(m)=mp_x/h_x$ for all $i\in\mathcal{S}$.
It now follows from the above claim on $p_x/h_x$ that $s:=s_x^-(m)$ is independent of $x\in\mathcal{S}$.
Hence, we have $\nu^{-m}(A)\cong A[s]$.
\end{proof}
Suppose $A$ is periodically $\nu$-formal.
We denote by $(h_A,-s_A)$ the twisted Calabi-Yau dimension of $A$.
Note that since $\nu$ is an auto-equivalence, $\nu^{-k}(A)\cong A[s]$ is equivalent to $\nu^k(A)\cong A[-s]$, so $h_A$ is equal to the minimal positive integer $k$ so that $\nu^{-k}(A)\cong A[s]$, and $s_A=s_x^-(h_A)=-s_x^+(h_A)$ for any $x\in\mathcal{S}$.
\begin{example}
If $A=kQ$ is representation-finite hereditary of Dynkin type $\Delta$, then $\sigma$ is the permutation on $Q_0$ induced by the graph automorphism $\nu_\Delta$ in the notation of the previous section.
This implies that $k_A=1$ or $k_A=2$, and the integer $m$ in the proof of Proposition \ref{prop-twistedCY} is $k_Ah_\Delta/2$.
If the orientation of $Q$ is stable under $\nu_\Delta$, then the twisted Calabi-Yau dimension is $(h_A,-s_A)=(h_\Delta/2,h_\Delta/2-1)$; otherwise, $(h_A,-s_A)=(h_\Delta,h_\Delta-2)$.
In particular, both $m$ and $h_A$ are not necessarily the Coxeter number associated to $A$.
\end{example}
We also remark that for periodically $\nu$-formal algebras $A$, there is no formula to express $k_A$, or even for the upper bound of $k_A$, in terms of $\operatorname{idim} A$ when $A$ is not hereditary.
For example, consider the algebra $A:= K\vec\mathbb{A}_n/I$, where $I$ is generated by the unique path of length $n-1$.
This is a 2-hereditary (and 2-representation-finite, see Subsection \ref{subsec-NF-naka}) algebra where $\sigma$ has order $k_A=n$ if $n$ odd; otherwise, $k_A=n/2$.
\subsection{Homological dimensions}\label{subsec-NF-dims}
We are now going to state the $\nu$-formal analogue of the results and their ``dual" of the previous section.
We will not give detailed proofs because they are exactly the same as in the hereditary case, but we will explain what extra ingredients and tweaks are needed in order to copy the proof directly.
The proof for the dual results (projective resolutions and projective dimensions of modules) will be completely omitted.
Moreover, our result in the previous section says that appearance of higher Auslander algebras as replicated algebras is equivalent to the base algebra being of Dynkin type.
The second aim of this section is to explain why the analogous result also holds for $\nu$-formal base algebras, i.e. the appearance of minimal Auslander-Gorenstein algebras as replicated algebras is equivalent to the base algebra being periodically $\nu$-formal.
Throughout, we assume $A$ is an indecomposable $\nu$-formal algebra with $\mathcal{S}$ indexing the simple $A$-modules.
Let us begin with the following analogue of Lemma \ref{lem-injcopres}.
\begin{lemma}\label{lem-injcopres2}
Let $m$ be a positive integer, $x\in \mathcal{S}$, and $k\geq 0$.
\begin{enumerate}
\item For any $0\leq i< m$, we have an exact sequence of $A^{(m)}$-modules:
\begin{align}
\xi(x,k,i):\qquad 0 \to [P_x^{\succ k}]_i \to I^0 \to I^1 \to \cdots \to I^p \to [P_x^{\succ k+1}]_{i+1} \to 0,\notag
\end{align}
where $p=s_x^-(k)-s_x^-(k+1)$ and the middle $p+1$ terms are projective-injective.
This sequence is a truncation of the minimal injective coresolution of $[P_x^{\succ k}]_i$ in $\mod A^{(m)}$.
\item For any $0< i\leq m$, we have an exact sequence of $A^{(m)}$-modules:
\begin{align}
\eta(x,k,i):\qquad 0 \to [I_x^{\prec k+1}]_{i-1} \to P_r \to \cdots \to P_1 \to P_0 \to [I_x^{\prec k}]_{i} \to 0,\notag
\end{align}
where $r=s_x^+(k+1)-s_x^+(k)$ and the middle $r+1$ terms are projective-injective.
This sequence is a truncation of the minimal projective coresolution of $[I_x^{\prec k}]_i$ in $\mod A^{(m)}$.
\end{enumerate}
\end{lemma}
\begin{proof}
As in the proof of Lemma \ref{lem-injcopres}, we take the minimal injective coresolution of $P_x^{\succ k}$ in $\mod A$, say, $0\to P_x^{\succ k}\to \overline{I}^0 \to \cdots \to \overline{I}^p\to 0$ where $p=s_x^-(k)-s_x^-(k+1)$, and take $I^j := [\overline{I}^j,\nu_A^-(\overline{I}^j)]_i$ for all $0\leq j\leq p$.
The exactness at $[P_x^{\succ k}]_i$ and at $I^j$ for $0<j<p$ follows from Lemma \ref{lem-vanish} (1).
The exactness at the last two non-zero terms follows from the fact that $P_x^{\succ k+1}\cong \nu^{-1}(P_x^{\succ k})[s_x^-(k)-s_x^-(k+1)]$.
\end{proof}
Similarly, we can define the analogues $\xi(x,k,m)$ of (\ref{eq-seq3}), which will be an exact sequence containing $s_x^-(k)-s_x^-(k+1)$ terms of injective non-projective $A^{(m)}$-modules.
The minimal injective coresolution of the $A^{(m)}$-module $[P_x^{\succ k}]_i$ can now be obtained by iteratively taking the Yoneda products of
\[
\xi(x,k,i), \xi(x,k+1,i+1), \cdots, \xi(x,k+m-1-i,m-1), \xi(x,k+m-i,m).
\]
Dually, we can define an exact sequence $\eta(x,k,0)$ which has $s_x^+(k+1)-s_x^+(k)+1$ terms of projective non-injective $A^{(m)}$-modules, and the minimal projective resolution of the $A^{(m)}$-module $[I_x^{\prec k}]_i$ can be obtained by iteratively taking the Yoneda products of
\[
\eta(x,k+i,0), \eta(x,k+i-1,1), \cdots, \eta(x,k+1,i-1), \eta(x,k,i).
\]
This gives the following analogue of Proposition \ref{prop-dim}; note that $s_x^-(k)$ here is equal to $s_{P_x}(k)-k$ in the notation of the previous section.
\begin{proposition}\label{prop-dim2}
Let $m$ be a positive integer, $x\in \mathcal{S}$, and $k\geq 0$.
Then we have the following formulae for all $0\leq i\leq m$.
\begin{align}
\operatorname{domdim}_{A^{(m)}}[P_x^{\succ k}]_i &= m-i+s_x^-(k)-s_x^-(k+m-i), \notag\\
\operatorname{idim}_{A^{(m)}}[P_x^{\succ k}]_i &= m-i+s_x^-(k)-s_x^-(k+m-i+1), \notag \\
\operatorname{codomdim}_{A^{(m)}}[I_x^{\prec k}]_i &= i+s_x^+(k+i)-s_x^+(k), \notag\\
\operatorname{pdim}_{A^{(m)}}[I_x^{\prec k}]_i &= i+s_x^+(k+i+1)-s_x^+(k). \notag
\end{align}
\end{proposition}
\begin{proof}
We prove the first two formulae; the later two are proved dually.
For each $0\leq j\leq m-i$, the sequence $\xi(x,k+j,i+j)$ contributes precisely $s_x^-(k+j)-s_x^-(k+j+1)+1$ injective terms in the minimal injective coresolution of $[P_x^{\succ k}]_i$.
Therefore, the injective dimension is given by
\[
-1+\sum_{j=0}^{m-i}\left( s_x^-(k+j)-s_x^-(k+j+1)+1\right) = m-i + s_x^-(k) - s_x^-(k+m-i+1).
\]
For each $j<m-i$, all of the middle terms in $\xi(P,k+j,i+j)$ are projective(-injective); whereas for the case when $j=m-i$, none of them are projective.
Therefore, the dominant dimension is just $\operatorname{idim}_{A^{(m)}} [P_x^{\succ k}]_i-(s_x^-(k+m-i)-s_x^-(k+m-i+1))$.
\end{proof}
The following is our main result concerning the replicated algebras of $\nu$-formal algebras.
In the special case of $d$-hereditary algebras, due to the regularity in the behaviour of the functions $s_i^\pm$, the corresponding results give nicer formulae.
For the convenience of reader, we will state them separately in subsection \ref{subsec-NF-d-hered}.
\begin{theorem}\label{thm-dims2}
Let $A$ be a $\nu$-formal algebra, then for all $m\geq 1$, we have
\begin{align*}
\operatorname{codomdim} DA^{(m)} = \operatorname{domdim} A^{(m)} & = m-\max\{s_x^-(m)\mid x\in \mathcal{S}\} = m+\min\{s_x^+(m) \mid x\in \mathcal{S}\}, \\
\text{ and } \quad \operatorname{pdim} DA^{(m)} = \operatorname{idim} A^{(m)} &= m-\min\{s_x^-(m+1)\mid x\in\mathcal{S}\} = m+\max\{s_x^+(m+1)\mid x\in\mathcal{S}\}.
\end{align*}
In particular, $A^{(m)}$ is Iwanaga-Gorenstein.
\end{theorem}
\begin{proof}
By definition, we have
\begin{align}
\operatorname{domdim} A^{(m)} & = \min\{ \operatorname{domdim} [P_x]_0 \mid x\in\mathcal{S} \}, & \operatorname{idim} A^{(m)} &= \max\{ \operatorname{idim} [P_x]_0 \mid x\in\mathcal{S} \},\notag \\
\operatorname{codomdim} DA^{(m)} & = \min\{ \operatorname{codomdim} [I_x]_0 \mid x\in\mathcal{S} \}, & \operatorname{pdim} DA^{(m)} &= \max\{ \operatorname{pdim} [I_x]_0 \mid x\in\mathcal{S} \}.\notag
\end{align}
Recall (see Section \ref{sec-Prelim}) that for any algebra $\Lambda$, we always have $\operatorname{domdim} \Lambda=\operatorname{codomdim} D\Lambda$, so the two values in the first column are equal; similarly, we always have $\operatorname{idim} \Lambda = \operatorname{pdim} D\Lambda$ whenever both of them are finite, which gives the equality of the two values in the second column.
Now the claim follows from applying Proposition \ref{prop-dim2} and the fact that $s_x^\pm(0)=0$ for all $x\in\mathcal{S}$.
\end{proof}
One easy consequence of this theorem is that
\[
\min\{-s_x^-(k)\mid x\in\mathcal{S}\} = \min\{s_x^+(k)|x\in\mathcal{S}\} \quad \text{and}\quad \max \{ -s_x^-(k) \mid x\in\mathcal{S}\} = \max\{ s_x^+(k)\mid x\in\mathcal{S}\}
\]
for all $k\geq 0$.
\begin{corollary}\label{cor-minAG}
Let $A$ be a $\nu$-formal algebra and fix some $m\geq 1$.
Then the following are equivalent
\begin{enumerate}[(1)]
\item the $m$-th replicated algebra $A^{(m)}$ is minimal Auslander-Gorenstein;
\item $s_x^-(m)=s_y^-(m+1)$ for all $x,y\in\mathcal{S}$;
\item $P_x^{\succ m}$ is injective for all $x\in\mathcal{S}$;
\item $s_x^+(m)=s_y^+(m+1)$ for all $x,y\in\mathcal{S}$;
\item $I_x^{\succ m}$ is projective for all $x\in\mathcal{S}$;
\item $\nu^{-(m+1)}(A)\cong A[n]$ for some $n\in\mathbb{Z}$.
\end{enumerate}
\end{corollary}
\begin{proof}
By Theorem \ref{thm-dims2}, we have equivalence between (1) and any of the following conditions.
\begin{enumerate}[(i)]
\item $\max\{s_x^-(m)\mid x\in \mathcal{S}\}=\min\{s_x^-(m+1)\mid x\in \mathcal{S}\}$.
\item $\min\{s_x^+(m)\mid x\in \mathcal{S}\}=\max\{s_x^+(m+1)\mid x\in \mathcal{S}\}$.
\end{enumerate}
Since $s_x^-$ is a decreasing function over $\mathbb{Z}_{\geq 0}$ for each $x\in \mathcal{S}$, we have
\[
\max\{s_x^-(m)\mid x\in\mathcal{S}\} \geq s_y^-(m) \geq s_y^-(m+1) \geq \min\{s_x^-(m+1)\mid x\in\mathcal{S}\},
\]
for all $y\in \mathcal{S}$.
Therefore, (i) is equivalent to the condition (2), which in turn is equivalent to (3).
Dually, (ii) is equivalent to (4), which in turn is also equivalent to (5).
Given (6), then (2) and (3) are satisfied with $s_x^-(m)=n=s_y^-(m+1)$ for all $x,y\in\mathcal{S}$.
Conversely, given (2) (and hence, (3)), then clearly we get (6) with $n=s_x^-(m)$.
\end{proof}
\begin{corollary}\label{cor-minAG2}
Suppose $A$ is a $\nu$-formal algebra.
Then the following are equivalent.
\begin{enumerate}[(1)]
\item $A$ is periodically $\nu$-formal.
\item There exists some $m\in \mathbb{Z}_+$ such that $A^{(m)}$ is minimal Auslander-Gorenstein.
\item There exists infinitely many $m\in \mathbb{Z}_+$ such that $A^{(m)}$ is minimal Auslander-Gorenstein.
\end{enumerate}
Moreover, when $A$ is periodically $\nu$-formal with twisted Calabi-Yau dimension $(h_A,-s_A)$, $A^{(m)}$ is minimal Auslander-Gorenstein if, and only if, $m=th_A-1$ for some $t>0$; in such a case, we have
\[
\operatorname{domdim} A^{(m)} = t(h_A-s_A)-1 = \operatorname{idim} A^{(m)}.
\]
\end{corollary}
\begin{proof}
The equivalence is just an immediate consequence of Proposition \ref{prop-twistedCY} and Corollary \ref{cor-minAG}.
By the minimality of $h_A$, $\nu^{-m}(A)\cong A[n]$ is equivalent to $m=th_A$ for some integer $t>0$; the final equivalence follows.
Moreover, since $\nu^{-th_A}(A)\cong A[ts_A]$, we have $s_x^-(th_A-1)=s_x^-(th_A)=ts_A$.
Applying this to Theorem \ref{thm-dims2} yields $\operatorname{idim} A^{(m)}=\operatorname{domdim} A^{(m)}=th_A-1-ts_A$ as stated.
\end{proof}
\subsection{The case of higher hereditary algebras}\label{subsec-NF-d-hered}
We specialise Theorem \ref{thm-dims2} and its corollaries for the case when $A$ is a ring-indecomposable non-simple $d$-hereditary algebra with $d\geq 1$, i.e. $\operatorname{gldim} A=d$ and $\nu^i(A) \in \operatorname{\mathrm{add}}\bigcup_{k\in \mathbb{Z}}(\mod A)[dk]$ for all $i\in \mathbb{Z}$.
Recall the following subclasses of $d$-hereditary algebras.
\begin{definition}\label{def-d-hered}
Let $d$ be a positive integer.
\begin{itemize}
\item An algebra $A$ is \emph{$d$-representation-finite} if $\operatorname{gldim} A\leq d$ and for each indecomposable projective $A$-module $P$, there is some $t_P\geq 0$ so that $\nu^{-t_P}(P[dt_P])$ is an indecomposable injective $A$-module.
\item An algebra $A$ is \emph{$d$-representation-infinite} if $\operatorname{gldim} A\leq d$ and $\nu^{-i}(A[di]) \in \mod A$ for all $i\geq 0$.
\end{itemize}
\end{definition}
In order to help comparing these algebras with the more general $\nu$-formal algebras, the definitions given here are not presented in the same way as the usual one in the literature.
Note that the usual definition of $d$-representation-finite algebra is stated in terms of the existence of $d$-cluster-tilting module and the requirement on global dimension; the two definitions are equivalent by \cite[Thm 1.23]{Iya2}.
It is clear from the definition that $d$-representation-finite algebras are periodically $\nu$-formal.
More generally, as shown in \cite[Prop 2.9, 3.3]{HIO}, any $d$-hereditary algebra is also $\nu$-formal.
Moreover, the statement of \cite[Lemma 3.5]{HIO} says precisely that $d$-representation-infinite algebras are aperiodically $\nu$-formal.
The following dichotomy theorem says that there does not exists any aperiodically $\nu$-formal algebra that is not $d$-representation-infinite.
\begin{theorem}[Dichotomy Theorem]{\rm \cite[Theorem 3.4]{HIO}}
Any ring-indecomposable $d$-hereditary algebra is either $d$-representation-finite or $d$-representation-infinite.
\end{theorem}
From now on, we assume that $A$ is a ring-indecomposable non-simple $d$-hereditary algebra.
As before, we use $\mathcal{S}$ to denote the indexing set of simple $A$-modules.
For $x\in \mathcal{S}$, the functions $s_x^\pm : \mathbb{Z}_{\geq 0}\to \pm \mathbb{Z}_{\geq 0}$ can be described more concretely relative to general $\nu$-formal algebras as follows.
\[
s_x^\pm(k) = \pm \left(k-r_x^\pm(k)\right)d,
\]
where $r_x^-(k)$ (resp. $r_x^+(k)$) is the number of (indecomposable) injective (resp. projective) modules appearing in the sequence $(P_x^{\succ i})_{0\leq i<k}$ (resp. $(I_x^{\prec i})_{0\leq i<k}$).
If $A$ is, moreover, $d$-representation-infinite, then $r_x^\pm(k)=0$ for all $x\in\mathcal{S}$ and $k\geq 0$.
For simplicity, we only consider the results involved on the $s_x^-$ side, since the other side can be treated dually.
Now Theorem \ref{thm-dims2} gives us the following.
\begin{corollary}\label{cor-dim-dhered}
If $A$ is an indecomposable non-simple $d$-hereditary algebra, then we have
\begin{align}
\operatorname{idim} A^{(m)} = \operatorname{pdim} DA^{(m)} &= (d+1)m+d-(\min\{r_x^-(m+1)\mid x\in\mathcal{S}\})d,\notag \\
\text{and}\quad \operatorname{domdim} A^{(m)} = \operatorname{codomdim} DA^{(m)} & = (d+1)m-(\max\{r_x^-(m)\mid x\in\mathcal{S}\})d. \notag
\end{align}
In particular, if $A$ is $d$-representation-infinite, then
\[
\operatorname{gldim} A^{(m)} = \operatorname{idim} A^{(m)} = (d+1)m+d \quad \text{ and }\quad \operatorname{domdim} A^{(m)} = \operatorname{codomdim} DA^{(m)} = (d+1)m.
\]
\end{corollary}
By Proposition \ref{prop-twistedCY} (or \cite[Theorem 1.1]{HerIya}), $A$ being $d$-representation-finite means that it is twisted fractionally Calabi-Yau of dimension $(h_A,-s_A)$.
Note that $h_A$ is given by $\ell_x+\ell_{\sigma(x)}+\cdots +\ell_{\sigma^{r-1}(x)}$ for some $0<r\leq k_A$ independent of $x\in\mathcal{S}$, and $s_A=s_x^-(h_A)=(r_x^-(h_A)-h_A)d=(r-h_A)d$.
Therefore, we have the following $d$-hereditary specialisation of Corollary \ref{cor-minAG2}.
\begin{proposition}\label{prop-RF-d}
Suppose $A$ is a ring-indecomposable non-simple $d$-representation-finite algebra of twisted Calabi-Yau dimension $(h,(h-r)d)$.
$A^{(m)}$ is higher Auslander algebra if, and only if, $m=th-1$ for some $t>0$.
Moreover, in such a case, we have
\[ \operatorname{gldim} A^{(m)} = ((d+1)h-dr)t-1 = \operatorname{domdim} A^{(m)} \]
\end{proposition}
\subsection{Further example I: Quotient of linear oriented path algebra of type A}\label{subsec-NF-naka}
Consider the quiver $\vec\mathbb{A}_n$ given by linearly oriented Dynkin quiver of type $\mathbb{A}_n$ (with $n>1$):
\[
\vec\mathbb{A}_n:= \xymatrix{ 1 \ar[r] & 2 \ar[r] & 3 \ar[r] & \cdots \ar[r]& {n-1} \ar[r] & {n}}.
\]
In this subsection, we will classify all the ring-indecomposable quotient algebras of $K\vec\mathbb{A}_n$ which are $\nu$-formal.
It turns out that $\nu$-formality is equivalent to higher representation-finiteness.
Let us recall some background information and fix some notations first.
Since $\vec\mathbb{A}_n$ is acyclic, the global dimension of any quotient $A$ of $K\vec\mathbb{A}_n$ is always finite.
In particular, it is Iwanaga-Gorenstein with $\operatorname{gldim} A = \operatorname{idim} A_A$.
Our calculation depends only on a combinatorial tool which uniquely determines the Morita equivalence classes of Nakayama algebras - the Kupisch series.
Suppose $A$ is a Nakyama algebra with (Ext-)quiver $Q$, then its \emph{Kupisch series} $[c_1,c_2,\ldots ,c_n]$ is given by $c_i$ being the length of the indecomposable projective $A$-module corresponding to the vertex $i\in Q_0$.
For example, the algebra $K\vec\mathbb{A}_n$ has Kupisch series $[n,n-1,n-2,\ldots ,2,1]$.
In our setting, since $A$ is a ring-indecomposable quotient of $\mathbb{A}_n$, the Kupisch series always satisfies the conditions $c_n=1$, $c_i \geq 2$ for $i < n$, and $c_i\leq c_{i+1}+1$.
Throughout, we will use $P_i, I_i, S_i$ to denote the indecomposable projective, injective, simple modules corresponding to $1\leq i\leq n$ respectively.
For $l\leq n$, we define $T_{n,l}:=K\vec\mathbb{A}_n/\operatorname{\mathrm{rad}}^l(K\vec\mathbb{A}_n)$ to be the Nakayama algebra with $n$ simple modules and Kupisch series $[l,l,l,\ldots,l,l-1,l-2,\ldots,2,1]$.
For the background on the computation of minimal projective resolutions and minimal injective coresolutions in Nakayama algebras, see for example \cite{Mar}.
\begin{lemma}\label{lem-naka-3types}
Suppose $A$ is a ring-indecomposable quotient algebra of $K\vec\mathbb{A}_n$ with Kupisch series $[c_1,c_2,\ldots, c_n]$. Then one of the following holds.
\begin{enumerate}[(1)]
\item There exists $1\leq i<n$ such that $c_i<c_{i+1}$.
\item There exists $1< i<n$ such that $c_{i-1}-1=c_i=c_{i+1}$.
\item There exists $l\leq n$ such that $A=T_{n,l}$.
\end{enumerate}
\end{lemma}
\begin{proof}
If (1) does not hold, then we have $c_1 \geq c_2 \geq \cdots \geq c_n=1$.
Since (3) is equivalent to having some $1\leq i<n$ with $c_1=c_2=\cdots =c_i>c_{i+1} >\cdots >c_n$, when both (1) and (3) fail to hold, then there exists some $1<i<n$ with $c_{i-1}>c_i=c_{i+1}$, which is exactly (2).
\end{proof}
\begin{lemma}\label{lem-non-nf-naka1}
Suppose $A$ is a ring-indecomposable quotient algebra of $K\vec\mathbb{A}_n$.
If there exists $1\leq i<n$ so that $c_i<c_{i+1}$, then $A$ is not $\nu$-formal.
\end{lemma}
\begin{proof}
Suppose on contrary that $A$ is $\nu$-formal.
By assumption on $c_i$, the path starting from $i$ and ending at $i+c_i$ in $\vec\mathbb{A}_n$ is zero in $A$.
This implies that the $\mathrm{top}(I_{i+c_i})\cong S_{i+1}$, but $c_i<c_{i+1}$ also implies that $\mathrm{soc}(P_{i+1})\neq S_{i+c_i}$, so $I_{i+c_i}$ is non-projective, and there is an exact sequence $P_{i+c_i+1}\to P_{i+1}\to I_{i+c_i}\to 0$.
Since $I_{i+c_i}$ being non-projective says that $\nu_A(I_{i+c_i}) = H^0(\nu I_{i+c_i}) = 0$, applying $\nu_A = -\otimes_ADA$ to the above exact sequence we get that an exact sequence $I_{i+c_i+1}\to I_{i+1}\to 0$.
As $\mathrm{top}(I_{i+c_i+1})=S_{i+1}$, this exact sequence implies that $I_{i+1}= S_{i+1}$, which contradicts the fact that $A$ is ring-indecomposable (i.e. $I_1$ is the unique simple injective module).
\end{proof}
\begin{lemma}\label{lem-non-nf-naka2}
Suppose $A$ is a ring-indecomposable quotient algebra of $K\vec\mathbb{A}_n$.
If there exists $1<i<n$ such that $c_{i-1}>c_i\leq c_{i+1}$, then $A$ is not $\nu$-formal.
\end{lemma}
\begin{proof}
As in the proof of Lemma \ref{lem-non-nf-naka1}, the assumption $c_i\leq c_{i+1}$ says that $\mathrm{top} I_{i+c_i}=S_{i+1}$.
On the other hand, the assumption $c_{i-1}>c_i$ says that $\mathrm{top} I_{i+c_i-1}=S_j$ for some $j\leq i-1$.
Therefore, the length of $I_{i+c_i-1}$ is larger than that of $I_{i+c_i}$.
This means that the opposite ring $A^\mathrm{op}$ is a ring-indecomposable quotient algebra of $K\vec\mathbb{A}_n$ satisfying the condition (1) of Lemma \ref{lem-naka-3types}, and we can apply Lemma \ref{lem-non-nf-naka1}.
\end{proof}
\begin{lemma}\label{lem-non-nf-naka3}
Consider $A= T_{n,l}$ with $n=lt+r$ for some $0\leq r< l$ and $t\geq 1$.
If $l\neq 2$ and $r\neq 1$, then $A$ is not $\nu$-formal.
\end{lemma}
\begin{proof}
First assume that $r\notin \{0,1\}$, which also forces $l\neq 2$.
Observe that by our assumption, the minimal injective coresolution of $P_{n-r+1}$ is of the form
\[
0\to P_{n-r+1} \to I_n \to I_{n-r} \to I_{n-l} \to I_{n-l-r} \to I_{n-2r}\to \cdots \to I_{l+r}\to I_l\to I_r\to 0.
\]
More precisely, the sequence $(I^k)_{k\geq 0}$ of injective modules in the coresolution is given by
\[
I^k = \begin{cases}
I_{n-lj}, & \text{if }k=2j\leq 2t; \\
I_{n-lj-r}, & \text{if }k=2j+1 < 2t;\\
0, & \text{otherwise.}
\end{cases}
\]
Applying $\nu_A^-$ to this exact sequence, we see that $\nu^{-1}(P_{n-r+1}) \cong M[-2t]$ where $M\cong P_r/\operatorname{\mathrm{rad}}^{l-r}P_r$.
Now we have an exact sequence $0\to M\to I_{l-1}\to I_{r-1}\to 0$ given by the minimal injective coresolution of $M$.
Applying $\nu_A^-$ we see that $H^i(\nu^{-1}(M))\neq 0$ for both $i=0$ and $i=1$.
Therefore, $\nu^{-2}(P_{n+r-1})[2t]=\nu^{-1}(M)$ is not a stalk complex and thus, $A$ is not $\nu$-formal.
Assume now that $r=0$ and $l\neq 2$.
We carry out a similar calculation starting with $P_n$.
The minimal injective coresolution of $P_n$ is given by
\[
0 \to P_n \to I_n \to I_{n-1} \to I_{n-l} \to I_{n-l-1} \to I_{n-2l} \to \cdots \to I_l \to I_{l-1} \to 0,
\]
i.e. for all $0\leq j< t$, the $2j$-th term is given by $I_{n-l(j-1)}$ and the $(2j+1)$-st term is given by $I_{n-lj-1}$.
Applying $\nu_A^-$ to this exact sequence, we see that $\nu^{-1}(P_n)\cong S_{l-1}[1-2t]$.
Now we have exact sequence $0\to S_{l-1}\to I_{l-1}\to I_{l-2}\to 0$.
Since $l\neq 2$, $\nu^{-1}(S_{l-1}) = (0\to P_{l-1}\to P_{l-2}\to 0)$ has non-zero cohomology at both degrees 0 and 1.
Hence, $\nu^{-2}(P_n)[2t-1] = \nu^{-1}(S_{l-1})$ is not a stalk complex, which means that $A$ is not $\nu$-formal.
\end{proof}
Finally, we note a recent result of Vaso, which classifies $d$-representation-finite quotients of $K\vec\mathbb{A}_n$.
\begin{theorem}{\rm \cite{Vas}}\label{thm-vaso}
A quotient algebra $A$ of $K\vec\mathbb{A}_n$ is $d$-representation-finite if, and only if, $A\cong T_{n,l}$ with $l=2$ or $l$ divides $n-1$; in which case, we have $d=2(n-1)/l$ respectively.
\end{theorem}
The following classification result now follows easily.
\begin{theorem}\label{thm-nu-formal-naka}
Suppose $A$ is a ring-indecomposable quotient algebra of $K\vec\mathbb{A}_n$.
Then the following are equivalent.
\begin{enumerate}[(1)]
\item $A$ is $\nu$-formal.
\item $A$ is periodically $\nu$-formal.
\item $A$ is $d$-hereditary for some $d\geq 1$.
\item $A$ is $d$-representation-finite for some $d\geq 1$.
\item $A\cong T_{n,l} = K\vec\mathbb{A}_n / \operatorname{\mathrm{rad}}^l(K\vec\mathbb{A}_n)$ with $l=2$ or $l$ divides $n-1$.
\end{enumerate}
\end{theorem}
\begin{proof}
The implication form (4) to (2) (hence, to (1)) is trivial.
Theorem \ref{thm-vaso} gives the equivalence between (3), (4), and (5).
The equivalence between (1) and (5) follows from combining Lemma \ref{lem-naka-3types} with Lemma \ref{lem-non-nf-naka1}, \ref{lem-non-nf-naka2}, and \ref{lem-non-nf-naka3}.
\end{proof}
\subsection{Further example II: Higher canonical algebra}\label{subsec-NF-canonical}
We give another class of $\nu$-formal algebras with finite global dimension, but not necessarily $d$-hereditary.
This class of algebra is called \emph{$d$-canonical algebra} (for some integer $d\geq 1$), and is first studied in \cite{HIMO} as a generalisation of the classical canonical algebras \cite{Rin,GeiLen}.
To keep exposition concise, we will give minimal detail about many of the results we use, but refer the reader to \cite{HIMO}.
Fix positive integers $n\geq 1$ and $d\geq -1$.
Let $R$ be a so-called Geigle-Lenzing complete intersection ring of Krull dimension $d+1$ with weights $p_1, p_2, \ldots, p_n$, where
each $p_i$ is a positive integer with $p_i\geq 2$.
Then there is an abelian category $\mathrm{coh}\mathbb{X}$ given by the coherent sheaves on the associated Geigle-Lenzing projective space $\mathbb{X}$.
Denote by $\mathcal{O}$ the structure sheaf of $\mathbb{X}$.
Define an abelian group $\mathbb{L}$ generated by $\vec x_1, \vec x_2, \ldots, \vec x_n, \vec c$ with relations $p_i\vec x_i - \vec c$ for all $1\leq i\leq n$.
This group acts on $\mathrm{coh}\mathbb{X}$ naturally, i.e. there are auto-equivalences $?(\vec x)$ on $\mathrm{coh}\mathbb{X}$ for any $\vec x\in \mathbb{L}$ with inverse $?(-\vec x)$.
There is a natural poset structure on the abelian group $\mathbb{L}$ generated by the underlying monoid, i.e. $\vec x \geq \vec y$ if $\vec x-\vec y$ is an element of the submonoid generated by $\vec x_i$'s and $\vec c$.
We use the usual notation $[\vec x,\vec y]$ to denote the interval $\{\vec z\mid \vec x\leq \vec z\leq \vec y\}$ under this poset structure.
By \cite[Theorem 6.1]{HIMO}, the \emph{$d$-canonical algebra} of $\mathbb{X}$ is the endomorphism ring $\operatorname{End}_\mathbb{X}(T)$ of a tilting object $T=\bigoplus_{\vec x\in [0, d\vec c]} \mathcal{O}(\vec x)$ in $\mathcal{D}$.
For $d=1$, this is precisely the canonical algebra of Ringel \cite{GeiLen}.
The following is the main result of this subsection.
\begin{theorem}\label{thm-canonical}
Any $d$-canonical algebra is $\nu$-formal.
\end{theorem}
Define an element $\vec\omega$ of $\mathbb{L}$ by
\[
\vec\omega := (n-d-1) \vec c - \sum_{i=1}^n \vec x_i.
\]
By \cite[Theorem 5.4]{HIMO}, $\mathrm{coh}\mathbb{X}$ has global dimension $d$ and the Serre functor $\nu$ on the bounded derived category $\mathcal{D}:=\mathrm{D}^{\mathrm{b}}(\mathrm{coh} \mathbb{X})$ is given by the composition of twist by $\vec \omega$ and shift by $d$, i.e. $\nu\cong (\vec\omega)[d]$.
Since there is an equivalence of triangulated categories $\mathrm{D}^{\mathrm{b}}(\mod A)\simeq\mathcal{D}$ given by $\operatorname{\mathsf{R}Hom}_\mathbb{X}(T,-)$, for the purpose of proving Theorem \ref{thm-canonical}, it suffices to understand various vanishing conditions of the Hom-space $\operatorname{Hom}_\mathcal{D}(\mathcal{O}, \mathcal{O}(\vec x)[l]) = \operatorname{Ext}_\mathbb{X}^l(\mathcal{O},\mathcal{O}(\vec x))$ for $\vec x\in \mathbb{L}$ and $l\in \mathbb{Z}_{\geq 0}$.
The following result will be useful.
\begin{proposition}\label{prop-canonical}{\rm \cite[Observation 3.1(c), Proposition 5.3]{HIMO}}
\begin{enumerate}
\item $\operatorname{Ext}_\mathbb{X}^i(\mathcal{O}(\vec x),\mathcal{O}(\vec y))\neq 0$ implies that $i=0$ or $d$.
\item $\operatorname{Hom}_\mathbb{X}(\mathcal{O},\mathcal{O}(\vec x))\neq 0$ if, and only if, $\vec x\geq 0$.
\item $\operatorname{Ext}_\mathbb{X}^d(\mathcal{O},\mathcal{O}(\vec x))\neq 0$ if, and only if, $\vec\omega-\vec x\geq 0$.
\end{enumerate}
\end{proposition}
\begin{proof}[Proof of Theorem \ref{thm-canonical}]
By construction, the equivalence $\mathcal{D} \simeq \mathrm{D}^{\mathrm{b}}(\mod A)$ given by $\operatorname{\mathsf{R}Hom}_\mathbb{X}(T,-)$ sends $\mathcal{O}(\vec x)$ to the indecomposable projective $A$-modules $P_{\vec x}$ for all $\vec x\in [0,d\vec c]$.
Hence, for any $k,p\in \mathbb{Z}$ and $\vec x\in [0,d\vec c]$, we have an isomorphism $H^p(\nu^k P_{\vec x})\cong \operatorname{Hom}_\mathcal{D}(T,\nu^k\mathcal{O}(\vec x)[p])$.
Therefore, we just need to show the this space is non-zero for at most one $p\in \mathbb{Z}$.
Note that we have isomorphisms
\[
\operatorname{Hom}_\mathcal{D}(T,\nu^k\mathcal{O}(\vec x)[p]) \cong \operatorname{Hom}_\mathcal{D}(T,\mathcal{O}(\vec x +k\vec\omega)[dk+p]) \cong \operatorname{Ext}_\mathbb{X}^{dk+p}(T,\mathcal{O}(\vec x+k\vec\omega)).
\]
It follows from Proposition \ref{prop-canonical} (1) that this space is non-zero for at most two $p$'s; namely, when $p=-dk$ or $-d(k-1)$.
In the case when the space is non-zero at $p=-dk$, by Proposition \ref{prop-canonical} (2), we have $\vec x+k\vec\omega \geq \vec y \geq 0$ for some $\vec y\in [0,d\vec c]$.
On the other hand, when the space is non-zero at $p=-d(k-1)$, then we have the following isomorphism given by the Serre duality:
\[
\operatorname{Hom}_\mathcal{D}(T,\mathcal{O}(\vec x+k\vec\omega)[d]) \cong D\operatorname{Hom}_\mathcal{D}(\mathcal{O}(\vec x+k\vec\omega)[d],T(\vec\omega)[d]).
\]
Hence, we get that $\operatorname{Hom}_\mathcal{D}(\mathcal{O}(\vec x+k\vec\omega),T(\vec\omega)) \cong \operatorname{Hom}_\mathbb{X}(\mathcal{O}(\vec x+(k-1)\vec\omega), T)$ is also non-zero.
By Proposition \ref{prop-canonical} (3),
we get that $\vec x+k\vec \omega\leq \vec z+\vec \omega$ for some $\vec z\in [0,d\vec c]$.
Thus, $\vec x+k\vec\omega \leq d\vec c+\vec\omega$.
Therefore, if $H^p(\nu^kP_{\vec x})$ is non-zero at more than one $p$, then we get that
\[
(n-1)\vec c-\sum_{i=1}^n\vec x_i=d\vec c+\vec\omega \geq \vec x+k\vec\omega\geq 0,
\]
which is a contradiction (see \cite[Observation 3.1(c)]{HIMO}).
\end{proof}
It is not difficult to say slightly more than just $\nu$-formality for $d$-canonical algebras.
\begin{proposition}
A $d$-canonical algebra is periodically $\nu$-formal if, and only if, $\mathbb{X}$ is Calabi-Yau, i.e. $\vec\omega$ is torsion in $\mathbb{L}$.
\end{proposition}
\begin{proof}
By Proposition \ref{prop-twistedCY}, we want to show that a $d$-canonical algebra $A$ is twisted fractionally Calabi-Yau precisely when $\mathbb{X}$ is Calabi-Yau.
From the equivalence $\operatorname{\mathsf{R}Hom}_\mathbb{X}(T,-):\mathcal{D}\xrightarrow{\sim}\mathrm{D}^{\mathrm{b}}(\mod A)$, we get that $A$ is twisted fractionally Calabi-Yau if, and only if, there is some non-zero $k\in \mathbb{Z}$ so that $\nu^k(T)$ is isomorphic to a shift of $T$.
Since there is a natural isomorphism $\nu\simeq (\vec\omega)[d]$, we have
\[
\nu^k(T)\cong \bigoplus_{\vec x\in [0,d\vec c]}\mathcal{O}(\vec x+k\vec\omega)[dk] \cong \bigoplus_{\vec x\in [k\vec\omega, d\vec c+k\vec\omega]}\mathcal{O}(\vec x)[dk]
\]
for all $k\in \mathbb{Z}$.
Therefore, $\nu^k(T)$ is isomorphic to a shift of $T$ if, and only if, $k\vec\omega=0$, i.e. $\vec\omega$ is torsion.
\end{proof}
\section{SGC extension algebras}\label{sec-SGC}
In \cite[Example 2]{Yam}, Yamagata considered an iterative construction of algebras, starting from hereditary algebras.
We call such a construction the \emph{SGC extensions} of the original algebra, where SGC stands for \underline{s}mallest \underline{g}enerator-\underline{c}ogenerator.
He observed that the global and dominant dimensions of the algebras in such a series is always finite (without detailed proof) - in fact, this inspires all the investigations presented in this article, and this section in particular.
We will first show a connection between SGC extension algebras and replicated algebras when the original algebra satisfies certain Hom-vanishing condition in subsection \ref{subsec-SGC-replicate}.
In particular, the two constructions almost always coincide when the base algebra is hereditary, which allows us to recover, and improve, the observation of Yamagata by utilising the results from the previous sections (namely, Theorem \ref{thm-dims}).
In subsection \ref{subsec-SGC-typeA}, we give explicit calculation on the homological dimensions of SGC extensions of certain algebras.
These calculations allow us to cover some of the ``bad" cases where Theorem \ref{thm-dims} and Theorem \ref{thm-dims2} fail to generalise to SGC extensions in a straightforward manner.
Finally, we give some further remarks about SGC extensions in subsection \ref{subsec-SGC-remarks}.
Let us start by defining SGC extensions properly.
\begin{definition}
Let $A$ be a finite dimensional $K$-algebra.
The \emph{SGC extension} of $A$ is the endomorphism ring $\operatorname{End}_A(A\oplus DA)$.
More generally, by taking $A^{[0]}:=A$, we define a series of algebras $\{A^{[m]}\}_{m\geq 0}$ iteratively by $A^{[m+1]}:=\operatorname{End}_{A^{[m]}}(A^{[m]}\oplus DA^{[m]})$.
The algebra $A^{[m]}$ is called the \emph{$m$-th SGC extension} of $A$.
For simplicity, the basic algebra of $A^{[m]}$ is called the \emph{$m$-th basic SGC extension} of $A$.
\end{definition}
\subsection{Relation to replicated algebras}\label{subsec-SGC-replicate}
The following result shows that, when the projective and injective modules satisfy a specific condition, then the $m$-th basic SGC extension algebra is isomorphic to an idempotent truncation of the $m$-th replicated algebra.
\begin{theorem}\label{thm-SGC-replicate}
Let $A=A^{[0]}$ be a finite dimensional algebra and $e$ be an idempotent of $A$ so that $\operatorname{\mathrm{add}} A \cap \operatorname{\mathrm{add}} DA = \operatorname{\mathrm{add}} (1-e)A$.
If $\operatorname{Hom}_A(DA,eA)=0$, then the $m$-th SGC extension $A^{[m]}$ of $A$ is Morita equivalent to an idempotent truncation
\[
e^{[m]}A^{(m)}e^{[m]} =
\left(\begin{array}{cccccc}
eAe & 0 & 0 & \cdots & 0 & 0 \\
e(DA)e & eAe & 0 & & & 0 \\
0 & \ddots & \ddots & & & \vdots \\
\vdots & & \ddots & \ddots & &\vdots \\
0 & & & e(DA)e & eAe &0 \\
0 & 0 & \cdots & 0 & (DA)e & A
\end{array}\right).
\]
of the $m$-replicated algebra $A^{(m)}$, where the idempotent $e^{[m]}$ is given by an $(m+1)\times(m+1)$ diagonal matrix with entries $(e,e,\ldots, e,1)$.
In particular, $A^{[m]}$ is Morita equivalent to
\[
\left(\begin{array}{cc}(eAe)^{(m)} & 0 \\ \ [(1-e)Ae]_m & (1-e)A(1-e)\end{array}\right).
\]
\end{theorem}
\begin{proof}
We prove the theorem by induction on $m$.
For $m=1$, by definition, $A^{[1]}$ is Morita equivalent to $\operatorname{End}_A(eA\oplus DA)$, which can be written as
\begin{align}
\left(\begin{array}{cc}\operatorname{Hom}_A(eA,eA) & \operatorname{Hom}_A(DA,eA) \\ \operatorname{Hom}_A(eA,DA) & \operatorname{Hom}_A(DA,DA)\end{array}\right)
& \cong \left(\begin{array}{cc}eAe & 0 \\ (DA)e & A\end{array}\right) \notag \\
& \cong \left(\begin{array}{cc}e & 0 \\ 0 & 1 \end{array}\right) \left(\begin{array}{cc}A & 0 \\ DA & A \end{array}\right) \left(\begin{array}{cc}e & 0 \\ 0 & 1 \end{array}\right) \cong e^{[1]} A^{(1)} e^{[1]}.\notag
\end{align}
Suppose the hypothesis is true for some $m\geq 1$, i.e. $A^{[m]}$ is Morita equivalent to the basic algebra $B:=e^{[m]} A^{(m)} e^{[m]}$.
Using the matrix form of $B$, it can be seen that the projective $B$-modules $[(DA)e,A]_{m-1}$ and $[e(DA)e,eAe]_i\cong [D(eAe),eAe]_i$ for $0\leq i\leq m-2$ are also injective.
Moreover, the projective $B$-module $[(DA)e, A]_{m-1}$ is isomorphic to $[e(DA)e,eA]_{m-1}\oplus [(1-e)(DA)e,(1-e)A]_{m-1}$.
Since $(1-e)(DA)e=0$ and $(1-e)A\in \operatorname{\mathrm{add}} DA$ by our assumption, we have $[(DA)e,A]_{m-1}$ being injective.
Thus, $[eAe]_0\oplus DB$ is the smallest generator-cogenerator and $A^{[m+1]}$ is Morita equivalent to $\operatorname{End}_{B}([eAe]_0\oplus DB)$, which can be written as
\[ \left(\begin{array}{cc}
\operatorname{Hom}_{B}\Big([eAe]_0,[eAe]_0\Big) & \operatorname{Hom}_{B}\Big(DB,[eAe]_0\Big) \\
\operatorname{Hom}_{B}\Big([eAe]_0,DB\Big) & \operatorname{Hom}_{B}\Big(DB, DB\Big)
\end{array}\right) \cong \left(\begin{array}{cc}
\operatorname{Hom}_{eAe}\Big(eAe,eAe\Big) & \operatorname{Hom}_{B}\Big(DB,[eAe]_0\Big) \\
\operatorname{Hom}_{B}\Big([eAe]_0,DB\Big) & B
\end{array}\right).
\]
Note that $DB = \left(\bigoplus_{i=0}^{m-2} [e(DA)e, eAe]_i\right)\oplus [e(DA)e,eA]_{m-1}\oplus [DA]_{m}$, so all direct summands of $DB$ apart from $[e(DA)e,eAe]_0$ has $0$-th entry being zero,
which means that the upper-right entry is
\[
\operatorname{Hom}_{B}\Big(DB,[eAe]_0\Big) \cong \operatorname{Hom}_{B}\Big([e(DA)e,eAe]_0,[eAe]_0\Big)
\]
By the description of maps between modules we gave in Section \ref{sec-Repl-Alg},
any map in this space is given by $(f,0):(e(DA)e,eAe)\to (eAe,0)$ with $f\circ \mathrm{id}=0$; hence, this space is always zero.
For the bottom-left entry, we have isomorphisms
\[
\operatorname{Hom}_{B}\Big([eAe]_0,DB\Big) \cong \operatorname{Hom}_{K}\Big([eAe]_0, K\Big) \cong [D(eAe)]_0^{\top} \cong [e(DA)e]_0^{\top},
\]
where $[X]_0^\top$ denotes the transpose of the row vector $[X]_0$.
Hence, $\operatorname{End}_{B}([eAe]_0\oplus DB)$ is isomorphic to $\left(\begin{array}{cc} eAe & 0 \\ \ [e(DA)e]_0^\top & B \end{array}\right)$.
If we write $B=e^{[m]}A^{(m)}e^{[m]}$ in its matrix form, then $e(DA)e$ in the bottom-left entry is regarded as a column vector with $e(DA)e$ in the top entry and everywhere else zero.
It is now clear that this is precisely $e^{[m+1]}A^{(m+1)}e^{[m+1]}$ as required.
For the final statement, we show the stated matrix algebra is isomorphic to $e^{[m]}A^{(m)}e^{[m]}$.
By our assumption, we have $\operatorname{Hom}_A((1-e)A,eA)\cong eA(1-e)=0$.
This means that we have
\[
(DA)e = \left( \begin{array}{c} e(DA)e \\ 0 \end{array}\right) \text{ and }A\cong \left( \begin{array}{cc}eAe&0 \\ (1-e)Ae & (1-e)A(1-e)A\end{array}\right)\]
So we can replace the bottom row of $e^{[m]}A^{(m)}e^{[m]}$ by a $2\times (m+2)$-matrix
\[
\left(\begin{array}{cccccc}
0 & \cdots & 0 & e(DA)e & eAe & 0\\
0 & \cdots & 0 & 0 & (1-e)Ae & (1-e)A(1-e)
\end{array}.\right)
\]
Thus, we obtain a $(m+2)\times (m+2)$-matrix with the top-left $(m+1)\times(m+1)$-submatrix being $(eAe)^{(m)}$.
This completes the proof.
\end{proof}
Note that if $A$ is $d$-hereditary or $d$-canonical, then the condition of Theorem \ref{thm-SGC-replicate} is satisfied.
If $A$ is self-injective, then Theorem \ref{thm-SGC-replicate} simply says that $A^{[m]}$, which is Morita equivalent to $A$, is the ``embedded in the corner entry" of $A^{(m)}$.
The following follows immediately from Theorem \ref{thm-SGC-replicate} by taking $e=1$.
\begin{corollary}\label{cor-e=1}
Let $A$ be a $\nu$-formal algebra with $\operatorname{Hom}_A(DA,A)=0$.
Then $A^{(m)}$ and $A^{[m]}$ are Morita equivalent for all $m\geq 1$.
\end{corollary}
\begin{example}\label{eg-naka-SGC}
Consider the Nakayama algebra $A=T_{n,l+1}=K\vec\mathbb{A}_n/\operatorname{\mathrm{rad}}^{l+1}(K\vec\mathbb{A}_n)$.
Then $A$ satisfies the condition of Theorem \ref{thm-SGC-replicate} with $e=e_{n-l+1}+e_{n-l+2}+\cdots +e_n$, where $e_i$ is the primitive idempotent corresponding to the vertex $i\in \vec\mathbb{A}_n$.
Then $A$ can be presented as a matrix algebra with $(i,j)$-entry $K$ for all $1\leq i\leq n$ and $i-l\leq j\leq i$; zero otherwise.
Then $eAe$ is given by the top-left $l\times l$-matrix with $(1-e)Ae$ being a $(n-l)\times l$-matrix and $(1-e)A(1-e)$ being the bottom-right $(n-l)\times (n-l)$-matrix.
Note that $eAe\cong K\vec \mathbb{A}_{l}$ and $(1-e)A(1-e)\cong K\vec \mathbb{A}_{n-l}$.
In particular, $(eAe)^{(m)}\cong T_{l(m+1),l+1}$ as we have mentioned in Example \ref{eg-typeA-Kronecker} (1).
It now follows from the theorem and the matrix presentation of $T_{l(m+1),l+1}$ that $A^{[m]}$ can be presented by a square matrix of size $l(m+1)+(n-l)=n+ml$ with $(i,j)$-entry $K$ for all $1\leq i\leq n+ml$ and $i-l\leq j\leq i$; zero otherwise.
In other words, the basic algebra of $T_{n,l+1}^{[m]}$ is $T_{n+ml,l+1}$.
\end{example}
We list two easy consequences of Corollary \ref{cor-minAG2}.
\begin{corollary}\label{cor-SGC-replicate}
The following are equivalent for a $\nu$-formal algebra $A$ satisfying $\operatorname{Hom}_A(DA,A)=0$.
\begin{enumerate}[(1)]
\item $A$ is periodically $\nu$-formal.
\item There exists some $m\in \mathbb{Z}_+$ such that $A^{[m]}$ is minimal Auslander-Gorenstein.
\item There exists infinitely many $m\in \mathbb{Z}_+$ such that $A^{[m]}$ is minimal Auslander-Gorenstein.
\end{enumerate}
\end{corollary}
In \cite{Iya1} (resp. \cite{IS}), it is shown that any $n$-Auslander algebra (resp. $n$-Auslander-Gorenstein algebra) $\Gamma$ is isomorphic to the endomorphism ring of a so-called \emph{$n$-cluster-tilting} (resp. \emph{$n$-precluster-tilting}) $\Lambda$-module $M$.
Moreover, the pair $(\Lambda,M)$ is unique up to Morita equivalence of $\Lambda$ and a choice of generator of $\operatorname{\mathrm{add}} M$.
We call $(\Lambda,M)$ the \emph{higher (pre)cluster-tilting pair corresponding to $A$}.
\begin{corollary}\label{cor-SGC-higherAus}
Suppose $A$ is a periodically $\nu$-formal algebra with $\operatorname{Hom}_A(DA,A)=0$ and has finite (resp. infinite) global dimension.
If the $(m+1)$-st replicated algebra $A^{(m+1)}$ is a higher Auslander (resp. minimal Auslander-Gorenstein) algebra, then the corresponding higher (pre)cluster-tilting pair is $(A^{[m]}, A^{[m]}\oplus DA^{[m]})$.
\end{corollary}
\begin{proof}
Under the assumption of the claim, $A^{(m+1)}$ is Morita equivalent to $A^{[m+1]}$ by Corollary \ref{cor-e=1}; hence, the later is higher Auslander (resp. minimal Auslander-Gorenstein).
The claim now follows from the higher Auslander correspondence (resp. Iyama-Solberg correspondence).
\end{proof}
When the condition $\operatorname{Hom}_A(DA,A)=0$ is replaced by the one in Theorem \ref{thm-SGC-replicate}, it is natural to ask whether we can compare the homological dimensions of the series $A^{[m]}$ and that of the series $A^{(m)}$; in particular, whether Corollary \ref{cor-SGC-replicate} still holds.
While a general method is desirable, we do not have much success.
We can, however, rescue the result in the case of hereditary algebras as follows.
\subsection{SGC extensions of certain Nakayama algebras}\label{subsec-SGC-typeA}
Theorem \ref{thm-SGC-replicate} implies that the SGC extensions of almost all hereditary algebras coincide with their replicated algebras up to Morita equivalence.
The only family of exceptions is given by hereditary Nakayama algebra (i.e. whose valued quiver is $\vec\mathbb{A}_n$).
In this subsection, we determine the formulae for the dominant and global dimensions of the SGC extensions of the Nakayama algebras with $n$ simple modules and Kupisch series $[l,l,\ldots, l,l-1,\ldots, 2, 1]$.
In particular, this gives a classification of higher Auslander algebras taking such a form.
Since the computation of the homological dimensions for Nakayama algebras depends only on the Kupisch series, we can assume that the algebras are given by $T_{n,l}:=K\vec \mathbb{A}_n/\operatorname{\mathrm{rad}}^l(K\vec \mathbb{A}_n)$.
We will also give some ring theoretic observations about these SGC extensions in comparison to Proposition \ref{prop-QF}.
\medskip
For convenience, we will use $T_{n,l}^{[m]}$ to denote the basic algebra of the $m$-th SGC extension throughout, instead of the original definition.
Since $T_{n,l}^{[m]}\cong T_{n+m(l-1),l}$ as we have shown in Example \ref{eg-naka-SGC}, we can just calculate $\operatorname{domdim} T_{n,l}$ and $\operatorname{idim} T_{n,l}$ for arbitrary $n,l$ to achieve our goal.
Note that the case of $l=2$ is already done in \cite[Example 2.4(a)]{Iya2}, but our strategy below works for the case $l=2$, too.
It turns out that there are recursive formulae determining these dimensions.
It is more convenient for us to work over an infinite dimensional algebra $A$, where $T_{n,l}$ appears as a quotient, as well as an idempotent truncation, of $A$, for all $n\geq 1$.
Let $Q$ be the linear oriented quiver of type $\mathbb{A}_\infty$, i.e.
\[
Q: \qquad 1 \to 2 \to 3 \to \cdots
\]
The infinite dimensional algebra $A$ is defined as $KQ/\operatorname{\mathrm{rad}}^l(KQ)$.
As usual, we denote by $e_i$ the primitive idempotent given by the trivial path at the vertex $i$, and $P_i, I_i, S_i$ the corresponding indecomposable projective, indecomposable injective, simple $A$-modules respectively.
The finite dimensional indecomposable $A$-modules are of the form $M_{i,s}:=e_iA/e_i\operatorname{\mathrm{rad}}^s(A)$ for some $i \geq 1$ and $s$ with $1 \leq s \leq l$.
Note that every $P_i$ is injective and isomorphic to $I_{i+l-1}$, but $I_i$ is projective only for $i\geq l$.
It is easy to see that the algebra $T_{n,l}$ is isomorphic to $eAe\cong A/A(1-e)A$ where $e=e_1+e_2+\cdots +e_n$; roughly speaking, $T_{n,l}$ is obtained from $A$ by deleting the vertices $i\in Q_0$ for all $i > n$.
Define $d_{i,s}:=\operatorname{domdim}_A(M_{i,s})$ and $g_{i,s}:=\operatorname{idim}_A(M_{i,s})$ for all $i\geq 1$ and $1 \leq s \leq l$.
For convenience, we set $d_{i,s}=0=g_{i,s}$ for any pair of integers $(i,s)$ with $i \leq 0$.
\begin{lemma}\label{lem-recursion}
Fix some $i\geq 1$.
If $1\leq s\leq l-i$, then we have
\[
d_{i,s}=0 \quad \text{ and }\quad g_{i,s}=\begin{cases}
0, & \text{if }i=1;\\ 1, & \text{if }i>1.
\end{cases}
\]
Otherwise, we have the following recursions
\[
d_{i,s}=1+d_{i+s-l,l-s} \quad \text{ and }\quad g_{i,s}=1+g_{i+s-l,l-s}.
\]
In particular, every indecomposable non-injective $A$-module $M$ satisfies $\operatorname{idim}_A M-\operatorname{domdim}_A M\leq 1$.
\end{lemma}
\begin{proof}
Recall that $M_{1,s}\cong I_s$ for $1\leq s\leq l-1$ are all the indecomposable injective non-projective $A$-modules, which means that its dominant and injective dimensions (i.e. $d_{1,s}$ and $g_{1,s}$) are both zero.
Since $I_i$ is projective is equivalent to $i\geq l$, the first term in the injective coresolution $M_{i,s}$ is projective if, and only if, $i+s-1\geq l$.
Suppose we have $s\leq l-i$.
We have already explained the case when $i=1$; for the case of $i>1$, we have $\Omega^{-1}(M_{i,s}) \cong I_{i-1}$, which yields $g_{i,s}=1$ and $d_{i,s}=0$.
In the case of $s>l-i$, we get that the injective envelop of $M_{i,s}$ is projective, and the cosyzygy $\Omega^{-1}(M_{i,s})$ is then isomorphic to $M_{i+s-l,l-s}$.
The two recursion formulae now follow.
For the last statement, observe that it holds for $M_{i,s}$ with $s<l-i$, then apply the recursions.
\end{proof}
We have the following formulae; note that $\operatorname{gldim} T_{n,l}$ is also calculated in \cite{Vas}.
\begin{proposition} \label{prop-naka-1}
Let $n\geq l$ be a positive integers and write $n=lt+r$ for some $0\leq r< l$ and $t\geq 1$.
Then we have
\begin{align}
\operatorname{gldim} T_{n,l}=\operatorname{idim} T_{n,l} = \begin{cases}
2t-1 & \text{if }r=0;\\
2t & \text{if }r=1;\\
2t+1 & \text{otherwise},
\end{cases} \quad \text{and}\quad
\operatorname{domdim} T_{n,l} = \begin{cases}
2t-1 & \text{if }r<l-1;\\
2t & \text{if }r=l-1.
\end{cases}\notag
\end{align}
In particular, $T_{n,l}$ is a higher Auslander algebra if, and only if, $l=2$ or $l$ divides $n$.
Moreover, in such a case the higher Auslander corresponding pair is $(T_{n-l+1,l},T_{n-l+1,l}\oplus D(T_{n-l+1,l}))$.
\end{proposition}
\begin{proof}
By construction, the indecomposable projective non-injective $T_{n,l}$-modules are given by $e_i T_{n,l}$ with $n-l+1< i\leq n$, which is isomorphic to $M_{i,n-l+1-i}$ as $A$-module, which means that
\begin{align}
\operatorname{idim}(T_{n,l})_{T_{n,l}} & = \max \{g_{n-i,i+1} \mid 0\leq i<l-1 \},\notag \\
\text{ and } \quad \operatorname{domdim}(T_{n,l})_{T_{n,l}} &= \min \{d_{n-i,i+1} \mid 0\leq i< l-1\}. \notag
\end{align}
It follows from Lemma \ref{lem-recursion} that $g_{n-i,i+1} \leq g_{n-j,j+1}$ and $d_{n-i,i+1}\leq d_{n-j,j+1}$ whenever $i \leq j$ for $i,j<l-1$.
In particular, it follows from the discussion above that we have $\operatorname{domdim} T_{n,k}=d_{n-l+2,l-1}$ and $\operatorname{idim} T_{n,k}=g_{n,1}$.
By the recursion in Lemma \ref{lem-recursion}, we get that
\[
g_{n,1} = g_{lt+r,1} = 1+g_{l(t-1)+r+1,l-1} = 2+g_{l(t-1)+r,1} = \cdots = 2(t-1)+1+g_{r+1,l-1}.
\]
If $r=0$, then we get that $g_{r+1,l-1}=g_{1,l-1}=0$ and $\operatorname{idim} T_{n,l} = g_{n,1} = 2(t-1)+1=2t-1$.
If $r\neq 0$, we get that
\[
\operatorname{idim} T_{n,l} = 2(t-1)+2+g_{r,1} = \begin{cases} 2t, &\text{if }r=1 \\ 2t+1, &\text{if }r>1.
\end{cases}
\]
Similarly, for the dominant dimension, we get that
\begin{align}
\operatorname{domdim} T_{n,l}&= d_{n-l+2,l-1} = d_{l(t-1)+r+2,l-1} = 1+d_{l(t-1)+r+1,1} = 2+d_{l(t-2)+r+2,l-1} = \cdots \notag \\
&= 2(t-1)+1+d_{r+1,1} = \begin{cases}
2t-1, & \text{if }r<l-1;\\
2t & \text{if }r=l-1.
\end{cases}\notag
\end{align}
The criteria of $T_{n,l}$ being higher Auslander follows immediately from these formulae.
Moreover, in such a case, it follows from Example \ref{eg-naka-SGC} (which says that $T_{n-l+1,l}^{[1]}\cong T_{n,l}$) that the higher Auslander corresponding pair is precisely as stated in our claim.
\end{proof}
Note that Proposition \ref{prop-naka-1} says that $T_{n,l}$ is a higher hereditary algebra when $l$ divides $n-1$ or $l=2$.
Hence, combining Proposition \ref{prop-naka-1} with Lemmas \ref{lem-non-nf-naka1}, \ref{lem-non-nf-naka2}, \ref{lem-non-nf-naka3} gives an alternative proof for Vaso's result (Theorem \ref{thm-vaso}).
\begin{theorem} \label{thm-naka-SGC}
Let $m$ be a positive integer.
\begin{enumerate}[(1)]
\item $T_{n,l}^{[m]}$ is a higher Auslander algebra if, and only if, $l=2$ or $l$ divides $|n-m|$.
\item $\operatorname{idim} M - \operatorname{domdim} M\leq 1$ for all $M\in \mod T_{n,l}^{[m]}$.
\item $T_{n,l}^{[m]}$ is a QF-13 algebra.
\end{enumerate}
\end{theorem}
\begin{proof}
(1): For $l>2$, apply Proposition \ref{prop-naka-1} to $T_{n,l}^{[m]}\cong T_{n+m(l-1),l}$.
For $l=2$, since $T_{n,l}^{[m]}\cong T_{n+m,2}$, the result is already shown in \cite[Example 2.4(a)]{Iya2}.
(2): Immediate from Lemma \ref{lem-recursion}.
(3):
By Proposition \ref{prop-yamagata-QF13}, we need to prove that every indecomposable $T_{n+m(l-1),l}$-module $M_{i,s}$ has $\operatorname{domdim} M_{i,s}\geq 1$ or $\operatorname{codomdim} M_{i,s}\geq 1$.
Note that the dominant dimension $d_{i,s}$ of $M_{i,s}$ is zero precisely when $i<l$ and $1\leq s\leq l-i$ (cf. Lemma \ref{lem-recursion}).
Clearly the projective cover (both as $A$- and $T_{n+m(l-1),l}$-module) $P_i$ of such $M_{i,s}$ is injective, which means that $\operatorname{codomdim} M_{i,s}\geq 1$ as required.
\end{proof}
Now, combining Corollary \ref{cor-SGC-replicate} with Theorem \ref{thm-naka-SGC}, we have the following relation between SGC extensions and representation-finiteness of hereditary algebras.
\begin{corollary}\label{cor-Yam-ext-high-Aus}
The following are equivalent for a hereditary algebra $A$.
\begin{enumerate}[(1)]
\item $A$ is representation-finite.
\item There exists some $m\in \mathbb{Z}_+$ such that $A^{[m]}$ is a higher Auslander algebra.
\item There exists infinitely many $m\in \mathbb{Z}_+$ such that $A^{[m]}$ is a higher Auslander algebra.
\end{enumerate}
\end{corollary}
\subsection{Further remarks}\label{subsec-SGC-remarks}
Recall from Proposition \ref{prop-repli-tensor-SI} that taking the $m$-replicated algebra is an operation which ``commutes" with taking the tensor product algebra with a self-injective algebra.
The analogous statement is also true for $m$-th SGC extensions.
\begin{proposition}\label{prop-tensor2}
Let $A$ be an algebra with $m$-th SGC extension $A^{[m]}$ and $B$ a basic and self-injective algebra.
The $m$-th SGC extension of $A \otimes_K B$ is Morita equivalent to $A^{[m]} \otimes_K B$.
\end{proposition}
\begin{proof}
For simplicity, we omit the subscript $K$ on $\otimes_K$.
Note that $D(A \otimes B) \cong DA \otimes DB$ and $DB \cong B$ since $B$ is assumed to be basic and self-injective. It suffices to prove the proposition for $m=1$:
\begin{align*}
(A \otimes B)^{[1]} &=\operatorname{End}_{A \otimes B}\big((A \otimes B) \oplus D(A \otimes B)\big) \cong \operatorname{End}_{A \otimes B}\big( (A \otimes B) \oplus (DA \otimes B)\big) \\
& \cong \operatorname{End}_{A \otimes B}\big( (A \oplus DA) \otimes B\big) \cong \operatorname{End}_A(A \oplus DA) \otimes \operatorname{End}_B(B) \cong A^{[1]} \otimes B. \qedhere
\end{align*}
\end{proof}
This gives us yet another new construction of minimal Auslander-Gorenstein algebras.
\begin{example}
Let $A=K\vec\mathbb{A}_2$, $B=K[x]/(x^2)$, and $C=A \otimes_K B$.
Then $C$ is a periodically $\nu$-formal algebra with
\begin{align*}
C^{[m]} \cong A^{[m]}\otimes_K B \cong T_{m+2,2}\otimes_KB,\;\;\text{ and }\;\; C^{(m)} \cong A^{(m)}\otimes_K B \cong T_{2m+2,3}\otimes_KB.
\end{align*}
By our results, $C^{[m]}$ (resp. $C^{(m)}$) is minimal Auslander-Gorenstein for all even $m$ (resp. $m=3t-1$ for some positive integer $t$).
One can show that $C^{[m]}$ is representation-finite and $C^{(m)}$ is representation-infinite for all $m \geq 1$.
\end{example}
The next example shows that SGC extensions need not preserve representation-type, Iwanaga-Gorenstein property, as well as non-Iwanaga-Gorenstein property.
\begin{example}
Let $K$ be an algebraically closed field.
Define $A:=KQ/I$ where
\[
Q:= \xymatrix{ & 1\ar@(ul,dl)_{a} \ar[r]^{b} & 2, }
\]
and $I$ is generated by $a^2, ab=0$.
Note that there is no projective-injective $A$-module but $\operatorname{Hom}_A(DA,A)$ is non-zero.
We write the associated data in the following table, where $|\mathrm{ind}\Lambda|$ means the number of isomorphism classes of indecomposable $\Lambda$-modules.
\begin{center}
\begin{tabular}{ccc}
Algebra $\Lambda$ & $|\mathrm{ind}\Lambda|$ & $\operatorname{idim} \Lambda, \operatorname{pdim} D\Lambda$ \\ \hline
$A$ & 5 & $\infty$ \\
$A^{[1]}$ & 24 & 3 \\
$A^{[2]}$ & $\infty$ & $\infty$
\end{tabular}
\end{center}
We will not detail the calculations, which can be obtained by manual calculations (or more realistically using a computer, such as the QPA package).
To see that $A^{[2]}$ is representation-infinite, one can calculate that $\dim_K A^{[2]}=28$ and there is a simple $A^{[2]}$-module $S$ with $\dim_K\tau^{10}(S)=59$.
Therefore, there is an indecomposable $A^{[2]}$-module $M$ so that $\dim_K M>\max\{30,2\dim_K A^{[2]}\}$.
It follows by a critera of Bongartz \cite[3.9]{Bo} that $A^{[2]}$ is representation-infinite.
\end{example}
|
\section{Introduction}
The AdS/CFT (or Gravity/Gauge theory) correspondence \cite{Maldacena:1997re}
is a promising tool to crack a hard problem in strongly coupled gauge theories (see \cite{Aharony:1999ti,Klebanov:2000me,DHoker:2002nbb} for some reviews), but
is still mysterious even after many pieces of evidence and application appeared after the first proposal.
Although the correspondence may be a manifestation of the closed string/open string duality, an alternative understanding might exist due to its holographic property.
One important mystery is the precise mechanism how the AdS radial direction emerges from conformal field theory (CFT). It may be a common sense that the AdS radial direction is emergent as a renormalization scale of dual CFT as both behave similarly under the dilatational symmetry \cite{Maldacena:1997re}, and
this viewpoint worked out for renormalization group flows triggered by relevant deformations \cite{Girardello:1998pd,Distler:1998gb}.
However a direct approach to search the Wilsonian cutoff corresponding to the sharp cutoff on AdS radial direction \cite{Heemskerk:2010hk} is still far from a clear answer, since the ordinary Wilsonian renormalization gives rise to non-local interaction in the bulk interpretation. See also \cite{Lee:2010ub}
\footnote{This situation may drastically change in the AdS$_3$/CFT$_2$ correspondence. A clear dictionary was proposed in \cite{McGough:2016lol}.}
While several approaches were developed to {\it grab the tail} of AdS radial direction and construct bulk dynamics from CFT \cite{Banks:1998dd,Balasubramanian:1999ri,Bena:1999jv,Lifschytz:2000bj} (see also \cite{VanRaamsdonk:2009ar,Nozaki:2012zj}),
one of the present authors, together with his collaborators, has proposed an alternative method to define a geometry from a quantum field theory and explicitly calculated the metric from several quantum field theories \cite{Aoki:2015dla,Aoki:2016ohw,Aoki:2016env}.
In Ref.~\cite{Aoki:2015dla}, the method was proposed and applied to the O$(n)$ non-linear sigma model in 2-dimensions, and the 3 dimensional metric in the large $n$ limit was shown to describe an AdS space in the massless limit.
In Ref.~\cite{Aoki:2016ohw},
it was shown in the large $n$ limit that
the induced metric describes an AdS$_{d+1}$ space with $d\ge 3$ in the UV limit and with $d \ge 1$ in the IR limit, if the method is applied to the massive O$(n)$ $\varphi^4$ model and
an appropriate normalization is introduced for the flowed field.
In Ref.~\cite{Aoki:2016env}, the large $n$ expansion was performed for the massless O$(n)$
$\varphi^4$ model in 3 dimensions. While the induced metric describes the AdS$_4$ space at the leading order, the next leading order corrections make the space asymptotically AdS only in UV and IR limits with different radii.
By observing that the induced metric always gives the AdS metric when the theory is conformal, it is natural to expect that symmetry plays a key role behind these results. In this paper, we investigate a direct relation between CFT and the AdS metric in this framework. The goal of this paper is to generalize the previous results to an arbitrary conformal field theory incorporating symmetry argument.
The rest of this paper is organized as follows.
In Sec.~\ref{sec:Metric}, after brief explanation of the proposal of Ref.~\cite{Aoki:2015dla},
we first show that our metric corresponds to the information metric stressing importance of the field normalization introduced in Ref.~\cite{Aoki:2016ohw}.
In Sec.~\ref{sec:Isometry}, we explicitly derive the AdS metric directly from the CFT.
We then prove that the induced metric in $d+1$ dimension possesses the isometry of the AdS space as a consequence of the conformal symmetry in $d$ dimensions.
Sec.~\ref{sec:Discussion} is devoted to summary and discussion.
\section{Gradient flow and information metric}
\label{sec:Metric}
\noindent
In this section, we briefly review the proposal in Ref.~\cite{Aoki:2015dla,Aoki:2016ohw} to define a $d+1$ dimensional induced metric from a $d$ dimensional quantum field theory.
We also show that the metric defined in this way with appropriate normalization can be interpreted as a quantum information metric,
called the Bures or Helstrom metric. \\
\subsection{Gradient flow and induced metric}
We consider an $n$ real component scalar field $\varphi(x)$ in $d$ dimensions, whose quantum dynamics is controlled by the action functional $S(\varphi)$. The flowed field $\phi$ is defined from $\varphi$ with the initial condition $\phi(x;0)=\varphi(x)$ through the flow equation as
\beqa
\frac{\partial \phi^a(x;t)}{\partial {t}} = - \left.\frac{\delta S_f(\varphi)}{\delta \varphi^a(x)}\right\vert_{\varphi(x)\to\phi(x;t)},
\eeqa
where
the flow time $t$ {has the (length)$^2$ dimension},
$x$ is the $d$ dimensional coordinate system, $a=1,2,\cdots, n$ labels a component of the scalar field, and $S_f(\varphi)$ is an appropriate action for $\varphi$, which is not necessarily related to the original action $S(\varphi)$ in general. Particularly when they coincide, the flow is called the gradient flow \cite{Narayanan:2006rf,Luscher:2010iy, Luscher:2009eq,Luscher:2013vga}. In the case of the free flow ({\it i.e.} $S_f$ is the free action), the flow equation becomes the heat equation. Thus the flow equation defines a procedure to smear the original field $\varphi$ into a smeared field $\phi$, correlation functions of which are all finite at $t > 0$.
A $d+1$ dimension metric operator is given by
\beqa
\hat g_{MN} (x;t) := R^2 \sum_{a=1}^n {\partial \sigma^a(x;t) \over \partial z^M} {\partial \sigma^a(x;t) \over \partial z^N},
\eeqa
where {$R$ is a constant with the length dimension, and}
$z^M=(x^\mu, \tau)$ with {$\tau=\sqrt{2 d t}$}, which is regarded as the $d+1$ dimensional coordinates after $d$ dimensional quantum averaging, and $\sigma^a(x;t)$ is the {(dimensionless)} normalized flowed field defined as
\beqa
\sigma^a(x;t) := \frac{\phi^a(x;t)}{\sqrt{\langle \sum_{a=1}^n \phi^a (x;t)^2 \rangle_S}} .
\eeqa
Here $\langle O(\varphi) \rangle_S$ denotes the quantum average with the $d$ dimensional action $S$, given by
\beqa
\langle O(\varphi) \rangle_S := \frac{1}{Z} \int {\cal D}\varphi\, O(\varphi)\, e^{-S(\varphi)}, \
Z = \int {\cal D}\varphi\, e^{-S(\varphi)}.
\eeqa
Since $\sigma^a(x;t)$ always satisfies $\displaystyle\sum_{a=1}^n \left\langle \sigma^a(x;t) \sigma^a(x;t) \right\rangle_S =1$, we call this normalization condition the non-linear sigma model (NLSM) normalization, whose importance becomes clear in the next subsection.
The vacuum expectation value {(VEV)} of the metric is denoted as $g_{MN}(z) := \langle \hat g_{MN}(x;t) \rangle_S$, {whose fluctuations are suppressed in the large $n$ limit.}
(See Ref.~\cite{Aoki:2015dla,Aoki:2016ohw,Aoki:2016env} for more details.)\\
\subsection{Information metric}
\label{InformationMetric}
In this subsection, we show that $g_{MN}(z)$ defined in the previous subsection is equivalent to a quantum information metric, called the Bures (or Helstrom) metric. The NLSM normalization is important to show this.
The Bures metric for the density matrix is defined from the infinitesimal distance between two density matrices $\rho$ and $\rho+d\rho$ as
\beqa
D(\rho,\rho+d\rho)^2 = \frac{1}{2}{\rm tr} (d\rho\, G),
\eeqa
where $G$ is the Hermitian 1-form operator implicitly given by
$
\rho\, G + G\, \rho = d\rho.
$
In particular for the density matrix $\rho$ of a pure state, the Hermitian operator is determined as $G=d\rho$ since $\rho^2=\rho$.
In order to apply to our case, we consider an eigenstate of the position operator as well as the flow time one denoted by $\ket{z}=\ket{(x, \tau)}$ and define the inner product of the state as
\beqa
\langle z \vert w \rangle :=
\sum_{a=1}^n \langle \sigma^a(x;t) \sigma^a(y;s)\rangle_S , \quad
w=(y,\sqrt{2d s}) .
\eeqa
Notice that the NLSM normalization guarantees $\langle z \vert z \rangle =1$.
Then the information metric for this pure state is computed as
\beqa
R^2 D(\rho_z, \rho_{z+dz})^2
&=& g_{MN}(z) dz^M dz^N,
\eeqa
where $\rho_z =\vert z\rangle \langle z\vert, $ and
we used $\langle\sigma(x;t)\cdot \partial_M \sigma(x;t) \rangle_S =0$ and
\beqa
\langle\sigma(x;t)\cdot \partial_M\partial_N \sigma(x;t) \rangle_S
= - \langle \partial_M \sigma(x;t) \cdot \partial_N \sigma(x;t)\rangle_S \notag
\eeqa
with $A\cdot B := \sum_{a=1}^n A^a B^a$.
Our metric $g_{MN}(z)$ defines a distance in the space of the density matrices made of the pure states (in unit of $R$).
In particular when the action $S$ of the original theory has O$(n)$ symmetry, such pure states defined above in an abstract way may be given by $\sum_{a}\sigma^a(x;t) \vert 0 \rangle$
\footnote{In the case of the $O(n)$ invariant mixed state $\rho_z :=\sum_a \sigma^a(x;t) \vert 0 \rangle \langle 0 \vert \sigma^a(x;t)$, we have $G = n d\rho_z$ from $\rho_z^2 =\rho_z/n$.}
\section{AdS isometry from conformal symmetry}
\label{sec:Isometry}
In this section,
we directly relate arbitrary conformal field theory in the flat $d$ dimensional space-time
with the AdS metric in $d+1$ dimensions.\footnote{The argument and calculation below hold just by changing the signature suitably when we consider the Euclidean flat space.}
We here assume that the CFT contains a real scalar primary operator $\varphi(x)$ of a general conformal dimension $\Delta$ without specifying any concrete models for CFT.
{In this section, we are interested in the VEV of the metric operator, which determines the classical geometry induced from this primary operator,
though its ``quantum" fluctuations around the VEV exist for $n=1$ in this case.} \\
\subsection{AdS metric from CFT}
We start with the free generally massive flow equation {with $n=1$}
\beqa
\frac{\partial \phi(x;t)}{\partial {t}} = {(\partial^2 - m^2)} \phi(x;t),
\label{eq:free_flow}
\eeqa
where $\partial^2 = \eta^{\mu\nu} \partial_\mu\partial_\nu$.
This is easily solved as
\beqa
\phi(x;t) = e^{ {t} {(\partial^2-m^2)}}\varphi(x) .
\eeqa
The two-point function of $\phi$ is evaluated as
\beqa
G_0(x;t|y;s)&:=& \langle \phi(x;t) \phi(y;s) \rangle_{\rm CFT} \nonumber \\
&=& {e^{-(t+s)m^2}} e^{({t}\partial^2_x + s\partial^2_y)} \langle\varphi(x)\varphi(y) \rangle_{\rm CFT},
\eeqa
where we used the subscript CFT instead of $S$, since the action is not specified.
The Poincare invariance and the scale transformation of $\varphi$ with $ \varphi(\lambda x) = \lambda^{-\Delta} \varphi(x)$
fix the form of $G_0$ such that
\beqa
G_0(x;t|y;s) = \frac{{e^{-(t+s)m^2}}}{{(t+s)^\Delta}} F_0\left(\frac{(x-y)^2}{{t+s}}\right)
\eeqa
where $F_0$ is a certain smooth function.
Furthermore, the flow equation (\ref{eq:free_flow}) implies
\beqa
\frac{\partial}{\partial t} G_0(x;t|y;s)
{({\partial^2_x} -m^2)} G_0(x;t|y;s),
\eeqa
which leads to $\Delta F_0(0) = -2 d F_0^\prime (0) $.
Thus the two-point function of the normalized field $\sigma$ {becomes $m$ independent as}
\beqa
G(x;t|y;s) &:=& \langle \sigma(x;t) \sigma(y;s) \rangle_{\rm CFT} \nonumber\\
&=& \left(\frac{2\sqrt{ts}}{t+s}\right)^\Delta F\left(\frac{(x-y)^2}{{t+s}}\right),
\label{2ptsigma}
\eeqa
where $F(x) \equiv F_0(x)/F_0(0)$.
Hence $F(0)=1$ and $2 d F^\prime(0) = -\Delta$.\footnote{
Explicitly, $F(x)$ is computed as
\beqa
F(x) = \frac{\Gamma(d/2)}{\Gamma(\Delta)\Gamma(d/2-\Delta)}&\displaystyle \int_0^1& dv\, v^{\Delta-1}(1-v)^{d/2-\Delta-1}e^{-x v/4}. \nonumber
\eeqa
}
Taking the $d+1$ dimensional coordinates as $z=(x,{\tau=\sqrt{2dt}})$,
the vacuum expectation value of the induced metric $\hat g_{MN}$ is
\beqa
g_{MN}(z)
= R^2 \langle \partial_M\sigma(x;t) \partial_N\sigma(x;t) \rangle_{\rm CFT}.
\eeqa
This is computed as ($g_{\mu\tau}(z)=g_{\tau \nu}(z)=0$)
\beqa
g_{\mu\nu}(z) &=&\eta_{\mu\nu} \frac{R^2 \Delta}{\tau^2}, \quad
g_{\tau\tau}(z) =\frac{R^2 \Delta}{\tau^2},
\label{eq:AdS_metric}
\eeqa
which is nothing but the AdS metric with $R\sqrt\Delta$ its radius.
\\
\subsection{Isometry from conformal transformation}
In the previous subsection,
we have explicitly calculated the induced metric $g_{MN}$ from CFT,
and have shown that it becomes the AdS metric in the Poincare patch.
In this subsection, we argue the relation between the AdS metric and the CFT only from the symmetry.
Namely we show that the induced metric $g_{MN}(z) =\aver{\hat g_{MN}(x;t)}_{\rm CFT}$ possesses the isometry of the AdS space.
This is necessary and sufficient since AdS is a maximally symmetric space, so that the metric is completely fixed by the isometry group $SO(2,d)$ up to an overall constant.
We first relate the conformal transformation to the isometry of AdS
\footnote{
{This question was addressed in the early stage after the Maldacena's proposal in \cite{Jevicki:1998qs}, where the mechanism was studied how the special conformal transformation of the adjoint scalar fields in ${\cal N}=4$ super Yang-Mills theory was {\it metamorphosed} into the corresponding isometry transformation in the bulk.
We would like to thank T.~Yoneya for his valuable comment given in KEK Theory workshop 2017 ``East Asia Joint Workshop on Fields and Strings 2017''. }
}
The infinitesimal conformal transformation and the response of the primary scalar operator are given by
\beqa
\delta x^\mu &=& a^\mu + \omega^\mu{}_\nu x^\nu + \lambda x^\mu + b^\mu x^2 -2 x^\mu (b_\nu x^\nu), \nonumber \\
\delta^{\rm conf} \varphi(x) &=& - {\delta x}^\mu\partial_\mu \varphi(x) - {\Delta \over d} (\partial_\mu \delta x^\mu) \varphi(x).
\label{conformalTr}
\eeqa
Here $a^\mu, \omega^{\mu\nu}, \lambda, b^\mu$ are parameters of the transformation.
Since the infinitesimal conformal transformation is quadratic in the coordinate $x$, the normalized field $\sigma$, given in terms of $\varphi$ by
\beqa
\sigma (x;t) = \frac{ \left(\sqrt{2t}\right)^{\Delta}}{\sqrt{F_0(0)}}\, e^{ {t \partial^2}}\, \varphi(x),
\eeqa
is transformed a
\beqa
\delta^{\rm conf}\sigma(x;t)
&=&- \left\{{t} (\partial^2 \delta x^\mu) + 2 {t^2} (\partial^\nu \partial^\rho \delta x^\mu) \partial_\nu \partial_\rho + 2{t} (\partial^\nu \delta x^\mu) \partial_\nu + \delta x^\mu\right\} \partial_\mu \sigma(x;t) \nn &&
- {\Delta \over d}\left\{2 {t} (\partial^\nu \partial_\mu {\delta x}^\mu) \partial_\nu + (\partial_\mu {\delta x}^\mu) \right\} \sigma(x;t) .
\label{conformalTrsigma}
\eeqa
Plugging \eqref{conformalTr} into \eqref{conformalTrsigma} we obtain
\be
\delta^{\rm conf}\sigma(x;t)
= -\{ 2\lambda - 4 (b_\mu x^\mu) \}t\partial_t \sigma(x;t) - \{\delta x^\mu+2 {t} (d-2 - \Delta ) b^\mu \}\partial_\mu \sigma(x;t)
+ 4 {t}^2 b^\mu \partial_\mu \partial_t \sigma (x;t),
\label{eq:conformalTrsigma2}
\ee
where we have used $\partial_t \sigma(x;t) =({\Delta \over 2t } + {\partial^2)} \sigma(x;t) $.
{Notice that a special conformal transformation of the normalized flow operator contains a higher derivative term as the last term in \eqref{eq:conformalTrsigma2}, which cannot be interpreted as an infinitesimal diffeomorphism in the bulk.
\footnote{
{Difficulty in the bulk interpretation of the special conformal transformation of a key operator was also observed in the bilocal field approach by using a vector model \cite{Das:2003vw}, where the special conformal transformation of the collective field mixes up fields with different spin. They speculated on the necessity of a suitable field redefinition to resolve this issue, though our answer may be different in the flow field approach as shown below.
We would like to thank S.~Das for his valuable comment given in KIAS-YITP joint workshop 2017 ``Strings, Gravity and Cosmology".}
}
{To deal with this we} rewrite eq.~(\ref{eq:conformalTrsigma2}) as
\beqa
\delta^{{\rm conf}} \sigma (x;t) &=& \delta^{{\rm diff}} \sigma(x;t) + \delta^{{\rm extra}} \sigma (x;t), \\
\delta^{{\rm diff}} \sigma(x;t) &=& -(\bar\delta t \partial_t + \bar\delta x^\mu \partial_\mu)\sigma (x;t), \quad
\delta^{{\rm extra}}\sigma(x;t)
= 4 {t^2} b^\nu \partial_\nu (\partial_t + {\Delta +2 \over 2t }) \sigma (x;t),~~~~~~
\label{extraterm}
\eeqa
with $ \bar\delta x^\mu =\delta x^\mu + 2d { t} b^\mu,
\bar\delta t =(2\lambda - 4 (b_\mu x^\mu) )t$.
By setting $\tau^2 =2 d {t}$, the transformation $\delta^{\rm diff}$ can be rewritten as
\beqa
\bar\delta x^\mu =\delta x^\mu + \tau^2 b^\mu , \quad
\bar\delta \tau = (\lambda - 2 (b_\mu x^\mu))\tau .
\eeqa
This is nothing but the isometry transformation of the AdS space whose metric is given by $ds^2_{{\rm AdS}} \propto \displaystyle { d\tau^2 + dx^\mu{}^2 \over \tau^2} .$
The conformal transformation of the induced metric operator is computed as
\be
\delta^{\rm conf} \hat g_{MN}(x;t)
=\delta^{\rm diff} \hat g_{MN}(x;t) +R^2 \lim_{(y;s)\to (x;t)} {\partial \over \partial z^M} {\partial \over \partial w^N}\left\{ \delta^{{\rm extra}}\sigma(x;t) \sigma(y;s) +\sigma(x;t) \delta^{{\rm extra}}\sigma(y;s)\right\}.
\label{eq:confTrMetric}
\ee
The first term is
\beqa
\delta^{\rm diff} \hat g_{MN}(x;t) = - \bar\delta z^K \partial_K \hat g_{MN}(x;t) - \partial_M \bar\delta z^K \hat g_{KN}(x;t) - \partial_N \bar\delta z^K \hat g_{MK}(x;t),
\eeqa
which is nothing but the diffeomorphism of the metric tensor in the $d+1$ dimensions.
Thus our task is to show that the second term in eq.~(\ref{eq:confTrMetric}) vanishes in the vacuum expectation value.
By using \eqref{2ptsigma} and \eqref{extraterm} the term is computed as
\be
\aver{\delta^{{\rm extra}}\sigma(x;t) \sigma(y;s) + \sigma(x;t)\delta^{{\rm extra}}\sigma(y;s) }
= -8 { (\sqrt{4 t s})^{\Delta} \over (t+s)^{\Delta+2}}(t -s) {b_\mu(x-y)^\mu }{(x-y)^2
F''\left({(x-y)^2 \over { t+s}}\right).
\ee
Then it is easy to see
\beqa
\lim_{(y;s)\to (x;t)} {\partial \over \partial z^M} {\partial \over \partial w^N}\aver{\delta^{{\rm extra}}\sigma(x;t) \sigma(y;s) + \sigma(x;t)\delta^{{\rm extra}}\sigma(y;s) } = 0.
\eeqa
{We stress that this happens only when the conformal transformation is decomposed as \eqref{extraterm}.}
Note that the quantum averaging and the differentiation commute since all correlation functions of $\sigma$ are finite as long as the flow time is non-zero.
Therefore we obtain
\be
\aver{\delta^{\rm conf} \hat g_{MN}(x;t)} =\aver{\delta^{\rm diff} \hat g_{MN}(x;t)}.
\ee
Since the conformal invariance of the two point function of the primary scalar operator implies that $\aver{\delta^{\rm conf} \hat g_{MN}(x;t)} =0$, it follows that $\aver{\delta^{\rm diff} \hat g_{MN}(x;t)}=0$. This means that the induced metric $g_{MN}(z)$ satisfies the Killing equation of the AdS space as
\beqa
\delta^{\rm diff} g_{MN}(z) = - \bar\delta z^K \partial_K g_{MN}(z) - \partial_M \bar\delta z^K g_{KN}(z) - \partial_N \bar\delta z^K g_{MK}(z) = 0,
\eeqa
which implies that the induced metric must be the AdS metric up to an overall constant.
This completes the proof of our claim.
\\
\section{Discussion}
\label{sec:Discussion}
In this paper an induced geometry by a flow equation from a quantum field theory was investigated.
The induced metric was shown to appear as a quantum information metric, which measures a distance in the space of the pure states constructed by scalar fields in a general quantum field theory.
In a conformally symmetric situation, it was shown that the induced metric matches the AdS one when the flow equation is free
\footnote{{
This result does not depend on mass $m$ in the flow equation, which violates the conformal symmetry,
thanks to the NLSM normalization.}}
This agreement was confirmed only by using symmetry without any explicit computations of the metric.
{An appearance of the AdS metric from CFT in our method relies on the following two facts.
(a) The field $\sigma$ used to define the metric operator is dimensionless thanks to the NLSM normalization. If $\phi$ were used instead, one would not obtain the AdS metric.
(b) The VEV of the metric operator is UV finite thanks to the free flow equation.
If the VEV were UV divergent, one would not obtain the AdS metric due to additional dimensionful quantities introduced through renormalization.
}
So far any relation between the induced metric formalism presented in this paper and other approaches to see dual geometry mentioned in the introduction is not known.
It may be reasonable to think that there is no relation between them since, for example, the procedure to renormalize fields in quantum field theory and that to smear operators are generally independent.
Still we expect that the results and technique developed in this paper, particularly the symmetry argument, will become useful to study the AdS geometry from CFT by {other methods}.
{For example, it may be possible to define an induced metric similarly in the Wilsonian renormalization approach to the dual gravity. Then it would be interesting to see whether the metric becomes the AdS one or not.}
In this paper we assume that the background of quantum field theory or conformal field theory is flat.
It would be interesting to extend the presented calculation to curved backgrounds.
In particular it would be curious to check whether the induced metric from CFT on a curved space-time is still of the AdS form in a different coordinate system from the Poincare patch.
A challenging but important issue is whether this formalism encodes the gravitational dynamics or not.
The first step toward this goal may be to see how a linearized Einstein gravity is encoded in this formulation, as shown in a different method to derive dual bulk dynamics by using the entanglement entropy \cite{Faulkner:2013ica,Bhattacharya:2013bna}.
For this analysis it will be necessary to specify a concrete model to test the proposal such as an O$(n)$ sigma model, since bulk dynamics is dependent on each CFT. {Note that the $1/n$ expansion becomes important to see the dual bulk dynamics beyond the geometry in AdS/CFT correspondence. }
A virtue of this formulation is that observables in this formalism are correlation functions of scalar fields in quantum field theory, which admit analytic computation by the ordinary technique of the $1/n$ expansion \cite{Aoki:2016env}, so that one can proceed by checking one's guesswork explicitly by hand.
We hope to report on resolutions on these issues in the near future. \\
\section*{Acknowledgment}
S. A. would like to thank Dr. G.~Ishiki for his comment on the information metric.
{We would also like to thank S.~Das and T.~Yoneya for valuable comments.}
S. A. is supported in part by the Grant-in-Aid of the Japanese Ministry of Education, Sciences and Technology, Sports and Culture (MEXT) for Scientific Research (No. JP16H03978),
by a priority issue (Elucidation of the fundamental laws and evolution of the universe) to be tackled by using Post ``K" Computer,
and by Joint Institute for Computational Fundamental Science (JICFuS).
\bibliographystyle{utphys}
|
\section{Admissible Bundles and Instanton Floer Homology}
\label{admissible}
Let $E$ be a hermitian vector bundle of rank $N$ over a 3-manifold $M$. Then the vector bundle $E$ is determined up to isomorphism by its first Chern class. The pair $(M,E)$ is called an {\it admissible pair} if there is an oriented embedded surface $S$ in $M$ such that the pairing of $c_1(E)$ and the fundamental class of $S$ is coprime to $N$. We will also write $\mathcal E$ for the ${\rm PU}(N)$-bundle associated to $E$. Note that $M$ in this case is not an integral homology sphere. Floer's instanton homology can be extended to the case that $(M,\mathcal E)$ arises from an admissible pair \cite{fl3,KM:YAFT}. We will write $I^*(M,\mathcal E)$ for this version of instanton Floer homology. (See \cite[Section 3.1]{DX:suture} for a review of the general properties of $I^*(M,\mathcal E)$.) The proposal of the previous section to define symplectic instanton Floer homology can be also specialized to admissible pairs. We shall keep using the notation $I^*_{symp}(M,\mathcal E)$ to denote this conjectural invariant. There is yet another approach to define symplectic instanton Floer homology of $(M,\mathcal E)$ in this context, temporarily denoted by $\widetilde I^*_{symp}(M,\mathcal E)$, which avoids the technical difficulties of equivariant Floer homology in divisor complements. The current section concerns the relationship between the invariants $I^*_{symp}(M,\mathcal E)$ and $\widetilde I^*_{symp}(M,\mathcal E)$.
The definition of $\widetilde I^*_{symp}(M,\mathcal E)$ follows a similar route as $I^*_{symp}(M,\mathcal E)$. Suppose $F$ is a hermitian vector bundle of rank $N$ over an oriented Riemann surface $\Sigma$ such that the evaluation of $c_1(F)$ is coprime to $N$. Then the pair $(\Sigma,F)$ is called an admissible pair. Let $\mathcal F$ be the ${\rm PU}(2)$-bundle associated to $F$. Then the moduli space of flat connections $R(\Sigma,\mathcal F)$, defined in Section \ref{sec:extendedmoduli}, is a smooth K\"ahler manifold for this choice of $\mathcal F$. Let $M$ be a 3-manifold with boundary $\Sigma$ and $E$ be a hermitian vector bundle on $M$ extending $F$. Then we define ${\widetilde {\mathcal R}}(M,\mathcal E)$ to be the space of all elements of $R(\Sigma,\mathcal F)$ represented by flat connections on $\mathcal F$ that can be extended to $\mathcal E$. This space can be perturbed to an immersed Lagrangian submanifold of $R(\Sigma,\mathcal F)$ which we still denote by ${\widetilde {\mathcal R}}(M,\mathcal E)$ \cite{He}. The moduli space of solutions to the mixed equation can be also used to define a bounding cochain $b_{(M,\mathcal E)}$ for this Lagrangian \cite[Theorem 1.1 (1)]{fu7}. Therefore, $({\widetilde {\mathcal R}}(M,\mathcal E),b_{(M,\mathcal E)})$ defines an object of ${\mathfrak {Fuk}}(R(\Sigma,\mathcal F))$.
Next, let $(M,E)$ be an admissible pair. There is an embedded Riemann surface $\Sigma$ in $M$ such that removing $\Sigma$ from $M$ gives a disconnected manifold, and the pair given by $\Sigma$ and $F:=E|_{\Sigma}$ is admissible. Let $M_1$ and $M_2$ be the closure of the connected components of $M\backslash \Sigma$ and $E_i:=E|_{M_i}$. We can assume that the Lagrangians ${\widetilde {\mathcal R}}(M_1,\mathcal E_1)$ and ${\widetilde {\mathcal R}}(M_2,\mathcal E_2)$ have clean intersection by applying holonomy perturbations to one of them. Then $\widetilde I^*_{symp}(M,\mathcal E)$ is defined to be the Lagrangian Floer homology of the two elements $({\widetilde {\mathcal R}}(M_1,\mathcal E_1),b_{(M_1,\mathcal E_1)})$ and $({\widetilde {\mathcal R}}(M_2,\mathcal E_2),b_{(M_2,\mathcal E_2)})$. This Lagrangian Floer homology is independent of the choice of $\Sigma$.
\begin{conjecture} \label{admissible-AF-synp-quo}
For an admissible pair $(M,E)$, the vector spaces $I^*(M,E)$, $I_{symp}^*(M,E)$ and ${\widetilde I}_{symp}^*(M,E)$
are isomorphic to each other.
\end{conjecture}
The part of the above conjecture about the relationship between $I^*(M,E)$ and ${\widetilde I}_{symp}^*(M,E)$ is what we previously referred as the {\it ${\rm SO}(3)$-analogue of the Atiyah-Floer conjecture}. In the case that the Lagrangians involved in the definition of ${\widetilde I}_{symp}^*(M,E)$ are embedded, the ${\rm SO}(3)$-Atiyah-Floer conjecture is addressed in \cite{DFL}. The more general case will be treated in another forthcoming paper following the strategy proposed in \cite{fu7}. In the following, we discuss some general results in symplectic Floer homology which are related to the part of Conjecture \ref{admissible-AF-synp-quo} about the existence of isomorphism between $I_{symp}^*(M,E)$ and ${\widetilde I}_{symp}^*(M,E)$. Once the definition of the invariant $I_{symp}^*(M,E)$ is fully developed, we hope that these general results give a proof for this part of the above conjecture.
We firstly need an extension of the category $\frak {Fuk}(X,\omega,{\mathfrak L})$ for a clean collection of immersed Lagrangian
submanifolds ${\mathfrak L}$ in a symplectic manifold $(X,\omega)$. Suppose ${\mathfrak b}\in H^{{\rm even}}(X;\Lambda_0)$ with $\frak b \equiv 0 \mod T^{\epsilon}$.\footnote{The condition $\frak b \equiv 0 \mod T^{\epsilon}$ is not necessary. However, we need a slightly delicate argument to prove the convergence of operators. See, for example, \cite[Definition 17.8]{fooospectr}. For our application in this paper it suffices to consider the case when this extra condition is satisfied.} Then the $A_{\infty}$ operations ${\mathfrak m}_k$ associated to ${\mathfrak L}$ can be deformed by $\frak b$ to $\frak m^{\frak b}_k$ as in \cite[Definition 3.8.38]{fooobook}. Such deformations of the $A_\infty$ structure of $\frak {Fuk}(X,\omega,{\mathfrak L})$ are called {\it Lagrangian Floer theory with bulk deformation}. Roughly speaking, we deform $\frak m_k$ to $\frak m^{\frak b}_k$ using the holomorphic disks which hit a cycle that is Poincar\'e dual to $\frak b$. Bounding cochains of this deformed structure are also defined in the same way as in \eqref{formMC}. Consequently, there is a (filtered) $A_{\infty}$-category $\frak {Fuk}(X,\omega,{\mathfrak L},{\mathfrak b})$ whose objects are pairs of a Lagrangian $L \in \frak {\mathfrak L}$ and a bounding cochain $b$ with respect to the $\frak b$-deformed (filtered) $A_{\infty}$ structure \cite[Definition 3.8.38]{fooobook}.
\begin{shitu}\label{situ55}
Suppose a Hamiltonian action of a Lie group $G$ on a symplectic manifold $(X,\omega)$ is given.
Let $\mu : X \to \frak g^*$ denote the moment map of this action. Let the action of $G$ on $\mu^{-1}(0)$ is free. Then
the quotient $Y=\mu^{-1}(0)/G$ is a symplectic manifold with a symplectic form $\overline \omega$ \cite{MaW}.
Let $\frak L$ be a clean collection of $G$-equivariant immersed Lagrangian submanifolds.
For each $(\widetilde L,\iota_{\widetilde L})\in {\mathfrak L}$, we assume that the $G$ action on $\widetilde L$ is free and
$\iota_{\widetilde L}(\widetilde L) \subset \mu^{-1}(0)$.
Then $\overline L := (\widetilde L/G,[i_{\widetilde L}])$ is an immersed Lagrangian submanifold of $Y$.
The collection of all such immersed Lagrangians of $Y$ is denoted by $\overline{{\mathfrak L}}$.
Finally we assume that the following Lagrangian:
\begin{equation}\label{quotcorr}
\{(x,y) \in X \times Y \mid x \in \mu^{-1}(0), y = [x]\}
\end{equation}
is spin.
\end{shitu}
\begin{theorem}\label{quotthem}
There exists $\frak b\in H^{{\rm even}}(Y;\Lambda_0)$ such that the two filtered $A_{\infty}$-categories
$\frak{Fuk}^G(X,\omega,{\mathfrak L})$ and $\frak{Fuk}(Y,\overline \omega,\overline{{\mathfrak L}},\frak b)$ are homotopy
equivalent\footnote{See \cite[Definition 8.5]{fu6}.}. If \eqref{quotcorr} is a monotone Lagrangian
with minimal Maslov number $>2$, then ${\mathfrak b}$ can be chosen to be zero.
\end{theorem}
\begin{remark}
The element $\frak b$ in Theorem \ref{quotthem} is related to the {\it quantum Kirwan map}
introduced by Woodward in \cite{Wo10,Wo2,Wo3}. Theorem \ref{quotthem} is also closely related to the results
of Tian and Xu, written or announced in a series of papers \cite{TX}.
Both the works of Woodward and Tian-Xu (as well as various other related works such as \cite{GW})
are based on the study of gauged sigma model \cite{Mu, CGS}.
On the other hand, the second author's proof, which will appear in \cite{FuFu6},
uses equivariant Kuranishi structures
and relies on the idea of employing Lagrangian correspondence and cobordism arguments in a similar way
as in \cite{fu3,LL}. We were informed by Max Lipyanskiy that he had similar ideas
to use Lagrangian correspondences and cobordism arguments instead of gauged sigma model.
\end{remark}
\begin{conjecture}\label{cong107}
Suppose $(X,\omega)$ and ${\mathfrak L}$ are given as in Situation \ref{situ55}. Moreover, assume that
there exists a $G$-invariant normal crossing divisor $D \subset X \setminus \mu^{-1}(0)$
such that $X \setminus D$ is monotone. Let \eqref{quotcorr} be a monotone Lagrangian
submanifold of $(X \setminus D) \times Y$.
Then the category of filtered $A_{\infty}$ category $\frak{Fuk}^G(X \setminus D,\omega,{\mathfrak L})$
is homotopy equivalent to $\frak{Fuk}(Y,\overline \omega,\overline {\mathfrak L})$.
\end{conjecture}
\begin{remark}
The above generalization of Theorem \ref{quotthem} is related to Conjecture \ref{admissible-AF-synp-quo}.
By picking $X=\widehat{R}(\Sigma,\mathcal N,\mathcal F,p,{\epsilon})$,
this conjecture implies the predicted relationship between
$I_{symp}^*(M,E)$ and ${\widetilde I}_{symp}^*(M,E)$ in Conjecture \ref{admissible-AF-synp-quo}.
The main difficulty with this
conjecture is to define $\frak{Fuk}^G(X \setminus D,\omega,{\mathfrak L})$ for the case that $D$ is a normal crossing divisor.
Existence of this $A_\infty$-category in the special case that $D$ is a
smooth divisor is the content of Theorem \ref{cat-div-comp}.
Combination of the techniques used in verifying Theorems \ref{cat-div-comp} and \ref{quotthem}
prove the above conjecture in the special case that $D$ is a smooth divisor.
\end{remark}
\section{The Atiyah-Floer Conjecture}
\label{Sec:AFori}
Floer's original instanton Floer homology is an invariant of 3-manifolds which have the same integral homology as the 3-dimensional sphere \cite{Fl:I}. Given an integral homology sphere $M$, he introduced a chain complex $C_*(M)$ with the differential $\partial$, whose homology is defined to be instanton Floer homology and is denoted by $I^*(M)$. The chain complex $C_*(M)$ is a free group generated by non-trivial flat ${\rm SU}(2)$-connections. Since any ${\rm SU}(2)$-bundle over a 3-manifold can be trivialized, these connections all have the same topological type.
The differential $\partial$ is defined by considering the moduli space of instantons on the trivial ${\rm SU}(2)$-bundle $P$ over $\bR \times M$. To be a bit more detailed, fix a product metric on $\bR \times M$ corresponding to a fixed metric on $M$. The Hodge $*$-operator decomposes the space of 2-forms to {\it anti-self-dual} and {\it self-dual} forms. Then an instanton on $\bR \times M$ is a connection $A$ on $P$ such that:
\begin{equation} \label{instanton-eq}
F^+(A)=0 \hspace{2cm} |\!|F(A)|\!|_2<\infty
\end{equation}
where $F^+(A)$ and $|\!|F(A)|\!|_2$ are respectively the self-dual part and the $L^2$ norm of the curvature of $A$. For any instanton $A$, there are flat connections $a_+$ and $a_-$ such that:
\begin{equation} \label{asymp}
\lim_{t \to \pm \infty} A|_{\{t\}\times M}=a_{\pm}
\end{equation}
Translation in the $\bR$-direction and ${\rm SU}(2)$-bundle automorphisms act on the space of instantons. The quotient space of instantons satisfying \eqref{asymp} with respect to these two actions is denoted by $\mathcal M(a_-,a_+;M)$. Moreover, if we require $|\!|F(A)|\!|_2^2$ to be equal to a fixed real number $E$, then the resulting space is denoted by $\mathcal M(a_-,a_+;E;M)$. The differential $\partial(a)$ for a non-trivial flat connection $a$ is defined as:
\begin{equation*}
\partial(a)=\sum \#\mathcal M(a,b; E;M)\cdot b
\end{equation*}
where the sum is over all $E$ and $b$ that $\mathcal M(a,b; E;M)$ is 0-dimensional. Here $\#\mathcal M(a,b; M)$ denotes the signed count of the points in the 0-dimensional space $\mathcal M(a,b;E; M)$. In general, we might need to perturb the equation in \eqref{instanton-eq} as the space of flat connections on $M$ and the space of instantons on $\bR \times M$ might not be cut down by transversal equations.
There are several other versions of instanton Floer homology in the literature. The trivial connection on an integral homology sphere $M$ does not play any role in the definition of $I^*(M)$. An alternative version of this invariant, constructed in \cite{Don:YM-Floer}, uses the moduli spaces $\mathcal M(a,b,E;M)$ where $a$ or $b$ could be the trivial connection. We will write $\overline I^*(M)$ for this invariant, which is an $H_{{\rm PU}(2)}^*$-module\footnote{The original notation for this invariant in \cite{Don:YM-Floer} is $\overline{\overline{HF}}(M)$.}.
Next, we aim to construct a version of symplectic instanton Floer homology which is conjecturally isomorphic to $I^*(M)$ for an integral homology sphere $M$. We shall apply a combination of the constructions of the previous two sections to the following pair of a symplectic manifold and a smooth divisor, introduced in \eqref{ex} and \eqref{decom-ex}:
\begin{equation} \label{symp-div}
(X,D):=({\widehat R}(\Sigma,p,\theta), \overline{\mu}^{-1}(\theta)/S^1)
\end{equation}
with $0<\theta < \frac{1}{2}$. This version of symplectic instanton Floer homology can be regarded as an equivariant version of a variation of the construction in \cite{MW}. (See Remark \ref{MW-AF}.) Fix a Heegaard splitting of the 3-manifold $M$:
\begin{equation}\label{form73}
M = H^0_g \cup_{\Sigma_g} H^1_g.
\end{equation}
According to Proposition \ref{handle-body-lag}, we can form the Lagrangian submanifolds $\tilde R(H^i_g,p)$ of $X$ associated to this Heegaard splitting. The following Lemma about the intersection of these Lagrangians can be proved using {\it holonomy perturbations} \cite{Cliff:Cas-invt,Fl:I,Don:YM-Floer,He}. We omit the details here:
\begin{lemma}\label{lem7474}
There are Hamiltonian isotopies of the Lagrangians $\tilde R(H^i_g,p)$ in $X\backslash D$ to submanifolds with clean intersection.
Moreover, we can assume that each connected component of the intersection of the perturbed Lagrangians is
either a point which consists of the trivial connection or a single ${\rm PU}(2)$-orbit.
\end{lemma}
Suppose $L_i$ denotes the perturbation of the Lagrangian $\tilde R(H^i_g,p)$ provided by Lemma \ref{lem7474}. Since $\tilde R(H^i_g,p)$ is monotone in $X\backslash D$ \cite{MW}, the Lagrangian $L_i$ is also monotone in $X\backslash D$. The manifold $L_i$ is diffeomorphic to the Cartesian product of $g$ copies of ${\rm SU}(2)$ \cite{MW}. In particular, it can be equipped with a spin structure. The intersection of $L_0$ and $L_1$ can be decomposed as:
\begin{equation}\label{decompLL}
L_0 \cap L_1=\{\theta\} \cup \bigcup_{a \in \mathcal A} R_{a}
\end{equation}
where $R_a \cong {\rm PU}(2)$. Here $\theta$ denotes the trivial connection. Let $\mathcal A_+ = \mathcal A \cup \{\theta\}$ and $R_\theta= \{\theta\}$.
For $a,b \in \mathcal A$, define $\overset{\circ}{\mathcal M}(a,b;\beta;L_0,L_1)$ to be the moduli space of maps $u : \bR \times [0,1] \to X \setminus D$ which satisfy the analogues of \eqref{CR}, \eqref{form38} and represent the homology class $\beta \in H(a,b)$. As before, we also identify two maps $u$ and $u'$ if $u'(\tau,t) = u(\tau+\tau_0,t)$ for some $\tau_0 \in \bR$. There is an obvious ${\rm PU}(2)$ action on this moduli space. We can also form the restriction maps:
\begin{equation}\label{form7474}
{\rm ev}_{-\infty} : \overset{\circ}{\mathcal M}(a,b;\beta;L_0,L_1) \to R_a,\quad
{\rm ev}_{+\infty} : \overset{\circ}{\mathcal M}(a,b;\beta;L_0,L_1)\to R_b.
\end{equation}
A combination of the proofs of Theorems \ref{existequiv} and \ref{mainthm} can be used to prove the following Proposition:
\begin{prop}\label{prop7575}
There exists a compactification of $\overset{\circ}{\mathcal M}(a,b;\beta;L_0,L_1)$, denoted by ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$,
which satisfies the following properties:
\vspace{-5pt}
\begin{enumerate}
\item[(i)]
This space has a Kuranishi structure with corners. The ${\rm PU}(2)$ action of $\overset{\circ}{\mathcal M}(a,b;\beta;L_0,L_1)$
extends to ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ and the Kuranishi structure is ${\rm PU}(2)$-equivariant.
The evaluation maps in \eqref{form7474} also extend to ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$
and are underlying maps of
${\rm PU}(2)$-equivariant weakly submersive maps.\footnote{See \cite[Definition 32.1]{fooospectr} for its definition.}
\item[(ii)]
Let $d(\beta)$ be the virtual dimension of ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$.
For any $d$, there are only finitely many choices of $\beta$ such that
${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ is nonempty and $d(\beta) = d$.
There also exists ${\rm deg}:\mathcal A \to \bZ/8\bZ$ such that $\deg_+(\theta)=0$ and for any $a \in \mathcal A$,
$b \in \mathcal A_+$, $\beta \in H(a,b)$,
we have:
\begin{equation} \label{mod-8-1}
d(\beta) \equiv {\rm deg}(b) - {\rm deg}(a) + 2 \mod 8.
\end{equation}
Moreover, if $b \in \mathcal A_+$ and $\beta \in H(\theta,b)$ then:
\begin{equation}\label{mod-8-2}
d(\beta) \equiv{\rm deg}(b) - 1 \mod 8.
\end{equation}
\item [(iii)]
The codimension one boundary components of ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ consists of fiber products
\begin{equation}\label{form7575}
{\mathcal M}^{\rm RGW}(a,c;\beta_1;L_0,L_1) \times_{R_c} {\mathcal M}^{\rm RGW}(c,b;\beta_2;L_0,L_1),
\end{equation}
where the union is taken over $c \in \mathcal A_+$ and $\beta_1 \in H(a,c),\beta_2 \in H(c,b)$ with $\beta_1 \#\beta_2 = \beta$.
\end{enumerate}
\end{prop}
\begin{remark}
Characterization of co-dimension one boundary components in \eqref{form7575} implies that if $c \ne \theta$, then:
\[
d(\beta_1 \#\beta_2) = d(\beta_1) + d(\beta_2) - 3
\]
and if $c = \theta$, then:
\[
d(\beta_1 \#\beta_2) = d(\beta_1) + d(\beta_2).
\]
This is consistent with the identities in \eqref{mod-8-1} and \eqref{mod-8-2}.
\end{remark}
Analogous to the construction of Section \ref{sec:equivariant}, we can use the compactification provided by this proposition to define a Lagrangian Floer homology group :
\begin{equation} \label{symp-equiv}
HF_{{\rm PU}(2)}(L_0,L_1;X \setminus D)
\end{equation}
for an integral homology sphere. This Lagrangian Floer homology group is a module over $H^*_{{\rm PU}(2)}= H^*_{{{\rm SU}}(2)}$. The following conjecuture states that this module is a 3-manifold invariant. This invariant can be regarded as a version of symplectic instanton Floer homology and is denoted by $\overline I^*_{symp}(M)$.
\begin{conjecture} \label{symp-equiv}
The $H^*_{{\rm PU}(2)}$-modules in \eqref{symp-equiv} for different choices of Heegaard splitting
are isomorphic to each other.
\end{conjecture}
\begin{remark} \label{inv}
We hope to address Conjecture \ref{symp-equiv} in the same way as in the proof of the corresponding result in \cite{MW}.
(The result of \cite{MW} can be regarded as a non-equivariant version of Conjecture \ref{symp-equiv}.)
Following the arguments in \cite{MW} requires us to consider quilted Floer homology of Lagrangian correspondences such that each
Lagrangian correspondence is from a pair $(X_1,D_1)$ of a symplectic manifold and a smooth divisor to another such pair $(X_2,D_2)$.
Consequently, we need to study the moduli space of holomorphic curves for the pair $(X_1\times X_2, (D_1 \times X_2) \cup (X_1 \times D_2))$.
The space $(D_1 \times X_2) \cup (X_1 \times D_2)$ is a normal crossing divisor in $X_1\times X_2$.
The extension of the theory of Section \ref{sec:divisor complement}
to normal crossing divisors is the content of a work in progress and its details have not been completely worked out yet.
(See Conjecture \ref{cong59} and Remark \ref{normal-crossing}.)
However, as it is explained in \cite[Section 12]{FuFu50},
we can use a different compactification of holomorphic discs whose target is the product $X_1\times X_2$.
This compactification is denoted by $\mathcal M^{\prime}$ and is discussed in \cite[Section 12]{FuFu50}.
In this compactification, the sphere bubbles on two factors are studied separately.
It is plausible that adapting this construction to our set up allows us to avoid the case of holomorphic discs
in the complement of normal crossing divisors and to work only with smooth divisors.
\end{remark}
There is an alternative version of symplectic instanton Floer homology constructed by the moduli spaces ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$. The ${\rm PU}(2)$ action on the space ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$
is free unless $a=b=\theta$. If $a=b=\theta$, then the action is still free unless $\beta=0$, which is the homology class of the constant map. The moduli space
${\mathcal M}^{\rm RGW}(\theta,\theta;0;L_0,L_1)$ consists of a single element. Therefore, the quotient space
\[
\overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1):={\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)/{\rm PU}(2)
\]
has an induced Kuranishi structure. Proposition \ref{prop7575} can be used to verify the following lemma:
\begin{lemma}\label{lem6767}
For $a,b \in \mathcal A$, the boundary of $\overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$
is the union of two types of spaces:
\begin{enumerate}
\item The direct product:
\[
\overline{\mathcal M}^{\rm RGW}(a,c;\beta_1;L_0,L_1) \times
\overline{\mathcal M}^{\rm RGW}(c,b;\beta_2;L_0,L_1)
\]
for $c \in \mathcal A$, $\beta_1 \in H(a,c)$ and $\beta_2 \in H(c,b)$ such that
$\beta_1 \#\beta_2 = \beta$.
\item The quotient of the union of direct products
\[
{\mathcal M}^{\rm RGW}(a,\theta;\beta_1;L_0,L_1) \times
{\mathcal M}^{\rm RGW}(\theta,b;\beta_2;L_0,L_1)
\]
by the diagonal ${\rm PU}(2)$ action.
Here the union is taken over $\beta_1 \in H(a,\theta)$ and $\beta_2 \in H(\theta,b)$
with $\beta_1 \#\beta_2 = \beta$.
\end{enumerate}
\end{lemma}
We pick a system of ${\rm PU}(2)$ invariant multi-sections over each moduli space ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ that is compatible with the description of the boundaries in \eqref{form7575}.
This is equivalent to choosing a system of
multi-sections over various
$\overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$
that is compatible with the
description of the boundaries in Lemma \ref{lem6767}. In the case that $d(\beta)=0$ and $a,b \ne \theta$, Lemma \ref{lem6767} and the compatibility of the multi-sections show that the zero set of the multisection
in the moduli space $\overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ is a compact 0-dimensional space. Therefore, we can count the number of points in this space (with signs) to define:
\begin{equation}\label{numberRGWkado}
\# \overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1).
\end{equation}
Now we are ready to define another version of symplectic instanton Floer homology for integral homology spheres. Define:
\begin{equation} \label{symp-chain-cx-1}
C^*_{symp}(M):= \bigoplus_{a \in \mathcal A} \bQ [a]
\end{equation}
and
\begin{equation}\label{symp-chain-cx-2}
\partial [a]:=\sum_{b \in \mathcal A, \beta\in H(a,b)}\left( \# \overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)\right) [b].
\end{equation}
where the sum is over all $b \in \mathcal A$ and $\beta\in H(a,b)$ that $d(\beta)=0$. Another application of Lemma \ref{lem6767} and the compatibility of the multi-sections show that $\partial^2=0$. To be a bit more detailed, the terms in $\partial^2(a)$, for a non-trivial flat connection $a$, are in correspondence with the boundary points of the 1-dimensional moduli spaces $\overline{\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ which are of type (1) in Lemma \ref{numberRGWkado} For a 1-dimensional moduli space, the space of boundary points of type (2) is empty, because each component of the space of type (2) boundary points has dimension at least 3.
The homology of the chain complex in \eqref{symp-chain-cx-1} and \eqref{symp-chain-cx-2} is denoted by $I^*_{symp}(M)$. Note that our definition of $I^*_{symp}(M)$ resembles Floer's instanton homology $I^*(M)$ in the sense that the trivial connection $\theta$ doss not enter into the definition of the corresponding chain complex. The following is the analogue of Conjecture \ref{symp-equiv}. The same comment as in Remark \ref{inv} applies to this conjecture.
\begin{conjecture} \label{symp-equiv-2}
The group $I^*_{symp}(M)$ is an invariant of the integral homology sphere $M$.
That is to say, the homology of the chain complex $(C^*_{symp}(M),\partial)$
is independent of the choices of the Heegaard splitting.
\end{conjecture}
\begin{remark}
In the course of defining $I^*_{symp}(M)$, we only need the moduli spaces of virtual dimension $1$ or $0$.
Therefore, we do not need to prove the smoothness of the coordinate change maps
of our Kuranishi structure.
We also do not need to study triangulation of the zero set
of our multi-sections. For example, we can discuss in the same way as in
\cite[Section 14]{foootech2}.
\end{remark}
The first part of the following conjecture can be regarded as a rigorous formulation of the original version of the Atiyah-Floer conjecture for integral homology spheres. In Section \ref{modulispace}, we sketch a plan for the proof of the first part of Conjecture \ref{conj7878}.
\begin{conjecture}\label{conj7878}
For any integral homology sphere $M$, the vector spaces
$I^*(M)$ and $I^*_{symp}(M)$ are isomorphic to each other.
The $H^*_{{\rm PU}(2)}$-modules $\overline I^*(M)$ and $\overline I_{symp}^*(M)$
are also isomorphic to each other.
\end{conjecture}
\begin{remark} \label{MW-AF}
One can forget the ${\rm PU}(2)$-action on ${\mathcal M}^{\rm RGW}(a,b;\beta;L_0,L_1)$ and
apply the construction of the previous section to define (non-equivariant) Lagrangian Floer homology
for the Lagrangians $L_0$ and $L_1$ in the complement of $D$. The resulting Floer homology is essentially
the same 3-manifold invariant as the version of symplectic instanton Floer homology that is constructed in \cite{MW}.
There is also an analogue of the Atiyah-Floer conjecture for this invariant.
It is conjectured in \cite{MW} that this invariant is isomorphic to an alternative version of instanton Floer homology,
defined in \cite{Don:YM-Floer} and denoted by $\widetilde {HF}(M)$.
\end{remark}
\section{Lagrangians and $A_\infty$-categories}
\label{Sec:ainfinity}
For a given symplectic manifold $(X,\omega)$, we can define an $A_{\infty}$-category over the universal Novikov ring $\Lambda_0^\bR$, which is usually denoted by $\frak{Fuk}(X,\omega,{\mathfrak L})$. The objects of this category $\frak{Fuk}(X,\omega,{\mathfrak L})$ are defined using immersed Lagrangian submanifolds, and ${\mathfrak L}$ denotes a collection of such Lagrangians.
Suppose $\widetilde L$ is an immersed Lagrangian submanifold of $X$ given by $\iota:\widetilde L \to X$ where the self-intersections of $X$ are transversal. Define $CF(\widetilde L,\widetilde L)$ to be the cohomology group $H^*(\widetilde L \times_X \widetilde L,\Lambda^\bR)$ where $\widetilde L \times_X \widetilde L$ is the fiber product of the map $\iota$ with itself\footnote{To be more precise, one needs to start with a chain model for this cohomology group. As it is shown in \cite{fooobook}, this chain model can be replaced with the cohomology groups by an algebraic argument.}. Therefore, $CF(\widetilde L,\widetilde L)$ is the direct sum of $H^*(\widetilde L)$ and a free abelian group generated by the self-intersection points of $\widetilde L$. For $\beta \in H_2(X,\iota(\widetilde L);\bZ)$, let $\mathcal M_{k+1}(\widetilde L;\beta)$ be the compactified moduli space of pseudo-holomorphic disks with $k+1$ boundary marked points. The elements of $\mathcal M_{k+1}(\widetilde L;\beta)$ are required to represent the homology class $\beta$ and need to satisfy the Lagrangian boundary condition. The boundary of an element of $\mathcal M_{k+1}(\widetilde L;\beta)$ has to be mapped to $\iota (\widetilde L)$, and away from the marked points, it can be lifted to $\widetilde L$. (See Figure \ref{zu2} for a schematic picture and \cite[Definition 2.1.27]{fooobook} and \cite[Section 4]{AJ} for the precise definitions of these moduli spaces.) The moduli space $\mathcal M_{k+1}(\widetilde L;\beta)$ can be used to form the following diagram:
\begin{equation}
\xymatrix{
(\widetilde L\times_X\widetilde L)^k&&&\mathcal M_{k+1}(\widetilde L;\beta) \ar[lll]_{({\rm ev}_1,\dots,
{\rm ev}_k)\hspace{.4cm}}\ar[rrr]^{\hspace{.8cm}{\rm ev}_0} &&&(\widetilde L\times_X\widetilde L)
}
\nonumber
\end{equation}
where ${\rm ev}_i$ for $0\leq i\leq k$, is the evaluation map at the $i^{\rm th}$ marked point. A standard `pull-up-push-down construction' applied to these diagrams for various choices of $\beta$ determines a map ${\mathfrak m}_k:CF(\widetilde L,\widetilde L)^{\otimes k} \to CF(\widetilde L,\widetilde L)$ for any $k\geq 0$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{FigureAinfty.eps}
\caption{The operation $\frak m_k$}
\label{zu2}
\end{figure}
Next, let ${\mathfrak L}$ be a finite family of immersed spin Lagrangian submanifolds of $X$. We say that this family is clean if for any two elements $(\widetilde L_1, \iota_1)$ and $(\widetilde L_2, \iota_2)$ of this family the fiber product $\widetilde L_1\times_X \widetilde L_2$ is a smooth manifold and the tangent space at each point of $\widetilde L_1\times_X \widetilde L_2$ is given by the fiber product of the tangent spaces of $\widetilde L_1$ and $\widetilde L_1$.
Here we include the case $\tilde L_1 =\tilde L_2$.
For any two such elements of ${\mathfrak L}$, we define $CF(\widetilde L_1,\widetilde L_2)$ to be the cohomology group $H^*(\widetilde L_1 \times_X \widetilde L_2)$. Then the construction of the previous paragraph can be modified to define the following maps for any $k \geq 0$ and for any sequence $\{(\widetilde L_i, \iota_i)\}_{0\leq i \leq k}$ of the elements of ${\mathfrak L}$ \cite{fooobook,AJ}:
\[
{\mathfrak m}_k : \bigotimes_{i=1}^k CF(\widetilde L_{i-1},\widetilde L_i) \to CF(\widetilde L_{0},\widetilde L_k)
\]
The maps ${\mathfrak m}_k$ satisfy the analogues of the $A_\infty$-relations in \eqref{formula25}. However, the map ${\mathfrak m}_0$ does not have to vanish in general\footnote{An $A_\infty$-category with a non-vanishing ${\mathfrak m}_0$ is called a {\it curved $A_\infty$-category.}}. Therefore, these maps cannot immediately be used to define an $A_\infty$-category. This issue can be fixed with the aid of {\it bounding cochains}. For an immersed Lagrangian $(\widetilde L,\iota)$, an element $b\in H^{odd}(\widetilde L,\Lambda_0)$ is called a bounding cochain if it is divisible by $T^\epsilon$ for a positive $\epsilon$, and it satisfies the following {\it Maurer-Cartan} equation:
\begin{equation}\label{formMC}
\sum_{k=0}^{\infty} {\mathfrak m}_{k}(b,\dots,b) = 0,
\end{equation}
An object of the category $\frak{Fuk}(X,\omega,{\mathfrak L})$ is a pair $(\widetilde L,b)$ where $\widetilde L$ is an element of ${\mathfrak L}$ and $b$ is a bounding cochain. The module of morphisms for two objects $(\widetilde L_0,b_0)$ and $(\widetilde L_1,b_1)$ is defined to be $CF(\widetilde L_0, \widetilde L_1)$. The structural map ${\mathfrak m}_k^{\vec b} : \bigotimes_{i=1}^k CF(\widetilde L_{i-1},\widetilde L_i) \to CF(\widetilde L_{0},\widetilde L_k)$ for a sequence of objects $\{(\widetilde L_i, \iota_i,b_i)\}_{0\leq i \leq k}$ is also defined as follows:
\begin{equation}\label{modified-m}
{\mathfrak m}_k^{\vec b}(p_1,\dots,p_k):=\sum_{l_0\geq, \dots, l_k\geq 0}{\mathfrak m}_{k+l_0+\dots+l_k}(b_0^{\otimes l_0},p_1,b_1^{\otimes l_1},\dots,
b_{k-1}^{\otimes l_{k-1}},p_k,b_{k}^{\otimes l_{k}})
\end{equation}
Using the results of \cite{fooobook,fooobook2}, it is shown in \cite{anchor,AFOOO,FuFu50} that $\frak{Fuk}(X,\omega,{\mathfrak L})$ is an $A_\infty$-category in the case that ${\mathfrak L}$ consists of only embedded Lagrangians. The more general case of immersed Lagrangians is treated in \cite{AJ}.
Suppose $L_0$ and $L_1$ are two monotone and embedded Lagrangians in $X$ that satisfy the condition (m.a) of Section \ref{sec:equivariant}. Then the map ${\mathfrak m}_0:\Lambda_0 \to CF(L_i,L_i)$ vanishes and we can associate the trivial bounding cochain to each of these Lagrangians. The map ${\mathfrak m}_1:CF(L_0,L_1) \to CF(L_0,L_1)$ defines a differential. The homology of this chain complex is the same as Oh's Lagrangian Floer homology for monotone Lagrangians \cite{Oh}.
We can also consider equivariant version of the category $\frak{Fuk}(X,\omega,{\mathfrak L})$. The following theorem provides the main ingredient for the equivariant construction:
\begin{theorem}\label{thm5151}
Let $G$ be a Lie group acting on $(X,\omega)$. Let ${\mathfrak L}$ be a clean collection of immersed Lagrangians which are equivariant with respect to the action of $G$.
For any sequence $\{(\widetilde L_i, \iota_i)\}_{0\leq i \leq k}$ of the elements of ${\mathfrak L}$,
there exists a $H_G^*$-linear homomorphism:
\[
{\mathfrak m}_k^G: \bigotimes_{i=1}^k H^*_G(\widetilde L_{i-1}\times_X \widetilde L_i,\Lambda_0) \to
H^*_G(\widetilde L_{0}\times_X \widetilde L_k,\Lambda_0)
\]
which satisfies the $A_{\infty}$-relations in \eqref{formula25}. Moreover, the following diagram commutes:
\begin{equation}
\begin{CD}
\bigotimes_{i=1}^k H^*_G(\widetilde L_{i-1}\times_X \widetilde L_i,\Lambda_0) @ >\frak m_k^G>>
H^*_G(\widetilde L_{0}\times_X \widetilde L_k,\Lambda_0)\\
@ V{}VV @ VV{}V\\
\bigotimes_{i=1}^k H^*(\widetilde L_{i-1}\times_X \widetilde L_i,\Lambda_0) @ >\frak m_k>>
H^*(\widetilde L_{0}\times_X \widetilde L_k,\Lambda_0).
\end{CD}
\end{equation}
Here the vertical arrows are canonical maps from equivariant cohomology to de-Rham cohomology.
\end{theorem}
\begin{proof}[Sketch of the proof]
For the simplicity of exposition, assume that the immersed Lagrangians are the same as each other.
We use the equivariant Kuranishi structure on the space
$\mathcal M_{k+1}(L;\beta)$ \cite{FuFu5} and an approximation of the universal principal $G$-bundle $EG \to BG$ to obtain:
\begin{equation}
\xymatrix{
&&\mathcal M_{k+1}(L;\beta) \times_G EG(N) \ar[dl]^{({\rm ev}_1,\dots,
{\rm ev}_k)}\ar[dr]_{{\rm ev}_0} \\
&(\widetilde L\times_X\widetilde L)^k \times_G EG(N) \ar[d]
&&
L \times_G EG(N) \\
&((\widetilde L\times_X\widetilde L) \times_G EG(N))^k
}
\nonumber
\end{equation}
where ${\rm ev}_i$ for $0\leq i\leq k$, is the evaluation map at the $i$-th marked point.
By a formula similar to (\ref{form31100}), we can define operations:
\[
\frak m^{G,N}_{k,\beta} : H(L \times_G EG(N))^{\otimes k} \to H(L \times_G EG(N)).
\]
Taking the limit $N\to \infty$, we obtain the operation $\frak m^{G}_{k,\beta}$ between the equivariant cohomology groups.
Then
$\frak m^G_k = \sum_{\beta} T^{\omega(\beta)}\frak m^{G}_{k,\beta}$ is the required $A_{\infty}$ operation.
\end{proof}
\begin{definition}
An element $b \in H^{{\rm odd}}_G(L;\Lambda_0)$ is a {\it $G$-equivariant bounding cochain},
if $b$ is divisible by $T^{\epsilon}$ for a positive $\epsilon$ and $b$ satisfies (\ref{formMC}), where
$\frak m_k$ is replaced with $\frak m_k^G$.
\end{definition}
The following theorem claims the existence of the $G$-equivariant analogue of $\frak{Fuk}(X,\omega,{\mathfrak L})$. The geometric content of this theorem is given in Theorem \ref{thm5151}:
\begin{theorem}\label{thm54}
Let ${\mathfrak L}$ be a clean collection of $G$-equivariant immersed Lagrangian submanifolds of $(X,\omega)$.
There exists a (filtered) $A_{\infty}$-category $\frak{Fuk}^G(X,\omega,{\mathfrak L})$ whose objects are pairs of the form
$(L,b)$ where $L \in {\mathfrak L}$ and $b$ is a $G$-equivariant bounding cochain. The structural maps ${\mathfrak m}_k^{\vec b,G}$
are also defined by applying the analogue of the formula of \eqref{modified-m} to the maps ${\mathfrak m}_k^G$.
\end{theorem}
The above theorem can be regarded as the generalization of the results of Section \ref{sec:equivariant} on $G$-equivariant Lagrangian Floer homology. Similarly, the techniques of \cite{DF} can be used to extend the results of Section \ref{sec:divisor complement}. More generally, we
can also consider $A_\infty$-category associated to $G$-invariant Lagrangians in the complement of a smooth divisor:
\begin{theorem}\label{cat-div-comp}
Let $(X,\omega)$ be a symplectic manifold with a Lie group $G$ acting on $X$ by symplectomorphisms.
Let $D$ be a $G$-invariant smooth divisor in $X$ such that $X$ admits a K\"ahler
structure in a neighborhood of $D$ compatible with the symplectic form
$\omega$. Let ${\mathfrak L}$ be a clean collection of $G$-equivariant immersed Lagrangian submanifolds of $(X,\omega)$.
\begin{enumerate}
\item[(i)] There are operations:
\begin{equation}\label{div-equiv-Ainfty}
{\mathfrak m}_k^G : \bigotimes_{i=1}^k CF(\widetilde L_{i-1},\widetilde L_i) \to CF(\widetilde L_{0},\widetilde L_k)
\end{equation}
for any sequence $\{(\widetilde L_{i},\iota_i)\}_{0\leq i \leq k}$ of the elements of ${\mathfrak L}$.
These operations satisfy $A_\infty$-relations.
\item[(ii)] There exists a (filtered) $A_{\infty}$-category $\frak{Fuk}^G(X\backslash D,\omega,{\mathfrak L})$ whose objects are pairs of the form
$(L,b)$ where $L \in {\mathfrak L}$ and $b$ is a bounding cochain in $CF(\widetilde L,\widetilde L)$ with respect to the operators defined in
Item {\rm (i)}. The structural operations of $\frak{Fuk}^G(X\backslash D,\omega,{\mathfrak L})$ are given by modifications of
the operators in \eqref{div-equiv-Ainfty} as in \eqref{modified-m}.
\end{enumerate}
\end{theorem}
\begin{conjecture}\label{cong59}
Theorems \ref{mainthm} and \ref{cat-div-comp} still hold in the case that $D$ is a normal crossing divisor with respect to a K\"ahler structure in
a neighborhood of $D$ which is compatible with $\omega$.
\end{conjecture}
\begin{remark} \label{normal-crossing}
As in Section \ref{sec:divisor complement}, we need to use a non-standard compactification of
the moduli spaces of pseudo-holomorphic disks in $X \setminus D$ to prove Theorem \ref{cat-div-comp}.
It is plausible that the compactification appearing in relative Gromov-Witten theory for the complements of normal crossing divisors \cite{groSie}
can be employed to prove Conjecture \ref{cong59}. The analysis of Gromov-Witten theory for the complement of normal crossing
divisors is much more subtle than the case of the complements of smooth divisors, and we would expect that a similar phenomena happens
in the construction of Lagrangian Floer homology and the category $\frak {Fuk}(X\backslash D,\omega)$.
\end{remark}
\section{Cut-down Extended Moduli Spaces for Other Lie Groups}
\label{sec:generalgroup}
In this section, we discuss a plan to address the conjectures in Section \ref{Sec:conjectures}. Fix a Lie group $G$ as in Section \ref{sec:extendedmoduli} and let $T$ be a maximal torus of $G$ whose Lie algebra is denoted by ${\mathfrak t}$. The Lie group $G$ acts on $\frak g$ and the quotient space can be identified with the quotient ${\mathfrak t}/W$ of $\frak t$ by the Weyl group $W$. Let ${\mathfrak t}^+\subset {\mathfrak t}$ be a Weyl chamber of $G$. Then ${\mathfrak t}^+$ is a fundamental domain for the action of the Weyl group on $\frak t$, i.e., we can identify ${\mathfrak t}/W$ with ${\mathfrak t}^+$. The quotient map from ${\mathfrak g}$ to ${\mathfrak t}^+$ is denoted by $Q$. We will also write ${\mathfrak t}_\bZ$ for the integer lattice in ${\mathfrak t}$. Thus $T$ is equal to the quotient ${\mathfrak t}/{\mathfrak t}_\bZ$. The dual lattice of ${\mathfrak t}_\bZ$ is denoted by ${\mathfrak t}_{\bZ}^*$. The action of the Weyl group on ${\mathfrak t}$ induces actions of this group on the lattices ${\mathfrak t}_\bZ$ and ${\mathfrak t}_{\bZ}^*$. For a finite subset $\mathcal N=\{\alpha_1,\dots,\alpha_n\}$ of $\frak t_{\bZ}^*$, we define:
\begin{equation} \label{Delta}
\overset{\circ}\Delta_{\mathcal N}(\epsilon) = \{\xi \in \frak t \mid \forall \alpha \in \mathcal N, \alpha(\xi) < \epsilon \}
\end{equation}
Let $\Delta_{\mathcal N}(\epsilon)$ be the closure of $\overset{\circ}\Delta_{\mathcal N}(\epsilon)$. The intersection of these open and closed polytopes with the Weyl chamber ${\mathfrak t}^+$ is denoted by $\Delta_{\mathcal N}^+(\epsilon)$ and $\overset{\circ\hspace{3mm}}{\Delta_{\mathcal N}^+}(\epsilon)$.
\begin{conds}\label{cond92}
The set $\mathcal N$ is required to satisfy the following conditions:
\begin{enumerate}
\item $\mathcal N$ is invariant with respect to the action of the Weyl group.
\item $\Delta_{\mathcal N}(\epsilon)$ is compact.
\item If $\alpha \in \mathcal N$, $\xi \in \Delta_{\mathcal N}(\epsilon)$, $\sigma \in W$
satisfy $\alpha(\xi) = \epsilon$ and $\sigma(\xi) = \xi$, then $\sigma(\alpha) = \alpha$.
\item For any vertex $v$ of $\Delta_{\mathcal N}(\epsilon)$ the set
$\{\alpha \in \mathcal N \mid \alpha(v) = \epsilon\}$ is a $\bZ$ basis of $\frak t_{\bZ}^*$.
\end{enumerate}
\end{conds}
\begin{example}
For $G = {\rm PU}(3)$, we can assume that ${\mathfrak t}$ is the set of diagonal matrices with diagonal entries
$(2\pi{\bf i}\theta_0,2\pi{\bf i}\theta_1,2\pi{\bf i}\theta_2)$ with $\sum \theta_i = 0$.
Suppose $\alpha_i\in {\mathfrak t}_\bZ$ is the map that assigns $\theta_i$ to a diagonal matrix of this form.
We may take $\mathcal N$ to be the set that consists of $\theta_i$ and $-\theta_i$ for $i=1,2,3$.
The Wely chamber and the set $\Delta_{\mathcal N}(\epsilon)$ is demonstrated in Figure \ref{zu3}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{FigureSec8.eps}
\caption{Wely chamber and symplectic cut}
\label{zu3}
\end{figure}
\end{example}
\begin{remark}
A set of similar conditions for polytopes in ${\mathfrak t}^+$ are introduced by Woodward in \cite{Wo20}.
For example, Condition \ref{cond92} (2) implies the requirements of \cite[Definition 1.1]{Wo20}.
Condition \ref{cond92} (4) also asserts that $\Delta_{\mathcal N}^+(\epsilon)$ is Delzant. (See \cite[page 5]{Wo20}.)
However, our requirements are more restrictive. For example, we only consider polytopes in ${\mathfrak t}^+$
which contain the origin. In fact, our definition is essentially the same as outward-positive condition in \cite{MT}.
\end{remark}
For the following proposition, let $\widehat{R}(\Sigma,\mathcal F,p)$ and $\mu : \widehat{R}(\Sigma,\mathcal F,p) \to \frak g$ be given as in \eqref{form2424}. This proposition is a consequence of well-established results on non-abelian symplectic cut \cite{Wo20,Mei,MT}:
\begin{prop}\label{prop93
Let $\epsilon$ be a positive real number. If $\epsilon$ is small enough, then
there exists a compact symplectic manifold $\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)$
with a Hamiltonian $G$ action and a moment map $\widehat \mu:\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon) \to {\mathfrak g}$,
which satisfies the following properties:
\begin{enumerate}
\item[(i)] The image of the map $Q \circ \widehat \mu$ is equal to $\Delta^+_{\mathcal N}(\epsilon)$.
\item[(ii)] The open subspaces $(Q\circ \mu)^{-1}(\overset{\circ\hspace{3mm}}{\Delta_{\mathcal N}^+}(\epsilon))$ and
$(Q\circ \widehat \mu)^{-1}(\overset{\circ\hspace{3mm}}{\Delta_{\mathcal N}^+}(\epsilon))$ are symplecomorphic.
\end{enumerate}
\end{prop}
\begin{proof}[Sketch of the proof]
Let $\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)_0$ denote the subspace
$(Q\circ \mu)^{-1}(\overset{\circ\hspace{3mm}}{\Delta_{\mathcal N}^+}(\epsilon))$ of $\widehat{R}(\Sigma,\mathcal N,\mathcal F,p)$.
According to Proposition \ref{symp-ext}, this space has a symplectic structure if $\epsilon$ is small enough.
We compactify this space into a closed symplectic manifold in the following way.
Given $x \in \mu^{-1}(\Delta_{\mathcal N}(\epsilon))$, we may assume without loss of generality that there is $k$ such that:
\[
\alpha_i(\mu(x)) = \epsilon
\]
if and only if $1\leq i\leq k$.
The map $\Phi_i:\mu^{-1}({\mathfrak t})\to \bR$ defined as:
\[
\Phi_i(y)=\alpha_i(\mu(y))
\]
can be extended as a smooth function to a neighborhood $U_x\subset \widehat{R}(\Sigma,\mathcal F,p)$ of $x$. We define
a real-valued function $f_i$ on $U_x \times \bC^k$ as follows:
\[
f_i(y,\xi_1,\dots,\xi_k) = \epsilon^2 - \phi_i(y) - \vert \xi_i\vert^2.
\]
Then the function $f:=(f_1,\dots,f_k):U_x \times \bC^k\to \bR^k$ is the moment map for a Hamiltonian action of $T^k$ on $U_x\times \bC^k$.
The symplectic quotient $f^{-1}(0)/T^k$ contains a dense subset which is symplectomorphic to
$\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)_0\cap U_x$. To be a bit more detailed, if we map
$y \in \widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)_0 \cap U_x$ to the equivalence class of
$(y,\xi_1\,\dots,\xi_k)$ where $\xi_i = \sqrt{\vert\epsilon^2 - \Phi_i(y)\vert}$, then we obtain an open embedding of
$\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)_0 \cap U_x$ into $f^{-1}(0)/T^k$.
Condition \ref{cond92} (4) shows that $\overline U_x:=f^{-1}(0)/T^k$ is a smooth manifold. (See \cite[Proposition 6.2]{Wo20}.)
We can glue $\overline U_x$ for various choices of $x \in \mu^{-1}(\Delta(\epsilon))$ to obtain the desired symplectic manifold
$\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)$.
\end{proof}
Let $D$ denote the complement of $\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)_0 $ in $X:=\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)$. We expect that $\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)$ admits a K\"ahler structure compatible with the symplectic structure of $X$ denoted by $\omega$ such that $D$ forms a normal crossing divisor in this neighborhood. Therefore, in the light of Conjecture \ref{cong59}, we make the following conjecture:
\begin{conjecture}
There is an $A_\infty$-category ${\frak Fuk}^G(X\backslash D,\omega)$ associated to $(X,\omega)$ and $D$ as above, where the objects of
this category form a family of $G$-equivariant immersed Lagrangian submanifolds of $X\backslash D$, and the morphisms of
this category are constructed by holomorphic maps to $X$. The homotopy equivalence of this category is independent of
$\mathcal N$ and $\epsilon$.
\end{conjecture}
\begin{conjecture}
Suppose $\Sigma$ is a connected Riemann surface. Then the $A_\infty$-category has the properties of the
category ${\mathfrak I}(\Sigma,\mathcal F)$ in Conjecture \ref{conj21}.
For a disconnected $\Sigma$, we can take the tensor product of categories associated to the connected components.
\end{conjecture}
To elaborate on this proposal, let $(M,\mathcal E)$ be a cobordism from the empty pair to $(\Sigma,\mathcal F)$. As in the case of handlebodies discussed in Section \ref{sec:extendedmoduli}, we can associate to $(M,\mathcal E)$ a subspace of $\widehat{R}(\Sigma,\mathcal F,p)$ which lives in $\mu^{-1}(0)$. Therefore, it can be also regarded as a subspace of $X=\widehat{R}(\Sigma,\mathcal N,\mathcal F,p;\epsilon)$. A holonomy perturbation of this subspace can be used to turn this space into an immersed Lagrangian submanifold $\widetilde L_{(M,\mathcal E)}$.
\begin{conjecture}\label{conj84}
There exists a bounding cochain $b_M$ in
\[
H_G(\widetilde L_{(M,\mathcal E)} \times_{X} \widetilde L_{(M,\mathcal E)} ).
\]
Together with $b_M$, the immersed Lagrangian submanifold $\widetilde L_{(M,\mathcal E)}$ determines an object of
${\frak Fuk}^G(X\backslash D,\omega)$.
\end{conjecture}
If the construction of various moduli space can be carried out as expected, then the proof of Conjecture \ref{conj84} is expected to be similar to the similar result in the case of ${\rm SO}(3)$-bundles. (See \cite[Theorem 1.1 (1)]{fu7} and \cite{takagi}.)
\par
Let $(M_i,\mathcal E_i)$ be a pair such that:
\[
\partial(M_1,\mathcal E_1) = (\Sigma, \mathcal F)= -\partial(M_2,\mathcal E_2).
\]
Therefore, we can glue these two pairs to form a closed manifold $M$ and a $G$-bundle $\mathcal E$ over $M$. Conjecture \ref{conj84} asserts that there are pairs $(\widetilde L_{(M_1,\mathcal E_1)},b_1)$ and $(\widetilde L_{(M_2,\mathcal E_2)},b_2)$ of the objects of ${\frak Fuk}^G(X\backslash D,\omega)$.
\begin{conjecture}\label{cong8585}
The Lagrangian Floer homology
\[
HF_G((\widetilde L_{(M_1,\mathcal E_1)},b_1),(\widetilde L_{(M_2,\mathcal E_2)},b_2);X \setminus D)
\]
is an imvariant of $(M,\mathcal E_M)$.
\end{conjecture}
\section{Lagrangian Floer Theory in a Smooth Divisor Complement}\label{sec:divisor complement}
Let $(X,\omega)$ be a compact
symplctic manifold and $D$ be a codimension $2$ submanimfold. We assume that $(X,D)$ is a K\"ahler manifold in a neighborhood of $D$, and $D$ is a smooth divisor in this neighborhood.
\begin{definition}
Let $L_1$ and $L_2$ be compact subsets of $X\setminus D$.
We say $L_1$ is {\it Hamiltonian isotopic} to $L_2$ relative to $D$
if there exists a compactly supported time dependent Hamiltonian
$H : (X\backslash D) \times [0,1] \to \bR$ so that the Hamiltonian diffeomorphism $\varphi : X \setminus D \to X \setminus D$
generated by $H$ sends $L_1$ to $L_2$, that is, $\varphi(L_1) = L_2$.
\end{definition}
\begin{definition}
We say $L \subset X \setminus D$ is {\it monotone} if
\eqref{ineq32} holds for $\beta \in H_2(X\setminus D,L)$.
The minimal Maslov number of $L$ relative to $D$ is defined as:
\[
\inf \{\mu_L(\beta) \mid \beta \in \pi_2(X\setminus D,L),\,\omega(\beta) >0 \}.
\]
\end{definition}
In general, $\Lambda_0^{R}$,
the universal Novikov ring with ground ring $R$, consists of formal sums
$\sum_i c_i T^{\lambda_i}$ where $c_i \in R$, $\lambda_i \in \bR_{\ge 0}$, $\lim_{i\to\infty} \lambda_i = +\infty$, and
$T$ is a formal parameter. Similarly,
$\Lambda^{R}$ consists of $\sum_i c_i T^{\lambda_i}$ where $c_i \in R$, $\lambda_i \in \bR$, $\lim_{i\to\infty} \lambda_i = +\infty$.
If $R$ is a field then $\Lambda^{R}$ is also a field.
\begin{theorem}\label{mainthm}
{\rm (\cite{DF})}
Let $L_0,L_1$ be compact, monotone and spin Lagrangian submanifolds of $X\setminus D$.
We assume that {\rm (m.a)} or {\rm (m.b)} holds for these Lagrangains.
Then there is a vector space $HF(L_0,L_1;X\setminus D)$ over $\Lambda^\bQ$ which is called
the Lagrangian Floer homology of $L_0$ and $L_1$ relative to $D$, and satisfies the following properties:
\vspace{-5pt}
\begin{enumerate}
\item[(i)]
If $L_0$ is transversal to $L_1$ then we have
$$
{\rm rank}_{\Lambda^\bQ} HF(L_0,L_1;X\setminus D) \le \# (L_0 \cap L_1).
$$
\item[(ii)]
If $L_i'$ is Hamiltonian isotopic to $L_i$ in $X \setminus D$
for $i=0,1$ then
$$
HF(L_0,L_1;X\setminus D) \cong HF(L'_0,L'_1;X\setminus D)
$$
\item[(iii)]
If either $L_1 = L_0$ or $\pi_1(L_0) = \pi_1(L_1) =0$, then
we can take $\bQ$ as a coefficient ring instead of the
Novikov filed $\Lambda^\bQ$.
\item[(iv)]
If $L_0 = L_1 =L$ holds, then there exists a
spectral sequence whose $E^2$ page is $H(L;\bQ)$
and which converges to $HF(L,L;X\setminus D)$. \end{enumerate}
\end{theorem}
\begin{remark}
The main point in Theorem \ref{mainthm} is that
we do {\it not} assume $L_i$ is a monotone Lagrangian submanifold
in $X$, for $i=1,2$. The general theory of \cite{fooobook,fooobook2}
says that there is an obstruction to define Floer homology $HF(L_0,L_1)$.
The Floer homology $HF(L_0,L_1;X\setminus D)$ uses only holomorphic disks
which `do not intersect' $D$. Therefore, the situation is similar to
monotone Lagrangian Floer homology due to Oh \cite{Oh}.
If $X\setminus D$ is convex at infinity, then \cite{fooobook,fooobook2} imply
that we can define Floer homology $HF(L_0,L_1;X\setminus D)$ satisfying the properties mentioned in Theorem \ref{mainthm}.
Note that in Theorem \ref{mainthm}, we do not impose any kind of convexity assumption for $X \setminus D$.
The specialization of the construction of Theorem \ref{mainthm} to the case that $L_i$
is exact and the homology class of each component of $D$ is proportional to the Poincar\'e dual
of $[\omega]$ is given in \cite{sh}.
\end{remark}
\begin{proof}[Sketch of the proof]
We assume that $L_0$ is transversal to $L_1$. Let $p,q \in L_0 \cap L_1$. We consider the moduli space
$\overset{\circ}{\mathcal M}(p,q;\beta;L_0,L_1)$ of pseudo-holomorphic maps to $X\backslash D$
which satisfy \eqref{CR} and \eqref{form38} for $G=\{1\}$. Following Floer \cite{Fl:LHF} and Oh \cite{Oh}
(see also \cite[Chapter2]{fooobook}), we can define $HF(L_0,L_1;X\setminus D)$
if we obtain a compactification ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$ of our moduli space
$\overset{\circ}{\mathcal M}(p,q;\beta;L_0,L_1)$ with the following properties:
\begin{enumerate}\label{formula31}
\item[(I)] The compactification ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
carries a Kuranishi structure with boundary and corner.
\item[(II)] The codimension one boundary of this moduli space is identified with the union of
\begin{equation}\label{splitend}
{\mathcal M}^{\rm RGW}(p,r;\beta_1;L_0,L_1) \times
{\mathcal M}^{\rm RGW}(r,q;\beta_2;L_0,L_1)
\end{equation}
for various $r \in L_1 \cap L_2$ and $\beta_1,\beta_2$
with $\beta_1 + \beta_2 = \beta$. The (virtual) dimension $d(\beta)$ of ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
is determined by the homology class $\beta$ and satisfies $d(\beta) = d(\beta_1) + d(\beta_2) + 1$
for the boundary component in \eqref{splitend}.
\end{enumerate}
We fix a {\it multisection}\footnote{See \cite[Definition A1.21]{fooobook2})} (or equivalently a multivalued perturbation)
which is transversal to $0$ and which is compatible with the description of the boundary as in \eqref{splitend}.
Note that transversality implies that its zero set is the empty set when the virtual dimension is negative.
Therefore, the zero set is a finite set in the case that the virtual dimension is $0$.
Assuming $d(\beta) = 0$, let $\# {\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
be the number (counted with sign and multiplicity) of the
points in the zero set of the perturbed moduli space. Then we define:
\[
\partial [p] = \sum_{q,\beta} \# {\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)[q].
\]
Here the sum is taken over all $q \in L_0 \cap L_1$ and the homology classes $\beta$ such that $d(\beta) = 0$.
In case $L_0 = L_1$ or $\pi_1(L_0) = \pi_1(L_1) = 0$ the right hand side is a finite sum. Otherwise we use appropriate Novikov ring and put the
weight $T^{\omega(\beta)}$ to each of the
terms of the right hand side, so that the right hand side converges in $T$ adic topology.
As it is customary with other Floer theories, we can show $\partial\circ \partial = 0$
using the moduli spaces associated to the homology classes $\beta$
with $d(\beta) =1$. (See, for example, \cite{Fl:LHF,Oh}.) The proofs of parts $(ii)$ and $(iii)$) are also similar to the
proof of the corresponding statements in the case of usual monotone Lagrangian Floer homology.
If $X\backslash D$ is convex at infinity, then we can let ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
to be the closure of $\overset{\circ}{\mathcal M}(p,q;\beta;L_0,L_1)$ in the moduli space of stable holomorphic maps
to $X$. In this case, monotonicity can be used to show that \eqref{formula31} gives all the configurations
appearing in the boundary of ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$.
In case we do not assume monotonicity, disk bubble can occur as the other type
of boundary components. (See \cite[Subsection 2.4.5]{fooobook} for example.)
The stable map compactification in \cite[Subsection 7.1.4]{fooobook2} does {\it not} give a compactification
${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$ with the required properties.
The issue is that in the stable map compactification a stable map with a sphere bubble which
is contained completely in the divisor $D$ is included.
At the point of such stable maps,
the limit of the following two kinds of sequence of stable maps are mixed up.
\begin{enumerate}
\item[(A)]
A limit of a sequence of pseudo-holomorphic disks $u_i : (D^2,\partial D^2) \to (X,L)$
such that $u_i(D^2) \cap D = \emptyset$.
\item[(B)]
A limit of a sequence of stable maps $u_i : (\Sigma_i,\partial \Sigma_i) \to (X,L)$,
where $\Sigma_i$ is a disk plus sphere bubbles,
and such that $u_i(\Sigma_i) \cap D \ne \emptyset$.
\end{enumerate}
We need to include (A) in our moduli space but (B) is not supposed to be
an element of the moduli space.
As it is shown in Figure \ref{zu0}, elements given as the limit points of type (A) and type (B) can be mixed up in the stable map compactification.
Here all the sphere bubbles in the figure are contained in $D$. The numbers written in the sphere bubbles $S^2$ are the intersection numbers $[S^2] \cap D$. The numbers written at the roots of the sphere bubble are the intersection multiplicities of the disk with the divisor $D$.
The configuration shown as (a) is a limit of disks as in (A) above since $2 + (-2) = 1 + (-1) =0$.
The configuration shown as (b) is {\it not} a limit of disks as in (A),since $2 + (-1) \ne 0 \ne 1 + (-2)$. However, these two configurations can intersect
in the limit, which is the stable map shown as (c) in the figure.
Note that a limit of the configuration (b) in the figure can split into two pieces as shown in the figure.Then the union of the disk component together with sphere bubble rooted on it is {\it not} monotone. Thus if we include (b), then there will be a trouble to show \eqref{splitend}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{Newfigure-1.eps}
\caption{(a) and (b) mixed up.}
\label{zu0}
\end{figure}
The idea to resolve this issue is to use a compactification
which is different from the stable map compactification (in $X$).
We use the compactification used in relative
Gromov-Witten theory, where the
limits of type (A) and type (B) are clearly separated. (See \cite{LR,IP,JL,JL2,groSie,pa,TZ}.)
Namely in this compactification configuration (c) in the
figure comes with additional information so that the
limit of type (A) and of type (B) becomes a different element
in this compactifiation.
Using this fact we can then show the above properties (I)(II).
\end{proof}
\section{Lagrangian Floer Theory in a Smooth Divisor Complement}\label{sec:divisor complement}
Let $(X,\omega)$ be a compact
symplctic manifold and $D$ be a codimension $2$ submanimfold. We assume that $(X,D)$ is a K\"ahler manifold in a neighborhood of $D$, and $D$ is a smooth divisor in this neighborhood.
\begin{definition}
Let $L_1$ and $L_2$ be compact subsets of $X\setminus D$.
We say $L_1$ is {\it Hamiltonian isotopic} to $L_2$ relative to $D$
if there exists a compactly supported time dependent Hamiltonian
$H : (X\backslash D) \times [0,1] \to \bR$ so that the Hamiltonian diffeomorphism $\varphi : X \setminus D \to X \setminus D$
generated by $H$ sends $L_1$ to $L_2$, that is, $\varphi(L_1) = L_2$.
\end{definition}
\begin{definition}
We say $L \subset X \setminus D$ is {\it monotone} if
\eqref{ineq32} holds for $\beta \in H_2(X\setminus D,L)$.
The minimal Maslov number of $L$ relative to $D$ is defined as:
\[
\inf \{\mu_L(\beta) \mid \beta \in \pi_2(X\setminus D,L),\,\omega(\beta) >0 \}.
\]
\end{definition}
\begin{theorem}\label{mainthm}
{\rm (\cite{DF})}
Let $L_0,L_1$ be compact, monotone and spin Lagrangian submanifolds of $X\setminus D$.
We assume that {\rm (m.a)} or {\rm (m.b)} holds for these Lagrangains.
Then there is a vector space $HF(L_0,L_1;X\setminus D)$ over $\bQ$ which is called
the Lagrangian Floer homology of $L_0$ and $L_1$ relative to $D$, and satisfies the following properties:
\vspace{-5pt}
\begin{enumerate}
\item[(i)]
If $L_0$ is transversal to $L_1$ then we have
$$
{\rm rank}_{\bQ} HF(L_0,L_1;X\setminus D) \le \# (L_0 \cap L_1).
$$
\item[(ii)]
If $L_i'$ is Hamiltonian isotopic to $L_i$ in $X \setminus D$
for $i=0,1$ then
$$
HF(L_0,L_1;X\setminus D) \cong HF(L'_0,L'_1;X\setminus D)
$$
\item[(iii)]
If $L_0 = L_1 =L$ and {\rm (m.a)} holds, then there exists a
spectral sequence whose $E^2$ page is $H(L;\bQ)$
and which converges to $HF(L,L;X\setminus D)$. \marginpar{\color{red} Should we state this part about clean intersections?}
\end{enumerate}
\end{theorem}
\begin{remark}
The main point in Theorem \ref{mainthm} is that
we do {\it not} assume $L_i$ is a monotone Lagrangian submanifold
in $X$, for $i=1,2$. The general theory of \cite{fooobook,fooobook2}
says that there is an obstruction to define Floer homology $HF(L_0,L_1)$.
The Floer homology $HF(L_0,L_1;X\setminus D)$ uses only holomorphic disks
which `do not intersect' $D$. Therefore, the situation is similar to
monotone Lagrangian Floer homology due to Oh \cite{Oh}.
If $X\setminus D$ is convex at infinity, then \cite{fooobook,fooobook2} imply
that we can define Floer homology $HF(L_0,L_1;X\setminus D)$ satisfying the properties mentioned in Theorem \ref{mainthm}.
Note that in Theorem \ref{mainthm}, we do not impose any kind of convexity assumption for $X \setminus D$.
The specialization of the construction of Theorem \ref{mainthm} to the case that $L_i$
is exact and the homology class of each component of $D$ is proportional to the Poincar\'e dual
of $[\omega]$ is given in \cite{sh}.
\end{remark}
\begin{proof}[Sketch of the proof]
We assume that $L_0$ is transversal to $L_1$. Let $p,q \in L_0 \cap L_1$. We consider the moduli space
$\overset{\circ}{\mathcal M}(p,q;\beta;L_0,L_1)$ of pseudo-holomorphic maps to $X\backslash D$
which satisfy \eqref{CR} and \eqref{form38} for $G=\{1\}$. Following Floer \cite{Fl:LHF} and Oh \cite{Oh}
(see also \cite[Chapter2]{fooobook}), we can define $HF(L_0,L_1;X\setminus D)$
if we obtain a compactification ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$ of our moduli space
$\overset{\circ}{\mathcal M}(p,q;\beta;L_0,L_1)$ with the following properties:
\begin{enumerate}
\item[(I)] The compactification ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
carries a Kuranishi structure with boundary and corner.
\item[(II)] The codimension one boundary of this moduli space is identified with the union of
\begin{equation}\label{splitend}
{\mathcal M}^{\rm RGW}(p,r;\beta_1;L_0,L_1) \times
{\mathcal M}^{\rm RGW}(r,q;\beta_2;L_0,L_1)
\end{equation}
for various $r \in L_1 \cap L_2$ and $\beta_1,\beta_2$
with $\beta_1 + \beta_2 = \beta$. The (virtual) dimension $d(\beta)$ of ${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
is determined by the homology class $\beta$ and satisfies $d(\beta) = d(\beta_1) + d(\beta_2) + 1$
for the boundary component in \eqref{splitend}.
\end{enumerate}
We fix a {\it multisection}\footnote{See \cite[Definition A1.21]{fooobook2})} (or equivalently a multivalued perturbation)
which is transversal to $0$ and which is compatible with the description of the boundary as in \eqref{splitend}.
Note that transversality implies that its zero set is the empty set when the virtual dimension is negative.
Therefore, the zero set is a finite set in the case that the virtual dimension is $0$.
Assuming $d(\beta) = 0$, let $\# {\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$
be the number (counted with sign and multiplicity) of the
points in the zero set of the perturbed moduli space. Then we define:
\[
\partial [p] = \sum_{q,\beta} \# {\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)[q].
\]
Here the sum is taken over all $q \in L_0 \cap L_1$ and the homology classes $\beta$ such that $d(\beta) = 0$.
The properties of compactification imply that the right hand side is actually a finite sum.
As it is customary with other Floer theories, we can show $\partial\circ \partial = 0$ using the moduli spaces associated to the homology classes $\beta$
with $d(\beta) =1$. (See, for example, \cite{Fl:LHF,Oh}.) The proofs of parts $(ii)$ and $(iii)$) are also similar to the
proof of the corresponding statements in the case of usual monotone Lagrangian Floer homology.
The stable map compactification in \cite[Subsection 7.1.4]{fooobook2} does {\it not} give a compactification
${\mathcal M}^{\rm RGW}(p,q;\beta;L_0,L_1)$ with the required properties.
The issue is that in the stable map compactification a stable map with a sphere bubble which
is contained completely in the divisor $D$ is included.
{\color{red} At the point of such stable maps,
the limit of the following two kinds of sequence of stable maps mixed up.
\begin{enumerate}
\item[(a)]
A limit of a sequence of pseudo-holomorphic disks $u_i : (D^2,\partial D^2) \to (X,L)$
such that $u_i(D^2) \cap D = \emptyset$.
\item[(b)]
A limit of a sequence of stable maps $u_i : (\Sigma_i,\partial \Sigma_i) \to (X,L)$,
where $\Sigma_i$ is a disk plus sphere bubbles,
and such that $u_i(\Sigma_i) \cap D \ne \emptyset$.
\end{enumerate}
We need to include (a) in our moduli space but (b) is not supposed to be
an element of the moduli space.
If $X\setminus D$ is convex the above phenomenon cannot occur.
We can avoid this trouble by dimension counting if, for any
nontr ivial holomorphic map $v : S^2 \to D$, the inequality
$v_*([S^2]) \cap D > 0$ holds. In general it is hard to avoid it using the stable
map compactification.
The idea to resolve this issue is to use a compactification
which is different from the stable map compactification (in $X$).
We use the compactification used in relative
Gromov-Witten theory, where the
limits of type (a) and type (b) are clearly separated. (See \cite{IP,LR,pa,TZ}.)
We can then show the above properties (I)(II).}
\end{proof}
\section{Equivariant Lagrangian Floer Homology} \label{sec:equivariant}
Let $(X,\omega)$ be a symplectic manifold. For simplicity, we assume that $X$ is a spin manifold. For any Lagrangian submanifold $L$ of $X$, there is a homomorphism $\mu_L : H_2(X;L;\bZ) \to \bZ$ which is called the {\it Maslov index}. (See, for example, \cite[Definition 2.1.15]{fooobook}.)
\begin{definition} \label{monotone}
A Lagrangian submanifold $L$ is {\it monotone}, if there exists $c>0$ such that the following identity holds for
all $\beta \in \pi_2(X,L)$:
\begin{equation}\label{ineq32}
c\cdot {\mu_L}(\beta) = \omega(\beta).
\end{equation}
The {\it minimal Maslov number of $L$} is defined to be:
\begin{equation*}
\inf \{ \mu_L(\beta) \mid \beta \in \pi_2(X,L),\,\omega(\beta) >0 \}.
\end{equation*}
\end{definition}
Following Floer's original definition \cite{Fl:LHF}, Oh constructed Lagrangian Floer homology for a pair $L_0$ and $L_1$ of monotone Lagrangians, if one of the following conditions holds \cite{Oh}:
\begin{enumerate}
\item[(m.a)] The minimal Maslov numbers of $L_0$ and of $L_1$ are both strictly greater than 2.
\item[(m.b)] The Lagrangian submanifold $L_1$ is Hamiltonian isotopic to $L_0$.
\end{enumerate}
Lagrangian Floer homology can be enriched when there is a group action on the underlying symplectic manifold. Such constructions have been carried out in various ways in the literature. (See Remark \ref{works-on-equiv-HF}). Let a compact Lie group $G$ acts on $X$, preserving the symplectic structure $\omega$. We fix a $G$-equivariant almost complex structure $J$ which is compatible with $\omega$. Note that the space of all such almost complex structures is contractible because the set of all $G$-invariant Riemannian metrics is convex. In the following, $H_G^*(M)$ for a $G$-space $M$ denotes the $G$-equivariant cohomology of $M$ with coefficients in $\bR$. In the case that, $M$ is just a point, this group is denoted by $H_G^*$. The group $H_G^*(M)$ has the structure of a module over $H_G^*$ \cite{Bo:equi-co}.
\begin{theorem}\label{existequiv}
Let $L_0$, $L_1$ be $G$-equivariant spin Lagrangian submanifolds of $X$.
Suppose they are both monotone and satisfy either {\rm (m.a)} or {\rm (m.b)}.
Then there is a $H_G^*$-module $HF_G(L_0,L_1)$, called $G$-equivariant Lagrangian Floer homology of
$L_0$ and $L_1$.
In the case that the intersection $L_0 \cap L_1$ is clean, there exists a spectral sequence whose $E_2$ page is
$H_G^*(L_0\cap L_1)$ and which converges to $HF_G(L_0,L_1)$.
\end{theorem}
Recall that two submanifolds $L_0$ and $L_1$ of a smooth manifold $M$ have {\it clean intersection}, if $N=L_0\cap L_1$ is a smooth submanifold of $M$ and for any $x\in N$, we have $T_xN=T_xL_0\cap T_xL_1$.
\begin{proof}[Sketch of the proof]
We assume that the intersection $L_0\cap L_1$ is a disjoint union of finitely many $G$-orbits $G\cdot p$ for $p \in \mathcal A$. A {\it pseudo-holomorphic strip} $u : \bR \times [0,1] \to X$ is a map that satisfies the following Cauchy-Riemann equation:
\begin{equation} \label{CR}
\partial_t u+J\partial_\tau u=0
\end{equation}
We are interested in the moduli space of pseudo-holomorphic maps $u$ which satisfy the following boundary condition:
\begin{equation}\label{form38}
\aligned
u(\bR \times \{0\}) &\subset L_0, \qquad
&u(\bR \times \{1\}) \subset L_1 \\
\lim_{t\to+\infty} u(t,\tau) &\in G\cdot p, \qquad
&\lim_{t\to-\infty} u(t,\tau) \in G\cdot q.
\endaligned
\end{equation}
We will denote the homology classes of all such maps by $H(p,q)$. For a fixed $\beta \in H(p,q)$, let $\overset{\circ}{\mathcal M}(p, q;\beta;L_0,L_1)$ be the moduli space of pseudo-holomorphic maps satisfying \eqref{form38} and representing $\beta$, where we identify two maps $u$ and $u'$ if $u'(t,\tau) = u(t+t_0,\tau)$ for some $t_0 \in \bR$. Note that $\overset{\circ}{\mathcal M}( p, q;\beta;L_0,L_1)$ is invariant with respect to the action of the the group $G$. We also assume that this space is cut-down transversely by Equation \eqref{CR}. This moduli space can be compactified to a cornered manifold ${\mathcal M}(p, q;\beta;L_0,L_1)$ using {\it stable map compactification} \cite[Definition 10.3]{FO}. Codimension one boundary components of this space can be identified with the union of the fiber products:
\begin{equation}\label{form3838}
{\mathcal M}(p, r;\beta_1;L_0,L_1)
\times_{G \cdot r}{\mathcal M}(r, q;\beta_2;L_0,L_1)
\end{equation}
where $r \in \mathcal A$ and $\beta_1 \#\beta_2 = \beta$. Here $\#:H(p,r)\times H(r,q)\to H(p,q)$ is the concatenation of homology classes. Monotonicity and (m.a) or (m.b) are the main ingredients to prove these claims about stable map compactification.
The classifying space $BG$ and the universal bundle $EG$ over $BG$ can be approximated by finite dimensional manifolds $BG(N)$, $EG(N)$. To be more precise, suppose $EG(N)$ is a principal $G$-bundle over a manifold $BG(N)$ such that the homotopy groups of $EG(N)$ vanish up to degree $N$. We consider the approximate Borel construction
${\mathcal M}(p,q;\beta;L_0,L_1)
\times_G EG(N)$.
Taking asymptotic value as $t \to \pm \infty$, we obtain two evaluation maps as below:
\begin{equation*}\label{form39}
\begin{CD}
G\cdot p \times_G EG(N)
@<{{\rm ev}_{-\infty}}<<
{\mathcal M}(p, q;\beta;L_0,L_1)
\times_G EG(N)
@>{{\rm ev}_{+\infty}}>> G\cdot q \times_G EG(N).
\end{CD}
\end{equation*}
If ${\rm ev}_{+\infty}$ is a submersion, then we can define an operator:
\begin{equation}\label{form311009}
d_{p,q;\beta} : \Omega^*(G\cdot p \times_G EG(N))
\to \Omega^*(G\cdot q \times_G EG(N))
\end{equation}
between the space of differential forms by:
\begin{equation}\label{form31100}
d_{p,q;\beta}(h) = ({\rm ev}_{+\infty})_!({\rm ev}_{-\infty}^*h),
\end{equation}
where $({\rm ev}_{+\infty})_!$ is integration along the fiber.
Characterization of codimension one boundary components in \eqref{form3838} implies that:
\begin{equation}\label{form311}
d \circ d_{p,q;\beta} \pm d_{p,q;\beta} \circ d
= \sum_r\sum_{\beta_1+\beta_2=\beta} \pm
d_{p,r;\beta_1} \circ d_{r,q;\beta_2}.
\end{equation}
Here $d$ is the usual de Rham differential. Therefore, the map $\delta_N= d + \sum d_{p,q;\beta}$ defines a differential, {\it i.e.}, it satisfies $\delta_N \circ \delta_N = 0$. Taking the limit $N \to \infty$, we obtain the
equivariant Floer homology as the limit.
In general, it might be the case that ${\mathcal M}(p,q;\beta;L_0,L_1)$ is not a smooth manifold or ${\rm ev}_{+\infty}$ is not a submersion. Then we can use the theory of {\it Kuranishi structures} and {\it continuous family of perturbations} on Kuranishi spaces to prove the same conclusion. In fact, following \cite{FuFu5}, we obtain a $G$-equivariant Kuranishi structure on ${\mathcal M}(p,q;\beta;L_0,L_1)$ and hence a Kuranishi structure on ${\mathcal M}(p, q;\beta;L_0,L_1)\times_G EG(N)$. Then we can define a system of perturbations on these Kuranishi structures which give rise to a map as in \eqref{form311009} between the spaces of differential forms that satisfy \eqref{form311}.
\end{proof}
The elements of the moduli space ${\mathcal M}(p,q;\beta;L_0,L_1)$ can be regarded as solutions of a Fredholm equation which is defined on an infinite dimensional space and takes values in another infinite dimensional space. Roughly speaking, a Kuranishi structure on this moduli space replaces these infinite dimensional spaces with the spaces of finite dimensions. To be a bit more detailed, a Kuranishi structure is a covering of the moduli space with {\it Kuranishi charts}. For a point $p$ in the moduli space, a Kuranishi chart in a neighborhood of $p$ is a quadruple $(V,E,s,\psi)$ such that $V$ is a manifold, $E$ is a vector bundle, $s$ is a section of $E$ and $\psi$ is a homeomorphism from $s^{-1}(0)$ to an open neighborhood of $p$ in the moduli space. In general, we might need to work in the case that $V$ and $E$ are orbifold and orbi-bundle. Another part of the data of a Kuranishi structure is the set of coordinate change maps which explain how to glue different Kuranishi charts together. In order to get smooth spaces, we need to perturb the zero sets of the sections of Kuranishi charts in a consistent way, and continuous family of perturbations give a systematic way to achieve this goal. For a more detailed definition of Kuranishi structures and continuous family of perturbations, we refer the reader to \cite[Definition A1.5]{fooobook} and \cite{foootech2}.
\begin{remark} \label{works-on-equiv-HF}
In this section, we discussed an approach to equivariant Lagrangian Floer homology which is
given by applying the Borel construction {\it after} taking finite dimensional reduction.
This approach was proposed independently by the second author in \cite[Section 7]{fu-1}
and Viterbo. There are alternative approaches to equivariant Lagrangian Floer homology which also use
Borel construction but avoid virtual techniques. These approaches give rise to similar results as Theorem \ref{existequiv}
under more restrictive assumptions. In the case that $G=\bZ/2\bZ$, Floer homology coupled with Morse homology on
$EG$ is used in \cite{SeiSmi:loc} by Seidel and Smith to define
equivariant Lagrangian Floer homology .
More recently, Hendricks, Lipshitz and Sarkar employed homotopy theoretic methods to define Lagrangian Floer homology
in the presence of the action of a Lie group \cite{HLS:equiv-finite,HLS:equiv-Lie}.
There are also various other equivariant theories for other Floer homologies (see, for example,
\cite{Don:YM-Floer,KM:monopoles-3-man,AB:equiv-I}).
\end{remark}
\section{The Extended Moduli Spaces of Flat Connections}
\label{sec:extendedmoduli}
The most serious difficulty to define the symplectic counterpart of instanton Floer homology lies in the fact that the space of flat ${\rm SU}(2)$-connections on a Riemann surface is singular. In this section, we review Manolescu and Woodward's approach to go around this difficulty \cite{MW}. The main idea is to replace the moduli space of flat connections with the {\it extended moduli space of flat connections} \cite{Je,Hu}, which is a more nicely-behaved space.
Fix a compact, simply connected and semi-simple Lie group $\widetilde G$, and let $G = \widetilde G/Z(\widetilde G)$ be the associated {\it adjoint group}. Here $Z(\widetilde G)$ denotes the center of $\widetilde G$, which is a finite group. The adjoint action of $\widetilde G$ on itself induces an action of $G$ on $\widetilde G$, which is denoted by $ad$. We will also write $\mathfrak g$ for the Lie algebra of $\widetilde G$ (or equivalently $G$). We identify ${\mathfrak g}$ with the dual of the Lie algebra ${\mathfrak g}^*$ using the inner product $-B(\cdot,\cdot)$ where $B(\cdot,\cdot)$ is the Killing form.
Suppose $\Sigma$ is an oriented and connected surface with genus $g \geq 1$ and $\mathcal F$ is a principal $G$-bundle over $\Sigma$. Define the {\it gauge group} $\mathcal G(\Sigma,\mathcal F)$ to be the space of all smooth sections of the bundle $\mathcal F \times_{ad} \tilde G$. Any element of the gauge group induces an automorphism of the bundle $P$. Fix a base point $p\in \Sigma$, an open neighborhood $U$ of $p$ and a trivialization of the fiber of $\mathcal F$ over $p$. Then the {\it based gauge group} is defined to be:
\begin{equation}
\mathcal G_0(\Sigma,\mathcal F,p)= \{ g \mid g \in \mathcal G(\Sigma,\mathcal F),\,\,\,
g(q)= [e,\widetilde e]\in \mathcal F|_q \times_{ad} \tilde G \text{ for $q\in V\subseteq U$}\},
\end{equation}
where $e$ and $\widetilde e$ are respectively the unit elements of $G$ and $\widetilde G$ and $[e,\widetilde e]$ is the element of $\mathcal F|_q \times_{ad} \tilde G$ which is induced by the trivialization of $\mathcal F|_{q}$. Here $V$ is another open neighborhood of $q$ which is contained in $U$.
Let $\mathcal A_{\rm fl}(\Sigma,\mathcal F)$ denote the space of all flat connections on $\mathcal F$. Since the elements of $\mathcal G(\Sigma,\mathcal F)$ induce isomorphisms of $P$, this gauge group acts on $\mathcal A_{\rm fl}(\Sigma,\mathcal F)$. We will write $R(\Sigma,\mathcal F)$ for the quotient space. This quotient space has singular points in general. Away from the singular points, $R(\Sigma,\mathcal F)$ admits a symplectic structure \cite{go}. In the next definition, we recall the definition of a larger space which is a smooth symplectic manifold and can be used to replace $R(\Sigma,\mathcal F)$ for our purposes:
\begin{definition}
Suppose $\mathcal A_{\rm ex}(\Sigma,F,p)$ consists of the pairs $(A,\xi)$ where $A$ is a flat connection on
$\Sigma\backslash \{p\}$ and $\xi \in \mathfrak g$. We require that there is an open neighborhood $V$ of $p$, included in $U$,
such that the restriction of $A$ to $V$ is equal to $\xi ds$
with respect to the chosen trivialization of $\mathcal F$. Here we assume that $U$ is identified with the punctured disc centered at the origin
in the 2-dimensional Euclidean space and $ds$ represents the angular coordinate.
The group $\mathcal G_0(\Sigma,\mathcal F,p)$ clearly acts on
$\mathcal A_{\rm ex}(\Sigma,F,p)$ and the quotient space is denoted by ${\widehat R}(\Sigma,\mathcal F,p)$, and is called
the {\it extended moduli space of flat connections on $\mathcal F$}.
\end{definition}
The action of the based gauge group fixes the ${\mathfrak g}$-valued component of the elements of $\mathcal A_{\rm ex}(\Sigma,F,p)$. Therefore, we can define a map:
\begin{equation}\label{form2424}
\mu :
{\widehat R}(\Sigma,\mathcal F,p) \to \frak g
\end{equation}
by:
\[
\mu([A,\xi]) = \xi.
\]
There is also an action of $G$ on the space ${\widehat R}(\Sigma,\mathcal F,p)$. Given any element $g\in \widetilde G$, we can construct an element of $\mathcal G(\Sigma,\mathcal F,p)$ which is given by $g$ in a neighborhood of $p$ that is contained in $U$ and is extended arbitrarily to the rest of $\Sigma$. This element of the gauge group acts on ${\widehat R}(\Sigma,\mathcal F,p)$ and this action only depends on the image of $g$ in the adjoint group $G$.
\begin{prop}[\cite{Je,Hu}]\label{symp-ext}
There exist a $G$-invariant neighborhood ${\mathfrak U}$ of $0$ in ${\mathfrak g}$ and a symplectic structure on
$\mu^{-1}({\mathfrak U})\subset {\widehat R}(\Sigma,\mathcal F,p)$
such that the action of $G$ is Hamiltonian and $\mu$ is the moment map.
Furthermore, we have $R(\Sigma,\mathcal F)= \mu^{-1}(0)/G$, i.e.,
$R(\Sigma,\mathcal F)$ can be identified with the symplectic quotient ${\widehat R}(\Sigma,\mathcal F,p)/\!\!/G$.
\end{prop}
\begin{remark}
The extended moduli space for a disconnected $\Sigma$ is defined to be the product of the extended moduli spaces associated
to the connected components of $\Sigma$. To be more precise, we need to fix a base point $p_i$ for each connected component
$\Sigma_i $ of $\Sigma$. These base points altogether are denoted by $p$. Then:
\[
{\widehat R}(\Sigma,\mathcal F,p):=\prod_{i}{\widehat R}(\Sigma_i,\mathcal F_i,p_i)
\]
where $\mathcal F_i$ is the restriction of $\mathcal F$ to $\Sigma_i$. Note that there is an action of $G^m$ on the
moduli space in this case where $m$ is the number of connected components.
\end{remark}
For the rest of this section, we assume that $\widetilde G={\rm SU}(2)$, and $\mathcal F$ is the trivial ${\rm PU}(2)$-bundle. In this case, the corresponding extended moduli space will be denoted by ${\widehat R}(\Sigma,p)$. Consider the following subspace of $\mathfrak{su}(2)$, the Lie algebra of ${\rm SU}(2)$:
\begin{equation}\label{form44}
\Delta(\theta_0) =
\left\{ \xi \in \mathfrak{su}(2)\mid \xi \in \mathcal O_\theta
,\,\, \theta \in [0,\theta_0)
\right\}
\end{equation}
Here $\mathcal O_\theta$ is the adjoint orbit of the following element of $\mathfrak{su}(2)$:
\[
\left(
\begin{matrix}
2\pi {\bf i} \theta &0 \\
0& -2\pi {\bf i} \theta
\end{matrix}
\right)
\]
In Proposition \ref{symp-ext} and for the case that $\mathcal F$ is the trivial ${\rm PU}(2)$-bundle, we can assume that ${\mathfrak U}=\Delta(\theta_0)$ where $\theta_0 \leq \frac{1}{2}$ \cite{MW}.
Let $Q:\mathfrak{su}(2)\to \bR^{\geq 0}$ be the map which assigns to $\xi$ in \eqref{form44} the value $\theta$. Define the function:
\[
\overline{\mu}:=Q \circ \mu: {\widehat R}(\Sigma,p)\to [0,\infty).
\]
This function is smooth on the complement of $\mu^{-1}(0)$, and defines a Hamiltonian vector field for an $S^1$-action on $\mu^{-1}(\Delta(\frac{1}{2}))$. This $S^1$-action extends to $\mu^{-1}(\frac{1}{2})$. Consider the map:
\[
\widetilde \mu: {\widehat R}(\Sigma,p) \times \bC \to \bR
\]
defined by $\widetilde \mu(x,z)=\overline \mu(x)+\frac{1}{2}|z|^2$. This map is also a moment map for a Hamiltonian $S^1$-action. For each $0<\theta\leq \frac{1}{2}$, we define:
\begin{equation} \label{ex}
{\widehat R}(\Sigma,p,\theta):=\widetilde \mu^{-1}(\theta)/S^1
\end{equation}
This space is a smooth manifold and is called the {\it cut down extended moduli space}. The manifold ${\widehat R}(\Sigma,p,\theta)$ can be decomposed as the disjoint union of the following spaces:
\begin{equation} \label{decom-ex}
\mu^{-1}(\Delta(\theta)) \hspace{1cm}\overline{\mu}^{-1}(\theta)/S^1
\end{equation}
This construction is an example of non-abelian symplectic cut \cite{Wo20}.
\begin{theorem}\label{thm3434}{\rm(\cite{MW})}
For $0<\theta< \frac{1}{2}$, the closed manifold ${\widehat R}(\Sigma,p,\theta)$ admits a symplectic form. The codimension two
submanifold ${(\overline \mu)}^{-1}(\theta)/S^1$ of ${\widehat R}(\Sigma,p,\theta)$ is a symplectic hypersurface and
${\widehat R}(\Sigma,p,\theta)$ is K\"ahler in a neighborhood of ${(\overline \mu)}^{-1}(\theta)/S^1$.
The ${\rm PU}(2)$-action on $ \mu^{-1}(\Delta(\theta))\subset {\widehat R}(\Sigma,p,\theta)$ extends to
a Hamiltonian action on ${\widehat R}(\Sigma,p,\theta)$ whose moment map $\widehat \mu$
is the smooth extension of the map $\mu$
defined on $\mu^{-1}(\Delta(\theta))$. The symplectic quotient ${\widehat R}(\Sigma,p,\theta)/\!\!/{\rm PU}(2)$
can be identified with the moduli space of flat ${\rm SU}(2)$-connections ${R}(\Sigma)$ on $\Sigma$.
\end{theorem}
\begin{proof}
Let $\bC$ be given the negative of the standard symplectic form and consider the induced product symplectic form on
${\widehat R}(\Sigma,p) \times \bC$.
The ${\rm PU(2)}$-action on ${\widehat R}(\Sigma,p)$ induces a Hamiltonian action on ${\widehat R}(\Sigma,p) \times \bC$.
The map $\widetilde \mu$ is also the moment map for an $S^1$-action on the complement of $\mu^{-1}(0)\times \bC$ in
${\widehat R}(\Sigma,p) \times \bC$ which commutes with the ${\rm PU}(2)$-action.
The manifold ${\widehat R}(\Sigma,p,\theta)$ is the symplectic quotient of ${\widehat R}(\Sigma,p) \times \bC$
with respect to the $S^1$-action. Therefore, ${\widehat R}(\Sigma,p,\theta)$ admits a natural symplectic structure and a
Hamiltonian ${\rm PU(2)}$-action whose moment map is denoted by $\widehat \mu$.
It is straightforward to check that this action is an extension of the ${\rm PU}(2)$-action on $\mu^{-1}(\Delta(\theta))$,
and the symplectic quotient $\widehat \mu^{-1}(0)/{\rm PU}(2)$ is symplectomorphic to ${R}(\Sigma)$.
It is shown in \cite[Proposition 4.6]{MW} that the hypersurface $\overline{\mu}^{-1}(\theta)/S^1$ is symplectomorphic to:
\begin{equation*}
P(\Sigma,p,\theta) \times \mathcal O_{\theta}
\end{equation*}
where $P(\Sigma,p,\theta)$ is equal to the following symplectic quotient:
\begin{equation*}
P(\Sigma,p,\theta):=\mu^{-1}(\mathcal O_\theta)/{\rm PU}(2).
\end{equation*}
The adjoint orbit $ \mathcal O_{\theta}$ is a 2-sphere equipped with the symplectic form $2\theta\cdot \omega_0$ with $\omega_0$ being
the standard volume form on a 2-sphere. The symplectic manifold $P(\Sigma,p,\theta)$ can be identified with a moduli space
of parabolic bundles, and hence it is K\"ahler.
Now the claim about the symplectic form in a neighborhood of $\overline{\mu}^{-1}(\theta)/S^1$
is a consequence of the standard neighborhood theorems in symplectic geometry.
\end{proof}
\begin{remark}
Theorem \ref{thm3434} to some degree can be extended to the case that $\theta=\frac{1}{2}$. The smooth manifold
${\widehat R}(\Sigma,p,\frac{1}{2})$ admits a closed 2-form which is non-degenerate on $\mu^{-1}(\Delta(\frac{1}{2}))$.
The ${\rm PU}(2)$-action also extends to this manifold and the map $\mu$ on $\mu^{-1}(\Delta(\frac{1}{2}))$ determines
an $\mathfrak{su}(2)$-valued map on ${\widehat R}(\Sigma,p,\frac{1}{2})$ which is the `moment' map for the ${\rm PU}(2)$-action.
Since the symplectic form of ${\widehat R}(\Sigma,p,\frac{1}{2})$ is degenerate, the moment map condition should be interpreted
carefully, and we refer the reader to \cite{MW} for more details.
\end{remark}
A handlebody $H_g$ of genus $g$ determines a Lagrangian submanifold of ${\widehat R}(\Sigma,p,\theta)$. Assume that the boundary of $H_g$ is identified with $\Sigma_g$ and $\mathcal U$ is an open neighborhood of $p$ in $H_g$ whose intersection with the boundary is equal to $U$. Let $\mathcal A_{\rm fl}(H_g,p)$ be the space of all flat connections on the trivial ${\rm SU}(2)$-bundle over $H_g$ which are equal to the trivial connection in a neighborhood of $p$ contained in $\mathcal U$. Then the analogue of the based gauge group for the trivial ${\rm SU}(2)$-bundle on $H_g$ acts on $\mathcal A_{\rm fl}(H_g,p)$, and the quotient space is denoted by ${\tilde R(H_g,p)}$.
\begin{prop}[\cite{MW}] \label{handle-body-lag}
The restriction to the boundary induces a smooth embedding of ${\tilde R(H_g,p)}$ into ${\widehat R}(\Sigma,p,\theta)$.
This submanifold of ${\widehat R}(\Sigma,p,\theta)$ is Lagrangian and invariant with respect to the action of ${\rm PU}(2)$.
\end{prop}
In \cite{MW}, the notion of monotone Lagrangians are extended to appropriate families of manifolds with degenerate symplectic forms\footnote{For the definition of monotonicity of Lagrangians in non-degenerate symplectic manifolds see Definition \ref{monotone}.}. In particular, it is shown there that ${\tilde R(H_g,p)}$ gives rise to a monotone Lagrangian submanifold of ${\widehat R}(\Sigma,p,\frac{1}{2})$. Thus, given a Heegaard splitting $H^0_g \cup_{\Sigma} H^1_g$ of a 3-manifold $M$, we can construct two monotone Lagrangian submanifolds $L_0$ and $L_1$ of ${\widehat R}(\Sigma,p,\frac{1}{2})$. The degeneracies of the symplectic form on ${\widehat R}(\Sigma,p,\frac{1}{2})$ is also studied carefully in \cite{MW} and subsequently the Lagrangian Floer homology of $L_0$ and $L_1$ is defined. This Lagrangian Floer homology group is an invariant of the 3-manifold $M$. One might hope to use this 3-manifold invariant in the formulation of the Atiyah-Floer conjecture. However, this invariant is not isomorphic to instanton Floer homology even in the simplest case that $M$ is equal to $S^3$. The right candidate for the the Atiyah-Floer conjecture should incorporate the action of ${\rm PU}(2)$ and should be defined as an appropriate ${\rm PU}(2)$-equivariant Lagrangian Floer homology of the Lagrangians $L_0$ and $L_1$.
As an alternative approach, one can try to define Lagrangian Floer homology for the Lagrangians associated to ${\tilde R(H^0_g,p)}$ and ${\tilde R(H^1_g,p)}$ in the non-degenerate symplectic manifold ${\widehat R}(\Sigma,p,\theta)$ where $\theta<\frac{1}{2}$. Unfortunately, these Lagrangians are not monotone anymore and one cannot use the general construction of \cite{Oh} to define Lagrangian Floer homology. It is not even clear to the authors whether these Lagrangians are unobstructed in the sense of \cite{fooobook}. Nevertheless, the general construction of \cite{DF} shows that one can replace the monotonicity of ${\tilde R(H^i_g,p)}$ in ${\widehat R}(\Sigma,p,\theta)$ with a weaker condition. We shall review this construction in Section \ref{sec:divisor complement}. The advantage of this construction is that it avoids the detailed analysis of the degeneracy of the symplectic form on ${\widehat R}(\Sigma,p,\frac{1}{2})$. In particular, we believe this framework can be used to define Lagrangian Floer homology for other choices of the Lie group $\widetilde G$. It also seems that this construction can be combined more easily with equivariant Floer homology. We shall review the definition of two such equivariant theories in Section \ref{Sec:AFori}.
\section{Introduction}
In a series of groundbreaking papers, Floer adapted methods of Morse homology to infinite dimensional settings and defined various homology theories. After the great success of Donaldson's work in 4-manifold topology, Floer employed Yang-Mills gauge theory to introduce an invariant of 3-manifolds, known as {\it instanton Floer homology} \cite{Fl:I}. In another direction, Floer built up on Gromov's theory of pseudo-holomorphic curves and defined {\it Lagrangian Floer homology} \cite{Fl:LHF}. The Atiyah-Floer conjecture \cite{At} concerns a relationship between instanton and Lagrangian Floer homologies.
Given an integral homology sphere $M$, instanton Floer homology determines an invariant of $M$, denoted by $I^*(M)$. This invariant is the homology of a chain complex $C_*(M)$ whose generators are identified with the non-trivial flat ${\rm SU}(2)$-connections on $M$. The differential of this chain complex is defined by an appropriate count of {\it anti-self-dual} connections on the 4-manifold $\bR \times M$ which are asymptotic to flat connections on $M$. To follow this scheme more rigorously, flat ${\rm SU}(2)$-connections on $M$ are required to satisfy some non-degeneracy conditions. In general, one can avoid this issue by perturbing flat connections.
Lagrangian Floer homology is an invariant of a pair of Lagrangian submanifolds in a symplectic manifold. If $X$ is a symplectic manifold and $L_1$, $L_2$ are Lagrangian submanifolds of $M$, satisfying certain restrictive assumptions, then one can define Lagrangian Floer homology of $L_1$ and $L_2$, denoted by $HF(L_1,L_2)$. This invariant of $L_1$ and $L_2$ is the homology group of a chain complex $CF(L_1,L_2)$ whose set of generators can be identified with the intersection points of $L_1$ and $L_2$. The boundary operator for this chain complex is defined by an appropriate count of the number of pseudo-holomorphic strips which are bounded by $L_1$, $L_2$
and are asymptotic to given points in $L_1 \cap L_2$.
Let $\Sigma_g$ be a Riemann surface of genus $g\geq 1$. Then $R(\Sigma_g)$, the space of flat ${\rm SU}(2)$-connections on $\Sigma_g$ up to isomorphism, has a symplectic structure \cite{go}\footnote{To be more precise, the open dense subset of irreducible connections is a symplectic manifold.}. Suppose $M$ is an integral homology sphere with the following Heegaard splitting along the surface $\Sigma_g$:
\begin{equation}\label{handle-body}
M = H_g^0 \cup_{\Sigma_g} H_g^1
\end{equation}
where $H^i_g$ is a handle-body of genus $g$. Then the space $R(H^i_g)$ of flat connections on $\Sigma_g$ which extend to flat connections of $H_g^i$ determines a `Lagrangian subspace' of $R(\Sigma_g)$. The Atiyah-Floer conjecture states that:
\begin{equation}\label{AF111}
I^*(M) \cong HF(R(H^1_g),R(H^2_g))
\end{equation}
where the right hand side is the Lagrangian Floer homology of $R(H^1_g)$ and $R(H^2_g)$, regarded as Lagrangians in $R(\Sigma_g)$. This Floer homology group is called the {\it symplectic instanton Floer homology} of $M$.
One of the main difficulties with the Atiyah-Floer conjecture is to give a rigorous definition of symplectic instanton Floer homology. The spaces $R(\Sigma_g)$, $R(H^1_g)$ and $R(H^2_g)$ are singular due to the existence of flat connections with non-trivial isotropy groups. One of the main goals of this article is to sketch the construction of a candidate for symplectic instanton Floer homology of integral homology spheres. Our method is mainly based on the construction of \cite{MW}.
In \cite{MW}, Manolescu and Woodward replace $R(\Sigma_g)$ with another space $X$, which is a smooth closed manifold with a closed 2-form $\omega$. The 2-form $\omega$ is symplectic in the complement of a codimension 2 subspace $D\subset X$. For a Heegaard splitting as in \eqref{handle-body}, Lagrangian submanifolds $L_0,L_1$ of $ X\backslash D$ are also constructed in \cite{MW}. These Lagrangians are monotone in an appropriate sense. A detailed study of the degenerate locus of $\omega$ in \cite{MW} allows Manolescu and Woodward to define Lagrangian Floer homology of $L_0$ and $L_1$. However, this version of symplectic instanton Floer homology is not expected to be isomorphic to $I^*(M)$.
A variation of the triple $(X,\omega,D)$ leads to $(X_\theta,\omega_\theta,D)$ for $0<\theta<\frac{1}{2}$, where $\omega_\theta$ is a symplectic form on $X_\theta$. The submanifold $D\subset X_\theta$ is a smooth divisor with respect to a K\"ahler structure in a neighborhood of $D$, compatible with the symplectic form $\omega_\theta$. In Section \ref{sec:extendedmoduli}, we review the definition of $X$, $X_\theta$, which are respectively denoted by ${\widehat R}(\Sigma,p,\frac{1}{2})$, ${\widehat R}(\Sigma,p,\theta)$ and are constructed using the {\it extended moduli spaces of flat connections} \cite{Je,Hu}.
For the Heegaard splitting in \eqref{handle-body}, we can also construct Lagrangian submanifolds $L_0, L_1\subset X_\theta\backslash D$. These Lagrangian manifolds are monotone only in the complement of $D$. Therefore, the general theory of \cite{Oh} or \cite{fooobook} is not strong enough to define Lagrangian Floer homology of $L_0$ and $L_1$. In Section \ref{sec:divisor complement}, we sketch the results of \cite{DF} on the construction of Floer homology of Lagrangians which are monotone in the complement of a divisor. There is also a group action of ${\rm PU}(2)$ on $X_\theta$, and the Lagrangians $L_0$ and $L_1$ are invariant with respect to this group action. Section \ref{sec:equivariant} concerns a review of equivariant Lagrangian Floer homology. In Section \ref{Sec:AFori}, we explain two different approaches that the action of ${\rm PU}(2)$ on $X_\theta$ can be combined with Floer homology of Lagrangians in divisor complements. In particular, we introduce two versions of symplectic instanton Floer homology $I^*_{symp}(M)$ and ${\overline I}^*_{symp}(M)$. We conjecture that $I^*_{symp}(M)$ is isomorphic to $I^*(M)$. This conjecture can be regarded as a rigorous version of the Atiyah-Floer conjecture. A possible approach to verify this version of the Atiyah-Floer conjecture is the content of Section \ref{modulispace}. There is a different approach to study the Atiyah-Floer conjecture proposed by Salamon. For this proposal and related works see \cite{Sa,Sawe, We, We2,Duncan}.
The Atiyah-Floer conjecture can be considered as a part of a broader program to define `instanton Floer homology of 3-manifolds with boundary'. Yang-Mills gauge theory in dimensions 2,3 and 4 is expected to produce the following structures:
\begin{enumerate}
\item
For a 4-dimensional manifold and a ${\rm PU}(2)$-bundle, it defines a numerical invariant \`a la Donaldson.
\item
For a 3-dimensional manifold and a ${\rm PU}(2)$-bundle, it defines a group which is instanton Floer homology.
\item
For a 2-dimensional manifold and a ${\rm PU}(2)$-bundle, it defines a category.
\item
The Donaldson invariants of a 4-manifold $N$ with boundary gives an element of the instanton Floer homology of the boundary of $N$.
\item
Instanton Floer homology of a 3-dimensional manifold $M$ with boundary is expected to give an object of the category associated to $\partial M$.
\end{enumerate}
In 1992 and in a conference at the University of Warwick, Donaldson proposed a `first approximation' to the category associated to a Riemann surface $\Sigma$. In his proposal, an object of this category is a Lagrangian submanifold $L$ of the symplectic space $R(\Sigma)$. In this category, the set of morphisms between $L_1$ and $L_2$ should be defined as a version of Lagrangian Floer homology of $L_1$ and $L_2$. Then the Atiyah-Floer conjecture in \eqref{AF111} can be regarded as a consequence of the axioms of topological field theories. Of course, this plan faces with the same issue as before; the space $R(\Sigma)$ is singular and $HF(L_1,L_2)$ cannot be defined in a straightforward way.
In Section \ref{Sec:conjectures}, we use the language of $A_\infty$-categories to state conjectures reagarding stronger versions of the above properties for instanton Floer homology in dimensions 2 and 3. In particular, we work with more arbitrary $G$-bundles on surfaces and 3-manifolds in this section. The conjectures of Section \ref{Sec:conjectures} are based on the second author's proposals in \cite{fu3}. In fact, his main motivation to introduce $A_\infty$-categories associated to Lagrangians in symplectic manifolds arises from studying this problem. In Section \ref{sec:generalgroup}, we will discuss how we expect that the extended moduli spaces of flat connections can be employed to produce these expected structure. The required results from symplectic geometry is discussed in Section \ref{Sec:ainfinity}. When working with a non-trivial $G$-bundle $\mathcal F$ over a Riemann surface, there are various cases that the moduli space $R(\Sigma,\mathcal F)$ of flat connections on $\mathcal F$ is smooth. For such choices of $\mathcal F$, there is an alternative approach to realize the above construction of 2- and 3-dimensional topological field theory based on Yang-Mills theory \cite{fu6,takagi,DFL}. The expected relationship between the two approaches is discussed in the last section of the paper.
In the present article, we only sketch the proofs of most theorems and a more detailed treatment will appear elsewhere. There are two main reasons for this style of writing. Firstly, these proofs contain some technical parts, and the limited available space in this proceedings journal does not allow us to include them in this article. Secondly, we hope that the current exposition together with the statements of the theorems and the conjectures gives the reader a better big picture of the projects related to this paper.
\subsubsection*{Acknowledgement.} We would like thank Max Lipyanskiy, Paul Seidel, Mohammad Tehrani, Michael Thaddeus and Aleksey Zinger for helpful discussions. We are very grateful to the support of the {\it Simons Center for Geometry and Physics}. We also thank the {\it IBS Center for Geometry and Physics} in Pohang for its hospitality during the authors' visit in the Summer of 2016.
\section{Atiyah-Floer Conjecture and Moduli Space of Solutions to the Mixed Equation}
\label{modulispace}
In this section, we propose a program to prove Conjecture \ref{conj7878}. The main geometrical input in this program is a moduli space which is a mixture of the moduli space of Anti-Self-Dual connections and pseudo-holomorphic curves. Here we describe the version introduced in \cite{Li}. Similar moduli spaces appeared in \cite{fu4}. Analogous mixed moduli spaces are also being used by Max Lipyanskiy and the authors to prove an ${\rm SO}(3)$-analogue of the Atiyah-Floer conjecture \cite{DFL}.
Suppose $M$ is an integral homology sphere and a Heegaard splitting as in \eqref{form73} is fixed for $M$. Therefore, we can form the symplectic manifold $X={\widehat R}(\Sigma,p,\theta)$ and the Lagrangian submanifolds $L_i=\tilde R(H^i_g,p)$. For the simplicity of exposition, we assume that $L_0$ and $L_1$ have clean intersection. Recall that Lemma \ref{lem7474} states that in general we can perturb these Lagrangians by Hamiltonian isotopies to ensure that this assumption holds.
Let the domain $W$ in the complex plane $\bC$ be given as in Figure \ref{zu1}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{zu1.eps}
\caption{The domain $W$}
\label{zu1}
\end{figure}
We also decompose this domain into two parts $W_-$ and $W_+$ as in the figure and let $C = W_- \cap W_+$. Using the coordinate $t,\tau$ in the figure, the line $C$ is the part $\tau = 0$. The domain $W$ has four boundary components, denoted by $\partial_0W_-$, $\partial_1W_-$, $\partial_0W_+$, $\partial_1W_+$, and four ends as below:
\begin{equation}
\aligned
&\{(t,\tau) \mid \tau \in [-1,1], t < -K_0\},
\quad
\{(t,\tau) \mid \tau \in [-1,1], t > K_0\},
\\
&\{(t,\tau) \mid t \in [-1,1], \tau < -K_0\},
\quad
\{(t,\tau) \mid t \in [-1,1], \tau > K_0\}.
\endaligned
\end{equation}
We fix a Riemannian metric $g_W$ on $W$ which coincides with the standard Riemannian metric on the complex plane where $\vert t\vert$ or $\vert \tau\vert$ is large and outside a small neighborhood of $\partial_0W_+ \cup \partial_1W_+$. We also require that the metric is isometric to $(-\epsilon,0] \times \bR$ on a small neighborhood of $\partial_0W_{+}$, $\partial_1W_{+}$.
Fix a product metric on the product 4-manifold $W_+ \times \Sigma_g$. We glue $H_g^0 \times \bR$ and $H_g^1 \times \bR$ to the boundary components $\Sigma_g \times \partial_0W_+$
and $\Sigma_g \times \partial_1W_+$ of $W_+ \times \Sigma_g$, respectively. We will denote the resulting 4-manifold with $Y_+$ (cf. Figure \ref{zu2}).
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{zu2.eps}
\caption{The 4-dimensional manifold $Y_+$}
\label{zu2}
\end{figure}
The manifold $Y_+$ has three ends and one boundary component which is $\Sigma_g \times C$. The three ends correspond to the part $t \to \pm \infty$ and $\tau \to + \infty$, and they can be identified with:
\begin{equation}\label{form7171}
H^0_g \times (-\infty,-K_0),\quad H^1_g \times (K_0,+\infty),\quad (H^0_g \cup_{\Sigma_g} H^1_g) \times (K_0,+\infty).
\end{equation}
We extend the product Riemannian metric on $\Sigma_g \times W_+$ to $Y_+$ so that the ends in \eqref{form7171} have the product Riemannian metric. Note that $H^0_g \cup_{\Sigma_g} H^1_g$ in \eqref{form7171} is the integral homology sphere $M$.
Consider the decomposition:
\[
L_0 \cap L_1 = \{\theta\} \cup \bigcup_{a \in \mathcal A} R_a
\]
as in \eqref{decompLL}. The set $\mathcal A$ is identified with the set of irreducible flat connections on the trivial $SU(2)$-bundle over $M$.
\begin{definition}\label{defn71}
Let $a,b \in \mathcal A$. We say the pair $(u,A)$ satisfies the {\it mixed equation}, if they satisfy the following properties.
The first two conditions are constraints on the map $u$:
\begin{enumerate}
\item[(1.1)] $u: W_- \to X \setminus D$ is a holomorphic map with finite energy.
Here $X$ and $D$ are given in \eqref{symp-div}, and the energy of $u$ is defined to be:
\[
\int_{W_0} u^* \omega
\]
with $\omega$ being the symplectic form of $X$.
\item[(1.2)] The map $u$ satisfies the boundary conditions $u(\partial_0W_-)\subset L_0$ and
$u(\partial_1W_-)\subset L_1$. Moreover, we require that for $t \in [-1,1]$, we have:
\[
\lim_{\tau\to -\infty} u(\tau,t) = \frak p \in R_a.
\]
Here $\frak p$ is an element of $R_a$ which is independent of $t$.
\end{enumerate}
The next two conditions are on the connection $A$:
\begin{enumerate}
\item[(2.1)] $A$ is a connection on the trivial ${\rm SU}(2)$-bundle over $Y_+$ which satisfies the anti-self-duality equation
\begin{equation*}
F^+(A)=0
\end{equation*}
and its energy given by:
\[
\int_{Y_+} \vert F_A\vert^2 \, dvol
\]
is finite.
\item[(2.2)]
For $\tau>K_0$, let $A_\tau$ denote the restriction of $A$ to $(H^0_g \cup_{\Sigma_g} H^+_g) \times \{\tau\} \cong M $.
The connection $A_{\tau}$ on $M$ converges to the flat connection $b$ as $\tau$ goes to $+\infty$.
\end{enumerate}
The last three conditions are matching conditions for $u$ and $A$ on the borderline $C$:
\begin{enumerate}
\item[(3.1)]
If $(t,0) \in C$, then $u(t,0) \in \widehat \mu^{-1}(0)$, where $\widehat \mu:X \to \frak{su}(2)$ is the moment map of
Theorem \ref{thm3434}.
\item[(3.2)]
The restriction of $A$ to $\Sigma_g \times \{(t,0)\} \subset \partial X_+$, denoted by $A_{(t,0)}$, is flat for any $(t,0) \in C$.
\item[(3.3)]
The gauge equivalence class of the flat connection $A_{(t,0)}$ coincides with the equivalence class $[u(t,0)]$
of $u(t,0)$ in ${\widehat \mu}^{-1}(0)/{\rm PU}(2) = R(\Sigma)$. (See Theorem \ref{thm3434}.)
\end{enumerate}
\end{definition}
\begin{definition}\label{defn72}
Suppose $(u,A)$, $(u',A')$ are two pairs that satisfy the mixed equation. These two elements are equivalent, if
there exists a gauge transformation $g$ on $Y_+$ and $h \in {\rm PU}(2)$ such that:
\[
A' = g_*A\qquad u' = h u.
\]
We will write $\overset{\circ}{\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ for the space of equivalence classes of the pairs
$(u,A)$ satisfying the mixed equation and the following energy constraint:
\[
E=\int_{W_0} u^* \omega + \int_{Y_+} \Vert F_A\Vert^2.
\]
\end{definition}
We wish to show that the moduli space $\overset{\circ}{\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ behaves nicely and it can be compactified in a way that we can use it to construct an isomorphism between $I^*(M)$ and $I_{symp}^*(M)$. This requires us to generalize the analytical results of \cite{Li, DFL}. The matching condition in Definition \ref{defn72} can be regarded as a Lagrangian boundary condition associated to a Lagrangian correspondence from the infinite dimensional space of ${\rm SU}(2)$ connections over $\Sigma$ to $X$. A similar infinite dimensional Lagrangian correspondence appears in \cite{Li, DFL}. However, the Lagrangian correspondence in the present context is singular. Therefore, proving the required analytical results for the moduli space $\overset{\circ}{\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ (such as Fredholm theory, regularity and compactness) seems to be more challenging. Nevertheless, we conjecture that this moduli space satisfies these properties and it can be compactified to a space ${\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$.
This compactification ${\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ is expected to have a virtual fundamental chain whose boundary is the union of the following two types of spaces. The first type is:
\begin{equation}\label{boundary73}
\overline{\mathcal M}^{\rm RGW}(a,c;\beta;L_0,L_1)\times {\mathcal M}(W_-,Y_+,L_0,L_1;c,b;E_2)
\end{equation}
with $c \in \mathcal A$ and $\omega[\beta]+E_2 = E$, and the second type is:
\begin{equation}\label{boundary74}
{\mathcal M}(W_-,Y_+,L_0,L_1;a,c;E_1)
\times
\mathcal M(c,b;E_2;M)
\end{equation}
with $c \in \mathcal A$ and $E_1 + E_2 = E$. Here $\overline{\mathcal M}^{\rm RGW}(a,c;\beta;L_0,L_1)$ and $\mathcal M(c,b;E_2;M)$ are the moduli spaces that appeared in Section \ref{Sec:AFori}.
Assuming the existence of compactification ${\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ with the above properties, we define a map $\Phi:C^*_{symp}(M) \to C^*(M)$ as follows:
\[
\Phi(a) = \sum_{b,E}\#{\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)[b].
\]
where the sum is over all choices of $E$ and $b$ such that ${\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ is 0-dimensional. The signed number of points in this 0-dimensional moduli space is denoted by $\#{\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$.
By a standard argument applying to the 1-dimensional moduli spaces ${\mathcal M}(W_-,Y_+,L_0,L_1;a,b;E)$ and using the description of the boundary of this moduli space in \eqref{boundary73}, \eqref{boundary74}, we can conclude this implies that $\Phi$ is a chain map. The energy $0$ part of the moduli space ${\mathcal M}(W_-,Y_+,L_0,L_1;a,b;0)$ is empty if $a \ne b$ and has one point if $a=b$. It implies that $\Phi$ induces an isomorphism between corresponding Floer homologies.
\begin{remark}
Note that in Definition \ref{defn71}, we do not assume any particular asymptotic boundary conditions on the ends where $t \to \pm\infty$.
In fact, the finiteness of the energy should imply that the pair $(u,A)$ converges to a constant map to a flat connection on $H_g^i$ as $t \to \pm\infty$.
Therefore, the choices of Definition \ref{defn71} imply asymptotic convergence to the fundamental class of ${\tilde R(H^0_g,p)}$ on these ends.
This particular choice of the asymptotic boundary condition at $t \to \pm\infty$ is very important to show that $\Phi$ induces an isomorphism in homology.
In fact, we use it to show that the contribution of the lowest energy part to $\Phi$ is the identity map.
\end{remark}
\begin{remark}
The map $\Phi$ is defined similar to some chain maps which appear in \cite{fu3}. In the definition of these chain maps the Lagrangian ${\tilde R(H^0_g,p)}$ is replaced with arbitrary Lagrangian submanifold of
the underlying symplectic manifold. However, the idea that such maps can be used to construct isomorphisms is inspired by Lekili and Lipyanskiy's work in \cite{LL}, where the methods of \cite{fu3} is revived
in a similar context.
\end{remark}
\begin{remark}
The special case of the ${\rm SO}(3)$-Atiyah-Floer conjecture for mapping tori of surface diffeomorphisms was proved
in the seminal work of Dostoglou and Salamon \cite{DS}. Their proof uses an adiabatic limit argument and is based on the following crucial observation.
Consider the 4-manifold $\Sigma_g \times W$, where $W$ is a surface, and let the metric on $\Sigma_g$ degenerate.
Then the ASD equation turns into the holomorphic curve equation
from $W$ to the space $R(\Sigma_g)$ of flat connections on $\Sigma_g$.
Later, Salamon proposed a program for the original version of the Atiyah-Floer conjecture using similar adiabatic limit argument \cite{Sa} and this
approach was pursued further by Salamon and Wehrheim \cite{Sawe, We, We2}.
The extension of the adiabatic limit argument to the general case of the ${\rm SO}(3)$-analogue of the Atiyah-Floer conjecture is also being investigated
by David Duncan \cite{Duncan}.
The adiabatic limit argument has the potential advantage of finding a relationship between the moduli spaces involved in gauge theory and symplectic geometry, and not only a relationship at the level of Floer homologies.
The drawback is one has to face complicated analytical arguments.
We believe the approach discussed in this section (and the corresponding one in \cite{DFL} for the
${\rm SO}(3)$-analogue of the Atiyah-Floer conjecture) has less analytical difficulties because it uses the functorial properties
of Floer homologies. A similar phenomenon appears in the proof of the connected sum theorem for instanton Floer homology
of integral homology spheres where the ``functorial'' approach \cite{fu15,Don:YM-Floer} seems to be easier than the adiabatic limit argument \cite{WL}.
\end{remark}
\section{Yang-Mills Gauge Theory and 3-Manifolds with Boundary}
\label{Sec:conjectures}
In Section \ref{Sec:AFori}, we sketched the construction of $\overline I_{symp}(M)$, as a module over $H^*_{{\rm PU}(2)}$, for an integral homology sphere $M$. This invariant is defined by considering Yang-Mills gauge theory on principal ${\rm SU}(2)$-bundles. It is natural to ask to what extent this construction can be generalized to arbitrary 3-manifolds and arbitrary choice of principal bundles. In the following conjecture, suppose $G$ is given as in Section \ref{sec:extendedmoduli}.
\begin{conjecture}
Suppose $\mathcal E$ is a $G$-bundle over a 3-manifold $M$. Then there is a $H^*_{G}$-module $\overline I^*_{symp}(M,\mathcal E)$
which is an invariant of the pair $(M,\mathcal E)$. In the case that $M$ is an integral homology sphere and $G={\rm PU}(2)$,
this invariant matches with the construction of Section \ref{Sec:AFori}.
\end{conjecture}
\noindent
To be more precise, we expect that the above invariant is defined using Lagrangian Floer homology on appropriate moduli spaces of flat $G$-connections over Riemann surfaces. We shall propose a plan for the construction of this invariant in Section \ref{sec:generalgroup}. The tools from symplectic topology required for this construction are discussed in the next section.
In another level of generalization, one can hope to define symplectic instanton Floer homology for 3-manifolds with boundary. In order to state the expected structure of symplectic instanton Floer homology for 3-manifolds with boundary, we need to recall the definition of $A_\infty$-categories:
\begin{definition}
Fix a commutative ring $R$. An $A_{\infty}$-category $\mathscr C$ consists of a set of objects $\mathscr{OB}(\mathscr C)$,
a graded $R$-module of {\it morphisms} $\mathscr C(c,c')$ for each pair of objects $c,c' \in \mathscr{OB}(\mathscr C)$,
and the structural operations $\frak m_k : \bigotimes_{i=1}^k \mathscr C(c_{i-1},c_i)\to \mathscr C(c_{0},c_k)$ of degree $k-2$
for each $k\geq 1$.
The multiplication maps $\frak m_k$ are required to satisfy the following relations:
\begin{equation}\label{formula25}
\sum_{k_1+k_2=k+1}\sum_{i=0}^{k_1-1}(-1)^* \frak m_{k_1}(x_1,\dots,x_i,\frak m_{k_2}(x_{i+1},\dots,x_{k_2}),\dots,x_k)=0
\end{equation}
where $* = i +\sum_{j=1}^i \deg x_j$.
\end{definition}
Let $M$ be a 3-dimensional manifold whose boundary is decomposed as below:
\begin{equation*}
\partial M = -\Sigma_1 \sqcup \Sigma_2.
\end{equation*}
where $-\Sigma_1$ denotes the 3-manifold $\Sigma_1$ with the reverse orientation. Suppose also $\mathcal E$ is a $G$-bundle on $M$ whose restriction to $\Sigma_i$ is denoted by $\mathcal F_i$. We shall say $(M,\mathcal E)$ is a {\it cobordism} from $(\Sigma_1,\mathcal F_1) $ to $(\Sigma_2,\mathcal F_2)$ and we shall write:
\begin{equation}\label{3mfdassum}
(M,\mathcal E): (\Sigma_1,\mathcal F_1) \to(\Sigma_2,\mathcal F_2).
\end{equation}
\begin{conjecture}\label{conj21}
\begin{enumerate}
\item[(A-1)]
For any $G$-bundle $\mathcal F$ over a Riemann surface,
there is a unital\footnote{A filtered $A_{\infty}$ category is unital if it has a strict unit.} filtered $A_{\infty}$-category
${\mathfrak I}(\Sigma,\mathcal F)$ over the ring $\Lambda_0^{H^*_{G}}$ The $A_\infty$-category associated to $(-\Sigma,\mathcal F) $ is
${\mathfrak I}(\Sigma,\mathcal F)^{\rm op}$, the opposite $A_{\infty}$ category of
${\mathfrak I}(\Sigma,\mathcal F)$\footnote{The opposite $A_{\infty}$ category is defined
by reversing the direction of arrows. See \cite[Definition 7.8]{fu6}.}.
Moreover, if $\Sigma$ is the disjoint union of two surfaces $\Sigma_1$, $\Sigma_2$, and
the restriction of $\mathcal E$ to $\Sigma_i$ is $\mathcal E_i$, then we have the identification:
\[
{\mathfrak I}(\Sigma,\mathcal F)\cong {\mathfrak I}(\Sigma_1,\mathcal F_1)\otimes {\mathfrak I}(\Sigma_2,\mathcal F_2).
\]
Here the right hand side is the tensor product of filtered $A_{\infty}$-categories. (See \cite{Limo,FuFu50}.)
\item[(A-2)]
For any pair as in \eqref{3mfdassum}, there is a filtered
$A_{\infty}$ functor\footnote{In the terminology of \cite{FuFu50},
${\mathfrak I}_{(M,\mathcal E)}$ is a {\it strict} filtered $A_{\infty}$ functor.}:
\[
{\mathfrak I}_{(M,\mathcal E)} : {\mathfrak I}(\Sigma_1,\mathcal F_1)\to
{\mathfrak I}(\Sigma_2,\mathcal F_2).
\]
The $A_\infty$-functor associated to $(-M,\mathcal E)$ is the adjoint functor of
${\mathfrak I}(M,\mathcal E)$\footnote{See \cite{FuFu50} for the definition of adjoint functor of a filtered $A_{\infty}$functor.}.
\item[(A-3)]
For $i=1,2$, let $(M_i,\mathcal E_i)$ be a 3-dimensional cobordism from $(\Sigma_i,\mathcal F_i)$ to
$(\Sigma_{i+1},\mathcal F_{i+1})$.
Let $(M,\mathcal E)$ be the result of composing these cobordisms along $(\Sigma_2,\mathcal F_2)$. Then:
\begin{equation}\label{form14}
{\mathfrak I}_{(M,\mathcal E)}\cong {\mathfrak I}_{(M_2,\mathcal E_2)}\circ {\mathfrak I}_{(M_1,\mathcal E_1)}.
\end{equation}
Here $\circ$ is the composition of filtered $A_{\infty}$-functors
and $\cong$ is the homotopy equivalence of filtered
$A_{\infty}$ functors\footnote{Two filtered $A_{\infty}$ functors are homotopy equivalent if they are
homotopy equivalent in the functor category. (See \cite[Theorem 7.55]{fu6}.)}
from ${\mathfrak I}(\Sigma_1,\mathcal F_1)$ to ${\mathfrak I}(\Sigma_2,\mathcal F_2)$.
\end{enumerate}
\end{conjecture}
The following conjecture extends Conjecture \ref{conj21} to the case that at least one of the ends of $(M,\mathcal E)$ is empty:
\begin{conjecture}
Let $(M,\mathcal E)$ be as in \eqref{3mfdassum}:
\begin{enumerate}
\item[(B-1)]
If $\Sigma_1=\emptyset$, then ${\mathfrak I}_{(M,\mathcal E)}$ is an object of ${\mathfrak I}(\Sigma_2,\mathcal F_2)$.
\item[(B-2)]
If $\Sigma_2=\emptyset$, then ${\mathfrak I}_{(M,\mathcal E)}$ is a filtered $A_{\infty}$ functor from ${\mathfrak I}(\Sigma_1,\mathcal F_1)$ to
$\mathcal{CH}$, where $\mathcal{CH}$ is the $DG$ category of chain complexes.
\item[(B-3)]
If $\Sigma_1=\Sigma_2=\emptyset$, then ${\mathfrak I}_{(M,\mathcal E)}$ is a chain complex over $\Lambda^{H^*_{G}}_0$.
\end{enumerate}
\end{conjecture}
The next conjecture is an extension of part (A-3) of Conjecture \ref{conj21} to the case that one of the boundary components is empty:
\begin{conjecture}
Let $(M_1,\mathcal E_1)$ and $(M_2,\mathcal E_2)$ be as in part {\rm (A-3)} of Conjecture \ref{conj21}:
\begin{enumerate}
\item[(C-1)] If $\Sigma_1 = \emptyset$, then:
\begin{equation}\label{eqC1}
{\mathfrak I}_{(M,\mathcal E)}\cong {\mathfrak I}_{(M_2,\mathcal E_2)}({\mathfrak I}_{(M_1,\mathcal E_1)})
\end{equation}
This is a homotopy equivalence of objects in the category ${\mathfrak I}(\Sigma_3,\mathcal F_3)$.
\item[(C-2)] If $\Sigma_2= \emptyset$, then \eqref{form14} as the homotopy equivalence of
$A_\infty$ functors from the category ${\mathfrak I}(\Sigma_1,\mathcal F_1)$ to the category $\mathcal{CH}$ holds.
\item[(C-3)] If $\Sigma_1 =\Sigma_2= \emptyset$, then \eqref{eqC1} as a chain homotopy equivalence between chain complexes
holds.
\end{enumerate}
\end{conjecture}
|
\section{Introduction}
Nonlinear dynamical systems exhibit a vast variety of behaviors, from simple effects such as harmonic generation to sophisticated multiple bifurcations, to pattern formation, and chaos \cite{Strogatz,Cross1993,Sagdeev}. Particularly interesting are strongly nonlinear phenomena in the quantum regime, which can be realized in low-dissipative microwave systems such as circuit quantum electrodynamics (cQED) devices. Such phenomena play a central role and are widely employed in quantum information technology for qubit read-out \cite{Siddiqi2004,Lin2014,Krantz2016}, photon entanglement \cite{Eichler2011,Bergeal2012,Flurin2012,Macklin2015} and generation of Schr\"odinger cat states \cite{Mirrahimi2014,Puri2017}.
Period-multiplying subharmonic oscillations \cite{Hayashi,JordanSmith} constitute a particular class of nonlinear phenomena. The oscillations appear as a nonlinear response at the oscillator frequency to an external drive at a multiple of the resonant frequency. In the quantum picture, the elementary process that underlies the subharmonic oscillations is a decay of a single photon into three, four, or more photons. The subharmonic oscillations are described by non-perturbative solutions to the dynamical equations, which appear abruptly and co-exist with the stable vacuum state. In this respect the subharmonic oscillations distinctly differ from conventional parametric oscillations, which gradually emerge as a result of the vacuum instability. This difference is analogous to the difference between a first and a second order phase transition \cite{Haken1975}. Furthermore, a symmetry breaking aspect of this difference has important implications for the quantum dynamics of the period-tripling oscillations \cite{Zhang2017}.
Although the period-multiplying phenomenon is theoretically explained in textbooks, experimental demonstrations are not common. A few early observations of subharmonic resonances in electromagnetic devices were performed on essentially classical electrical circuits with saturable inductors \cite{Spitzer1945} or varactors \cite{Linsay1981}. More recent reports concern subharmonic resonances in lasers \cite{Ngai1993,SotoCrespo2004}. In Josephson circuits, the period-multiplying phenomenon has not received much attention; instead, research was focused on transition to chaos \cite{Huberman1980,Kautz1981} and lately on bifurcation phenomena \cite{Siddiqi2005,Vijay2009} and parametric oscillations \cite{Wilson2010,Wustmann2013,Lin2014,Krantz2016,Yamamoto2016}. Only recently, subharmonic oscillations in the quantum regime were theoretically discussed in the context of cQED \cite{Guo2013,Denisenko2016,Zhang2017}.
In this paper we report the first experimental observation of period-tripling subharmonic oscillations in a driven superconducting resonator in the quantum regime, $k_BT\ll \hbar\omega$ \cite{Sandberg2008,Wallquist2006,PalaciosLaloy2008}, where the thermal energy is much smaller than the energy of a single photon. We drive the nonlinear resonator with a harmonic signal, with power $P_d$, at frequency $3\omega$, approximately equal to three times the fundamental resonator mode frequency, and observe a strong response at $\omega$. The output microwave signal consists of three correlated beams with equal amplitudes and different phases, shifted by $2\pi/3$ radians. The oscillations are detected within a certain window of the driving field amplitude: they start at finite threshold detuning within the resonator bandwidth and persist deep into the red detuning region.
Our observations can be qualitatively understood from the theory of a nonlinear oscillator \cite{Hayashi,JordanSmith}. When a driving force with off-resonant frequency $3\omega$ is applied, it generates a linear response at the same frequency, which is downconverted to frequency $\omega$ due to nonlinearity. The downconversion has the highest efficiency when the detuning, $\delta_1= \omega-\omega_1$, from the oscillator resonant frequency $\omega_1$ is small, $\delta_1 \ll \omega_1$.
However, application of this single-mode scenario to the resonator setting requires additional considerations.
In our experiment, we drive the resonator close to its second mode, $\omega_2$, such that $\delta_2 = 3\omega - \omega_2 \ll \omega_1$. As a result, the response at the driving frequency becomes strongly enhanced and nonlinear, and the system dynamics is well described by two strongly interacting modes. This situation is different from the single mode oscillator model explored in previous works on subharmonic oscillations. In fact, this resonant enhancement of the external drive, by more than three orders of magnitude, is crucial for the possibility to access the subharmonic oscillation regime in experiments.
\section{Experiment}
\begin{figure}
\center
\includegraphics{Setup_a.pdf}
\caption{\label{fig:SampleAndSetup}(a) Optical micrograph of one of the samples. The four coplanar waveguide resonators meander between the on-chip flux line on the left and the contact pads on the right. (b) Zoom-in on the SQUID that terminates the bottom resonator. The SQUID is designed with two identical Josephson junctions. (c) Measurement setup. Microwave signals are applied via attenuated coaxial cables, one for direct driving by an external AC current, and one for flux modulation. To separate input and output, a directional coupler is used to route the signal. A static magnetic flux is induced by a superconducting coil. The output signal is amplified in a $4-8\,$GHz bandwidth by a cryogenic amplifier as well as a room temperature amplifier. The quadrature voltages are acquired by heterodyne detection followed by digital demodulation.}
\end{figure}
To observe subharmonic oscillations we use a frequency-tunable coplanar waveguide microwave resonators\cite{Sandberg2008,Wallquist2006,PalaciosLaloy2008}. The resonator is capacitively coupled to a $50\,\Omega$ transmission line in one end, and grounded via a superconducting quantum interference device (SQUID) in the other end, see Fig.~\ref{fig:SampleAndSetup}. In practice, the SQUID acts as a tunable nonlinear inductance controlled by the magnetic flux $\Phi$ threading its loop and the current $I_s$, $L_J(\Phi_{},I_s)=\Phi_0/\left(2\pi|\cos(\pi\Phi_{}/\Phi_0)|\sqrt{I_c^2-I_s^2}\right)$, where $\Phi_0=h/2e$ is the magnetic flux quantum, and $I_s$ and $I_c$ denote the current flowing through the device and its critical current, respectively. The Josephson nonlinearity of the SQUID induces, at weak excitation, a Duffing-Kerr nonlinearity in the resonator.
The samples are fabricated using standard processes: Josephson junctions are deposited by two-angle evaporation of aluminum. The rest of the circuit is etched out of a sputtered layer of niobium on a sapphire wafer. A micrograph of one chip with four resonators of different lengths is shown in Fig.~\ref{fig:SampleAndSetup}(a). The chip is anchored to the mixing chamber stage of a dilution refrigerator with a base temperature of $12\,$mK.
The measurement setup is sketched in Fig.~\ref{fig:SampleAndSetup}(c). For static magnetic flux biasing of the SQUID we use a superconducting coil mounted close to the sample box. The SQUID nonlinearity can be modulated by applying a microwave signal as an external drive of the current through the SQUID $I_s$ or by flux modulation. We focus mainly on the external driving. The resonator output signal is amplified by a low-noise cryogenic amplifier at the $3\,$K stage, as well as an additional amplifier at room temperature, before being sampled by a digitizer. To maintain phase coherence, a $10\,$MHz signal is used to lock the signal generator and the digitizer together. The digitizer downconverts both the in- and out-of-phase quadratures, $I(t)$ and $Q(t)$, with a local oscillator before digitizing the data at an effective sampling rate $f_s$ during a time $t_s$. From the individual quadratures the total output power after amplification can be calculated as $P_{out}=\langle I^2\rangle+\langle Q^2\rangle$.
\begin{table*}
\caption{\label{tab:resonators} Resonator and SQUID parameters: $d$, resonator length; $I_c$, SQUID critical current; $C_J$, SQUID capacitance; $\gamma_0=L_J^0/L_0d$, the inductive participation ratio; $L_0$ and $C_0$, inductance and capacitance per unit length; $\omega_1(0)$, resonant frequency of the first mode at $\Phi=0$; $3\omega_1(0)-\omega_2(0)$, spectrum anharmonicity; $2\Gamma_1(0)$, damping of the first resonator mode; $Q_{1,int}$ and $Q_{1,ext}$, quality factors of the fundamental mode.}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$d$ & $I_c$ & $C_J$ & $\gamma_0$ & $L_0$ & $C_0$ & $\omega_1(0)/2\pi$ & $(3\omega_1(0)-\omega_2(0))/2\pi$ & $2\Gamma_1(0)/2\pi$ & $Q_{1,int}(0)$ & $Q_{1,ext}(0)$\\
$[\mu$m] & [$\mu$A] & [fF] & $[\%]$ & [$\mu$H/m] & [nF/m] & [GHz] & [MHz] &[MHz] & [$10^3$] & [$10^3$]\\ \hline
5080 & $1.90$ & $86.1$ & $7.7$ & $0.44$ & $0.16$ & $5.504$ & 136 & 0.38 & 61.1 & 19 \\ \hline
\end{tabular}
\end{center}
\end{table*}
We directly probe the first resonator mode in a reflection measurement. Higher modes can only be measured indirectly due to the 4-8$\,$GHz bandwidth limitation of our setup. To detect the second mode we use parametric up-conversion \cite{ZakkaBajjani2011,Wustmann2017}: we modulate the magnetic flux penetrating the SQUID loop at the difference frequency of the first and the second modes, $\omega_2-\omega_1$, while simultaneously applying a weak drive tone at $\omega_1$. The flux pump converts photons from the first to the second mode, resulting in an avoided crossing, as shown in Fig.~\ref{fig:BSexample}, from which we can determine the difference frequency.
\begin{figure}
\center
\includegraphics{BeamsplittingExample_a.pdf}
\caption{\label{fig:BSexample}
Determination of the second mode frequency via parametric up-conversion. A weak probe tone (horizontal axis), near the resonant frequency of the first mode $\omega_1$, is up-converted by the flux pump tone (vertical axis), near the difference frequency $\omega_2-\omega_1$. The process results in an avoided level crossing visible in the reflected probe tone. The data presented shows the reflected phase response for a magnetic flux bias, $\Phi=0.17\,\Phi_0$, where $\omega_1/2\pi$ = 5.432$\,$GHz and $(\omega_2 - \omega_1)/2\pi$ = 10.70$\,$GHz, yielding $\omega_2/2\pi$ = 16.132$\,$GHz.
}
\end{figure}
\begin{figure}
\center
\includegraphics{DCtuningFitsAndAnharm_a.pdf}
\caption{\label{fig:DCtuning}(a) Spectroscopy of resonator modes 1 and 2, fitted using \Eq{spectrum}. Extracted resonator and SQUID parameters are found in Table \ref{tab:resonators}. Inset: anharmonicity of the spectrum, $3\omega_1-\omega_2$. (b) An illustration of the frequency spectrum. Since the spectrum is non-equidistant the drive signal around $3\omega$ falls slightly above mode 2 while the measurement frequency $\omega$ is slightly below mode 1.}
\end{figure}
In Fig.~\ref{fig:DCtuning}(a) we present the extracted frequencies of the lowest mode for our sample together with a fit to the spectral dispersion, \Eq{spectrum}. The extracted parameters are presented in Table~\ref{tab:resonators}. The resonator is overcoupled and have a narrow bandwidth, less than $1\,$MHz.
\section{Subharmonic oscillations}
\subsection{Observations}
We observe period-tripling subharmonic oscillations at $\omega \approx \omega_1$, by applying an external signal at $3\omega$. This effect has been observed in several samples though in the paper we only present data for one sample. As can be seen in Fig.~\ref{fig:Region}(a), we observe subharmonic oscillations in a range of red detuning, $\delta_1=\omega-\omega_1<0$, and for signal generator drive powers above $\sim-13\,$dBm. This range of drive powers and detunings forms region II, where the subharmonic oscillations are visible. Above and below, in regions I and III, respectively, no subharmonic oscillations are observed.
\begin{figure*}
\center
\includegraphics{SubharmonicRegion_a.pdf}
\caption{\label{fig:Region}
(a) Intensity of the subharmonic oscillation output signal as function of drive power and detuning, for $\Phi_{}=0$. The oscillations are detected in region II. The white line corresponds to the maximum output signal for each measurement frequency; the white star at the end of the white line represents the subharmonic oscillation frequency threshold. The dashed white line that separates regions I and II corresponds to the theoretical boundary of existence for the subharmonic oscillations. In region III, the oscillations, although existing as a solution to Eqs.~\eqref{EOMtext}, are not visible because the oscillator switches to the ground state. (b) The dots represent three linecuts from (a), at detuning values indicated by the white arrows; a theory fit to these three data traces is represented by the black lines. (c-f) Histograms of the detected radiation quadrature voltages from the operating points indicated by the white circles in (a). These histograms reveal three dynamic states: (c) the ground state, \textit{i.e.} the background noise level; (d) the ground state (in the middle) and the three excited states with equal amplitudes and with phases differing by $2\pi/3$; (e) and (f) the three excited states. All histograms are sampled at a rate of $f_s=100\,$kHz. The features in (f) are connected by faint lines, an artifact from the averaging that indicates that the switching rate at this point is higher than in (d) and (e), and not negligible compared to $f_s$. These histograms can be compared to the histograms sampled with $f_s=10\,$kHz that are presented in Fig.~\ref{fig:fs10kHz}, and were measured with parameters as indicated by the white squares in (a).}
\end{figure*}
We also investigated the quadratures of the oscillations using histograms of the $I(t)$ and $ Q(t)$ signals. In Fig.~\ref{fig:Region}(c) we show a background histogram (peaked at $ I= Q =0$) illustrating the system noise level. This histogram represents the system ground state. At higher drive power, inside region II, the histograms feature three well-defined stable states forming a regular triangle, see Fig.~\ref{fig:Region}(e) and (f). At the low-power edge of region II the system shows four states, see Fig.~\ref{fig:Region}(d). This observation is in full agreement with the phase portrait of the subharmonic oscillator, Fig.~\ref{phaseportrait}, featuring four coexisting stable steady states: the silent ground state and the three excited states.
The system switching rates between the states are different for different operating points, $\delta_1$ and $P_d$. Analysis of the underlying time-domain data yields a switching rate of $1\,$kHz in Fig.~\ref{fig:Region}(d) and $15\,$kHz in (f). In (d) and (e) the histograms show clearly separated states while in (f) the states are connected by faint lines. These lines indicate enhancement of the stochastic switching between the steady states. When the system switching rate becomes comparable to the sampling rate, $f_{sw}\sim f_s$, the states are averaged together.
Lowering the sampling rate to $10\,$kHz, makes the switching processes between the stationary states more visible, see Fig.~\ref{fig:fs10kHz}. In region II the transitions occur only between the excited states forming a triangle configuration, while at the border of regions II and III the transitions connect the ground and excited states forming a star configuration.
\begin{figure}
\center
\includegraphics{SwitchingRate10kHz_a.pdf}
\caption{\label{fig:fs10kHz} Histograms sampled with a rate $f_s=10\,$kHz, ten times slower than in Fig.~\ref{fig:Region}(c-f). Here the lines between the steady states are clearly seen. (a) The histogram measured well inside region II reveals a triangle configuration of transitions between the excited state features. (b) The histogram measured at the border between region II and III exhibits a ``star" configuration of transitions between the ground and excited states. These histograms are measured at $\delta_1/2\pi=1.5\,$MHz and drive power as indicated by the white square markers in Fig.~\ref{fig:Region}(a).
}
\end{figure}
\subsection{Theory}
To explain the experimental observations and establish a basis for quantitative comparison, we perform a theoretical analysis based on the theory for two-mode resonant dynamics in a frequency-tunable resonator\cite{Wustmann2013,Wustmann2017}. The two-mode equations for slowly varying Heisenberg operators of the coupled modes, $a_1$ and $a_2$, in the doubly rotating frame with frequencies $\omega$ and $3\omega$, have the form
\begin{eqnarray}\label{EOMtext}
&&i\dot a_1 + (\delta_1 + i\Gamma_1 + \alpha_1 a_1^\dag a_1 + 2\alpha \,a_2^\dag a_2) a_1 + \tilde\alpha \,
a_1^{\dag 2} a_2 = 0
\nonumber\\
&& i\dot a_2 + (\delta_2 + i\Gamma_2 + \alpha_2\, a_2^\dag a_2 + 2\alpha\, a_1^\dag a_1 ) a_2 + {\tilde\alpha\over 3}
\,a_1^{3} \nonumber\\
&&\hspace{5cm} = \sqrt{2\Gamma_{2,ext}} \, B_2 \,.
\end{eqnarray}
Here, the amplitude of the fundamental mode, $a_1$, describes the subharmonic oscillator, while that of the second mode, $a_2$, acts as an effective parametric pump. $B_2$ is the complex amplitude of the external drive:
$\Gamma_n$ is the mode damping. Explicit equations for the external damping, $\Gamma_{2,ext}$, and the Kerr coefficients, $\alpha_n$, are presented in \Eqs{Gammas}-(\ref{alphas}). The cross-Kerr coefficients are related to the Kerr coefficients,
$\alpha=\sqrt{\alpha_1\alpha_2}$ and $\tilde\alpha=\sqrt[4]{\alpha_1^3\alpha_2}$.
Eqs. \eqref{EOMtext} are associated with and can be derive from the quantum Hamiltonian,
\begin{eqnarray}\label{QHamiltonian}
&& {H}/\hbar = -\sum_{n=1,2} \left(\delta_n {a}_n^\dag {a}_n + {\alpha_n\over 2}{a}_n^{\dag 2} {a}_n^2 \right) - 2\alpha \,{a}_1^\dag {a}_1 {a}_2^\dag {a}_2 \nonumber\\
&& - {\tilde\alpha\over 3}\left({a}_1^{\dag 3}{a}_2 + {a}_1^{ 3}{a}_2^\dag \right) + \sqrt{2\Gamma_{2,ext}}\left(B_2{a}_2^{\dag} + B_2^* {a}_2\right).
\end{eqnarray}
Subharmonic oscillations are essentially a classical phenomenon since a large number of photons is generated in the resonator. Thus we restrict the analysis to the quasiclassical solutions of \Eq{EOMtext}, neglecting quantum effects. A detailed description of the
eigenfunctions and tunneling rates can be found in Ref.~\cite{Zhang2017} for the Hamiltonian \eqref{QHamiltonian} in the single-mode case.
The trivial quasiclassical solution to \Eq{EOMtext}, $a_1=0$, describes a silent oscillator state. It is always stable, see Appendix, \Eq{stability1}. The nontrivial solutions describing stable steady states of the excited oscillator consist of a phase-degenerate triad, the states being stable within the region of existence, \Eq{lambda}. In terms of a polar parametrization of the quasiclassical field amplitudes,
%
\begin{eqnarray}\label{parametrization}
a_1 = r_1e^{i\phi_1}, \quad a_2 = {r_2\over \beta} e^{i\phi_2}, \quad \theta = 3\phi_1- \phi_2 \,,
\end{eqnarray}
with $\beta = \sqrt[4]{\alpha_2/ \alpha_1}$,
the stable solution has the form,
\begin{eqnarray}\label{r1text}
r_1^2 = {|\delta_1|\over\alpha_1} - {3r_2^2\over 2} + \sqrt{ {r_2^2|\delta_1|\over\alpha_1} - {7 r_2^4\over4} -
{\Gamma_1^2\over \alpha_1^2}}\,,
\end{eqnarray}
\begin{eqnarray}\label{sincos_text}
\sin\theta ={\Gamma_1 \over \alpha_1 r_2r_1}, \quad
\theta \in ( \pi/2,\; \pi) \;{\rm mod} (2\pi) \,.
\end{eqnarray}
The solution \eqref{r1text} exists within an interval of the effective pump intensity, $r_{2\pm}^2$,
\begin{eqnarray}\label{rpm_text}
r_2^2 \in (r_{2-}^2 , \; r_{2+}^2), \quad r_{2\pm}^2 = {2|\delta_1| \over 7\alpha_1} \left[1 \pm \sqrt {1 - {7 \Gamma_1^2\over \delta_1^2}} \,\right] ,
\end{eqnarray}
and at negative, red detuning from the fundamental resonator mode,
%
\begin{eqnarray}\label{threshold_text}
\delta_1 \;\leq \; - \sqrt7 \,{\Gamma_1 }.
\end{eqnarray}
The solution \eqref{r1text} is finite at the boundaries of existence \eqref{rpm_text}, \textit{i.e.}, the subharmonic oscillations emerge abruptly when the boundaries are crossed. The oscillations achieve a maximum intensity,
\begin{eqnarray}\label{r1max_text}
r_{1,max}^2 = {4\over 7\alpha_1} \left(|\delta_1|+ \sqrt{|\delta_1| ^2 - 7\Gamma_1^2}\right)\,,
\end{eqnarray}
that grows linearly with the detuning far from the threshold, $|\delta_1|\gg \Gamma_1$. The maximum is achieved in this region at $r_2^2 = |\delta_1|/ 14\alpha_1$.
The effective pump strength, $r_2$, is defined by a nonlinear response to the external drive $B_2$, \Eq{EOMclass2}. The response exhibits instability at a weak drive, $|B_{2}|^2 \lesssim \beta^2|\delta_1|^3/(18\Gamma_{2,ext}\alpha_1)$, as shown in Figs. \ref{r2_vs_delta} and \ref{r2_vs_B}, but has a regular monostable behavior at larger drive, up to the maximum value given by \Eq{B2max},
\begin{flalign}\label{B2max_text}
\!\!|B_{2,max}|^2 \approx {(3\omega_1-\omega_2)^2 |\delta_1|\over 7\beta^2\Gamma_{2,ext}\alpha_1}, \; \delta_1\gg \Gamma_1, \Phi < 0.4\Phi_0\,.
\end{flalign}
The phase $\phi_2$ is defined by the phase of the drive, $\phi_B$, and for the stable branch and large detuning, $\delta_1 \gg \Gamma_{2}$, it is approximately $\pi$-shifted from the latter (see Appendix).
This situation persists within a wide interval of magnetic flux bias, $0<\Phi<0.4\Phi_0$, as long as the anharmonicity of the resonator spectrum exceeds the detuning, $3\omega_1-\omega_2 \gg \delta_1$, see the inset in Fig~\ref{fig:DCtuning}.
\subsection{Analysis}
\begin{figure*}
\center
\includegraphics{RegionInfo_a.pdf}
\caption{\label{fig:RegionInfo}(a) Growth of the intensity of subharmonic oscillations with red detuning for five different flux bias values. (b) Slopes of the data traces in (a) plotted versus magnetic flux (blue circles); the red line is the theory prediction, \Eqs{r1max_text} and \eqref{maxima}. (c) Threshold detunings for subharmonic oscillations at different magnetic flux values; blue dots - experimental data, red line - theory, see comment in the text.
}
\end{figure*}
Using the outlined theoretical results we are able to quantitatively analyze the details in Fig.~\ref{fig:Region}. The (b)-panel displays three linecuts of the subharmonic oscillation region taken at three different values of detuning. The oscillation amplitudes, represented by dots, show sharp onsets at the low-power edge of region II and smoother decays towards the high-power edge. The solid lines correspond to a theoretical fit. The power of the external drive is in linear units
\begin{eqnarray}\label{Pd}
P_d = 3\hbar\omega\, |B_2|^2 10^{Att/10}\,,
\end{eqnarray}
where $Att$ is the attenuation (in dB) between the generator and the resonator for the $3\omega$ drive signal and $|B_2|^2$ is given by Eq.~\eqref{B2r2simple}. The measured output power is
%
\begin{eqnarray}\label{Pout}
P_{out}=\hbar\omega |a_1|^2 2\Gamma_{1,ext} 10^{G/10}.
\end{eqnarray}
Here $G$ denotes the overall gain of the measurement signal $\omega$ between the resonator and digitizer. The relation between the amplitudes of mode one and two is given in Eq.~\eqref{r1text}. The fit is done by adjusting only one fitting parameter, $X = Q_{2,ext}10^{Att/10}$, which is found in Eqs.~\eqref{Pd} and \eqref{B2r2simple}, where $Q_{2,ext}=\omega_2/2\Gamma_{2,ext}$. The other parameters are measured independently and listed in Table \ref{tab:resonators}; the Kerr coefficient $\alpha_1/2\pi= 85\,$kHz assumes the theory value, and the gain is estimated to be $G=66\pm 0.5\,$dB. The best fit is achieved for $X= (9.97 \pm 0.03) \cdot 10^{11}$. From this we can calculate the photon population of the first resonator mode, $|a_1|^2$. It is found that $2.4\,$nW output power in Fig.~\ref{fig:Region}(a) corresponds to roughly 100 photons.
As seen in Fig.~\ref{fig:Region}(b) the observed oscillations reach a maximum and disappear (the oscillator switches to the ground state) before they reach the theoretical maximum, \Eq{r1max_text}. A comparison of the experimental and theoretical maxima reveals the scaling
\begin{eqnarray}\label{maxima}
|a_{1,max}|^2_{exp} = 0.7 \, |a_{1,max}|^2\,.
\end{eqnarray}
Using \Eq{B2max} and fitting parameters extracted from Fig.~\ref{fig:Region}(b), we evaluate the boundaries of existence and the stability of the subharmonic oscillations. The upper boundary is presented by the dashed white line in Fig.~\ref{fig:Region}(a). Above the dashed line, in region I, the oscillations do not exist; below this line the theory predicts the existence of oscillations and stability within the whole region II+III (the oscillation lower boundary, \Eq{rpm_text}, lies far below the edge of the panel). However, the oscillations are only visible in the narrow region II but not in III. This can be explained by a competition between the excited states and the stable ground state. At the boundary between regions II and III, the system explores all four available states, as indicated by Fig.~\ref{fig:Region}(d), and in region III the system preferentially stays in the ground state, see Fig.~\ref{fig:Region}(c). Quantitative evaluation of the lower boundary of visibility of the subharmonic oscillations requires a dynamical analysis including the effect of noise, which goes beyond the scope of the present study.
In Fig.~\ref{fig:Region} all data is taken at zero magnetic flux, $\Phi = 0$. However, the subharmonic oscillations are detected also at nonzero flux up to $\Phi\approx 0.4\Phi_0$.
In Fig.~\ref{fig:RegionInfo}(a) we present the maximum output power as a function of detuning for different flux bias values. At $\Phi=0$ this corresponds to the white solid line in Fig.~\ref{fig:Region}(a). The output power is proportional to the maximum population of the first resonator mode and grows linearly with the detuning in good agreement with the theory, \Eq{r1max_text}. Furthermore, the flux dependence of the line slopes in Fig.~\ref{fig:RegionInfo}(a) is in nice agreement with the theory prediction given by the flux dependence of the Kerr coefficient \eqref{alphas}, in \Eq{r1max_text}, and making use of the scaling, \Eq{maxima}, as illustrated in Fig.~\ref{fig:RegionInfo}(b).
The subharmonic oscillations are predicted to start at a threshold at small red detuning, \Eq{threshold_text}. Experimentally, this threshold is defined as the endpoint of the white curve in Fig.~\ref{fig:Region}(a), marked with a white star. Experimental data for the frequency thresholds at different $\Phi$ is presented in Fig.~\ref{fig:RegionInfo}(c) (blue dots). For smaller flux values, $|\Phi| \lesssim 0.4\Phi_0$, the threshold values of the output radiation, \Eq{Pout} evaluated for $\delta_1=-\sqrt{7}\Gamma_1$, exceed the noise level, $P_n \approx 0.44$ nW, and therefore the measurement procedure identifies the true threshold, Eq.~\eqref{threshold_text}. However, at the edges of this region, $|\Phi| \approx 0.4\Phi_0$, the output power rapidly decreases, as indicated in Fig.~\ref{fig:Region}(a)-(b), and therefore the visible oscillation threshold shifts to larger detuning. Quantitatively, the shifted position of the threshold is defined by $P_{out}(|a_{1,max}|^2) = P_n$. We compute the solution to this equation using the parameters in Table~\ref{tab:resonators}, and the flux dependence of $\Gamma_{1,ext}$ in \Eq{Gammas}. The internal losses and noise power are assumed flux independent. The theory plot of this solution, the red line in Fig.~\ref{fig:RegionInfo}(c), excellently reproduces the data. We note that no fitting parameters were used in Fig.~\ref{fig:RegionInfo}(b) and (c).
\section{Conclusion}
We have observed period-tripling subharmonic oscillations in a driven nonlinear multimode microwave resonator in the quantum regime. When an external drive tone is applied at a frequency $3\omega$, we observe output oscillations at $\omega$, demonstrating period tripling. The output signal consists of three correlated beams having the same amplitudes but with their phases shifted by $2\pi/3$ radians with respect to each other. The oscillations are observed at red detuning from the resonator fundamental mode, and in a finite interval of drive power. Due to the proximity of the second resonator mode to the drive tone, the downconversion efficiency is strongly enhanced, enabling access to the subharmonic oscillation regime. A theory for the two-mode subharmonic resonance was developed to explain the observations. The theoretical predictions are in good quantitative agreement with the experimental observations regarding the boundary of existence of oscillations, maximum output power, and frequency threshold.
Our successful implementation of an intermode interaction of the $a_1^{\dagger 3}a_2$ type may in the future be used to create multi-photon entanglement and multi-component macroscopic cat states \cite{Puri2017}.
\section{Acknowledgments}
We thank Waltraut Wustmann, Mark Dykman and G\"oran Johansson for useful discussions. We gratefully acknowledge financial support from the European Research Council, the European project PROMISCE, the Swedish research council, and the Wallenberg Foundation. J.B. acknowledges partial support by the EU under REA grant agreement no. CIG-618353.
|
\section{Introduction}
\IEEEpubidadjcol
\IEEEPARstart{F}{innish} and Estonian are agglutinative languages, meaning that
words are formed by concatenating smaller linguistic units, and a great deal of
grammatical information is conveyed by inflection. Modeling these inflected
words correctly is important for automatic speech recognition, to produce
understandable transcripts. Recognizing a suffix correctly can also help to
predict the other words in the sentence. By collecting enough training data, we
can get a good coverage of the words in one form or another---perhaps names and
numbers being an exception---but we are far from having enough training data to
find examples of all the inflected word forms.
\IEEEpubidadjcol
Another common feature of Finnish and Estonian is that the orthography is
phonemic. Consequently, the spelling of a word can be altered according to the
pronunciation changes in conversational language. Especially Finnish
conversations are written down preserving the variation that happens in
colloquial pronunciation \cite{Enarvi:2013}. Modeling such languages as a
sequence of complete word forms becomes difficult, as most of the forms are very
rare. In our data sets, most of the word forms appear only once in the training
data.
Agglutination has a far more limited impact on the vocabulary size in English.
Nevertheless, the vocabularies used in English language have grown as larger
corpora are used and computers are able to store larger language models in
memory. Moreover, as speech technology improves, we start to demand better
recognition of e.g. proper names that do not appear in the training data.
Modern automatic speech recognition (ASR) systems can handle vocabularies as
large as millions of words with simple n-gram language models, but a second
recognition pass with a neural network language model (NNLM) is now necessary
for achieving state-of-the-art performance. Vocabulary size is much more
critical in NNLMs, as neural networks take a long time to train, and training
and inference times depend heavily on the vocabulary size. While computational
efficiency is the most important reason for finding alternatives to word-based
modeling, words may not be the best choice of language modeling unit with regard
to model performance either, especially when modeling agglutinative languages.
Subword models have been successfully used in Finnish ASR for more than a decade
\cite{Hirsimaki:2006}. In addition to reducing the complexity of the language
model, subword models bring the benefit that even words that do not occur in the
training data can be predicted. However, subwords have not been used for
modeling \emph{conversational} Finnish or Estonian before. Our earlier attempts
to use subwords for conversational Finnish ASR failed to improve over word
models. In this paper, we show how subword models can be used in the FST-based
Kaldi speech recognition toolkit and obtain the best results to date by
rescoring subword lattices using subword NNLMs, 27.1~\% WER for spontaneous
Finnish conversations, and 21.9 \% WER for spontaneous Estonian conversations.
This is the first published evaluation of subwords in conversational Finnish and
Estonian speech recognition tasks.
Our conclusions are slightly different from those earlier published on standard
Finnish and Estonian tasks, where n-gram models based on statistical morphs have
provided a large improvement to speech recognition accuracy
\cite{Siivola:2003,Hirsimaki:2006,Kurimo:2006}. An important reason is that we
are able to use very large vocabularies (around two million words) in the
word-based n-gram models. Recently it has been noticed that the gap between
subword and word models becomes quite small when such a large word vocabulary is
used \cite{Varjokallio:2016}. In our conversational Finnish and Estonian
experiments, word and subword n-gram models performed quite similarly in terms
of evaluation set word error rate. Our new observation is that neural networks
are especially beneficial for modeling subwords---subword NNLMs are clearly
better than word NNLMs trained using the full vocabulary.
Another approach for very large vocabulary speech recognition
is using word classes in the language models. We evaluate different
algorithms for clustering words into classes. Recent comparisons have shown an
advantage in perplexity for the exchange algorithm over Brown clustering, while
clusterings created from distributed word representations have not worked as
well \cite{Botros:2015,Dehdari:2016,Song:2017}. We present additionally a novel
rule-based algorithm that clusters colloquial Finnish word forms, and also
evaluate word error rate. Surprisingly, class-based n-gram models perform better
than word models in terms of perplexity and speech recognition accuracy in
conversational Finnish and Estonian.
Word classes and subword units are especially attractive in NNLMs, because the
vocabulary size has a great impact on the memory consumption and computational
complexity. The size of the input layer projection matrix and the output layer
weight matrix, as well as the time required to normalize the output
probabilities using softmax, have a linear dependency on the vocabulary size.
The output normalization can also be made more efficient by using one of the
several methods that try to approximate the full softmax, either by modifying
the network structure or the training objective. So far the only comparison of
these approximations for large-vocabulary NNLMs that we are aware of is in
\cite{Chen:2016}. They found hierarchical softmax to perform best in terms of
perplexity with a vocabulary of 800,000 words and a feedforward network.
We compare hierarchical softmax, sampling-based softmax, class-based models, and
subword models in speech recognition on languages that are known for very large
vocabularies. Both data sets contain around two million unique word forms. In
our experiments where the training time was limited to 15 days, class-based
NNLMs clearly exceeded the performance of word-based NNLMs in terms of
perplexity and recognition accuracy. The best results were from subword models.
In the Estonian task, the best subword vocabularies were quite large, and the
best result was from a class-based subword model. We also test two methods for
weighting separate language modeling data sets: weighted sampling, which has
already been introduced in \cite{Schwenk:2005} and update weighting, which is a
novel method.
All the neural network language modeling techniques presented in this paper have
been implemented in the open-source toolkit TheanoLM \cite{Enarvi:2016}, which
we hope to lower the threshold of using neural network language models in speech
recognition research.\footnote{\url{https://github.com/senarvi/theanolm}} We
implemented hierarchical softmax \cite{Goodman:2001}, noise-contrastive
estimation \cite{Gutmann:2010}, and BlackOut \cite{Ji:2016} training criteria,
and a lattice decoder that takes advantage of parallel computation using a GPU.
We use a fairly complex recurrent model consisting of an LSTM layer and a
highway network to obtain state-of-the-art results, and run the experiments on
a high-end GPU. Our experiments show that class and subword models are more
attractive than word models for several reasons. They are efficient
computationally and in terms of memory consumption, and they can achieve better
performance than word models. Usually subword vocabularies include all the
individual letters, meaning that any word that uses the same letters can be
constructed. Class models are restricted to a certain vocabulary, but the
efficiency is not limited by the vocabulary size, so very large vocabularies can
be used.
To summarize, this is the first time Finnish and Estonian subword models have
outperformed word models in conversational speech recognition, even without
limiting the word vocabulary size. We compare word clustering techniques and
show that class-based models outperform full-vocabulary word models in these
tasks. We also present the first comparison of word, class, and subword NNLMs
trained using different softmax approximations, applied to speech recognition.
Finally, we test a novel method for weighting NNLM training corpora.
\section{Class-Based Language Models}
Finnish and Estonian are highly agglutinative languages, so the number of
different word forms that appear in training corpora is huge. The pronunciation
variation in colloquial Finnish is also written down, making it very difficult
to reliably estimate the probability of the rare words in new contexts. If we
can cluster word forms into classes based on in which contexts they appear, we
can get more reliable estimates for the class n-gram probabilities. In a
class-based language model, the probability of a word within its class is
usually modeled simply as the unigram probability of the word in the training
data \cite{Brown:1992}:
\begin{equation}
\label{eq:class-ngram}
\begin{split}
&P(w_t \mid w_{t-n+1} \ldots w_{t-1}) = \\
&P(c(w_t) \mid c(w_{t-n+1}) \ldots c(w_{t-1})) P(w_t \mid c(w_t)),
\end{split}
\end{equation}
\noindent where $c(w)$ is a function that maps a word to a class. This is also
the model that we use in this article.
\subsection{Statistical Methods for Clustering Words into Classes}
A common cost function for learning the word classes is the perplexity of a
class bigram model, which is equivalent to using the log probability objective:
\begin{equation}
\label{eq:class-bigram}
\mathcal{L} = \sum_t [\log P(c(w_t) \mid c(w_{t-1})) + \log P(w_t \mid c(w_t))]
\end{equation}
Finding the optimal clustering is computationally very challenging. Evaluating
the cost involves summation over all adjacent classes in the training data
\cite{Brown:1992}. The algorithms that have been proposed are suboptimal.
Another approach that can be taken is to use knowledge about the language to
group words that have a similar function.
Brown et al. \cite{Brown:1992} start by assigning each word to a distinct class,
and then merge classes in a greedy fashion. A naive algorithm would evaluate the
objective function for each pair of classes. One iteration of the naive
algorithm would on average run in $\mathcal{O}(N_V^4)$ time, where $N_V$ is the
size of the vocabulary. This involves a lot of redundant computation that can be
eliminated by storing some statistics between iterations, reducing the time
required to run one iteration to $\mathcal{O}(N_V^2)$.
To further reduce the computational complexity, they propose an approximation
where, at any given iteration, only a subset of the vocabulary is considered.
Starting from the most frequent words, $N_C$ words are assigned to distinct
classes. On each iteration, the next word is considered for merging to one of
the classes. The running time of one iteration is $\mathcal{O}(N_C^2)$. The
algorithm stops after $N_V - N_C$ iterations, and results in all the words being
in one of the $N_C$ classes.
The exchange algorithm proposed by Kneser and Ney \cite{Kneser:1993} starts from
some initial clustering that assigns every word to one of $N_C$ classes. The
algorithm iterates through all the words in the vocabulary, and evaluates how
much the objective function would change by moving the word to each class. If
there are moves that would improve the objective function, the word is moved to
the class that provides the largest improvement.
By storing word and class bigram statistics, the evaluation of the objective
function can be done in $\mathcal{O}(N_C)$, and thus one word iterated in
$\mathcal{O}(N_C^2)$ time \cite{Martin:1995}. The number of words that will be
iterated is not limited by a fixed bound. Even though we did not perform the
experiments in such a way that we could get a fair comparison of the training
times, we noticed that our exchange implementation needed less time to converge
than what the Brown clustering needed to finish.\footnote{We are using a
multithreaded exchange implementation and stop the training when the cost stops
decreasing. Our observation that an optimized exchange implementation can be
faster than Brown clustering is in line with an earlier comparison
\cite{Botros:2015}.}
These algorithms perform a lot of computation of statistics and evaluations over
pairs of adjacent classes and words. In practice the running times are better
than the worst case estimates, because all classes and words do not follow each
other. The algorithms can also be parallelized using multiple CPUs, on the
expense of memory requirements. Parallelization using a GPU would be difficult,
because that would involve sparse matrices.
The exchange algorithm is greedy so the order in which the words are iterated
may affect the result. The initialization may also affect whether the
optimization will get stuck in a local optimum, and how fast it will converge.
We use the exchange\footnote{\url{https://github.com/aalto-speech/exchange}}
tool, which by default initializes the classes by sorting the words by frequency
and assigning word $w_i$ to class $i \mod N_C$, where $i$ is the sorted index.
We compare this to initialization from other clustering methods.
\subsection{Clustering Based on Distributed Representation of Words}
Neural networks that process words need to represent them using real-valued
vectors. The networks learn the word embeddings automatically. These
\textit{distributed representations} are interesting on their own, because the
network tends to learn similar representation for semantically similar
words \cite{Mikolov:2013:NAACL}. An interesting alternative to statistical
clustering of words is to cluster words based on their vector representations
using traditional clustering methods.
Distributed word representations can be created quickly using shallow networks,
such as the Continuous Bag-of-Words (CBOW) model \cite{Mikolov:2013:ICLR}. We
use word2vec\footnote{\url{https://code.google.com/archive/p/word2vec/}}
to cluster words by creating word embeddings using the CBOW model and clustering
them using k-means.
\subsection{A Rule-Based Method for Clustering Finnish Words}
Much of the vocabulary in conversational Finnish text is due to colloquial
Finnish pronunciations being written down phonetically. There are often several
ways to write the same word depending on how colloquial the writing style is.
Phonological processes such as reductions (``miksi'' $\rightarrow$ ``miks''
[\textit{why}]) and even word-internal sandhi (``menenpä'' $\rightarrow$
``menempä'' [\textit{I will go}]) are often visible in written form. Intuitively
grouping these different phonetic representations of the same word together
would provide a good clustering. While the extent to which a text is colloquial
does provide some clues for predicting the next word, in many cases these word
forms serve exactly the same function.
This is closely related to normalization of imperfect text, a task which is
common in all areas of language technology. Traditionally text normalization is
based on hand-crafted rules and lookup tables. In the case that annotated data is
available, supervised methods can be used for example to expand abbreviations
\cite{Sproat:2001}. When annotations are not available, candidate expansions for
a non-standard word can be found by comparing its lexical or phonemic form to
standard words \cite{Han:2011}. The correct expansion often depends on the
context. A language model can be incorporated to disambiguate between the
alternative candidates when normalizing text. We are aware of one earlier work
where colloquial Finnish has been translated to standard Finnish using both
rule-based normalization and statistical machine translation
\cite{Listenmaa:2015}.
Two constraints makes our task different from text normalization: A word needs
to be classified in the same way regardless of the context, and a word cannot be
mapped to a word sequence. Our last clustering method, \textit{Rules}, is based
on a set of rules that describe the usual reductions and alternations in
colloquial words. We iterate over a standard Finnish vocabulary and compare the
standard Finnish word with every word in a colloquial Finnish vocabulary. If the
colloquial word appears to be a reduced pronunciation of the standard word,
these words are merged into a class. Because all the words can appear in at most
one class, multiple standard words can be merged into one class, but this is
rare. Thus, there will be only a handful of words in each class. Larger classes
can be created by merging the classes produced by this algorithm using some
other clustering technique.
\section{Subword Language Models}
Subword modeling is another effective technique to reduce vocabulary size. We
use the Morfessor method \cite{Creutz:2002,Creutz:2007:1}, which has been
successfully applied in speech recognition of many agglutinative languages
\cite{Kurimo:2006,Creutz:2007:2}. Morfessor is an unsupervised method that uses
a statistical model to split words into smaller fragments. As these fragments
often resemble the surface forms of morphemes, the smallest information-bearing
units of a language, we will use the term ``morph'' for them.
Morfessor has three components: a model, a cost function, and the training and
decoding algorithm. The model consists of a lexicon and a grammar. The lexicon
contains the properties of the morphs, such as their written forms and
frequencies. The grammar contains information of how the morphs can be combined
into words. The Morfessor cost function is derived from MAP estimation with the
goal of finding the optimal parameters $\theta$ given the observed training data
$D_W$:
\begin{equation}
\begin{aligned}
\theta_{MAP} &= \operatorname*{arg\,max}_{\theta}P(\theta \mid D_W) \\
&= \operatorname*{arg\,max}_{\theta}P(\theta)P(D_W \mid \theta)
\end{aligned}
\end{equation}
The objective function to be maximized is the logarithm of the product
$P(\theta)P(D_W \mid \theta)$. In a semisupervised setting, it is useful to add
a hyperparameter to control the weight of the data likelihood
\cite{Kohonen:2010}:
\begin{equation}
\mathcal{L}(\theta, D_W) = \log P(\theta) + \alpha \log P(D_W \mid \theta)
\end{equation}
We use the hyperparameter $\alpha$ to control the degree of segmentation in a
heuristic manner. This allows for optimizing the segmentation, either for
optimal perplexity or speech recognition accuracy or to obtain a specific size
lexicon. A greedy search algorithm is used to find the optimal segmentation of
morphs, given the training data. When the best model is found, it is used to
segment the language model training corpus using the Viterbi algorithm.
We apply the Morfessor 2.0
implementation\footnote{\url{https://github.com/aalto-speech/morfessor}} of the
Morfessor Baseline algorithm with the hyperparameter extension
\cite{Virpioja:2013}. In the output segmentation, we prepend and append the
in-word boundaries of the morph surface forms by a ``+'' character. For example
the compound word ``luentokalvoja'' is segmented into ``luento kalvo ja'' and
then transformed to ``luento+ +kalvo+ +ja'' [\textit{lecture+ +slide+ +s}]
before language model training. All four different variants of a subword (e.g.
``kalvo'', ``kalvo+'', ``+kalvo'', and ``+kalvo+'') are treated as separate
tokens in language model training. As high-order n-grams are required to provide
enough context information for subword-based modeling, we use variable-length
n-gram models trained using the VariKN
toolkit\footnote{\url{https://github.com/vsiivola/variKN}} that implements the
Kneser-Ney growing and revised Kneser pruning algorithms \cite{Siivola:2007}.
In the speech recognition framework based on weighted finite-state transducers
(FSTs), we restrict the lexicon FST in such a way that only legal sequences
(meaning that a morph can start with ``+'' if and only if the previous morph
ends with a ``+'') are allowed \cite{Smit:2017}. After decoding the ASR results,
the morphs are joined together to form words for scoring.
\section{Neural Network Language Models}
Recurrent neural networks are known to work well for modeling language, as they
can capture the long-term dependencies neglected by n-gram models
\cite{Mikolov:2010}. Especially the subword-based approach should benefit from
this capability of modeling long contexts. In this article we experiment with
language models that are based on LSTMs and highway networks.
These are layer types that use \textit{sigmoid gates} to control information
flow. The gates are optimized along with the rest of the neural network
parameters, and learn to pass the relevant activations over long distances.
LSTM \cite{Hochreiter:1997} is a recurrent layer type. Each gate can be seen as
an RNN layer with two weight matrices, $W$ and $U$, a bias vector $b$, and
sigmoid activation. The output of a gate at time step $t$ is
\begin{equation}
g(x_t, h_{t-1}) = \sigma(W x_t + U h_{t-1} + b),
\end{equation}
\noindent where $x_t$ is the output vector of the previous layer and $h_{t-1}$
is the LSTM layer state vector from the previous time step. When a signal is
multiplied by the output of a sigmoid gate, the system learns to discard
unimportant elements of the vector depending on the gate's input.
An LSTM layer uses three gates to select what information to pass from the
previous time step to the next time step unmodified, and what information to
modify. The same idea can be used to select what information to pass to the next
layer. Highway networks \cite{Srivastava:2015} use gates to facilitate
information flow across many layers. At its simplest, only one gate is needed.
In the feedforward case, there is only one input, $x_t$, and the gate needs only
one weight matrix, $W_\sigma$. The gate learns to select between the layer's
input and its activation:
\begin{equation}
\begin{aligned}
\label{eq:highway-network}
g(x_t) &= \sigma(W_\sigma x_t + b_\sigma) \\
y_t &= g(x_t) \odot \tanh(W x_t + b) + (1 - g(x_t)) \odot x_t
\end{aligned}
\end{equation}
While LSTM helps propagation of activations and gradients in recurrent networks,
deep networks benefit from highway connections. We did not notice much
improvement by stacking multiple LSTM layers on top of each other. While we did
not have the possibility to systematically explore different network
architectures, one LSTM layer followed by a highway network seemed to perform
well. The architecture used in this article is depicted in Figure
\ref{fig:network-architecture}. Every layer was followed by Dropout
\cite{Srivastava:2014} at rate $0.2$.
\begin{figure}[t]
\centering
\begin{tikzpicture}[
font=\sffamily\scriptsize,
every matrix/.style={ampersand replacement=\&,column sep=7mm,row sep=6mm},
input/.style={draw,thick,circle,fill=blue!20},
output/.style={inner sep=.1cm},
layer/.style={draw,thick,fill=yellow!20,inner sep=.2cm},
arrow/.style={->,>=stealth,shorten >=1pt,semithick},
arrow/.style={->,>=stealth,shorten >=1pt,semithick},
every node/.style={align=center}]
\matrix {
\node[output] (y0) {$P(w_1 \mid w_0)$}; \&
\node[output] (y1) {$P(w_2 \mid w_1, w_0)$}; \&
\node[output] (y2) {$P(w_3 \mid w_2, w_1, w_0)$}; \&
\node[coordinate] (y3) {}; \\
\node[layer] (output0) {$softmax$}; \&
\node[layer] (output1) {$softmax$}; \&
\node[layer] (output2) {$softmax$}; \&
\node[coordinate] (output3) {}; \\
\node[layer] (hidden40) {highway}; \&
\node[layer] (hidden41) {highway}; \&
\node[layer] (hidden42) {highway}; \&
\node[coordinate] (hidden43) {}; \\
\node[layer] (hidden30) {highway}; \&
\node[layer] (hidden31) {highway}; \&
\node[layer] (hidden32) {highway}; \&
\node[coordinate] (hidden33) {}; \\
\node[layer] (hidden20) {highway}; \&
\node[layer] (hidden21) {highway}; \&
\node[layer] (hidden22) {highway}; \&
\node[coordinate] (hidden23) {}; \\
\node[layer] (hidden10) {highway}; \&
\node[layer] (hidden11) {highway}; \&
\node[layer] (hidden12) {highway}; \&
\node[coordinate] (hidden13) {}; \\
\node[coordinate] (h0) {}; \&
\node[coordinate] (h1) {}; \&
\node[coordinate] (h2) {}; \&
\node[coordinate] (h3) {}; \\
\node[layer] (hidden00) {LSTM}; \&
\node[layer] (hidden01) {LSTM}; \&
\node[layer] (hidden02) {LSTM}; \&
\node[coordinate] (hidden03) {}; \\
\node[layer] (projection0) {projection}; \&
\node[layer] (projection1) {projection}; \&
\node[layer] (projection2) {projection}; \&
\node[coordinate] (projection3) {}; \\
\node[input] (x0) {$w_0$}; \&
\node[input] (x1) {$w_1$}; \&
\node[input] (x2) {$w_2$}; \&
\node[coordinate] (x3) {}; \\
};
\draw[arrow] (x0) -- (projection0);
\draw[arrow] (x1) -- (projection1);
\draw[arrow] (x2) -- (projection2);
\draw[arrow] (projection0) -- (hidden00);
\draw[arrow] (projection1) -- (hidden01);
\draw[arrow] (projection2) -- (hidden02);
\draw[arrow] (hidden00) -- node[midway,above] {$C_0$} (hidden01);
\draw[arrow] (hidden01) -- node[midway,above] {$C_1$} (hidden02);
\draw[arrow] (hidden02) -- node[midway,above] {$C_2$} (hidden03);
\draw[arrow] (h0) -| node[near start,above] {$h_0$} (hidden01.145);
\draw[arrow] (h1) -| node[near start,above] {$h_1$} (hidden02.145);
\draw[arrow] (h2) -- node[midway,above] {$h_2$} (h3);
\draw[arrow] (hidden00) -- (hidden10);
\draw[arrow] (hidden01) -- (hidden11);
\draw[arrow] (hidden02) -- (hidden12);
\draw[arrow] (hidden10) -- (hidden20);
\draw[arrow] (hidden11) -- (hidden21);
\draw[arrow] (hidden12) -- (hidden22);
\draw[arrow] (hidden20) -- (hidden30);
\draw[arrow] (hidden21) -- (hidden31);
\draw[arrow] (hidden22) -- (hidden32);
\draw[arrow] (hidden30) -- (hidden40);
\draw[arrow] (hidden31) -- (hidden41);
\draw[arrow] (hidden32) -- (hidden42);
\draw[arrow] (hidden40) -- (output0);
\draw[arrow] (hidden41) -- (output1);
\draw[arrow] (hidden42) -- (output2);
\draw[arrow] (output0) -- (y0);
\draw[arrow] (output1) -- (y1);
\draw[arrow] (output2) -- (y2);
\end{tikzpicture}
\caption{The recurrent network architecture used in our experiments, unrolled
three time steps. The cell state $C_t$ of the LSTM layer conveys information
over time steps until the gates choose to modify it. One gate selects part of
this information as the output of the layer, $h_t$, which is also passed to the
next time step. A highway network uses a gate to select which parts of the
output are passed to the next layer unmodified and which parts are modified.}
\label{fig:network-architecture}
\end{figure}
The input of the network at time step $t$ is $w_t$, an index that identifies the
vocabulary element. The output contains the predicted probabilities for every
vocabulary element, but only the output corresponding to the target word is
used. The vocabulary can consist of words, word classes, subwords, or subword
classes. The choice of vocabulary does not make any difference with regard to
the neural network training, except that large vocabularies require more memory
and are slower to train.
Usually word vocabularies are limited to a \emph{shortlist} of the most frequent
words. A special token such as \unk can be used in place of any out-of-shortlist
(OOS) words, which is necessary with RNN language models in particular. The NNLM
can be combined with a large-vocabulary n-gram model to obtain a probability for
every training word. The \unk probability represents the total probability mass
of all OOS words, which can be distributed according to n-gram language model
probabilities. An n-gram language model is convenient to integrate with a
feedforward NNLM, which is a particular kind of n-gram model itself, but less
trivial in our RNN decoder. It also becomes computationally demanding to
normalize the resulting probability distribution correctly using a large n-gram
model \cite{Park:2010}.
In our baseline shortlist NNLMs we distribute the OOS probability according to a
unigram model. When rescoring lattices, the output does not have to be a proper
probability distribution. Assuming that the probability mass that the NNLM and
n-gram models allocate to the OOS words are close to each other, a reasonable
approximation is to replace the OOS probabilities with n-gram probabilities
\cite{Park:2010}. We tried this, but did not get good results because the
assumption was too far from the truth. Our class NNLMs are similar to Equation
\ref{eq:class-ngram}, except that RNNs do not fix the context to $n$ previous
words---the length of the history used to predict the next word is limited only
by the mini-batch size.
Memory consumption becomes a problem when using GPUs for training, since current
GPU boards typically have no more than 12 GB of memory. Each layer learns a
weight matrix whose dimensionality is input size by output size. For example, an
NNLM with one hidden layer of size 1,000 and a vocabulary of size 100,000
requires a 1,000 by 100,000 matrix on the input and output layer. Assuming
32-bit floats are used, such a matrix uses 400 MB of memory. In addition,
temporary matrices are needed when propagating the data through the network.
Memory required to store the weight matrices can be reduced by using a small
projection layer and a small layer before the output layer, or factorizing a
large weight into the product of two smaller matrices \cite{Sainath:2013}.
Another possibility is to divide weight matrices to multiple GPUs. The size of
the temporary data depends also on the mini-batch size.
We did the experiments with quite a complex model to see how good speech
recognition accuracy we are able to achieve in these tasks. The projection layer
maps words to 500-dimensional embeddings. Both the LSTM and highway network
layers have 1500 outputs. With vocabularies larger than 100,000 elements we
added a 500-unit feedforward layer before the output layer to reduce memory
consumption. With class models mini-batches were 32 sequences and with other
models 24 sequences of length 25. We also explored the possibility of training
models with larger vocabularies using hierarchical softmax and sampling-based
softmax. These approximations are explained in more detail in the following
sections.
\subsection{Output Normalization}
The final layer normalizes the output to provide a valid probability
distribution over the output classes. Normally the softmax function is used:
\begin{equation}
\label{eq:softmax}
y_{t,i} = \frac{\exp(x_{t,i})}{\sum_j \exp(x_{t,j})}
\end{equation}
At each time step $t$, the cross-entropy cost function requires computing the
conditional probability of the target word only, $P(w_{t+1} \mid w_0 \ldots w_t)
= y_{t,w_{t+1}}$. Still all the activations $x_{t,j}$ are needed to explicitly
normalize the probability distribution. This becomes computationally expensive,
because vocabularies can be very large, and the cost of computing the
normalization term scales linearly with the vocabulary size.
There has been a great deal of research on improving the speed of the softmax
function by various approximations. Hierarchical NNLM is a class-based model
that consists of a neural network that predicts the class and separate neural
networks that predict the word inside a class \cite{Kuo:2012}. This can reduce
training time of feedforward networks considerably, because different n-grams
are used to train each word prediction model. Hierarchical softmax is a single
model that factors the output probabilities into the product of multiple softmax
functions. The idea has originally been used in maximum entropy training
\cite{Goodman:2001}, but exactly the same idea can be applied to neural networks
\cite{Morin:2005}. SOUL combines a shortlist for the most frequent words with
hierarchical softmax for the out-of-shortlist words \cite{Le:2011}. Adaptive
softmax \cite{Grave:2016} is a similar approach that optimizes the word cluster
sizes to minimize computational cost on GPUs.
Another group of methods do not modify the model, but use sampling during
training to approximate the expensive softmax normalization. These methods speed
up training, but use normal softmax during evaluation. Importance sampling is a
Monte Carlo method that samples words from a distribution that should be close
to the network output distribution \cite{Bengio:2003:AISTATS}. Noise-contrastive
estimation (NCE) samples random words, but instead of optimizing the
cross-entropy cost directly, it uses an auxiliary cost that learns to classify a
word as a training word or a noise word \cite{Gutmann:2010}. This allows it to
treat the normalization term as a parameter of the network. BlackOut continues
this line of research, using a stochastic version of softmax that explicitly
discriminates the target word from the noise words \cite{Ji:2016}.
Variance regularization modifies the training objective to encourage the network
to learn an output distribution that is close to a real probability distribution
even without explicit normalization \cite{Shi:2014}. This is useful for example
in one-pass speech recognition, where evaluation speed is important but the
output does not have to be a valid probability distribution. The model can also
be modified to predict the normalization term along with the word probabilities
\cite{Sethy:2015}. NCE objective also encourages the network to learn an
approximately normalized distribution, and can also be used without softmax e.g.
for speech recognition \cite{Chen:2015}.
\subsection{Hierarchical Softmax}
\label{sec:hierarchical-softmax}
Hierarchical softmax factors the output probabilities into the product of
multiple softmax functions. At one extreme, the hierarchy can be a balanced
binary tree that is $log_2(N)$ levels deep, where $N$ is the vocabulary size.
Each level would differentiate between two classes, and in total the
hierarchical softmax would take logarithmic time. \cite{Morin:2005}
We used a two-level hierarchy, because it is simple to implement, and it
does not require a hierarchical clustering of the vocabulary. The first level
performs a softmax between $\sqrt{N}$ word classes and the second level performs
a softmax between $\sqrt{N}$ words inside the correct class:
\begin{equation}
\label{eq:hierarchical-softmax}
\begin{split}
&P(w_t \mid w_0 \ldots w_{t-1}) = \\
&P(c(w_t) \mid w_0 \ldots w_{t-1}) P(w_t \mid w_0 \ldots w_{t-1}, c(w_t))
\end{split}
\end{equation}
\noindent This already reduces the time complexity of the output layer to the
square root of the vocabulary size.
The clustering affects the performance of the resulting model, but it is not
clear what kind of clustering is optimal for this kind of models. In earlier
work, clusterings have been created from word frequencies
\cite{Mikolov:2011:ICASSP}, by clustering distributed word representations
\cite{Mnih:2009}, and using expert knowledge \cite{Morin:2005}.
Ideally all class sizes would be equal, as the matrix product that produces the
preactivations can be computed efficiently on a GPU when the weight matrix is
dense. We use the same word classes in the hierarchical softmax layer that we
use in class-based models, but we force equal class sizes; after running the
clustering algorithm, we sort the vocabulary by class and split it into
partitions of size $\sqrt{N}$. This may split some classes unnecessarily into
two, which is not optimal. On the other hand it is easy to implement and even
as simple methods as frequency binning seem to work \cite{Mikolov:2011:ICASSP}.
An advantage of hierarchical softmax compared to sampling based output layers is
that hierarchical softmax speeds up evaluation as well, while sampling is used
only during training and the output is properly normalized using softmax
during inference.
\subsection{Sampling-Based Approximations of Softmax}
Noise-contrastive estimation \cite{Gutmann:2010} turns the problem from
classification between $N$ words into binary classification. For each training
word, a set of noise words (one in the original paper) is sampled from some
simple distribution. The network learns to discriminate between training words
and noise words. The binary-valued class label $C_w$ is used to indicate whether
the word $w$ is a training or noise word. The authors derive the probability
that an arbitrary word comes from either class, $P(C_w \mid w)$, given the
probability distributions of both classes. The objective function is the
cross entropy of the binary classifier:
\begin{equation}
\begin{aligned}
\mathcal{L} = &\sum_w [C_w \log P(C_w=1 \mid w) \\
&+ (1 - C_w) \log P(C_w=0 \mid w)]
\end{aligned}
\end{equation}
The expensive softmax normalization can be avoided by making the normalization
term a network parameter that is learned along the weights during training. In a
language model, the parameter would be dependent on the context words, but it
turns out that it can be fixed to a context-independent constant without harming
the performance of the resulting model \cite{Mnih:2012}. In the beginning of the
training the cost will be high and the optimization may be unstable, unless the
normalization is close to correct. We use one as the normalization constant and
initialize the output layer bias to the logarithmic unigram distribution, so
that in the beginning the network corresponds to the maximum likelihood unigram
distribution.
BlackOut \cite{Ji:2016} is also based on sampling a set of noise words, and
motivated by the discriminative loss of NCE, but the objective function
directly discriminates between the training word $w_T$ and noise words $w_N$:
\begin{equation}
\mathcal{L} = \sum_{w_T} [\log P(w_T) + \sum_{w_N} \log (1 - P(w_N))]
\end{equation}
\noindent Although not explicitly shown, the probabilities $P(w)$ are
conditioned on the network state. They are computed using a weighted softmax
that is normalized only on the set of training and noise words. In addition to
reducing the computation, this effectively performs regularization in the output
layer similarly to how the Dropout \cite{Srivastava:2014} technique works in the
hidden layers.
Often the noise words are sampled from the uniform distribution, or from the
unigram distribution of the words in the training data \cite{Mnih:2012}. Our
experiments confirmed that the choice of \textit{proposal distribution} is
indeed important. Using uniform distribution, the neural network optimization
will not find as good parameters. With unigram distribution the problem is that
some words may be sampled very rarely. Mikolov et al. \cite{Mikolov:2013:NIPS}
use the unigram distribution raised to the power of $\beta$. Ji et al.
\cite{Ji:2016} make $\beta$ a tunable parameter. They also exclude the correct
target words from the noise distribution.
We used the power distribution with $\beta = 0.5$ for both BlackOut and NCE.
We did not modify the distribution based on the target words, however, as that
would introduce additional memory transfers by the Theano computation library
used by TheanoLM. We observed also that random sampling from a multinomial
distribution in Theano does not work as efficiently as possible with a GPU. We
used 500 noise words, shared across the mini-batch. These values were selected
after noting the speed of convergence with a few values. Small $\beta$ values
flatten the distribution too much and the optimal model is not reached. Higher
values approach the unigram distribution, causing the network to not learn
enough about the rare words. Using more noise words makes mini-batch updates
slower, while using only 100 noise words we noticed that the training was barely
converging.
These methods seem to suffer from some disadvantages. Properly optimizing the
$\beta$ parameter can take a considerable amount of time. A large enough set of
noise words has to be drawn for the training to be stable, diminishing the speed
advantage in our GPU implementation. While we did try a number of different
parameter combinations, BlackOut never finished training on these data sets
without numerical errors.
\subsection{Decoding Lattices with RNN Language Models}
While improving training speed is the motivation behind the various softmax
approximations, inference is also slow on large networks. Methods that modify
the network structure, such as hierarchical softmax, improve inference speed as
well. Nevertheless, using an RNN language model in the first pass of
large-vocabulary speech recognition is unrealistic. It is possible to create a
list of $n$ best hypothesis, or a word lattice, during the first pass, and
rescore them using an NNLM in a second pass. We have implemented a word lattice
decoder in TheanoLM that produces better results than rescoring n-best lists.
Conceptually, the decoder propagates tokens through the lattice. Each token
stores a network state and the probability of the partial path. At first one
token is created at the start node with the initial network state. The algorithm
iterates by propagating tokens to the outgoing links of a node, creating new
copies of the tokens for each link. Evaluating a single word probability at a
time would be inefficient, so the decoder combines the state from all the tokens
in the node into a matrix, and the input words into another matrix. Then the
network is used to simultaneously compute the probability of the target word in
all of these contexts.
Rescoring a word lattice using an RNN language model is equivalent to rescoring
a huge n-best list, unless some approximation is used to limit the dependency of
a probability on the earlier context. We apply three types of pruning, before
propagation, to the tokens in the node \cite{Sundermeyer:2014:LatticeDecoding}:
\begin{itemize}
\item \textbf{n-gram recombination}. \enspace If there are multiple tokens,
whose last $n$ context words match, keep only the best. We use $n = 22$.
\item \textbf{cardinality pruning}. \enspace Keep at most $c$ best tokens.
We use $c = 62$.
\item \textbf{beam pruning}. \enspace Prune tokens whose probability is low,
compared to the best token. The best token is searched from all nodes that
appear at the same time instance, or in the future. (Tokens in the past have a
higher probability because they correspond to a shorter time period.) We prune
tokens if the difference in log probability is larger than 650.
\end{itemize}
We performed a few tests with different pruning parameters and chose large
enough $n$ and $c$ so that their effect in the results was negligible. Using a
larger beam would have improved the results, but the gain would have been small
compared to the increase in decoding time.
\section{Experiments}
\subsection{Data Sets}
We evaluate the methods on difficult spontaneous Finnish and Estonian
conversations. The data sets were created in a similar manner for both
languages. For training acoustic models we combined spontaneous speech corpora
with other less spontaneous language that benefits acoustic modeling.
For training language models we combined transcribed conversations with web data
that has been filtered to match the conversational speaking style
\cite{Kurimo:2016}.
For the Finnish acoustic models we used 85 hours of training data from
three sources. The first is the complete Finnish SPEECON \cite{Iskra:2002}
corpus. This corpus includes 550 speakers in different noise conditions that all
have read 30 sentences and 30 words, numbers, or dates, and spoken 10
spontaneous sentences. Two smaller data sets of better matching spontaneous
conversations were used: DSPCON \cite{Aalto:2017} corpus, which consists of
short conversations between Aalto University students, and FinDialogue part of
the FinINTAS \cite{Lennes:2009} corpus, which contains longer spontaneous
conversations. For language modeling we used 61,000 words from DSPCON and 76
million words of web data. We did not differentiate between upper and lower
case. This resulted in 2.4 million unique words.
For the Estonian acoustic models we used 164 hours of training data,
including 142 hours of broadcast conversations, news, and lectures collected at
Tallinn University of Technology \cite{Meister:2012}, and 23 hours of
spontaneous conversations collected at the University of Tartu\footnote{Phonetic
Corpus of Estonian Spontaneous Speech. For information on distribution, see
\url{http://www.keel.ut.ee/et/foneetikakorpus}.}. These transcripts contain 1.3
million words. For language modeling we used additionally 82 million words
of web data. The language model training data contained 1.8 million unique
words, differentiating between upper and lower case. One reason why the Estonian
vocabulary is smaller than the Finnish vocabulary, even though the Estonian data
set is larger, is that colloquial Estonian is written in a more systematic way.
Also standard Estonian vocabulary is smaller than standard Finnish vocabulary
\cite{Creutz:2007:2}, probably because standard Finnish uses more inflected word
forms.
\begin{table}[t]
\caption{{\it Out-of-vocabulary word rates (\%) of the evaluation sets,
excluding start and end of sentence tokens. The last row is the full training
set vocabulary, which applies also for the class models.}}
\label{tab:oov-rates}
\begin{center}
\begin{tabular}{lcc}
\hline
Vocabulary Size & Finnish & Estonian \\
\hline
100,000 & 6.67 & 3.89 \\
500,000 & 3.36 & 1.59 \\
2.4M (Fin) / 1.8M (Est) & 2.31 & 1.01 \\
\hline
\end{tabular}
\end{center}
\end{table}
We use only spontaneous conversations as development and evaluation data.
As mentioned earlier, Finnish words can be written down in as many different
ways as they can be pronounced in colloquial speech. When calculating Finnish
word error rates we accept the different forms of the same word as
correct, as long as they could be used in the particular context. Compound words
are accepted even if they are written as separate words. However, we compute
perplexities on transcripts that contain the phonetically verbatim word forms,
excluding out-of-vocabulary (OOV) words. The perplexities from n-gram and neural
network word and class models are all comparable to one another, because they
model the same vocabulary consisting of all the training set words. Subwords can
model also unseen words, so the perplexities in subword experiments are higher.
OOV word rates of the evaluation sets are reported in Table \ref{tab:oov-rates}
for different vocabulary sizes.
The Estonian web data is the filtered data from \cite{Kurimo:2016}. The same
transcribed data is also used, except that we removed from the acoustic
training set three speakers that appear in the evaluation set. The evaluation
data is still 1236 sentences or 2.9 hours. The Finnish data is what we used in
\cite{Enarvi:2016}, augmented with 2016 data of DSPCON and read speech from
SPEECON. While we now have more than doubled the amount of acoustic training
data, we have only a few more hours of spontaneous conversations. The switch to
neural network acoustic models had a far greater impact on the results than the
additional training data. We still use the same Finnish evaluation set of 541
sentences or 44 minutes. The Finnish development and evaluation sets and
reference transcripts that contain the alternative forms are included in the
latest DSPCON release, without a few sentences that we could not license.
\subsection{Models}
The word based n-gram models were 4-grams, trained using the Modified Kneser-Ney
implementation of SRILM toolkit \cite{Stolcke:2002}. Class-based models did not
use Kneser-Ney smoothing, because the class n-gram statistics were not suitable
for computing the Modified Kneser-Ney discount parameters. The quality of our
web data is very different from the transcribed conversations, and simply
pooling all the training data together would cause the larger web data to
dominate the model. Instead we created separate models from different data sets,
and combined them by interpolating the probabilities of the observed n-grams
from the component models using weights that were optimized on the development
data. In the Finnish task we created a mixture from two models, a web data model
and a transcribed data model. In the Estonian task we created a mixture from
three models, separating the transcribed spontaneous conversations from the
broadcast conversations.
The mixture weights were optimized independently for each language model on the
development data, using expectation maximization (EM). In the Finnish
experiments this gave the transcribed data a weight slightly less than 0.5. In
the Estonian experiments the weights of the spontaneous conversations and the
web data were typically around 0.4, while the broadcasts were given a weight
less than 0.2. Morph models were similarly combined from component models, but
the EM optimization failed to give good weights. We used initially those
optimized for the word-based models, and after the other parameters were fixed,
we optimized the mixture weights for development set perplexity using a grid
search with steps of $0.05$.
The word clustering algorithms do not support training data weighting, so we
simply concatenated the data sets. There are many parameters that can be
tweaked when creating distributed word representations with word2vec. We tried
clustering words using a few different parameters, and report only the best
n-gram model for each class vocabulary size. Within the set of values that we
tried, the best performance was obtained with continuous bag of words (CBOW),
window size 8, and layer size 300 to 500.
For the subword language models, we trained Morfessor on a word list combined
from all training corpora; the difference to other options such as token-based
training was negligible. For each language, four segmentations were trained with
$\alpha$-values 0.05, 0.2, 0.5, and 1.0. This resulted in respective vocabulary
sizes of 42.5k, 133k, 265k, and 468k for Finnish, and 33.2k, 103k, 212k, and
403k for Estonian. The sizes include the different morph variants with ``+''
prefix and affix. When training the subword n-gram models with the VariKN
toolkit, the growing threshold was optimized on the development set, while
keeping the pruning threshold twice as large as the growing threshold.
Word-based neural network models were trained on two shortlist sizes: 100k and
500k words. With 500k words we added a normal 500-unit layer with hyperbolic
tangent activation before the output layer, which reduced memory consumption and
speeded up training. The neural networks were trained using Adagrad
\cite{Duchi:2011} optimizer until convergence or until the maximum time limit of
15 days was reached. All neural network models were trained on a single NVIDIA
Tesla K80 GPU and the training times were recorded.
We tried two different approaches for weighting the different data sets during
neural network training: by randomly sampling a subset of every data set in the
beginning of each epoch \cite{Schwenk:2005}, and by weighting the parameter
updates depending on from which corpus each sentence comes from. In the latter
approach, the gradient is scaled by both the learning rate and a constant that
is larger for higher-quality data, before updating the parameters. We are not
aware that this kind of update weighting would have been used before.
Optimizing the weights for neural network training is more difficult than
for the n-gram mixture models. As we do not have a computational method for
optimizing the weights, we tried a few values, observing the development set
perplexity during training. Sampling 20 \% of the web data on each iteration, or
weighting the web data by a factor of 0.4 seemed to work reasonably well. We
used a slightly higher learning rate when weighting the web data to compensate
for the fact that the updates are smaller on average. More systematic tests were
performed using these weights with the five vocabularies in Table
\ref{tab:subset-processing-results}.
It would be possible to train separate neural network models from each data set,
but there are no methods for merging several neural networks in the similar
fashion that we combine the n-gram models. Often the best possible NNLM results
are obtained by interpolating probabilities from multiple models, but that kind
of system is cumbersome in practice, requiring multiple models to be trained and
used for inference. The layer sizes and other parameters would have to be
optimized for each model separately.
We combined the NNLMs with the nonclass word or subword n-gram model by
log-linear interpolation. We did not notice much difference to linear
interpolation, so we chose to do the interpolation in logarithmic space, because
the word probabilities may be smaller than what can be represented using 64-bit
floats. We noticed that optimization of the interpolation parameters was quite
difficult with our development data, so we gave equal weight to both models. In
some cases it could have been beneficial to give a larger weight to the neural
network model. Development data was used to select a weight for combining
language model and acoustic scores from four different values.
\subsection{Speech Recognition System}
We use the Kaldi \cite{Povey:2011} speech recognition system for training our
acoustic models and for first-pass decoding. The TDNN acoustic models were
trained on a pure sequence criterion using Maximum Mutual Information (MMI)
\cite{Povey:2016}. The data sets were cleaned and filtered using a Gaussian
Mixture Model recognizer and augmented through speed and volume perturbation
\cite{Ko:2015}. The number of layers and parameters of the TDNN were optimized
to maximize development set accuracy on the word model.
First-pass decoding was very fast with real-time factor less than $0.5$. The
accuracy of the first-pass recognition exceeded our earlier results on both data
sets \cite{Enarvi:2016,Kurimo:2016}, due to the new neural network acoustic
models.
Kaldi does not, at this moment, directly support class-based decoding. Instead
we created lattices using regular n-gram models, and rescored them with class
n-gram and neural network models. Using the methods described in
\cite{Allauzen:2003} it is possible to construct a word FST that represents the
probabilities of a class-based language model and use this in first-pass
decoding or rescoring, without explicitly using the classes. Kaldi does not have
any restrictions on vocabulary size, but compiling the FST-based decoding graph
from the language model, pronunciation dictionary, context dependency
information, and HMM structure did consume around 60 GB of memory. The memory
requirement can be lowered by reducing the number of contexts in the first-pass
n-gram model.
\subsection{Results}
\begin{table}[t]
\caption{{\it Results from word, class, and subword n-gram language models on
development data. Includes perplexity, word error rate (\%), and word error
rate after interpolation with the nonclass model. Exchange algorithm is
initialized using classes created with the other algorithms, in addition to the
default initialization. Perplexities from word and subword models are not
comparable.}}
\label{tab:ngram-dev-results}
\begin{center}
\begin{tabular}{lccc}
\hline
Classes & Perplexity & WER & +Nonclass \\
\hline
\noalign{\vskip 1mm}
\multicolumn{4}{l}{\textbf{Finnish, 2.4M words}} \\
Nonclass & 736 & 30.5 & \\
Brown 2k & 705 & 29.9 & 29.0 \\
CBOW 2k & 1017 & 33.2 & 30.2 \\
Exchange 2k & 698 & 29.7 & 29.1 \\
Brown+Exchange 2k & 701 & 30.0 & 29.0 \\
CBOW+Exchange 2k & 695 & 29.8 & 29.3 \\
Rules+Exchange 2k & 700 & \bf 29.3 & \bf 28.9 \\
Brown 5k & 694 & 29.9 & 29.3 \\
CBOW 5k & 861 & 31.4 & 29.8 \\
Exchange 5k & \bf 683 & 29.9 & 29.3 \\
Brown+Exchange 5k & 688 & 29.8 & 29.2 \\
CBOW+Exchange 5k & 684 & 29.8 & 29.3 \\
Rules+Exchange 5k & 688 & 29.8 & 29.4 \\
CBOW 10k & 801 & 31.1 & 29.9 \\
Exchange 10k & 691 & 29.9 & 29.4 \\
CBOW+Exchange 10k & 691 & 29.9 & 29.5 \\
Rules+Exchange 10k & 690 & 29.9 & 29.4 \\[1mm]
\multicolumn{4}{l}{\textbf{Finnish, 42.5k subwords}} \\
Nonclass & 1127 & 29.9 & \\
Exchange 5k & 1433 & 31.2 & 29.8 \\[1mm]
\multicolumn{4}{l}{\textbf{Finnish, 133k subwords}} \\
Nonclass & 1135 & 30.1 & \\
Exchange 5k & 1412 & 31.4 & 30.2 \\[1mm]
\multicolumn{4}{l}{\textbf{Finnish, 265k subwords}} \\
Nonclass & 1128 & 30.2 & \\
Exchange 5k & 1334 & 30.6 & 29.2 \\[1mm]
\multicolumn{4}{l}{\textbf{Finnish, 468k subwords}} \\
Nonclass & 1100 & 30.0 & \\
Exchange 5k & 1252 & 30.2 & 29.1 \\[1mm]
\hline
\noalign{\vskip 1mm}
\multicolumn{4}{l}{\textbf{Estonian, 1.8M words}} \\
Nonclass & 447 & 23.4 & \\
Brown 2k & 438 & 22.8 & \bf 22.5 \\
Exchange 2k & 439 & 22.9 & 22.6 \\
Brown+Exchange 2k & 438 & \bf 22.6 & \bf 22.5 \\
Brown 5k & 432 & 22.7 & \bf 22.5 \\
Exchange 5k & 432 & 23.0 & 22.6 \\
Brown+Exchange 5k & \bf 430 & 22.8 & 22.7 \\[1mm]
\multicolumn{4}{l}{\textbf{Estonian, 33.2k subwords}} \\
Nonclass & 591 & 23.4 & \\
Exchange 5k & 707 & 24.3 & 23.9 \\[1mm]
\multicolumn{4}{l}{\textbf{Estonian, 103k subwords}} \\
Nonclass & 582 & 23.7 & \\
Exchange 5k & 689 & 24.1 & 23.5 \\[1mm]
\multicolumn{4}{l}{\textbf{Estonian, 212k subwords}} \\
Nonclass & 577 & 23.1 & \\
Exchange 5k & 659 & 23.6 & 23.2 \\[1mm]
\multicolumn{4}{l}{\textbf{Estonian, 403k subwords}} \\
Nonclass & 582 & 23.4 & \\
Exchange 5k & 644 & 23.4 & 23.0 \\[1mm]
\hline
\end{tabular}
\end{center}
\end{table}
Table \ref{tab:ngram-dev-results} lists perplexities and word error rates given
by n-gram language models on the development data. The baseline word model
performance, 30.5 \% on Finnish and 23,4 \% on Estonian can be compared to the
various word class and subword models. We have also included results from
subword classes created using the exchange algorithm for reference. The word
error rates were obtained by rescoring lattices that were created using the
nonclass word or subword model.
In the Finnish task, 2,000, 5,000, and 10,000 class vocabularies were compared.
The worst case running times of the exchange and Brown algorithms have quadratic
dependency on the number of classes. With 10,000 classes, even using 20 CPUs
Brown did not finish in 20 days, so the experiment was not continued. The
exchange algorithm can be stopped any time, so there is no upper limit on the
number of classes that can be trained, but the quality of the clustering may
suffer if it is stopped early. However, with 5,000 classes and using only 5
CPUs, it seemed to converge in 5 days. Increasing the number of threads
increases memory consumption. Training even 40,000 classes was possible in 15
days, but the results did not improve, so they are not reported. The most
promising models were evaluated also in the Estonian task.
CBOW is clearly the fastest algorithm, which is probably why it has gained some
popularity. These results show, however, that the clusters formed by k-means
from distributed word representations are not good for n-gram language models.
CBOW does improve, compared to the other clusterings, when the number of classes
is increased. Other than that, the differences between different classification
methods are mostly insignificant, but class-based models outperform word models
on both languages. This result suggests that class-based softmax may be a viable
alternative to other softmax approximations in neural networks. The performance
of the Estonian subword models is close to that of the word model, and the
Finnish subword models are better than the word model. Subword classes do not
work as well, but the difference to nonclass subword models gets smaller
when the size of the subword vocabulary increases.
Mostly the differences between the different initializations of the exchange
algorithm seemed insignificant. However, our rule-based clustering algorithm
followed by running the exchange algorithm to create 2,000 classes
(Rules+Exchange 2k) gave the best word error rate on Finnish. In the NNLM
experiments we did not explore with different clusterings, but used the ones
that gave the smallest development set perplexity in the n-gram experiments. For
Finnish the 5,000 classes created using the exchange algorithm was selected
(Exchange 5k). On Estonian, initialization using Brown classes gave slightly
better perplexity (Brown+Exchange 5k) and was selected for neural network
models.
\begin{table}[t]
\caption{{\it Comparison of uniform data processing, random sampling of web
data by 20~\%, and weighted parameter updates from web data by a factor of 0.4,
in NNLM training. The models were trained using normal softmax. Includes
development set perplexity, word error rate (\%), and word error rate after
interpolation with the n-gram model.}}
\label{tab:subset-processing-results}
\begin{center}
\begin{tabular}{lcccc}
\hline
Subset & Training \\
Processing & Time & Perplexity & WER & +NGram \\
\hline
\noalign{\vskip 1mm}
\multicolumn{5}{l}{\textbf{Finnish, 5k classes}} \\
Uniform & 143 h & 511 & 26.0 & 25.6 \\
Sampling & 128 h & 505 & 26.2 & 25.6 \\
Weighting & 101 h & 521 & 26.4 & 25.5 \\[1mm]
\multicolumn{5}{l}{\textbf{Finnish, 42.5k subwords}} \\
Uniform & 360 h & 679 & 25.2 & \bf 24.6 \\
Sampling & 360 h & 671 & 25.5 & 25.0 \\
Weighting & 360 h & 672 & \bf 25.1 & \bf 24.6 \\[1mm]
\multicolumn{5}{l}{\textbf{Finnish, 468k subwords, 5k classes}} \\
Uniform & 141 h & 790 & 26.0 & 25.0 \\
Sampling & 119 h & 761 & 25.9 & 25.1 \\[1mm]
\hline
\noalign{\vskip 1mm}
\multicolumn{5}{l}{\textbf{Estonian, 5k classes}} \\
Uniform & 86 h & 339 & \bf 19.8 & 19.9 \\
Sampling & 87 h & 311 & 20.2 & 19.9 \\
Weighting & 105 h & 335 & 20.0 & \bf 19.6 \\[1mm]
\multicolumn{5}{l}{\textbf{Estonian, 212k subwords, 5k classes}} \\
Uniform & 187 h & 424 & 20.0 & 19.7 \\
Sampling & 130 h & 397 & 20.0 & 19.8 \\
Weighting & 187 h & 409 & 19.9 & \bf 19.6 \\[1mm]
\hline
\end{tabular}
\end{center}
\end{table}
Table \ref{tab:subset-processing-results} compares training time, perplexity,
and word error rate in NNLM training, when different processing is applied to
the large web data set. \textit{Uniform} means that the web data is processed
just like other data sets, \textit{sampling} means that a subset of web data is
randomly sampled before each epoch, and \textit{weighting} means that the
parameter updates are given a smaller weight when the mini-batch contains web
sentences. Sampling seems to improve perplexity, but not word error rate.
Because sampling usually speeds up training considerably and our computational
resources were limited, the rest of the experiments were done using sampling.
\begin{table}[t]
\caption{{\it NNLM results on development data. Word models were trained using
class and shortlist vocabularies. Subword models were trained using the full
subword vocabulary and on classes created from the subwords. Includes
perplexity, word error rate (\%), and word error rate after interpolation with
the n-gram model. Perplexities from word and subword models are not
comparable.}}
\label{tab:nn-dev-results}
\begin{center}
\begin{tabular}{l@{\hskip 0.3cm}l@{\hskip 0.3cm}c@{\hskip 0.3cm}c@{\hskip 0.3cm}
c@{\hskip 0.3cm}c@{\hskip 0.1cm}c}
\hline
Network & & & Training \\
Output & Vocabulary & Parameters & Time & PPL & WER & +NGram \\
\hline
\noalign{\vskip 1mm}
\multicolumn{7}{l}{\textbf{Finnish, 2.4M words}} \\
HSoftmax & 100k short & 231M & 360 h & 535 & 26.9 & 25.9 \\
NCE & 100k short & 230M & 360 h & 531 & 26.8 & 25.7 \\
HSoftmax & 500k short & 532M & 360 h & 686 & 28.4 & 27.0 \\
Softmax & 5k classes & 40M & 128 h & 505 & 26.2 & 25.6 \\[1mm]
\multicolumn{7}{l}{\textbf{Finnish, 42.5k subwords}} \\
Softmax & 42.5k full & 115M & 360 h & 671 & \bf 25.5 & \bf 25.0 \\
HSoftmax & 42.5k full & 116M & 360 h & 700 & 25.7 & \bf 25.0 \\
Softmax & 5k classes & 40M & 146 h & 857 & 26.2 & 25.5 \\[1mm]
\multicolumn{7}{l}{\textbf{Finnish, 133k subwords}} \\
HSoftmax & 133k full & 164M & 360 h & 742 & 26.5 & 25.4 \\
Softmax & 5k classes & 40M & 190 h & 811 & 26.1 & 25.4 \\[1mm]
\multicolumn{7}{l}{\textbf{Finnish, 265k subwords}} \\
HSoftmax & 265k full & 296M & 360 h & 849 & 27.0 & 25.6 \\
Softmax & 5k classes & 40M & 133 h & 813 & 26.2 & 25.3 \\[1mm]
\multicolumn{7}{l}{\textbf{Finnish, 468k subwords}} \\
HSoftmax & 468k full & 500M & 360 h & 1026 & 28.6 & 26.9 \\
Softmax & 5k classes & 40M & 119 h & 761 & 25.9 & 25.1 \\[1mm]
\hline
\noalign{\vskip 1mm}
\multicolumn{7}{l}{\textbf{Estonian, 1.8M words}} \\
HSoftmax & 100k short & 231M & 360 h & 321 & 20.6 & 19.9 \\
NCE & 100k short & 230M & 142 h & 384 & 22.4 & 21.4 \\
HSoftmax & 500k short & 532M & 360 h & 380 & 21.0 & 20.2 \\
Softmax & 5k classes & 40M & 87 h & 311 & 20.2 & 19.9 \\[1mm]
\multicolumn{7}{l}{\textbf{Estonian, 33.2k subwords}} \\
Softmax & 33.2k full & 97M & 360 h & 357 & 20.4 & 20.2 \\
HSoftmax & 33.2k full & 97M & 293 h & 370 & 20.7 & 20.2 \\
Softmax & 5k classes & 40M & 116 h & 418 & 20.9 & 20.2 \\[1mm]
\multicolumn{7}{l}{\textbf{Estonian, 103k subwords}} \\
HSoftmax & 103k full & 134M & 306 h & 393 & 20.8 & 20.2 \\
Softmax & 5k classes & 40M & 126 h & 410 & 20.5 & 19.9 \\[1mm]
\multicolumn{7}{l}{\textbf{Estonian, 212k subwords}} \\
HSoftmax & 212k full & 243M & 360 h & 411 & 20.9 & 20.2 \\
Softmax & 5k classes & 40M & 130 h & 397 & \bf 20.0 & 19.8 \\[1mm]
\multicolumn{7}{l}{\textbf{Estonian, 403k subwords}} \\
HSoftmax & 403k full & 434M & 360 h & 463 & 21.4 & 20.7 \\
Softmax & 5k classes & 40M & 124 h & 395 & 20.3 & \bf 19.6 \\[1mm]
\hline
\end{tabular}
\end{center}
\end{table}
Table \ref{tab:nn-dev-results} lists perplexities and word error rates
given by neural network models on the development data. The word error rates
were obtained by rescoring the same lattices as in Table
\ref{tab:ngram-dev-results}. The shortlist and word class models can predict all
training set words, so the perplexities can be compared. Subword models can
predict also new words, so their perplexities cannot be compared with word
models. The percentage of evaluation set words that are not in the shortlist and
words that are not in the training set can be found in Table
\ref{tab:oov-rates}.
The class-based models were clearly the fastest to converge, 5 to 8 days on
Finnish data and 4 to 6 days on Estonian data. The experiments include
shortlists of 100k and 500k words. Other shortlist models, except the Estonian
100k-word NCE, did not finish before the 360 hour limit. Consequently,
improvement was not seen from using a larger 500k-word shortlist.
Our NCE implementation required more GPU memory than hierarchical softmax and we
were unable to run it with the larger shortlist. With the smaller shortlist NCE
was better on Finnish and hierarchical softmax was better on Estonian. We
experienced issues with numerical stability using NCE with subwords, and decided
to use only hierarchical softmax in the subword experiments. BlackOut training
was slightly faster than NCE, but even less stable, and we were unable to finish
the training without numerical errors. With hierarchical softmax we used the
same classes that were used in the class-based models, but the classes were
rearranged to have equal sizes as described in Section
\ref{sec:hierarchical-softmax}. This kind of class arrangement did not seem to
improve from simple frequency binning, however.
In terms of word error rate and perplexity, class-based word models performed
somewhat better than the shortlist models. The best results were from subword
models. On both languages it can be seen that class-based subword models improve
compared to the nonclass subword models when the vocabulary size grows. In the
Finnish task, the smallest 42.5k-subword vocabulary worked well, which is small
enough to use normal softmax without classes. In the Estonian task, larger
subword vocabularies performed better, provided that the subwords were clustered
into classes. The best result was obtained by clustering 403k subwords into
5,000 classes using the exchange algorithm.
\begin{table}[t]
\caption{{\it Performance of best n-gram and NNLM models on evaluation data. All
NNLMs except the shortlist models were using softmax output. Includes
perplexity, word error rate (\%), and word error rate after interpolation with
the nonclass or n-gram model. Perplexities from word and subword models are not
comparable.}}
\label{tab:results-summary}
\begin{center}
\begin{tabular}{lcc@{\hskip 0.1cm}ccc@{\hskip 0.0cm}c}
\hline
& \multicolumn{3}{@{}c@{}}{N-Gram} & \multicolumn{3}{@{}c}{NNLM} \\
Vocabulary & PPL & WER & +Nonclass & PPL & WER & +NGram \\
\hline
\noalign{\vskip 1mm}
\textbf{Finnish} & \multicolumn{3}{@{}c@{}}{Full 2.4M} & \multicolumn{3}{@{}c}{Shortlist 100k (NCE)} \\
Word & 785 & 31.7 & & 618 & 28.1 & 27.9 \\[1mm]
& \multicolumn{3}{@{}c@{}}{Rules+Exchange 2k} & \multicolumn{3}{@{}c}{Exchange 5k} \\
Class & 760 & \bf 31.4 & \bf 31.1 & 589 & 29.2 & 27.9 \\[1mm]
& \multicolumn{3}{@{}c@{}}{Morfessor 42.5k} & \multicolumn{3}{@{}c}{Morfessor 42.5k} \\
Subword & 1313 & 31.7 & & 846 & \bf 27.3 & \bf 27.1 \\[1mm]
& \multicolumn{3}{@{}c@{}}{Morfessor 468k} & \multicolumn{3}{@{}c}{Morfessor 468k} \\
Subword Class & 1499 & 32.1 & 31.3 & 942 & 28.2 & 27.4 \\[1mm]
\hline
\noalign{\vskip 1mm}
\textbf{Estonian} & \multicolumn{3}{@{}c@{}}{Full 1.8M} & \multicolumn{3}{@{}c}{Shortlist 100k (HSoftmax)} \\
Word & 483 & 26.1 & & 344 & 23.1 & 22.6 \\[1mm]
& \multicolumn{3}{@{}c@{}}{Brown+Exchange 2k} & \multicolumn{3}{@{}c}{Brown+Exchange 5k} \\
Class & 465 & \bf 25.3 & \bf 25.2 & 324 & 22.2 & 22.2 \\[1mm]
& \multicolumn{3}{@{}c@{}}{Morfessor 212k} & \multicolumn{3}{@{}c}{Morfessor 33.2k} \\
Subword & 628 & 26.0 & & 377 & 22.7 & 22.6 \\[1mm]
& \multicolumn{3}{@{}c@{}}{Morfessor 403k} & \multicolumn{3}{@{}c}{Morfessor 403k} \\
Subword Class & 682 & 25.8 & 25.5 & 403 & \bf 22.1 & \bf 21.9 \\[1mm]
\hline
\end{tabular}
\end{center}
\end{table}
Table \ref{tab:results-summary} compares the best models on evaluation data. The
best word, class, subword, and subword class n-gram and NNLM models were
selected based on development set word error rate. The evaluation set results
show that the advantage that the Finnish subword n-gram models had on the
development set was due to the optimization of the morph segmentations on
development data. Word classes are the best choice for n-gram modeling. However,
neural networks seem to benefit especially subword modeling, because the overall
best results are from subword NNLMs. Classes work well also in NNLMs, although
the best Finnish shortlist model, 100k-word NCE, performed exceptionally well in
speech recognition. Interpolation with the n-gram model gives a small but
consistent improvement.
\section{Conclusions}
Our experiments show that class-based models are very attractive for
conversational Finnish and Estonian speech recognition. When the vocabulary
contains millions of words, class-based n-gram models perform better than normal
word models. A class-based NNLM can be trained in less than a week and when the
training time is limited, often performs better than word-shortlist models.
In previous work, class-based models did not outperform word-based models in
recognizing standard Finnish and Estonian, such as broadcast news
\cite{Varjokallio:2016}. Improvement was made only when interpolating word and
class models. One reason why word classes are especially beneficial in the
conversational tasks may be that in the absence of large conversational corpora,
most of our training data is from the Internet. Web data is noisy and there are
many ways to write the same word.
One would expect less to be gained from using subword models, when a word model
is trained from full vocabulary of millions of words. This seems to be the case,
but RNNs are good at learning the structure of the language from a text that has
been segmented into subwords. Subwords can also solve the vocabulary size
problem with neural network models. In the Finnish task, the best results were
from an NNLM trained on a relatively small 42.5k-subword vocabulary with full
softmax output. In the Estonian task, the best results are from a large
403k-subword vocabulary that was clustered into 5,000 classes.
We explored the possibility of using NCE, BlackOut, or hierarchical softmax to
overcome the problem of training neural networks with large output
dimensionality. Generally they were slower than class-based training, and did
not converge to as good a model in the 15-day time constraint, but Finnish
100k-word NCE training gave good results on the evaluation set. The mixed
results could mean that some details have been overlooked in our implementation
of sampling-based softmax.
In both tasks we obtained the best word error rates from a subword NNLM
interpolated with a subword n-gram model. In the Finnish task the best result
was 27.1 \%, which is a 14.5 \% relative improvement from the 31.7 \% WER given
by our baseline 4-gram model. The best result in the Estonian task, 21.9 \%, is
a 16.1 \% relative improvement from our 26.1 \% baseline WER. These are the best
results achieved in these tasks, and better than our previously best results by
a large margin. The best previously published results are 48.4 \% WER in the
Finnish task \cite{Enarvi:2016} and 52.7 \% WER in the Estonian task
\cite{Kurimo:2016}.
The corpus weighting methods that we used in NNLM training showed potential for
improvement, but more thorough research should be done on how to select optimal
weights.
\section{Acknowledgements}
Computational resources were provided by the Aalto Science-IT project.
\bibliographystyle{IEEEtran}
|
\subsection{Quantifying predictive power}
There is little consensus on quantifying a channel's predictive capability.
Common metrics are designed to quantify \emph{memory} rather than
\emph{prediction} \cite{white2004short}, but even when adapted for measuring
prediction, one can choose different types of readout function. We focus on
what we call the \emph{total predictable information}:
\begin{align*}
\I_\text{fut} := I[\mathcal{R};\Future]
~.
\end{align*}
This is the mutual information between present channel state and the input's
future. It is the amount of information that is predictable about the input
future from the present channel state. Due to the feedforward nature of the
channel-input setup---that is, the channel's state does not affect the
input---and the Markov chains given earlier, we have the Markov chain
$\mathcal{R}\rightarrow\St^+\rightarrow\St^-\rightarrow\Future$. As a result: \begin{align*}
\I_\text{fut} = \mathbf{I}[\mathcal{R};\St^-] = \mathbf{I}[\mathcal{R};\MeasSymbol_-,\mathcal{T}_-]
~,
\end{align*}
which decomposes into:
\begin{align*}
\I_\text{fut} = \mathbf{I}[\mathcal{R};\MeasSymbol_-] + \mathbf{I}[\mathcal{R};\mathcal{T}_-|\MeasSymbol_-]
~.
\end{align*}
The term $\mathbf{I}[\mathcal{R};\MeasSymbol_-]$ is called the \emph{instantaneous memory} $\I_\text{mem}$
\cite{Still2012}, since it is the amount of information available from the
channel state about the just-seen input symbol. We find that:
\begin{align*}
\I_\text{fut} = \I_\text{mem} + \mathbf{I}[\mathcal{R};\mathcal{T}_-|\MeasSymbol]
~.
\end{align*}
Thus, the total predictable information is the sum of instantaneous memory and
information that is truly about the future, which here is the time to next
symbol.
The Supplementary Material justifies $\I_\text{fut}$ on generational timescales as a metric via an extension of
Kelly's classic bet-hedging argument \cite{Cove06a}. In a discrete-time
setting, increases in growth rate via increases in sensory information is equal
to the instantaneous predictable information $\mathbf{I}[\mathcal{R}_0;\MeasSymbol_{\Delta t}]$. The
total predictable information $\I_\text{fut}$ is an upper bound on this increase in
growth rate. On ontogenetic timescales, we merely assert that total predictable information might increase concurrently with energetic rewards.
Next, we need to quantify the power consumed by the sensor system. Assuming
access to a temperature-normalized ``energy function'' $\beta E(\meassymbol,r)$,
the temperature-normalized power $\beta P$ is given by:
\begin{align}
\beta P
& = \lim_{\Delta t\rightarrow 0}
\frac{\langle \beta E(\meassymbol_{t+\Delta t},r_t)\rangle -\langle \beta E(\meassymbol_t,r_t) \rangle}{\Delta t}
~.
\label{eq:betaP0}
\end{align}
If determining an energy function is not possible, we can calculate a lower
bound using a continuous-time adaptation of the inequality in Ref.
\cite{Still2012}:
\begin{align}
\dot{\I}_\text{np} := \lim_{\Delta t\rightarrow 0}
\frac{\mathbf{I}[\mathcal{R}_t;\MeasSymbol_t]-\mathbf{I}[\mathcal{R}_t;\MeasSymbol_{t+\Delta t}]}{\Delta t} \leq \beta P
~,
\label{eq:ToP}
\end{align}
with an alternate equivalent definition in Ref. \cite{Horowitz2014}; see the
Supplementary Material.
$\dot{\I}_\text{np}$ is called the \emph{nonpredictive information rate} since it loosely
corresponds to how much of the instantaneous memory is useless for predicting
the next input. Reference \cite{barato2014efficiency} viewed $\dot{\I}_\text{np}/\beta P$
as a learning efficiency. We take the view that $\dot{\I}_\text{np}$ is a potentially
useful lower bound on temperature-normalized power consumption, and use $\I_\text{mem}$ and $\I_\text{fut}$
instead to characterize learning. Any differences between the formulae shown
here and Ref. \cite[Eq. (2)]{Still2012} are superficial; we merely adapted the
derivation for continuous-time processes. Unfortunately, no one has yet given a
guarantee that the nonpredictive information rate is a tight lower bound on
temperature-normalized power.
\paragraph*{Closed-form metrics}
As stated earlier, we wish to find closed-form expressions for $\I_\text{fut}$,
$\I_\text{mem}$, $\dot{\I}_\text{np}$, and $\beta P$ in terms of input
properties---$\phi_{\meassymbol}(\tau)$, $\epsilon^+(g,\meassymbol)$, $p(\meassymbol|g)$---and channel
properties---$M(\meassymbol)$. Their derivations are too lengthy for here and so are
relegated to the Supplementary Materials (SM). The appropriate equations there
are referenced here, where relevant.
Calculating $\I_\text{fut}$ and $\I_\text{mem}$ can be accomplished once $\rho(\st^+,r) =
\Prob(\St^+=\st^+,\mathcal{R}=r)$ is obtained. This follows, in turn, by
manipulating a Chapman-Kolmogorov equation, shown in the SM. Set any ordering
on the pairs $(g,\meassymbol)$; e.g., the ordering
$(g_1,\meassymbol_1),(g_1,\meassymbol_2),\ldots,(g_{|\mathcal{G}|},\meassymbol_{|\MeasAlphabet|})$. An
expression for $p(r|\st^+) = p(r|g,\meassymbol,\tau)$ is given by a combination of
Eqs.~(\ref{eq:5}) and (\ref{eq:init_conds}):
\begin{align*}
p(r|g,\meassymbol,\tau)
= \left(
e^{M(\meassymbol)\tau} \text{eig}_1(\textbf{C})_{(g,\meassymbol)}/\mu_{g} p(g)
\right)_{r}
~,
\end{align*}
where $\textbf{C}$ is a block matrix with entries
$\textbf{C}_{(g,\meassymbol),(g',\meassymbol')} = \delta_{g,\epsilon^+(g',\meassymbol')} p(\meassymbol'|g')
\int_0^{\infty} \phi_{g'}(t) e^{M(\meassymbol')t} dt$. And so,
$\text{eig}_1(\textbf{C})$ is a block vector. Normalization forces
$\vec{1}^{\top} \text{eig}_1(\textbf{C})_{(g,\meassymbol)} = \mu_{g} p(g)$.
We then find the joint probability distribution as $\rho(r,\st^+) =
p(r|\st^+) \rho(\st^+)$, which enables computation of all predictive metrics.
Instantaneous memory is given by $\I_\text{mem} = \mathbf{I}[\MeasSymbol;\mathcal{R}]$, whereas $\I_\text{fut} =
\mathbf{I}[\mathcal{R};\St^-]$. All the relevant distributions---namely, $p(\meassymbol,r)$ and
$\rho(\st^-,r)$---are obtained from the previously derived
$\rho(\st^+,r)$. For instance, to calculate $\rho(\st^-,r)$, we employ the
previously stated Markov chain to find:
\begin{align*}
\rho(\st^-,r) = \sum_{\st^+} \rho(\st^-|\st^+) p(r|\st^+) \rho(\st^+)
~.
\end{align*}
And, to calculate $p(\meassymbol,r)$, we recall that $\st^- = (\meassymbol,\tau_-)$, so we
only need marginalize the joint distribution of $\rho((\meassymbol,\tau_-),r)$.
Calculation of dissipative metrics can be additionally accomplished once:
\begin{align*}
\frac{\delta p(\meassymbol,r)}{\delta t}
& = \lim_{\Delta t\rightarrow 0}
\Big(\Prob(\MeasSymbol_{t+\Delta t}=\meassymbol,\mathcal{R}_t=r) \nonumber \\
& \quad - \Prob(\MeasSymbol_t=\meassymbol,\mathcal{R}_t=r) \Big)/\Delta t
\end{align*}
is obtained. An expression for $\frac{\delta p}{\delta t}$ in terms of input
and channel properties is given in Eq.~(\ref{eq:deltap_deltat1}):
\begin{align*}
\frac{\delta p}{\delta t}
& = \sum_{g',\meassymbol'\neq\meassymbol}
\int d\tau'
~p(\meassymbol|\epsilon^+(g',\meassymbol')) p(\meassymbol'|g') \phi_{g'}(\tau') \nonumber \\
& \quad \times \left(e^{M(\meassymbol')\tau'}
\text{eig}_1(\textbf{C})_{(g',\meassymbol')}\right)_{r} \nonumber \\
& \quad - \sum_{g'} \int d\tau' ~p(\meassymbol|g') \phi_{g'}(\tau')
\Big(e^{M(\meassymbol)\tau'} \text{eig}_1(\textbf{C})_{(g',\meassymbol)}\Big)_{r}
,
\end{align*}
where normalization again requires $\vec{1}^{\top} \text{eig}_1(\textbf{C})_{(g,\meassymbol)} = \mu_{\meassymbol} p(g)$. Then, from earlier, we find that:
\begin{align}
\dot{\I}_\text{np} & = \lim_{\Delta t\rightarrow 0}
\frac{H[\mathcal{R}_t,\MeasSymbol_{t+\Delta t}]-H[\mathcal{R}_t,\MeasSymbol_t]}{\Delta t} \\
& = \lim_{\Delta t\rightarrow 0}
\Big(
\sum_{\meassymbol,r} \left(p(\meassymbol,r)+\frac{\delta p}{\delta t}\Delta t\right)
\log \frac{1}{p(\meassymbol,r)+\frac{\delta p}{\delta t}\Delta t}\nonumber \\
& \qquad - \sum_{\meassymbol,r} p(\meassymbol,r) \log \frac{1}{p(\meassymbol,r)}\Big)/\Delta t \\
& = -\sum_{\meassymbol,r} \frac{\delta p(\meassymbol,r)}{\delta t} \log p(\meassymbol,r)
~.
\label{eq:Inps}
\end{align}
When there are no nondecaying oscillations in $\vec{p}(r,t)$ \cite{schnakenberg1976network}, we can find $\beta P$ despite lacking direct access to an energy function by calculating the steady-state distribution over channel states with fixed input:
\begin{align*}
\vec{p}_{eq}(r|\meassymbol)
= \text{eig}_0(M(\meassymbol)) = \frac{e^{-\beta E(\meassymbol,r)}}{Z_{\beta}(\meassymbol)}
~,
\end{align*}
where the partition function is $Z_{\beta}(\meassymbol) := \sum_{r} e^{-\beta E(\meassymbol,r)}$. Hence:
\begin{align}
\beta E(\meassymbol,r)
= \log \frac{1}{p_{eq}(r|\meassymbol)} - \log Z_{\beta}(\meassymbol)
~.
\label{eq:energy_function}
\end{align}
Recalling Eq.~(\ref{eq:betaP0}) and invoking stationarity---that
$\Prob(\MeasSymbol_t=\meassymbol)=\Prob(\MeasSymbol_{t+\Delta t}=\meassymbol)$---yields:
\begin{align}
\beta P
& = \lim_{\Delta t\rightarrow 0}
\Big( \langle \log\frac{1}{p_{eq}(\meassymbol,r)}
\rangle_{\Prob(\MeasSymbol_{t+\Delta t}=\meassymbol,\mathcal{R}_t=r)} \nonumber \\
& \qquad - \langle \log\frac{1}{p_{eq}(\meassymbol,r)}
\rangle_{\Prob(\MeasSymbol_t=\meassymbol,\mathcal{R}_t=r)}\Big)/\Delta t \\
& = \sum_{\meassymbol,r}
\frac{\delta p(\meassymbol,r)}{\delta t} \log \frac{1}{p_{eq}(r|\meassymbol)}
~.
\label{eq:betaP}
\end{align}
The distributions $\Prob(\MeasSymbol_{t+\Delta t}=\meassymbol,\mathcal{R}_t=r)$ and
$\Prob(\MeasSymbol_t=\meassymbol,\mathcal{R}_t=r)$ can be obtained from $M(\meassymbol)$. In other words,
when there are no recurrent cycles, we can calculate $\beta P$ directly from
the kinetic rates $k_{r\rightarrowr'}(\meassymbol)$ and input generator
($\phi_g(\tau),~\epsilon^+(g,\meassymbol),~p(\meassymbol|g)$) alone.
\paragraph*{Effect of cooperativity on prediction and dissipation}
The Hill molecule is a common fixture in theoretical biology, as it is the
simplest mechanistic model of cooperativity \cite{marzen2013statistical}.
Recall Fig.~\ref{fig:HillMoleculeConception}. A Hill molecule can be in one of
two states, open or closed. When open, $n$ ligand molecules are bound; when
closed, no ligand molecules are bound. Hence, the state of the Hill molecule
carries information about the number of bound ligand molecules. In other words,
such molecules are sensors of their environment.
Let us outline a simple dynamical model of the Hill molecule. The rate of
transition from closed $C$ to open $O$ given a ligand concentration $x$ is:
\begin{align}
k_{C\rightarrow O} = k_O x^n
~.
\label{eq:Hill1}
\end{align}
While the transition rate from open $O$ to closed $C$ is:
\begin{align}
k_{O\rightarrow C} = k_C
~.
\label{eq:Hill2}
\end{align}
The steady-state distribution given fixed ligand concentration is the familiar:
\begin{align*}
\text{Pr}_{\text{eq}} (\mathcal{R}=O|\MeasSymbol=\meassymbol) & = \frac{k_O x^n}{k_C + k_O x^n} \\
& = \frac{x^n}{(k_C/k_O) + x^n}
~.
\end{align*}
Although the mechanistic model makes sense only when $n$ is a nonnegative
integer, this model is often used when $n$ is any nonnegative real number;
increases in $n$ can still be thought of as increases in cooperativity.
Equations~(\ref{eq:Hill1})-(\ref{eq:Hill2}) constitute a complete
characterization of channel properties.
Increasing the cooperativity $n$ increases the steepness of the molecule's ``binding
curve''---the probability of being ``on'' as a function of concentration. In
other words, the sensor becomes more switch-like and less a proportionately
responding transducer of the input. If the concentration is greater than $\left(k_C/k_O\right)^{1/n}$,
the switch is essentially ``on'' if $n$ is high. A more switch-like sensor is
useful if the optimal phenotype depends only upon the condition ``ligand
concentration greater than X''. Whereas, a less switch-like, smoother
responding sensor helps if the optimal phenotype depends on ligand
concentration in a more graded manner.
The concentration scale is set by $\left(k_C/k_O\right)^{1/n}$, while the time
scale is set by $1/k_C$; as such, we set both to $k_O=k_C=1$ without loss of
generality. We imagine that the ligand concentration alternates between two
values: $\meassymbol_l$ and $\meassymbol_h$. When there is less ligand ($\meassymbol_l$), we will tend
to see the Hill molecule revert to and stay in the closed state. When there is
ligand ($\meassymbol_h>\meassymbol_l$), we will tend to see the Hill molecule revert to and
stay in the open state. With no particular application in mind, we imagine
that the dwell-time distributions take the form $\phi_{\meassymbol}(\tau) =
\lambda(\meassymbol)^2\tau e^{-\lambda(\meassymbol)\tau}$ with $\lambda(\meassymbol_l)=5$ and
$\lambda(\meassymbol_h)=4$.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{pred_piness}
\includegraphics[width=.96\columnwidth]{diss_piness}
\caption{(Top) Ligand concentration can take one of two values, $\meassymbol_l = 0.5$
and $\meassymbol_h = 2.0$, in units of $\left(k_C/k_O\right)^{1/n}$. Dwell-time
distributions are parametrized as $\phi_{\meassymbol}(t) = \lambda(\meassymbol)^2 t
e^{-\lambda(\meassymbol)t}$ with $\lambda(\meassymbol_l) = 5,~\lambda(\meassymbol_h)=4$ in units of
$1/k_C$. Hill molecule parameters are set to $k_O=k_C=1$, with varying
cooperativity $n$. (Bottom) prediction-related metrics: instantaneous
memory $I_{mem}$ and total predictive power $I_{fut}$, as functions of $n$
in nats. At right, temperature-normalized dissipation-related metrics:
non-predictive information rate $\dot{I}_{np}$ and temperature-normalized
power $\beta P$, both in nats per unit time.
}
\label{fig:hill}
\end{figure}
We now deploy the earlier formulae to study the predictive capabilities and
dissipative tendencies of a Hill molecule subject to semi-Markov input.
Previous studies of biological sensors found that increases in cooperativity
accompanied increases in channel capacity \cite{tkavcik2009optimizing,
walczak2010optimizing, martins2011trade}. Others studied the thermodynamics of
prediction of cooperative biological sensors \cite{barato2014efficiency,
becker2015optimal}, but did not use the more general class of semi-Markov input
and did not calculate the full suite of metrics here, leaving much sensor
operation untouched.
An example---$\meassymbol_l=0.5,~\meassymbol_h=2.0$ and $k_O=k_C=1.0,~n=2$---illustrates
that roughly $99\%$ of $\I_\text{fut}$ is devoted to instantaneous memory $\I_\text{mem}$ and
roughly $25\%$ of $\beta P$ is devoted to $\dot{\I}_\text{np}$. That is, the inefficiency
in choosing what information to store about the present input contributes
greatly to energetic inefficiency. These results hold qualitatively even when
the dwell-time distributions are log-normal, i.e., are heavier-tailed.
Fig.~\ref{fig:hill} shows that increased cooperativity---that is, increases in
$n$---lead to increases in predictive performance, qualitatively in line with
Ref. \cite{martins2011trade}. Additionally, the larger the cooperativity, the
higher the fraction of $\I_\text{mem}/\I_\text{fut}$. Larger cooperativity, however, leads to
roughly linear increases in the power consumption and the nonpredictive
information rate, whereas increases in predictive power take a more sigmoidal
shape. We therefore might prefer intermediate values of cooperativity (e.g.,
$n\approx 5$) to larger values of cooperativity (e.g., $n\geq 10$). This is
qualitatively similar to the results of Refs. \cite{tkavcik2009optimizing,
walczak2010optimizing}, in that physical constraints can force optimal
information transmission at intermediate levels of cooperativity.
In the absence of a reward function, we assert that energetic rewards are
proportional to $I_{fut}$ \cite{Stil07c, tishby2011information}, so that the
total energy budget is $\alpha \I_\text{fut}-\beta P$. The proportionality constant
$\alpha$ is set by the type of environment in which one finds oneself. Then, the
total energy budget $\alpha I_{fut}-\beta P$ is optimized by the cooperativity
$\widehat{n} = \arg\max_n \left( \alpha I_{fut}-\beta P \right)$. As both
$I_{fut}$ and $\beta P$ increase monotonically with $n$, there is generically
only one such cooperativity $\widehat{n}$. At lower $\alpha$, though, there are
two local maxima of the function of $n$ given by $\alpha I_{fut}-\beta P$, as
shown at Fig. \ref{fig:Efficiency}(Top). Let's pursue the consequences of this
regime dependence.
There are rules for how sensor biochemical parameters adapt to
the present environment. If adaptation rules for cooperativity of a Hill
molecule increase the total energy budget by gradient descent then, for a range
of $\alpha$, we expect that the cooperativity of the Hill molecule to be in
either of the two local maxima of $\alpha \I_\text{fut}-\beta P$ just noted. We assume a separation of timescales---namely, that cooperativity adapts much more slowly than the longest environmental timescale. Figure
\ref{fig:Efficiency}(Bottom) shows the optimal cooperativity $\widehat{n}$ as a
function of conversion factor $\alpha$. As expected from the presence of two
local maxima at lower $\alpha$, there is a discontinuity (supercritical
bifurcation \cite{Stro94a}) in the function of $\alpha$ given by $\arg\max_n
\alpha I_{fut}-\beta P$. Thus, initially, if $\alpha$ (energetic reward for
prediction) increases, the cooperativity discontinuously jumps to a higher
$\widehat{n}$ at a critical $\alpha_h$. From there, if one decreases $\alpha$,
optimal cooperativity slowly decreases, but stays high well below $\alpha_h$,
suddenly decreasing to zero cooperativity at the lower value $\alpha_l$. Thus,
there is a substantial hysteresis loop built into the optimal trade-off between
energy and sensitivity.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Lagrangian2}
\includegraphics[width=\columnwidth]{Optimaln_2}
\caption{(Top) ``Lagrangian'' $\alpha I_{fut} + \beta P$ for several arbitrary
$\alpha$---that represent the energy rewards of $I_{fut}$---as a function of
$n$. Notice that a particular $\alpha$ singles out a particular optimal
$\widehat{n}$.
(Bottom) Optimal cooperativity $\widehat{n}$ as a function of conversion
factor $\alpha$.
Circles (solid blue) are the values of the global maximum $\widehat{n}$
and circles (solid green) are local maxima.
Arrows indicate the directionality in the hysteresis loop:
$\widehat{n}$ jumps up at the upper value $\alpha_h$, if starting at low
$\alpha$, and jumps down at the lower value $\alpha_l$, if starting at high
$\alpha$.
}
\label{fig:Efficiency}
\end{figure}
Recall that in switching circuits hysteresis is essential to adding stability
to a switch's response. Hysteresis stops ``race'' conditions in which the
switch oscillates wildly just as the threshold is passed, amplifying any noise
in the control and internal dynamics. In the Hill molecule, hysteresis is
helpful if a memory of past environmental conditions ($\alpha$) provides
insight into future conditions (future $\alpha$). For example, the environment
might shift $\alpha$ suddenly to being low, but there is a replenishment
mechanism for the available energy that will soon increase $\alpha$ again.
Thus, we see that robustness to environmental noise emerges naturally as the
sensor adapts to and anticipates changing external conditions.
\paragraph*{\textbf{Conclusion}}
We provided closed-form expressions for instantaneous memory, total predictable
information, nonpredictive information rate, and power consumption for a
conditionally Markovian channel subject to unifilar hidden semi-Markov input.
In motivating these metrics for prediction and dissipation on general
timescales, we appealed to an extension of Kelly's classic bet-hedging that
arises from his information-theoretic analysis of trading-off the benefits of
risky, but highly profitable resource investment against the costs of sudden
loss \cite{Cove06a, bergstrom2004shannon}. Here, a sensor faces an analogous
challenge of high sensitivity but at an energy cost that might be suddenly
wasted when the environmental conditions fluctuate. Similar bet-hedging
strategies have been implicated in other biological systems, such as in seed
germination in annual plants \cite{Bulm84a} and bacteriophages \cite{Masl15a}
and in population biology \cite{Cohe66a,Metz92a} and evolution
\cite{bergstrom2004shannon,Sege87a} more generally. The present setting,
though, implicates such strategy optimization in a substantially more
elementary and primitive biological subsystem.
Finally, we used these formulae to calculate the predictive performance and
energetic inefficiency of a simple model of a biological sensor---a Hill
molecule. We found that increases in cooperativity yield increases in both
predictive performance and energy consumption and that the relative balance
between those increases naturally leads to sensor robustness to environmental
fluctuations supported by dynamical hysteresis. Given the Hill molecule's
simplicity as a model sensor, we expect to find hysteresis and the resulting
robustness in more complex biological sensory systems.
The ease with which these various metrics were calculated masks the difficulty
of obtaining the necessary closed-form expressions. (Cf. Supplementary
Materials.) We provided universal estimators for various predictive and
dissipative metrics for conditionally Markovian channels, as unifilar hidden
semi-Markov processes are that general. One practical consequence it that those
wishing to study the relationship between prediction and dissipation need not
simulate arbitrarily long trajectories. Instead, they can validate or
invalidate predictive learning rules and sensor designs using the universal
estimators of these predictive and dissipative metrics. Then they can
efficiently search through parameter space for ``optimal'' (predictive and
energy-efficient) sensors. In addition, given that the theories of random
dynamical systems and of input-dependent dynamical systems are still under
development \cite{arnold2013random}, we believe the formulae presented here
will lead in those domains to a precise generalization of time-scale matching
for nonlinear systems \cite{becker2015optimal}.
\acknowledgments
We thank N. Ay, A. Bell, W. Bialek, S. Dedeo, C. Hillar, I. Nemenman, and S.
Still for useful conversations and the Santa Fe Institute for its hospitality
during visits, where JPC is an External Faculty member. This material is based
upon work supported by, or in part by, the John Templeton Foundation grant
52095, the Foundational Questions Institute grant FQXi-RFP-1609, the U. S.
Army Research Laboratory and the U. S. Army Research Office under contract
W911NF-13-1-0390. S.E.M. was funded by an MIT Physics of Living Systems
Fellowship.
|
\section*{Acknowledgments}
We thank David Duff, KC Ng, Ramakrishna Padmanabhan,
and Paul Saab for their assistance and helpful comments.
\section{Limitations \& Future Work}
In offline operation, {\em e.g.,} anonymizing addresses of clients
that access a worldwide CDN as we report here, the guarantee of
anonymization is strong because every address observed is used as
input to the anonymization. However, in online operation, {\em e.g.,}
anonymizing addresses on the fly, the guarantee is unclear because
the anonymous aggregate set was determined {\em a priori,}
based on addresses active in the past. Thus, online operation of
this method entails {\em forecasting}, wherein the anonymizer likely
assumes that past activity is suitably representative.
Lets consider attacks and situations that might
call our claimed privacy guarantee into question.
In this work, we treat an address' /64 prefix and anything more specific,
{\em e.g.,} the IID, as private.
While it's common for ISPs to provide a /64 prefix to a
customer, some ISPs will honor requests for a larger prefix,
{\em e.g.,} a /60 or /56~\cite{RFC7550,JohnB,LeeH}. Then,
the customer's router can advertise SLAAC prefix(es) to the their hosts.
In this case, it is possible for an individual customer to have a
set of simultaneously-assigned /64 prefixes, resulting in an
anonymous aggregate where the number of distinct {\em customers}
could be much less than $k$.
To combat this, an anonymizer wants to know the customer's prefix length,
so that it might increase $k$ accordingly. Discovering this prefix
length automatically (via the activity matrix) is ongoing work.
Similarly, if a malicious party generates traffic
from what would be quiescent source addresses in many unique /64 prefixes, they
might cause $k$IP-anonymization to report more specific anonymous
aggregates allowing them to determine what their neighbors' nearest
active prefixes might be. For this reason, it may be important to keep time
series of simultaneously assigned address counts (as we do here),
so that anomalous counts, {\em e.g.,} during flash crowds, or
attacks, can be identified and/or ignored.
\section{Motivation \& Introduction}
Protecting personally identifiable information (PII) in the form
of IP addresses warrants special attention with IPv6 due both
to nascent privacy concerns and mandates, {\em e.g.,} in the
European Union, and to increased IPv6 use, worldwide.
Given today's significant IPv6 deployment and dual-stack operation,
the IPv6 address may be the identifier most likely to be unique to a client on
the increasingly encrypted World-Wide Web (WWW). While individual
IPv4 addresses are increasingly shared due to address exhaustion,
this is neither intended nor commonplace with IPv6 which offers
unique, globally-routed addresses end-to-end.
In this work we investigate but one Internet privacy
measure: IP address anonymization by {\em truncation}. Address
truncation means simply to delete a set of contiguous low
(rightmost) bits, {\em i.e.,} to remove a suffix from an
input address. Typically the suffix' bits are replaced with zeroes
so that the anonymized output is an address-sized value. While
more complex anonymization techniques have been implemented and are
well-studied~\cite{DBLP:conf/imw/XuFAM01,DBLP:journals/cn/FanXAM04},
they anonymize addresses in a way that prevents the result from
being used for standard security, operations, and research tasks.
Specifically, they prevent correlation with network topology,
routing, service providers, and locations. For these purposes,
trun\-cation-based anonymization is ideal {\em if, and only if,}
it can be guaranteed to improve privacy.
Such anonymization is typically performed by truncating input
addresses to one fixed length. Consider, for instance, a WWW
analytic system employing trun\-cat\-ion\--based
IP address anonymization; {\em e.g.,} zeroing the last 8 bits of
a user's IPv4 IP address and the last 80 bits of an IPv6
address~\cite{googleAnalytics}. Essentially, this is equivalent
to masking or aggregating to /24 and /48 prefixes, respectively,
perhaps combining information about as many as 256 IPv4 addresses or
64K IPv6 /64 prefixes.
Of
course, the utilization of the IPv4 and IPv6 address spaces
differ dramatically. While someone might believe that an IPv4 /24
prefix would aggregate individual users' addresses~\footnote{Evaluating
anonymization of IPv4 addresses by truncation is warranted as
well, but it is not the subject of this work.}, we ask two
questions. First, can passive measurements inform
decisions about anonymization? Second, is
there reason to believe that any one IPv6 prefix length would
perform satisfactorily?
The key problem is how to decide at what prefix (bit) length(s) real
addresses should be cleaved into a ``public,'' suitably anonymous prefix to be
preserved and reported as is and a private suffix to be discarded
or obscured. Note that prefix preservation in truncation-based
anonymization, differs from ``prefix-preserving anon\-ymiz\-ation'' in
the literature~\cite{DBLP:conf/imw/XuFAM01,DBLP:journals/cn/FanXAM04}
which preserves prefix {\em lengths} amongst anonymized addresses
but not the original prefix {\em value.} To tackle the problem
of determining whether truncated prefixes or aggregates might
effectively provide anonymity, {\em i.e.,} to make an individual appear
indistinguishable amongst a set of individuals
(see Section 6.1.1.~\cite{RFC6973}),
we count active addresses to
determine how many they actually aggregate. Then, we use
such counting as the basis for anonymization by {\em variable length}
truncation or aggregation, resulting in different lengths to
anonymize different areas of the address space.
The key to our technique is to count a subset of simultaneously
assigned, active IPv6 addresses given the likelihood that a given
temporary privacy address must still be assigned in between two times at
which the address' activity was observed from a (possibly
remote) vantage point. We rely on the
ostensibly unique identifier present in temporary addresses that
employ Stateless Address Autoconfiguration (SLAAC) with privacy
extensions~\cite{rfc4941}. Far and away, this is the most common
address assignment mechanism for World-Wide Web (WWW) clients. As of
March 2017 by the 3-day stable definition, {\em i.e.,} ``3d-stable
(-7d,+7d)''~\cite{plonka2015temporal}, we find 684 million
(93\%) ephemeral IPv6 WWW client addresses are active per day and
4.33 billion (97\%) per week. While temporary privacy
addressing aims to improve privacy by complicating the tracking of user
activity beyond hours or days, we manage to use logs of sporadic
activity of these short-lived addresses to calculate a lower bound
on the number of simultaneously-{\em assigned} IPv6 addresses in a
given prefix {\em even during times when those addresses' seem
inactive}, {\em e.g.,} from a remote CDN vantage point.
To the best of our knowledge, prior works have not reported the
privacy concerns we have with IPv6 address anonymization by truncation nor
have they proposed privacy guarantees with such methods.
Our goal is to develop IPv6 address anonymization that yields
precise, useful results for operations and research while
guaranteeing address privacy for users. Although this is a work
in progress, we offer the following contributions: $(i)$
an evaluation of IPv6 address anonymization by truncation; $(ii)$
our {\em kIP anonymization} method
(inspired by $k$-anonymity~\cite{Samarati98protectingprivacy})
and early results on its performance.
\section{Methods}
\label{sec:method}
kIP anonymization involves three operations, namely:
address classification,
address activity matrix analysis, and
anonymous aggregate (prefix) synthesis.
\subsection{Address Classification}
We employ address classification to identify SLAAC privacy
addresses, {\em i.e.,} those having pseudorandom values in their
64-bit IID. To do so, every input address is preprocessed
by the {\tt addr6} tool which performs an initial stateless
classification~\cite{ipv6toolkit}. For example, consider
the 16 IPv6 addresses in Figure~\ref{fig:firstmatrix}; {\tt
addr6} reports each of them as having a {\tt randomized} IID
because they do not have some other easily recognized IID type,
{\em e.g.,} EUI-64, nor an easily recognized pattern, {\em e.g.,}
only low-bytes being non-zero~\cite{gont2016network}. Next,
we perform a stateful classification using our {\tt dendracron}
tool~\cite{plonka2015temporal,dendrachronology}. By ``stateful,'' we
mean that we classify each address in: $(a)$~{\em space}, relative
to others addresses in a set, {\em e.g.,} those within the same /64
prefix, and in $(b)$~{\em time}, throughout an observation timeframe,
{\em e.g.,}
a week. This yields two classification metrics for each address
that we will use below: {\em (1)} its Discriminating Prefix Length
(DPL) and {\em (2)} its number of Stable Days (SD) during which
we've observed that address to be active and throughout which the
address might have remained assigned. The DPL is simply the position
of the first
(left-most) bit at which the address differs from its nearest
(observed) address. The SD is the number of days across which
the address has been seen active. The smaller an address' DPL
(and SD) value, the more likely it is to have a randomized IID
(and a temporary one at that) by the following rationale.
\textbf{\em Identifying Plausible Randomness:}
Given a set of addresses in a /64 prefix,
the following test for randomness in IIDs complements those above;
it is based
on the likelihood that a subset of bits at a given position
is distinct across all of the IIDs
assuming the bits were chosen randomly.
For example, suppose there are 2 addresses, and consider the leading 6 bits
of their IIDs.
If these 6 bits were chosen randomly~\cite{rfc4941}, then, out of the $2^6 = 64$ possible 6-bit strings, the likelihood that these two IIDs have different values
for these 6 bits is $63/64 \approx 0.98$.
More generally, given $A$ addresses with candidate random IIDs
in a /64 prefix, and a bit string of length $N$ at a given position in the IIDs,
then the number of possible bit strings is $S = 2^N$, and the probability that
the bit strings in those IIDs are distinct is:
\begin{equation*}
\frac{S}{S} ~*~ \frac{S-1}{S} ~*~ \frac{S-2}{S} ~*~ \cdots ~*~ \frac{S-(A-1)}{S}
\end{equation*}
For classification, we start with the number of addresses and a desired probability, {\em i.e.,} 0.99, and compute the number of bits.
In particular, given $A$ addresses with candidate random IIDs, we compute the smallest number of bits, $N$, such that
the probability is at least 0.99 that all $N$-bit strings at a given position in the IIDs are distinct.
Given this $N$, we examine the IIDs in the /64 to see whether the given bit strings
are all distinct.
If all bit strings are distinct, then we infer that there is further evidence
that the IIDs are pseudorandom. If they are not all distinct, then the IIDs may still be pseudorandom, but we choose to not make the inference.
Lets consider the candidate random bit string to begin at the first bit
of the IID. RFC4941~\cite{rfc4941} dictates that the 7th bit
be set to zero in an otherwise randomized IID,
so, if the bit string spans that bit, we need to allow an additional DPL bit.
Conveniently, this makes $64+1+N$ the 99\%-probable
{\em maximum} DPL of each address in a set of addresses, of size $A$, having candidate random IIDs.
We implement this additional test for randomness using a precomputed lookup
table in {\tt dendracron}, for $A$ ranging from $2$ to about $1$ million.
For example, for 16 addresses such as those in Figure~\ref{fig:firstmatrix},
the 99\%-likely maximum DPL is $64+1+14=79$; since all the actual DPL values (in the
second column of the figure) are less than $79$, we classify these addresses
as having plausibly random IIDs as a basis for assignment inferences, next.
\subsection{Address Activity Matrix Analysis}
Richter {\em et al.}~\cite{IMC2016Beyond} employed an IP address
activity matrix, with time on the horizontal axis and address space
on the vertical axis, to visualize daily activity over months and to
calculate IPv4 address space utilization. At finer timescales, we
use an enhanced address matrix to handle the sheer size and sparsity
resulting from IPv6 assignment practices, {\em e.g.,} SLAAC with
privacy extensions. Figure~\ref{fig:firstmatrix} is an activity
matrix capturing activity per hour for 16 active IPv6 addresses, sorted
by address value, in
one /64 prefix during 24 hours. An address is active sometime during
each hour marked with ``{\tt \#}'' and inactive other hours from the CDN's
vantage; see legend in
Figure~\ref{fig:matrixlegend}. The address matrix has temporal
parameters that allow the analysis and method to operate on other
timescales: $i$, the time interval used to aggregate activity and $w$
the time window of observation. Here $i=1$ and $w=24$ in hour units.
Figure~\ref{fig:secondmatrix} is a resorted activity matrix that
contains the same activity information, but with the address' rows
in order of initial activity time (earliest first) rather than by
address value. This makes clearer which of the addresses {\em might}
be active simultaneously. (Also, when the IID is random,
sorting by address value is meaningless.) We see, for instance,
that two addresses were both active in interval (hour) 3, but we
can not conclude that they were active {\em simultaneously,} since
interval-binned summaries {\em e.g.,} hourly,
do not record durations of transactions.
In Figure~\ref{fig:thirdmatrix}, we rewrite
Figure~\ref{fig:secondmatrix}'s activity in four ways: address'
activity in just one interval $i$ (within window $w$) is marked
``{\tt X}''; address' activity in multiple, contiguous intervals
have the first interval marked ``{\tt >}'' and the last ``{\tt
<}''; intervals between those at which we {\em infer} that the given
temporary privacy address is {\em assigned} throughout and are marked
``{\tt @}''.
\textbf{\em Address Assignment Inference:}
The ability to infer address assignment between moments of activity
is the key to our method. Critically, we assume that IPv6
host implementations that support privacy extensions choose good
pseudorandom values when building their IIDs. This allows one to
{\em infer} that a given host's temporary SLAAC address with randomized
IID must still be assigned between (any) two instances of observed
activity since it is ridiculously unlikely that host or any other
will choose the
same pseudorandom value for the IID on any subsequent reconfiguration
of its network interface(s).
Now that we have inferred the intervals in which each address is
assigned, we can count the simultaneously assigned addresses. To do
so, we perform special arithmetic to ``sum'' the marks in each
hour interval (column):
{\bf{\em (1)}}
Each of a column's ``{\tt @}'' marks increment its total by 1; this
is because those addresses were assigned at every moment during
the interval and the moments between the previous and next intervals.
{\bf{\em (2)}}
Either $(a)$, each of a column's ``{\tt >}'' marks can increment its
total by 1 because we know that the moment between this interval and
the next had that additional address assigned; or $(b)$ each of a
column's ``{\tt <}'' marks can increment its total by 1 because we
know that the moment between this interval and the next had that
additional address assigned. Of $(a)$ or $(b)$, we choose whichever
column total would be larger.
{\bf{\em (3)}}
All of a column's ``{\tt X}'' marks, taken together, increment its
total by 1 only in column's having no ``{\tt >}'' or ``{\tt <}'' marks;
this is because we know there was some specific moment amidst that
interval when (at least) one of those addresses was assigned that
wasn't a moment between the previous and next intervals. This
process is performed for each column to come to a sum, for example,
shown totaled below the matrix in Figure~\ref{fig:thirdmatrix}.
These totals, single digits with whitespace removed here: ``{\tt
000100011112332321122100}'', are the lower bounds on the number of
simultaneously assigned addresses in each interval (hour). Their
minimum is $0$ and maximum is $3$, meaning we are
confident at least 3 addresses were simultaneously assigned within
this /64 prefix at some moment during this day.
The last step in our matrix analysis is to infer precise
moments that the /64 prefix,
itself: {\tt 2001:db8::/64}, {\em must} be assigned to some hosts'
interface, given when the addresses it covers were known assigned. We do
so by inferring address assignment at ``fenceposts,'' {\em i.e.,}
the moments between our ``fence sections'' (intervals) in time.
Because these are the moments between intervals, they number 1
fewer than intervals in the window: $f=w-1=23$. The /64 prefix
is inferred to have been assigned at the fencepost trailing each
interval where there is a {\tt !} (exclamation point) mark.
We now have a time series array of size $f$, temporarily encoded:
``{\tt --------!!!!!!!!-!!!!--?},'' that indicates the precise moments
(between intervals) when the prefix must have been assigned.
Translating the {\tt !} marks to $1$ (and others to $0$) makes them
time series values suitable for accumulating across an entire network
(or the entire Internet) to compute $f$ (hourly) lower bounds on the count of
simultaneously assigned /64 prefixes.
\lstset{keepspaces=true,basicstyle=\scriptsize\ttfamily}
\newsavebox{\firstmatrix}
\begin{lrbox}{\firstmatrix}
\begin{lstlisting}
D
P S hour of day
IPv6 address L D 0 1 2
----------------------------- -- - 012345678901234567890123
2001:db8::117a:e091:b2bd:ca65 67 0 |-------+-------+--##---
2001:db8::21ad:6d24:641a:1314 68 0 |--#----+-------+-------
2001:db8::3454:ae0d:20a0:df4d 68 0 |-------+--#----+-------
2001:db8::4974:fa8b:465d:4c2a 68 0 |-------+-------+#---#--
2001:db8::503c:a91d:be00:9a63 68 0 |-------##-###--+-------
2001:db8::6867:8a64:5417:e731 70 0 |-------+---##--+-------
2001:db8::6d35:ee11:ec45:f658 70 0 |-------+-------+#------
2001:db8::7070:a7fc:47d5:02ba 70 0 |------#+-------+-------
2001:db8::7554:b66a:a983:9665 70 0 |-------+--#----+-------
2001:db8::7939:1bd6:fec2:85bb 70 0 |-------+------#+-------
2001:db8::7ccc:3977:7c76:bdef 70 0 |-------+-------+---#---
2001:db8::890b:1f0d:14e2:0ccb 67 0 |-------+----#--+-------
2001:db8::a0fc:1e18:48aa:eb2e 67 0 |-------+---#---#-------
2001:db8::f930:9833:f8c5:3926 74 0 |-------+----#--#-------
2001:db8::f94d:fcec:6b8e:d61f 74 0 |-------#-------+-------
2001:db8::fd28:50fe:8445:83e7 70 0 |--#----+-------+-------
2001:db8::/64 16; Temporary SLAAC: 100.00%
2001:db8::/64 -------------------------------------------->
\end{lstlisting}
\end{lrbox}
\newsavebox{\secondmatrix}
\begin{lrbox}{\secondmatrix}
\begin{lstlisting}
hour of day
0 1 2
012345678901234567890123
|--#----+-------+-------
|--#----+-------+-------
|------#+-------+-------
|-------##-###--+-------
|-------#-------+-------
|-------+--#----+-------
|-------+--#----+-------
|-------+---##--+-------
|-------+---#---#-------
|-------+----#--#-------
|-------+----#--+-------
|-------+------#+-------
|-------+-------+#---#--
|-------+-------+#------
|-------+-------+--##---
|-------+-------+---#---
----------------------->
\end{lstlisting}
\end{lrbox}
\newsavebox{\thirdmatrix}
\begin{lrbox}{\thirdmatrix}
\begin{lstlisting}
hour of day
0 1 2
012345678901234567890123
|--X----+-------+-------
|--X----+-------+-------
|------X+-------+-------
|------->@@@@<--+-------
|-------X-------+-------
|-------+--X----+-------
|-------+--X----+-------
|-------+---><--+-------
|-------+--->@@@<-------
|-------+---->@@<-------
|-------+----X--+-------
|-------+------X+-------
|-------+-------+>@@@<--
|-------+-------+X------
|-------+-------+--><---
|-------+-------+---X---
000100011112332321122100
|-------+-------+-------
--------!!!!!!!!-!!!!--?
\end{lstlisting}
\end{lrbox}
\newsavebox{\legend}
\begin{lrbox}{\legend}
\begin{lstlisting}
| = every 24th interval in the matrix, e.g., first hour of day, UTC
+ = every 8th interval in the matrix, e.g., 8 hours
# = activity logged during the given hour
X = activity started and ended during the given hour, i.e., a "short" episode
> = activity started during the given hour, i.e., beginning of an episode
< = activity ended during the given hour, i.e., end of an episode
@ = infer address assigned throughout the given interval, e.g., hour
! = infer /64 prefix assigned at trailing edge of given hour, i.e., the "fencepost" moments between intervals
? = the last "fencepost" moment is discarded since address assignment can't be determined (yet)
\end{lstlisting}
\end{lrbox}
\begin{figure*}
\begin{center}
\subfigure[Initial Activity Matrix: 16 addresses in 1 /64 prefix]{\label{fig:firstmatrix}\usebox{\firstmatrix}}
\subfigure[Time Sorted Activity]{\label{fig:secondmatrix}\usebox{\secondmatrix}}
\subfigure[Inferred Assignment]{\label{fig:thirdmatrix}\usebox{\thirdmatrix}}
\subfigure[matrix legend]{\label{fig:matrixlegend}\usebox{\legend}}
\vspace{-3mm}
\caption{$(a)$ Address classifications (DPL, SD) and activity matrix for one /64 prefix: \texttt{2001:db8::/64}, here, having 16 SLAAC addresses observed as active during 24 hours time.
In the matrices, space is represented vertically and time progresses horizontally, left to right.
The matrix' addresses are then sorted $(b)$ by their initial activity times. Finally, we infer $(c)$ the number of simultaneously assigned IPv6 addresses and /64 prefix(es) on-off times.
\label{fig:matrix}}
\end{center}
\end{figure*}
\newsavebox{\meetingmatrix}
\begin{lrbox}{\meetingmatrix}
\begin{lstlisting}
/64 Inferred Assignment Matrix
------------------------
hour of day
0 1 2
012345678901234567890123
2001:db8:370::/64 !!!!!!!!!!!!!!!!!!!!!!!
2001:db8:370:128::/64 !!!!!!!!!!!!!!!!!!!!!!!
2001:db8:370:228::/64 !------------!---------
\end{lstlisting}
\end{lrbox}
\begin{figure*}
\begin{center}
\subfigure[]{
\label{fig:meetingtree}
\includegraphics[width=0.75\textwidth]{Meeting_Network_1day_64_ptCrop}
}
\subfigure[]{
\label{fig:meetingtreetoo}
\includegraphics[width=1.0\textwidth]{Meeting_Network_1day_64_too_ptCrop}
}
\subfigure[]{
\label{fig:meetinganon}
\includegraphics[width=0.8\textwidth]{Meeting_Network_1day_64_agurify-k2_min_ptCrop}
}
\vspace{-3mm}
\caption{
A 3-step method to calculate anonymous aggregates
using a binary PATRICIA tree, for example, the
Meeting Network in Table~\ref{tab:dataSummary}. (a) First, a tree is
populated with active /64 prefixes (the 3 nodes with solid lines), each
with a time series showing assignment, on or off. (b) Next, the tree is
aggregated and the arrays added to their parent until the min count
of $k=2$ is reached. Here, the first two /64s have been aggregated to their
parent, but the third has not yet been visited. (c) Finally, after
aggregation is complete, we report only the anonymous aggregates
with (at least) the desired number of simultaneous addresses ($2$):
only the one prefix shown in a solid box.
\label{fig:meeting}}
\end{center}
\end{figure*}
\subsection{Synthesizing Anonymous Aggregates}
Our synthesis of aggregates, {\em i.e.,} larger, covering prefixes
that coalesce individuals' SLAAC prefixes for privacy, is inspired by
Cho {\em et al.}~\cite{DBLP:conf/qofis/ChoKK01,plonka2015temporal}
who developed ``aguri'' which recursively aggregates prefixes based
on activity counts until a stopping condition is met. We augment
the aguri tree with an array (size $f$) of time series counters
in each node and with a new stopping condition where a statistical
function of those arrays' (hourly) values, {\em i.e.,} either minimum
(min) or maximum (max) meets or exceed $k$, the desired number of
active addresses that can not be distinguished once anon\-ym\-ized.
We use min and max to show the range of outcomes, and choose median
as a robust metric to avoid outliers, {\em e.g.} due to flash crowds,
quiescent hours of day, or sustained outages.
Figure~\ref{fig:meeting} shows a simple example for active
address in a Meeting Network, which had only 3 active /64 prefixes covering
all the attendees active addresses (as shown in Table~\ref{tab:dataSummary}),
despite the meeting running for a week and providing IPv6 connectivity,
both wired and wireless, to its $1000+$ attendees.
With minimum $k=2$, our method reports one anonymous aggregate,
\texttt{2001:db8:370::/55}, because this network has few active /64 prefixes.
\section{Results}
We offer two early results based on logged IPv6
addresses of active WWW clients from real networks: {\em
(1)}~an evaluation of current practice for truncation-based
anony\-miz\-ation and {\em (2)}~a characterization of the anonymous aggregates
produced by $k$IP anony\-miz\-ation. Both these use real WWW client addresses
found in the activity logs of a large CDN as input data.
Consider Table~\ref{tab:dataSummary}
which characterizes our active WWW client address data sets.
We chose these specific networks for the variety of address assignment
practices they demonstrate, rather than, {\em e.g.,} their country.
We first show the Meeting Network
used in Figure~\ref{fig:meeting} in Section~\ref{sec:method}.
Note the lower bounds for simultaneously assigned /64 prefixes and addresses
and their plausibility with 1000 attendees,
a subset of whose wired and wireless hosts (apparently) use SLAAC.
Table~\ref{tab:dataSummary} also shows three ISP networks,
one each from Europe (EU), Japan (JP), and the United States (US).
Seen over the 7 days of observation,
the three ISPs have wildly varying numbers of active /64 prefixes
(up to $\sim 10\times$ different) and /48 covering prefixes
(up to $\sim 300\times$ different). This strongly suggests that either
{\em (a)} their subnet and address assignment practices differ greatly or
{\em (b)} their WWW client population sizes differ greatly.
However, note that they share similar lower bounds on number of
simultaneously assigned /64 (SLAAC) prefixes and addresses calculated
by our method, each being in the low single-digit millions.
These intermediate results, {\em i.e.,} estimated counts of
each network's IPv6-capable WWW clients, suggests that our method
can mitigate bias in counting caused by some networks' address
assignment practices and is a basis of ongoing work. In networks
whose IPv6 addresses' network identifiers (in addition to IIDs) contain
pseudorandom segments or have short assignment periods from very large
pools, simple counts, {\em e.g.,} active prefixes per day or week,
can far exceed the number of IPv6-capable clients that actually
exist in a given network. Thus, more careful counting is warranted
as the basis for guaranteed anonymity in aggregates.
In our results below, we used the full ISP data sets in
Table~\ref{tab:dataSummary}; meaning $w=168$ (overall window
of time) and $f=167$ (fencepost moments) in units of $i=1$ hour
(intervals). Thus, min is the minimum
of the 167 lower bounds counts of simultaneously-assigned /64
prefixes, {\em i.e.,} 167 hours (one week's time). Likewise,
max and median are of the 167 (hourly) lower bound counts.
\subsection{Evaluation of /48 Aggregation}
Today, a not uncommon practice for IPv6 address anonymization
is to truncate the low 80 bits of the address, preserving only the
address' /48 (covering) prefix as the supposed anonymous aggregate.
Based simply on counting, in Table~\ref{tab:dataSummary}, we see
that the JP ISP has the same count (2.46 million) of active /64
prefixes and /48 prefixes. This means, by and large, the JP ISP does
not use the bits 48-63 to differentiate individual SLAAC-prefixed
subscribers. (Those bits happen to be zeroes, but they could use
any value to the same effect.) Thus, truncating the last 80 bits
of the JP ISP addresses does nothing to obscure the SLAAC-prefix
ostensibly associated with an individual customer. This proves 80
bit truncation, and /48 aggregation, is ineffective at anonymizing
a customer's SLAAC prefix in this network.
To explore this more broadly, Figure~\ref{fig:k2} plots the
distribution of prefix lengths necessary to aggregate 2 simultaneously
assigned active /64 prefixes together.
(IPv6 hosts have a 64-bit subnet prefix~\cite{RFC4291,RFC7421},
thus an ISP commonly provides at least a /64 prefix to each customer
or subscriber.)
This is based on our
empirical measurements and lower bounds on numbers of
simultaneously-assigned IPv6 /64 prefixes as described in
Section~\ref{sec:method}.
In this figure, note the JP ISP's (red) $k=2$ min and max prefix lengths
always plot left of bit 48 on the horizontal axis; as above, this confirms
that aggregation to /48 would {\em never} aggregate any individual
customer's SLAAC prefixes together, thus truncation to /48 does not
improve privacy to subscribers. The figure also shows (solid
blue) that the EU ISP's $k=2$ min more than 25\% of the prefixes
needed prefix lengths less than 48, {\em i.e.,} not even two /64s are
aggregated by /48s therein.
Moreover, we presume no one would settle for such a weak notion
of anonymity as $k=2$, which would be equivalent to truncating only 1
bit in IPv4, or aggregating to /31. Both in the past and today with IPv4,
it is common to truncate or aggregate to /24 prefixes, presumably with
the intention of aggregating up to 256 (max) individuals' addresses
together. We consider more reasonable values of $k$ for privacy, next.
\subsection{Anonymous Aggregates}
We now apply $k$IP anon\-ymization to addresses for each of the
three ISPs. Figure~\ref{fig:aggCDF} characterizes the resulting
prefixes in histograms for $k=32$ and $k=256$. (Essentially, these
are equivalent to aggregating a fully-utilized IPv4 network to /27 or
/24, respectively.) At these levels of anonymization, we find that,
the JP ISP and EU ISP almost always required more aggregation
than /48 (more than 80 bits truncated) for us to {\em guarantee}
that the aggregation meets our desired $k$ on lower bounds for
median counts of simultaneously-assigned addresses. The $k$IP
anonymous aggregate prefixes reported here vary from /25 to /58.
In CDF plots (not shown) for the US ISP's customer's, we find
that /48 aggregation guarantees $k=32$ anonymization for 90-95\%
(min-max) of those customers, but guarantees $k=256$ anonymization
for only 30-40\% of those customers. By comparing the resulting
$k$IP anonymous aggregate prefixes counts to fixed-length /48
prefix counts, as shown parenthetically in the legends in Figure~\ref{fig:aggCDF},
by and large, we see that $k$IP anonymization (with $k=256$ or even
$k=32$) can yield a much smaller set of anonymous prefixes while
guaranteeing significant aggregation of individuals' /64 prefixes.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.34]{dendracron_64_agurify-kK2_prefix_len_cdfSquashedCrop}
\caption{CDF of aggregate prefix lengths, $k=2$.} \label{fig:k2}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[histogram: median $k=32$]{\includegraphics[scale=0.34]{dendracron_64_agurify-k32-p50_prefix_len_histSquashedCrop} \label{fig:k32hist}}\\
\subfigure[histogram: median $k=256$]{\includegraphics[scale=0.34]{dendracron_64_agurify-k256-p50_prefix_len_histSquashedCrop} \label{fig:k256hist}}
\caption{Anonymous aggregate prefix lengths.} \label{fig:aggCDF}
\end{figure}
|
\section{Introduction}
The electronic properties of a DNA molecule is one of main subjects in biophysics, which occupies many researchers. This is because that each type of defects or damage to a DNA has a direct effect on its electronic properties. In the past, there have been many studies on this subject. For example, some authors have argued about the charge transport in a single-stranded DNA molecule in the direction perpendicular to the backbone axis and discussed that the different electronic structure of the four bases causes a difference in conductance of the nucleotides \cite{e1}. In another paper, the electronic transfer properties of a DNA molecule have been investigated in terms of localization lengths and a novel enhancement of localization lengths have been obtained at some energies for an progressing binary backbone disorder \cite{e2}. In another research effort, authors have discussed the possibility of a mechanism for fast sequencing based on the distributions of transverse electrical currents in single-stranded DNA. Their method gives this opportunity to study, predict and cure diseases from the perspective of the genetic compound of each individual \cite{e3}. In yet another study, authors have observed the high spin selectivity and the length-dependent spin polarization in double-stranded DNA \cite{e4}. Also, they have shown that the spin polarization in double-stranded DNA is significant even in the case of weak spin-orbit coupling, while no spin polarization can be seen in single-stranded DNA \cite{e5}. Now, the question arises: could we consider the evolution of DNA by detecting and analyzing their electromagnetic radiation?
In 2009, Montagnier and his collaborators have investigated about the capacity of some bacterial DNA sequences to radiate very low frequency electromagnetic signals in high aqueous dilutions. They have argued that the genomic DNA of most pathogenic bacteria contains sequences which are able to produce such signals \cite{DNA1}. In another research, authors have explained the experimental conditions by which electromagnetic waves (EMS) of low frequency can be radiated by diluted aqueous solutions of some bacterial and viral DNAs. Also, they have detected this transduction process in living human cells exposed to EMS irradiation and proposed a quantum field theory analysis of the phenomenon \cite{DNA2}. In parallel investigations, some other authors have considered the responses of DNA to electromagnetic fields in different frequency ranges, and specified the properties of DNA molecules as antennas \cite{DNA3}. Motivated by such research, in this paper we suggest a mathematical model for DNA and obtain its the energy and the current using the concepts of M-theory. We design an electrical circuit for communication with DNA and for the prevention of disease progression. We will show that by analyzing the radiated waves from DNA molecules, we can consider their development and defects. Then, by sending electromagnetic signals to DNA molecules, we can cancel the effect of damaged DNAs on the normal DNAs.
The outline of this paper is as follows. In section \ref{oo1}, we propose a mathematical model for calculating the energy of bases in a DNA molecule. In section \ref{o1}, we design a model for the electrical circuit of a normal DNA. In section \ref{o2}, we study the origin of the radiated waves from a damaged DNA and suggest a model for preventing disease progression. In section \ref{o3}, we suggest a mechanism to use the difference between radiated waves from DNAs in men and women to cure the cancer. The last section is devoted to summary and conclusions.
\section{A mathematical model for calculating the energy of bases in a DNA }\label{oo1}
In this section, first, we introduce polygonal molecules and construct an $N$-dimensional polygonal manifold. To this aim, we use concepts of string theory. We assume that electrons are similar to strings and the hexagonal shape of molecules is similar to the shape of polygonal manifolds. We will show that at first stage, there are molecules (point-like polygonal manifolds) in the human body (see Figure 1) and electrons (strings) are attached to them. Polygonal manifolds are manifolds with a $q$-gonal shape which are similar to molecules of DNA. These manifolds have only one dimension, in the direction of time. All interactions between electrons are simulated by interactions between strings. The potential of these interactions can be given by a delta function and thus, the energy of manifold shrinks to one. We can write:
\begin{eqnarray}
&& V(\sqrt{C(\tilde{X}^{I})})= \delta(\sqrt{C(\tilde{X}^{I})}) \nonumber\\&& E_{M^{0}}=1=\int_{M^{0}} d \sqrt{C(\tilde{X}^{I})} V(\sqrt{C(\tilde{X}^{I})})=\int_{M^{0}} d\sqrt{C(\tilde{X}^{I})} \delta(\sqrt{C(\tilde{X}^{I})})= \nonumber\\&& \int_{M^{0}} d \sqrt{C(\tilde{X}^{I})} e^{-\sqrt{C(\tilde{X}^{I})}\sqrt{C(\tilde{X}_{I})}} \label{spm1}
\end{eqnarray}
where $M^{0}$ represents the point-like manifold, $\tilde{X}^{I}$'s are strings hat are attached to them and $y$ is the length of the point, which shrinks to zero. Also, $C(\tilde{X}^{I})$ is the equation of the polygonal manifold. In the above equation, $C(\tilde{X}^{I})$ has the following relation with $\tilde{X}^{I}_{i}$:
\begin{eqnarray}
&& C(\tilde{X}^{I})=\Sigma_{i=1}^{q} \frac{(\tilde{X}^{I}_{i})^{2}}{2\pi y \tilde{a}_{i}^{2}} \label{ww1}
\end{eqnarray}
where $\tilde{a}_{i}$ is the coefficient of the $i$-th side of a point-like polygonal manifold. For example, for a sphere, we can write:
\begin{eqnarray}
&& C(\tilde{X}^{I})= \frac{(\tilde{X}^{I}_{1})^{2}}{2\pi y R^{3}} + \frac{(\tilde{X}^{I}_{2})^{2}}{2\pi y R^{3}} + \frac{(\tilde{X}^{I}_{3})^{2}}{2\pi y R^{3}} \label{ww2}
\end{eqnarray}
Using a new redefinition of string fields $\tilde{X}^{I}\longrightarrow \sqrt{2\pi y} X^{I}$, we obtain:
\begin{eqnarray}
&& E_{M^{0}}=1= \int_{M^{0}} d \sqrt{C(X^{I})} e^{-\pi \sqrt{C(X^{I})}\sqrt{C(X_{I})}} \label{spm2}
\end{eqnarray}
By calculating the integral, we can obtain a solution for strings ($X^{I}$):
\begin{eqnarray}
&& \int_{M^{0}} d\sqrt{C(X^{I})} e^{-\pi \sqrt{C(X^{I})}\sqrt{C(X_{I})}}=1\longrightarrow \nonumber\\&& \frac{1}{2}erf(\sqrt{ \sqrt{C(X^{I})}\sqrt{C(X_{I})}\pi})=1 \longrightarrow \nonumber\\&& C(X^{I})\approx I\label{spm3}
\end{eqnarray}
where $I$ is a unitary matrix. This equation shows that at the first stage, there is no interaction between strings and they are the same. In fact, there is a very high symmetry for the early stages of system and all matters have the same origin. For example, for a sphere, the above equation yields:
\begin{eqnarray}
&& \frac{(X^{I}_{1})^{2}}{R^{3}} + \frac{(X^{I}_{2})^{2}}{R^{3}} + \frac{(X^{I}_{3})^{2}}{R^{3}}\approx I\label{ww3}
\end{eqnarray}
The above example shows that initial strings constitute a spherical manifold with radius $R$. By changing $\tilde{a}_{i}$ and number of sides, the shape of a point-like manifold can be changed.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{ff1.jpg}
\end{tabular}
\end{center}
\caption{ One example of point-like manifolds and attached strings. }
\end{figure*}
We can extend this mechanism to $N$-dimensional manifolds (see Figure 2). Each $N$-dimensional manifold is similar to an $N$-dimensional DNA. This manifold can be built by joining $(N+1)$ point-like manifolds. The potential of interactions between strings on them can be described by a delta function. In fact, all interactions on a point-like manifold are concentrated at a point and thus the total potential is approximately infinite. This potential, which is zero at other points and infinite at an special point, can be represented by a delta function. Integrating over all these potentials, we can obtain the total energy of system:
\begin{eqnarray}
&& E_{M_{i}^{0}}=\int_{M_{i}^{0}} d C (\tilde{ X_{i}}^{I}) V( C(\tilde{X_{i}}^{I}))=\int_{M_{i}^{0}} d C(\tilde{X_{i}}^{I}) \delta(C (\tilde{X_{i}}^{I}))=1\Rightarrow \nonumber\\&&\nonumber\\&& \nonumber\\&& E_{M_{1}^{0}+...+M_{N+1}^{0}}=1=\int_{M_{1}^{0}+...+M_{N+1}^{0}} d C(\tilde{X_{1}}^{I})...d C(\tilde{X_{N+1}}^{I}) \delta(C(\tilde{X_{1}}^{I}))...\delta(C (\tilde{X_{N+1}}^{I}))=\nonumber\\&& \int_{M_{1}^{0}+...+M_{N+1}^{0}} d C(\tilde{X_{1}}^{I})...d C(\tilde{X_{N+1}}^{I})e^{-C(\tilde{X_{1}}^{I}) C(\tilde{X_{1}}_{I})}...e^{-C(\tilde{X_{N+1}}^{I}) C(\tilde{X_{N+1}}_{I})}\nonumber\\&& \label{spm24}
\end{eqnarray}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{ff8.jpg}
\end{tabular}
\end{center}
\caption{ Constructing an $N$-dimensional manifold by joining point-like manifolds. }
\end{figure*}
In the above equation, the energy of the manifold is normalized to one. It is concluded that for an $N$-dimensional manifold, there are $N+1$ integrations. An extra dimension corresponds to time and in totality, we have an $(N+1)$-dimensional manifold. Using a new definition of the string fields $\tilde{X}^{I}\longrightarrow \sqrt{2\pi y} X^{I}$, we can obtain:
\begin{eqnarray}
&& E_{M^{N}}=1= \int_{M_{1}^{0}+....+M_{N+1}^{0}} d \sqrt{C (X_{1}^{I})}... d \sqrt{C(X_{N+1}^{I})} e^{-\pi \sqrt{C (X_{1}^{I})}\sqrt{C (X_{1 I})}}....e^{-\pi \sqrt{C( X_{N+1}^{I})}\sqrt{C (X_{N+1 I})}} \label{spm25}
\end{eqnarray}
By joining point-like manifolds and building an $N$-dimensional manifold, strings will be functions of coordinates of $(N+1)$-manifolds ($X^{I}(y_{1}....y_{N+1})$). Thus, equation (\ref{spm25}) can be written in the form
\begin{eqnarray}
&& E_{M^{N}}=1= \int_{M_{1}^{0}+...+M_{N+1}^{0}}dy_{1}^{I}...dy_{N+1}^{I} (\Sigma_{i_{1},...i_{N+1}=1...N+1}\frac{d \sqrt{C (X_{i_{2}}^{I})}}{dy_{i_{1}}^{I}}\frac{d \sqrt{C(X_{i_{1}}^{I})}}{dy_{i_{2}}^{I}}...\frac{d \sqrt{C (X_{i_{N+1}}^{I})}}{dy_{i_{N}}^{I}}\frac{d \sqrt{C(X_{i_{N}}^{I})}}{dy_{i_{N+1}}^{I}} \nonumber\\&& e^{-\pi \sqrt{C(X_{i_{1}}^{I})} \sqrt{C (X_{i_{1} I})}}...e^{-\pi \sqrt{C(X_{i_{N+1}}^{I})} \sqrt{C(X_{i_{N+1} I})}} + \nonumber\\&& \int_{M_{1}^{0}+...+M_{N+1}^{0}}dy_{i_{1}}^{I}...dy^{i_{1}I} (\frac{d \sqrt{C (X_{i_{1}}^{I})}}{dy_{i_{1}}^{I}}....\frac{d \sqrt{C(X_{i_{N+1} I})}}{dy_{i_{1} I}}) e^{-\pi \sqrt{C (X_{i_{1}}^{I}}) \sqrt{C (X_{i_{1} I})}}...e^{-\pi \sqrt{C (X_{i_{N+1}}^{I})} \sqrt{C (X_{i_{N+1} I})}}+\nonumber\\&&...+\nonumber\\&& \int_{M_{1}^{0}+...+M_{N+1}^{0}}dy_{i_{N+1}}^{I}...dy^{i_{N+1}I} (\frac{d \sqrt{C (X_{i_{1}}^{I})}}{dy_{i_{N+1}}^{I}}....\frac{d \sqrt{C (X_{i_{N+1} I})}}{dy_{i_{N+1} I}}) e^{-\pi\sqrt{ C (X_{i_{1}}^{I})} \sqrt{C (X_{i_{1} I})}}...e^{-\pi \sqrt{C (X_{i_{N+1}}^{I})} \sqrt{C (X_{i_{N+1} I})}}\label{spm26}
\end{eqnarray}
Taylor-expanding the exponential functions over the crossed points, we obtain
\begin{eqnarray}
&& E_{M^{N}}=1=
\int_{M_{1}^{0}+...+M_{N+1}^{0}}dy_{1}^{I}...dy_{N+1}^{I} (\Sigma_{i_{1},...i_{N+1}=1...N+1}\frac{d \sqrt{C (X_{i_{2}}^{I})}}{dy_{i_{1}}^{I}}\frac{d \sqrt{C (X_{i_{1}}^{I})}}{dy_{i_{2}}^{I}}...\frac{d \sqrt{C(X_{i_{N+1}}^{I})}}{dy_{i_{N}}^{I}}\frac{d \sqrt{C(X_{i_{N}}^{I})}}{dy_{i_{N+1}}^{I}}\times \nonumber\\&& \pi^{N+1} (\Sigma_{i_{1},...,i_{N+1}=1,,,,N+1}(\frac{\partial}{\partial y_{i_{1} I}}(\sqrt{C (X_{i_{1}}^{I})}\sqrt{ C (X_{i_{1} I})})+....+\frac{\partial}{\partial y_{i_{1} I}}...\frac{\partial}{\partial y_{i_{N+1}}^{I}}(\sqrt{C (X_{i_{1}}^{I}}) \sqrt{C (X_{i_{1} I})})+1)\times \nonumber\\&&....\times \nonumber\\&&(\frac{\partial}{\partial y_{i_{1} I}}(\sqrt{C (X_{i_{N+1}}^{I})} \sqrt{C (X_{i_{N+1} I})})+...+\frac{\partial}{\partial y_{i_{1} I}}...\frac{\partial}{\partial y_{i_{N+1}}^{I}}(\sqrt{C (X_{i_{N+1}}^{I})} \sqrt{C (X_{i_{N+1} I})})+1)\times \nonumber\\&&\Big(\Sigma_{i_{1},...i_{N+1}=0...N+1}(y_{i_{N+1}}^{I}-y_{0}^{I})^{i_{N+1}}....(y_{1 I}-y_{0 I})^{i_{1}}\Big) + ...\label{spm27}
\end{eqnarray}
We now wish to obtain the relation between different derivatives of strings respect to coordinates and various types of the matter, such as scalars, fermions, gauge fields and gravitons. For this, we compare the action of branes in string theory with the action of matter in field theory. We can write \cite{D3}:
\begin{eqnarray}
S_{Gravity-Matter}&=& S_{D3} \label{D4}
\end{eqnarray}
where
\begin{eqnarray}
S_{D3}&=& -T_{D3}\int d^{4}y \sqrt{1+ g_{ij}\partial_{a}X^{i}\partial^{a}X^{j} -4\pi^{2} l_{s}^{4}F_{ab}F^{ab}},\nonumber\\ S_{Gravity-Matter}&=& \int d^{4}y \sqrt{-g}\Big(R +g_{ab}\partial^{a}\phi\partial^{b}\phi -i\bar{\psi}\gamma^{a}\partial_{a}\psi+A_{i}A^{i}+\frac{1}{2}\phi^{2}+1\Big), \label{D2}
\end{eqnarray}
Here, $A_{b}$ is the gauge field, $F_{ab}$ is the field strength, $X^{\mu}$ is the string, $g_{\mu\nu}$ is the metric, $T_{D3}$ is tension and $l_{s}$ is the string length. Also, $\phi$ is the scalar field and $\psi$ is the fermionic field. Using equation (\ref{D4}), we can write the following relations between strings and fields (see Figures 3,4 and 5):
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{ff3.jpg}
\end{tabular}
\end{center}
\caption{ Strings with both ends located on a point-like manifold form scalar fields ($\phi$). }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{ff4.jpg}
\end{tabular}
\end{center}
\caption{ Strings with one end on a point-like manifold and the other on another point-like manifold form gauge fields ($F$-fields ($F_{ij}$)). }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{ff5.jpg}
\end{tabular}
\end{center}
\caption{ Strings with two ends near a linked point, yielding a symmetric shape, form graviton fields ($g_{ij}$). }
\end{figure*}
\begin{eqnarray}
&& X_{1}^{I}X_{1 I}=....= X_{j}^{I}X_{j I} \nonumber\\&& \nonumber\\&& \frac{dX_{i}^{I}}{dy_{j}^{I}}=e^{i}_{j} \quad i,j=1,2 \quad F_{12}\rightarrow \Big( \frac{\partial}{\partial y_{2 I}}\frac{\partial}{\partial y_{1}^{I}}(X_{1}^{I}X_{1 I})- \frac{\partial}{\partial y_{1 I}}\frac{\partial}{\partial y_{2}^{I}}(X_{1}^{I}X_{1 I})\Big) \nonumber\\&& g_{12}\rightarrow \Big( \frac{\partial}{\partial y_{2 I}}\frac{\partial}{\partial y_{1}^{I}}(X_{1}^{I}X_{1 I})+ \frac{\partial}{\partial y_{1 I}}\frac{\partial}{\partial y_{2}^{I}}(X_{1}^{I}X_{1 I})\Big) \nonumber\\&&\nonumber\\&& \phi \rightarrow \frac{\partial}{\partial y_{1 I}}\frac{\partial}{\partial y_{1}^{I}}(X_{1}^{I}X_{1 I})\quad \sqrt{-g}g^{12} \rightarrow (\frac{dX_{1}^{I}}{dy_{2}^{I}}\frac{dX_{2 I}}{dy_{2 I}}) \quad \phi\rightarrow \psi_{\downarrow}\psi_{\uparrow} \nonumber\\&& \nonumber\\&& \Sigma_{ijkf}\frac{\partial}{\partial y_{f}^{I}}\frac{\partial}{\partial y_{k I}}\frac{\partial}{\partial y_{j I}}\frac{\partial}{\partial y_{i}^{I}}(X_{i}^{I}X_{i I})\approx \frac{1}{2}R_{ijkf} + \epsilon^{ijkf}\frac{\partial}{\partial y_{f}^{I}}\frac{\partial}{\partial y_{k I}}(F_{ij}))\label{spm28}
\end{eqnarray}
We suppose that point-like manifolds are very close to each other and thus, we can obtain:
\begin{eqnarray}
&& y_{i}^{I}-y_{0}^{I}= ...y_{3}^{I}-y_{0}^{I}= y_{2}^{I}-y_{0}^{I}=y_{1}^{I}-y_{0}^{I}=\sigma\longrightarrow \frac{1}{\pi} \label{spmtt16}
\end{eqnarray}
Using equations ( \ref{spm28} and \ref{spmtt16} ) and equation (\ref{spm27}), we obtain:
\begin{eqnarray}
&& E_{M^{N}}=1= \int_{M^{N}}d^{N+1}y\sqrt{-g} [\Pi_{n'=1}^{N} \Sigma_{\alpha_{n'}=1}^{q_{n'}}\frac{1}{ \tilde{a}_{\alpha_{n'}}^{2}}\Big(R^{\alpha_{n'}}-\frac{1}{2}\partial_{i}\phi^{\alpha_{n'}}\partial^{i}\phi^{\alpha_{n'}} - \frac{1}{4}\epsilon^{ijkm} F_{ij}^{\alpha_{n'}}F_{km}^{\alpha_{n'}}\nonumber\\&&+i\gamma^{i}\bar{\psi}\partial_{i}\psi.....-\nonumber\\&&\frac{1}{N(N-1)..1}\epsilon^{i_{1}i_{2}...i_{N}}( F_{i_{1}i_{2}}^{\alpha_{n'}}...F_{i_{N-3}i_{N-2}}^{\alpha_{n'}}F_{i_{N-1}i_{N}}^{\alpha_{n'}})... + \frac{1}{N(N-1)..1}\epsilon^{i_{1}i_{2}...i_{N}} (R_{i_{1}i_{2}}^{\alpha_{n'}}...R_{i_{N-3}i_{N-2}}^{\alpha_{n'}}R_{i_{N-1}i_{N}}^{\alpha_{n'}})-\nonumber\\&&.....-\nonumber\\&&\frac{1}{N(N-1)..1}\epsilon^{i_{1}i_{2}...i_{N}}( \partial_{i_{1}}...\partial_{i_{N-2}}F_{i_{N-1}i_{N}}^{\alpha_{n'}})... + \frac{1}{N(N-1)..1}\epsilon^{i_{1}i_{2}...i_{N}} (\partial_{i_{1}}...\partial_{i_{N-2}}R_{i_{N-1}i_{N}}^{\alpha_{n'}})+ V(\phi^{\alpha_{n'}}) +... \Big)] \nonumber\\&& V(\phi)=1+ \frac{\phi^{2}}{2}..+\frac{\phi^{N}}{N(N-1)...1} \label{spm29}
\end{eqnarray}
In a DNA molecule, $N$ is the number of bases, $q_{n'}$ is the number of corners in each base, $F$ is the field strengths of photons which are exchanged between bases and $R$ is the curvature of gravity which is exchanged between DNA molecules. This equation shows that by joining point-like polygonal manifolds and building $N$-dimensional ones, the related energy in usual quantum field theory is produced. This energy contains all interactions that may be produced between bases in a DNA molecule.
This energy is seemed to be very distant of results of refs \cite{R7,R8}. However, if we use of the relation between gauge fields, curvatures, scalars and fermions in equation (\ref{spm29}), we can achieve the same results. To this end, we generalize our calculations to M-theory and use of the following replacements for gauge fields and scalars \cite{R7,R8}:
\begin{eqnarray}
&&F_{ab} \text { in string theory} \rightarrow F_{abc} \text { in M-theory},\nonumber\\ && F_{abc}=\partial_{a} A_{bc}-\partial_{b} A_{ca}+\partial_{c}
A_{ab}.\nonumber\\ &&\partial_{a}\phi\partial^{a}\phi \text { in string theory} \rightarrow \partial_{b}\partial_{a}\phi\partial^{a}\partial^{b}\phi \text { in M-theory}\label{wj2}
\end{eqnarray}
These gauge fields and also scalars have a direct relation with spinor fields like electrons \cite{R7,R8}:
\begin{eqnarray}
&& A_{ab}\rightarrow \psi^{U}_{a}\psi^{L}_{b}-\psi^{L}_{a}\psi^{U}_{b}\nonumber \\
&& \phi \rightarrow \psi^{U}_{a}A^{ab}\psi^{L}_{b}-\psi^{L}_{a}A^{ab}\psi^{U}_{b}+\Psi^{U}_{a}A^{ab}\Psi^{L}_{b}-\Psi^{L}_{a}A^{ab}\Psi^{U}_{b}+\Psi^{U}_{a}A^{ab}\psi^{L}_{b}-\psi^{L}_{a}A^{ab}\Psi^{U}_{b}\nonumber \\
&&\partial_{a} =\partial_{a}^{U}+\partial_{a}^{L} \nonumber \\
&&\partial_{a}^{U}\psi^{U}_{a}=1,\;\partial_{a}^{L}\psi^{L}_{a}=1\label{wj4}
\end{eqnarray}
where the indices $L$ and $U$ refer to lower and upper indices. On the other hand, gauge fields and spinors have a directional relation with curvature:
\begin{eqnarray}
&&A^{ab}=g^{ab}=h^{ab}=h_{1}^{ab}\otimes h_{2}^{ab}-h_{2}^{ab}\otimes h_{1}^{ab} ~ ~ and ~ ~ a,b,c=\mu,\nu,\lambda
\Rightarrow \nonumber\\&& F_{abc}=\partial_{a} A_{bc}-\partial_{b}
A_{ca}+\partial_{c}
A_{ab}=2(\partial_{\mu}g_{\nu\lambda}+\partial_{\nu}g_{\mu\lambda}-\partial_{\lambda}g_{\mu\nu})=
2\Gamma_{\mu\nu\lambda}\nonumber\\&&\nonumber\\&&\langle
F^{\rho}\smallskip_{\sigma\lambda},F^{\lambda}\smallskip_{\mu\nu}\rangle=\nonumber\\&&\partial_{\nu}
\Gamma^{\rho}_{\sigma\mu}-\partial_{\mu}\Gamma^{\rho}_{\sigma\nu}+\Gamma^{\rho}_{\lambda\nu}
\Gamma^{\lambda}_{\sigma\mu}-\Gamma^{\rho}_{\lambda\mu}\Gamma^{\lambda}_{\sigma\nu}
=R^{\rho}_{\sigma\mu\nu}\nonumber\\&&\langle
F_{abc},F_{a'}^{bc}\rangle=R_{aa'}^{anti-parallel}-R_{aa'}^{parallel}\nonumber\\&& R_{MN}=R_{aa'}+R_{ia'}+R_{ij'}=R_{Free-Free}^{anti-parallel}+R_{Free-Bound}^{anti-parallel}+R_{Bound-Bound}^{anti-parallel}-\nonumber\\&&R_{Free-Free}^{parallel}-R_{Free-Bound}^{parallel}-R_{Bound-Bound}^{parallel}\label{wj7}
\end{eqnarray}
and
\begin{eqnarray}
&&\langle \partial^{b}\partial^{a}\phi,\partial_{b}\partial_{a}\phi\rangle = \Big(\Psi^{\dag a,U}R_{aa'}^{anti-parallel} \Psi^{a',L}+\Psi^{\dag a,L}R_{aa'}^{anti-parallel}\Psi^{a',U}-\nonumber \\
&&\Psi^{\dag a,U}R_{aa'}^{parallel}\Psi^{a',U}-\Psi^{\dag a,L}R_{aa'}^{parallel} \Psi^{a',L}+\nonumber \\
&&\Psi^{\dag a,L}\Psi^{\dag d,U}\partial_{d}\partial^{d'}(R_{aa'}^{parallel}+R_{aa'}^{anti-parallel})\Psi^{a',L}\Psi^{U}_{d'}-\nonumber \\
&&\Psi^{\dag a,L}\Psi^{\dag d,U}\partial_{d}(R_{aa'}^{parallel}+R_{aa'}^{anti-parallel}) \Psi^{a',L}-\nonumber \\
&&\Psi^{\dag a,U}\Psi^{\dag d,L}\partial_{d}(R_{aa'}^{parallel}+R_{aa'}^{anti-parallel})\Psi^{a',U}\Big)+\nonumber \\
&& \Big(\psi^{\dag i,U}R_{ii'}^{anti-parallel} \psi^{i',L}+\psi^{\dag i,L}R_{ii'}^{anti-parallel}\psi^{i',U}-\nonumber \\
&&\Big(\psi^{\dag i,U}R_{ii'}^{parallel}\psi^{i',U}-\psi^{\dag i,L}R_{ii'}^{parallel}\psi^{i',L}+\nonumber \\
&&\psi^{\dag i,L}\psi^{\dag m,U}\partial_{m}\partial^{m'}(R_{ij'}^{parallel}+R_{ij'}^{anti-parallel})\psi^{i',L}\psi^{U}_{m'}-\nonumber \\
&&\psi^{\dag i,L}\psi^{\dag m,U}\partial_{m}(R_{ij'}^{parallel}+R_{ij'}^{anti-parallel})\psi^{i',L}-\nonumber \\
&&\psi^{\dag i,U}\psi^{\dag m,L}\partial_{m}(R_{ij'}^{parallel}+R_{ij'}^{anti-parallel})\psi^{i',U}\Big)+\nonumber \\
&& \Big(\Psi^{\dag a,U}R_{ai'}^{anti-parallel}\psi^{i',L}+\Psi^{\dag a,L}R_{ai'}^{anti-parallel}\rangle\psi^{i',U}-\nonumber \\
&&\Psi^{\dag a,U}R_{ai'}^{parallel}\psi^{i',U}-\Psi^{\dag a,L}R_{ai'}^{parallel}\psi^{i',L}+\nonumber \\
&&\Psi^{\dag a,L}\Psi^{\dag d,U}\partial_{d}\partial^{i'}(R_{ai'}^{parallel}+R_{ai'}^{anti-parallel})\psi_{j'}^{L}\psi^{i',U}-\nonumber \\
&&\Psi^{\dag a,L}\Psi^{\dag d,U}\partial_{d}(R_{ai'}^{parallel}+R_{ai'}^{anti-parallel})\psi^{i',L}-\nonumber \\
&&\Psi^{\dag a,U}\Psi^{\dag d,L}\partial_{d}(R_{ai'}^{parallel}+R_{ai'}^{anti-parallel})\psi^{i',U}\Big)\approx \nonumber \\
&&( R_{Free-Free}^{parallel})^{2}+( R_{Free-Free}^{anti-parallel})^{2}+( R_{Free-Bound}^{parallel})^{2}+\nonumber \\&&( R_{Free-Bound}^{anti-parallel})^{2}+( R_{Bound-Bound}^{parallel})^{2}+( R_{Bound-Bound}^{anti-parallel})^{2}+\nonumber \\&& (R_{Free-Free}^{parallel} R_{Free-Free}^{anti-parallel})\partial^{2}(R_{Free-Free}^{parallel}+ R_{Free-Free}^{anti-parallel} )+\nonumber \\&&(R_{Free-Bound}^{parallel} R_{Free-Bound}^{anti-parallel})\partial^{2}(R_{Free-Bound}^{parallel}+ R_{Free-Bound}^{anti-parallel} )+\nonumber \\&&(R_{Bound-Bound}^{parallel} R_{Bound-Bound}^{anti-parallel})\partial^{2}(R_{Bound-Bound}^{parallel}+ R_{Bound-Bound}^{anti-parallel} )\label{wj8}
\end{eqnarray}
where the index $Bound-Bound$ denotes the interaction between two bound electrons, $Free-Bound$ refers to the interaction of one free electron with one bound electron and $Free-Free$ denotes the interaction between two free electrons. Free electrons are electrons that aren't bound to specific atoms, while bound electrons are electrons that are bound to specific atoms. Substituting equations (\ref{wj2},\ref{wj4},\ref{wj7},\ref{wj8}) in equation (\ref{spm29}), putting $N=4$ and assuming that all corners are the same, we obtain the following result for one base \cite{R7,R8}:
\begin{eqnarray}
&& E_{system}=\int d^{4}y \times\nonumber \\&&
\Biggl[\sqrt{-g}\Bigl(\Big(-(1-(\frac{1}{ \tilde{a}_{\alpha_{n'}}^{2}})^{2})
[( R_{Free-Free}^{parallel})^{2}+( R_{Free-Free}^{anti-parallel})^{2}+( R_{Free-Bound}^{parallel})^{2}+\nonumber \\&&( R_{Free-Bound}^{anti-parallel})^{2}+( R_{Bound-Bound}^{parallel})^{2}+( R_{Bound-Bound}^{anti-parallel})^{2}+\nonumber \\&& (R_{Free-Free}^{parallel} R_{Free-Free}^{anti-parallel})\partial^{2}(R_{Free-Free}^{parallel}+ R_{Free-Free}^{anti-parallel} )+\nonumber \\&&(R_{Free-Bound}^{parallel} R_{Free-Bound}^{anti-parallel})\partial^{2}(R_{Free-Bound}^{parallel}+ R_{Free-Bound}^{anti-parallel} )+\nonumber \\&&(R_{Bound-Bound}^{parallel} R_{Bound-Bound}^{anti-parallel})\partial^{2}(R_{Bound-Bound}^{parallel}+ R_{Bound-Bound}^{anti-parallel} )]+\nonumber\\&&
(\frac{1}{ \tilde{a}_{\alpha_{n'}}^{2}})^{2}\lambda^{2}\delta_{\rho_{1}\sigma_{1}}^{\mu_{1}\nu_{1}}
((R^{anti-parallel,\rho_{1}\sigma_{1}}_{Free-Free,\mu_{1}\nu_{1}}+R^{anti-parallel,\rho_{1}\sigma_{1}}_{Bound-Bound,\mu_{1}\nu_{1}}+R^{anti-parallel,\rho_{1}\sigma_{1}}_{Free-Bound,\mu_{1}\nu_{1}})-\nonumber\\&&(R^{parallel,\rho_{1}\sigma_{1}}_{Free-Free,\mu_{1}\nu_{1}}+R^{parallel,\rho_{1}\sigma_{1}}_{Bound-Bound,\mu_{1}\nu_{1}}+R^{parallel,\rho_{1}\sigma_{1}}_{Free-Bound,\mu_{1}\nu_{1}})) \Bigr)\Big)\label{wj10}
\end{eqnarray}
Above equation shows that the energy of bases in each DNA depends on the curvature, which is created by the interaction of two bound electrons, a free and a bound electron and also two free electrons. For the small size of a DNA molecule, this curvature is very strong and is a signature of the communication of bases with very distant bases in other DNA molecules.
\section{Modeling the electrical circuit of a DNA }\label{o1}
In this section, we show that the structure of a DNA molecule is very similar to the structure of a radio receiver. A simple radio receiver contains an antenna, an inductor or a coil, a capacitor, a detector or a diode and a headphone (see Figure 6). A diode is a two-terminal electronic component that conducts primarily in one direction; it has low (ideally zero) resistance to the current in one direction, and high (ideally infinite) resistance in the other direction. A capacitor stores energy in the electric field (E) between its plates. The amount of this energy depends on the voltage between its plates. An inductor stores energy in its magnetic field (B). The amount of this energy depends on the current that passes through it. The inductor and the capacitor together form a tuned or resonant circuit. This circuit can act as an electrical resonator, which stores energy. This energy is oscillating at the circuit's resonant frequency. The resonant frequency of this circuit is $f_{0}=\frac{1}{2\pi \sqrt{LC}}$ where $L$ is the inductance and $C$ is the capacitance. At this frequency, the energy oscillates back and forth between the capacitor and the inductor.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{1.jpg}
\end{tabular}
\end{center}
\caption{ The circuit of a simple radio receiver \cite{R1} }
\end{figure*}
We now compare the structure of a DNA molecule with the circuit of a radio receiver. To this aim, we begin with the structure of base pairs in DNA. It is clear that each A (Adenine)-T(Thymine) base pair contains two electric dipoles between adenine and thymine, which have opposite directions with respect to each other. Also, each G (Guanine)-C (Cytosine) pair has three electric dipoles between the guanine and cytosine bases, and the direction of one of them is opposite with respect to the direction of the two other dipoles \cite{R2}. The electric field of this dipole is:
\begin{eqnarray}
&& d\sim 2\times 10^{-9} \quad q=\delta \times 1.6 \times 10^{-19} \quad r=\frac{d}{2} \quad \delta \approx 10^{-3} \nonumber\\&& E\sim \frac{\delta qd}{4\pi\varepsilon r^{3}}\sim 1.152 \times 10^{8} \label{D1}
\end{eqnarray}
where $d$ is the separation distance between positive and negative charges within a base pair, $q$ is the electric charge and $r$ is the place of one point between two bases. The base pair has a length of 1--3 nanometers \cite{R3}. This results in a large amount of energy stored in the electric field of dipoles between two bases and thus the dipoles act like a capacitor. In AT base pairs, two capacitors act approximately in opposition to each other and the total electric field is zero (see Figure 7). However, in a GC base pair, one capacitor acts in opposition to two capacitors and the total electric field is not zero (see Figure 8). On the other hand, these strong electric fields conduct electrons in one direction and prevent their motion in the other direction and thus sometimes, base pairs act as a diode.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{2.jpg}
\end{tabular}
\end{center}
\caption{ The A-T base pair contains two capacitors that act in opposition to each other. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{3.jpg}
\end{tabular}
\end{center}
\caption{ The G-C base pair contains three capacitors that act in opposition to each other. }
\end{figure*}
Each DNA molecule contains two biopolymer strands, which are coiled around each other to build a double helix. The nitrogenous bases of the two separate strands are bounded together, according to base pairing rules (A with T, and C with G), with hydrogen bounds to make a double-stranded DNA molecule. Each base pair contains two or three electric dipoles, which produce an strong electric field along the coil. The electric fields of all bases are summed with each other and create a pure electric field along the coil. This electric field induces a force on electrons and creates a current along the coil of DNA. This leads to the emergence of an inductor in a DNA molecule (see Figure 9).
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{4.jpg}
\end{tabular}
\end{center}
\caption{ Dipoles between bases in a DNA molecule form an inductor (coil). }
\end{figure*}
The length of a DNA is much greater than the size of the nucleus and for this reason, the DNA has to be packed tightly in a chromosome. The degree to which DNA is compressed is called its packing ratio \cite{R4,R5}. In a chromosome, this packing occurs during several stages. In the first stage of packing, DNA winds around a protein core to produce a ``bead-like'' structure termed a nucleosome. This results in a packing ratio of about 6. The second stage of packing is the coiling of beads in a helical structure called the 30 nm fiber that is found in all types of chromosomes. This structure increases the packing ratio to about 40. The final packaging happens when the fiber is organized in loops, scaffolds and domains that give a final packing ratio of about 1000 to 10000 in different types of chromosomes. During these stages, four types of inductor or coil appear. One coil emerges because of the electric dipoles in the structure of a DNA. A second type is produced by coiling DNA bases around the histone in a nucleosome (see Figure 10). A third type is created by the formation of loops in a chromatin fiber (see Figure 11) and finally, a fourth type of inductor emerges along the supercoil within a chromosome (see Figure 12). Each of these inductor types produces one type of magnetic field and plays a main role in a resonant circuit.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{C1.jpg}
\end{tabular}
\end{center}
\caption{ Formation of an inductor or a coil along the nucleosome with two magnetic fields. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{C2.jpg}
\end{tabular}
\end{center}
\caption{ Formation of an inductor or a coil along the chromatin fiber with three magnetic fields.}
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{C3.jpg}
\end{tabular}
\end{center}
\caption{ Formation of an inductor or a coil along the supercoil with four magnetic fields.}
\end{figure*}
A chromosome has different parts, like a telomere, a satellite, a centromere and arms \cite{R6}. A telomere is the region of DNA at the end of the linear eukaryotic chromosome that is required for the replication and stability of the chromosome. This region is one of the main parts of a chromosome that has a direct effect on sending and receiving electromagnetic waves. For a human, the sequence of nucleotides in telomeres is TTAGGG, with the complementary DNA strand being AATCCC, with a single-stranded TTAGGG overhang. In a telomere, each guanine has at least two positive and one negative induced charges. As we previously considered, a dipole can form between positive and negative charges and a capacitor emerges. Also, extra positive induced charges may have the role of free charges which move along the antenna (see Figure 13). If two guanine bases join each other, two extra positive charges lead to the emergence of electrical resistance between two capacitors (see Figure 14). In a thymine base, there is only one positive and one negative induced charge. These charges produce strong electric fields in one direction and induce a force to charges to move in this direction and prevent their motion in the opposite direction, just like a diode (see Figure 15). If we put electric devices that are similar to the bases of a telomore near each other, we can build a part of radio circuit like figure 16.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=12cm]{C4.jpg}
\end{tabular}
\end{center}
\caption{ Guanine in a telomere may play the role of capacitor + antenna in a circuit. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=16cm]{C5.jpg}
\end{tabular}
\end{center}
\caption{ Two guanine bases in a telomere may play the role of capacitor + resistance + capacitor in a circuit. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=10cm]{C6.jpg}
\end{tabular}
\end{center}
\caption{ Thymine in a telomere may play the role of a diode in a circuit. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=16cm]{C7.jpg}
\end{tabular}
\end{center}
\caption{ Modeling the electrical circuit of a telomere.}
\end{figure*}
If we replace bases with electrical devices like inductors and capacitors, we can construct at least four tuned or resonant circuits. Thus, each DNA has at least four resonant frequencies and many diodes, resistors and antennae (Figure 17 gives only an approximate representation of a DNA molecule when its bases are replaced with electric devices). This circuit is very similar to the circuit of an FM radio. In each DNA, there are some genes that have been constructed from a special number of AT and CG-base pairs. However, in each cell, all genes aren't active and thus, the number of genes and base pairs that are playing the role in electrical circuit in each chromosome is different from chromosomes in other cells. This causes the resonant frequencies of chromosomes in each cell to be different. This difference ensures that each chromosome receives only special electromagnetic messages from the brain, neurons, other cells and proteins. Also, each chromosome can emit special electromagnetic signals to other cells and inform them of its own activity. This long distance communication offers an opportunity for all cells in a human body to be in unison with each other.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=16cm]{C8.jpg}
\end{tabular}
\end{center}
\caption{ Modeling a part of the electrical circuit of DNA when bases are replaced by electric devices. }
\end{figure*}
\section{Damaged DNA, teleportation and cancer}\label{o2}
When a DNA molecule is damaged, its electric charges change and its radiated electromagnetic fields become different when compared to the signals of a normal DNA molecule. For example, if one atom is added to one of bases in a DNA molecule, this base becomes a heptagonal molecule. Also, if a DNA molecule loses one of its atoms, the hexagonal shape of its base changes and a pentagonal molecule is created (see Figures 18, 19 and 20). Previously, for graphene, it has been shown that the properties of a pentagonal molecule are different from hexagonal and heptagonal ones. For example, a pentagonal molecule absorbs electrons, while a heptagonal molecule repels them \cite{R7, R8}.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{C9.jpg}
\end{tabular}
\end{center}
\caption{ The hexagonal shape of a molecule like molecules of bases in a DNA or molecules of a graphene. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{C10.jpg}
\end{tabular}
\end{center}
\caption{ The heptagonal shape of a molecule like molecules of bases in a DNA when one atom is added to them. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{C11.jpg}
\end{tabular}
\end{center}
\caption{ The pentagonal shape of a molecule like molecules of bases in a DNA when lose one atom. }
\end{figure*}
In some recent research \cite{R7,R8}, the electric current for each of these molecules in graphene has been calculated. In these models, it has been asserted that the origin of gravitational waves and electromagnetic waves are the same and for this reason, all currents have been given in terms of curvature that is produced between electrons \cite{R7,R8}:
\begin{eqnarray}
&& J \approx \nonumber \\&&\sqrt{-g}\Bigl(\Big(-(1-m_{g}^{2})
[( R_{Free-Free}^{parallel})^{2}-( R_{Free-Free}^{anti-parallel})^{2}+( R_{Free-Bound}^{parallel})^{2}-\nonumber \\&&( R_{Free-Bound}^{anti-parallel})^{2}+( R_{Bound-Bound}^{parallel})^{2}-( R_{Bound-Bound}^{anti-parallel})^{2}+\nonumber \\&& (R_{Free-Free}^{parallel} R_{Free-Free}^{anti-parallel})\partial^{2}(R_{Free-Free}^{parallel}- R_{Free-Free}^{anti-parallel} )+\nonumber \\&&(R_{Free-Bound}^{parallel} R_{Free-Bound}^{anti-parallel})\partial^{2}(R_{Free-Bound}^{parallel}- R_{Free-Bound}^{anti-parallel} )+\nonumber \\&&(R_{Bound-Bound}^{parallel} R_{Bound-Bound}^{anti-parallel})\partial^{2}(R_{Bound-Bound}^{parallel}- R_{Bound-Bound}^{anti-parallel} )]+\nonumber\\&&
m_{g}^{2}\lambda^{2}\delta_{\rho_{1}\sigma_{1}}^{\mu_{1}\nu_{1}}
((R^{anti-parallel,\rho_{1}\sigma_{1}}_{Free-Free,\mu_{1}\nu_{1}}+R^{anti-parallel,\rho_{1}\sigma_{1}}_{Bound-Bound,\mu_{1}\nu_{1}}+R^{anti-parallel,\rho_{1}\sigma_{1}}_{Free-Bound,\mu_{1}\nu_{1}})+\nonumber\\&&(R^{parallel,\rho_{1}\sigma_{1}}_{Free-Free,\mu_{1}\nu_{1}}+R^{parallel,\rho_{1}\sigma_{1}}_{Bound-Bound,\mu_{1}\nu_{1}}+R^{parallel,\rho_{1}\sigma_{1}}_{Free-Bound,\mu_{1}\nu_{1}})) \Bigr)\Big) \label{w1}
\end{eqnarray}
where $R_{Free-Free}^{parallel}$ is the curvature that is produced between two parallel free electrons in a DNA molecule, $R_{Free-Free}^{anti-parallel}$ is the curvature that is created between two anti-parallel free electrons, $R_{Free-Free}^{parallel}$ is the curvature that emerges between two parallel free electrons in a DNA molecule, $R_{Bound-Bound}^{parallel}$ is the curvature that is produced between two parallel bound electrons to atoms in a DNA molecule and $R_{Bound-Bound}^{anti-parallel}$ is the curvature that is created between two anti-parallel bound electrons to atoms in a DNA molecule. Also, $m_{g}$ is the graviton mass and $\lambda\sim \frac{1}{m_{g}}$. In additional, curvature is obtained from \cite{R7,R8}:
\begin{eqnarray}
&& R_{Free/Bound-Free/Bound}^{anti-parallel}=l_{1}^{1-1}+l_{1}^{1-2}+l_{1}^{1-3}+\nonumber\\&&
(l'_{1})^{1-1}+(l'_{1})^{1-2}+(l'_{1})^{1-3}\nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{parallel}=l_{2}^{1-1}+l_{2}^{1-2}+l_{2}^{1-3}-\nonumber\\&&
(l'_{2})^{1-1}-(l'_{2})^{1-2}-(l'_{2})^{1-3}\label{w2}
\end{eqnarray}
where $l_{1}$ is the coupling between two anti-parallel spins and $l_{2}$ is the coupling between parallel spins.
For the corner of carbon-oxygen in a thymine, with hexagonal shape, we can measure these couplings in terms of the angle between axes in a molecule and electric charges of atoms (see Figure 15 and Figure 18):
\begin{eqnarray}
&& l_{1}^{1-1}\approx 2 q_{0}q_{1}\cos(0)=2 q_{0}q_{1} \quad l_{1}^{1-2}\approx q_{0}q_{2}\cos(120)=-\frac{1}{2}q_{0}q_{2} \quad l_{1}^{1-3}\approx q_{0}q_{3}\cos(240)=-\frac{1}{2}q_{0}q_{3}\nonumber\\&& (l'_{1})^{1-1}\approx \sin(0)=0 \quad (l'_{1})^{1-2}\approx q_{0}q_{2}\sin(120)=\frac{\sqrt{3}}{2}q_{0}q_{2} \quad (l'_{1})^{1-3}\approx q_{0}q_{3}\sin(240)=-\frac{\sqrt{3}}{2}q_{0}q_{3}\nonumber\\&& \nonumber\\&& l_{2}^{1-1}\approx q_{0}q_{1} \cos(60)=\frac{1}{2}q_{0}q_{1} \quad l_{2}^{1-2}\approx q_{0}q_{2}\cos(180)=-q_{0}q_{2} \quad l_{2}^{1-3}\approx q_{0}q_{3}\cos(300)=\frac{1}{2}q_{0}q_{3}\nonumber\\&& (l'_{2})^{1-1}\approx q_{0}q_{1}\sin(60)=\frac{\sqrt{3}}{2}q_{0}q_{1} \quad (l'_{2})^{1-2}\approx q_{0}q_{2}\sin(180)=0 \quad (l'_{2})^{1-3}\approx q_{0}q_{3}\sin(300)=-\frac{\sqrt{3}}{2}q_{0}q_{3}\label{w3}
\end{eqnarray}
where $q_{0}$ is the charge of the electron, $q_{1}$ is the charge of the oxygen atom, $q_{2}$ is the charge of the carbon atom, $q_{3}$ is the charge of the nitrogen atom. We have assumed that half of electrons are parallel with each other and the other half are antiparallel with each other. Substituting these values in equation (\ref{w2}), we can calculate the curvature for parallel and anti-parallel spins:
\begin{eqnarray}
&& R_{Free/Bound-Free/Bound}^{anti-parallel}=l_{1}^{1-1}+l_{1}^{1-2}+l_{1}^{1-3}+\nonumber\\&&
(l'_{1})^{1-1}+(l'_{1})^{1-2}+(l'_{1})^{1-3}=\nonumber\\&& 2 q_{0}q_{1} -\frac{1}{2}q_{0}q_{2}-\frac{1}{2}q_{0}q_{3} +\frac{\sqrt{3}}{2}q_{0}q_{2}-\frac{\sqrt{3}}{2}q_{0}q_{3}\nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{parallel}=l_{2}^{1-1}+l_{2}^{1-2}+l_{2}^{1-3}-\nonumber\\&&
(l'_{2})^{1-1}-(l'_{2})^{1-2}-(l'_{2})^{1-3}=\nonumber\\&& \frac{1}{2}q_{0}q_{1}-q_{0}q_{2}+\frac{1}{2}q_{0}q_{3}-\frac{\sqrt{3}}{2}q_{0}q_{1}+\frac{\sqrt{3}}{2}q_{0}q_{3} \nonumber\\&& \label{w4}
\end{eqnarray}
By substituting the density of electrons for oxygen, nitrogen and carbon in equations (\ref{w1} and \ref{w4}), we can obtain the curvatures and current density:
\begin{eqnarray}
&& q_{0}=\frac{1.6 \times 10^{-19}}{V_{Atom}} \quad q_{1}= \frac{8 q_{0}}{V_{Atom}} \quad q_{2}= \frac{6 q_{0}}{V_{Atom}} \quad q_{3}=\frac{7 q_{0}}{V_{Atom}}\nonumber\\&& V_{Atom}=\frac{4\pi}{3} R^{3} \quad R\sim 10^{-10} \quad m_{g}\sim 0 \nonumber\\&& \nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{anti-parallel}=88.576 \times 10^{22} \nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{parallel}=5.22496 \times 10^{22} \nonumber\\&& \nonumber\\&&\nonumber\\&& J_{carbon-oxygen}\sim 7.818 \times 10^{48} \label{w5}
\end{eqnarray}
This result shows that the current density is positive and thus, free electrons are repelled by this corner of thymine. In parallel, we can calculate the current density in the corner of nitrogen-hydrogen:
\begin{eqnarray}
&& J_{Nitrogen-Hydrogen}\sim -J_{carbon-oxygen}\sim -7.818 \times 10^{48} \label{w6}
\end{eqnarray}
This equation indicates that the amount of current that goes outward of a thymine is equal to the amount of current that enters into a thymine and the system is stable.
We now assume that a thymine base loses one of its atoms (For example a carbon atom) and assumes a pentagonal shape. For carbon-oxygen corner in a thymine base, with hexagonal shape, we can repeat our calculations and obtain couplings as a function of the angle between axes in a molecule and electric charges of atoms (see Figures 15 and 20):
\begin{eqnarray}
&& l_{1}^{1-1}\approx 2q_{0}q_{1} \cos(0)=2q_{0}q_{1} \quad l_{1}^{1-2}\approx q_{0}q_{3}\cos(126)=-0.587q_{0}q_{3} \quad l_{1}^{1-3}\approx q_{0}q_{3}\cos(234)=-.587q_{0}q_{3}\nonumber\\&& (l'_{1})^{1-1}\approx \sin(0)=0 \quad (l'_{1})^{1-2}\approx q_{0}q_{3}\sin(126)=0.809q_{0}q_{3} \quad (l'_{1})^{1-3}\approx q_{0}q_{3}\sin(234)=-0.809q_{0}q_{3}\nonumber\\&& \nonumber\\&& l_{2}^{1-1}\approx 2q_{0}q_{1}\cos(60)=q_{0}q_{1} \quad l_{2}^{1-2}\approx q_{0}q_{3}\cos(186)=-0.994q_{0}q_{3} \quad l_{2}^{1-3}\approx q_{0}q_{3}\cos(294)=0.406 q_{0}q_{3}\nonumber\\&& (l'_{2})^{1-1}\approx 2q_{0}q_{1}\sin(60)=1.732q_{0}q_{1} \quad (l'_{2})^{1-2}\approx q_{0}q_{3}\sin(186)=-0.104q_{0}q_{3}\quad (l'_{2})^{1-3}\approx q_{0}q_{3}\sin(294)=-0.913q_{0}q_{3}\label{w7}
\end{eqnarray}
Substituting these values in equation (\ref{w2}), we can calculate the curvature for parallel and anti-parallel spins:
\begin{eqnarray}
&& R_{Free/Bound-Free/Bound}^{anti-parallel}=l_{1}^{1-1}+l_{1}^{1-2}+l_{1}^{1-3}+\nonumber\\&&
(l'_{1})^{1-1}+(l'_{1})^{1-2}+(l'_{1})^{1-3}=\nonumber\\&& 2q_{0}q_{1} -0.587q_{0}q_{3}-.587q_{0}q_{3} \nonumber\\&& +0.809q_{0}q_{3} -0.809q_{0}q_{3} \nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{parallel}=l_{2}^{1-1}+l_{2}^{1-2}+l_{2}^{1-3}-\nonumber\\&&
(l'_{2})^{1-1}-(l'_{2})^{1-2}-(l'_{2})^{1-3}=\nonumber\\&& q_{0}q_{1} -0.994q_{0}q_{3}+0.406 q_{0}q_{3} \nonumber\\&& - 1.732q_{0}q_{1} +0.104q_{0}q_{3}+0.913q_{0}q_{3} \label{w8}
\end{eqnarray}
Again, by substituting the density of electrons for oxygen, nitrogen and carbon atoms in equations (\ref{w1} and \ref{w8}), we can derive the curvature and current density:
\begin{eqnarray}
&& q_{0}=\frac{1.6 \times 10^{-19}}{V_{Atom}} \quad q_{1}= \frac{8 q_{0}}{V_{Atom}} \quad q_{3}=\frac{7 q_{0}}{V_{Atom}}\nonumber\\&& V_{Atom}=\frac{4\pi}{3} R^{3} \quad R\sim 10^{-10} \quad m_{g}\sim 0 \nonumber\\&& \nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{anti-parallel}=4.98048 \times 10^{22} \nonumber\\&&\nonumber\\&& R_{Free/Bound-Free/Bound}^{parallel}=-1.82692 \times 10^{22} \nonumber\\&& \nonumber\\&&\nonumber\\&& J_{carbon-oxygen}\sim 21.4747833 \times 10^{44} \label{w5}
\end{eqnarray}
This current density is much smaller than the current density that is produced by this corner (carbon-oxygen) in a normal DNA , however the neighboring atoms at other corners (nitrogen-hydrogen) aren't changed and thus, they produce the previous current density:
\begin{eqnarray}
&& J_{Nitrogen-Hydrogen}\sim -7.818 \times 10^{48}\nonumber\\&& \nonumber\\&&\nonumber\\&& J_{induced}\sim J_{Nitrogen-Hydrogen}-J_{carbon-oxygen}\sim 7.818 \times 10^{48}-21.4747833 \times 10^{44}\neq 0 \label{w6}
\end{eqnarray}
This extra current density produces many problems not only for this DNA molecule and the cell that contains it but also for other DNAs. In fact, a change in the current of this DNA occurs and some extra signals are produced. These signals are received by inductors in other DNA molecules and an extra current density is created that leads to the destruction of other DNA molecules (see Figure 21). This event may be a reason for the production and duplication of cancer cells.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{C12.jpg}
\end{tabular}
\end{center}
\caption{ The emitted signals by the inductor of one DNA can be detected by the inductor in other DNA. }
\end{figure*}
Now, we can suggest a mechanism for preventing of the progression of some diseases like cancer. If we add a third inductor to this system with properties that are the same as those of inductors of DNA and produce a current in the opposite direction with respect to the current of the damaged DNA, an extra magnetic field emerges. This field cancels the effects of the magnetic field that is induced by the damaged DNA in normal DNA and prevents its destruction (see Figure 22). However, this is not a very simple method. To accomplish this, we need to know the exact form of the circuit in a DNA and design an exact circuit for communication with it.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{C13.jpg}
\end{tabular}
\end{center}
\caption{ The emitted signals by the inductor of a damaged DNA can be canceled by the artificial inductor . }
\end{figure*}
\section{The difference between radiated waves from DNA molecules in men and women}\label{o3}
First, we design a simple circuit that measures radiated waves by DNA molecules of a man and a woman. This circuit contains a ferromagnet like an iron core, with two antiparallel magnetic fields are applied to it. These magnetic fields cause two regions to appear, with the spin of electrons in each of them parallel with each other and antiparallel with the spin of electrons in the other region (see Figure 23). This ferromagnet is connected to a wire, resistance and ammeter. Our ammeter can measure voltages, too. Under normal conditions, electrons with upper spin in one region are attracted by electrons with lower spin in another region and the total current in the circuit becomes zero. If a single man comes close to this system, a negative current will be measured by the ammater, whereas the proximity of a single woman to it induces a positive current.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{f2.jpg}
\end{tabular}
\end{center}
\caption{ The circuit of a simple circuit for the measurement of the difference between radiated waves from DNAs in men and women }
\end{figure*}
To discover the origin of the difference between radiated signals from man and women, we consider the difference in their chromosomes. Each human cell contains 23 pairs of chromosomes (22 pairs of autosomes or body chromosome(s) and one pair of sex chromosomes). This gives 46 chromosomes in total. Autosomes emerge in pairs, whose members have the same form but differ from other pairs in a human cell. To have a stable system, it is necessary for the magnetic field in each member of one body chromosome or autosome to be equal in magnitude but opposite in direction to the magnetic field in the other member, such that they form a pair without any magnetic field (see Figure 24).
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=5cm]{f11.jpg}
\end{tabular}
\end{center}
\caption{ The magnetic field of each member in an autosome pair is in the opposite direction to the magnetic field of the other member. }
\end{figure*}
Each cell contains one pair of sex chromosomes, one X and one Y chromosome in men and two X chromosomes in women. One of X chromosomes in a man's cell is inactive. Also, the Y chromosome in a woman's cell with 59 million base pairs of DNA is shorter of the X chromosome in a man's cell with 155 million base pairs. These differences lead to the breakdown of stability in pairs and the radiation of two different signals. In the XX pair of a woman's cell, the magnetic field of one of chromosomes is turned off and thus, this pair radiates a magnetic field (see Figure 25). In the XY pair of a man's cell, the magnetic field of the X chromosome cannot be canceled by the magnetic field of the Y chromosome and a magnetic field is created in a direction that is opposite to that of the XX pair in a woman's cell (see Figure 26).
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{f12.jpg}
\end{tabular}
\end{center}
\caption{ In the XX pair of a woman's cell, the magnetic field of one of chromosomes is turned off and thus, this pair radiates a magnetic field. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{f13.jpg}
\end{tabular}
\end{center}
\caption{In the XY pair of a man's cell, the magnetic field of the X chromosome cannot be canceled by the magnetic field of the Y chromosome and a magnetic field is created in a direction opposite to that of the XX pair in a woman's cell. }
\end{figure*}
When a DNA molecule is damaged, an extra current appears along its inductor, which produces an extra electromagnetic field. This extra wave induces an extra current in other DNA molecules and leads to their destruction and disease progression. This event can be seen in some diseases like cancer. If we put a damaged DNA of a male or a female near the damaged DNA of a female or a male, their radiated waves cancel the effect of each other and disease progression is stopped (see Figure 27).
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{f14.jpg}
\end{tabular}
\end{center}
\caption{To remove extra signals of a damaged DNA in males, we can use of damaged DNA in females }
\end{figure*}
Another model to cure cancer is that we send some electromagnetic signals and force cancer cells to return to their initial stem cells. Then, by controlling their differentiating, we can produce some new types of non-cancer cells.
Another application of the difference between the radiated electromagnetic signals of men and women is the determination of the gender of a fetus in pregnant women. When the fetus is female, its radiated waves have the same sign and oscillation as the radiated waves of the mother and amplify her signals, while if the fetus is male, its radiated waves have the opposite sign and oscillation compared to those of the mother and thus they destroy her signal. This method helps us determine the gender of a fetus much sooner than normal methods like sonography. Also, this mechanism is more harmless when compared to sonography.
Another interesting result that comes out of these calculations is the difference in the electromagnetic signals of sperms and eggs. In fact, sperms that produce males have different signals respect to the sperms that produce females. This may help us to determine the gender of a fetus which is produced by the fusion of an special sperm with an special egg.
Dad cell and Mam cell are stable and constructed from chromosomes that their radiated waves cancel the effect of each other. By dividing Dad cell, two sperms are emerged that their radiated waves have opposite signs. Also, by dividing Mam cell, two eggs are produced that their waves have opposite signs. By joining an sperm and an egg with opposite waves, a diploid cell is created. It is clear that the created cell isn't stable and may emit waves with positive sign or negative sign. To determine the gender of a fetus, we can put two types of sperms near it. Sperms which radiates waves with the same sign of fetus, repel it and becomes far from it and sperms that radiates waves with opposite sign of fetus, become close to it. This method is more easier than normal mechanisms like sonography.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=10cm]{wire.jpg}
\end{tabular}
\end{center}
\caption{The Female Virus Wire (Superconductor) for Males and the Male Virus Wire (Superconductor) for Females. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=10cm]{VMI-blood.jpg}
\end{tabular}
\end{center}
\caption{The Virus Medical Imaging Technique: Viruses in blood take signals of viruses outside the blood in VMI machine and controll the process of differentiation. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{reverse.jpg}
\end{tabular}
\end{center}
\caption{The radiated waves of two cells in neighbour is in reverse and total wave is zero. For cancers, total wave isnot zero. }
\end{figure*}
We can generalize this discussion to viruses. We can show that viruses which are attracted by cells of a female have different signals respect to viruses which are absorbed by cells of a male. Although, type of these viruses are the same, however, packing of DNA or the location of chromosomes is different which leads to the radiation of two different signals. In fact, some of viruses are male and some others are female. On the other hand, the reason for the emergence of some cancers are viruses. Thus, we can remove the effects of viruses which lead to the cancer with viruses of opposite gender. Another interesting application of viruses is the constructing a wire that transmit information very fast like wormhole in cosmology or superconductor in nano. We can call them virus wormhole or virus superconductor (See Figure 28). These wires consist of different type of viruses and penetrate in human's body.
Another interesting point about viruses is their application as the receiver or sender of electromagnetic waves. We can use of some viruses as the wireless devises for communications with DNAs in various cells. Also, they can help us to communicate with specialized cells and force them to return to initial stem cell or even initial sperm and egg.
Above results help us to introduce two types of medical imaging techniques:1.Sperm Medical Imaging (SMI)2.Virus Medical Imaging (VMI). In Sperm Medical Imaging technique, we use of the communications between sperms and cells. In this method, by decoding the radiated signals of sperms and considering their evolutions, we can imagine the events interior of a human's body. In Virus Medical Imaging, we use of the communications between Viruses and cells interior of human's body. In this model, by considering the evolutions of viruses out of a body, we can analyse the events interior of a human's body.
To construct a device that communicate with DNAs of cells interior of human's body, we use of the MRI machin which it's protons are replaced by sperms or viruses. In Sperm/Virus Medical Imaging, signals from sperms/Viruses are received by some electronic devices or some special antennas and then applied to the patient. This leads to the occurrence of a communication between sperms/Viruses and cells. The DNAs in cells interior of human's body or tissue response to sperms and send radio frequency waves, which are taken by a radio receiver. These signals may be decoded and the exact information about the evolutions of cells interior of a body can be obtained (See Figure 29).
In human's body, the radiated waves of cells cancel the effect of each other and total radiation of body is approximately zero (See figure 30). If these waves couldn't remove the effect of each other, some extra information is transmitted between cells and cancer may be emerged. In these conditions, we can communicate with cancer cells via electromagnetic cells and force them to transform to stem cell. By this the problem of cancer may be solved.
One of interesting subject in biology is the ability of DNA for saving information. We can show that water has the same ability and we can define a coding for it like the genetic coding of DNA. Also, electromagnetic waves have an special type of coding and act similar to DNA or protein. When electromagnetic waves achieve to molecules of water, give their information to water and take it's information. The process is very the same of transcription and translation in biological systems. Sometimes, electromagnetic waves act like DNA and water acts like the protein or RNA and sometimes, reversely, water acts like the DNA and electromagnetic waves act like the protein. On the other hand, water and electromagnetic waves have special genders. For example, molecules of water attract waves with opposite gender.
In a DNA teleportation, two tubes or container of water become located near each other. The first tube emit some signals that contain some packages. These packages carry the information about structures which are formed in initial water. When these packages achieve to second tube of water, open and give their information to second water (See figure 31). We can write:
\begin{eqnarray}
&& H^{EM}= H^{tot}= \Sigma_{j=3}^{N-3}\Sigma_{i=1}^{N}P_{ij}H_{j,i}^{tot} = \nonumber\\ && \Sigma_{j=3}^{N-3}\Sigma_{i=1}^{N}EM_{j,i}^{tot}
\label{u4}
\end{eqnarray}
where $EM_{j,i}^{tot}$ is the package of information in an electromagnetic wave which produces j number of i-gonal structures with the Humiltonian of $H_{j,i}^{tot}$ and probability of $P_{ij}$. Molecules of water have trigonal shapes, however in a tube, they join each other and form other structures like hexagonal shapes. This type of writing for the Humiltonian of electromagnetic waves helps us to introduce a new information codding for waves.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{em.jpg}
\end{tabular}
\end{center}
\caption{ Coding of water and electromagnetic waves }
\end{figure*}
If we put DNA in initial tube or container of water, this water emit some electromagnetic waves that carry the information about the changes produced by the DNA in the structure of initial pure water . We can write:
\begin{eqnarray}
&& H^{EM, DNA-Water}= H^{tot, DNA-Water}= \Sigma_{j=3}^{N-3}\Sigma_{i=1}^{N}\Sigma_{k=1}^{M}P_{ij,k}H_{j,i,k}^{tot, DNA-Water} = \nonumber\\ && \Sigma_{j=3}^{N-3}\Sigma_{i=1}^{N}\Sigma_{k=1}^{M}EM_{j,i,k}^{tot, DNA-Water}
\label{sdu4}
\end{eqnarray}
where $EM_{j,i,k}^{tot, DNA-Water}
$ is the package of information in an electromagnetic wave which produces j number of i-gonal structures and k extra corners corresponded to k atoms of DNA. Also, $H_{j,i,k}^{tot, DNA-Water}$ is the Hamiltonian of these structures and $P_{ij,k}$ is the probability for producing them. When these packages of waves achieve to another pure water, form structures of initial water with some extra corners or defects which are the signature of the existence of DNA. For example, assume that a tube or container of water contains DNA molecules. These molecules change the hexagonal and pentagonal structures of water. Signals which are produced by this water contain packages which carry the information of structures of initial water. When these packages achieve to second tube of pure water, give their information to water, change structures and produce the same of structures of initial water. We can detect the structures in initial water by using the PCR or other methods.
We can use of the gender of water in medical imaging. First, we separate two waters one with the gender of male which interacts with men and another with the gender of female which interacts with women. If a woman eat water wit the gender of female,
all of it repel with her body. For a pregnant women with a fetus of male, some of this water are absorbed, while for a fetus of female, all of this water go out. This method is completely harmless respect to other methods like the Sonography. Another interesting pint is separating waves by using water. Water with the gender of male attract waves with the gender of female and water with the gender of male, attract waves with the gender of female. Using this property, we can analyse the events interior of human's body by considering the exchanged waves between waters inside calls and outside of body.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{topoisomerases.jpg}
\end{tabular}
\end{center}
\caption{Waves (Topoisomerases) which participate in the overwinding or underwinding of DNA in extra dimensions }
\end{figure*}
In additional to electromagnetic waves, DNAs emit some new types of waves which have the shape of topoisomerases. Topoisomerases are enzymes that participate in the overwinding or underwinding of DNA. We generalise this concept to waves which participate in the overwinding or underwinding of DNA in extra dimensions (See figure 32). These wave topoisomerases have 4 + n -dimensional shapes. This is because that DNA is very compacted in 3 + 1 dimensional space-time. However if we see topology of DNA in extra dimension, we can open packings of it and obtain a simpler shape. In fact DNA acts like a 4 + n-dimensional coil which has been packed to 3 + 1-dimensions. Only, waves could interact with DNA, which could move in 4 + n dimensions and see the un-packed structure of it in extra dimensions. These waves couldn't be classified as the electromagnetic or gravitational waves. However, by reducing dimensions to 4 or 3, they obtain some properties similar to electromagnetic waves.
When, a DNA is packed, some extra fields are emerged that play the role of extra photons in 4-dimensions. These fields change the action and Hamiltonian of DNAs and produce some new wormhole-like tunnels. For example, the interaction of two atoms can be written as:
\begin{eqnarray}
&& S_{atom-atom}=-T_{tri} \int d^{11}\sigma \sqrt{\eta^{ab}
g_{MN}\partial_{a}X^{M}\partial_{b}X^{N} +2\pi
l_{s}^{2}U(F)+ +2\pi
l_{s}^{2}\bar{U}(\bar{F})}\nonumber\\&&
U=(\sum_{n=1}^{Z}\frac{1}{n!}(-\frac{F_{1}..F_{n}}{\beta^{2}}))
\nonumber\\&&
\bar{U}=(\sum_{n=1}^{\bar{z}}\frac{1}{n!}(-\frac{\bar{F}_{1}..\bar{F}_{n}}{\beta^{2}}))
\nonumber\\&& F=F_{\mu\nu}F^{\mu\nu}\quad
F_{\mu\nu}=\partial_{\mu}A_{\nu}-
\partial_{\nu}A_{\mu}\label{7}
\end{eqnarray}
where z is the number of paired electrons, $\bar{z}$ is the number of photons which are produced by the packings of DNA, $g_{MN}$ is the background metric, $ X^{M}(\sigma^{a})$'s
are scalar fields which are constructed from paring two electrons with opposite spins, $\sigma^{a}$'s are
the manifold coordinates, $a, b = 0, 1, ..., 3$ are world-volume
indices of the manifold and $M,N=0, 1, ..., 11$ are eleven dimensional
spacetime indices. Also, $G$ is the nonlinear field
and $A$ is the photon which exchanges between
atoms. Also, $\bar{A}$ is the photon which is emerged by te packings of DNA. We can obtain:
\begin{eqnarray}
&&H_{atom-atom}=4\pi T_{tri}\int d\sigma_{11}... d\sigma_{1}[(1 + Q_{z,atom-atom}) (1 + \bar{Q}_{\bar{z},atom-atom})]\times\nonumber\\ && \sqrt{1+\eta^{ab}g_{MN}g_{MN}\psi^{M,\uparrow}_{atom1}\partial_{a}\psi^{M,\downarrow}_{atom2}\psi^{N,\uparrow\uparrow}_{atom2}\partial_{b}\psi^{N,\downarrow}_{atom1}}\times\nonumber\\ && \sqrt{1+\eta^{ab}g_{MN}g_{MN}\bar{\psi}^{M,\uparrow}_{atom1}\partial_{a}\bar{\psi}^{M,\downarrow}_{atom2}\bar{\psi}^{N,\uparrow\uparrow}_{atom2}\partial_{b}\bar{\psi}^{N,\downarrow}_{atom1}}
\nonumber\\ && \nonumber\\ &&
\bar{Q}_{\bar{z},atom-atom}=\bar{Q}_{\bar{z}-1,atom-atom}\sqrt{1+\frac{k^{2}_{2}}{\bar{Q}_{\bar{z}-1,atom-atom}\theta^{4}_{\bar{z},atom-atom}}}....\sqrt{1+\frac{k^{2}_{2}}{\bar{Q}_{1,atom-atom}\theta^{4}_{1,atom-atom}}}
\nonumber\\
&& \bar{Q}_{1,atom-atom}= \sqrt{1+\frac{k^{2}_{1}}{\theta^{4}_{1,atom-atom}}}
\nonumber\\ &&
Q_{z,atom-atom}=Q_{z-1,atom-atom}\sqrt{1+\frac{k^{2}_{2}}{Q_{z-1,atom-atom}\sigma^{4}_{z,atom-atom}}}....\sqrt{1+\frac{k^{2}_{2}}{Q_{1,atom-atom}\sigma^{4}_{1,atom-atom}}}
\nonumber\\
&& Q_{1,atom-atom}= \sqrt{1+\frac{k^{2}_{1}}{\sigma^{4}_{1,atom-atom}}}
\label{8}
\end{eqnarray}
where $\sigma$ is the separation distance between two atoms and $\theta$ is the angle of packings. Also, packing of DNA changes the couplings between fermions and produce some new couplings which are presented by $bar{\psi}$. Generalizing these calculations to all molecules of DNA and following the method in previous section, we can obtain:
\begin{eqnarray}
&&H_{DNA}=4\pi T_{tri}\int d\sigma_{11}... d\sigma_{1} Q_{DNA}E_{DNA}\nonumber\\ && Q_{DNA} =\Sigma_{M,N,X,Y=1}^{W} Q_{un-packed,DNA,N,M}\bar{Q}_{packed,DNA,X,Y}\nonumber\\ && E_{DNA} =\Sigma_{M,N,X,Y=1}^{W} E_{un-packed,DNA,N,M}\bar{E}_{packed,DNA,X,Y} \nonumber\\ && Q_{un-packed,DNA, M,N}= \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + Q_{6,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + Q_{5,i-j} \big)]\times \nonumber\\ && \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + Q_{6-5,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + Q_{5-6,i-j} \big)]\nonumber\\ &&E_{un-packed,DNA,M,N}= \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + E_{6,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + E_{5,i-j} \big)]\times \nonumber\\&& \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + E_{6-5,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + E_{5-6,i-j} \big)] \nonumber\\ && \bar{Q}_{packed,DNA,X,Y}= \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + \bar{Q}_{6,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + \bar{Q}_{5,i-j} \big)]\times \nonumber\\ && \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + \bar{Q}_{6-5,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + \bar{Q}_{5-6,i-j} \big)]\nonumber\\ && \bar{E}_{packed,DNA, X,Y}= \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + \bar{E}_{6,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + \bar{E}_{5,i-j} \big)]\times \nonumber\\&& \Pi_{n=1}^{N}[\Sigma_{i,j=1}^{n} \big(1 + \bar{E}_{6-5,i-j} \big)] \Pi_{m=1}^{M}[\Sigma_{l,k=1}^{m} \big(1 + \bar{E}_{5-6,i-j} \big)]
\label{9}
\end{eqnarray}
where $w = M + N + X + Y$ is the number of total fields which are emerged in a packed DNA. M, N are numbers of fields in un-packed DNA and X, Y are number of fields which are produced by packings of DNA. This system is very complicated and waves couldn't read it's information easily (See figure 4). They should open packings of DNA in extra dimensions and make it's topology simple. For this reason, it is needed that the waves give an appropriated energy to DNA, excited and open it's packings. In these conditions, total energy and topology system tend to a constant number. We can write:
\begin{eqnarray}
&& 1= H_{DNA} + H_{topoisomerase-like wave} = \nonumber\\ && 4\pi T_{tri}\int d\sigma_{11}... d\sigma_{1} [Q_{DNA}E_{DNA} + Q_{topoisomerase}E_{topoisomerase}]= \nonumber\\ && \frac{4\pi^{w}}{V_{11-w}}\int d\sigma_{11}... d\sigma_{1}\Pi_{n=1}^{w} [\delta (\sigma_{n} + \theta_{n})]
\label{10}
\end{eqnarray}
where we have replaced $4\pi T_{tri}$ with $\frac{4\pi^{w}}{V_{11-w}}$. Here, $V_{11-w}$ is the volume of space which is empty od DNA. To obtain this delta function, we can use of waves that number of packing fields in them is equal to number of un-packing fields of DNA and also, number of un-packing fields in them is equal to number of packings in DNA. To this aim, we put $X,Y$ instead of $M,N$ and also $M,N$ instead of $X,Y$ in Hamiltonian of DNA (equation (\ref{9})) and write:
\begin{eqnarray}
&&H_{topoisomerase-like wave}=4\pi T_{tri}\int d\sigma_{11}... d\sigma_{1} Q_{topoisomerase}E_{topoisomerase}\nonumber\\ && Q_{topoisomerase} =\Sigma_{M,N,X,Y=1}^{W} Q_{un-packed,topoisomerase,X,Y}\bar{Q}_{packed,topoisomerase,M,N}\nonumber\\ && E_{topoisomerase} =\Sigma_{M,N,X,Y=1}^{W} E_{un-packed,topoisomerase,X,Y}\bar{E}_{packed,DNA,M,N} \label{11}
\end{eqnarray}
Using equations (\ref{9} and \ref{11}), we can obtain the explicit form of delta function in equation (\ref{11}):
\begin{eqnarray}
&& \delta(\sigma_{n} -\theta_{n})= \frac{1}{\pi} \frac{[K_{\theta_{n}} + K_{\sigma_{n}}]^{2}}{[K_{\theta_{n}} + K_{\sigma_{n}}]^{2} + [\theta_{n} + \sigma_{n}]^{2}} \label{12}
\end{eqnarray}
Above results show that to exchange information between DNAs, it is needed to some waves which their structures are similar to topoisomerases in biology. In these systems, number of packed manifolds is equal to number of un-packed manifolds of DNA and number of un-packed manifolds is equal to number of packed manifolds in DNA. By joining these waves to DNA, Hamiltonian and topology of system tends to a constant number. In these conditions, all information of DNA can be recovered and exchanged.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{water-extra.jpg}
\end{tabular}
\end{center}
\caption{Extra atoms of water in extra dimensions. }
\end{figure*}
To understand the reason for water memory,we should go to extra dimensions and consider the real structure of water (See figure 33). In fact, water has a coding structure that only part of it can be seen in 4-dimensions. Also,waves have coding structures that only part of their structures can be detected in 4-dimensions. However,by considering amount of memory of water and waves,we can achieve to their real structure in extra dimensions. Also,in extra dimensions,the differences between two genders of water can be cleared more and their structures could be different. Also, an observer in extra dimensions can see two types of gender for waves.Each water can interact with water of opposite gender. Reducing dimensions from 11 or more to 4, some of information about structures of water and waves is lost.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{34.jpg}
\end{tabular}
\end{center}
\caption{Teleportation of sperms into eggs via extra dimensions. }
\end{figure*}
One of examples of teleportation via waters is the teleportation of sperms into the egg (See figure 34). When water of men which includes sperms become near to the water of women which includes egg, they become entangled. When two matters be very entangled, repel other object that be in opposed with this entanglement. For this reason, one of sperms is repelled by waters and transfer into eggs. Another cause of this teleportation is the entanglement between water inside the egg and water outside it. Sperms radiate some topoisomerse-like waves. These waves create some holes in egg. To fill these holes, sperm should be transferred to egg.
Another main subject is the magnetic fields of earth. It seems that magnetic fields of earth are more complicated of what they are seemed and have structures like topoisomerse-like waves. For this reason, they can contribute in teleportation. They produce some holes appropriated with objects that for filling them, objects should be transferred.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{two-water.jpg}
\end{tabular}
\end{center}
\caption{Two structures for water in extra dimensions:waters are the same in four dimensions and different in extra dimensions }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{teleporting.jpg}
\end{tabular}
\end{center}
\caption{Teleportation of particles into holes in entangled sstems of waters of earth and sky. }
\end{figure*}
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{DNAwater.jpg}
\end{tabular}
\end{center}
\caption{DNA-like structures of waters in extra dimensions. }
\end{figure*}
Another application of two structures for water in extra dimensions(See figure 35) and waves is radiating two types of topoisomerase-like waves from earth and sun or other objects in the sky. The origin of these waves are unclear. About earth, it seems that the liquid in it's core produce these types of waves. These waves are needed for activity of cells and growing plants. Also, these waves help in exchanging information between DNAs and contribute in transcription and translation. About radiated topoisomerase-like waves from sun, we guess that there are some waters in the sky that their waves mix with radiated waves of sun. Although,the structure of sun itself is yet questionable. The radiated waves from sun and earth create a medium that some points of it are holes and empty of fields. To fill these holes, particles could be teleported to them. In fact,waters of earth and sky are entangled. Each particle enters into these waters, break the entanglement and produces some holes. To fill these holes, particles are teleported into these holes.From this point,we can use for transferring of objects and fields in future (See figure 36).
Another question arises that what is the real structure of waters in extra dimensions (See figure 37). we can guess that water has the structure like DNA or protein or topoisomerases in biology. The DNA-model for water is more probable. This is because that water like DNA has the ability for storing informations.
We will show that in some conditions, topoisomerase-like waves interact with water, extract the structure of DNA from it and produce life (See figure 37). This is because that the structure of water in extra dimension includes the structure of DNA. If some waves could reverse the structure of water and replace the structure of water in four dimension by the structure of water in extra dimensions, they can produce DNA and the life.We can obtain the below equations for waters:
\begin{eqnarray}
&& H_{water,extra,women}=4\pi T_{tri} \int d\sigma_{11}.. d\sigma_{1}\Sigma_{n.m,X,Y}^{w}\Pi_{a,b,c=1}^{N}\times \nonumber\\ &&[ \Big(P^{water,4-dimension}(a,b,c,n,m,X,Y)H^{water,4-dimension}(a,b,c,n,m,X,Y)\Big)^{-1}\times \nonumber\\ && \Pi_{i=1}^{n}\Pi_{j=1}^{m}\Pi_{k=1}^{X}\Pi_{l=1}^{Y}\delta (\sigma_{nX}^{2}-\theta_{mY}^{2})-\nonumber\\ && P^{DNA,women}(a,b,c,n,m,X,Y) \Pi_{i=1}^{n}Q_{6,un-packed,i,women}\times \nonumber\\&&\Pi_{j=1}^{m}\bar{Q}_{6,packed,j,women}\Pi_{k=1}^{X}E_{6,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{6,packed,l,women}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{5,un-packed,i,women}\Pi_{j=1}^{m}\bar{Q}_{5,packed,j}\times \nonumber\\&&\Pi_{k=1}^{X}E_{5,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{5,packed,l,women}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{3,un-packed,i,women}\Pi_{j=1}^{m}\bar{Q}_{3,packed,j,women}\Pi_{k=1}^{X}E_{3,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{3,packed,l,women}]
\label{wa7}
\end{eqnarray}
\begin{eqnarray}
&& H_{water,extra, Men}=4\pi T_{tri} \int d\sigma_{11}.. d\sigma_{1}\Sigma_{n.m,X,Y}^{w}\Pi_{a,b,c=1}^{N}\times \nonumber\\ &&[ \Big(P^{water,4-dimension}(a,b,c,n,m,X,Y)H^{water,4-dimension}(a,b,c,n,m,X,Y)\Big)^{-1}\times \nonumber\\ && \Pi_{i=1}^{n}\Pi_{j=1}^{m}\Pi_{k=1}^{X}\Pi_{l=1}^{Y}\delta (\sigma_{nX}^{2}-\theta_{mY}^{2})-\nonumber\\ && P^{DNA, Men}(a,b,c,n,m,X,Y) \Pi_{i=1}^{n}Q_{6,un-packed,i, Men}\Pi_{j=1}^{m}\bar{Q}_{6,packed,j, Men} \times \nonumber\\&&\Pi_{k=1}^{X}E_{6,un-packed,k, Men}\Pi_{l=1}^{Y}\bar{E}_{6,packed,l, Men}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{5,un-packed,i, Men}\Pi_{j=1}^{m}\bar{Q}_{5,packed,j, Men}\Pi_{k=1}^{X}E_{5,un-packed,k, Men}\Pi_{l=1}^{Y}\bar{E}_{5,packed,l, Men}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{3,un-packed,i, Men}\Pi_{j=1}^{m}\bar{Q}_{3,packed,j, Men}\Pi_{k=1}^{X}E_{3,un-packed,k, Men}\Pi_{l=1}^{Y}\bar{E}_{3,packed,l, Men}]
\label{wa8}
\end{eqnarray}
In equations (\ref{wa7},\ref{wa8}), we have shown that the structure of waves and waters have a direct relation with the structure of DNA. Thus, if a wave could reverse the structure of water, the structure of DNA is appeared in four dimension and life is emerged. We have:
\begin{eqnarray}
&& H_{water,women} + H_{wave,reverse,women}= H_{DNA, women}
\label{LI1}
\end{eqnarray}
and
\begin{eqnarray}
&& H_{water,men} + H_{wave,reverse,men}= H_{DNA,men}
\label{LI2}
\end{eqnarray}
Solving above equations, we can obtain the Hamiltonian of waves as:
\begin{eqnarray}
&& H_{wave,reverse,women}=4\pi T_{tri} \int d\sigma_{11}.. d\sigma_{1}\Sigma_{n.m,X,Y}^{w}\Pi_{a,b,c=1}^{N}\times \nonumber\\ && \Big([ P^{DNA,women}(a,b,c,n,m,X,Y) \Pi_{i=1}^{n}Q_{6,un-packed,i,women}\times \nonumber\\&&\Pi_{j=1}^{m}\bar{Q}_{6,packed,j,women}\Pi_{k=1}^{X}E_{6,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{6,packed,l,women}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{5,un-packed,i,women}\Pi_{j=1}^{m}\bar{Q}_{5,packed,j}\times \nonumber\\&&\Pi_{k=1}^{X}E_{5,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{5,packed,l,women}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{3,un-packed,i,women}\Pi_{j=1}^{m}\bar{Q}_{3,packed,j,women}\Pi_{k=1}^{X}E_{3,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{3,packed,l,women}]-1]|\times \nonumber\\ &&\Pi_{i=1}^{n}\Pi_{j=1}^{m}\Pi_{k=1}^{X}\Pi_{l=1}^{Y}\delta (\sigma_{nX}^{2}-\theta_{mY}^{2})Big)-\nonumber\\ &&[\Big(P^{water,4-dimension}(a,b,c,n,m,X,Y)H^{water,4-dimension}(a,b,c,n,m,X,Y)\Big)^{-1}]\times\nonumber\\ && [P^{DNA,women}(a,b,c,n,m,X,Y) \Pi_{i=1}^{n}Q_{6,un-packed,i,women}\times \nonumber\\&&\Pi_{j=1}^{m}\bar{Q}_{6,packed,j,women}\Pi_{k=1}^{X}E_{6,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{6,packed,l,women}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{5,un-packed,i,women}\Pi_{j=1}^{m}\bar{Q}_{5,packed,j}\times \nonumber\\&&\Pi_{k=1}^{X}E_{5,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{5,packed,l,women}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{3,un-packed,i,women}\Pi_{j=1}^{m}\bar{Q}_{3,packed,j,women}\Pi_{k=1}^{X}E_{3,un-packed,k,women}\Pi_{l=1}^{Y}\bar{E}_{3,packed,l,women}]]
\label{LI3}
\end{eqnarray}
and
\begin{eqnarray}
&& H_{wave,reverse,men}=4\pi T_{tri} \int d\sigma_{11}.. d\sigma_{1}\Sigma_{n.m,X,Y}^{w}\Pi_{a,b,c=1}^{N}\times \nonumber\\ && \Big([ P^{DNA,men}(a,b,c,n,m,X,Y) \Pi_{i=1}^{n}Q_{6,un-packed,i,men}\times \nonumber\\&&\Pi_{j=1}^{m}\bar{Q}_{6,packed,j,men}\Pi_{k=1}^{X}E_{6,un-packed,k,men}\Pi_{l=1}^{Y}\bar{E}_{6,packed,l,men}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{5,un-packed,i,men}\Pi_{j=1}^{m}\bar{Q}_{5,packed,j}\times \nonumber\\&&\Pi_{k=1}^{X}E_{5,un-packed,k,men}\Pi_{l=1}^{Y}\bar{E}_{5,packed,l,men}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{3,un-packed,i,men}\Pi_{j=1}^{m}\bar{Q}_{3,packed,j,men}\Pi_{k=1}^{X}E_{3,un-packed,k,men}\Pi_{l=1}^{Y}\bar{E}_{3,packed,l,men}]-1]|\times \nonumber\\ &&\Pi_{i=1}^{n}\Pi_{j=1}^{m}\Pi_{k=1}^{X}\Pi_{l=1}^{Y}\delta (\sigma_{nX}^{2}-\theta_{mY}^{2})Big)-\nonumber\\ &&[\Big(P^{water,4-dimension}(a,b,c,n,m,X,Y)H^{water,4-dimension}(a,b,c,n,m,X,Y)\Big)^{-1}]\times\nonumber\\ && [P^{DNA,men}(a,b,c,n,m,X,Y) \Pi_{i=1}^{n}Q_{6,un-packed,i,men}\times \nonumber\\&&\Pi_{j=1}^{m}\bar{Q}_{6,packed,j,men}\Pi_{k=1}^{X}E_{6,un-packed,k,men}\Pi_{l=1}^{Y}\bar{E}_{6,packed,l,men}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{5,un-packed,i,men}\Pi_{j=1}^{m}\bar{Q}_{5,packed,j}\times \nonumber\\&&\Pi_{k=1}^{X}E_{5,un-packed,k,men}\Pi_{l=1}^{Y}\bar{E}_{5,packed,l,men}\times\nonumber\\ && \Pi_{i=1}^{n}Q_{3,un-packed,i,men}\Pi_{j=1}^{m}\bar{Q}_{3,packed,j,men}\Pi_{k=1}^{X}E_{3,un-packed,k,men}\Pi_{l=1}^{Y}\bar{E}_{3,packed,l,men}]]
\label{LI4}
\end{eqnarray}
Above equations show that for extracting the structure of DNA from pure waters, we need to some special waves. These waves send the structure of waters into extra dimension and recover it's structure from extra dimension. This can be the main reason for the emergence of life.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{newborn.jpg}
\end{tabular}
\end{center}
\caption{Producing new born of chicken without needing to egg byy using topoisomerase like waves. }
\end{figure*}
Now, the question arises that how we could store these main waves . We can show that there are some mechanism for storing topoisomerase like waves in nature. For example, trees could store all topoisomerase like waves and then apply them for special aims. If we put new born chicken under some trees, they exchange topoisomerase like waves and trees could act like artificial mother of these chickens. Using the stored topoisomerase-like waves, we could control the growth of cells of fetus of chicken without needing to the egg and produce chicken before 21 days (See figure 38). This can be generalized to human and animals so.
We have proposed some experimental reasons for programming molecules of two tubes of water by DNAs of hen and cock (rooster) separately. Then, we have extracted DNAs and poured some milk in tubes of water. We have observed that bacteria in milk like lactic acid bacteria interact differently with two types of waters. Then, we have put a metal barrier between two types of water and milk. We expect that electromagnetic waves dont pass this barrier. However, we have observed two different evolutions in milk. This means that some new waves are exchanged between molecules of waters and bacteria in milk. These waves act like topoisomerases in biology and store information in water and DNA.
\begin{figure*}[thbp]
\begin{center}
\begin{tabular}{rl}
\includegraphics[width=8cm]{rooster-hen.jpg}
\end{tabular}
\end{center}
\caption{Producing fetus of chicken by mixing blood cells of a rooster and a hen and returning them to stem ones in matterial of an egg. }
\end{figure*}
Another interesting result is producing a fetus in material of an egg by mixing stem cells of a hen and cock(Rooster) (See figure 39). We can take blood of a hen and a cock and mix them in some thermal conditions. Then, we pour material of an egg without any fetus in a container. We pour mixtures of cells of hen and rooster in that material. After some days, a fetus is emerged in that container.
\section{Summary} \label{sum}
Recently, Montagnier and his collaborators have claimed that a DNA molecule can have communication with other DNA molecules by radiating electromagnetic signals. Using the concepts of string theory and M-theory, we suggested a mathematical model for their idea and calculated the energy and the current of a normal DNA molecule and a damaged DNA molecule. In our model, each part of a DNA molecule acts like an electrical device. For example, some base pairs act like capacitors and some bases behave like diodes or resistors. We have discussed that the length of a DNA molecule is much greater than the size of the nucleus and for this reason, a DNA molecule has to be packed in a chromosome. During this packing, at least four types of inductors or coils appear. One coil emerges due to the electric dipoles in the structure of DNA. A second type is produced by coiling DNA bases around the histone in a nucleosome. A third type is created by the formation of loops in a chromatin fiber and finally, a fourth type of inductor emerges along the supercoil within a chromosome. Each type of these inductors produces one type of magnetic field and plays a main role in a resonant circuit. If we replace bases with electrical devices like inductors and capacitors, we can construct at least four tuned or resonant circuits. Thus, each DNA has at least four resonant frequencies and many diodes, resistors and antennae. This circuit is very similar to the circuit of an FM radio. We then argued that when a DNA molecule is damaged, its electric charges change and its emitted electromagnetic fields become different respect to the signals of a normal DNA molecule. For example, if an atom is added to one of the bases, it becomes a heptagonal molecule. If a DNA molecule loses one of its atoms, the hexagonal shape of its base changes and a pentagonal base is produced. This causes a change in the current in the DNA's inductors and an extra signal is produced. These signals are received by inductors in other DNA molecules and an extra current is induced in those DNA molecules, which leads to their destruction. This event may be a reason for the production and replication of cancer cells. If we add a third inductor to this system, which produces an extra current in the opposite direction with respect to the current of the DNA molecule, an extra magnetic field emerges. This field cancels the effect of the magnetic field which is induced by a damaged DNA molecule in a normal DNA molecule and prevents its destruction. If we put a damaged DNA molecule of a male or a female near a damaged DNA of a female or male, their radiated waves cancel the effect of each other and disease progression is stopped. This is because the types of packing of DNA are different in some chromosomes of men and women and consequently their signals have opposite signs and can cancel the effect of each other in a pair. Finally, the model of this paper lets us to consider the difference in the electromagnetic signals of sperms and eggs. We can show that the radiated waves of sperms which fuse with eggs and produce males are different respect to the waves of sperms that produce females.
These results can be extended to viruses. One can observe that viruses which are attracted by cells of a female radiate different waves respect to viruses which are absorbed by cells of a male. These viruses are the same, however, type of packing of DNA or the location of chromosomes is different which causes that their signals become different. In fact, some of viruses are male and some others are female. On the other hand, some cancers are produced by viruses. Thus, we can use of the interactions between viruses with different gender to cure cancers. Also, we can construct a wire of viruses that transmit information very fast like wormhole in cosmology or superconductor in nano. These wires includes different type of viruses and can pass the skin and achieve to the special cell.
Another interesting point of viruses is their application as the receiver or sender of electromagnetic waves. We can use of some viruses as the wireless devises for communications with DNAs in various cells. Also, they can help us to communicate with specialized cells and force them to return to initial stem cell or even initial sperm and egg. We can also use of viruses as devices for communications with cells and doing indirect surgery without any touch with human's body.
Above results help us to introduce two types of medical imaging techniques:1.Sperm Medical Imaging (SMI)2.Virus Medical Imaging (VMI). In Sperm Medical Imaging technique, we use of the communications between sperms and cells. In this method, by decoding the radiated signals of sperms and considering their evolutions, we can imagine the events interior of a human's body. In Virus Medical Imaging, we use of the communications between Viruses and cells interior of human's body. In this model, by considering the evolutions of viruses out of a body, we can analyse the events interior of a human's body.
To construct a device that communicate with DNAs of cells interior of human's body, we use of the MRI machin which it's protons are replaced by sperms or viruses. In Sperm/Virus Medical Imaging, signals from sperms/Viruses are received by some electronic devices or some special antennas and then applied to the patient. This leads to the occurrence of a communication between sperms/Viruses and cells. The DNAs in cells interior of human's body or tissue response to sperms and send radio frequency waves, which are taken by a radio receiver. These signals may be decoded and the exact information about the evolutions of cells interior of a body can be obtained.
In cancers that have another reason, we can send electromagnetic waves to them and force them to convert to their initial stem cell. Then by the management of their differentiation, we can obtain the non-cancer cell. We can show that in a human's body, two cells in neighbourhood, have reverse signals and cancel the effect of each other. If these waves couldn't remove the effect of each other, the stability of system vanishes and diseases like the cancer are emerged.
One of interesting subject in biology is the ability of DNA for saving information. We can show that water has the same ability and we can define a coding for it like the genetic coding of DNA. Also, electromagnetic waves have an special type of coding and act similar to DNA or protein. When electromagnetic waves achieve to molecules of water, give their information to water and take it's information. The process is very the same of transcription and translation in biological systems. Sometimes, electromagnetic waves act like DNA and water acts like the protein and sometimes, reversely, water acts like the DNA and electromagnetic waves act like the protein. On the other hand, water and electromagnetic waves have special genders. For example, molecules of water attract waves with opposite gender.
We can use of the gender of water in medical imaging. First, we separate two waters one with the gender of male which interacts with men and another with the gender of female which interacts with women. If a woman eat water wit the gender of female,
all of it repel with her body. For a pregnant women with a fetus of male, some of this water are absorbed, while for a fetus of female, all of this water go out. This method is completely harmless respect to other methods like the Sonography. Another interesting pint is separating waves by using water. Water with the gender of male attract waves with the gender of female and water with the gender of male, attract waves with the gender of female. Using this property, we can analyse the events interior of human's body by considering the exchanged waves between waters inside calls and outside of body.
In additional to electromagnetic waves, DNAs emit some new types of waves which have the shape of topoisomerases. Topoisomerases are enzymes that participate in the overwinding or underwinding of DNA. We generalise this concept to waves which participate in the overwinding or underwinding of DNA in extra dimensions (See figure 32). These wave topoisomerases have 4 + n -dimensional shapes. This is because that DNA is very compacted in 3 + 1 dimensional space-time. However if we see topology of DNA in extra dimension, we can open packings of it and obtain a simpler shape. In fact DNA acts like a 4 + n-dimensional coil which has been packed to 3 + 1-dimensions. Only, waves could interact with DNA, which could move in 4 + n dimensions and see the un-packed structure of it in extra dimensions. These waves couldn't be classified as the electromagnetic or gravitational waves. However, by reducing dimensions to 4 or 3, they obtain some properties similar to electromagnetic waves.
To understand the reason for water memory,we should go to extra dimensions and consider the real structure of water (See figure 33). In fact, water has a coding structure that only part of it can be seen in 4-dimensions. Also,waves have coding structures that only part of their structures can be detected in 4-dimensions. However,by considering amount of memory of water and waves,we can achieve to their real structure in extra dimensions. Also,in extra dimensions,the differences between two genders of water can be cleared more and their structures could be different. Also, an observer in extra dimensions can see two types of gender for waves.Each water can interact with water of opposite gender. Reducing dimensions from 11 or more to 4, some of information about structures of water and waves is lost.
One of examples of teleportation via waters is the teleportation of sperms into the egg (See figure 34). When water of men which includes sperms become near to the water of women which includes egg, they become entangled. When two matters be very entangled, repel other object that be in opposed with this entanglement. For this reason, one of sperms is repelled by waters and transfer into eggs. Another cause of this teleportation is the entanglement between water inside the egg and water outside it. Sperms radiate some topoisomerse-like waves. These waves create some holes in egg. To fill these holes, sperm should be transferred to egg.
Another main subject is the magnetic fields of earth. It seems that magnetic fields of earth are more complicated of what they are seemed and have structures like topoisomerse-like waves. For this reason, they can contribute in teleportation. They produce some holes appropriated with objects that for filling them, objects should be transferred.
Another application of two structures for water in extra dimensions(See figure 35) and waves is radiating two types of topoisomerase-like waves from earth and sun or other objects in the sky. The origin of these waves are unclear. About earth, it seems that the liquid in it's core produce these types of waves. These waves are needed for activity of cells and growing plants. Also, these waves help in exchanging information between DNAs and contribute in transcription and translation. About radiated topoisomerase-like waves from sun, we guess that there are some waters in the sky that their waves mix with radiated waves of sun. Although,the structure of sun itself is yet questionable. The radiated waves from sun and earth create a medium that some points of it are holes and empty of fields. To fill these holes, particles could be teleported to them. In fact,waters of earth and sky are entangled. Each particle enters into these waters, break the entanglement and produces some holes. To fill these holes, particles are teleported into these holes.From this point,we can use for transferring of objects and fields in future (See figure 36).
Another question arises that what is the real structure of waters in extra dimensions (See figure 37). we can guess that water has the structure like DNA or protein or topoisomerases in biology. The DNA-model for water is more probable. This is because that water like DNA has the ability for storing informations.
We have shown that in some conditions, topoisomerase-like waves interact with water, extract the structure of DNA from it and produce life. This is because that the structure of water in extra dimension includes the structure of DNA. If some waves could reverse the structure of water and replace the structure of water in four dimension by the structure of water in extra dimensions, they can produce DNA and the life.
Now, the question arises that how we could store these main waves. We can show that there are some mechanism for storing topoisomerase like waves in nature. For example, trees could store all topoisomerase like waves like battery and then apply them for special aims. If we put new born chicken under some trees, they exchange topoisomerase like waves and trees could act like artificial mother of these chickens. Using the stored topoisomerase-like waves, we could control the growth of cells of fetus of chicken without needing to the egg and produce chicken before 21 days (See figure 38). This can be generalized to human and animals so.
We have proposed some experimental reasons for programming molecules of two tubes of water by DNAs of hen and cock (rooster) separately. Then, we have extracted DNAs and poured some milk in tubes of water. We have observed that bacteria in milk like lactic acid bacteria interact differently with two types of waters. Then, we have put a metal barrier between two types of water and milk. We expect that electromagnetic waves dont pass this barrier. However, we have observed two different evolutions in milk. This means that some new waves are exchanged between molecules of waters and bacteria in milk. These waves act like topoisomerases in biology and store information in water and DNA.
Another interesting result is producing a fetus in material of an egg by mixing stem cells of a hen and cock(Rooster) (See figure 39). We can take blood of a hen and a cock and mix them in some thermal conditions. Then, we pour material of an egg without any fetus in a container. We pour mixtures of specialised cells of hen and rooster in that material. After some days and under some conditions, a fetus is emerged in that container.
\section*{Acknowledgments}
\noindent The work of Alireza Sepehri has been supported financially by Research Institute for biotech development in Tehran. We are designing a circuit for communication with DNA and we hypothesize that cancer can be cured via this method. We also obtained the first results for producing first stages of fetus of chicken without needing to egg. We are producing a fetus by returning specialized cells of a hen (Chicken) and a rooster to stem ones and then mixing them in material of an egg. Comments are welcome.
|
\section{\bf Introduction}
Asymptotic freedom is a key feature of non-Abelian gauge theories \cite{Gross:1973id,Politzer:1973fx}. It predicts that interactions weaken with growing energy due to quantum effects, thereby reaching a free ultraviolet (UV) fixed point under the renormalisation group.
Asymptotic safety, on the other hand, stipulates that running couplings may very well asymptote into an interacting UV fixed point at highest energies \cite{Wilson:1971bg,Weinberg:1980gg}.
The most striking difference between asymptotically free and asymptotically safe theories relates to residual interactions in the UV. Canonical power counting is modified, whence
establishing asymptotic safety in a reliable manner becomes a challenging task \cite{Falls:2013bv}.
Rigorous results for asymptotic safety at weak coupling have been known since long for models including either scalars, fermions, gauge fields or gravitons, and away from their respective critical dimensionality
\cite{Gastmans:1977ad,Tomboulis:1977jk,Christensen:1978sc,Tomboulis:1980bs,
Weinberg:1980gg,Peskin:1980ay,Smolin:1981rm,Bardeen:1983rv,Gawedzki:1985uq,
Gawedzki:1985ed,deCalan:1991km,Kazakov:2007su}.
In these toy models asymptotic safety arises through the cancellation of tree level and leading order quantum terms.
Progress has also been made to substantiate the asymptotic safety conjecture beyond weak coupling \cite{Falls:2013bv}. This is of particular relevance for quantum gravity where good evidence has arisen in a variety of different settings~\cite{Reuter:1996cp,Litim:2003vp,Litim:2006dx,Niedermaier:2006ns,Litim:2008tt,
Codello:2008vh,Litim:2011cp,Falls:2014tra,Biemans:2016rvp,Folkerts:2011jz,Christiansen:2012rx,Dona:2013qba,Meibohm:2015twa,Christiansen:2017cxa,Falls:2017lst}.
An important new development in the understanding of asymptotic safety has been initiated in \cite{Litim:2014uca} where it was shown that certain four-dimensional quantum field theories involving $SU(N)$ gluons, quarks, and scalars can develop weakly coupled UV fixed points.
Results have been extended beyond classically marginal interactions \cite{Buyukbese:2017ehm}. Structural insights into the renormalisation of general gauge theories have led to necessary and sufficient conditions for asymptotic safety, alongside strict no go theorems
\cite{Bond:2016dvk,Bond:2017sem}.
Asymptotic safety invariably arises as a quantum critical phenomenon through cancellations at loop level for which
all three types of elementary degrees of freedom --- scalars, fermions and gauge fields --- are required.
Findings have also been extended to cover supersymmetry \cite{Bond:2017suy} and UV conformal windows \cite{Bond:2017tbw}.
Throughout, it is found that suitable Yukawa interactions are pivotal \cite{Bond:2016dvk,Bond:2017sem}.
In this paper, we are interested in fixed points of
semi-simple gauge theories. Our primary motivation is the semi-simple nature of the Standard Model, and the prospect for asymptotically safe extensions thereof \cite{Bond:2017wut}. We are particularly interested in semi-simple theories where interacting fixed points and asymptotic safety can be established rigorously \cite{Bond:2016dvk}.
More generally, we also wish to
understand
how low- and high-energy fixed points are generated dynamically, what their features are, and whether novel phenomena arise owing to the semi-simple nature of the underlying gauge symmetry.
Understanding the stability of a Higgs-like ground state at interacting fixed points is also of interest in view of the ``near-criticality" of the Standard Model vacuum \cite{EliasMiro:2011aa,Buttazzo:2013uya}.
We investigate these questions for quantum field theories with
$SU(N_C)\times SU(N_c)$ local gauge symmetry coupled to massless fermionic and singlet scalar matter. Our models also have a global $U(N_{\rm F})_L \times U(N_{\rm F})_R\times U(N_{\rm f})_L \times U(N_{\rm f})_R$ flavor symmetry, and are characterised by up to nine independent couplings. Matter field multiplicities serve as free parameters. We obtain rigorous results from the leading orders in perturbation theory by adopting a Veneziano limit. We then provide a comprehensive classification of quantum field theories according to their UV and IR limits, their fixed points, and eigenvalue spectra.
Amongst these, we find semi-simple gauge theories with exact asymptotic safety in the UV.
We also find a large variety of theories with crossover- and low-energy fixed points. Further novelties include theories with inequivalent yet fully attractive IR conformal fixed points, theories with weakly interacting fixed points in both the UV and the IR, and massless theories with a non-trivial gauge sector which is UV free and IR free.
We illustrate our results by providing general phase diagrams for simple and semi-simple gauge theories with and without Yukawa interactions.
The paper is organised as follows. General aspects of weakly interacting fixed points in $4d$ gauge theories are laid out in Sec.~\ref{FPgeneral}, together with first results and expressions for universal exponents. In Sec.~\ref{MM} we introduce concrete families of semi-simple gauge theories coupled to elementary singlet ``mesons'' and suitably charged massless fermions. Perturbative RG equations for all gauge, Yukawa and scalar couplings and masses in a Veneziano limit are provided to the leading non-trivial orders in perturbation theory. Sec.~\ref{FPs} presents our results for all interacting perturbative fixed points and their universal scaling exponents. Particular attention is paid to new effects which arise due to the semi-simple nature of the models. Sec.~\ref{stab} provides the corresponding fixed points in the scalar sector. It also establishes stability of the quantum vacuum whenever a physical fixed point arises in the gauge sector. Using field multiplicities as free parameters, Sec.~\ref{class} provides a complete classification of distinct models with asymptotic freedom or asymptotic safety in the UV, or without UV completions, together with their scaling in the deep IR. In Sec.~\ref{PD}, the generic phase diagrams for simple and semi-simple gauge theories with and without Yukawas are discussed. The phase diagrams, UV -- IR transitions, and aspects of IR conformality are analysed in more depth for sample theories with asymptotic freedom and asymptotic safety. Further reaching topics such as exact perturbativity, extensions beyond the Veneziano limit, and conformal windows are discussed in Sec.~\ref{disc}. Sec.~\ref{sum} closes with a brief summary.
\section{\bf Fixed points of gauge theories}\label{FPgeneral}
In this section, we discuss general aspects of interacting fixed points in semi-simple gauge theories which are weakly coupled to matter, with or without Yukawa interactions, following \cite{Bond:2016dvk,Bond:2017sem}. We also introduce some notation and conventions.
\subsection{Fixed points in perturbation theory}
We are interested in the renormalisation of general gauge theories coupled to matter fields, with or without Yukawa couplings. The running of the gauge couplings $\alpha_i=g_i^2/(4\pi)^2$ with the renormalisation group scale $\mu$ is determined by the beta functions of the theory. Expanding them perturbatively up to two loop we have
\begin{align}\label{eq:RGE}
\mu \partial_\mu \alpha_i \equiv \beta_i= \alpha_i^2(-B_i + C_{ij}\alpha_j
- 2\,Y_{4,i}) + O(\alpha^4)\,,
\end{align}
where a sum over gauge group factors $j$ is implied.
The one- and two-loop gauge contributions $B_i$ and $C_{ij}$ and the two-loop Yukawa contributions $Y_{4,i}$ are known for general gauge theories, see~\cite{Machacek:1983tz,Machacek:1983fi,Machacek:1984zw,Luo:2002ti,Bond:2016dvk} for explicit expressions. While $B_i$ and $C_{ii}$ may take either sign, depending on the matter content, the Yukawa contribution $Y_{4,i}$ and the off-diagonal gauge contributions $C_{ij}$ $(i\neq j)$ are strictly positive in any quantum field theory.
Scalar couplings do not play any role at this order in perturbation theory.
The effect of Yukawa couplings is incorporated by projecting the gauge beta functions \eq{eq:RGE} onto the Yukawa nullclines $(\beta_Y = 0)$, leading to explicit expressions for $Y_{4,i}$ in terms of the gauge couplings $g_j$.
Moreover, for many theories the Yukawa contribution along nullclines can be written as $Y_{4,i}=D_{ij}\,\alpha_j$ with $D_{ij} \ge 0$ \cite{Bond:2016dvk}. We can then go one step further and express the net effect of Yukawa couplings as a shift of the two loop gauge contribution,
$C_{ij} \rightarrow C_{ij}' = C_{ij} - 2D_{ij} \leq C_{ij}$. Notice that the shift will always be by some negative amount provided at least one of the Yukawa couplings is non-vanishing. It leads to the reduced gauge beta functions
\begin{align}\label{eq:RGE2}
\beta_i = \alpha_i^2(-B_i + C'_{ij}\alpha_j) + O(\alpha^4)\,.
\end{align}
Fixed points solutions of \eq{eq:RGE2} are either free or interacting and $\alpha^*=0$ for some or all gauge factors is always a self-consistent solution.
Consequently, interacting fixed points are solutions to
\begin{align}\label{eq:matFP}
B_i &= C_{ij}'\,\alpha_j^*\,,\quad\text{subject to}\quad\alpha_i^*> 0\,,
\end{align}
where only those rows and columns are retained where gauge couplings are interacting.
Next we discuss the role of Yukawa couplings for the fixed point structure. In the absence of Yukawa couplings, the two-loop coefficients remain unshifted $C_{ij}' = C_{ij}$. An immediate consequence of this is that any interacting fixed point must necessarily be IR. The reason is as follows: for an interacting fixed point to be UV, asymptotic freedom cannot be maintained for all gauge factors, meaning that some $B_i < 0$. However, as has been established in \cite{Bond:2016dvk}, $B_i \leq 0$ necessarily entails $C_{ij} \geq 0$ in any $4d$ quantum gauge theory. If the left hand side of \eqref{eq:matFP} is negative, if only for a single row, positivity of $C_{ij}$ requires that some $\alpha_j^*$ must take negative values for a fixed point solution to arise. This, however, is unphysical \cite{Dyson:1952tj} and we are left with $B_i > 0$ for each $i$, implying that asymptotic freedom remains intact in all gauge sectors. Besides the Gaussian, the theory may have weakly interacting infrared Banks-Zaks fixed points in each gauge sector, as well as products thereof, which arise as solutions to \eq{eq:matFP} with the unshifted coefficients.
In the presence of Yukawa couplings, the coefficients $C_{ij}'$ can in general take either sign. This has far reaching implications. Firstly, the theory can additionally display gauge-Yukawa fixed points where both the gauge and the Yukawa couplings take interacting values.
Most importantly, solutions to \eqref{eq:matFP} are then no longer limited to theories with asymptotic freedom. Instead, interacting fixed points can be infrared, ultraviolet, or of the crossover type. In general we may expect gauge-Yukawa fixed points for each independent Yukawa nullcline. In summary, perturbative fixed points are either $(i)$ free and given by the Gaussian, or $(ii)$ free in the Yukawa but interacting in the gauge sector (Banks-Zaks fixed points), or $(iii)$ simultaneously interacting in the gauge and the Yukawa sector (gauge-Yukawa fixed points), or $(iv)$ combinations and products of $(i)$, $(ii)$ and $(iii)$. Banks-Zaks fixed points are always IR, while the Gaussian and gauge-Yukawa fixed points can be either UV or IR. Depending on the details of the theory and its Yukawa structure, either the Gaussian or one of the interacting gauge-Yukawa fixed points will arise as the ``ultimate'' UV fixed point of the theory and may serve to define the theory fundamentally \cite{Bond:2017sem}.
The effect of scalar quartic self-couplings on the fixed point is strictly sub-leading in terms of the values of the fixed points, as they do not affect the running of gauge couplings at this order of perturbation theory. However, as to have a true fixed point we must acquire one in all couplings, they provide additional constraints on the physicality of candidate gauge-Yukawa fixed points, as we additionally require that the quartic couplings take fixed points which are both real-valued, and lead to a bounded potential which leads to a stable vacuum state.
\subsection{Gauge couplings}
Let us now consider a semi-simple gauge-Yukawa theory with non-Abelian gauge fields under the semi-simple gauge group ${\cal G}_1\otimes {\cal G}_2$ coupled to fermions and scalars. We have two non-Abelian gauge couplings $\al 1$ and $\al 2$, which are related to the fundamental gauge couplings via $\alpha_i=g_i^2/(4\pi)^2$.
The running of gauge couplings within perturbation theory is given by
\begin{equation}\label{beta12}
\begin{array}{rcl}
\beta_1&=&
\displaystyle
-B_1\, \alpha_1^2+C_1\,\alpha_1^3 +G_1\,\alpha_1^2\,\alpha_2\,,\\[1ex]
\beta_2&=&
\displaystyle
-B_2\, \alpha_2^2+C_2\,\alpha_2^3 +G_2\,\alpha_2^2\,\alpha_1\,.
\end{array}
\end{equation}
Here, $B_i$ are the well known one-loop coefficients. In theories without Yukawa interactions, or where Yukawa interactions take Gaussian values, the numbers $C_i$ and $G_i$ are the two-loop coefficients which arise owing to the gauge loops and owing to the mixing between gauge groups, meaning $C_i\equiv C_{ii}$ (no sum), and $G_{1}\equiv C_{12}$, $G_{2}\equiv C_{12}$, see \eq{eq:RGE}. In this case, we also have that $C_i, G_i\ge 0$ as soon as $B_i<0$.\footnote{General formal expressions of loop coefficients in the conventions used here are given in \cite{Bond:2016dvk}.} For theories where Yukawa couplings take interacting fixed points
the numbers $C_i$ and $G_i$ receive corrections due to the Yukawas, $C_i\equiv C'_{ii}$ (no sum), and $G_{1}\equiv C'_{12}$, $G_{2}\equiv C'_{12}$, see \eq{eq:RGE2}. Most notably, strict positivity of $C_i$ and $G_i$ is then no longer guaranteed \cite{Bond:2016dvk}.
\begin{table*}
\begin{center}
\begin{tabular}{cccc}
\toprule
\rowcolor{LightGreen}
&&&\\[-4mm]
\rowcolor{LightGreen}
\multicolumn{2}{c}{\bf fixed point}
&${}\quad\bm\alpha_{\rm \bf gauge}{}\quad$
&${}\quad\bm\alpha_{\rm \bf Yukawa}{}\quad$
\\
\midrule
\rowcolor{LightGray}
Gauss&G&$=0$&$=0$\\
Banks-Zaks&BZ& $\neq 0$&$=0$\\
\rowcolor{LightGray}
gauge-Yukawa\ \ &GY& $\neq 0$ & $\neq 0$\\
\bottomrule
\end{tabular}
\caption{Conventions to denote the basic fixed points (Gaussian, Banks-Zaks, or gauge-Yukawa) of simple gauge theories weakly coupled to matter.}
\label{tFPdef}
\end{center}
\end{table*}
In either case, the fixed points of the combined system are determined by the vanishing of \eq{beta12}. For a general semi-simple gauge theory with two gauge factors, one finds four different types of fixed points. The Gaussian fixed point
\begin{equation}
\label{1}
(\al1^*,\al2^*)=(0,0)
\end{equation}
always exist (see Tab.~\ref{tFPdef} for our conventions). It is the UV fixed point of the theory as long as the one-loop coefficients obey $B_i>0$. The theory may also develop partially interacting fixed points,
\arraycolsep .1em \begin{eqnarray}
\label{2}
&&(\al1^*,\al2^*)=\left(0,\frac{B_2}{C_2}\right)\,,\\
\label{3}
&&(\al1^*,\al2^*)=\left(\frac{B_1}{C_1},0\right)\,.
\end{eqnarray}
Here, one of the gauge coupling is taking Gaussian values whereas the other one is interacting. The interacting fixed point is of the Banks-Zaks type \cite{Caswell:1974gg,Banks:1981nn}, provided Yukawa interactions are absent. This then also implies that the gauge coupling is asymptotically free. Alternatively, the interacting fixed point can be of the gauge-Yukawa type, provided that Yukawa couplings take an interacting fixed point themselves. In this case, and depending on the details of the Yukawa sector, the fixed point can be either IR or UV.
Finally, we also observe fully interacting fixed points
\begin{equation}
\label{4}
(\al1^*,\al2^*)=\left(\frac{C_2B_1-B_2G_1}{C_1C_2-G_1G_2},\frac{C_1B_2-B_1G_2}{C_1C_2-G_1G_2}\right)\,.
\end{equation}
As such, fully interacting fixed points \eq{4} can be either UV or IR, depending on the specific field content of the theory. In all cases we will additionally require that the couplings obey
\begin{equation}\label{pos}
\begin{array}{rcl}
\alpha_1&\ge& 0\,,\\[1ex]
\alpha_2&\ge& 0\,.
\end{array}
\end{equation}
to ensure they reside in the physical regime of the theory \cite{Dyson:1952tj}.
\begin{center}
\begin{table}
\begin{tabular}{ccccc}
\toprule
\rowcolor{LightBlue}
{}\ \ \ \bf coupling\ \ \ &\multicolumn{4}{c}{\ \ order in perturbation theory\ \ }\\
\midrule
\rowcolor{LightGray}
$\beta_{\rm gauge}$&1&2&2&$n+1$ \\
\rowcolor{white}
$\beta_{\rm Yukawa}$ &0&1&1&$n$\\
\rowcolor{LightGray}
$\beta_{\rm scalar}$ &0&0&1&$n$\\
\midrule
\rowcolor{LightGreen}
\ \ \bf approximation\ \ &{}\ \ \ \ \ LO\ \ \ \ \ &\ \ NLO\ \ &\ NLO${}^\prime$\ &\ $n$NLO${}^\prime$\ \\
\bottomrule
\end{tabular}
\caption{\label{Tab} Relation between approximation level and the loop order up to which couplings are retained in perturbation theory, following the terminology of \cite{Litim:2014uca,Litim:2015iea}.}\label{tNLO}
\end{table}
\end{center}
\subsection{Yukawa couplings}
In order to proceed, we must specify the Yukawa sector.
We assume three types of non-trivially charged fermions
with charges under ${\cal G}_1$ and ${\cal G}_2$. Some or all of the fermions which are only charged under ${\cal G}_1$ (${\cal G}_2$)
also couple to scalar fields via Yukawa couplings $\al Y$ ($\al y$), respectively.
The scalars may or may not be charged under the gauge symmetries.
They will have quartic self couplings which play no primary role for the fixed point analysis at weak coupling \cite{Bond:2016dvk}.
Within perturbation theory, the beta functions for the gauge and Yukawa couplings are of the form
\begin{equation}\label{12Yy}
\begin{array}{rcl}
\beta_1&=&
\displaystyle
-B_1\, \alpha_1^2+C_1\,\alpha_1^3 -D_1\,\alpha_1^2\,\al{Y}+G_1\,\al1^2\,\al2\,,\\[.5ex]
\beta_Y&=&
\displaystyle
\ \ E_1\,\al{Y}^2-F_1\,\al{Y}\,\al 1\,,\\[.5ex]
\beta_2&=&
\displaystyle
-B_2\, \al2^2+C_2\,\al2^3 -D_2\,\al2^2\,\al y +G_2\,\al2^2\,\al1\,,\\[.5ex]
\beta_y&=&
\displaystyle
\ \ E_2\,\al{y}^2-F_2\,\al{y}\,\al 2\,.
\end{array}
\end{equation}
The RG flow is given up to two-loop in the gauge couplings, and up to one-loop in the Yukawa couplings.
We refer to this as the NLO approximation, see Tab.~\ref{Tab} for the terminology.
We are interested in the fixed points of the theory, defined implicitly via the vanishing of the beta functions for all couplings. The Yukawa couplings can display either a Gaussian or an interacting fixed point
\begin{equation}\label{Y}
\begin{array}{rcl}
\displaystyle
\al Y^*&=&
\displaystyle
0\,,\quad\al Y^*=\frac{F_1}{E_1}\,\al 1^*\,,\\[2ex]
\al y^*&=&
\displaystyle
0\,,\quad\al y^*=\frac{F_2}{E_2}\,\al 2^*\,.
\end{array}
\end{equation}
Depending on whether none, one, or both of the Yukawa couplings take an interacting fixed point, the system \eq{12Yy} reduces to \eq{beta12} whereby the two-loop coefficients $C_i$ of the gauge beta functions are shifted according to
\begin{equation} \label{C'}
\begin{array}{rcl}
\alpha_Y^*\neq 0:\quad C_1\to C_1'&=&
\displaystyle
C_1-D_1\,\0{F_1}{E_1}\le C_1\,,\\[2ex]
\displaystyle
\alpha_y^*\neq 0:\quad C_2\to C_2'&=&
\displaystyle
C_2-D_2\0{F_2}{E_2}\le C_2\,.
\end{array}
\end{equation}
Notice also that in this model the values for the mixing terms $G_i$ do not depend on whether the corresponding Yukawa couplings vanish, or not, due to the fact that no fermions charged under both groups are involved in Yukawa interactions. Owing to the fixed point structure of the Yukawa sector \eq{Y}, the formal fixed points \eq{1}, \eq{2}, \eq{3} and \eq{4} have the multiplicity $1, 2, 2$ and $4$, respectively. In total, we end up with nine qualitatively different fixed points FP${}_1$ -- FP${}_9$, summarised in Tab.~\ref{tFPs}: FP${}_1$ denotes the unique Gaussian fixed point. FP${}_2$ and FP${}_3$ correspond to a Banks-Zaks fixed point in one of the gauge couplings, and a Gaussian in the other. They can therefore be interpreted effectively as a ``product'' of a Banks-Zaks with a Gaussian fixed point. Similarly, at FP${}_4$ and FP${}_5$, one of the Yukawa couplings remains interacting, and they can therefore effectively be viewed as the product of a gauge-Yukawa (GY) type fixed point in one gauge coupling with a Gaussian fixed point in the other. The remaining fixed points FP${}_6$ -- FP${}_9$ are interacting in both gauge couplings. These fixed points are the only ones which are sensitive to the two-loop mixing coefficients $G_1$ and $G_2$. At FP${}_6$, both Yukawa couplings vanish meaning that it is effectively a product of two Banks-Zaks type fixed points.
At FP${}_7$ and FP${}_8$, only one of the Yukawa couplings vanish, implying that these are products of a gauge-Yukawa with a Banks-Zaks fixed point. Finally, at FP${}_9$, both Yukawa couplings are non-vanishing meaning that this is effectively the product of two gauge-Yukawa fixed points.
\begin{table*}
\begin{center}
\begin{tabular}{cccccc}
\toprule
\rowcolor{LightBlue}
&
\multicolumn{2}{c}{\cellcolor{LightRed} \bf gauge couplings}
&
\multicolumn{2}{c}{\cellcolor{LightYellow} \bf Yukawa couplings}
&\cellcolor{LightGreen} \bf \ fixed point \
\\
\rowcolor{LightBlue}
\multirow{-2}{*}{\bf \ fixed point\ }
&\cellcolor{LightRed}
${}\quad \bm{\al 1^*}\quad$
&\cellcolor{LightRed}
${}\quad \bm{\al 2^*}\quad$
&\cellcolor{LightYellow}
${}\quad\quad \bm{\al Y^*}\quad \quad$
&\cellcolor{LightYellow}
${}\quad \quad \bm{\al y^*}\quad \quad$
&\cellcolor{LightGreen}
{\bf \ \ \ type\ \ \ }\\
\midrule
\rowcolor{LightGray}
&&&&& \\[-2.5ex]
\rowcolor{LightGray}
\bf FP${}_{\bf 1}$
&0&0&0&0
&\bf G $\bm \cdot$ G\\[.5ex]
&&&&& \\[-2.5ex]
\bf FP${}_{\bf 2}$
&$\displaystyle\0{B_1}{C_1}$&0&0&0
&\bf BZ $\bm \cdot$ G\\[1.5ex]
\rowcolor{LightGray}
&&&&& \\[-2.5ex]
\rowcolor{LightGray}
\bf FP${}_{\bf 3}$
&0&$\displaystyle\0{B_2}{C_2}$&0&0
&\bf G $\bm \cdot$ BZ\\[1.5ex]
&&&&& \\[-2.5ex]
\bf FP${}_{\bf 4}$
&$\displaystyle\0{B_1}{{C'_1}}$&0&$\displaystyle\0{F_1}{E_1}\,\al 1$&0
&\bf GY $\bm \cdot$ G\\[1.5ex]
\rowcolor{LightGray}
&&&&& \\[-2.5ex]
\rowcolor{LightGray}
\bf FP${}_{\bf 5}$
&0&$\displaystyle\0{B_2}{{C'_2}}$&0&$\displaystyle\0{F_2}{E_2}\,\al 2$
&\quad\bf G $\bm \cdot$ GY\\[1.5ex]
&&&&& \\[-2.5ex]
\bf FP${}_{\bf 6}$
&$\ \displaystyle\frac{C_2B_1-B_2G_1}{C_1C_2-G_1G_2}\ $
&$\ \displaystyle\frac{C_1B_2-B_1G_2}{C_1C_2-G_1G_2}\ $&0&0
&\bf BZ $\bm \cdot$ BZ\\[1.5ex]
\rowcolor{LightGray}&&&&& \\[-2.5ex]
\rowcolor{LightGray}\bf FP${}_{\bf 7}$
&$\displaystyle\frac{C_2B_1-B_2G_1}{{C'_1}C_2-G_1G_2}$
&$\displaystyle\frac{{C'_1}B_2-B_1G_2}{{C'_1}C_2-G_1G_2}$&
$\displaystyle\frac{F_1}{E_1}\,\al 1$&0
&\bf GY $\bm \cdot$ BZ\\[1.5ex]
&&&&& \\[-2.5ex]
\bf FP${}_{\bf 8}$
&$\displaystyle\frac{{C'_2}B_1-B_2G_1}{C_1{C'_2}-G_1G_2}$
&$\displaystyle\frac{C_1B_2-B_1G_2}{C_1{C'_2}-G_1G_2}$&0&
$\displaystyle\frac{F_2}{E_2}\,\al 2$
&\bf BZ $\bm \cdot$ GY\\[1.5ex]
\rowcolor{LightGray}
&&&&& \\[-2.5ex]
\rowcolor{LightGray}
\bf FP${}_{\bf 9}$
&$\displaystyle\frac{{C'_2}B_1-B_2G_1}{{C'_1C'_2}-G_1G_2}$
&$\displaystyle\frac{{C'_1}B_2-B_1G_2}{{C'_1C'_2}-G_1G_2}$
&$\displaystyle\frac{F_1}{E_1}\,\al 1$
&$\displaystyle\frac{F_2}{E_2}\,\al 2$
&\bf GY $\bm \cdot$ GY\\[1.5ex]
\bottomrule
\end{tabular}
\caption{The various types of fixed points in gauge-Yukawa theories with semi-simple gauge group ${\cal G}_1\otimes {\cal G}_2$ and \eq{12Yy}, \eq{C'}.
We also indicate how the nine qualitatively different fixed points can be interpreted as products of the Gaussian (G), Banks-Zaks (BZ) and gauge-Yukawa (GY) fixed points as seen from the individual gauge group factors (see main text).}
\label{tFPs}
\end{center}
\end{table*}
In theories where none of the fermions carries gauge charges under both gauge groups, we have that $G_1= 0=G_2$. In this limit,
and at the present level of approximation,
the gauge sectors do not communicate with each other and the ``direct product'' interpretation of the fixed points as detailed above becomes exact. For the purpose of this work
we will find it useful to refer to the effective ``product'' structure of interacting fixed points even in settings with $G_1,G_2\neq 0$.
Whether any of the fixed points is factually realised in a given theory crucially depends on the explicit values of the various loop coefficients. We defer an explicit investigation for certain ``minimal models'' to Sec.~\ref{MM}.
\subsection{Scalar couplings}
In \cite{Bond:2016dvk}, it has been established that scalar self-interactions play no role for the primary occurrence of weakly interacting fixed points in the gauge- or gauge-Yukawa sector. On the other hand, for consistency, scalar couplings must nevertheless take free or interacting fixed points on their own. The necessary and sufficent conditions for this to arise have been given in \cite{Bond:2016dvk}. Firstly, scalar couplings must take physical (real) fixed points. Secondly, the theory must display a stable ground state at the fixed point in the scalar sector. Below, we will analyse concrete models and show that both of these conditions are non-empty.
\subsection{Universal scaling exponents}
We briefly comment on the universal behaviour and scaling exponents at the interacting fixed points of Tab.~\ref{tFPs}. Scaling exponents
arise as the eigenvalues $\vartheta_i$
of the stability matrix
\begin{equation}\label{M} M_{ij}=\partial\beta_i/\partial\alpha_j|_*\end{equation}
at fixed points. Negative or positive eigenvalues correspond to relevant or irrelevant couplings respectively. They imply that couplings approach the fixed point following a power-law behaviour in RG momentum scale,
\begin{equation} \alpha_i(\mu)-\alpha_i^*=\sum_n c_n\, V^n_i \,\left(\frac{\mu}{\Lambda}\right)^{\vartheta_n}+{\rm subleading}\,.\end{equation}
Classically, we have that $\vartheta\equiv 0$. Quantum-mechanically, and at a Gaussian fixed point, eigenvalues continue to vanish and the behaviour of couplings is determined by higher order effects. Then couplings are either exactly marginal $\vartheta\equiv 0$ or marginally relevant $\vartheta\to 0^-$ or marginally irrelevant $\vartheta\to 0^+$. In a slight abuse of language we will from now on denote relevant and marginally relevant ones as $\vartheta\le 0$, and vice versa for irrelevant ones.
Given that the scalar couplings do not feed back to the gauge-Yukawa sector at the leading non-trivial order in perturbation theory, we may neglect them for a discussion of the eigenvalue spectrum
\begin{equation}\{\vartheta_i,i=1,\cdots 4\}\,,
\end{equation}
related to the two gauge and Yukawa couplings.
The fixed point FP${}_1$ is Gaussian in all couplings, and the scaling of couplings are either marginally relevant or marginally irrelevant. Only if $B_i>0$ trajectories can emanate from the Gaussian, meaning that it is a UV fixed point iff the theory is asymptotically free in both couplings. Furthermore, asymptotic freedom in the gauge couplings entails asymptotic freedom in the Yukawa couplings leading to four marginally relevant couplings with eigenvalues
\begin{equation}\label{4R}
\vartheta_1, \vartheta_2,\vartheta_3,\vartheta_4 \le 0
\end{equation}
The fixed points FP${}_2$ and FP${}_3$ are products of a Banks-Zaks in one gauge sector with a Gaussian fixed point in the other. Scaling exponents are then of the form
\begin{equation}\label{3R}
\vartheta_1, \vartheta_2,\vartheta_3\le 0< \vartheta_4
\end{equation}
provided the gauge sector with Gaussian fixed point is asymptotically free. For IR free gauge coupling, we instead have the pattern
\begin{equation}\label{1R}
\vartheta_1<0\le \vartheta_2,\vartheta_3,\vartheta_4\,.
\end{equation}
At the fixed points FP${}_4$ and FP${}_5$, the theory is the product of a Gaussian and a gauge-Yukawa fixed point. Consequently, four possibilities arise: Provided that the theory is asymptotically safe at the gauge-Yukawa fixed point and asymptotically or infrared free at the Gaussian, scaling exponents are of the form \eq{3R} or \eq{1R}, respectively. Conversely, if the gauge Yukawa fixed point is IR, the eigenvalue spectrum reads
\begin{equation}\label{2R}
\vartheta_1, \vartheta_2\le0\le \vartheta_3,\vartheta_4
\end{equation}
if the Gaussian is asymptotically free. Finally, if the Gaussian is IR free and the gauge-Yukawa fixed point IR, all couplings are UV irrelevant and
\begin{equation}\label{0R}
0\le \vartheta_1, \vartheta_2,\vartheta_3,\vartheta_4\,.
\end{equation}
More work is required to determine the scaling exponents at the fully interacting fixed points FP${}_6$ -- FP${}_9$. To that end, we write the characteristic polynomial of the stability matrix as
\begin{equation}
\sum_{n=0}^4 T_n\,\vartheta^n=0\,.
\end{equation}
The coefficients $T_n$ are functions of the loop coefficients. Introducing $B=|B_1|$ and $B_2=P\,B_1$, with $P$ some free parameter, we can make a scaling analysis in the limit $B\ll 1$. Normalising the coefficient $T_4$ to unity, $T_4=1$, it then follows from the structure of the beta functions that $T_0={\cal O}(B^6), T_1={\cal O}(B^4), T_2={\cal O}(B^2)$ and $T_3={\cal O}(B)$ to leading order in $B$. In the limit where $B\ll 1$ we can deduce exact closed expressions for the leading order behaviour of the eigenvalues from solutions to two quadratic equations,
\begin{equation}\label{dtheta}
\begin{array}{rcl}
0&=&\ \ \ \,\vartheta^2+T_3\,\vartheta+T_2\\[1ex]
0&=&T_2\, \vartheta^2+T_1\,\vartheta+T_0\,.
\end{array}
\end{equation}
The general expressions are quite lengthy and shall not be given here explicitly. We note that the four eigenvalues of the four couplings at the four fully interacting fixed points FP${}_6$ -- FP${}_9$ are the four solutions to \eq{dtheta}. Irrespective of their signs, and barring exceptional numerical cancellations, we conclude that
two scaling exponents are quadratic and two are linear in $B$,
\begin{equation}\label{theta}
\begin{array}{rcl}
\vartheta_{1,2}&=&
\displaystyle
-\frac{1}{2}\left(T_3\pm\sqrt{T^2_3-4T_2}\right)={\cal O}(B^2)\\[2ex]
\vartheta_{3,4}&=&
\displaystyle
-\frac{1}{2 T_2}\left(T_1\pm\sqrt{T^2_1-4T_0\,T_2}\right)={\cal O}(B)\,.
\end{array}
\end{equation}
This is reminiscent of fixed points in gauge-Yukawa theories with a simple gauge group. The main reason for the appearance of two eigenvalues of order ${\cal O}(B^2)$ relates to the gauge sector, where the interacting fixed point arises through the cancellation at two-loop level. Conversely, two eigenvalues of order ${\cal O}(B)$ relate to the Yukawa couplings, as they arise from a cancellation at one-loop level.
This completes the discussion of fixed points in general weakly coupled semi-simple gauge theories.
\section{\bf Minimal models}\label{MM}
In this section we introduce in concrete terms a family of semi simple gauge theories whose interacting fixed points will be analysed exactly within perturbation theory in the Veneziano limit.
\subsection{Semi-simple gauge theory}
We consider families of massless four-dimensional quantum field theories with a semi-simple gauge group
\begin{equation}\label{gaugegroup}
SU(N_{\rm C}) \times SU(N_{\rm c})
\end{equation}
for general non-Abelian factors with $N_{\rm C}\ge 2$ and $N_{\rm c}\ge 2$.
Specifically, our models contains $SU(N_{\rm C})$ gauge fields $A_\mu$ with field strength $F_{\mu\nu}$,
and $SU(N_{\rm c})$ gauge fields $a_\mu$ with field strength $f_{\mu\nu}$.
The gauge fields are coupled to $N_{\rm F}$ flavors of fermions $Q_i$, $N_{\rm f}$ flavors of fermions $q_i$, and $N_\psi$ flavors of fermions $\psi_i$. The fermions $(Q,q,\psi)$ transform in the fundamental representation
of the first, the second, and both gauge group(s) \eq{gaugegroup}, respectively, as summarised in Tab.~\ref{tCharges}. The Dirac fermions $\psi$ are responsible for the semi-simple character of the theory and
serve as messengers to communicate between gauge sectors. All fermions are Dirac to guarantee anomaly cancellation.
The fermions $(Q,q)$ additionally couple via Yukawa interactions to an $N_{\rm F}\times N_{\rm F}$ matrix scalar field $H$ and an $N_{\rm f}\times N_{\rm f}$ matrix scalar field $h$, respectively. The scalars $H$ and $h$ are invariant under $U(N_{\rm F})_L \times U(N_{\rm F})_R$ and $U(N_{\rm f})_L \times U(N_{\rm f})_R$ global flavor rotations, respectively, and
singlets under the gauge symmetry. They can be viewed as elementary mesons in that they carry the same global quantum numbers as the singlet scalar bound states $\sim \langle Q\bar Q\rangle$ and $\sim \langle q\bar q\rangle$. The fermions $\psi$ are not furnished with Yukawa interactions.
The fundamental action is taken to be the sum of the individual Yang-Mills actions, the fermion kinetic terms, the Yukawa interactions, and the scalar kinetic and self-interaction Lagrangeans $L=L_{\rm YM}+L_F+L_Y+L_S+L_{\rm pot}$, with
\begin{equation} \label{L}
\begin{array}{rcl}
L_{\rm YM}&=&
\displaystyle
- \frac{1}{2} {\rm Tr} \,F^{\mu \nu} F_{\mu \nu}
- \frac{1}{2} {\rm Tr} \,f^{\mu \nu} f_{\mu \nu}\\[1.5ex]
L_F&=&
\displaystyle
{\rm Tr}\left(\overline{Q}\, i\slashed{D}\, Q \right)
+{\rm Tr}\left(\overline{q}\, i\slashed{D}\, q\right)
+{\rm Tr}\left(\overline{\psi}\, i\slashed{D}\, \psi\right)
\\[1.5ex]
L_Y&=&
\displaystyle
Y \,{\rm Tr}\left(\overline{Q}_L H Q_R + \overline{Q}_R H^\dagger Q_L\right)
+y \,{\rm Tr}\left(\overline{q}_L h q_R + \overline{q}_R h^\dagger q_L\right)
\\[1.5ex]
L_S&=&
\displaystyle
{\rm Tr}\,(\partial_\mu H ^\dagger\, \partial^\mu H) +{\rm Tr}\,(\partial_\mu h ^\dagger\, \partial^\mu h) \\[1.5ex]
L_{\rm pot}&=&
\displaystyle
-U\,{\rm Tr}\,(H ^\dagger H )^2 -V\,({\rm Tr}\,H ^\dagger H )^2 \\[1.5ex]
&&
-u\,{\rm Tr}\,(h ^\dagger h )^2 -v\,({\rm Tr}\,h ^\dagger h )^2
-w\,{\rm Tr}\,H ^\dagger H \,{\rm Tr}\,h ^\dagger h \,.
\end{array}
\end{equation}
The trace ${\rm Tr}$ denotes the trace over both color and flavor indices, and the decomposition $Q=Q_L+Q_R$ with $Q_{L/R}=\frac 12(1\pm \gamma_5)Q$ is understood for all fermions $Q$ and $q$.
Mass terms are neglected at the present stage as their effect is subleading to the main features developed below.
In four dimensions, the theory is renormalisable in perturbation theory.
\begin{table}[t]
\begin{center}
\begin{tabular}{ c c c c cccc}
\toprule
\rowcolor{LightBlue}
& \multicolumn{3}{c}{\cellcolor{LightYellow}\bf fermions}
& \multicolumn{2}{c}{\cellcolor{LightGreen} \bf scalars}
& \multicolumn{2}{c}{\cellcolor{LightRed} \bf gauge fields}\\
\rowcolor{LightBlue}\multirow{-2}{*}{\bf\ representation\ } &
\cellcolor{LightYellow} ${}\quad Q{}\quad$ &
\cellcolor{LightYellow}${}\quad q{}\quad$ &
\cellcolor{LightYellow} ${}\quad\psi{}\quad$ &
\cellcolor{LightGreen} ${}\quad H{}\quad$ &
\cellcolor{LightGreen} ${}\quad h{}\quad$ &
\cellcolor{LightRed} ${}\quad\ \ A_\mu{}\ \ \quad$ &
\cellcolor{LightRed} ${}\quad \ \ a_\mu{}\ \ \quad$\\
\midrule
\rowcolor{LightGray}
under~$SU(N_{C})$& $N_{\rm C}$ & 1 & $N_{\rm C}$ & 1 & 1 & $N_{\rm C}^2-1$ & 1\\
under~$ SU(N_{c})$& 1 & $N_{\rm c}$ & $N_{\rm c}$ & 1 & 1 & 1 & $N_{\rm c}^2-1$ \\
\rowcolor{LightGray}
multiplicity
& $N_{\rm F}$ & $N_{\rm f}$ & $N_\psi$ & $N_{\rm F}^2$ & $N_{\rm f}^2$ & 1 & 1\\
\bottomrule
\end{tabular}
\caption{Representation under the semi-simple gauge symmetry \eq{gaugegroup} together with flavour multiplicities of all fields. Gauge (fermion) fields are either in the adjoint (fundamental) or trivial representation.}
\label{tCharges}
\end{center}
\end{table}
The theory has nine classically marginal coupling constants given by the two gauge couplings, the two Yukawa couplings, and five quartic scalar couplings. We write them as
\begin{equation}
\label{couplings}
\begin{array}{lcr}
\displaystyle
\al 1=\frac{g_1^2\,N_{\rm C}}{(4\pi)^2}\,,\quad
\al 2=\frac{g_2^2\,N_{\rm c}}{(4\pi)^2}\,,\quad
\al Y=\frac{Y^{2}\,N_{\rm C}}{(4\pi)^2}\,,\quad
\al y=\frac{y^{2}\,N_{\rm c}}{(4\pi)^2}\,,
&&\\[2ex]
\displaystyle
\al U=\frac{{u}\,N_{\rm F}}{(4\pi)^2}\,,\quad
\al V=\frac{{v}\,N^2_{\rm F}}{(4\pi)^2}\,,\quad
\al u=\frac{{u}\,N_{\rm f}}{(4\pi)^2}\,,\quad
\al v=\frac{{v}\,N^2_{\rm f}}{(4\pi)^2}\,,
\end{array}
\end{equation}
where we have normalized the couplings with the appropriate loop factor and powers of $N_C, N_c, N_{\rm F}$ and $N_{\rm f}$ in view of the Veneziano limit to be adopted below. Notice the additional power of $N_{\rm F}$ and $N_{\rm f}$ in the definitions of the scalar double-trace couplings. We normalise the quartic ``portal'' coupling as
\begin{equation}\label{w}
\al w=\frac{{w}\,N_{\rm F}N_{\rm f}}{(4\pi)^2}\,.
\end{equation}
It is responsible for a mixing amongst the scalar sectors starting at tree level.
Below, we use the shorthand notation $\beta_i\equiv\partial_t\alpha_i$ with $i=(1,2,Y,y,U,u,V,v,w)$ to indicate the $\beta$-functions for the couplings \eq{couplings}. To obtain explicit expressions for these, we exploit the formal results summarised in \cite{Machacek:1983tz,Machacek:1983fi,Machacek:1984zw}.
The semi-simple character of the theory is switched off if the $N_\psi$ messenger fermions (which carry charges under both gauge groups) are replaced by
$N_1$ and $N_2$ Yukawa-less fermions in the fundamental of $SU(N_C)$ and $SU(N_c)$, respectively,
with
\begin{equation}\label{Relax}
\begin{array}{l}
N_1=N_c\,N_\psi\,,
\\
N_2=N_C\,N_\psi\,.
\end{array}
\end{equation}
If in addition $\alpha_w=0$, the theories \eq{L} reduce to a ``direct product'' of simple gauge Yukawa theories with \eq{Relax}.
Also, in the limit where one of the gauge groups is switched off, $\alpha_1\equiv 0$ (or $\alpha_2\equiv 0$), one gauge sector and the scalars decouples straightaway, and we are left with a simple gauge theory. Finally, if $N_1=0=N_2$, we recover the models of \cite{Litim:2014uca} in each gauge sector (displaying asymptotic safety for certain field multiplicities). Below, we will find it useful to contrast results with those from the ``direct product'' limit.
\subsection{Free parameters and Veneziano limit}
We now discuss the set of fundamentally free parameters of our models. On the level of the Lagrangean, the free parameters of the theory are the matter field multiplicities
\begin{equation}\label{Ns}
N_{\rm C},\quad N_{\rm c},\quad N_{\rm F},\quad N_{\rm f},\quad N_\psi\,.
\end{equation}
Notice that the $N_\psi$ fermions $\psi$ are centrally responsible for interactions between the gauge sectors. In the limit
\begin{equation}\label{N=0}
N_\psi=0
\end{equation}
the interaction between gauge sectors reduces to effects mediated by the portal coupling $\alpha_w\neq 0$, which are strongly loop-suppressed. In this limit, results for fixed points and running couplings fall back to those for the individual gauge sectors \cite{Litim:2014uca}.
Results for fixed points for general $N_\psi$ are deferred to App.~\ref{AppN}. Here, we will set $N_\psi$ to a finite value,
\begin{equation}\label{N=1}
N_\psi=1\,.
\end{equation}
This leaves us with four free parameters. In order to achieve exact perturbativity, we perform a Veneziano limit \cite{Veneziano:1979ec} by sending the number of colors and the number of flavors $(N_{\rm C}, N_{\rm c}, N_{\rm F},N_{\rm f})$ to infinity but keeping their ratios fixed. This reduces the set of free parameters of the model down to three, which we chose to be
\begin{equation}
\label{RST}
R=\frac{N_{\rm c}}{N_{\rm C}}\,,\quad
S=\frac{N_{\rm F}}{N_{\rm C}}\,,\quad
T=\frac{N_{\rm f}}{N_{\rm c}}\,.
\end{equation}
The ratio
\begin{equation}
\label{F}
F=\frac{N_{\rm f}}{N_{\rm F}}
\end{equation}
is then no longer a free parameter, but fixed as $F=R\,T/S$ from \eq{RST}. By their very definiton, the parameters \eq{RST} are positive semi-definite and can take values $0\le F,R,S,T\le \infty$.
However, we will see below that their values are further constrained if we impose perturbativity for all couplings.\subsection{Perturbativity to leading order}
The RG evolution of couplings is analysed within the perturbative loop expansion. To leading order (LO), the running of the gauge couplings reads $\beta_i=-B_i\,\al i^2$ (no sum), with $B_i$ the one-loop gauge coefficients for the gauge coupling $\alpha_i$. In the Veneziano limit, the one-loop coefficients take the form \begin{equation}\label{Bi}
B_i=-\frac43 \epsilon_i\,.
\end{equation}
In terms of \eq{RST} and in the Veneziano limit, the parameters $\epsilon_i$ are given by
\begin{equation} \label{eps12}
\begin{array}{rcl}
\epsilon_1 &=&\displaystyle S+ R- \frac{11}{2}\,,\\[2ex]
\epsilon_2 &= &\displaystyle T+ \frac{1}{R}- \frac{11}{2}\,.
\end{array}
\end{equation}
We can therefore trade the free parameters $(S,T)$ defined in \eq{RST} for $(\epsilon_1,\epsilon_2)$ and consider the set \begin{equation}\label{eps12R}
(\epsilon_1,\epsilon_2,R)
\end{equation}
as free parameters which characterise the matter content of the theory. Under the exchange of gauge groups we have
\begin{equation}\label{eps21}
(\epsilon_1,\epsilon_2,R)\to (\epsilon_2,\epsilon_1,R^{-1})\,.
\end{equation}
For fixed $R$, we observe that $R-\frac{11}{2}\le \epsilon_1<\infty$ and $1/R-\frac{11}{2}\le \epsilon_2<\infty$.
Perturbativity in either of the gauge couplings requires that both one-loop coefficients $B_i$ are parametrically small compared to unity. Therefore we impose
\begin{equation}\label{eps}
0<|\epsilon_i|\ll1\,.
\end{equation}
This requirement of exact perturbativity in both gauge sectors entails the important constraint
\begin{equation}\label{R}
\frac{2}{11}< R<\frac{11}{2}\,.
\end{equation}
Outside of this range, no physical values for $S$ and $T$ can be found such that \eq{eps} holds true. Inside this range, physical values are constrained within $0\le S,T\le \0{11}2-\02{11}$.
The parameters \eq{eps12R} have a simple interpretation. The small paramaters $\epsilon_i$ control the perturbativity within each of the gauge sectors, whereas the parameter $R$ controls the ``interactions'' between the two gauge sectors. It is the presence of $R$ which makes these theories intrinsically semi-simple, rather than being the direct product of two simple gauge theories. Perturbativity is no longer required in the limit where one of the gauge sectors is switched off, and the constraint \eq{R} is relaxed into
\begin{equation}\label{Relax2}
\begin{array}{rcc}
\displaystyle
0\le R<\frac{11}{2}&\quad\text{if}&\quad \alpha^*_2\equiv 0\,,\\[1ex]
\displaystyle
\frac{2}{11}< R<\infty&\quad\text{if}&\quad \alpha^*_1\equiv 0\,.
\end{array}
\end{equation}
The parametrisation \eq{eps12R} is most convenient for expressing the relevant RG beta functions for all couplings.
Finally, for some of the subsequent considerations we replace the two small parameters $(\epsilon_1,\epsilon_2)$ by $(\epsilon,P)$, a single small parameter $\epsilon$ proportional to $\epsilon_1$ together with a parameter $P$ related to the ratio between $\epsilon_1$ and $\epsilon_2$. Specifically, we introduce
\begin{equation}
\label{P}
\begin{array}{rcl}
\epsilon_1&=&\, R\,\epsilon\,,\\[1ex]
\epsilon_2&=&\displaystyle
P\,\0{\epsilon}{R}\,.
\end{array}
\end{equation}
which is equivalent to $P= R^2\,{\epsilon_2}/{\epsilon_1}$ together with $\epsilon=\epsilon_1/R$ and $\epsilon=P\,R\,\epsilon_2$.\footnote{The choice \eq{P} can be motivated by dimensional analysis of \eq{eps12} which shows that $\epsilon_1$ and $\epsilon_2$ formally scale as $\sim R$ and $\sim1/R$ for large or small $R$, respectively, whereby their ratio $\epsilon_1/\epsilon_1$ scales as $\sim R^2$. The large-$R$ behaviour is factored-out by our parametrisation.} Since $R$ can only take finite positive values, the additional rescaling with $R$ does not affect the relative sign between $\epsilon_1$ and $\epsilon$. In this manner we have traded the free parameters $(\epsilon_1,\epsilon_2,R)$ for
\begin{equation}
\label{epsPR}
(R,P,\epsilon)\,.
\end{equation}
Notice that the parameter $P$ can be expressed as
\begin{equation}\label{Pexplicit}
P=\frac{1+(N_{\rm f}-\0{11}{2} N_{\rm c})/N_{\rm C}}{1+(N_{\rm F}-\0{11}{2}N_{\rm C})/N_{\rm c}}
\end{equation}
in terms of the field multiplicities \eq{Ns}. It thus may take any real value of either sign with $-\infty<P<\infty$, whereas $R$ must take values within the range \eq{R}. Moreover,
\begin{equation}
\epsilon=1+\frac{N_{\rm F}-\s0{11}{2}N_{\rm C}}{N_{\rm c}}\,.
\end{equation}
In this parametrisation, the ratio of fermion flavour multiplicities \eq{F} becomes
\begin{equation}\label{epsF}
F=\frac{11 R-2}{11-2R}
+\frac{2R}{11-2R}\left(\frac{P}{R}-\frac{11R-2}{11-2R}\right)\epsilon
+{\cal O}(\epsilon^2)\,.
\end{equation}
We also observe that the substitution
\begin{equation}\label{exchange}
(R,P,\epsilon)\to \left(R^{-1},P^{-1},P\epsilon\right)
\end{equation}
relates to the exchange of gauge groups. The parametrisation \eq{epsPR} is most convenient for analysing the various interacting fixed points and their scaling exponents (see below).
This completes the definition of our models.
\subsection{Anomalous dimensions}\label{anomalous}
We provide results for the anomalous dimensions associated to the fermions and scalars. Furthermore, if mass terms are present, their renormalisation is induced through the RG flow of the gauge, Yukawa, and scalar couplings. Following \cite{Litim:2014uca}, we define the scalar anomalous dimensions as
$\Delta_S=1+\gamma_S$, where $\gamma_S\equiv \frac{1}{2}{d\ln Z_S}/{d\ln\mu}$
and $S=H,h$. Within perturbation theory, the one and two loop contributions read
\begin{equation}
\label{gammaH}
\begin{array}{rcl}
\gamma_H&=&
\displaystyle
\alpha_Y
-\032\left(\0{11}{2}
-\epsilon_1-R\right)\alpha_Y^2
+\052\alpha_Y\,\alpha_1
+2\alpha_U^2 +{\cal O}(\alpha^3) \,,
\\[2ex]
\gamma_h&=&
\displaystyle
\,\alpha_y
-\032\left(\0{11}{2}+\epsilon_2-\01R\right)\alpha_y^2
+\052\alpha_y\,\alpha_2
+2\alpha_u^2 +{\cal O}(\alpha^3) \,.
\end{array}
\end{equation}
For the fermion anomalous dimensions $\gamma_F\equiv {d\ln Z_F}/{d\ln\mu}$
with $F=Q,q,\psi$, we find
\begin{equation}
\label{gammaH}
\begin{array}{rcl}
\gamma_Q&=&
\displaystyle
\left(\0{11}{2}+\epsilon_1-R\right)\alpha_Y
+\xi_1\,\alpha_1 +{\cal O}(\alpha^2) \,,
\\[2ex]
\gamma_q&=&
\displaystyle
\left(\0{11}{2}+\epsilon_2-\01R\right)\alpha_Y
+\xi_2\,\alpha_2 +{\cal O}(\alpha^2) \,,
\\[2ex]
\gamma_\psi&=&
\displaystyle
\xi_1\,\alpha_1
+\xi_2\,\alpha_2 +{\cal O}(\alpha^2) \,,
\end{array}
\end{equation}
where $\xi_1$ and $\xi_2$ denote the gauge fixing parameters for the first and second gauge group respectively.
The anomalous dimension for the scalar mass terms can be derived from the composite operator $\sim M^2\,{\rm Tr}\, H^\dagger H$ and $\sim m^2\,{\rm Tr}\, h^\dagger h$. Introducing the mass anomalous dimension $\gamma_{M}= d\ln M^2/d \ln \mu$, and similarly for $m$,
one finds
\begin{equation}\label{gammam}
\begin{array}{rcl}
\gamma_M&=&8\al U+4 \al V+2\al Y+{\cal O}(\alpha^2) \\[1ex]
\gamma_m&=&8\al u\,+4 \al v\; +2\al y\ +{\cal O}(\alpha^2) \,,
\end{array}
\end{equation}
to one-loop order. We also compute the running of the mass terms for the scalars
\begin{equation}\label{massRunS}
\begin{array}{rcl} \beta_{M^2} &=& \gamma_M\, M^2
+ 2 F \, m^2\, \alpha_w
+ {\cal O}(\alpha^2,\alpha\,m^2_F) \,,\\[1ex]
\beta_{m^2} &=& \gamma_m\, m^2
+ 2F^{-1}\,M^2\,\alpha_w + {\cal O}(\alpha^2,\alpha\,m^2_F) \,,
\end{array}
\end{equation}
where the parameter $F\equiv N_{\rm f}/N_{\rm F}$ solely depends on $R$ to leading order in $\epsilon$, see \eq{epsF}. Notice that the coupling $\alpha_w$ induces a mixing amongst the different scalar masses already at one-loop level.
Analogously, the anomalous dimension for the fermion mass operator is defined as
$\Delta_F=3+\gamma_{M_F}$ with $\gamma_{M_F}\equiv {d\ln M_F}/{d\ln\mu}$,
and $M_F$ stands for one of the fermion masses with $F=Q,q$ or $\psi$. Within perturbation theory, the one loop contributions read
\begin{equation}
\label{gammaF}
\begin{array}{rcl}
\gamma_{M_Q}&=&
\displaystyle
\alpha_Y\, \left(\0{13}{2}+\epsilon_1-R\right)-3\, \alpha_1-+{\cal O}(\alpha^2) \ ,\\[2ex]
\gamma_{M_q}&=&
\displaystyle
\alpha_y\, \left(\0{13}{2}+\epsilon_2-\01R\right)-3\, \alpha_2+{\cal O}(\alpha^2) \\[2.5ex]
\gamma_{M_\psi}&=&
-3\, (\alpha_1+\alpha_2)+{\cal O}(\alpha^2) \,.
\end{array}
\end{equation}
For the fermion masses we have the running
\begin{align}
\nonumber
\beta_{m_Q} &= \gamma_{M_Q}\,m_Q
\,,\\
\label{massRunF}
\beta_{m_q} &= \gamma_{M_q}\,m_q
\,,\\
\nonumber
\beta_{m_\psi} &= \gamma_{M_\psi}\,m_\psi
\,.
\end{align}
We note that $\gamma_{M_\psi}$ is manifestly negative. For $\gamma_{M_Q}$ and $\gamma_{M_q}$ we observe that the gauge and Yukawa contributions arise with manifestly opposite signs in the parameter regime \eq{eps}, \eq{R}. Hence either of these may take either sign, depending on whether the gauge or Yukawa contributions dominate.
\subsection{Running couplings beyond the leading order}
We now go beyond the leading order in perturbation theory and provide the complete, minimal set of RG equations which display exact and weakly interacting fixed points. To that end, we must retain terms up to two loop order in the gauge coupling, or else an interacting fixed point cannot arise. At the same time, in order to explore the feasibility of asymptotically safe UV fixed points we must retain the Yukawa couplings \cite{Bond:2016dvk}, minimally at the leading non-trivial order which is one loop. Following \cite{Litim:2014uca} we refer to this level of approximation in the gauge-Yukawa sector as next-to-leading order (NLO). In the presence of scalar fields, we also must retain the quartic scalar couplings at their leading non-trivial order. We refer to this approximation of the gauge-Yukawa-scalar sector as NLO${}^\prime$ \cite{Litim:2015iea}, see Tab.~\ref{Tab}. This is the minimal order in perturbation theory at which a fully interacting fixed point can be determined in all couplings with canonically vanishing mass dimension.
In general, the RG flow for the gauge and Yukawa couplings at NLO${}^\prime$ is strictly independent of the scalar couplings owing to the fact that scalar loops only arise starting from the two loop order in the Yukawa sector, and at three (four) loop order in the gauge sector, if the scalars are charged (uncharged). Furthermore, the scalar sector at NLO${}^\prime$ depends on the Yukawa couplings, but not on the gauge couplings owing to the fact that the scalars are uncharged. Consequently, we observe a partial decoupling of the gauge-Yukawa sector
$(\alpha_1,\alpha_2,\alpha_Y,\alpha_y)$ and the scalar sector $(\alpha_U,\alpha_V,\alpha_u,\alpha_v,\alpha_{w})$. This structure will be exploited systematically below to identify all interacting fixed points.
We begin with the gauge-Yukawa sector where we find the coupled beta functions \eq{12Yy} which are characterised by ten loop coefficients $C_i, D_i, E_i, F_i$ and $G_i$ $(i=1,2)$, together with the coefficients $B_i$ given in \eq{Bi} or, equivalently, the perturbative control parameters \eq{eps12}. The one-loop coefficients arise in the Yukawa sector and take the values
\begin{equation} \label{EF}
\begin{array}{lcl}
E_1 = 13 + 2\left(\epsilon_1 - R\right)\,,
\quad \quad &&
F_1=6\,,
\\[1ex]
\displaystyle
E_2 = 13 + 2\left(\epsilon_2 - \frac{1}{R}\right)\,,\quad\quad &&
F_2=6\,.
\end{array}
\end{equation}
At the two-loop level we have six coefficients related to the gauge, Yukawa, and mixing contribution, which are found to be
\begin{equation} \label{CDG}
\begin{array}{lcl}
&&
C_1 =
\displaystyle
25 + \frac{26}{3}\epsilon_1\,,\quad
D_1 = 2\left(\epsilon_1-R+ \frac{11}{2}\right)^2\,,\quad
G_1=2 R\\[2ex]
&&
\displaystyle
C_2 = 25 + \frac{26}{3}\epsilon_2\,,\quad
D_2 = 2\left(\epsilon_2- \frac{1}{R}+ \frac{11}{2}\right)^2\,,\quad
G_2=\frac{2}{R}
\end{array}
\end{equation}
A few comments are in order.
Firstly, the loop coefficients $D_i, E_i, F_i, G_i>0$ as they must for any quantum field theory.
Additionally we confirm that $C_i>0$ \cite{Bond:2016dvk}, provided the parameters $\epsilon_i$ are in the perturbative regime \eq{eps}.
Secondly, provided that $R=0$ in the expressions for $\epsilon_1, E_1$ and $G_1$,
and $1/R=0$ in those for
$\epsilon_2, E_2$ and $G_2$,
the system \eq{12Yy} falls back onto a direct product of simple gauge-Yukawa theories, each of the type discussed in \cite{Litim:2014uca}. Notice that this limit cannot be achieved parametrically in $R$. The reason for this is the presence of $N_\psi$ fermions which are charged under both gauge groups. They contribute with reciprocal multiplicity $R\leftrightarrow 1/R$ to the Yukawa-induced loop terms $D_i$ and $E_i$ as well as to the mixing terms $G_i$. Exact decoupling of the gauge sectors then becomes visible only in the parametric limit where $N_\psi\to 0$ whereby all terms involving $R$ or $1/R$ drop out.
Finally, we note that the exchange of gauge groups ${\cal G}_1\leftrightarrow {\cal G}_2$ corresponds to $R\leftrightarrow 1/R$ and $S\leftrightarrow T$, implying $\epsilon_1\leftrightarrow \epsilon_2$ and $P\leftrightarrow 1/P$, respectively. Evidently, at the symmetric point $R=1$ and $\epsilon_1=\epsilon_2$ (or $P=1$) we have exact exchange symmetry between gauge group factors.
Inserting \eq{EF}, \eq{CDG} and \eq{Bi} into the general expression \eq{12Yy}, we obtain the perturbative RG flow for the gauge-Yukawa system at NLO accuracy
\begin{equation}\label{12Yyexplicit}
\begin{array}{rcl}
\beta_1&=&
\displaystyle
\043\epsilon_1\, \alpha_1^2+\left(25 + \frac{26}{3}\epsilon_1\right)\alpha_1^3 -2\left(\epsilon_1-R+ \frac{11}{2}\right)^2\alpha_1^2\,\al{Y}+2R\,\al1^2\,\al2\,,\\[1.5ex]
\beta_2&=&
\displaystyle
\043\epsilon_2\, \al2^2+\left(25 + \frac{26}{3}\epsilon_2\right)\al2^3 -2\left(\epsilon_2- \frac{1}{R}+ \frac{11}{2}\right)^2\al2^2\,\al y +\02R\,\al2^2\,\al1\,,\\[2.5ex]
\beta_Y&=&
\displaystyle
\left[13 + 2\left(\epsilon_1 - R\right)\left]\al{Y}^2\right.\right.-6\,\al{Y}\,\al 1\,,\\[2ex]
\beta_y&=&
\displaystyle
\left[13 + 2\left(\epsilon_2 - \s0{1}{R}\right)\right]\al{y}^2-6\,\al{y}\,\al 2\,.
\end{array}
\end{equation}
We observe that the running of Yukawa couplings at one loop is determined by the fermion mass anomalous dimension \eq{gammaF},
\begin{equation}\label{YyQq}
\begin{array}{rcl}
\beta_Y&=&2\,\gamma_{M_Q}\,\alpha_Y\,,\\
\beta_y&=&2\,\gamma_{M_q}\,\alpha_y\,.
\end{array}
\end{equation}
The result for the mass anomalous dimensions \eq{gammaF} can also be derived diagrammatically from the flow of the Yukawa vertices \eq{12Yyexplicit}, thus offering an independent confirmation for the link \eq{YyQq}.
Next, we turn to the scalar sector and the running of quartic couplings to leading order in perturbation theory, which is one loop.
At NLO${}^\prime$ accuracy, we have \eq{12Yyexplicit} together with the beta functions for the quartic scalar couplings which are found to be
\begin{equation}\label{HVhv}
\begin{array}{rcl}
\beta_U&=&-[11+2(\epsilon_1-R)]\alpha_Y^2+4\alpha_U(\alpha_Y+2\alpha_U)\,,\\[1ex]
\beta_V&=&\ \ 12\alpha_U^2+4\alpha_V(\alpha_V+4 \alpha_U +\alpha_Y)+\alpha_w^2\,,\\[1ex]
\beta_u&=&-[11+2(\epsilon_2-\s01R)]\alpha_y^2+4\alpha_u(\alpha_y+2\alpha_u)\,,\\[1ex]
\beta_v&=&\ \ 12\alpha_u^2+4\alpha_v(\alpha_v+4 \alpha_u +\alpha_y)+\alpha_w^2\,,\\[1ex]
\beta_w&=&\ \ \alpha_w\left[8(\alpha_U+\alpha_u)+4(\alpha_V+\alpha_v)
+2(\alpha_Y+\alpha_y)\right]\,.\end{array}
\end{equation}
Their structure is worth a few remarks: Firstly, in the Veneziano limit, $\beta_w$ contains no term quadratic in the coupling $\alpha_w$ as the coefficient is of the order ${\cal O}(N^{-1}_{\rm F} N^{-1}_{\rm f})$ and suppressed by inverse powers in flavour multiplicities. Secondly, we notice that $\beta_w$ comes out proportional to $\alpha_w$.
Consequently, $\alpha_w$ is a technically natural coupling according to the rationale of \cite{'tHooft:1979bh}, unlike all the other quartic interactions, implying that
\begin{equation}\label{wFP}
\alpha_w^*=0
\end{equation}
constitutes an exact fixed point of the theory. Comparison with \eq{gammam} shows that the proportionality factor is the sum of the scalar mass anomalous dimensions, $\beta_w=\alpha_w(\gamma_M+\gamma_m)$. The quartic coupling $\alpha_w$ would be promoted to a free parameter characterising a line of fixed points with exactly marginal scaling provided that its beta function vanishes identically at one loop. This would require the vanishing of the sum of scalar anomalous mass dimensions at the fixed point,
\begin{equation}
\gamma^*_M+\gamma^*_m=0\,.
\end{equation}
Below, however, we will establish that such scenarios are incompatible with vacuum stability (see Sect.~\ref{stab}).
Moreover, at interacting fixed points we invariably find that
\begin{equation}\label{gammaPos}
\gamma^*_M+\gamma^*_m>0
\end{equation}
as a consequence of vacuum stability. This implies that
$\alpha_w$
constitutes an infrared free coupling at any interacting fixed point with a stable ground state.
For the purpose of the present study, we therefore limit ourselves to fixed points with \eq{wFP}.
We then observe that the running of the remaining scalar couplings is solely fuelled by the Yukawa couplings. Furthermore, the scalar subsectors associated to the different gauge groups are disentangled in our approximation.\footnote{The degeneracy is lifted as soon as the quartic coupling $\alpha_w\neq 0$, see \eq{L}, \eq{couplings}.} Interestingly, the beta functions for $(\alpha_U,\alpha_V)$ and $(\alpha_u,\alpha_v)$ are related by the substitution $R\leftrightarrow 1/R$ and $\epsilon_1\leftrightarrow \epsilon_2$. Moreover, the double trace scalar couplings do not couple back into any of the other couplings and their fixed points are entirely dictated by the corresponding single trace scalar and the Yukawa coupling \cite{Litim:2014uca}. This structure allows for a straightforward systematic analysis of all weakly coupled fixed points of the theory to which we turn next.
\section{\bf Interacting fixed points}\label{FPs}
In this section, we present our results for exact fixed points in the Veneziano limit, corresponding to interacting conformal field theories, and the universal scaling exponents in their vicinity.
\subsection{Parameter space}\label{AppX}
In Tab.~\ref{tFPeps} we state our results for the gauge and Yukawa couplings to leading order in \eq{eps} at all fixed points, following the nomenclature of Tab.~\ref{tFPs}. Expressions are given as functions of the parameters $(P,R,\epsilon)$,
\begin{equation}\label{PRepsN}
\begin{array}{rcl}
P&=&
\displaystyle
\frac{1+(N_{\rm f}-\0{11}{2} N_{\rm c})/N_{\rm C}}{1+(N_{\rm F}-\0{11}{2}N_{\rm C})/N_{\rm c}}\\[2.5ex]
R&=&
\displaystyle
\frac{N_{\rm c}}{N_{\rm C}}\\[2.5ex]
{\rm sgn}\,\epsilon&=&{\rm sgn}\left(N_{\rm c}+N_{\rm F}-\frac{11}{2}N_{\rm C}\right)\,,\\[.5ex]
\end{array}
\end{equation}
which only depend on the matter and gauge field multiplicities \eq{Ns}, and $N_\psi=1$. Results for general $N_\psi$ are given in App.~\ref{AppN}. We also observe \eq{R}, unless stated otherwise.
Constraints on the parameters $(R,P,\epsilon)$
and other information is summarised Figs.~\ref{pFP23},~\ref{pFP45},~\ref{pFP6},~\ref{pFP78} and~\ref{pFP9} and in Tabs.~\ref{tBprime}~\ref{tPara15},~\ref{tPara69} for the various fixed points.
Below, certain characteristic values $\frac{2}{11}<R_1<R_2<R_3<R_4<\frac{11}2$ for the parameter $R$ are of particular interest, namely
\begin{equation}\label{R1234}
\begin{array}{rcl}
R_1 &= &
\displaystyle
\frac{343 - 3\sqrt{9361}}{100} \approx 0.53\,, \\ [1.2ex]
R_2 &=& \displaystyle
\frac{43 - 9\sqrt{5}}{38} \approx 0.60\,,\\[1.2ex]
R_3 &=&\displaystyle
\frac{43 + 9\sqrt{5}}{38} \approx 1.66\,, \\[1.2ex]
R_4 &=&\displaystyle
\frac{343 + 3\sqrt{9361}}{100}\approx 1.90\,.
\end{array}
\end{equation}
Their origin is explained in App.~\ref{AppX}. After these preliminaries we are in a position to analyse the fixed point spectra.
\begin{table*}[b]
\begin{center}
\begin{tabular}{cllcc}
\toprule
\rowcolor{LightBlue}
&
\cellcolor{LightRed}
&\cellcolor{LightYellow}
&\cellcolor{LightGreen}
\\[-1mm]
\cellcolor{LightBlue}
\multirow{-2}{*}{$\bm{\#}$}&
\multirow{-2}{*}{\cellcolor{LightRed} \quad\quad\quad\quad\quad\bf gauge couplings}
&\multirow{-2}{*}{\cellcolor{LightYellow} \quad\quad\bf Yukawa couplings}
&
\multirow{-2}{*}{\cellcolor{LightGreen} \bf type}
\\
\midrule
\rowcolor{LightGray}
\bf \ FP${}_{\bf 1}\ $
&$\alpha^*_1=0\,,\quad\quad\quad\alpha^*_2=0\,,$
&$\alpha^*_Y=0\,,\quad\quad\quad\alpha^*_y=0\,,$
& \bf G $\bm \cdot$ G
\\[1ex]
&&& \\[-2.5ex]
\bf \ FP${}_{\bf 2}\ $
&
$\alpha^*_1=-\s0{4}{75}R\epsilon\,,\ \, \alpha^*_2=0\,,$
&
$\alpha^*_Y=0\,,\quad\quad\quad \alpha^*_y=0\,,$&
\bf BZ $\bm \cdot$ G
\\[1ex]
\rowcolor{LightGray}
&&& \\[-2.5ex]
\rowcolor{LightGray}
\bf FP${}_{\bf 3}$
&$\alpha^*_1=0\,,\quad\quad\quad\alpha^*_2=-\s0{4}{75}\frac{P\epsilon}{R}\,,$
&
$\alpha^*_Y=0\,,\quad\quad\quad \alpha^*_y=0\,,$
& \bf G $\bm \cdot$ BZ
\\[1ex]
&&& \\[-2.5ex]
\bf FP${}_{\bf 4}$&
$\alpha^*_1=\s0{2}3 \s0{(13-2R)R\epsilon}{(2R-1)(3R-19)}\,,\quad\quad\ \ \alpha^*_2=0\,,$
&
$\alpha^*_Y=\s0{4R\epsilon}{(2R-1)(3R-19)}\,,\quad \alpha^*_y=0\,,$
& \bf GY $\bm \cdot$ G
\\[1ex]
\rowcolor{LightGray}
&&& \\[-2.5ex]
\rowcolor{LightGray}
\bf FP${}_{\bf 5}$
&$ \alpha^*_1=0\,,\quad \alpha^*_2=\0{2}3 \frac{(13-2/R)}{(2/R-1)(3/R-19)} \frac{P\epsilon}{R}\,,$
&
$\alpha^*_Y=0\,,\ \alpha^*_y=\frac{4\,P\epsilon/R}{(2/R-1)(3/R-19)}\,,$
&\bf G $\bm \cdot$ GY
\\[1ex]
&&& \\[-2.5ex]
\bf FP${}_{\bf 6}$&
$\alpha^*_1= \0{-4(25 - 2P/R)}{1863} R\epsilon\,,\quad
\alpha^*_2=\0{-4(25 - 2R/P)}{1863}\frac{P\epsilon}{R}$
&$\alpha^*_Y=0\,,\quad\quad\quad \alpha^*_y=0\,,$
&\bf BZ $\bm \cdot$ BZ
\\[1ex]
\rowcolor{LightGray}&&& \\[-2.5ex]
\rowcolor{LightGray}
&
$\alpha^*_1=\0{2}9\frac{(13-2R)(25 - 2P/R)}{50R^2-343R+167} R\epsilon$
&
$\alpha^*_Y=\0{4}3 \frac{25 - 2P/R}{50R^2-343R+167} R\epsilon$
&
\\[1ex]
\rowcolor{LightGray}
\multirow{-2}{*}{\bf FP${}_{\bf 7}$}
&
$\alpha^*_2= \0{4}9\frac{(13-2R)R/P + (2R-1)(19-3R)}{50R^2 - 343R + 167} \frac{P\epsilon}{R}$
&$\alpha^*_y=0\quad\quad\quad\quad\quad\quad\quad\ \ $&
\multirow{-2}{*}{\bf GY $\bm \cdot$ BZ}
\\[1ex]
&&& \\[-2.5ex]&
$\alpha^*_1=\0{4}9\frac{(13-2/R)P/R + (2/R-1)(19-3/R)}{50/R^2 - 343/R + 167} R\epsilon$
&$\alpha^*_Y=0\quad\quad\quad\quad\quad\quad\quad\ \ $
\\[1ex]
\multirow{-2}{*}{\bf FP${}_{\bf 8}$}
&
$\alpha^*_2=\0{2}9\frac{(13-2/R)(25 - 2R/P)}{50/R^2-343/R+167} \frac{P\epsilon}{R}
\quad\quad\ \ $
&
$\alpha^*_y=\0{4}3 \frac{25 - 2R/P}{50/R^2-343/R+167} \frac{P\epsilon}{R}$
&
\multirow{-2}{*}{\bf BZ $\bm \cdot$ GY}
\\[1ex]
\rowcolor{LightGray}&&& \\[-2.5ex]
\rowcolor{LightGray}
&
$\alpha^*_1=\0{2}9\0{(13-2R)[(13-2/R)P/R+(\s02R-1)(3/R-19)]}{(19R^2 - 43R + 19)(2/R^2 -13/R + 2)}R\epsilon\ $
&
$\alpha^*_Y=\0{6\,\alpha_1^*}{13-2R}$
&\\[1ex]
\rowcolor{LightGray}
\multirow{-2}{*}{\bf FP${}_{\bf 9}$}
&
$\alpha^*_2=
\0{2}9\0{(13-2/R)[(13-2R)R/P+(2R-1)(3R-19)]}{(19/R^2 - 43/R + 19)(2R^2 -13R + 2)} \frac{P\epsilon}{R}$
&
$\alpha^*_y=
\0{6\,\alpha_2^*}{13-2/R}$
&
\multirow{-2}{*}{\bf GY $\bm \cdot$ GY\ }\\[1ex]
\bottomrule
\end{tabular}
\caption{Gauge and Yukawa couplings at interacting fixed points following Tab.~\ref{tFPs} to the leading order in $\epsilon$ and in terms of $(R,P,\epsilon)$. Valid domains for $(\epsilon,P,R)$ in \eq{PRepsN} are detailed in Tab.~\ref{tPara15},~\ref{tPara69}.\\ }
\label{tFPeps}
\end{center}
\end{table*}
\subsection{Partially and fully interacting fixed points}
\label{pifp}
Gauge theories with \eq{12Yyexplicit}, \eq{HVhv} can have two types of interacting fixed points: partially interacting ones where one gauge coupling takes the Gaussian fixed point (\fp2,\fp3,\fp4,\fp5), and fully interacting ones where both gauge sectors remain interacting (\fp6,\fp7,\fp8,\fp9), see Tab.~\ref{tFPs}. At partially interacting fixed points, one gauge sector decouples and the semi-simple theory with \eq{12Yyexplicit}, \eq{HVhv} effectively reduces to a simple gauge theory. Simple gauge theories have three possible types of perturbative fixed points: the Gaussian (G), the Banks-Zaks (BZ), and gauge-Yukawa (GY) fixed points for each independent linear combination of the Yukawa couplings \cite{Bond:2016dvk}. In our setting, at \fp2 and \fp4 we have that $\alpha_2^*\equiv 0$, and the theory reduces to a simple gauge theory with
\begin{equation}\label{Simple}
\begin{array}{rcl}
\beta_1&=&
\displaystyle
\ \ \043\epsilon_1\, \alpha_1^2+\left(25 + \frac{26}{3}\epsilon_1\right)\alpha_1^3 -2\left(\epsilon_1-R+ \frac{11}{2}\right)^2\alpha_1^2\,\al{Y}\\[2ex]
\beta_Y&=&\displaystyle
\ \ \left[13 + 2\left(\epsilon_1 - R\right)\left]\al{Y}^2\right.\right.-6\,\al{Y}\,\al 1\\[1.5ex]
\displaystyle
\beta_U&=&-[11+2(\epsilon_1-R)]\alpha_Y^2+4\alpha_U(\alpha_Y+2\alpha_U)\,,\\[1ex]
\beta_V&=&\ \ 12\alpha_U^2+4\alpha_V(\alpha_V+4 \alpha_U +\alpha_Y)\,,
\end{array}
\end{equation}
at NLO${}^\prime$ accuracy, where the parameter $R$ with
\begin{equation}\label{Rrelaxed}
0\le R=\frac{N_1}{N_{\rm C}}<\frac{11}{2}
\end{equation}
measures the number of Yukawa-less Dirac fermions $N_1$ in the fundamental representation in units of $N_{\rm C}$. Notice that $N_1$ is related to $N_\psi$ via \eq{Relax} in the theories \eq{L}.
On the other hand, $N_1$ can be viewed as an independent parameter (counting the Yukawa-less fermions in the fundamental representation of the gauge group) if one were to switch off the semi-simple character of the theory.
For $R=0$ the theory \eq{Simple} reduces to the one investigated in \cite{Litim:2014uca}. The lower bound on $R$ \eq{R} is relaxed in \eq{Rrelaxed}, because the requirement of perturbativity for an interacting fixed point in the other gauge sector has become redundant.
We observe the $R$-independent Banks-Zaks (BZ) fixed point $\alpha_1^*=\frac{4}{75}\epsilon_1$ which is, invariably, IR.
To leading order in $\epsilon_1$ we also find a gauge-Yukawa (GY) fixed point
\begin{equation}\label{GYsimple}
\begin{array}{rcl}
\alpha_g^*&=&
\displaystyle
\frac{26-4R}{57-9 R}\,\frac{\epsilon_1}{1-2 R}\\[2ex]
\alpha_Y^*&=&
\displaystyle
\frac{4}{19-3 R}\frac{\epsilon_1}{1-2 R}\\[2ex]
\alpha_U^*&=&
\displaystyle
\frac{\sqrt{23-4R}-1}{19-3 R}\,\frac{\epsilon_1}{1-2 R}\\[2ex]
\alpha_V^*&=&
\displaystyle
\frac{-2\sqrt{23-4R}+
\sqrt{20-4R+6\sqrt{23-4R}}}{19-3 R}\frac{\epsilon_1}{1-2 R}\,.
\end{array}
\end{equation}
For $\epsilon_1>0$, the GY fixed point is UV and physical as long as $0\le R<\frac12$. It can be interpreted as a ``deformation'' of the UV fixed point analysed in \cite{Litim:2014uca} owing to the presence of charged Yukawa-less fermions. Once $R>\frac12$, however, the fixed point is physical iff $\epsilon_1<0$ where it becomes an IR fixed point. This new regime is entirely due to the Yukawa-less fermions and does not arise in the model of \cite{Litim:2014uca}. This pattern can also be read off from the scaling exponents, which, at the gauge Yukawa fixed point and to the leading non-trivial order in $\epsilon_1$, are given by
\begin{equation}\label{evGYsimple}
\begin{array}{rcl}
\vartheta_g&=&
\displaystyle
-\frac{104}{171}\,
\frac{1-\s02{13}R}{1-\s03{19}R}\,
\frac{\epsilon_1^2}{1-2 R}\\[2.5ex]
\vartheta_y&=&
\displaystyle
\ \ \frac{4}{19}\,\frac{1}{1-\s03{19}R}\frac{\epsilon_1}{1-2 R}\\[2.5ex]
\vartheta_u&=&\displaystyle
\ \ \frac{16\sqrt{23-4R}}{19-3 R}\,\frac{\epsilon_1}{1-2 R}\\[2.5ex]
\vartheta_v&=&\displaystyle
\ \ \frac{8\sqrt{20+6\sqrt{23-4R}-4R}}{19-3 R}\,\frac{\epsilon_1}{1-2 R}\,.
\end{array}
\end{equation}
For $\epsilon_1>0$ and $R<\s012$ asymptotic safety is guaranteed with
$\vartheta_g<0<\vartheta_y, \vartheta_u, \vartheta_v\,,$
showing that the UV fixed point has one relevant direction. The scaling exponents reduce to those in \cite{Litim:2014uca} for $R=0$. Conversely, for $\epsilon_1<0$ and $R>\s012$ the theory is asymptotically free and the interacting fixed point is fully IR attractive with
$0<\vartheta_g,\vartheta_y, \vartheta_u, \vartheta_v\,.$
Results straightforwardly translate to the partially interacting fixed points \fp3 and \fp5 where $\alpha_1^*\equiv 0$. The explicit $\beta$-functions in the other gauge sector are found from \eq{Simple} -- \eq{evGYsimple} via the replacements $\epsilon_1\leftrightarrow\epsilon_2$ and $R\leftrightarrow 1/R$, leading to
\begin{equation}\label{Simple2}
\begin{array}{rcl}
\beta_2&=&
\displaystyle
\ \ \; \043\epsilon_2\, \alpha_2^2+\left(25 + \frac{26}{3}\epsilon_2\right)\alpha_2^3 -2\left(\epsilon_2-\frac1R+ \frac{11}{2}\right)^2\alpha_2^2\,\al{y}\\[2ex]
\beta_y&=&\displaystyle
\ \ \; \left[13 + 2\left(\epsilon_2 - \frac1R\right)\right]\al{y}^2-6\,\al{y}\,\al 2\\[2ex]
\displaystyle
\beta_u&=&
\displaystyle
-\left[11+2(\epsilon_2-\frac1R)\right]\alpha_y^2+4\alpha_u(\alpha_y+2\alpha_u)\,,\\[2.5ex]
\beta_v&=&\ \ \; 12\alpha_u^2+4\alpha_v(\alpha_v+4 \alpha_u +\alpha_y)\,.
\end{array}
\end{equation}
Evidently, the coordinates of the fully interacting gauge-Yukawa fixed point and the corresponding universal scaling exponents of \eq{Simple2} are given by \eq{GYsimple}, \eq{evGYsimple} after obvious replacements.
Moreover, in \eq{Simple2} the parameter $R$ with
\begin{equation}\label{Rrelaxed2}
0\le \frac1R=\frac{N_2}{N_{\rm c}}<\frac{11}{2}
\end{equation}
measures the number of Yukawa-less Dirac fermions $N_2$ in the fundamental representation in units of $N_{\rm c}$, see \eq{Relax}.
The only direct communication between the different gauge sectors in \eq{L} is through the off-diagonal two-loop gauge contributions $G_i$. Were it not for the fermions $\psi$ which are charged under both gauge groups, the theory \eq{L} with \eq{12Yyexplicit}, \eq{HVhv} would be the ``direct product'' of the simple model \eq{Simple}, \eq{Rrelaxed} with its counterpart \eq{Simple2}, \eq{Rrelaxed2}. In this limit we will find nine ``direct product'' fixed points with scaling exponents from each pairing of the possibilities (G, BZ, GY) in each sector.
Below, we contrast findings for the full semi-simple setting \eq{12Yyexplicit}, \eq{HVhv} with those from the ``direct product'' limit in order to pin-point effects which uniquely arise from the semi-simple character of the theories \eq{L}.
At any of the partially interacting fixed points, the semi-simple character of the theory becomes visible in the non-interacting sector. In fact,
contributions from the $\psi$ fermions modify the effective one-loop coefficient $B_i\to B_i'$ according to
\begin{equation}\label{Bprime}
\begin{array}{l}
\alpha_1^*=0:\quad B_1\to B_1'=B_1+G_1\,\alpha^*_2\\[1ex]
\alpha_2^*=0:\quad B_2\to B_2'=B_2+G_2\,\alpha^*_1\,.
\end{array}
\end{equation}
No such effects can materialize in a ``direct product'' limit. Moreover, these contributions always arise with a positive coefficient ($B'>B)$ and are absent if $N_\psi=0$ (where $G_i=0)$. For $N_\psi\neq 0$, asymptotic freedom can thereby be changed into infrared freedom, but not the other way around.
This result is due to the fact that the Yukawa couplings are tied to individual gauge groups, and so by this structure we cannot have any Yukawa contributions to $B'$.
In principle, the opposite effect can equally arise: it would require Yukawa couplings which contribute to both gauge coupling $\beta$-functions, and would therefore have to involve at least one field which is charged under both gauge groups \cite{Bond:2016dvk}.
Tab.~\ref{tBprime} shows the effective one loop coefficients at partially interacting fixed points as a function of field multiplicities.
\begin{table}[t]
\begin{center}
\begin{tabular}{cl}
\rowcolor{LightGreen}
\toprule
\rowcolor{LightGreen}
\cellcolor{LightBlue}
$\bm{\#}\ $
& ${}$
\quad\quad\quad$\bm{B'}$ \bf coefficient
\\[.4mm]
\midrule
\rowcolor{LightGray}&\\[-4mm]
\rowcolor{LightGray}
\FP2 & $\ \ \displaystyle B_2' = -\frac{4}{3}\left(1 - \frac{2}{25}R/P\right)\0{P\epsilon}R\ \ $
\\[.4mm]
\rowcolor{white}
&\\[-3mm]
\rowcolor{white}
\FP3 & $\ \ B_1' = -\frac{4}{3}\left(1 - \frac{2}{25}P/R\right)R\epsilon$
\\[.4mm]
\rowcolor{LightGray}
&\\[-3mm]
\rowcolor{LightGray}
\FP4 & $\ \ \displaystyle B_2' = -\frac{4}{3}\left(1 - X(R)/{P}\right)\0{P\epsilon}R$
\\[.4mm]
\rowcolor{white}
&\\[-3mm]
\rowcolor{white}
\FP5 & $\ \ B_1' = -\frac{4}{3}\left(1 - P/{\widetilde{X}(R)}\right)R \epsilon$
\\[.4mm]
\bottomrule
\end{tabular}
\end{center}
\caption{Shown are the effective one-loop coefficients $B'$ for the non-interacting gauge coupling at \fp2, \fp3, \fp4 and \fp5, and their dependence on model parameters. $B'>0$ corresponds to asymptotic freedom. Notice that $B'$ changes sign across the boundaries $P=2R/25, 25R/2,X(R)$, and $\widetilde{X}(R)$, respectively, with $X$ and $\widetilde{X}$ given in \eq{P1234}.}
\label{tBprime}
\end{table}
\begin{table}[b]
\begin{center}
\begin{tabular}{cccccc}
\rowcolor{LightGreen}
\toprule
\rowcolor{LightGreen}
\cellcolor{LightBlue}
&\multicolumn{3}{c}{\cellcolor{LightGreen}\bf \ \ \ parameter range\ \ \ }
&
&
\\
\rowcolor{LightGreen}
\cellcolor{LightBlue}
\multirow{-2}{*}{$\bm{\#}\ $}
& \ sign$\ \epsilon\ $
& $R$
& $P$
&\multirow{-2}{*}{\cellcolor{LightGreen}\bf eigenvalue spectrum}
&\multirow{-2}{*}{\cellcolor{LightGreen}\bf info}
\\[.4mm]
\rowcolor{white}
\midrule
\rowcolor{white}
\FP1
& $\pm$
& $\ \left(\frac{2}{11},\frac{11}{2}\right)\ $
&$\ \ (-\infty,+\infty)\ \ $
& \eq{4R}, \eq{2R}, or \eq{0R}
&\cellcolor{LightRed} Gaussian
\\[.4mm]
\midrule
\rowcolor{LightGray} &&&&& \cellcolor{LightRed} Fig.~\ref{pFP23} (upper panel) \\
\rowcolor{LightGray}
& $-$
&$\left(\02{11},\frac{11}{2}\right)$
&$(\frac{2}{25}R,+\infty)$
& $\vartheta_{1,2,3}\le0<\vartheta_{4}$
& region 1
\\[.4mm]
\rowcolor{LightGray}
& $-$
&$\left(\frac{2}{11},\frac{11}{2}\right)$
&$(0,\frac{2}{25}R)$
& $\quad\quad\ \vartheta_{1}<0\le\vartheta_{2,3,4}$
& region 2
\\[.4mm]
\rowcolor{LightGray}
\multirow{-4}{*}{\cellcolor{LightGray}\FP2}
& $-$
&$\left(\frac{2}{11},\frac{11}{2}\right)$
&$(-\infty,0)$
& $\quad\quad\ \vartheta_{1}<0\le\vartheta_{2,3,4}$
& region 3
\\[.4mm]
\midrule
\rowcolor{white} &&&&&\cellcolor{LightRed} Fig.~\ref{pFP23} (lower panel)\\
&$-$
&$\left(\frac{2}{11},\frac{11}{2}\right)$
& $\left(0,\frac{25}{2}R\right)$
& $\vartheta_{1,2,3}\le0<\vartheta_{4}$
& region 1
\\[.4mm]
\rowcolor{white}
&$-$
&$\left(\frac{2}{11},\frac{11}{2}\right)$
& $\left(\frac{25}{2}R,+\infty\right)$
& $\quad\quad\ \vartheta_{1}<0\le\vartheta_{2,3,4}$
& region 2
\\[.4mm]
\multirow{-4}{*}{\FP3}
&$+$
&$\left(\frac{2}{11},\frac{11}{2}\right)$
& $\left(-\infty,0\right)$
&$\quad\quad\ \vartheta_{1}<0\le\vartheta_{2,3,4}$
& region 3
\\[.4mm]
\midrule
\rowcolor{LightGray} &&&&& \cellcolor{LightRed} Fig.~\ref{pFP45} (upper panel)\\
\rowcolor{LightGray}
& $-$
& $\left(\frac{1}{2},\frac{11}{2}\right)$
&$(-\infty,X(R))$
& $\ \quad\quad\quad \quad\quad 0\le\vartheta_{1,2,3,4}$
& region 1 \& 3
\\[.4mm]
\rowcolor{LightGray}
& $-$
& $\left(\frac{1}{2},\frac{11}{2}\right)$
&$(X(R),+\infty)$
& $\quad\, \vartheta_{1,2}\le0<\vartheta_{3,4}$
& region 2
\\[.4mm]
\rowcolor{LightGray}
&$+$
& $\left(\frac{2}{11},\frac{1}{2}\right)$
&$(X(R),+\infty)$
& $\quad\quad\ \vartheta_{1}<0\le\vartheta_{2,3,4}$
& region 4 \& 6
\\[.4mm]
\rowcolor{LightGray}
\multirow{-5}{*}{\cellcolor{LightGray}\FP4}
&$+$
& $\left(\frac{2}{11},\frac{1}{2}\right)$
&$(-\infty,X(R))$
& $\vartheta_{1,2,3}\le0<\vartheta_{4}$
& region 5
\\[.4mm]
\midrule &&&&& \cellcolor{LightRed} Fig.~\ref{pFP45} (lower panel) \\
&$-$
&$\big(\frac{2}{11},2\big)$
& $\big(\widetilde{X}(R),+\infty\big)$
& $\quad\quad\ \quad\quad\quad 0\le\vartheta_{1,2,3,4}$
®ion 1
\\[.4mm]
&$-$
& $\big(\frac{2}{11},2\big)$
&$\big(0,\widetilde{X}(R)\big)$
& $\quad\ \vartheta_{1,2}\le0<\vartheta_{3,4}$
®ion 2
\\[.4mm]
&$+$
&$\left(\frac{2}{11},2\right)$
&$\big(-\infty,0\big)$
&$\quad\quad\quad\quad\quad\ \; 0\le\vartheta_{1,2,3,4}$
®ion 3
\\[.4mm]
&$-$
&$\big(2,\frac{11}{2}\big)$
& $\big(-\infty,\widetilde{X}(R)\big)$
&$\quad\quad\; \; \vartheta_{1}<0\le\vartheta_{2,3,4}$
& region 4
\\[.4mm]
&$-$
&$\big(2,\frac{11}{2}\big)$
& $\big(\widetilde{X}(R),0\big)$
&$\vartheta_{1,2,3}\le0<\vartheta_{4}$
®ion 5
\\[.4mm]
\multirow{-7}{*}{\FP5}
&$+$
&$\big(2,\frac{11}{2}\big)$
& $\big(0,+\infty\big)$
&$\quad\quad\ \vartheta_{1}<0\le\vartheta_{2,3,4}$
®ion 6
\\[.4mm]\bottomrule
\end{tabular}
\end{center}
\caption{Parameter regions where the partially interacting fixed points \fp1 -- \fp5 exist, along with regions of relevancy for eigenvalues and effective one-loop terms, where applicable. The boundary functions $X(R)$ and $\widetilde{X}(R)$ are given in \eq{P1234}. The coefficient $B'$ for the gauge coupling at the Gaussian fixed point is given in Tab.~\ref{tBprime}.
}\label{tPara15}
\end{table}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.25]{pFP2n.pdf}
\\
\vskip.3cm
\hskip-.2cm
\includegraphics[scale=.25]{pFP3n.pdf}
\caption{The phase space of parameters \eq{PRepsN} for the partially interacting fixed points FP${}_2$ (upper panel) and FP${}_3$ (lower panel) where one of the two gauge sectors remains interacting and all Yukawa couplings vanish. The inset indicates the different parameter regions and conditions for existence, including whether the non-interacting gauge sector is asymptotically free $(B'>0)$ or infrared free $(B'<0)$, see Tab.~\ref{tBprime},\ref{tPara15}.
}\label{pFP23}
\end{center}
\end{figure}
\subsection{Gauss with Banks-Zaks}
Next, we discuss all fixed points one-by-one, and determine the valid parameter regimes $(R,P,\epsilon)$ for each of them. We recall that $N_\psi=1$ in our models. Whenever appropriate, we also compare results with the ``direct product'' limit, whereby the diagonal contributions from the Yukawa-less $\psi$-fermions are retained but their off-diagonal contributions to the other gauge sectors suppressed (see Sect.~\ref{pifp}). This comparison allows us to quantify the effect related to the semi-simple nature of the models \eq{L}.
For convenience and better visibility, we scale the axes in Figs.~\ref{pFP23}~\ref{pFP45},~\ref{pFP6},~\ref{pFP78} and~\ref{pFP9} as
\begin{equation}
\label{rescale}
X\to \frac{X}{1+|X|}\quad {\rm where}\quad X=P\ {\rm or}\ R\,,
\end{equation}
and within their respective domains of validity $R\in(\s02{11},\s0{11}{2})$ and $P\in (-\infty,\infty)$. The rescaling permits easy display of the entire range of parameters.
Fig.~\ref{pFP23} shows the results for \fp2 ({BZ$\cdot$G}, upper) and \fp3 ({G$\cdot$BZ}, lower panel), and parameter ranges are given in Tab.~\ref{tPara15}.
We observe that the Banks-Zaks fixed point always requires an asymptotically free gauge sector. Hence, \fp2 exists for any $R$ as long as $\epsilon<0$. Provided that $P\epsilon <0$, the other gauge sector either remain asymptotically free (region 1) or becomes infrared free (region 2). On the other hand, if $P\epsilon>0$, the other gauge sector is invariably infrared free. This is a consequence of \eq{Bprime} which states that the interacting gauge sector can turn asymptotic freedom of the non-interacting gauge sector into infrared freedom (region 2), but not the other way around. The existence of the parameter region 2 is thus entirely due to the semi-simple character of the theory which cannot arise from a ``direct product''.
The Banks-Zaks fixed point is invariably attractive in the gauge coupling, and repulsive in the Yukawa coupling. The eigenvalue spectrum in the gauge-Yukawa sector is therefore of the form \eq{3R} or \eq{1R}, depending on whether the free gauge sector is asymptotically free or infrared free, see Tab.~\ref{tPara15}.
Under the exchange of gauge groups we have $(R,P,\epsilon)\leftrightarrow \left(R^{-1},P^{-1},P\epsilon\right)$, see \eq{exchange}. On the level of Fig.~\ref{pFP23} this corresponds to a simple rotation by $180$ degree around the symmetric points $(R,P)=(1,1)$ (for $P>0$) and $(R,P)=(1,-1)$ (for $P<0$), owing to the rescaling of parameters. Consequently, the results for \fp3 can be deduced from those at \fp2 by a simple rotation, see Fig.~\ref{pFP23}.
More generally, this exchange symmetry relates the partially interacting fixed points
\fp2 $\leftrightarrow$ \fp3 (Fig.~\ref{pFP23}), \fp4 $\leftrightarrow$ \fp5 (Fig.~\ref{pFP45}), and the fully interacting fixed points \fp7 $\leftrightarrow$ \fp8 (Fig.~\ref{pFP78}).
The exchange symmetry is manifest at the fully interacting fixed points \fp6 (Fig.~\ref{pFP6}) and \fp9 (Fig.~\ref{pFP9}).
\begin{figure}[t]
\begin{center}
\vskip-.5cm
\hskip-.7cm
\includegraphics[scale=.275]{pFP4n.pdf}
\\
\vskip-.3cm
\includegraphics[scale=.275]{pFP5n.pdf}
\caption{The phase space of parameters for the partially interacting fixed points FP${}_4$ and FP${}_5$, where the gauge and Yukawa coupling in one gauge sector take interacting fixed points while those of the other sector remain trivial. The insets indicate the different parameter regions and conditions for existence,
and whether the non-interacting gauge sector is asymptotically free $(B'>0)$ or infrared free $(B'<0)$, see Tab.~\ref{tBprime},\ref{tPara15}.
}\label{pFP45}
\end{center}
\end{figure}
\subsection{Gauss with Gauge-Yukawa}
In Fig.~\ref{pFP45} we show the domains of existence for \fp4 ({GY\,$\cdot$\,G}, upper) and \fp5 ({G\,$\cdot$\,GY}, lower panel). We observe that the fixed point exists for any parameter choice though its features vary with matter multiplicities. Specifically, for \fp4, six qualitatively different parameter regions are found. If the interacting gauge coupling is asymptotically free $(\epsilon<0)$ and provided that $P\epsilon <0$, the other gauge sector either remains asymptotically free (region 2) or becomes infrared free (region 1), whereas for $P\epsilon >0$ the other gauge sector invariably remains infrared free (region 3).
Conversely, if the interacting gauge coupling is infrared free $(\epsilon>0)$ and provided that $P\epsilon <0$, the other gauge sector either remains asymptotically free (region 5) or becomes infrared free (region 4), whereas for $P\epsilon >0$ the other gauge sector invariably remains infrared free (region 6). Moreover, as explained in Tab.~\ref{tBprime}, the interacting gauge sector can turn asymptotic freedom of the non-interacting gauge sector into infrared freedom (region 1 and 4).
The eigenvalue spectrum in the gauge-Yukawa sector has therefore no relevant eigendirection \eq{0R} in region 1 and 3, one relevant eigendirection \eq{1R} in region 4 and 6, two relevant eigendirections \eq{2R} in region 2, and three relevant eigendirections \eq{3R} in region 5, see Tab.~\ref{tPara15}.
We make the following observations. Firstly, we note that \fp4 in region 1 and 3 corresponds to a fully attractive IR fixed point with all RG trajectories terminating in it. The fixed point then acts as an infrared ``sink'' for massless trajectories and all canononicaly marginal couplings of the theory. Once scalar masses are switched on, RG flows may run away from the hypercritical surface of exactly massless theories, leading to massive phases with or without spontaneous breaking of symmetry. The quantum phase transition at \fp4 in region 1 and 3 is of the second order. Notice that in the ``direct product'' limit only models with $P\epsilon>0>\epsilon$ and $R>\s012$ (analogous to region 3) would lead to a fully infrared attractive ``sink''. Hence, the availability of region 1 is an entirely new effect, solely due to the $\psi$ fermions and the semi-simple nature of our models. We conclude that the semi-simple structure opens up new types of fixed points which cannot be achieved through a product structure. In region 2, we find that \fp4 has two relevant eigendirections as it would in ``direct product'' settings.
Secondly, in regions 4 and 6, \fp4 shows a single relevant eigendirection. In the ``direct product'' limit, only models with $P,\epsilon>0$ and $R<\s012$ (analogous to region 6) would lead to a single relevant direction. Again, the availability of region 4 is a novel feature, and solely due to the $\psi$ fermions and thus a consequence of the semi-simple nature of the model.
In the parameter region 5 the fixed point shows the largest number of UV relevant directions as it would without the $\psi$ fermions. Moreover, in this parameter regime the Gaussian fixed point has only two relevant directions ($\epsilon>0, P\epsilon <0$). Therefore \fp4 in region 5 qualifies as an asymptotically safe UV fixed point. On the other hand, in region 2,4 and 6, it takes the role of a cross-over fixed point.
Results for \fp5 (Fig.~\ref{pFP45}, lower panel) follow from those for \fp4 via \eq{exchange}, and the distinct regions stated for \fp5 relate to the same physics as those for \fp4.
\begin{table}[b]
\begin{center}
\begin{tabular}{cccccc}
\rowcolor{LightGreen}
\toprule
\rowcolor{LightGreen}
\cellcolor{LightBlue}
&\multicolumn{3}{c}{\cellcolor{LightGreen}\bf \ \ \ parameter range\ \ \ }
&&
\\
\rowcolor{LightGreen}
\cellcolor{LightBlue}
\multirow{-2}{*}{$\bm{\#}\ $} & \ sign$\ \epsilon\ $ & $R$ & $P$
&\multirow{-2}{*}{\cellcolor{LightGreen}\bf eigenvalue spectrum}
&\multirow{-2}{*}{\cellcolor{LightGreen}\bf info}
\\[.4mm]
\rowcolor{white}
\midrule
\rowcolor{white}
\FP6
&$-$
& $\left(\frac{2}{11},\frac{11}{2}\right)$
& $\left(\frac{2}{25}R, \frac{25}{2}R\right)$
& $\vartheta_{1,2} < 0 < \vartheta_{3,4}$
&\cellcolor{LightRed}
Fig.~\ref{pFP6}\\[.4mm]
\midrule
\rowcolor{LightGray} &&&&& \cellcolor{LightRed} Fig.~\ref{pFP78} (upper panel)
\\
\rowcolor{LightGray}
&$-$
&$\left(\frac{2}{11},\frac{1}{2}\right)$
&$ \left(\frac{25}{2}R ,+\infty\right)$
&$\vartheta_{1,2} < 0 < \vartheta_{3,4}$
®ion 1
\\[.4mm]
\rowcolor{LightGray}
&$-$
& $\left(\frac{1}{2},R_1\right)$
& $\left(\frac{25}{2}R , X(R)\right)$
& $\vartheta_{1,2} < 0 < \vartheta_{3,4}$
®ion 1
\\[.4mm]
\rowcolor{LightGray}
&$-$
& $\left(R_1,\frac{11}{2}\right)$
& $\left(X(R) , \frac{25}{2}R\right)$
& $\quad\ \vartheta_1 < 0 < \vartheta_{2,3,4}$
®ion 2
\\[.4mm]
\rowcolor{LightGray}
\multirow{-5}{*}{\cellcolor{LightGray}\FP7}
&$+$
& $\left(\frac{2}{11},\frac{1}{2}\right)$
&$ \left(-\infty, X(R)\right)$
&$\vartheta_{1,2} < 0 < \vartheta_{3,4}$
& region 3
\\[.4mm]
\midrule
\rowcolor{white} &&&&& \cellcolor{LightRed} Fig.~\ref{pFP78} (lower panel)
\\
\rowcolor{white}
&$-$
& $\left(R_4,\frac{11}{2}\right)$
& $\big(\widetilde{X}(R),\frac{2}{25}R\big)$
& $\vartheta_{1,2} < 0 < \vartheta_{3,4}$
®ion 1 \& 3
\\[.4mm]
\rowcolor{white}
\multirow{-3}{*}{\FP8}
&$-$
& $\left(\frac{2}{11},R_4\right)$
& $\big(\frac{2}{25}R , \widetilde{X}(R)\big)$
& $\quad\ \vartheta_1 < 0 < \vartheta_{2,3,4}$
& region 2
\\[.4mm]
\midrule
\rowcolor{LightGray} &&&&& \cellcolor{LightRed} Fig.~\ref{pFP9}
\\
\rowcolor{LightGray}
&$-$
& $(\frac{2}{11},\frac{1}{2})$
&$ (\widetilde{X}(R) ,+\infty) $
& $\quad\ \vartheta_1 < 0 < \vartheta_{2,3,4}$
& region 1
\\[.4mm]
\rowcolor{LightGray}
&$-$
& $\left(\frac{1}{2},R_2\right)$
&$ \big(\widetilde{X}(R) , X(R)\big)$
& $\quad\ \vartheta_1 < 0 < \vartheta_{2,3,4}$
®ion 1
\\[.4mm]
\rowcolor{LightGray}
&$-$
& $\left(R_3,\s0{11}2\right)$
&$ \big(\widetilde{X}(R) , X(R)\big)$
& $\quad\ \vartheta_1 < 0 < \vartheta_{2,3,4}$
& region 1 \& 4
\\[.4mm]
\rowcolor{LightGray}
&$-$
& $\left(R_2,R_3\right)$
&$\ \ \big(X(R) , \widetilde{X}(R)\big)\ \ $
& $\quad\quad\quad\quad\ 0 < \vartheta_{1,2,3,4}$
®ion 2
\\[.4mm]
\rowcolor{LightGray}
\multirow{-6}{*}{\cellcolor{LightGray}\FP9}
&$+$
& $\left(\frac{2}{11},\frac{1}{2}\right)$
&$ \left(-\infty, X(R)\right)$
& $\quad\ \vartheta_1 < 0 < \vartheta_{2,3,4}$
& region 3
\\[.4mm]
\bottomrule
\end{tabular}
\end{center}
\caption{Parameter regions where the fully interacting fixed points \fp6 -- \fp9 exist, along with the eigenvalue spectrum for the various parameter regions.
}\label{tPara69}
\end{table}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.95]{pFP6.pdf}
\caption{The phase space of parameters for the interacting fixed point \fp6 (red) where both gauge sectors take interacting and physical fixed points while all Yukawa couplings vanish. The eigenvalue spectrum at the fixed point always displays exactly two relevant eigenvalues of $O(\epsilon)$ and two irrelevant eigenvalues of order $O(\epsilon^2)$,
see Tab.~\ref{tPara69}. Note that this fixed point invariably requires asymptotic freedom for both gauge sectors
(see main text).
}\label{pFP6}
\end{center}
\end{figure}
\subsection{Banks-Zaks with Banks-Zaks}
Next, we turn to fully interacting fixed points where both gauge couplings are non-vanishing, see Tab.~\ref{tPara69}. In general, the eigenvalue spectrum is determined through \eq{dtheta} with solutions \eq{theta}, with $\epsilon$ taking the role of the parameter $B$.
In the ``direct product'' limit, fully interacting fixed points reduce to direct products from each pairing of the possibilities (BZ, GY) in each of the simple gauge sectors. For $N_\psi\neq 0$, the fermions $\psi$ introduce a direct mixing between the gauge groups and we may then expect to find something close to a product structure, potentially modified by new effects parametrised via $R$ in fixed points not involving Gaussian factors.
The first such fixed point is \fp6 ({BZ$\cdot$BZ}), where each gauge sector achieves a Banks-Zaks fixed point. Yukawa couplings play no role, see Fig.~\ref{pFP6}. The fixed point invariably requires $\epsilon<0$ and $P\epsilon<0$ and entails an eigenvalue spectrum with two relevant directions of order ${\cal O}(\epsilon^2)$, and two irrelevant directions of order ${\cal O}(\epsilon)$ associated to the Yukawas,
\begin{equation}\label{12-34}
\vartheta_{1},\vartheta_{2}<0<\vartheta_{3},\vartheta_{4}\,.
\end{equation}
The quartics are marginally irrelevant.
The Gaussian is necessarily the UV fixed point in these settings which makes \fp6 a cross-over fixed point. The accessible parameter region, shown in Fig.~\ref{pFP6}, is invariant under the exchange of gauge groups \eq{exchange}. The ``direct product'' limit has qualitatively the same spectrum \eq{12-34}. The main effect due to the semi-simple character of the theory relates to the exclusion of certain parameter regions (white regions). We conclude that the semi-simple nature of the theory leads to parameter restrictions without otherwise changing the overall appearance of the fixed point.
\begin{figure}[t]
\begin{center}
\hskip-.5cm
\includegraphics[scale=0.285]{pFP7n.pdf}
\\
\vskip-.3cm
\hskip.5cm
\includegraphics[scale=0.285]{pFP8n.pdf}
\caption{The phase space of parameters for the fixed points FP${}_7$ and FP${}_8$ where two gauge and one of the Yukawa couplings take interacting and physical fixed points, while the other Yukawa coupling remains trivial. The inset indicates the signs for $\epsilon$ and $P\epsilon$, together with the sign for the eigenvalue $\vartheta_2$, Tab.~\ref{tPara69}
(see main text). }
\label{pFP78}
\end{center}
\end{figure}
\subsection{Banks-Zaks with Gauge-Yukawa}
At the interacting fixed points \fp7 ({BZ\,$\cdot$\,GY}, upper panel), and \fp8 ({GY\,$\cdot$\,BZ}, lower panel), we have that both gauge and one of the Yukawa couplings are non-trivial. Our results for the condition of existence and the eigenvalue spectra are displayed in Fig.~\ref{pFP78}. By definition, this type of fixed point requires that either $\epsilon<0$ or $P\epsilon<0$, or both, meaning that at least one of the gauge sectors is asymptotically free. In Fig.~\ref{pFP78}, this relates to three different parameter regions (see inset for the colour coding). In region 1 and 2, the theory is asymptotically free in both gauge sectors, whereas in region 3 the theory is asymptotically free in only one gauge sector. We observe that large regions of parameter space are excluded. Valid parameter regions are further distinguished by their eigenvalue spectrum which either takes the form \eq{2R} or \eq{1R}, meaning that minimally one and maximally two eigenoperators constructed out of the gauge kinetic terms and the Yukawa interactions are UV relevant, $\vartheta_1 <0<\vartheta_3,\vartheta_4$.
The sign of $\vartheta_2$ depends on the matter field multiplicities.
In region 1 and 3, and for either of \fp7 and \fp8, we find that
\begin{equation}\label{1-fp78}
\vartheta_1 , \vartheta_2<0<\vartheta_3,\vartheta_4\,.
\end{equation}
In region 2, conversely, we have
\begin{equation}\label{2-fp78}
\vartheta_1 <0< \vartheta_2,\vartheta_3,\vartheta_4\,.
\end{equation}
Hence, at \fp7 and in the regime where both gauge sectors are asymptotically free $(P > 0 > \epsilon)$, two types of valid fixed points are found. For sufficiently low $R<R_1$ \eq{R1234}, and large $P$, the fixed point has two relevant directions (region 1). Increasing $R>R_1$ at fixed $P$ may lead to a second type of IR fixed point with a single relevant direction (region 2). On the other hand, in the regime $\epsilon>0>P$ only one type of fixed point exists with two relevant directions (region 3).
It is worth comparing these results with the ``direct product'' limit. For $P > 0 > \epsilon$ the latter leads to the eigenvalue spectrum \eq{2-fp78}, as found in region 2. Also, for $\epsilon>0>P$ the ``direct product'' fixed point has the eigenvalue spectrum \eq{1-fp78}, which is qualitatively in accord with findings in region 3.
We conclude that the semi-simple nature of the interactions plays a minor quantitative role in region 2 and 3.
On the other hand, in region 1 where $P > 0 > \epsilon$, the semi-simple nature of the theory leads to an important qualitative modification: an eigenvalue spectrum with two relevant directions at \fp7 cannot be achieved through a direct product setting; rather, it necessarily requires matter fields charged under both gauge groups. We conclude that the semi-simple nature of interactions play a key qualitative role in region 1.
Analogous results hold true for \fp8 after the substitutions \eq{exchange} and the replacement $R_1\to R_4=1/R_1$, see \eq{R1234}.
\subsection{Gauge-Yukawa with Gauge-Yukawa}
At the fully interacting fixed point \fp9 ({GY\,$\cdot$\,GY}), we have that both gauge and both Yukawa couplings are non-trivial. We find that the eigenvalue spectrum in the gauge-Yukawa sector reads either \eq{1R} or \eq{0R}, meaning that at least three of the four eigenoperators constructed out of the gauge and fermion fields
are strictly irrelevant, $0<\vartheta_2,\vartheta_3,\vartheta_4$.
The sign of the eigenvalue $\vartheta_1$ depends on the matter field multiplicities of the model.
Our results for the condition of existence and the eigenvalue spectrum are stated in Fig.~\ref{pFP9}.
We observe four qualitatively different parameter regions (see inset for the colour coding).
For $P > 0 >\epsilon$, the theory is asymptotically free in both gauge sectors and we find two types of valid parameter regions, depending on whether $R$ takes values below $R_2$ or above $R_3$ (region 1), or in between (region 2); see \eq{R1234}. Moreover, in region 2, we find that the fixed point is strictly IR attractive in all couplings, owing to
\begin{equation}\label{2-fp9}
0<\vartheta_1\,,\vartheta_2,\vartheta_3,\vartheta_4\,.\end{equation}
Hence, the fixed point \fp9 in region 2 corresponds to a fully attractive IR fixed point acting as an infrared ``sink'' for massless trajectories and all canononicaly marginal couplings of the theory. Ultimately it describes a second order quantum phase transition between a symmetric and a symmetry broken phase, characterised by the vacuum expectation value of the scalar field. Qualitatively, the same type of result is achieved in the ``direct product'' limit.
Hence, the main effect due to semi-simple interactions
is to have generated a boundary in parameter space.
In region 1 we find
\begin{equation}\label{1-fp9}
\vartheta_1<0<\vartheta_2,\vartheta_3,\vartheta_4\,.
\end{equation}
This type of eigenvalue spectrum cannot be achieved without semi-simple interactions mediated by the $\psi$ fields and is therefore a novel feature, entirely due to the semi-simple nature of the theory. In this regime, \fp9 corresponds to a cross-over fixed point (the Gaussian is the UV fixed point) with a single unstable direction where trajectories escape either towards a weakly coupled IR fixed point, or towards a regime of strong coupling with (chiral) symmetry breaking, confinement, or infrared conformality.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.3]{pFP9n.pdf}
\vskip-.5cm
\caption{The phase space of parameters for the fully interacting fixed point FP${}_9$ where all gauge and all Yukawa couplings are non-trivial.
The coloured regions relate to the portions of parameter space where the fully interacting fixed point is physical. The inset provides additional information including the sign for the eigenvalue $\vartheta_4$ (see main text).
}\label{pFP9}
\end{center}
\end{figure}
For $\epsilon>0>P$ or $P\epsilon>0>\epsilon$, the theory is asymptotically free in one and infrared free in the other gauge sector. Valid fixed points then correspond to region 3 or region~4, respectively. In either of these cases, the eigenvalue spectrum shows a single relevant direction, \eq{1-fp9}. This agrees qualitatively with the eigenvalue spectrum in the ``direct product'' limit. We conclude, once more, that the main impact of the $\psi$ fields relates to the boundaries in parameter space which restrict the fixed point's domain of availability.
Finally, for $\epsilon,P>0$, the theory is infrared free in both gauge sectors. We observe that no such interacting fixed point arises, irrespective of matter multiplicities. Interestingly though, such fixed points do exist in the ``direct product'' limit with spectrum
\begin{equation}\label{thetafp9put}
\vartheta_1,\vartheta_2<0<\vartheta_3,\vartheta_4\,.
\end{equation}
The reason for their non-existence in our models is the presence of the $\psi$ fermions. The requirement of perturbativity in both gauge couplings then leads to limitations on the parameter $R$ which cannot be satisfied at \fp9 with eigenvalue spectrum \eq{thetafp9put}. This result provides us with an example where the semi-simple nature of the theory
``disables'' a fixed point. This completes the overview of interacting fixed points in the gauge-Yukawa sector and their key properties.
\section{\bf Scalar fixed points and vacuum stability}\label{stab}
In this section, we analyse the scalar sector and establish conditions for stability of the quantum vacuum. We also provide results for all scalar couplings at all interacting fixed points, Tab.~\ref{tsFPeps}.
\subsection{Yukawa and scalar nullclines}\label{YSnull}
Following \cite{Bond:2016dvk}, we begin by exploiting the results \eq{EF} to express the Yukawa nullclines in terms of the gauge couplings and the parameter $R$. To leading order in the small expansion parameters \eq{eps}, and using \eq{EF} together with \eq{12Yy}, the non-trivial Yukawa nullclines $\beta_Y=0$ and $\beta_y=0$ take the explicit form
\begin{equation}\label{Ynull}
\frac{\alpha_Y}{\alpha_1}=\frac{6}{13-2R}\,,\quad
\frac{\alpha_y}{\alpha_2}=\frac{6}{13-2/R}\,.
\end{equation}
For fixed gauge couplings, we observe that the Yukawa couplings are enhanced over their values in the absence of the fermions $\psi$. The relevance of nullcline solutions \eq{Ynull} is as follows. By their very definition, the Yukawa couplings no longer run with the RG scale when taking the values \eq{Ynull}. If at the same time the gauge couplings take fixed points on their own, the nullcline relations then provide us with the correct fixed point values for the Yukawa couplings. Evidently, \eq{Ynull} together with \eq{R} guarantees that the Yukawa fixed points are physical as long as the gauge fixed points are. Note also that the slope of the nullcline remains positive and finite for all $R$ within the domain \eq{R}. Hence strict perturbativity in the Yukawa couplings follows from strict perturbativity in the gauge couplings, in accord with the general discussion in \cite{Bond:2016dvk} based on dimensional analysis.
Next we turn to the scalar nullclines. Since the beta functions for the two scalar sectors decouple at this order, we may analyse their nullclines individually.\footnote{This simplification solely arises provided the mixing coupling $\alpha_w$ takes its exact Gaussian fixed point \eq{wFP}. For non-trivial $\alpha_w$ the nullclines take more general forms.} All results for the subsystem $(\alpha_U,\alpha_V)$ can straightforwardly be translated to the subsystem $(\alpha_u,\alpha_v)$ by substituting $R\leftrightarrow 1/R$, also using \eq{eps}. Furthermore, since the scalars are uncharged, their one loop beta functions are independent of the gauge coupling. Dimensional analysis then shows that all non-trivial scalar nullclines are proportional to the corresponding Yukawa coupling \cite{Bond:2016dvk}. The scalar nullclines represent exact fixed points of the theory provided the Yukawa couplings take interacting fixed points. Perturbativity of scalar couplings at an interacting fixed point then follows from the perturbativity of Yukawa couplings which, in turn, follows from perturbativity in the gauge couplings.
Specifically, the nullclines for the single trace scalar couplings are found from \eq{HVhv} by resolving $\beta_U=0$ for $\alpha_U$. We find two solutions
\begin{equation}\label{Hnull}
\frac{\alpha_{U\pm}}{\alpha_Y}=\frac14\left(-1\pm\sqrt{23-4R}\right)\,.
\end{equation}
Note that the double trace coupling does not couple back into the running of the single trace coupling. Within the parameter range \eq{R} we observe that $\alpha_{U+}>0>\alpha_{U-}$.
Next, we consider the nullclines for the double-trace quartic coupling $\alpha_V$. Inserting $\alpha_{U+}$ into $\beta_V=0$, we find a pair of nullclines given by
\begin{equation}\label{Vnull}
\frac{\alpha_{V\pm}}{\alpha_Y}=\frac14\left(-2\sqrt{23-4R}\pm
\sqrt{20-4R+6\sqrt{23-4R}}\right)\,.
\end{equation}
Both nullclines take real values for all $R$ within the range \eq{R}, and we end up with $\alpha_{U+}\ge 0$ together with $0>\alpha_{V+}>\alpha_{V-}$. Analogously, inserting $\alpha_{U-}$ into $\beta_V=0$, we find a second pair of nullclines given by
\begin{equation}\label{Vnull2}
\frac{\alpha_{V2\pm}}{\alpha_Y}=\frac14\left(2\sqrt{23-4R}\pm
\sqrt{20-4R-6\sqrt{23-4R}}\right)\,.
\end{equation}
In this case, however, the result \eq{Vnull2} comes out complex within the parameter range \eq{R}, meaning that even if $\alpha_Y^*$ takes a real positive fixed point the corresponding scalar fixed point is invariably unphysical.
The replacement $R\to 1/R$ in \eq{Hnull} and \eq{Vnull}, \eq{Vnull2} allows us to obtain explicit expressions for the nullclines for ${\alpha_{u\pm}}/{\alpha_y}$ and ${\alpha_{v\pm}}/{\alpha_y}$. The real solutions are given by
\begin{equation}\label{hnull}
\frac{\alpha_{u\pm}}{\alpha_y}=\frac14\left(-1\pm\sqrt{23-4/R}\right)
\end{equation}
with $\alpha_{u+}\ge 0>\alpha_{u-}$. The solution $\alpha_{u+}$ leads to real nullclines for the double-trace coupling $\alpha_{v}$ given by
\begin{equation}\label{vnull}
\frac{\alpha_{v\pm}}{\alpha_y}=\frac14\left(-2\sqrt{23-4/R}\pm
\sqrt{20-4/R+6\sqrt{23-4/R}}\right)\,,
\end{equation}
and we end up with $\alpha_{u+}\ge 0$ together with $0>\alpha_{v+}>\alpha_{v-}$. On the other hand, the solution $\alpha_{u-}$ does not lead to real solutions for $\alpha_{v\pm}$.
This completes the overview of Yukawa and scalar nullcline solutions.
\subsection{Stability of the vacuum}
We are now in a position to reach firm conclusions concerning the stability of the ground state at interacting fixed points. The reason for this is that this information is encoded in the scalar nullclines. The explicit form of the fixed point in the gauge-Yukawa sector is not needed.
To that end, we recall the stability analysis for potentials of the form
\begin{align}\label{potential}
W \propto \alpha_U\,{\rm Tr}\,(H ^\dagger H )^2 +\alpha_V/N_{\rm F}\,({\rm Tr}\,H ^\dagger H )^2 \,,
\end{align}
In the limit where $\alpha_w=0$ the scalar field potential in our models \eq{L} are given by \eq{potential} together with its counterpart $(H,\alpha_U,\alpha_V)\leftrightarrow (h,\alpha_u,\alpha_v)$. For potentials of the form \eq{potential}, the general conditions for vacuum stability read \cite{Paterson:1980fc,Litim:2015iea}
\begin{equation}\label{stability}
\begin{array}{rcl}
{\it a)}\quad \alpha^*_{U}>0&{\rm \ \ and\ \ }&\alpha^*_{U}+\alpha^*_{V}> 0\\
{\it b)}\quad \alpha^*_{U}<0&{\rm \ \ and\ \ }&\alpha^*_{U}+\alpha^*_{V}/N_{\rm F}>0
\end{array}
\end{equation}
and similarly for $(\alpha_U,\alpha_V)\leftrightarrow(\alpha_u,\alpha_v)$.
In the Veneziano limit, case $b)$ effectively becomes void and cannot be satisfied for any $\alpha_U^*$, irrespective of the sign of $\alpha_V^*$.
Inserting the fixed points into \eq{stability} we find
\begin{equation}\label{vac11}
\begin{array}{rcl}
\alpha^*_{U+}+\alpha^*_{V+}&=&
\displaystyle
\frac{\alpha^*_Y}4
\left(+\sqrt{20-4R+6\sqrt{23-4R}}-\sqrt{23-4R} -1\right)\ge 0\,,\\[2ex]
\displaystyle
\alpha^*_{U+}+\alpha^*_{V-}&=&
\displaystyle
\frac{\alpha^*_Y}4
\left(-\sqrt{20-4R+6\sqrt{23-4R}}-\sqrt{23-4R} -1\right)\le -\alpha^*_Y\,,
\end{array}
\end{equation}
Stability of the quantum vacuum is evidently achieved at the fixed point $(\alpha^*_{U+},\alpha^*_{V+})$ following case $a)$ and irrespective of the value for the Yukawa fixed point as long as $\alpha_Y^*>0$.
The potential \eq{potential} becomes exactly flat at the fixed point iff $R=\s0{11}{2}$. In this case, higher order or radiative corrections must be taken into consideration to guarantee stability in the presence of flat directions.
Stability is not achieved at the fixed point $(\alpha^*_{U+},\alpha^*_{V-})$,
for any $R$.
Turning to the second scalar sector, we find
\begin{equation}\label{vac21}
\begin{array}{rcl}
\alpha^*_{u+}+\alpha^*_{v+}&=&
\displaystyle
\frac{\alpha^*_y}4
\left(+\sqrt{20-4/R+6\sqrt{23-4/R}}-\sqrt{23-4/R} -1\right)\ge 0\,,\\[2ex]
\alpha^*_{u+}+\alpha^*_{v-}&=&
\displaystyle
\frac{\alpha^*_y}4
\left(-\sqrt{20-4/R+6\sqrt{23-4/R}}-\sqrt{23-4/R} -1\right)\le -\alpha^*_y\,,
\end{array}
\end{equation}
where the bounds refer to $R$ varying within the range \eq{R}.
This part of the potential becomes exactly flat at the fixed point iff $R=\s02{11}$.
The result establishes vacuum stability at the fixed point $(\alpha^*_{u+},\alpha^*_{v+})$.
We also confirm that the fixed point $(\alpha^*_{u+},\alpha^*_{v-})$
does not lead to a stable ground state.
We conclude that vacuum stability is guaranteed at the interacting fixed points $(\alpha^*_{U+},\alpha^*_{V+})$ and $(\alpha^*_{u+},\alpha^*_{v+})$, together with $\alpha_w=0$, irrespective of the fixed points in the gauge Yukawa sector, as long as the later is physical. Out of the a priori $2^3$ different fixed point candidates in the scalar sector at one loop (half of which lead to real fixed points) the additional requirement of vacuum stability has identified a {\it unique} viable solution.
In this light, vacuum stability
dictates that the anomalous dimensions \eq{gammam} are strictly positive at interacting fixed points, \eq{gammaPos}, with
\begin{equation}\label{gammam*}
\begin{array}{rcl}
\gamma^*_M&=&\displaystyle\al Y^*\, \sqrt{20-4R+6\sqrt{23-4R}}>0\,,\\[1ex]
\gamma^*_m&=&\displaystyle\al y^*\, \sqrt{20-4/R+6\sqrt{23-4/R}}>0 \,,
\end{array}
\end{equation}
and provided that \eq{R} is observed.
\subsection{Portal coupling}
Now we clarify whether the stability of the vacuum is affected by the presence of the ``portal'' coupling $\alpha_w\neq 0$ which induces a mixing between the scalar sectors.
In this case the scalar potential is given by $W=-L_{\rm pot}$ in \eq{L},
\arraycolsep .1em \begin{eqnarray}
W&=& U\, {\rm Tr}(H^\dagger H)^2 + V\, ({\rm Tr} H^\dagger H )^2
+ u\, {\rm Tr}(h^\dagger h)^2 + v\, ({\rm Tr} h^\dagger h )^2 + w\, {\rm Tr}(H^\dagger H){\rm Tr}(h^\dagger h)\,,\quad
\end{eqnarray}
where $H$ and $h$ are $N_{\rm F} \times N_{\rm F}$ and $N_{\rm f} \times N_{\rm f}$ matrices, respectively. Following the reasoning of \cite{Paterson:1980fc,Litim:2015iea}, we observe that the potential has a global $U(N_{\rm F})_L \otimes U(N_{\rm F})_R\otimes U(N_{\rm f})_L \otimes U(N_{\rm f})_R$
symmetry which allows us to bring each of the fields into a real diagonal configuration, $H = \operatorname{diag}(\Phi_1,\Phi_2,\dots)$ and $h = \operatorname{diag}(\phi_1,\phi_2,\dots)$. As the potential is homogeneous in either field, $W(c\,H, c\,h) = c^4 W(H,h)$,
it suffices to guarantee positivity on a fixed surface $\sum_i \Phi_i^2 =1= \sum_j \phi_j^2$ which is implemented using Lagrange multipliers $\Lambda$ and $\lambda$.
From
\begin{equation}
\begin{array}{rl}
\displaystyle\frac{\partial W}{\partial \Phi_i}
&= 2 \Phi_i (2U \Phi_i^2 + 2V + w - 2\Lambda)\,,\\[2ex]
\displaystyle \frac{\partial W}{\partial \phi_i} &= 2 \phi_i (2u \phi_i^2 + 2v + w - 2\lambda)\,,
\end{array}
\end{equation}
it follows that
extremal field configurations are those where all non-zero fields take equal values. If we have $M$ non-zero $\Phi$ fields and $m$ non-zero $\phi$ fields, the extremal field values are
$\Phi_i^2 = 0$ or $\Phi_i^2 = \frac{1}{M}$ alongside with
$\phi_i^2 = 0$ or $\phi_i^2= \frac{1}{m}$.
Three non-trivial cases arise. If $m = 0$ the extremal potential is
$W_e = {U}/{M} + V$. Likewise if $M = 0$ we have $W_e = {u}/{m} + v$.
Lastly, if both $m, M \neq 0$, we have $W_e = {U}/{M} + V + {u}/{m} + v + w$.
The values of $M,m$ for which these extremal potentials are minima depend on the signs of the couplings $U,u$, leaving us with the four possible cases $U,u>0$, $u>0>U$, $U>0>u$, and $0>U,u$. We thus obtain four distinct sets of conditions for vacuum stability which we summarise as follows:
\begin{equation}\label{conditionsW}
\begin{array}{llccccl}
a)\ \
&\alpha_u, \alpha_U\ge 0\,,
&\alpha_U + \alpha_V \geq 0\,,
&
&\alpha_u + \alpha_v \geq 0\,,
&
&\displaystyle F\,(\alpha_U + \alpha_V) + \frac{\alpha_u
+ \alpha_v}{F}+ \alpha_w \geq 0\,,
\\[2ex]
b)
&\alpha_u>0>\alpha_U\,,
&\displaystyle \alpha_U + \frac{\alpha_V}{N_{\rm F}} \geq 0 \,,
&
&\alpha_u + \alpha_v \geq 0\,,
&
&\displaystyle
\alpha_U + \frac{\alpha_V}{N_{\rm F}} + \frac{\alpha_u + \alpha_v}{F\,N_{\rm f}}
+ \frac{\alpha_w}{N_{\rm f}}\geq 0\,,
\\[2ex]
c)
& \alpha_U>0>\alpha_u\,,
&\alpha_U + \alpha_V \geq 0\,,
&
&\displaystyle \alpha_u + \frac{\alpha_v}{N_{\rm f}} \geq 0\,,
&
&\displaystyle
\alpha_u + \frac{\alpha_v}{N_{\rm f}} + F\,\frac{\alpha_U + \alpha_V}{N_{\rm F}}
+ \frac{\alpha_w}{N_{\rm F}}\geq 0\,,
\\[2ex]
d)
&0\ge\alpha_u,\alpha_U\,,
&\displaystyle \alpha_U + \frac{\alpha_V}{N_{\rm F}} \geq 0\,,
&
&\displaystyle \alpha_u + \frac{\alpha_v}{N_{\rm f}} \geq 0\,,
&
&\displaystyle \alpha_u + \frac{\alpha_v}{N_{\rm f}}
+ F\left(\alpha_U + \frac{\alpha_V}{N_{\rm F}}\right)
+ \frac{\alpha_w}{N_{\rm F}} \geq 0\,.
\end{array}
\end{equation}
Notice that we have rescaled the couplings as in \eq{couplings} and \eq{w} to make contact with the notation used in this paper. The parameter $F\equiv N_{\rm f}/N_{\rm F}>0$ can be expressed in terms of the parameter $R$ to leading order in $\epsilon\ll 1$, see \eq{epsF}.
We make the following observations. In all four cases, the additional condition owing to the mixing coupling \eq{w} takes the form of a lower bound for $\alpha_w$. Furthermore, $\alpha_w$ is allowed to be negative without destroying the stability of the potential, provided it does not become too negative.
We also note that none of the three cases $b)$, $c)$ or $d)$ in \eq{conditionsW} can have consistent solutions in the Veneziano limit where $N_{\rm F}, N_{\rm f} \rightarrow \infty$. This uniquely leaves the case $a)$ as the only possibility for vacuum stability in the parameter regions considered here. These solutions neatly fall back onto the solutions discussed previously in the limit $\alpha_w\to 0$. As long as the auxiliary condition
\begin{equation}
\label{aux}
\alpha_w \geq -[F\left(\alpha_U + \alpha_V\right) + F^{-1}\left(\alpha_u + \alpha_v\right)]
\end{equation}
is satisfied, we can safely conclude that a non-vanishing $\alpha_w\neq 0$ does not spoil vacuum stability, not even for negative portal coupling $\alpha_w$.
\begin{table*}[t]
\begin{center}
\begin{tabular}{cc}
\toprule
\rowcolor{LightBlue}
&\cellcolor{LightGreen}
\\[-1mm]
\cellcolor{LightBlue}
\multirow{-2}{*}{
$\bm{\#}$
}
&\cellcolor{LightGreen}
\multirow{-2}{*}{\bf quartic scalar couplings}
\\
\midrule
\rowcolor{LightGray}
\FP{1-3}
&
$
\alpha^*_U=0\,,\quad
\alpha^*_V=0\,,\quad
\alpha^*_u=0\,,\quad
\alpha^*_v=0\,,
$\\[1ex]
& \\[-2.5ex]
\FP4
&
$
\alpha^*_U=\s0{4F_1(R)R\epsilon}{(2R-1)(3R-19)}\,,\quad
\alpha^*_V=\s0{4F_2(R)R\epsilon}{(2R-1)(3R-19)}\,,\quad
\alpha^*_u=0\,,\quad
\alpha^*_v=0\,,
$
\\[1ex]
\rowcolor{LightGray}
& \\[-2.5ex]
\rowcolor{LightGray}
\multirow{1}{*}{\bf FP${}_{\bf 5}$}
&$
\alpha^*_U=0\,, \quad
\alpha^*_V=0\,, \quad
\alpha^*_u=\frac{4\,F_1(1/R)P\epsilon/R}{(2/R-1)(3/R-19)}\,,\quad
\alpha^*_v=\frac{4\,F_2(1/R)P\epsilon/R}{(2/R-1)(3/R-19)}\,,
$\\[1ex]
& \\[-2.5ex]
\FP6
&$\alpha^*_U=0\,,
\quad\alpha^*_V=0\,,
\quad \alpha^*_u=0\,,
\quad\alpha^*_v=0\,,
$\\[1ex]
\rowcolor{LightGray}
& \\[-2.5ex]
\rowcolor{LightGray}
\FP7
&
$\alpha^*_U=\0{4}3 \frac{(25 - 2P/R)F_1(R)}{50R^2-343R+167} R\epsilon\,,\quad
\alpha^*_V= \0{4}3 \frac{(25 - 2P/R)F_2(R)}{50R^2-343R+167} R\epsilon\,,\quad
\alpha^*_u=0\,,\,\quad
\alpha^*_v=0\,,
$\quad\quad
\\[1ex]
& \\[-2.5ex]
\multirow{1}{*}{\FP8}
&$
\alpha^*_U=0\,, \quad
\alpha^*_V=0\,, \quad
\alpha^*_u=\0{4}3 \frac{(25 - 2R/P)F_1(1/R)}{50/R^2-343/R+167} \frac{P\epsilon}{R}\,,\quad
\alpha^*_v=\0{4}3 \frac{(25 - 2R/P)F_2(1/R)}{50/R^2-343/R+167} \frac{P\epsilon}{R}\,,$
\\[1ex]
\rowcolor{LightGray}
& \\[-2.5ex]
\rowcolor{LightGray}
&
$\alpha^*_U=
\0{4}3\0{[(13-2/R)P/R+(2/R-1)(3/R-19)]F_1(R)}{(19R^2 - 43R + 19)(2/R^2 -13/R + 2)}
R\epsilon,$
$\alpha^*_u=
\0{4}3\0{[(13-2R)R/P+(2R-1)(3R-19)]F_1(1/R)}{(19/R^2 - 43/R + 19)(2R^2 -13R + 2)} \frac{P\epsilon}{R},$
\\[1ex]
\rowcolor{LightGray}
\multirow{-3}{*}{\FP9}
&
$\alpha^*_V=
\0{4}3\0{[(13-2/R)P/R+(2/R-1)(3/R-19)]F_2(R)}{(19R^2 - 43R + 19)(2/R^2 -13/R + 2)}
R\epsilon,$
$\alpha^*_v=
\0{4}3\0{[(13-2R)R/P+(2R-1)(3R-19)]F_2(1/R)}{(19/R^2 - 43/R + 19)(2R^2 -13R + 2)} \frac{P\epsilon}{R}\,. $ \\[1ex]
\bottomrule
\end{tabular}
\caption{Quartic scalar couplings at all weakly interacting fixed points to leading order in $\epsilon$ following Tab.~\ref{tFPs} using the auxiliary functions \eq{F12}. Same conventions as in Tab.~\ref{tFPeps}. Within the admissible parameter ranges (Tab.~\ref{tPara15},~\ref{tPara69}) we observe vacuum stability.
}
\label{tsFPeps}
\end{center}
\end{table*}
\subsection{Unique scalar fixed points}
In Tab.~\ref{tsFPeps}, we summarise our results for the quartic scalar couplings at all weakly interacting fixed points to leading order in $\epsilon$ following Tab.~\ref{tFPs}, using \eq{PRepsN}.
We also introduce the auxiliary functions
\begin{equation}\label{F12}
\begin{array}{l}
F_1(x)=\s014\left(\sqrt{23-4x}-1\right)\,,\\[1.5ex]
F_2(x)=\s014\left(\sqrt{20-4x+6\sqrt{23-4x}}-2\sqrt{23-4x}\right)
\end{array}
\end{equation}
which originate from the scalar nullclines. The main result is that vacuum stability together with a physical fixed point in the gauge-Yukawa sector singles out a {\it unique} fixed point in the scalar sector. The scalar fixed points do not offer further parameter constraints other than those already stated in Tabs.~\ref{tPara15} and~\ref{tPara69}. Within the admissible parameter ranges
we invariably find that the scalar couplings are either strictly irrelevant (at interacting fixed points) or marginally irrelevant (at the Gaussian fixed point).
\section{\bf Ultraviolet completions}\label{class}
In this section, we discuss interacting fixed points and the weak coupling phase structure of minimal models \eq{L} in dependence on matter field multiplicities. Differences from the viewpoint of their high- and low-energy behaviour are highlighted.
\subsection{Classification}
In Figs.~\ref{pFPall},~\ref{pAll_AF},~\ref{pAll_AS} and~\ref{pAll_EFT} we summarise results for the qualitatively different types of quantum field theories with Lagrangean \eq{L} in view of their fixed point structure at weak coupling, together with their behaviour in the deep UV and IR. Theories differ primarily through their matter multiplicities \eq{RST}, which translate to the parameters $(P,R)$ and the sign of $\epsilon$, \eq{PRepsN}. As such, the ``phase space'' shown in Fig.~\ref{pFPall} arises as the overlay of Figs.~\ref{pFP23},~\ref{pFP45},~\ref{pFP6},~\ref{pFP78} and~\ref{pFP9}. Distinctive parameter regions are separated from each other by the seven characteristic curves $P=0\,, X\,, \widetilde{X}\,, Y$ or $\widetilde{Y}$ and $R=\s0{1}{2}$ or $2$. The functions $X(R)\,, \widetilde{X}(R)\,, Y(R)$ and $\widetilde{Y}(R)$ are given explicitly in \eq{P1234}.
Overall, this leads to the 22 distinct regions shown in Fig.~\ref{pFPall} and denoted by capital letters.
Together with the sign of $\epsilon$ this leaves us with 44 different cases. Some of these are redundant and related under the exchange of gauge groups, see \eq{exchange}. In fact, for $P>0$ and for either sign of $\epsilon$, we find nine fundamentally independent cases corresponding to the parameter regions
\begin{equation}\label{++}
\begin{array}{c}
\text{A\,,\ B\,,\ C\,,\ D\,,\ E\,,\ F\,,\ G\,,\ H\,,\ I}
\end{array}
\end{equation}
given in Fig.~\ref{pFPall}. Theories with parameters in the regime
\begin{equation}\label{++b}
\begin{array}{c}
\text{Ab\,,\ Bb\,,\ Cb\,,\ Db\,,\ Eb\,,\ Fb\,,\ Gb\,,\ Hb}\,,
\end{array}
\end{equation}
are ``dual'' to those in \eq{++}
under the exchange of gauge groups (X $\leftrightarrow$ Xb) and for the same sign of $\epsilon$, except for the theories within (I, $\epsilon$), which are ``selfdual'' and mapped onto themselves under \eq{exchange}.
For $P<0$ we find five parameter regions for either sign of $\epsilon$,
\begin{equation}\label{+-}
\text{J\,,\ K\,,\ L\,,\ Kb\,,\ Jb}\,.
\end{equation}
For these, the manifest ``duality'' under exchange of gauge groups involves a change of sign for $\epsilon$
with (X, $\epsilon<0$) being dual to (Xb, $-\epsilon>0$) except for the parameter region L which is selfdual. In total, we end up with $2\times 9+5=23$ fundamentally distinct scenarios underneath the $2\times (9+8+5)=44$ cases tabulated in Figs.~\ref{pAll_AF},~\ref{pAll_AS} and~\ref{pAll_EFT} and discussed more extensively below.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.35]{pAllb.pdf}
\vskip-.75cm
\caption{The ``phase space'' of quantum field theories with fundamental action \eq{L} expressed as a function of field multiplicities and written in terms of $(P,R)$, see \eq{PRepsN}. The 22 different parameter regions are indicated by roman letters.
Theories with parameters in region X are dual to those in region Xb under the exchange of gauge groups following the map \eq{exchange}. Further details on fixed points and their eigenvalue spectra per parameter region are summarised in Figs.~\ref{pAll_AF},~\ref{pAll_AS} and~\ref{pAll_EFT}.}\label{pFPall}
\end{center}
\end{figure}
A comment on the nomenclature: in each row of Figs.~\ref{pAll_AF},~\ref{pAll_AS} and~\ref{pAll_EFT}, we indicate the parameter region $(P,R)$ as in Fig.~\ref{pFPall} together with the sign of $\epsilon$ (if required), followed by the set of fixed points. For each of these, the (marginally) relevant and irrelevant eigenvalues in the gauge-Yukawa sector are indicated by a $-$ and $+$ sign. For the Gaussian fixed point \fp1, the signs relate pairwise to the $SU(N_{\rm C})$ and $SU(N_{\rm c})$ gauge sector, respectively; for all other fixed points eigenvalues are sorted by magnitude. Red shaded slots indicate eigenvalue spectra which uniquely arise due to the semi-simple character of the theory. The column `` UV'' states the UV fixed point, differentiating between complete asymptotic freedom (AF), asymptotic safety (AS), asymptotic freedom in one sector without asymptotic safety in the other (pAF), asymptotic safety in one sector without asymptotic freedom in the other (pAS), or none of the above. The column ``IR'' states the fully attractive IR fixed point (provided it exists), distinguishing the cases where none (0), one (Y) or (y), or both (Yy) Yukawa couplings are non-trivial at the fixed point; a hyphen indicates that the IR regime is strongly coupled.
\subsection{Asymptotic freedom}\label{secAF}
We discuss main features of the different quantum field theories \eq{L} starting with those where each gauge sector is asymptotically free from the outset $(P > 0 > \epsilon)$, corresponding to the cases $1-17$ in Fig.~\ref{pAll_AF}. The Gaussian fixed point \fp1 is always the UV fixed point. Any other weakly interacting fixed point displays a lower number of relevant directions. All weakly interacting fixed points can be reached from the Gaussian. Another point in common is that all theories are completely asymptotically free meaning that-- besides the gauge and the Yukawa couplings -- all quartics reach the Gaussian UV fixed point.
Differences arise as to the set of interacting fixed points, summarised in Fig.~\ref{pAll_AF}. Overall, theories display between three and eight distinct weakly interacting fixed points. The partial Banks-Zaks fixed points (\fp2, \fp3) are invariably present in all 17 cases. This is a consequence of a general theorem established in \cite{Bond:2016dvk}, which states that the two loop gauge coefficient is strictly positive for any gauge theory in the limit where the one-loop coefficient vanishes. This guarantees the existence of a partial Banks-Zaks fixed point in either gauge sector. At least one of the partial gauge-Yukawa fixed points (\fp4, \fp5) also arises in all cases. Moreover, the fully interacting Banks-Zaks (\fp6) as well as the fully interacting gauge-Yukawa fixed points (\fp7, \fp8, \fp9) are present in many, though not all, cases. All nine distinct fixed points are available in the ``most symmetric'' parameter region I (case 9).
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.24]{pAll_AF.pdf}
\caption{
Shown are the fixed points and eigenvalue spectra
of quantum field theories with Lagrangean \eq{L} for the 17 parameter regions with $\epsilon<0$ and $P >0$ in Fig.~\ref{pFPall}. Scalar selfinteractions are irrelevant at fixed points. All cases display complete asymptotic freedom in the UV. Red shaded slots indicate eigenvalue spectra which arise due to the semi-simple character of the theory.
In the deep IR, various types of interacting conformal fixed points are achieved depending on whether both, one, or none of the Yukawa couplings $Y$ and $y$ vanish (from left to right). Regimes with ``strong coupling only'' in the IR are indicated by a hyphen.
}
\label{pAll_AF}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.23]{pAll_AS.pdf}
\caption{
Same as Fig.~\ref{pAll_AF}, covering the 10 parameter regions with $P <0$ of Fig.~\ref{pFPall}.
Notice that \fp6 is absent throughout. Exact asymptotic safety (AS) is realised in the cases 22 and 23. Red shaded slots indicate eigenvalue spectra which arise due to the semi-simple character of the theory. For the cases $18-21$ and $24-27$, partial asymptotic freedom (pAF) or partial asymptotic safety (pAS) is observed whereby one gauge sector decouples entirely at all scales. The latter theories are only UV complete in one of the two gauge sectors and must be viewed as effective rather than fundamental.
}
\label{pAll_AS}
\end{center}
\end{figure}
It is noteworthy that many theories
display a fully IR attractive ``sink'', invariably given by an IR gauge-Yukawa fixed point in one (\fp4, \fp5) or both gauge sectors (\fp9). In Fig.~\ref{pFPall}, this happens for matter field multiplicities in the regions A, B, C, E, F, G, I and their duals (cases 1, 2, 3, 5, 6, 7, 9, 11, 12, 13, 15, 16 and 17 of Fig.~\ref{pAll_AF}).
At \fp9, the fully IR attractive fixed point is largely a consequence of IR attractive fixed points in each gauge sector individually. This is not altered qualitatively by the semi-simple nature of the model. As such, a fully IR attractive fixed point \fp9 also arises in the ``direct product'' limit where the $\psi$ fermions are removed.
At \fp4 and \fp5, in contrast, the IR sink is a direct consequence of the semi~simple nature of the theory in that it would be strictly absent as soon as the messenger fermions $\psi$ are removed. Most importantly, the IR gauge~Yukawa fixed point in one gauge sector changes the sign of the effective one loop coefficient in the other, mediated via the $\psi$ fermions. This secondary effect means that one gauge sector becomes IR free dynamically, rather than remaining UV free. Overall, the fixed point becomes IR attractive in all canonically marginal couplings (including the quartic couplings).
In most cases the IR sink is unique except in parameter regions B and F (case 2, 6, 12 and 16) where we find two competing and inequivalent IR sinks (\fp4 versus \fp5).
Provided that one or both Yukawa couplings take Gaussian values, other fixed points may take over the role of IR ``sinks''. In these settings, one or both of the elementary ``meson'' fields remain free for all scales and decouple from the outset. Specifically, the IR sink is given by \fp6 provided that $y=0=Y$ (cases 5 -- 13); by \fp2 or \fp7 provided that $y=0$ (cases 14 or 7 -- 9, respectively); and by \fp3 or \fp8 provided that $Y=0$ (cases 4 or 9 -- 11). We note that \fp6, \fp7 and \fp8 are natural IR sinks, with or without $\psi$ fermions, provided that all Yukawa couplings of those fermions which interact with the Banks-Zaks fixed point(s) vanish. On the other hand, the result that \fp2 and \fp3 may become IR sinks is a strict consequence of the $\psi$ fermions and would not arise otherwise. Once more, one of the gauge sectors becomes IR free owing to the BZ fixed point in the other, an effect which is mediated via the $\psi$ fermions.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.225]{pAll_EFT.pdf}
\caption{Same as Figs~\ref{pAll_AF} and~\ref{pAll_AS},
covering the 17 parameter regions where
$\epsilon>0$ and $P>0$ in Fig.~\ref{pFPall}. Asymptotic freedom is absent in both gauge sectors implying that \fp2, \fp3, \fp6, \fp7 and \fp8 cannot arise.
Partial asymptotic safety (in one gauge sector) is observed in case $28, 31, 32, 35, 37, 40, 41$ and $44$, whereby the other gauge sector remains free at all scales (pAS). All models must be viewed as effective rather than fundamental theories. All theories become trivial in the IR.
}
\label{pAll_EFT}
\end{center}
\vskip-.5cm
\end{figure}
In the presence of non-trivial Yukawa couplings, no fully IR stable fixed point arises for theories with field multiplicities in the parameter regions D and H (case 4, 8, 10 and 14). Generically, trajectories will then run towards strong coupling with $e.g.$~confinement or strongly-coupled IR conformality. Analogous conclusions hold true in settings with fully attractive IR fixed points provided their basins of attraction do not include the Gaussian.
Finally, another interesting feature which is entirely due to the semi~simple nature of the theory are models where \fp9 has a single relevant direction (cases 1, 2, 5, 6, 12, 13, 16 and 17). Whenever this arises, the theory also always displays a fully IR attractive fixed point (\fp4, \fp5, or both).
\subsection{Asymptotic safety}
We now turn to quantum field theories with \eq{L} where asymptotic safety is realised. Asymptotic safety relates to settings where some or all couplings take non-zero values in the UV \cite{Bond:2016dvk}. A prerequisite for this is the absence of asymptotic freedom in at least one of the gauge sectors. We find two such examples provided $P<0$ (cases 22 and 23 in Fig.~\ref{pAll_AS}), corresponding precisely to settings where one gauge sector is QCD-like whereas the other is QED-like. For these theories, we furthermore find that all other interacting fixed points are also present, except those of the Banks-Zaks type involving the QED-like gauge sector. More specifically, in case 22 the role of the asymptotically safe UV fixed point is now taken by \fp5. The UV critical surface is three-dimensional, in distinction to asymptotically free settings where it is four-dimensional. This reduction, ultimately a consequence of an interacting fixed point in one of the Yukawa couplings, leads to enhanced predictivity of the theory. The Gaussian necessarily becomes a cross-over fixed point with both attractive and repulsive directions, similar to the interacting \fp8. Also, \fp2 and \fp9 are realised with a one-dimensional critical surface. The fully IR attractive \fp4 -- the counterpart of the UV fixed point \fp5 -- takes the role of an IR ``sink''.
In the low energy limit, the theory displays free $SU(N_c)$ ``gluons'' in one gauge sector and weakly interacting $SU(N_C)$ ``gluons'' in the other. Moreover, the spectrum includes both free and weakly interacting mesons related to the former and the latter sectors, as well as free and weakly interacting fermions. Qualitatively, a similar result arises in the ``direct product'' limit, showing that the semi-simple nature of \eq{L} is not crucial for this scenario.
A noteworthy feature of semi-simple theories with asymptotic safety is that they connect an interacting UV fixed point with an interacting IR fixed point. Hence, our models offer examples of quantum field theories with exact UV and IR conformality, strictly controlled by perturbation theory for all scales. In the massless limit, the phase diagram has trajectories connecting the interacting UV fixed point with the interacting IR fixed point. Some trajectories may escape towards the regime of strong coupling where the theory is expected to display confinement, possibly infrared conformality. The same picture arises in case 23 after exchange of gauge groups.
No asymptotically safe fixed point arises if both gauge sectors are IR free $(P, \epsilon > 0)$. This result is in marked contrast to findings in the ``direct product'' limit where models with an interacting UV fixed points exist -- simply because it exists for the simple gauge factors \eq{Simple} and \eq{Simple2}, given suitable matter field multiplicities. We conclude that it is precisely the semi-simple nature of the speific set of theories \eq{L} which disallows asymptotic safety for settings with $P,\epsilon>0$, see \eq{PRepsN}.
\subsection{Effective field theories}\label{eff}
We now turn to quantum field theories with \eq{L} which are not UV complete semi-simple gauge theories and, as such, must be seen as effective field theories. We find three different types of these. Firstly, we find models with partial asymptotic freedom (pAF), where one gauge sector remains asymptotically free whereas the other stays infrared free. These models always realise a Banks-Zaks fixed point (as they must), and some also realise an IR gauge-Yukawa fixed point. When viewed as a fundamental theory, the IR free sector decouples exactly, for all RG scales, and the theory becomes a simple asymptotically free gauge theory (which is UV complete). The IR-free sector can be interacting when viewed as an effective theory, very much like the $U(1)_Y$ sector of the Standard Model. This setting requires $P<0$ and is realised in cases 18 -- 21 and 23 -- 27.
Secondly, we find models with partial asymptotic safety (pAS), where one gauge sector becomes asymptotically safe whereas the other remains free at all scales. All such models display a UV gauge-Yukawa fixed point.
When viewed as a fundamental theory, these semi-simple gauge theories in fact reduce to a simple asymptotically safe gauge theory (which is UV complete). The IR-free sector can be interacting when viewed as a non-UV complete effective theory.
This setting mostly requires $P,\epsilon>0$ and is realised in cases 28, 31, 32, 35, 37, 40, 41 and 44. Curiously, pAS is also realised in cases 21 and 24 where $P<0$ alongside pAF in the other gauge sector --- such models have two disconnected UV scenarios, where we can choose to have either asymptotic freedom in one sector, or asymptotic safety in the other, in each case with the remaining sector decoupling at all scales. Once more, if both gauge sectors are interacting these models must be viewed as (non-UV complete) effective theories.
Finally, we find models with none of the above. In these settings (cases 29, 30, 33, 34, 36, 38, 39, 42 and 43), both gauge sectors are IR free and no other weakly coupled fixed points are realised, leaving us with no perturbative UV completion.
In the cases 28 -- 44, the Gaussian acts as in IR ``sink'' for RG trajectories. Along these, the long-distance behaviour is trivial, characterised by free massless non-Abelian gauge fields, quarks, and elementary mesons.
In summary, the semi-simple gauge Yukawa theories \eq{L} have a well-defined UV limit with either asymptotic freedom or asymptotic safety in 9+1=10 cases out of the 23 fundamentally distinct parameter settings covered in Fig.~\ref{pFPall}. The remaining 4+9=13 parameter settings do not offer a well-defined UV limit at weak coupling. This completes the classification of the models with \eq{L}.
\begin{figure}[t]
\begin{center}
\vskip-1cm
\includegraphics[scale=.4]{pSemiSimpleBZs.pdf}
\caption{Phase diagrams of asymptotically free semi-simple gauge theories (two gauge groups) coupled to matter without Yukawas, covering $a)$ asymptotic freedom and the Gaussian (G) without interacting fixed points and trajectories running towards strong coupling and confinement, $b)$ the same, with an additional Banks-Zaks (BZ) fixed point, $c)$ two BZ fixed points, one of which turned into an IR sink for all trajectories, or $d)$ three BZ fixed points, the fully interacting one now becoming the IR sink. Axes show the running gauge couplings, fixed points (black) are connected by separatrices (red), and red-shaded areas cover all UV free trajectories with arrows pointing from the UV to the IR.
}\label{pSemiSimpleBZs}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\vskip-1cm
\includegraphics[scale=.4]{pSimpleCasesPDred.pdf}
\caption{Phase diagrams of UV complete and weakly interacting simple gauge theories coupled to matter with a single Yukawa coupling, covering $a)$ asymptotic freedom with the Gaussian UV fixed point and no other weakly interacting fixed point, $b)$ asymptotic freedom with a Banks-Zaks (BZ) fixed point, $c)$ asymptotic freedom with a Banks-Zaks and an IR gauge-Yukawa (GY) fixed point, and $d)$ asymptotic safety with an UV gauge-Yukawa fixed point. Axes display the running gauge and Yukawa couplings, fixed points (black) are connected by separatrices (red), and red-shaded areas cover all UV free trajectories with arrows pointing from the UV to the IR \cite{Bond:2016dvk,Bond:2017sem}. Examples are given by \eq{Simple}, \eq{Simple2} (see main text).
}\label{pSimpleCasesPD}
\end{center}
\end{figure}
\begin{figure}[t]
\hskip1cm
\includegraphics[scale=.33]{pPrimitiveSimple_v2.pdf}
\caption{``Primitives'' for phase diagrams of simple gauge-Yukawa theories with asymptotic freedom (AF) or asymptotic safety (AS), corresponding to the different setting shown in Fig.~\ref{pSimpleCasesPD}. Arrows point from the UV to the IR and connect the different fixed points. Open arrows point towards strong coupling in the IR.
The number of outgoing red arrows gives the dimensionality of the UV critical surface. The separate UV safe trajectory towards strong coupling in case $d)$ is not indicated. Yukawa-induced IR unstable directions in $a,b)$ or gauge Yukawa fixed points in $c,d)$
are absent as soon as Yukawa interactions are switched off from the outset. }\label{pPrimitivesSimple}
\end{figure}
\section{\bf Phase diagrams of gauge theories}\label{PD}
In this section, we discuss the phase diagrams of UV complete theories of the type \eq{L}, particularly in view of theories with asymptotic freedom or asymptotic safety.
\subsection{Semi-simple gauge theories without Yukawas}
We begin with settings where Yukawa couplings are switched off. In these cases, interacting fixed points can only arise for asymptotically free gauge sectors, and fixed points are of the Banks-Zaks type or products thereof \cite{Bond:2016dvk,Bond:2017sem}. Qualitatively different cases realised amongst the theories \eq{L} are summarised in
Fig.~\ref{pSemiSimpleBZs} for semi-simple gauge theories with two gauge groups $G_1\times G_2$. Results generalise to more gauge groups in an obvious manner.
Specifically, Fig.~\ref{pSemiSimpleBZs}$a)$ shows theories with asymptotic freedom but without any BZ fixed points. UV free trajectories emanate out of the Gaussian fixed point and invariably escape towards strong coupling where the theory is expected to display confinement, or IR conformality. Similarily, Fig.~\ref{pSemiSimpleBZs}$b)$ shows theories with asymptotic freedom and a BZ fixed point in one of the gauge sectors. The other gauge coupling remains an IR relevant perturbation even at the BZ. Therefore UV free trajectories will again escape towards strong coupling in the IR.
Fig.~\ref{pSemiSimpleBZs}$c)$ shows asymptotic freedom with a BZ fixed point in both gauge sectors individually. Here, and much unlike Fig.~\ref{pSemiSimpleBZs}$b)$, one of the BZ fixed points has turned into an exact IR ``sink'', and both BZ fixed points are connected by a separatrix. As we have already noticed in Sect.~\ref{secAF}, the presence of an interacting fixed point in one gauge sector can turn the other gauge sector from UV free to IR free.
This new type of phenomenon has become possibe owing to the $\psi$ fermions and is once again due to the semi-simple nature of the theory. Therefore, all UV free trajectories invariably are attracted into the IR sink. In the deep IR, the theory approaches a conformal fixed point with massless and unconfined free and weakly coupled gluons and quarks. Regimes of strong coupling cannot be reached.
Fig.~\ref{pSemiSimpleBZs}$d)$ shows asymptotic freedom with a (partial) BZ fixed point in either gauge sector individually, as well as a fully interacting BZ fixed point. Most notably, all UV free trajectories are attracted by the later, which acts as an IR sink. No trajectories can escape towards strong coupling. The long distance physics is characterised by an interacting conformal field theory with massless weakly coupled gauge fields and fermions. Here, and unlike in Fig.~\ref{pSemiSimpleBZs}$c)$, all fields remain weakly coupled in the IR.
In the scenarios of Fig.~\ref{pSemiSimpleBZs}$a)$ and $b)$ UV free trajectoires run towards strong coupling and confinement in the IR, in one or both gauge sectors. In contrast, the scenarios in Fig.~\ref{pSemiSimpleBZs}$c)$ and $d)$ show that all UV free trajectories are attracted by an IR-stable conformal fixed point. These theories remain unconfined and perturbative at all scales. All four scenarios in Fig.~\ref{pSemiSimpleBZs} are realised for our template of semi-simple gauge theories with Lagrangean \eq{L}. Explicit examples are given for models without Yukawa couplings $(Y=0=y)$ and for field multiplicities in the parameter regions
$a)$ $\epsilon_1, \epsilon_2<-75/26$,
$b)$ $\epsilon_1<-75/26$ and $-75/26<\epsilon_2<0$, or $(\epsilon_1\leftrightarrow \epsilon_2)$,
$c)$ the cases 1 -- 4 and 14 -- 17 of Fig.~\ref{pAll_AF}, and
$d)$ the cases 5 -- 13 of Fig.~\ref{pAll_AF}.
\subsection{Simple gauge theories with Yukawas}
We continue the discussion of phase diagrams with simple gauge theories with gauge group $G$ and a single Yukawa coupling. Four distinct cases can arise \cite{Bond:2016dvk,Bond:2017sem},
summarised in Fig.~\ref{pSimpleCasesPD}. For asymptotically free settings, the theory either shows $a)$ only the Gaussian UV fixed point, $b)$ the Gaussian together with the Banks-Zaks, or $c)$ the Gaussian together with the Banks-Zaks and an IR gauge-Yukawa fixed point. Simple gauge theories can also become asymptotically safe, in which case $d)$ a UV gauge-Yukawa fixed point arises. Trajectories are directed towards the IR. The red-shaded areas indicate the set of UV complete trajectories emanating out of the UV fixed point. We genuinely observe a two-dimensional area of trajectories for asymptotically free settings, which is reduced to a one-dimensional set in the asymptotically safe scenario. The IR regime is characterised by either strong interactions and confinement such as in Fig.~\ref{pSimpleCasesPD}$a,b,d)$, or by an interacting conformal field theory with weakly coupled gluons and fermions alongside free or interacting scalar mesons ---corresponding to the BZ fixed points in Fig.~\ref{pSimpleCasesPD}$b)$ and $c)$, or the IR GY fixed point in Fig.~\ref{pSimpleCasesPD}$c)$, respectively---, or by Gaussian scaling, Fig.~\ref{pSimpleCasesPD}$d)$.
All four scenarios in Fig.~\ref{pSimpleCasesPD} are realised for simple gauge theories with \eq{Simple} corresponding to the parameter regions
$a)$ $\epsilon_1<-75/26$, $b)$ $-75/26<\epsilon_1<0$ and $R>\s012$, $c)$ $-75/26<\epsilon_1<0$ and $R<\s012$, or $d)$ $\epsilon_1>0$ and $R<\s012$, respectively, with $R$ additionally bounded by \eq{Rrelaxed}.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.45]{pAF_BZ_IR.pdf}
\caption{
Schematic phase diagram for asymptotically free semi-simple gauge theories \eq{L} with Banks-Zaks type fixed points without Yukawas and exact IR conformality. Field multiplicities correspond to the cases $a)$ 1 -- 4, $b)$ 5 -- 13, and $c)$ 14 -- 17 of Fig.~\ref{pAll_AF}, respectively, with scalars decoupled.
RG flows point from the UV to the IR (top to bottom). At each fixed point, the dimensionality of the UV critical surface is given by the number of outgoing red arrows. All UV free trajectories terminate at \fp2, \fp6 and \fp3, respectively, which act as fully attractive IR ``sinks''.
The topology of the phase diagram $b)$ is the ``square'' of Fig.~\ref{pPrimitivesSimple}$b)$, representing Fig.~\ref{pSemiSimpleBZs}$d)$. The phase diagrams $a)$ and $c)$, representing Fig.~\ref{pSemiSimpleBZs}$c)$, cannot be constructed from the primitives in Fig.~\ref{pPrimitivesSimple}.
}\label{pPD_BZ_IR}
\end{center}
\end{figure}
An economic way to display phase diagrams for semi-simple theories with or without Yukawas is achieved by introducing a schematic diagrammatic language, see Fig.~\ref{pPrimitivesSimple}. Each of the four basic phase diagrams in Fig.~\ref{pSimpleCasesPD} are represented by a ``primitive'' diagram, Fig.~\ref{pPrimitivesSimple}, where full dots indicate (free or interacting) fixed points, red arrows indicate the outgoing trajectories, and RG flows schematically run ``top-down'' from the UV to the IR. Also, at each fixed point the number of outgoing arrows indicates the dimensionality of the fixed point's ``UV critical surface''. Fixed points are connected by separatrices. We use straight lines to indicate separatrices involving the BZ fixed point, curved lines to indicate separatrices connecting GY fixed points with the Gaussian, and open-ended lines to denote RG trajectories running towards strong coupling without reaching any weakly coupled fixed points.
Specifically, in case $a)$, a two-dimensional array of RG flows are running out of the Gaussian UV fixed point towards strong coupling, with no weakly interacting fixed points. In case $b)$, we additionally observe a Banks-Zaks fixed point. It is connected with the Gaussian by a separatrix shown in red. Arrows invariably point towards the IR. Yukawa couplings act as an unstable direction at both fixed points.
In case $c)$, we additionally observe a gauge-Yukawa fixed point besides the Gaussian and the BZ. All three fixed points are connected by separatrices. Note that two lines emanate from the Gaussian, reflecting that the UV critical surface is two dimensional. The GY fixed point arises as an IR sink, which attracts all UV-free trajectories emanating out of the Gaussian.
In case $d)$, the model is asymptotically safe and the GY fixed point has become the interacting UV fixed point. A Banks-Zaks fixed point can no longer arise \cite{Bond:2016dvk}. The theory has a one-dimensional UV critical surface connecting the GY fixed point with the IR Gaussian fixed point via a separatrix. A second UV safe trajectory which leaves the GY fixed point towards strong coupling is not depicted. Finally, we note that the Yukawa-induced IR unstable directions in $a)$ and $b)$ or gauge Yukawa fixed points in $c)$ and $d)$ are absent as soon as Yukawa interactions are switched off from the outset.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.35]{pPD_AF.pdf}
\caption{
Asymptotic freedom and
schematic phase diagram for semi simple gauge-Yukawa theories with field multiplicities as in
case 9 of Fig.~\ref{pAll_AF}.
RG flows point from the UV to the IR (top to bottom). Besides the Gaussian UV fixed point (FP${}_1$), the theory displays all eight weakly interacting fixed points, see Tab.~\ref{tFPs}.
At each fixed point, the dimensionality of the UV critical surface is given by the number of outgoing red arrows. FP${}_9$ is fully attractive and acts as an IR ``sink''. The topology of the phase diagram is the ``square'' of Figs.~\ref{pSimpleCasesPD},\ref{pPrimitivesSimple}$c)$; see main text.
}\label{pPD_AF}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.35]{pPD_AF_pt2.pdf}
\caption{
Asymptotic freedom and schematic phase diagrams for semi-simple gauge-Yukawa theories
with field multiplicities as in case 8 of Fig.~\ref{pAll_AF}.
Flows point from the UV to the IR (top to bottom). The theories display five weakly interacting fixed points besides the Gaussian UV fixed point (FP${}_1$). The unavailability of \fp5, \fp8 and \fp9 implies that some trajectories escape towards
strong coupling (short arrows), and none of the fixed points acts as a complete IR attractor.
The topology of the phase diagram is the ``direct product'' of Fig.~\ref{pSimpleCasesPD},~\ref{pPrimitivesSimple}$c)$ with Fig.~\ref{pSimpleCasesPD},~\ref{pPrimitivesSimple}$b)$; see main text.
The IR unstable direction is removed provided that the Yukawa coupling $y\equiv 0$, in which case the singlet mesons $h$ decouple.
}\label{pPD_AF_pt2}
\end{center}
\end{figure}
\subsection{Semi-simple gauge theories with asymptotic freedom}
We consider phase diagrams for semi-simple theories \eq{L} with complete asymptotic freedom, exemplified by all models in Fig.~\ref{pAll_AF}. When Yukawa couplings are absent, the meson-like scalar degrees of freedom remain free at all scales and decouple from the theory. In the regime with asymptotic freedom solely Banks-Zaks fixed points can arise in the IR. Fig.~\ref{pSemiSimpleBZs}$d)$ and Fig.~\ref{pPD_BZ_IR}$b)$ shows settings where all Banks-Zaks fixed points are present, corresponding to
the cases 5 -- 13 of Fig.~\ref{pAll_AF}.
RG flows point from the UV to the IR (top to bottom) and connect the Gaussian UV fixed point (FP${}_1$) with either of the partially (\fp2 and \fp3) and the fully interacting (\fp6) Banks-Zaks fixed points. The latter is fully attractive and acts as an IR sink. The topology of the phase diagram is the ``square'' of Figs.~\ref{pSimpleCasesPD},\ref{pPrimitivesSimple}$b)$. In the deep IR the theory is unconfined yet weakly interacting, and the elementary gauge fields $A, a$ and fermions $Q, q$ and $\psi$ appear as massless particles at the IR conformal fixed point.
The phase diagrams in Figs.~\ref{pPD_BZ_IR}$a)$ and $c)$
cannot be constructed out of the simple primitives, Fig.~\ref{pPrimitivesSimple}. The reason for this is that the eigenvalue spectrum at one of the fixed points deviates from the ``direct product'' spectrum due to interactions.
Next we include Yukawa interactions.
We have already concluded from Fig.~\ref{pAll_AF} that the eigenvalue spectrum in the cases 8, 9 and 10 agrees qualitatively, for all fixed points, with the eigenvalue spectrum in the corresponding ``direct product'' limit. In these settings, we may then use the primitives in Fig.~\ref{pPrimitivesSimple} to find the semi-simple phase diagrams.
We consider the case where the parameters \eq{PRepsN} take values within the range I of Fig.~\ref{pFPall} and for $\epsilon<0$, corresponding to case 9 of Fig.~\ref{pAll_AF}. This family of theories includes the ``symmetric'' setup $(R,P)=(1,1)$ where symmetry under the exchange of gauge groups is manifest. The UV fixed point is given by the Gaussian (FP${}_1$), and the UV critical surface at the Gaussian is four-dimensional, owing to the marginal UV relevancy of the two gauge and the two Yukawa couplings. All scalar couplings are irrelevant in the UV and can be expressed in terms of the gauge and the Yukawa couplings along UV-free trajectories.
Moreover, each gauge sector displays the Banks-Zaks and a gauge-Yukawa fixed point individually, and all nine fixed points are realised in the full theory.
Since the sign pattern of the eigenvalue spectra at all fixed points
is equivalent to the ``direct product'' limit, the topology of the semi-simple phase diagram is the ``square'' of Fig.~\ref{pPrimitivesSimple}$c)$ -- shown in Fig.~\ref{pPD_AF}. Fixed points are connected by separatrices (red lines), and arrows always point towards the IR.
From top to bottom, the fixed points FP${}_1$ (FP${}_{2,3})$ [FP${}_{4,5,6}$] (FP${}_{7,8})$ and FP${}_9$ have a 4 (3) [2] (1) and 0-dimensional UV critical surface, respectively, corresponding to the number of outgoing red arrows. FP${}_9$ acts as an IR attractor for all trajectories within its basin of attraction. Consequently, the elementary quarks and gluons are not confined and the theory corresponds to a conformal field theory of weakly interacting massless gluons, fermions and mesons in the deep IR. For certain fine-tuned settings, the IR limit would, instead, correspond to one of the other interacting fixed points FP${}_{2}$ -- FP${}_{8}$, relating to different conformal field theories. Also, while all other fixed points can be reached from the Gaussian FP${}_1$ (whose UV critical surface has the largest dimensionality), it is not true in general that a fixed point with a smaller UV critical dimension can be reached from a fixed point with a larger one. Fixed points are also not connected ``horizontally''.
As a further example we consider a less symmetrical setting given by models with \eq{PRepsN} in the parameter range H (or Hb) of Fig.~\ref{pFPall}, and for $\epsilon<0$. In these theories, only one of the two gauge sectors can achieve a gauge-Yukawa fixed point. Consequently, six different types of fixed points are realised. The sign pattern of the eigenvalue spectrum (cases 8 or 10, Fig.~\ref{pAll_AF}) ensures
that the topology of the semi-simple phase diagram obtains as the direct product of Fig.~\ref{pPrimitivesSimple}$b)$ with Fig.~\ref{pPrimitivesSimple}$c)$, shown in Fig.~\ref{pPD_AF_pt2}. From top to bottom, the fixed points FP${}_1$ (FP${}_{2,3})$ [FP${}_{4,6}$] and FP${}_7$ have a 4 (3) [2] and 1-dimensional UV critical surface, respectively.
Fixed points are connected by separatrices. The absence of \fp5, \fp8 and \fp9 implies that some trajectories escape towards
strong coupling, indicated by short arrows, from each of the fixed points. The unstable direction relates to the Yukawa coupling $y$ in \eq{L}. Provided it is switched off, \fp7 would become the fully attractive IR ``sink''. In this case, the elementary mesons $h$ are spectators and remain free at all scales. Also, the elementary quarks and gluons remain unconfined. In the deep IR, the theory corresponds to a conformal field theory of weakly interacting massless gluons $A$, fermions $Q, \psi$ and mesons $H$, together with free and massless gluons $a$, fermions $q$ and mesons $h$, see Tab.~\ref{tCharges}. For certain fine-tuned settings, the IR limit would, instead, correspond to one of the other interacting fixed points FP${}_{2}$ -- FP${}_{8}$, relating to different conformal field theories.
The phase diagrams of asymptotically free theories in the cases 1 -- 7 and 11 -- 17 of Fig.~\ref{pAll_AF} cannot be constructed out of the simple primitives, Fig.~\ref{pPrimitivesSimple}. The reason for this is that their eigenvalue spectrum at some of the interacting fixed points deviates from the ``direct product'' spectrum. Once again this effect is due to the semi-simple nature of the theory. A more detailed study of these cases is left for future work.
\begin{figure}[t]
\includegraphics[scale=0.35]{pPD_AS.pdf}
\begin{center}
\caption{
Asymptotic safety and schematic phase diagram of semi simple gauge-Yukawa theories with field multiplicities as in case 22 of Fig.~\ref{pAll_AS}).
Besides the partially interacting UV fixed point (FP${}_5$), the theory displays five weakly interacting fixed points. The Gaussian (FP${}_1$) takes the role of a crossover fixed point and
FP${}_4$ takes the role of an IR sink. The topology of the phase diagram is the ``direct product'' of Fig.~\ref{pSimpleCasesPD},~\ref{pPrimitivesSimple}$c)$ with Fig.~\ref{pSimpleCasesPD},~\ref{pPrimitivesSimple}$d)$; see main text.}
\label{pPD_AS}
\end{center}
\end{figure}
\subsection{Semi-simple gauge theories with asymptotic safety}
We finally turn to the phase diagram of semi-simple gauge theories with exact asymptotic safety. From Figs.~\ref{pAll_AS} and~\ref{pAll_EFT} we conclude that asymptotic safety arises through a partially interacting UV fixed point where one gauge sector is interacting whereas the other gauge sector is free. This is achieved for matter field multiplicities \eq{PRepsN} taking values within the range J or Jb of Fig.~\ref{pFPall}, corresponding to cases 22 or 23 of Fig.~\ref{pAll_AS}. Once more, the eigenvalue spectra at all fixed points are equivalent to the ones in the direct product limit, implying that the phase diagram arises as the direct product
of the corresponding ``simple factors'' Fig.~\ref{pSimpleCasesPD},~\ref{pPrimitivesSimple}$c)$ and Fig.~\ref{pSimpleCasesPD},~\ref{pPrimitivesSimple}$d)$.
Fig.~\ref{pPD_AS} shows the schematic phase diagram for case 22,
where the asymptotically safe UV fixed point FP${}_5$ is of the {G$\,\cdot\,$GY}~type (see Tab.~\ref{tFPs} and~\ref{tFPeps}). Unlike the cases with asymptotic freedom, here, the UV hypercritical surface is three rather than four dimensional. The reason for this is that one of the Yukawa couplings is taking an interacting UV fixed point.
At each fixed point, the number of outgoing directions indicate the dimensionality of the fixed point's critical hypersurface. From top to bottom, the fixed points FP${}_5$ (FP${}_{1,8})$ [FP${}_{2,9}$] and FP${}_4$ have a 3 (2) [1] and 0-dimensional UV critical surface.
UV finite trajectories connect FP${}_5$ via intermediate cross-over fixed points with the fully IR attractive fixed point FP${}_4$, which is of the {GY$\,\cdot\,$G} type. At weak coupling, all UV-IR connecting trajectories proceed either via the Gaussian FP${}_1$ ({G$\,\cdot\,$G}) and FP${}_2$ ({BZ$\,\cdot\,$G}), or via FP${}_8$ ({BZ$\,\cdot\,$GY}) and FP${}_2$ or FP${}_9$ ({GY$\,\cdot\,$GY}). The Gaussian fixed point is IR free in one of the gauge couplings meaning that is necessarily arises as a cross-over fixed point. There are no trajectories connecting the fixed points FP${}_1$ with FP${}_9$ because the sole relevant direction at the latter is an irrelevant direction at the former. \fp4 acts as an IR ``sink'' for RG trajectories. While all other fixed points can be reached from the interacting UV fixed point FP${}_5$ (whose UV critical surface has the largest dimensionality), it is not true in general that a fixed point with a smaller UV critical dimension can be reached from a fixed point with a larger one (e.g.~\fp9 cannot be reached from \fp1). Fixed points are also not connected ``horizontally''.
An intriguing novelty of our models with asymptotic safety is that both the deep UV and the deep IR limits are characterised by weakly interacting conformal field theories. For example, in the deep UV the theories of case 22 correspond to conformal field theories of weakly interacting massless gluons $a$, fermions $q, \psi$ and mesons $h$, together with free and massless gluons $A$, fermions $Q$ and mesons $H$. Along the UV -- IR transition, the fields $(A,Q,H)$ and $(a,q,h)$ effectively ``interchange'' their roles, ultimately approaching conformal field theories of weakly interacting massless gluons $A$, fermions $Q, \psi$, and mesons $H$, together with free and massless gluons $a$, fermions $q$ and mesons $h$ in the IR. Hence, one may say that IR conformality in the $SU(N_{\rm c})$ gauge sector arises from UV conformality in the $SU(N_{\rm C})$ gauge sector through a ``see-saw'' mechanism transmitted via the $\psi$ fermions, $i.e.$ the only fields which are interacting at all scales including the UV and the IR limits. For certain fine-tuned settings, the IR limit would, instead, correspond to one of the other interacting fixed points FP${}_{1}$, \fp2, FP${}_{8}$ or \fp9, relating to different conformal field theories. Also, for certain UV parameters, theories may escape towards strong coupling in the IR.
\subsection{Mass deformations and phase transitions}
In the vicinity of fixed points phase transitions between different phases arise once mass terms are switched on. At weak coupling mass anomalous dimensions are perturbatively small (Sec.~\ref{anomalous}). The running of scalar or fermion mass terms, once switched on, will then be dominated by their canonical mass dimensions -- modulo small quantum corrections. Consequently, mass terms add additional relevant directions at all fixed points (e.g.~Figs.~\ref{pPD_AF} -- \ref{pPD_AS}). Each of the eight interacting UV fixed points relates to a quantum phase transition between phases with and without spontaneous breaking of symmetry where the vacuum expectation value of the scalar fields serves as an order parameter.
In particular, fixed points which act as IR sinks for the canonically marginal interactions (such as \fp9 in Fig.~\ref{pPD_AF} and \fp4 in Fig.~\ref{pPD_AS}) develop new unstable directions driven by the mass. Scalar fields may or may not develop vacuum expectation values leading to symmetric and symmetry broken phases, respectively. Also, fermions may acquire masses spontaneously. Thereby a variety of different phases may arise, connected by first and higher order quantum phase transitions. Close to interacting fixed points, phase transitions are continuous and, in some cases, of the Wilson-Fisher type with a single relevant parameter. We leave a more detailed investigation of phase transitions for a future study.
\section{\bf Discussion}\label{disc}
In this section, we address further aspects of interacting fixed points covering
universality and operator ordering, triviality bounds, perturbativity in and beyond the Veneziano limit, conformal symmetry, and conformal windows.
\subsection{Gap, universality, and operator ordering}
At partially or fully interacting fixed points, the degeneracy of the nine classically marginal couplings \eq{couplings}, \eq{w} is partly or fully lifted. We have computed scaling exponents to the leading non-trivial order in $\epsilon$. Interacting fixed points have non-trivial exponents of order $\sim \epsilon$, except if a gauge coupling is involved in which case one of the exponents is parametrically smaller $\sim\epsilon^2$. Hence, the eigenvalue spectrum opens up $\sim \epsilon$ because eigenvalues of order $\epsilon$ are invariably present at any of the interacting fixed points. It is convenient to denote the difference between the smallest negative eigenvalue and the smallest positive eigenvalue as the ``gap'' in the eigenvalue spectrum, which serves as an indicator for interaction strength \cite{Falls:2013bv,Falls:2014tra}. Simple $SU(N)$ gauge theories in the Veneziano limit such as \eq{Simple} display a gap of order $\sim \epsilon$ $(\epsilon^2)$ at the Banks-Zaks or the UV gauge Yukawa (IR gauge Yukawa) fixed point, respectively \cite{Litim:2014uca}.
In semi-simple theories, and depending on the specifics of the fixed point, we again find that the gap is either of order $\epsilon$ or of order $\epsilon^2$. (The gap trivially vanishes if one of the gauge sectors is asymptotically free and takes Gaussian values.) The gap still depends on the remaining free parameters $(P,R)$.
Also, all results for fixed points and scaling exponents are universal and independent of the RG scheme, although we have used a specific scheme (MS bar) throughout. This is obviously correct for dimensionless couplings at one loop where divergences are logarithmic. We have checked that it also holds at two loop level both for the gauge sectors, and for the Yukawa contributions to the running of the gauge coupling(s) \cite{Litim:2014uca}. The field strengths and the Yukawa couplings are marginally relevant operators at asymptotically free Gaussian UV fixed points (case 1 -- 17 of Fig.~\ref{pAll_AF}). At asymptotically safe UV fixed points, one of the field strengths becomes relevant and the corresponding Yukawa coupling irrelevant (case 22, 23 of Fig.~\ref{pAll_AS}). There is no UV fixed point where both gauge sectors remain interacting. The scalar selfinteractions are (marginally) irrelevant at any fixed point.
\subsection{Elementary gauge fields and scalars}
Triviality bounds relate to perturbative UV Landau poles of infrared free interactions. They limit the predictivity of theories to a maximal UV extension \cite{Callaway:1988ya}.
For theories with action \eq{L}, perturbative UV Landau poles can arise for gauge couplings
in the absence of asymptotic freedom or asymptotic safety.
Examples for this are given in cases 18 -- 21 and 24 -- 27 of Fig.~\ref{pAll_AS} where one gauge sector is IR free, as well as in cases 28 -- 44 of Fig.~\ref{pAll_EFT}
where both gauge sectors are IR free. In these cases the theories can at best be treated as effective rather than fundamental (see Sect.~\ref{eff}). Conversely, triviality in gauge sectors is trivially avoided in settings with asymptotic freedom (such as in cases 1 -- 17), and non-trivially in settings with asymptotic safety (case 22 and 23). In the latter cases, the loss of asymptotic freedom is compensated through an interacting fixed point in the Yukawa and scalar couplings, which enabled a fixed point for the gauge coupling \cite{Litim:2014uca}. We stress that scalar fields and Yukawa interactions play a key role. Without them, triviality of any QED-like gauge theories cannot be avoided \cite{Bond:2016dvk,Bond:2017sem}.
Triviality also relates to the difficulty of defining elementary self-interacting scalar quantum fields in four dimensions \cite{Wilson:1973jj,Luscher:1987ek,Hasenfratz:1987eh}. It is interesting to notice that the quartic scalar couplings always take a unique physical fixed points as soon as the gauge and Yukawa coupling take weakly coupled fixed points. Hence, in theories with \eq{L} scalar fields can be viewed as elementary and triviality is evaded in all settings with asymptotic freedom and asymptotic safety.
In either case gauge fields play an important role, albeit for different reasons \cite{Litim:2014uca}. For gauge interactions with asymptotic freedom, the running of gauge couplings dictates the running for Yukawa and scalar couplings, and conditions for complete asymptotic freedom have been derived \cite{Coleman:1973sx} which ensure that gauge theories coupled to matter reach the free UV fixed point \cite{Chang:1974bv}. For theories with asymptotic safety, scalars are required to help generate a combined fixed point in the gauge, Yukawa, and quartic scalar couplings. This leads invariably to a ``reduction of couplings'' and enhanced predictivity over models with asymptotic freedom through a reduced UV critical surface.
\subsection{Veneziano limit and beyond}
Our findings, throughout, rely on the existence of exact small parameters $\epsilon_1\ll 1$ and $\epsilon_2\ll 1$ \eq{eps12} [or $\epsilon\ll1 $ see \eq{P}] in the Veneziano limit, which relate to the gauge one loop coefficients.
Consequently, an iterative solution of perturbative beta functions becomes exact and interacting fixed points arise as exact power series in the small parameters. More specifically, the leading non-trivial approximation which is NLO${}^\prime$ (Tab.~\ref{tNLO}) retains the gauge beta functions up to two loop, and the Yukawa and scalar beta functions up to one loop. The parametric smallness of the gauge one-loop coefficients allows an exact cancellation of one and two loop terms implying that interacting fixed points for the gauge couplings must be of the order of the one loop coefficient $\sim\epsilon$. The Yukawa nullclines at one loop imply that Yukawa couplings are necessarily proportional to the gauge couplings, and
the scalar nullcline impose that scalar couplings are proportional to the Yukawas (see Sect.~\ref{YSnull}); hence either of these come out $\sim\epsilon$. Higher order loop approximations $n$NLO${}^\prime$ starting with $n= 2$ then correspond to retaining $n+1$ loops in the gauge, and $n$ loops in the Yukawa and scalar beta functions respectively, see Tab.~\ref{tNLO}. Hence, solving the beta functions for interacting fixed points order-by-order in perturbation theory $(n\to n+1)$ we have that
\begin{equation}\label{nNLO}
\alpha_i^*=\alpha_i^*\Big|_{n{\rm NLO'}}
+{\cal O}(\epsilon^{n+1})
\end{equation}
for all couplings \eq{couplings}, \eq{w} and all fixed points, with corrections from the $(n+1)$NLO${}^\prime$ level being at least one power in $\epsilon$ smaller than those from the preceding level.
We conclude that the expressions for the interacting fixed points $\alpha_i^*|_{n{\rm NLO'}}$ are accurate polynomials in $\epsilon$ up to including terms of order $\epsilon^{n}$, for all $n$.
Beyond the Veneziano limit, the parametrically small control parameter $\epsilon$ is no longer available. Instead, $\epsilon$ will take finite, possibly large, values dictated by the (finite) field multiplicities. Still, for sufficiently large matter field multiplicities, $\epsilon$ remains sufficiently small and perturbativity remains in reach \cite{Bond:2017wut}. It is then conceivable that the fixed points found in the Veneziano limit persist even for finite $N$.\footnote{An example for a conformal window with asymptotic safety is given in \cite{Bond:2017tbw} for the model introduced in \cite{Litim:2014uca}.}
At finite $N$, however, we stress that the $n$NLO${}^\prime$ approximations and \eq{nNLO} are no longer exact order-by-order. It then becomes important to check numerical convergence of higher loop approximations, including non-perturbative resummations. In this context it would be particularly useful to know the radius of convergence of beta functions (in $\epsilon$) in the Veneziano limit. A finite radius of convergence has been established rigorously in certain large-$N_F$ limits of
gauge theories without Yukawa interactions \cite{PalanquesMestre:1983zy,Holdom:2010qs} which makes it conceivable that the radius of convergence might be finite here as well.\footnote{Results for resummed beta functions of large-$N$ gauge theories with Yukawa couplings are presently not available.} If so, this would offer additional indications for the existence of interacting fixed points beyond the Veneziano limit.
\subsection{Conformal symmetry and conformal windows}
By their very definition, the gauge-Yukawa theories investigated here are scale-invariant at (interacting) fixed points. Conditions under which scale invariance entails exact conformal invariance have been discussed by Polchinski \cite{Polchinski:1987dy} (see also \cite{Luty:2012ww}). Applied to the theories \eq{L} at weak coupling, it implies that exact conformal invariance is realised at all interacting fixed points discovered here.
It would then be interesting to find the full conformally invariant effective action beyond the classically marginal invariants retained in \eq{L}.
First steps into these directions have been reported in \cite{Buyukbese:2017ehm}.
Moreover, for a quantum theory to be compatible with unitarity, scaling dimension of (primary) scalar fields must be larger than unity. This is confirmed for all fixed points by using the results of Sect.~\ref{anomalous} for the anomalous dimensions of fields and composite scalar operators, together with the results for fixed points at NLO${}'$ accuracy (Tab.~\ref{tPara15} and~\ref{tPara69}). We conclude that the residual interactions are compatible with unitarity.
Away from the Veneziano limit, findings for the various interacting conformal fixed points persist once $\epsilon$ is finite. One may then think of keeping the parameters in the gauge sectors $(N_{\rm C}, N_{\rm c}$) fixed and finite while varying the matter field content $(N_{\rm F}, N_{\rm f}, N_\psi)$. Then, the domain of existence for each of the interacting fixed points (Tab.~\ref{tPara15},~\ref{tPara69}) turns into a ``conformal window'' as a function of the matter field multiplicities. The fixed point ceases to exist outside the conformal window. The conformal window for asymptotic safety with a simple $SU(N)$ gauge factor has been determined in \cite{Bond:2017tbw}. Boundaries of conformal windows can be estimated within perturbation theory though more accurate results invariably require non-perturbative tools.\footnote{See \cite{DelDebbio:2010zz} for lattice studies of conformal windows in QCD with fermionic matter (Banks-Zaks fixed points).}
\section{\bf Summary}\label{sum}
We have used perturbation theory and large-$N$ techniques for a
rigorous and
comprehensive investigation of weakly interacting fixed points
of gauge theories coupled to fermionic and scalar matter.
For concrete families of simple and semi-simple
gauge theories with action \eq{L} and following the classification of fixed points put forward in \cite{Bond:2016dvk,Bond:2017sem}, we have discovered a large variety of exact high- and low-energy fixed points (Tab.~\ref{tFPs},~\ref{tPara15},~\ref{tPara69}).
These include partially interacting ones (Tab.~\ref{tPara15}) where one gauge sector remains free, and
fully interacting ones (Tab.~\ref{tPara69}) where both gauge sectors are interacting.
We have determined the domains of existence for all of them
(Fig.~\ref{pFP23}--~\ref{pFP9}). Interestingly, we also find that the requirement of vacuum stability always singles out a unique viable fixed point in the scalar sector.
As a function of field multiplicities, the phase space of distinct quantum field theories (Fig.~\ref{pFPall}) includes models with asymptotic safety and asymptotic freedom, and effective theories without UV completion (Figs.~\ref{pAll_AF},~\ref{pAll_AS} and~\ref{pAll_EFT}). In the IR, theories display either strong coupling and confinement, or weakly coupled fixed points where the elementary gauge fields and fermions are unconfined and appear as massless particles. Many features are a consequence of the semi-simple nature and would not arise in simple (or ``direct products'' of simple) gauge theories. Highlights include massless semi-simple gauge-matter theories where one gauge sector can be both UV free and IR free owing to a fixed point in the other, Fig.~\ref{pSemiSimpleBZs} $c)$, and theories with inequivalent scaling limits in the IR. Semi-simple effects are particularly pronounced for asymptotically free theories where they enhance the diversity of different IR scaling regimes (Fig.~\ref{pAll_AF}).
Another central outcome of our study is the first explicit ``proof of existence'' for asymptotic safety in semi-simple quantum field theories with elementary gauge fields, scalars and fermions. It establishes the important result that asymptotic safety is not limited to simple gauge factors \cite{Litim:2014uca}, fully in line with general theorems and structural results \cite{Bond:2016dvk}. Our findings, together with their supersymmetric counterparts in \cite{Bond:2017suy}, make it conceivable that semi-simple theories display interacting UV fixed points even beyond the Veneziano limit, thus further paving the way for asymptotic safety beyond the Standard Model \cite{Bond:2017wut}.
The stability of the vacuum (Sect.~\ref{stab}) in all models studied here suggests
that the near-criticality of the Standard Model Higgs \cite{EliasMiro:2011aa,Buttazzo:2013uya} can very well expand into full criticality at an interacting UV fixed point \cite{Bond:2017wut}.
In addition, we have investigated phase diagrams for simple and semi-simple gauge theories
with and without Yukawa interactions, continuing an analysis initiated in \cite{Bond:2016dvk,Bond:2017sem}.
We find that transitions from the UV to the IR can proceed from free or interacting fixed points to confinement and strong coupling. We also find transitions from free to interacting (Figs.~\ref{pSemiSimpleBZs},~\ref{pSimpleCasesPD}, \ref{pPrimitivesSimple}, \ref{pPD_BZ_IR}, \ref{pPD_AF}, \ref{pPD_AF_pt2})
or from interacting to other interacting conformal fixed points (Fig.~\ref{pPD_AS}).
In the latter cases,
theories display a variety of exact
``IR sinks'', meaning free or interacting IR conformal fixed points which are fully attractive in all classically marginal interactions. Once more, many new features have come to light beyond those observed in simple gauge theories \cite{Bond:2016dvk,Bond:2017sem}.
Our study used minimal models with a low number of Yukawa and gauge couplings. Already at this basic level, an intriguing diversity of fixed points and scaling regimes has emerged, with many novel characteristics both at high and low energies.
We believe that these findings warrant more extensive studies in view of rigorous results \cite{Bond:2016dvk,Bond:2017suy}, extensions towards strong coupling \cite{Buyukbese:2017ehm}, and its exciting potential for physics beyond the Standard Model \cite{Bond:2017wut}.\\[3ex]
\centerline{\bf Acknowledgements}
${}$\\ Some of the results have been presented at the workshops {\it The Exact Renormalisation Group}, Trieste (Sept 2016), and {\it Understanding the LHC}, Bad Honnef (Feb 2017). This work is supported by the Science and Technology Research Council (STFC) under the Consolidated Grant [ST/G000573/1] and by an STFC studentship.
|
\section{Introduction}
The objective of many problems that can be modeled as graphs is finding a group
of vertices that together satisfy some property.
In this respect, one of the concepts that has been quite extensively studied is
the notion of a defensive alliance
\cite{kristiansen2004alliances,kristiansen2002introduction}, which is a set of
vertices such that for each element $v$ at least half of its neighbors are also
in the alliance.
The name ``defensive alliance'' stems from the intuition that the neighbors of
an element $v$ that are also in the alliance can help out in case $v$ is
attacked by its other neighbors.
Notions like this can be applied to finding groups of nations, companies or
individuals that depend on each other, but also to more abstract situations
like finding groups of websites that form
communities~\cite{DBLP:journals/computer/FlakeLGC02}.
Another possible application for defensive alliances are computer networks,
where a defensive alliance represents computers that can provide a certain
desired resource; any computer in an alliance can then, with the help of its
neighbors that are also in the alliance, allow access to this resource from all
of its neighbors simultaneously~\cite{DBLP:journals/combinatorics/HaynesHH03}.
Several variants of defensive alliances have also been studied.
The papers that originally proposed defensive alliances also propose related
notions like offensive and powerful alliances.
An offensive alliance is a set $S$ of vertices such that every \emph{neighbor}
of an element of $S$ has at least half of its neighbors in $S$, and a powerful
alliance is both a defensive and an offensive alliance.
Any of these alliances is called \emph{global} if it is at the same time a
dominating set.
Another variant is to consider alliances $S$ where, for each vertex $v \in S$,
the difference between the number of neighbors of $v$ in $S$ and the number of
other neighbors of $v$ is at most a given integer~\cite{ShafiqueD2003}.
For comprehensive overviews of different kinds of alliances in graphs, we refer
to the surveys~\cite{yero2013defensive, fernau2014survey}.
The \prob{Defensive Alliance}{} problem can be specified as follows:
Given a graph $G$ and an integer $k$, is there a defensive alliance $S$
in $G$ such that $1\leq \size{S} \leq k$?
It is known that this problem is
\complexityclass{NP}-complete~\cite{Jamieson2007,jamieson2009algorithmic}, and so is the
corresponding problem for global defensive alliances~\cite{CamiBDD06}.
However, if we restrict ourselves to trees, \prob{Defensive Alliance}{} becomes trivial and in fact
the corresponding problems for several non-trivial variants become solvable in
linear time~\cite{Jamieson2007}.
There has also been some work on the parameterized complexity of alliance
problems.
In particular, determining whether a defensive, offensive and powerful alliance
of a given (maximum) size exists is fixed-parameter tractable when
parameterized by the solution size~\cite{sofsem:FernauR07, Enciso2009}.
Also structural parameters have been considered to some extent.
Recently, \cite{dam:KiyomiO17} proved that these problems can be solved in
polynomial time if the clique-width of the instances is bounded by a constant.
The authors also provide an FPT algorithm when the parameter is the size of the
smallest vertex cover.
Moreover, \cite{Enciso2009} showed that the decision problems for defensive
alliances and global defensive alliances are fixed-parameter tractable when
parameterized by the combination of treewidth and maximum degree.
Despite these advances regarding, the question of whether or not \prob{Defensive Alliance}{}
parameterized by treewidth is fixed-parameter tractable has so far remained
open.
Treewidth~\cite{DBLP:journals/jct/RobertsonS84,DBLP:journals/actaC/Bodlaender93,DBLP:conf/sofsem/Bodlaender05}
is one of the most extensively studied structural parameters and indicates how
close a graph is to being a tree.
It is particularly attractive because many hard problems become tractable on
instances of bounded treewidth, and in several practical applications it has
been observed that the considered problem instances exhibit small
treewidth~\cite{DBLP:journals/actaC/Bodlaender93,iandc:Thorup98,dam:KornaiT92}.
Hence it would be very appealing to obtain an FPT algorithm for the \prob{Defensive Alliance}{}
problem using this parameter.
The main contribution of this paper is a parameterized complexity analysis of
\prob{Defensive Alliance}{} with treewidth as the parameter.
The question of whether or not this problem is fixed-parameter tractable when
parameterized by treewidth has so far been unresolved~\cite{dam:KiyomiO17}.
In the current chapter, we provide a negative answer to this question:
We show that the problem is hard for the class \complexityclass{W[1]}{}, which rules out
fixed-parameter tractable algorithms under commonly held complexity-theoretic
assumptions.
This result is rather surprising for two reasons:
First, the problem is tractable on trees~\cite{Jamieson2007} and quite often
problems that become easy on trees turn out to become easy on graphs of bounded
treewidth.%
\footnote{To be precise, \cite{Jamieson2007,ho2009rooted,ChangCHKLW12} show
that some variants of \prob{Defensive Alliance}{} are tractable on trees, since \prob{Defensive Alliance}{} on trees is
trivial.}
Second, this puts \prob{Defensive Alliance}{} among the very few ``subset problems'' that are
fixed-parameter tractable w.r.t.\ solution size but not w.r.t.\ treewidth.
Problems with this kind of behavior are rather rare, as observed
by Dom et~al.~\cite{DBLP:conf/iwpec/DomLSV08}.
We show \complexityclass{W[1]}-hardness of the problem by first reducing a problem known to be
\complexityclass{W[1]}-hard to a variant of \prob{Defensive Alliance}{}, where vertices can be forced to be in or out
of every solution, and pairs of vertices can be specified to indicate that
every solution must contain exactly one element of each such pair.
In order to prove the desired complexity result, we then successively reduce
this variant to the standard \prob{Defensive Alliance}{} problem.
At the same time, we show \complexityclass{W[1]}{}-hardness for the exact variants of these
problems, where we are interested in defensive alliances \emph{exactly} of a
certain size.
Note that a set may lose the property of being a defensive alliance by adding
or removing elements, so these are non-trivial problem variants.
Indeed, exact versions of alliance problems have also been mentioned as
interesting variants in~\cite{sofsem:FernauR07} because some algorithms that
work for the non-exact case stop to work for the exact case:
A graph has a defensive alliance of size at most $k$ if and only if it has a
\emph{connected} defensive alliance of size at most $k$ since every component
of a defensive alliance is itself a defensive alliance.
Algorithms that exploit this by looking only for connected solutions hence fail
for the exact versions.
(In fact, we will also study the complexity of other problem variants where
this connectedness property does not apply even in the non-exact case.)
This paper is organized as follows:
We first introduce our problems of interest and describe preliminary concepts
in Section~\ref{sec:background}.
In Section~\ref{sec:da-wone-treewidth} we then show that the \prob{Defensive Alliance}{} problem is
\complexityclass{W[1]}-hard when parameterized by treewidth.
Section~\ref{sec:conclusion} concludes the paper with a discussion.
The reductions in the current work are based on ideas used in the
paper~\cite{DBLP:conf/wg/BliemW15}, which analyzed the complexity of a problem
related to \prob{Defensive Alliance}{} called \prob{Secure Set}{}.
That paper has since been extended by a \complexityclass{W[1]}-hardness proof for the \prob{Secure Set}{}
problem parameterized by treewidth~\cite{corr:BliemW14}.
In the current paper, we take up the ideas behind this hardness proof and apply
them to the \prob{Defensive Alliance}{} problem.
Due to the different nature of these two problems, the reductions and proofs
for \prob{Secure Set}{} do not work directly for \prob{Defensive Alliance}{} but require substantial
modifications.
\section{Background}
\label{sec:background}
All graphs are undirected and simple unless stated otherwise.
We denote the set of vertices and edges of a graph $G$ by $\vertices{G}$ and $\edges{G}$,
respectively.
We denote an undirected edge between vertices $u$ and $v$ as $(u,v)$ or
equivalently $(v,u)$.
It will be clear from the context whether an edge $(u,v)$ is directed or
undirected.
Given a graph $G$, the \emph{open neighborhood} of a vertex $v \in \vertices{G}$,
denoted by $N_G(v)$, is the set of all vertices adjacent to $v$, and
$N_G[v] = N_G(v) \cup \{v\}$ is called the \emph{closed neighborhood} of $v$.
If it is clear from the context which graph is meant, we write $N(\cdot)$ and
$N[\cdot]$ instead of $N_G(\cdot)$ and $N_G[\cdot]$, respectively.
The intuition behind defensive alliances is the following:
If we consider a set $S$ of vertices as ``good'' vertices and all other
vertices as ``bad'' ones, then $S$ being a defensive alliance means that each element of $S$
has at least as many ``good'' neighbors as ``bad'' neighbors.
\begin{definition}
Given a graph $G$, a set $S \subseteq \vertices{G}$ is a \emph{defensive
alliance in $G$} if for each $v \in S$ it holds that
$\size{N[v] \cap S} \geq \size{N[v] \setminus S}$.
\end{definition}
We often write ``$S$ is a defensive alliance'' instead of ``$S$ is a defensive
alliance in $G$'' if it is clear from the context which graph is meant.
By definition, the empty set is a defensive alliance in any graph. Thus, in the
following decision problems we ask for a defensive alliances of size at
least~1.
\input{figures/da-example-figure}
For example, in Figure~\ref{fig:da-example}, the set $S = \{a,b\}$ is
a defensive alliance as $\size{N[v] \cap S} \geq \size{N[v] \setminus S}$
holds for each $v \in S$.
Note that, for instance, $\{a,d\}$ is no defensive alliance since $d$ is
attacked by three vertices but only has the neighbor $a$ to help defend itself.
Next we introduce several variants of \prob{Defensive Alliance}{} that we require in our proofs.
The problem \prob{Defensive Alliance\textsuperscript{F}}{} generalizes \prob{Defensive Alliance}{} by designating some ``forbidden''
vertices that may never be in any solution.
This variant can be formalized as follows:
\decisionProblem{\prob{Defensive Alliance\textsuperscript{F}}}{A graph $G$, an integer $k$ and a set $V_\square \subseteq \vertices{G}$}{%
Does there exist a set $S \subseteq \vertices{G} \setminus V_\square$ with $1 \leq \size{S} \leq k$ that is a defensive alliance?%
}
\prob{Defensive Alliance\textsuperscript{FN}}{} is a further generalization that, in addition, allows ``necessary'' vertices to be specified that must occur in every solution.
\decisionProblem{\prob{Defensive Alliance\textsuperscript{FN}}}{A graph $G$, an integer $k$, a set $V_\square \subseteq \vertices{G}$ and a set $V_\triangle \subseteq \vertices{G}$}{%
Does there exist a set $S \subseteq \vertices{G} \setminus V_\square$ with $V_\triangle \subseteq S$ and $1 \leq \size{S} \leq k$ that is a defensive alliance?%
}
Finally, we introduce the generalization \prob{Defensive Alliance\textsuperscript{FNC}}{}.
Here we may state pairs of ``complementary'' vertices where each solution must contain exactly one element of every such pair.
\decisionProblem{\prob{Defensive Alliance\textsuperscript{FNC}}}{A graph $G$, an integer $k$, a set $V_\square \subseteq \vertices{G}$, a set $V_\triangle \subseteq \vertices{G}$ and a set $C \subseteq \vertices{G}^2$}{%
Does there exist a set $S \subseteq \vertices{G} \setminus V_\square$ with $V_\triangle \subseteq S$ and $1 \leq \size{S} \leq k$ that is a defensive alliance and, for each pair $(a,b) \in C$, contains either $a$ or $b$?%
}
For our results on the parameter treewidth, we need a way to represent the
structure of a \prob{Defensive Alliance\textsuperscript{FNC}}{} instance by a graph that augments $G$ with the
information in $C$:
\begin{definition}
Let $I$ be a \prob{Defensive Alliance\textsuperscript{FNC}}\ instance, let $G$ be the graph in $I$ and let $C$ the set
of complementary vertex pairs in $I$.
By the \emph{primal graph} of $I$ we mean the undirected graph $G'$ with
$\vertices{G'} = \vertices{G}$ and
$\edges{G'} = \edges{G} \cup C$.
\end{definition}
When we speak of the treewidth of an instance of \prob{Defensive Alliance}{}, \prob{Defensive Alliance\textsuperscript{F}}{} or \prob{Defensive Alliance\textsuperscript{FN}}{},
we mean the treewidth of the graph in the instance.
For an instance of \prob{Defensive Alliance\textsuperscript{FNC}}{}, we mean the treewidth of the primal graph.
While the \prob{Defensive Alliance}{} problem asks for defensive alliances of size \emph{at most} $k$, we also consider the \prob{Exact Defensive Alliance}{} problem that concerns defensive alliances of size \emph{exactly} $k$.
Analogously, we also define exact versions of the three generalizations of
\prob{Defensive Alliance}{} presented above.
In this paper's figures, we often indicate necessary vertices by means of a triangular node shape, and forbidden vertices by means of either a square node shape or a superscript square in the node name.
If two vertices are complementary, we often express this in the figures by putting a $\neq$ sign between them.
\input{figures/da-example-figure2}
For example, in Figure~\ref{fig:da-example2}, the vertices $b$ and $c$ are complementary and occur in no solution together; $a$ and the ``anonymous'' vertex adjacent to $c$ are necessary and occur in every solution; $d^\square$ and the ``anonymous'' vertex adjacent to $e$ are forbidden and occur in no solution.
In this figure, the unique minimum non-empty defensive alliance satisfying the conditions of forbidden, necessary and complementary vertices consists of $a$, $b$ and the ``anonymous'' necessary vertex adjacent to $c$.
The following terminology will be helpful:
We often use the terms \emph{attackers} and \emph{defenders} of an element $v$ of a defensive alliance candidate $S$.
By these we mean the sets $N[v] \setminus S$ and $N[v] \cap S$, respectively.
To show that an element $v$ of a defensive alliance candidate $S$ is \emph{not} a counterexample
to $S$ being a solution, we sometimes employ the notion of a \emph{defense} of
$v$ w.r.t.\ $S$, which assigns to each attacker a dedicated defender:
If we are able to find an injective mapping $\mu: N[v] \setminus S \to N[v] \cap S$, then obviously $\size{N[v] \setminus S} \leq \size{N[v] \cap S}$, and we call $\mu$ a \emph{defense} of $v$ w.r.t.\ $S$.
Given such a defense $\mu$, we say that a defender $d$ \emph{repels} an attack on $v$ by an attacker $a$ whenever $\mu(a) = d$.
Consequentially, when we say that a set of defenders $D$ \emph{can repel} attacks on $v$ from a set of attackers $A$, we mean that there is a defense that assigns to each element of $A$ a dedicated defender in $D$.
To warm up, we make some easy observations that we will use in our proofs.
First, for every set $R$ consisting of a majority of neighbors of a vertex $v$,
whenever $v$ is in a defensive alliance, also some element of $R$ must be in
it:
\begin{observation}
\label{obs:one-in-s}
Let $S$ be a defensive alliance in a graph,
let $v \in S$ and
let $R \subseteq N(v)$.
If $\size{R} > \frac{1}{2} N[v]$,
then $S$ contains an element of $R$.
\end{observation}
\begin{proof}
Suppose that $\size{R} > \frac{1}{2} \size{N[v]}$ and
$S$ contains no element of $R$.
Since all elements of $R$ attack $v$,
$\size{N[v] \setminus S} > \frac{1}{2} \size{N[v]}$.
Hence
$2 \size{N[v] \setminus S} > \size{N[v]} =
\size{N[v] \cap S} + \size{N[v] \setminus S}$, and we obtain
the contradiction
$\size{N[v] \setminus S} > \size{N[v] \cap S}$.
\end{proof}
Next, if one half of the neighbors of an element $v$ of a defensive alliance
attacks $v$, then the other half of the neighbors must be in the defensive
alliance:
\begin{observation}
\label{obs:all-in-s}
Let $S$ be a defensive alliance in a graph,
let $v \in S$ and
let $N(v)$ be partitioned into two equal-sized sets $A,D$.
If $A \cap S = \emptyset$, then $D \subseteq S$.
\end{observation}
\begin{proof}
Since $N(v)$ is partitioned into $A$ and $D$ such that
$A \cap S = \emptyset$,
we get
$N(v) \cap S = D \cap S$.
If some element of $D$ is not in $S$, then
$D \cap S \subset D$ and
$A \subset N[v] \setminus S$.
By $\size{D} = \size{A}$, we get
$\size{D \cap S} + 2 \leq \size{N[v] \setminus S}$.
From
$\size{N[v] \cap S} = 1 + \size{N(v) \cap S} = 1 + \size{D \cap S}$
we now obtain the contradiction
$\size{N[v] \cap S} < \size{N[v] \setminus S}$.
\end{proof}
In particular, if half of the neighbors of $v$ are forbidden, then $v$ can only
be in a defensive alliance if all non-forbidden neighbors are also in the
defensive alliance.
Finally, we recapitulate some background from complexity theory.
In parameterized complexity
theory~\cite{downey1999parameterized,flum2006parameterized,b:Niedermeier06,Cygan15},
we study problems that consist not only of an input and a question, but also of
some parameter of the input that is represented as an integer.
A problem is in the class \complexityclass{FPT}{} (``fixed-parameter tractable'') if it can be
solved in time $f(k) \cdot n^c$, where $n$ is the input size, $k$ is the
parameter, $f$ is a computable function that only depends on $k$, and $c$ is a
constant that does not depend on $k$ or $n$.
We call such an algorithm an \emph{FPT algorithm}, and we call it
\emph{fixed-parameter linear} if $c=1$.
Similarly, a problem is in the class \complexityclass{XP}{} (``slice-wise polynomial'') if it
can be solved in time $f(k) \cdot n^{g(k)}$, where $f$ and $g$ are computable
functions.
Note that here the degree of the polynomial may depend on $k$, so such
algorithms are generally slower than FPT algorithms.
For the class \complexityclass{W[1]}{} it holds that $\complexityclass{FPT} \subseteq \complexityclass{W[1]} \subseteq \complexityclass{XP}$,
and it is commonly believed that the inclusions are proper, i.e.,
\complexityclass{W[1]}{}-hard problems do not admit FPT algorithms.
\complexityclass{W[1]}{}-hardness of a problem can be shown using FPT reductions,
which are reductions that run in FPT time and produce an equivalent instance
whose parameter is bounded by a function of the original parameter.
For problems whose input can be represented as a graph, one important parameter
is \emph{treewidth}, which is a structural parameter that, roughly speaking,
measures the ``tree-likeness'' of a graph. It is defined by means of tree
decompositions, originally introduced in~\cite{DBLP:journals/jct/RobertsonS84}.
The intuition behind tree decompositions is to obtain a tree from a
(potentially cyclic) graph by subsuming multiple vertices under one node and
thereby isolating the parts responsible for cyclicity.
\begin{definition}
\label{def:td}
A \emph{tree decomposition} of a graph $G$
is a pair $\mathcal{T} = (T,\chi)$ where $T$ is a (rooted) tree and
$\chi : \vertices{T} \to 2^{\vertices{G}}$ assigns to each node of $T$ a set of vertices of $G$
(called the node's \emph{bag}), such that the following conditions are met:
\begin{enumerate}
\item For every vertex $v \in \vertices{G}$, there is a node $t \in \vertices{T}$ such that $v
\in \chi(t)$.
\item For every edge $(u,v) \in \edges{G}$, there is a node $t \in \vertices{T}$ such that
$\{u,v\} \subseteq \chi(t)$.
\item For every $v \in \vertices{G}$, the subtree of $T$ induced by $\{t \in \vertices{T} \mid
v \in \chi(t)\}$ is connected.
\end{enumerate}
We call $\max_{t \in \vertices{T}} \lvert \chi(t) \rvert - 1$ the \emph{width} of $\mathcal{T}$.
The \emph{treewidth} of a graph is the minimum width over all its tree
decompositions.
\end{definition}
In general, constructing an optimal tree decomposition (i.e., a tree
decomposition with minimum width) is
intractable~\cite{Arnborg:1987:CFE:37170.37183}.
However, the problem is solvable in linear time on graphs of bounded treewidth
(specifically in time
$w^{\O(w^3)} \cdot n$, where $w$ is the
treewidth)~\cite{siamcomp:Bodlaender96} and
there are also heuristics that offer good performance in
practice~\cite{DBLP:conf/micai/DermakuGGMMS08,DBLP:journals/iandc/BodlaenderK10}.
In this paper we will consider so-called \emph{nice} tree decompositions:
\begin{definition}
\label{def:nice-td}
A tree decomposition $\mathcal{T} = (T,\chi)$ is \emph{nice} if each node $t
\in \vertices{T}$ is of one of the following types:
\begin{enumerate}
\item Leaf node: The node $t$ has no child nodes.
\item Introduce node: The node $t$ has exactly one child node $t'$ such that
$\chi(t) \setminus \chi(t')$ consists of exactly one element.
\item Forget node: The node $t$ has exactly one child node $t'$ such that
$\chi(t') \setminus \chi(t)$ consists of exactly one element.
\item Join node: The node $t$ has exactly two child nodes $t_1$ and $t_2$ with
$\chi(t) = \chi(t_1) = \chi(t_2)$.
\end{enumerate}
Additionally, the bags of the root and the leaves of $T$ are empty.
\end{definition}
A tree decomposition of width $w$ for a graph with $n$ vertices can be
transformed into a nice one of width $w$ with $\O(wn)$ nodes in fixed-parameter
linear time~\cite{kloks1994treewidth}.
\input{figures/td-example-figure}
For any tree decomposition $\mathcal{T}$ and an element $v$ of some bag in $\mathcal{T}$,
we use the notation $t_v^\mathcal{T}$ to denote the unique ``topmost node'' whose bag
contains $v$ (i.e., $t_v^\mathcal{T}$ does not have a parent whose bag contains $v$).
Figure~\ref{fig:td-example} depicts a graph and a nice tree decomposition,
where we also illustrate the $t^\mathcal{T}_v$ notation.
When we speak of the treewidth of an instance of \prob{Defensive Alliance}{}, \prob{Defensive Alliance\textsuperscript{F}}{}, \prob{Defensive Alliance\textsuperscript{FN}}{}, \prob{Exact Defensive Alliance}{}, \prob{Exact Defensive Alliance\textsuperscript{F}}{} or
\prob{Exact Defensive Alliance\textsuperscript{FN}}{}, we mean the treewidth of the graph in the instance.
For an instance of \prob{Defensive Alliance\textsuperscript{FNC}}{} or \prob{Exact Defensive Alliance\textsuperscript{FNC}}{}, we mean the treewidth of the primal graph.
\section{Hardness of Defensive Alliance Parameterized by Treewidth}
\label{sec:da-wone-treewidth}
In this section, we prove the following theorem:
\begin{theorem}
\label{thm:da-variants-wone}
The following problems are all $\complexityclass{W[1]}$-hard when parameterized by treewidth:
\prob{Defensive Alliance}{},
\prob{Exact Defensive Alliance}{},
\prob{Defensive Alliance\textsuperscript{F}}{},
\prob{Exact Defensive Alliance\textsuperscript{F}}{},
\prob{Defensive Alliance\textsuperscript{FN}}{},
\prob{Exact Defensive Alliance\textsuperscript{FN}}{},
\prob{Defensive Alliance\textsuperscript{FNC}}{}, and
\prob{Exact Defensive Alliance\textsuperscript{FNC}}{}.
\end{theorem}
We prove hardness by providing a chain of FPT reductions from
a \complexityclass{W[1]}-hard problem to the problems under consideration.
Under the widely held assumption that $\complexityclass{FPT} \neq \complexityclass{W[1]}$, this rules out
fixed-parameter tractable algorithms for these problems.
\subsection{Hardness of Defensive Alliance with Forbidden, Necessary and Complementary Vertices}
To show \complexityclass{W[1]}-hardness of \prob{Defensive Alliance\textsuperscript{FNC}}{}, we reduce from the following
problem~\cite{DBLP:journals/dam/AsahiroMO11}, which is known to be
\complexityclass{W[1]}-hard~\cite{DBLP:journals/corr/abs-1107-1177} parameterized by the
treewidth of the graph:
\decisionProblem{\prob{Minimum Maximum Outdegree}}{%
A graph $G$,
an edge weighting $w: \edges{G} \to \mathbb{N}^+$ given in unary
and a positive integer $r$}{%
Is there an orientation of the edges of $G$ such that, for each $v \in
\vertices{G}$, the sum of the weights of outgoing edges from $v$ is at most
$r$?%
}
\begin{lemma}
\prob{Defensive Alliance\textsuperscript{FNC}}{} and \prob{Exact Defensive Alliance\textsuperscript{FNC}}{}, both parameterized by the treewidth of the primal
graph, are \complexityclass{W[1]}-hard.
\end{lemma}
\begin{proof}
Let an instance of \prob{Minimum Maximum Outdegree}\ be given by a graph $G$, an edge weighting $w:
\edges{G} \to \mathbb{N}^+$ in unary and a positive integer $r$.
From this we construct an instance of both \prob{Defensive Alliance\textsuperscript{FNC}}{} and \prob{Exact Defensive Alliance\textsuperscript{FNC}}{}.
An example is given in Figure~\ref{fig:minmaxoutdegree-reduction-example}.
For each $v \in \vertices{G}$, we define the set of new vertices
$H_v = \{h^v_1, \dots, h^v_{2r - 1}\}$, and for each $(u,v) \in \edges{G}$, we
define the sets of new vertices
$V_{uv} = \{u^v_1, \dots, u^v_{w(u,v)}\}$,
$V_{uv}^\square = \{u^{v\square}_1, \dots, u^{v\square}_{w(u,v)}\}$,
$V_{vu} = \{v^u_1, \dots, v^u_{w(u,v)}\}$ and
$V_{vu}^\square = \{v^{u\square}_1, \dots, v^{u\square}_{w(u,v)}\}$.
We now define the graph $G'$ with
\begin{align*}
\vertices{G'} ={}& \vertices{G} \cup \bigcup_{v \in \vertices{G}} H_v
\cup \bigcup_{(u,v) \in \edges{G}}
(V_{uv} \cup V_{uv}^\square \cup V_{vu} \cup V_{vu}^\square),\\
\edges{G'} ={}& \{(v,h) \mid v \in \vertices{G},\; h \in H_v\}\\
{}\cup{}&\{(u,x) \mid (u,v) \in \edges{G},\;
x \in V_{uv} \cup V_{uv}^\square\}\\
{}\cup{}& \{(x,v) \mid (u,v) \in \edges{G},\;
x \in V_{vu} \cup V_{vu}^\square\}.
\end{align*}
We also define the set of complementary vertex pairs
$C = \{(u^v_i,v^u_i) \mid (u,v) \in \edges{G},\; 1 \leq i \leq w(u,v)\} \cup
\{(v^u_i,u^v_{i+1}) \mid (u,v) \in \edges{G},\; 1 \leq i < w(u,v)\}$.
Finally, we define the set of necessary vertices
$V_\triangle = \vertices{G} \cup \bigcup_{v \in \vertices{G}} H_v$,
the set of forbidden vertices
$V_\square = \bigcup_{(u,v) \in \edges{G}} (V_{uv}^\square \cup
V_{vu}^\square)$
and
$k = \size{V_\triangle} + \sum_{(u,v) \in \edges{G}} w(u,v)$.
We use $I$ to denote $(G',k,C,V_\triangle,V_\square)$, which is an instance of
\prob{Defensive Alliance\textsuperscript{FNC}}\ and also of \prob{Exact Defensive Alliance\textsuperscript{FNC}}.
\input{figures/minmaxoutdegree-reduction-example}
Clearly $I$ can be computed in polynomial time.
We now show that the treewidth of the primal graph of $I$ depends only on the treewidth of $G$.
We do so by modifying an optimal tree decomposition $\mathcal{T}$ of $G$ as follows:
\begin{enumerate}
\item For each $(u,v) \in \edges{G}$, we take an arbitrary node whose bag $B$ contains
both $u$ and $v$
and add to its children
a chain of nodes $N_1, \dots, N_{w(u,v)-1}$ such that the
bag of $N_i$ is $B \cup \{u^v_i, u^v_{i+1}, v^u_i, v^u_{i+1}\}$.
\item For each $(u,v) \in \edges{G}$, we take an arbitrary node whose bag $B$
contains $u$ and add to its children a chain of nodes $N_1, \dots,
N_{w(u,v)}$ such that the bag of $N_i$ is $B \cup \{u^{v\square}_i\}$.
\item For each $(u,v) \in \edges{G}$, we take an arbitrary node whose bag $B$
contains $v$ and add to its children a chain of nodes $N_1, \dots,
N_{w(u,v)}$ such that the bag of $N_i$ is $B \cup \{v^{u\square}_i\}$.
\item For each $v \in \vertices{G}$, we take an arbitrary node whose bag $B$ contains $v$
and add to its children
a chain of nodes
$N_1, \dots, N_{r-1}$
such that the bag of $N_i$ is
$B \cup \{h^v_i\}$.
\end{enumerate}
It is easy to verify that the result is a valid tree decomposition of the
primal graph of $I$ and its width is at most the treewidth of $G$ plus four.
It remains to show that our reduction is correct.
Obviously $I$ is a positive instance of \prob{Defensive Alliance\textsuperscript{FNC}}\ iff it is a positive instance
of \prob{Exact Defensive Alliance\textsuperscript{FNC}}\ because the forbidden, necessary and complementary vertices make
sure that every solution of the \prob{Defensive Alliance\textsuperscript{FNC}}\ instance $I$ has exactly $k$ elements.
Hence we only consider \prob{Defensive Alliance\textsuperscript{FNC}}.
The intention is that for each orientation of $G$ we have a solution
candidate $S$ in $I$ such that
an edge orientation from $u$ to $v$ entails
$V_{vu} \subseteq S$ and
$V_{uv} \cap S = \emptyset$,
and the other orientation entails $V_{uv} \subseteq S$ and $V_{vu} \cap S = \emptyset$.
For each vertex $v \in \vertices{G}$ and every incident edge $(v,u) \in
\edges{G}$ regardless of its orientation, the vertex $v$ is attacked by the
forbidden vertices $V_{vu}^\square$.
So every vertex $v \in \vertices{G}$ has as least as many attackers as the sum
of the weights of all incident edges.
If in the orientation of $G$ all edges incident to $v$ are incoming edges, then
each attack on $v$ from $V_{vu}^\square$ can be repelled by $V_{vu}$, since
$V_{vu} \subseteq S$.
Due to the fact that the helper vertices $H_v$ consist of exactly $2r-1$
elements, $v$ can afford to have outgoing edges of total weight at most $r$.
We claim that $(G,w,r)$ is a positive instance of \prob{Minimum Maximum Outdegree}{} iff
$I$ is a positive instance of \prob{Defensive Alliance\textsuperscript{FNC}}{}.
\medskip
\noindent
\emph{``Only if'' direction.}
Let $D$ be the directed graph given by an orientation of the edges of $G$ such that for each vertex the sum of weights of outgoing edges is at most $r$.
The set
$S = V_\triangle \cup \{v^u_1, \dots, v^u_{w(u,v)} \mid (u,v) \in \edges{D}\}$
is a defensive alliance in $G'$:
Let $x$ be an arbitrary element of $S$.
If $x$ is an element of a set $H_v$ or $V_{uv}$, then the only neighbor of $x$
in $G'$ is a necessary vertex, so $x$ can trivially defend itself; so suppose
$x \in \vertices{G}$.
Let the sum of the weights of outgoing and incoming edges be denoted by
$w^x_{\mathrm{out}}$ and $w^x_{\mathrm{in}}$, respectively.
The neighbors of $x$ that are also in $S$ consist of the elements of $H_x$ and
all elements of sets $V_{xv}$ such that $(v,x) \in \edges{D}$.
Hence, including itself, $x$ has $2r + w^x_{\mathrm{in}}$ defenders in $G'$.
The attackers of $x$ consist of all elements of sets $V_{xv}$ such that $(x,v)
\in \edges{D}$ (in total $w^x_{\mathrm{out}}$) and all elements of sets
$V_{xv}^\square$ such that either $(v,x) \in \edges{D}$ or $(x,v) \in
\edges{D}$ (in total $w^x_{\mathrm{in}} + w^x_{\mathrm{out}}$).
Hence $x$ has $w^x_{\mathrm{in}} + 2 w^x_{\mathrm{out}}$ attackers in $G'$.
This shows that $x$ has at least as many defenders as attackers, as by
assumption $w^x_{\mathrm{out}} \leq r$.
Finally, it is easy to verify that
$\size{S} = k$,
$V_\square \cap S = \emptyset$,
$V_\triangle \subseteq S$,
and exactly one element of each pair of complementary vertices is in $S$.
\medskip
\noindent
\emph{``If'' direction.}
Let $S$ be a solution of $I$.
For every $(u,v) \in \edges{G}$, either $V_{uv} \subseteq S$ or $V_{vu} \subseteq S$ due to the complementary vertex pairs.
We define a directed graph $D$ by $\vertices{D} = \vertices{G}$ and
$\edges{D} = \{(u,v) \mid V_{vu} \subseteq S\} \cup \{(v,u) \mid V_{uv} \subseteq S\}$.
Suppose there is a vertex $x$ in $D$ whose sum of weights of outgoing edges is greater than $r$.
Clearly $x \in S$.
Let the sum of the weights of outgoing and incoming edges be denoted by
$w^x_{\mathrm{out}}$ and $w^x_{\mathrm{in}}$, respectively.
The defenders of $x$ in $G'$ beside itself consist of the elements of $H_x$ and
of $w^x_{\mathrm{in}}$ neighbors due to incoming edges in $D$.
These are in total $2r + w^x_{\mathrm{in}}$ defenders.
The attackers of $x$ in $G'$ consist of $2w^x_{\mathrm{out}}$ elements (of the
form $x^v_i$ as well as $x^{v\square}_i$) due to outgoing edges in $D$ and
$w^x_{\mathrm{in}}$ elements (of the form $x^{v\square}_i$) due to incoming
edges.
These are in total $2w^x_{\mathrm{out}} + w^x_{\mathrm{in}}$ attackers.
But then $x$ has more attackers than defenders, as by assumption
$w^x_{\mathrm{out}} > r$.
\end{proof}
\subsection{Hardness of Defensive Alliance with Forbidden and Necessary
Vertices}
Next we present a transformation $\ensuremath{\tau^\text{FNC}}$ that eliminates complementary vertex
pairs by turning a \prob{Defensive Alliance\textsuperscript{FNC}}{} instance into an equivalent \prob{Defensive Alliance\textsuperscript{FN}}{} instance.
Along with $\ensuremath{\tau^\text{FNC}}$, we define a function $\sFNC{I}$, for each \prob{Defensive Alliance\textsuperscript{FNC}}{} instance
$I$,
such that the solutions of $I$ are in a one-to-one correspondence with those of
$\ensuremath{\tau^\text{FNC}}(I)$ in such a way that any two solutions of $I$ have the same size iff
the corresponding solutions of $\ensuremath{\tau^\text{FNC}}(I)$ have the same size.
We use these functions to obtain a polynomial-time reduction from \prob{Defensive Alliance\textsuperscript{FNC}}{} to
\prob{Defensive Alliance\textsuperscript{FN}}{} as well as from \prob{Exact Defensive Alliance\textsuperscript{FNC}}{} to \prob{Exact Defensive Alliance\textsuperscript{FN}}{}.
Before we formally define our reduction, we briefly describe the intuition
behind the used gadgets.
The gadget in Figure~\ref{fig:complementary-reduction-gadget1} adds neighbors
$a_1,\dots,a_n,a_1^\square,\dots,a_n^\square$ to every vertex $a$, which are so
many that $a$ can only be in a solution if some of the new neighbors are also
in the solution.
The new vertices are structured in such a way that every solution must in fact
either contain all of $a,a_1,\dots,a_n$ or none of them.
Next, the gadget in Figure~\ref{fig:complementary-reduction-gadget2} is added
for every complementary pair $(a,b)$.
This gadget is constructed in such a way that every solution must either
contain all of $a_n,a^{ab},a^{ab}_1,\dots,a^{ab}_{n^2+n}$ or none of them, and
the same holds for $b_n,b^{ab},b^{ab}_1,\dots,b^{ab}_{n^2+n}$.
By making the vertex $\triangle^{ab}$ necessary, every solution must contain
one of these two sets.
At the same time, the bound on the solution size makes sure that we cannot
afford to take both sets for any complementary pair.
\begin{definition}
We define a function \ensuremath{\tau^\text{FNC}}{}, which assigns a \prob{Defensive Alliance\textsuperscript{FN}}{} instance to each \prob{Defensive Alliance\textsuperscript{FNC}}{}
instance
$I = (G,k,V_\square,V_\triangle,C)$.
For this, we use
$n$ to denote $\size{\vertices{G}}$
and first define a function
\[\sFNC{I} : x \mapsto x \cdot (n+1) + \size{C} \cdot (n^2+n+2).\]
For each $v \in \vertices{G}$, we introduce the following sets of new vertices.
\begin{align*}
Y_v^{\Circle} &{}=
\{v_1, \dots, v_n\}
&
Y_v^\square &{}=
\{v^\square_1, \dots, v^\square_n\}
\end{align*}
Next, for each $(a,b) \in C$, we introduce new vertices
$a^{ab}$, $b^{ab}$ and $\triangle^{ab}$ as well as,
for any $x \in \{a,b\}$, the following sets of new vertices.
\begin{align*}
Z^{ab}_{x{\Circle}} &{}=
\{x^{ab}_1, \dots, x^{ab}_{n^2+n}\}
&
Z^{ab}_{x\square} &{}=
\{x^{ab\square}_1, \dots, x^{ab\square}_{n^2+n}\}
\end{align*}
We use the notation
$u \oplus v$
to denote the set of edges
$\{(u,v),$ $(u,u^\square),$ $(v,v^\square),$ $(u,v^\square),$ $(v,u^\square)\}$.
\input{figures/complementary-reduction-gadget1}
\input{figures/complementary-reduction-gadget2}
Now we define
the \prob{Defensive Alliance\textsuperscript{FN}}{} instance
$\ensuremath{\tau^\text{FNC}}(I) = (G',k',V_\square',V_\triangle')$, where
$k' = \sFNC{I}(k)$,
$V_\square' = V_\square \cup
\bigcup_{v \in \vertices{G}} Y_v^\square \cup
\bigcup_{(a,b) \in C} (Z^{ab}_{a\square} \cup Z^{ab}_{b\square})$,
$V_\triangle' = V_\triangle \cup \bigcup_{(a,b) \in C} \{\triangle^{ab}\}$
and $G'$ is the graph defined by
\begin{align*}
\vertices{G'} = \vertices{G} &{}\cup
\bigcup_{v \in \vertices{G}} (Y_v^{\Circle} \cup Y_v^\square) \cup{}\\
&{}\cup \bigcup_{(a,b) \in C} \big(\{\triangle^{ab}, a^{ab}, b^{ab}\} \cup
Z^{ab}_{a{\Circle}} \cup Z^{ab}_{b{\Circle}} \cup Z^{ab}_{a\square} \cup
Z^{ab}_{b\square}\big),
\end{align*}
\begin{align*}
\edges{G'} = \edges{G} &{}\cup
\bigcup_{v \in \vertices{G}}
\big(
(\{v\} \times Y_v^{\Circle})
\cup
(\{v\} \times Y_v^\square)
\cup
\bigcup_{1\leq i < n} v_i \oplus v_{i+1}
\big)
\cup{}\\&{}\cup
\bigcup_{(a,b) \in C}
\bigcup_{x \in \{a,b\}} \big(
\{(\triangle^{ab}, x^{ab})\}
\cup
(\{x^{ab}\} \times Z^{ab}_{x{\Circle}})
\cup{}\\
&\phantom{{}\cup \bigcup_{(a,b) \in C}\bigcup_{x \in \{a,b\}} \big(}
{}\cup
x_n \oplus x_1^{ab}
\cup
\bigcup_{1 \leq i < n^2+n}
x_i^{ab} \oplus x_{i+1}^{ab}
\big).
\end{align*}
We illustrate our construction in
Figures~\ref{fig:complementary-reduction-gadget1} and
\ref{fig:complementary-reduction-gadget2}.
\label{def:dafnc-to-dafn}
\end{definition}
\begin{lemma}
Let $I = (G,k,V_\square,V_\triangle,C)$ be a \prob{Defensive Alliance\textsuperscript{FNC}}{} instance,
let $A$ be the set of solutions of $I$ and let $B$ be the set of solutions of
the \prob{Defensive Alliance\textsuperscript{FN}}{} instance $\ensuremath{\tau^\text{FNC}}(I)$.
There is a bijection $f: A \to B$ such that
$\size{f(S)} = \sFNC{I}(\size{S})$ holds for every $S \in A$.
\label{lem:dafnc-to-dafn-correct}
\end{lemma}
\begin{proof}
We use the same auxiliary notation as in Definition~\ref{def:dafnc-to-dafn}
and we define $f$ as
$S \mapsto S \cup
\bigcup_{v \in S} Y_v^{\Circle}
\cup
\bigcup_{(a,b) \in C,\, x \in S \cap \{a,b\}} (\{\triangle^{ab}, x^{ab}\} \cup
Z^{ab}_{x{\Circle}})$.
For every $S \in A$,
we thus obtain
$\size{f(S)} = \sFNC{I}(\size{S})$,
and we first show that indeed $f(S) \in B$.
Let $S \in A$ and let $S'$ denote $f(S)$.
Obviously $S'$ satisfies
$V_\square' \cap S' = \emptyset$ and
$V_\triangle' \subseteq S'$.
To see that $S'$ is a defensive alliance in $G'$,
let $x$ be an arbitrary element of $S'$.
If $x \notin S$, then $x$ clearly has as least as many neighbors in $S'$ as
neighbors not in $S'$ by construction of $f$, so suppose $x \in S$.
There is a defense
$\mu: N_G[x] \setminus S \to N_G[x] \cap S$
since $S$ is a defensive alliance in $G$.
We use this to construct a defense
$\mu': N_{G'}[x] \setminus S' \to N_{G'}[x] \cap S'$.
For any attacker $v$ of $x$ in $G'$, we distinguish two cases.
\begin{itemize}
\item If $v$ is some $x_i^\square \in Y_x^\square$ for some $x \in
\vertices{G}$, we set $\mu'(v) = x_i$.
This element is in $N_{G'}[x]$ by construction.
\item Otherwise $v$ is in $N_G[x] \setminus S$ (by our construction of $S'$).
Since the codomain of $\mu$ is a subset of the codomain of $\mu'$, we may set $\mu'(v) = \mu(v)$.
\end{itemize}
Since $\mu'$ is injective, each attack on $x$ in $G'$ can be repelled by $S'$.
Hence $S'$ is a defensive alliance in $G'$.
Clearly $f$ is injective.
It remains to show that $f$ is surjective.
Let $S'$ be a solution of $\ensuremath{\tau^\text{FNC}}(I)$.
First we make the following observations for each $v \in \vertices{G}$:
\begin{itemize}
\item
If $v \in S'$, then $Y_v^{\Circle} \cap S' \neq \emptyset$
due to Observation~\ref{obs:one-in-s}, since
$Y_v^{\Circle} \cup Y_v^\square$ contains a majority of neighbors of $v$, and
the vertices in $Y_v^\square$ are forbidden.
\item
For each $v^{ab} \in S'$, where $(a,b) \in C$ such that $v=a$ or $v=b$, it
holds that $Z^{ab}_{v{\Circle}} \cap S' \neq \emptyset$ again due to
Observation~\ref{obs:one-in-s}.
\item
If $S'$ contains an element of $Y_v^{\Circle}$, then
$\{v\} \cup Y_v^{\Circle} \cup \bigcup_{(v,z) \in C} Z^{vz}_{v{\Circle}} \cup
\bigcup_{(z,v) \in C} Z^{zv}_{v{\Circle}} \subseteq S'$
by repeated applications of Observation~\ref{obs:all-in-s}.
To see this, note in particular that $N(v_n)$ can be partitioned into the two
equal-sized sets
$\{v,v_{n-1}\} \cup \{v^{vz}_1 \mid (v,z) \in C\} \cup \{v^{zv}_1 \mid (z,v)
\in C\}$
and
$\{v_{n-1}^\square, v_n^\square\} \cup \{v^{vz\square}_1 \mid (v,z) \in C\}
\cup \{v^{zv\square}_1 \mid (z,v) \in C\}$,
and all vertices in the latter set are forbidden.
\item
If $S'$ contains an element of $Z^{ab}_{v{\Circle}}$, where $(a,b) \in C$ such
that $v=a$ or $v=b$, then $\{v^{ab}\} \cup Y_v^{\Circle} \cup \bigcup_{(v,z)
\in C} Z^{vz}_{v{\Circle}} \cup \bigcup_{(z,v) \in C} Z^{zv}_{v{\Circle}}
\subseteq S'$ for similar reasons.
\end{itemize}
It follows that for each $v \in \vertices{G}$, $S'$ contains either all or none
of
$\{v\} \cup Y_v^{\Circle} \cup \bigcup_{(v,z) \in C} \big(\{v^{vz}\} \cup
Z^{vz}_{v{\Circle}}\big) \cup \bigcup_{(z,v) \in C} \big(\{v^{zv}\} \cup
Z^{zv}_{v{\Circle}}\big)$.
For every $(a,b) \in C$,
$S'$ contains $a^{ab}$ or $b^{ab}$,
since $\triangle^{ab} \in S'$,
whose neighbors are $a^{ab}$ and $b^{ab}$.
It follows that $\size{S'} > \size{C} \cdot (n^2+n+2)$ even if $S'$ contains
only one of each $(a,b) \in C$.
If, for some $(a,b) \in C$, $S'$ contained both $a$ and $b$, we could derive a
contradiction to $\size{S'} \leq \sFNC{I}(k) = k \cdot (n+1) + \size{C} \cdot
(n^2+n+2)$ because then $\size{S'} > (\size{C}+1) \cdot (n^2+n+2) >
\sFNC{I}(k)$.
So $S'$ contains either $a$ or $b$ for any $(a,b) \in C$.
We construct $S = S' \cap \vertices{G}$ and observe that
$S' = f(S)$,
$V_\triangle \subseteq S$, $V_\square \cap S = \emptyset$, and
$\size{S \cap \{a,b\}} = 1$ for each $(a,b) \in C$.
It remains to show that $S$ is a defensive alliance in $G$.
Let $x$ be an arbitrary element of $S$.
We observe that $N_{G'}[x] \cap S' = (N_G[x] \cap S) \cup Y_x^{\Circle}$ and
similarly
$N_{G'}[x] \setminus S' = (N_G[x] \setminus S) \cup
Y_x^\square$.
Since the cardinality of each set $Y_x^{\Circle}$ is equal to the cardinality of
$Y_x^\square$,
this implies
$\size{N_{G'}[x] \cap S'} - \size{N_G[x] \cap S} = \size{N_{G'}[x] \setminus
S'} - \size{N_G[x] \setminus S}$.
Since $S'$ is a defensive alliance in $G'$ and $x \in S'$,
it holds that $\size{N_{G'}[x] \cap S'} \geq \size{N_{G'}[x] \setminus S'}$.
We conclude
that
$\size{N_G[x] \cap S} \geq \size{N_G[x] \setminus S}$.
Hence $S$ is a defensive alliance in $G$.
\end{proof}
To obtain the hardness result for \prob{Defensive Alliance\textsuperscript{FN}}{} parameterized by treewidth, it
remains to show that the reduction specified by $\ensuremath{\tau^\text{FNC}}$ preserves bounded
treewidth.
\begin{lemma}
\prob{Defensive Alliance\textsuperscript{FN}}{}, parameterized by the treewidth of the graph, is \complexityclass{W[1]}{}-hard.
\label{lem:dafn-wone-hard}
\end{lemma}
\begin{proof}
Let $I$ be a \prob{Defensive Alliance\textsuperscript{FNC}}{} instance whose primal graph we denote by $G$.
We obtain an equivalent \prob{Defensive Alliance\textsuperscript{FN}}{} instance $\ensuremath{\tau^\text{FNC}}(I)$, whose graph we denote by
$G'$.
This reduction is correct, as shown in Lemma~\ref{lem:dafnc-to-dafn-correct}.
It remains to show that the treewidth of $G'$ is bounded by a function of the treewidth of $G$.
Let $\mathcal{T}$ be an optimal nice tree decomposition of $G$.
We build a tree decomposition $\mathcal{T}'$ of $G'$ by modifying a copy of $\mathcal{T}$
in the following way:
For each vertex $v \in \vertices{G}$, we add $v_n$ and $v_n^\square$ to every
bag containing $v$.
Then we pick an arbitrary node $t$ in $\mathcal{T}$ whose bag contains $v$, and we
add new children $N_1,\dots,N_{n-1}$ to $t$ such that the bag of $N_i$ is
$\{v,v_i,v_i^\square,v_{i+1},v_{i+1}^\square\}$.
Next, for every pair $(a,b)$ of complementary vertices, we pick an arbitrary
node $t$ in $\mathcal{T}$ whose bag $B$ contains both $a_n$ and $b_n$, and we add a
chain of nodes $N_1, \dots, N_{2n^2+2n-1}$ between $t$ and its parent such
that,
for $1 \leq i < n^2+n$, the bag of $N_i$ is $B \cup \{a^{ab}, a_i^{ab},
a_i^{ab\square}, a_{i+1}^{ab}, a_{i+1}^{ab\square}\}$,
the bag of $N_{n^2+n}$ is
$B \cup \{a^{ab}, b^{ab}, \triangle^{ab}\}$, and the bag of $N_{n^2+n+i}$ is $B
\cup \{b^{ab}, b_{n^2+n+1-i}^{ab}, b_{n^2+n+1-i}^{ab\square}, b_{n^2+n-i}^{ab},
b_{n^2+n-i}^{ab\square}\}$.
It is easy to verify that $\mathcal{T}'$ is a valid tree decomposition of $G'$.
Furthermore, the width of $\mathcal{T}'$ is at most three times the width of $\mathcal{T}$
plus five.
\end{proof}
The instances of \prob{Defensive Alliance\textsuperscript{FNC}}{} are identical to the instances of the exact variant,
so $\ensuremath{\tau^\text{FNC}}$ is also applicable to the exact case.
In fact it turns out that this gives us also a reduction from \prob{Exact Defensive Alliance\textsuperscript{FNC}}{} to
\prob{Exact Defensive Alliance\textsuperscript{FN}}{}.
\begin{lemma}
\prob{Exact Defensive Alliance\textsuperscript{FN}}{}, parameterized by the treewidth of the graph, is \complexityclass{W[1]}{}-hard.
\label{lem:edafn-wone-hard}
\end{lemma}
\begin{proof}
Let $I$ and $I' = \ensuremath{\tau^\text{FNC}}(I)$ be our \prob{Exact Defensive Alliance\textsuperscript{FNC}}{} and \prob{Exact Defensive Alliance\textsuperscript{FN}}{} instances,
respectively, and let
$k$ and $k'$ denote their respective solution sizes.
By Lemma~\ref{lem:dafnc-to-dafn-correct},
there is a bijection $f$ between the
solutions of $I$ and the solutions of $I'$
such that,
for every solution $S$ of $I$, $f(S)$ has $\sFNC{I}(k) = k'$ elements,
and for every solution $S'$ of $I'$,
$f^{-1}(S')$ has $k$ elements since $\sFNC{I}$ is invertible.
We can derive the bound on the treewidth of $I'$ as in the proof of
Lemma~\ref{lem:dafn-wone-hard}.
\end{proof}
\subsection{Hardness of Defensive Alliance with Forbidden Vertices}
Now we present a transformation $\ensuremath{\tau^\text{FN}}$ that eliminates necessary vertices.
Our transformation not only operates on a problem instance, but also requires
an ordering $\preceq$ of the non-forbidden vertices of the graph.
Our reductions based on $\ensuremath{\tau^\text{FN}}$ will be correct for every ordering ${\preceq}$,
but in order to keep the treewidth of the resulting problem instance bounded by
the treewidth of the original problem instance, we must choose a suitable
ordering.
We will describe this in detail later; for now we can consider ${\preceq}$ to
be an arbitrary ordering of the non-forbidden vertices.
\input{figures/necessary-reduction-example-figure}
Before formally defining the transformation $\ensuremath{\tau^\text{FN}}$, we refer to
Figure~\ref{fig:necessary-reduction-example-figure}, which shows the result
for a simple example graph with only two vertices $a$ and $b$, of which $b$ is
necessary.
The basic idea is that the vertex $a'$ must be in every solution $S$:
If $a$ or any vertex to the left of $a$ is in $S$, it eventually forces $a'$ to
be in $S$ as well.
Likewise, if $b$ or any vertex to the right of $b$ is in $S$, it also forces
$a'$ to be in $S$.
Once $a' \in S$, the construction to the right of $a'$ makes sure that $b \in
S$.
We will generalize this to instances containing more vertices so that every
necessary vertex as well as the primed copy of each non-necessary vertex is in
every solution.
\begin{definition}
We define a function \ensuremath{\tau^\text{FN}}{}, which assigns a \prob{Defensive Alliance\textsuperscript{F}}{} instance to each pair $(I,{\preceq})$,
where
$I = (G,k,V_\square,V_\triangle)$ is a \prob{Defensive Alliance\textsuperscript{FN}}{} instance and
${\preceq}$ is an ordering of the non-forbidden elements of $\vertices{G}$.
For this,
let $V_{\Circle}$ denote $\vertices{G} \setminus (V_\square \cup V_\triangle)$.
We use
$n$ to denote $\size{\vertices{G}}$,
and we first define a function
$\sFN{I} : x \mapsto (n+3) \cdot (x + \size{V_{\Circle}}) - \size{V_\triangle}$.
We use $H$ to denote the set of new vertices $\{v', g_v, h_v, g_v^\square,
h_v^\square \mid v \in V_{\Circle}\}$.
The intention is for each $g_v^\square$ and $h_v^\square$ to be forbidden, for
each $v'$ and $h_v$ to be in every solution, and for $g_v$ to be in a solution
iff $v$ is in it at the same time.
We write
$V^+$ to denote $V_\triangle \cup V_{\Circle} \cup \{v' \mid v \in
V_{\Circle}\}$;
for each $v \in V^+$, we use $A_v$ to denote the set of new vertices
$\{v_1, \dots, v_{n+1}, v_1^\square, \dots, v_{n+1}^\square \}$,
and we use shorthand notation $A_v^{\Circle} = \{v_1, \dots, v_{n+1}\}$ and
$A_v^\square = \{v_1^\square, \dots, v_{n+1}^\square\}$.
The intention is for each $v_i^\square$ to be forbidden and for each $v_i$ to
be in a solution iff $v$ is in it at the same time.
We use the notation
$u \oplus v$
to denote the set of edges
$\{(u,v), (u,u^\square), (v,v^\square), (u,v^\square), (v,u^\square)\}$.
For any vertex $v \in V_{\Circle} \cup V_\triangle$, we define $p(v) = v$ if $v
\in V_\triangle$ and $p(v) = v'$ if $v \in V_{\Circle}$.
Let $P$ be the set consisting of all pairs $(p(u),p(v))$ such that $v$ is the
direct successor of $u$ according to ${\preceq}$.
Now we define $\ensuremath{\tau^\text{FN}}(I,{\preceq}) = (G',k',V_\square')$, where
$V'_\square = V_\square \cup \{g_v^\square, h_v^\square \mid v \in V_{\Circle}\}
\cup \bigcup_{v \in V^+} A_v^\square$,
$k' = \sFN{I}(k)$, and
$G'$ is the graph defined by
\begin{align*}
\vertices{G'} ={}& \vertices{G} \cup H \cup \bigcup_{v \in V^+} A_v,\\
\edges{G'} ={}& \edges{G} \cup \{ (v, v_i), (v,v_i^\square)
\mid v \in V^+,\; 1 \leq i \leq n+1 \}\\
{}\cup{}& \bigcup_{v \in V^+,\; 1 \leq i \leq n} v_i \oplus v_{i+1
\cup \bigcup_{(u,v) \in P} u_{n+1} \oplus v_1\\
{}\cup{}& \bigcup_{v \in V_{\Circle}} v_{n+1} \oplus g_
\cup \{ (v',g_v), (v',h_v), (g_v,h_v), (g_v,h_v^\square) \mid v \in
V_{\Circle} \}.
\end{align*}
We illustrate our construction in Figure~\ref{fig:necessary-reduction-gadget1} and~\ref{fig:necessary-reduction-gadget2}.
\label{def:dafn-to-daf}
\end{definition}
\input{figures/necessary-reduction-gadget1}
\input{figures/necessary-reduction-gadget2}
We now prove that $\ensuremath{\tau^\text{FN}}$ yields a correct reduction for any ordering $\preceq$.
\begin{lemma}
Let $I = (G,k,V_\square,V_\triangle)$ be a \prob{Defensive Alliance\textsuperscript{FN}}{} instance,
let ${\preceq}$ be an ordering of $\vertices{G} \setminus V_\square$,
let $A$ be the set of solutions of $I$ and
let $B$ be the set of solutions of the \prob{Defensive Alliance\textsuperscript{F}}{} instance $\ensuremath{\tau^\text{FN}}(I,{\preceq})$.
There is a bijection $f: A \to B$ such that
$\size{f(S)} = \sFN{I}(\size{S})$
holds for every $S \in A$.
\label{lem:dafn-to-daf-correct}
\end{lemma}
\begin{proof}
We use the same auxiliary notation as in Definition~\ref{def:dafn-to-daf}
and we define $f$ as
\[f(S) = S
\cup \bigcup_{v \in S} A_v^{\Circle}
\cup \{v', h_v \mid v \in V_{\Circle}\}
\cup \bigcup_{v \in V_{\Circle}} A_{v'}^{\Circle}
\cup \{g_v \mid v \in S \cap V_{\Circle}\}
.\]
For every $S \in A$,
we thus obtain
$\size{f(S)}
= \size{S} + \size{S} (n+1) + 2 \size{V_{\Circle}} + \size{V_{\Circle}} \cdot
(n+1) + (\size{S}-\size{V_\triangle})
= \sFN{I}(\size{S})$,
and we first show that indeed $f(S) \in B$.
Let $S \in A$ and let
$S'$ denote $f(S)$.
Obviously $S'$ satisfies
$V_\square' \cap S' = \emptyset$.
To see that $S'$ is a defensive alliance in $G'$,
let $x$ be an arbitrary element of $S'$.
If $x \in S$, then there is a defense
$\mu: N_G[x] \setminus S \to N_G[x] \cap S$
since $S$ is a defensive alliance in $G$.
We use this to construct a defense
$\mu': N_{G'}[x] \setminus S' \to N_{G'}[x] \cap S'$.
For any attacker $a$ of $x$ in $G'$, we distinguish the following cases:
\begin{itemize}
\item If $a$ is some $v_i^\square \in A_v^\square$ for some $v \in V^+$,
then $x$ is either $v_i$ or a neighbor of $v_i$, all of which are in $S'$,
and we set $\mu'(a) = v_i$.
\item Similarly, if $a$ is $g_v^\square$ for some $v \in V_{\Circle}$,
then we set $\mu'(a) = g_v$.
\item If $a$ is $h_v^\square$ for some $v \in V_{\Circle}$,
then $x = g_v$ and we set $\mu'(a) = h_v$.
\item If $a$ is $g_v$ for some $v \in V_{\Circle}$,
then $x$ is either $v'$ or $h_v$,
which is not used for repelling any other attack because $h_v^\square$ cannot
attack $x$,
so we set $\mu'(a) = h_v$.
\item Otherwise $a$ is in $N_G[x] \setminus S$ (by our construction of $S'$).
Since the codomain of $\mu$ is a subset of the codomain of $\mu'$, we may set $\mu'(a) = \mu(a)$.
\end{itemize}
Since $\mu'$ is injective, each attack on $x$ in $G'$ can be repelled by $S'$.
Hence $S'$ is a defensive alliance in $G'$.
Clearly $f$ is injective.
It remains to show that $f$ is surjective.
Let $S'$ be a solution of $\ensuremath{\tau^\text{FN}}(I,{\preceq})$.
We first show that $V_\triangle \cup \{v', h_v \mid v \in V_{\Circle}\}
\subseteq S'$:
\begin{itemize}
\item If $S'$ contains some $v \in V^+$, then
$S'$ contains an element of $A_v^{\Circle}$
by Observation~\ref{obs:one-in-s}.
\item If $S'$ contains an element of $A_v^{\Circle}$ for some $v \in V^+$, then
$\{v\} \cup A_v^{\Circle} \subseteq S'$
by Observation~\ref{obs:all-in-s}.
\item If $v_{n+1} \in S'$ for some $v \in V_{\Circle}$, then
$g_v \in S'$ for the same reason.
\item Furthermore, if $S'$ contains an element of $A_v^{\Circle}$ for some $v
\in V_\triangle \cup \{v' \mid v \in V_{\Circle}\}$, then
also $A_{u}^{\Circle} \subseteq S'$ for every $u \in V_\triangle \cup \{v' \mid
v \in V_{\Circle}\}$
for the same reason.
\item If $g_v \in S'$ for some $v \in V_{\Circle}$, then
$\{h_v,v',v_{n+1}\} \subseteq S'$ by Observation~\ref{obs:all-in-s}.
\item If $h_v \in S'$ for some $v \in V_{\Circle}$, then
$a' \in S'$ because at least $g_v$ or $v'$ must be in $S'$
and the former implies $v' \in S'$ as we have seen.
\item
Since $S'$ is nonempty, the previous observations show that
for every $v \in V_\triangle \cup \{v' \mid v \in V_{\Circle}\}$
it holds that
$\{v\} \cup A_v^{\Circle} \subseteq S'$.
Finally, we show that $\{h_v \mid v \in V_{\Circle}\} \subseteq S'$.
Suppose, for the sake of contradiction, that there is some
$v \in V_{\Circle}$ such that
$h_v \notin S'$.
We have seen that the latter can only be the case if $g_v \notin S'$, and we
know that $v' \in S'$.
We obtain the contradiction that $v'$ is attacked by $g_v$, $h_v$ and
$A_{v'}^\square$, whereas its only defenders are $v'$ itself and
$A_{v'}^{\Circle}$.
\end{itemize}
Let $S = S' \cap \vertices{G}$.
By the previous observations, it is easy to see that $S' = f(S)$.
It remains to show that $S$ is a defensive alliance in $G$.
Let $x$ be an arbitrary element of $S$.
We observe that $N_{G'}[x] \cap S' = (N_G[x] \cap S) \cup
A_x^{\Circle}$ and similarly
$N_{G'}[x] \setminus S' = (N_G[x] \setminus S) \cup
A_v^\square$.
Since $\size{A_x^{\Circle}} = \size{A_x^\square}$, this implies
$\size{N_{G'}[x] \cap S'} - \size{N_G[x] \cap S} = \size{N_{G'}[x] \setminus
S'} - \size{N_G[x] \setminus S}$.
Since $S'$ is a defensive alliance in $G'$ and $x \in S'$,
it holds that $\size{N_{G'}[x] \cap S'} \geq \size{N_{G'}[x] \setminus S'}$.
We conclude
that
$\size{N_G[x] \cap S} \geq \size{N_G[x] \setminus S}$.
Hence $S$ is a defensive alliance in $G$.
\end{proof}
Given an ordering ${\preceq}$, clearly $\ensuremath{\tau^\text{FN}}(I,\preceq)$ is computable in
polynomial time.
We can thus easily obtain a reduction from \prob{Defensive Alliance\textsuperscript{FN}}{} to \prob{Defensive Alliance\textsuperscript{F}}{}
by first computing an arbitrary ordering ${\preceq}$ of the non-forbidden
vertices.
We next show that by choosing $\preceq$ appropriately, this amounts to an FPT
reduction that preserves bounded treewidth.
\begin{lemma}
\prob{Defensive Alliance\textsuperscript{F}}{}, parameterized by the treewidth of the graph, is \complexityclass{W[1]}{}-hard.
\label{lem:daf-wone-hard}
\end{lemma}
\begin{proof}
Let $I = (G,k,V_\square,V_\triangle)$ be a \prob{Defensive Alliance\textsuperscript{FN}}{} instance
and let $\mathcal{T}$ be an optimal nice tree decomposition of $G$.
We can compute such a tree decomposition in
FPT time~\cite{siamcomp:Bodlaender96}.
Let ${\preceq}$ be
the
ordering of the elements of $V_\triangle \cup V_{\Circle}$
that is obtained in linear time by doing a post-order traversal of $\mathcal{T}$ and
sequentially recording the elements that occur for the last time in the current
bag.
We obtain the \prob{Defensive Alliance\textsuperscript{F}}{} instance $\ensuremath{\tau^\text{FN}}(I,\preceq)$, whose graph we denote by $G'$.
This reduction is correct, as shown in Lemma~\ref{lem:dafn-to-daf-correct},
and computable in FPT time.
It remains to show that the treewidth of $G'$ is bounded by a function of the treewidth of $G$.
To this end, we use $\mathcal{T}$ to build a tree decomposition $\mathcal{T}'$ of $G'$.
We initially set $\mathcal{T}' := \mathcal{T}$ and modify it by the following steps:
\begin{enumerate}
\item For each $v \in V_{\Circle}$, we add $g_v$, $g_v^\square$, $h_v$,
$h_v^\square$ and $v'$ to the bag of $t_v^{\mathcal{T}'}$.
Note that afterwards $t_v^{\mathcal{T}'} = t_{v'}^{\mathcal{T}'}$.
After this step we increased the width of $\mathcal{T}'$ by at most five.
\item For each $v \in V^+$, we use $B_v$ to denote the bag of $t_v^{\mathcal{T}'}$
and replace $t_v^{\mathcal{T}'}$ by a chain of nodes $N_1, \dots, N_n$, where
$N_n$ is the topmost node and the bag of $N_i$ is $B_v \cup \{v_i,
v_i^\square, v_{i+1}, v_{i+1}^\square\}$.
After this step we increased the width of $\mathcal{T}'$ by at most nine.
Note that the bag of the new node $t_v^{\mathcal{T}'}$ now contains $v_{i+1}$ and
$v_{i+1}^\square$.
We have so far covered all edges except the ones connecting elements of two
different sets $A_x$ and $A_y$ for $(x,y) \in P$.
\item For every $(u,v) \in P$, we add $v_1$ and $v_1^\square$ into the bag of
every node between (and including) $t_u^{\mathcal{T}'}$ and $t_{v_1}^{\mathcal{T}'}$.
Note that this preserves connectedness and afterwards the bag of
$t_u^{\mathcal{T}'}$ contains $u_{i+1}$, $u_{i+1}^\square$, $v_1$ and
$v_1^\square$, thus covering the remaining edges.
After this step we increased the width of $\mathcal{T}'$ by at most 13.
(Since the number of children of each tree decomposition node is at most
two, this step enlarges every bag at most twice.)
\end{enumerate}
It is easy to verify that $\mathcal{T}'$ is a valid tree decomposition of $G'$.
Furthermore, the width of $\mathcal{T}'$ is at most the width of $\mathcal{T}$ plus 13.
\end{proof}
We again get an analogous result for the exact variant.
\begin{corollary}
\prob{Exact Defensive Alliance\textsuperscript{F}}{}, parameterized by the treewidth of the graph, is \complexityclass{W[1]}{}-hard.
\label{cor:edaf-wone-hard}
\end{corollary}
\subsection{Hardness of Defensive Alliance}
\label{sec:da-wone-tw}
We now introduce a transformation $\ensuremath{\tau^\text{F}}$ that eliminates forbidden vertices.
The basic idea is that we ensure that a forbidden vertex $f$ is never part of a
solution by adding so many neighbors to $f$ that we could only defend $f$ by
exceeding the bound on the solution size.
\begin{definition}
We define a function \ensuremath{\tau^\text{F}}{}, which assigns a \prob{Defensive Alliance}{} instance to each \prob{Defensive Alliance\textsuperscript{F}}{}
instance
$I = (G,k,V_\square)$.
\label{def:daf-to-da}
For each $f \in V_\square$, we
introduce new vertices $f', f_1, \dots, f_{2k}$.
Now we define $\ensuremath{\tau^\text{F}}(I) = (G',k)$, where
$G'$ is the graph defined by
\begin{align*}
\vertices{G'} ={}& \vertices{G} \cup \{f', f_1, \dots, f_{2k} \mid f \in V_\square\},\\
\edges{G'} ={}& \edges{G} \cup \{ (f,f_i),\; (f',f_i) \mid f \in V_\square,\; 1 \leq i \leq 2k\}.
\end{align*}
\end{definition}
We now prove that $\ensuremath{\tau^\text{F}}$ yields a correct reduction from \prob{Defensive Alliance\textsuperscript{F}}{} to \prob{Defensive Alliance}{}.
\begin{lemma}
Every \prob{Defensive Alliance\textsuperscript{F}}{} instance $I$ has the same solutions as the \prob{Defensive Alliance}{} instance
$\ensuremath{\tau^\text{F}}(I)$.
\label{lem:daf-to-da-correct}
\end{lemma}
\begin{proof}
Let $I = (G,k,V_\square)$ and $\ensuremath{\tau^\text{F}}(I) = (G',k)$.
Each solution $S$ of $I$ is also a solution of $\ensuremath{\tau^\text{F}}(I)$ because the subgraph of
$G$ induced by $N_G[S]$ is equal to the subgraph of $G'$ induced by
$N_{G'}[S]$.
Now let $S'$ be a solution of $\ensuremath{\tau^\text{F}}(I)$.
For every $f \in V_\square$,
neither $f$ nor $f'$ are in $S'$
because each of these vertices has at least $2k$ neighbors,
and $S'$ cannot contain any $f_i$ because $N_{G'}(f_i) = \{f,f'\}$.
Hence $S'$ is also a solution of $I$ as the subgraphs induced by the respective
neighborhoods are again equal.
\end{proof}
We use $\ensuremath{\tau^\text{F}}$ to show \complexityclass{W[1]}{}-hardness of \prob{Defensive Alliance}{} by reducing from \prob{Defensive Alliance\textsuperscript{F}}{} while
preserving bounded treewidth.
\begin{lemma}
\prob{Defensive Alliance}{}, parameterized by the treewidth of the graph, is \complexityclass{W[1]}{}-hard.
\label{lem:da-wone-hard}
\end{lemma}
\begin{proof}
Let $I = (G,k,V_\square)$ be a \prob{Defensive Alliance\textsuperscript{F}}{} instance,
let $G'$ denote the graph of $\ensuremath{\tau^\text{F}}(I)$
and let $\mathcal{T}$ be an optimal nice tree decomposition of $G$.
We build a tree decomposition $\mathcal{T}'$ of $G'$ by modifying a copy of $\mathcal{T}$ in the following way:
For every $f \in V_\square$, we pick an arbitrary node $t$ in $\mathcal{T}$ whose bag $B$ contains $f$,
and we add a chain of nodes $N_1, \dots, N_{2k}$ between $t$ and its parent such that,
for $1 \leq i \leq 2k$, the bag of $N_i$ is
$B \cup \{f', f_i\}$.
It is easy to verify that $\mathcal{T}'$ is a valid tree decomposition of $G'$.
Furthermore, the width of $\mathcal{T}'$ is at most the width of $\mathcal{T}$ plus two.
\end{proof}
We again get an analogous result for the exact variant.
\begin{corollary}
\prob{Exact Defensive Alliance}{}, parameterized by the treewidth of the input graph, is \complexityclass{W[1]}{}-hard.
\label{cor:eda-wone-hard}
\end{corollary}
\section{Conclusion}
\label{sec:conclusion}
In this work, we proved that the problem of deciding whether a given graph
possesses a nonempty defensive alliance whose size is at most a given integer
is \complexityclass{W[1]}-hard when the parameter is the treewidth of the graph.
This means that no fixed-parameter tractable algorithm exists under the common
complexity-theoretic assumption $\complexityclass{W[1]} \neq \complexityclass{FPT}$.
Still, recent work has shown the problem to be solvable in polynomial time on
graphs of bounded clique-width~\cite{dam:KiyomiO17}, which implies that there
is a polynomial-time algorithm for graphs of bounded treewidth.
Our result proves that, for any such algorithm, the degree of this polynomial
must necessarily depend on the treewidth unless $\complexityclass{W[1]} = \complexityclass{FPT}$.
For future research it may be interesting to study related problems
corresponding to other alliance notions such as offensive alliances.
|
\section{Introduction}
Next generation wireless communication networks are expected to rely on low latency and high spectral efficiency cooperative relaying nodes \cite{Orikumhi2017}. This is due to their capability to serve as a virtual multi-antenna system to combat fading and be able to extend communication coverage\cite{Laneman2004,Azarian2005,Shadmand2010}. However, conventional cooperative relaying networks limit relay nodes on half-duplex (HD) mode, where the HD relay nodes can either receive or transmit data symbols at a given time-instant. With such kind of design, the cooperative system can avoid inter-frame interference and exploit spatial diversity gain. On the other hand, it suffers from spectral efficiency loss since transmission of one data frame needs to occupy two independent time slots or frequency bands.
Recent research works in \cite{Sabharwal2014,Liu2015,Naslcheraghi2017} show that full-duplex (FD) relaying becomes feasible for simultaneous transmission and reception when advanced self-interference cancellation methods can be exploited \cite{Hong2014,Ahmed2015}. However, in practical environments, self-interference cannot be completely cancelled due to inaccurate self-interference channel modeling and/or limited dynamic range of analog-to-digital converter\cite{Kim2015}. Consequently, the residual self-interference increases burden on decoding process at the FD relay node and results in error propagation. In \cite{Oechtering2004,Fan2007,Lehmann2016}, a promising technique named virtual FD relaying (or two-path successive relaying) has been proposed, where two HD relay nodes mimic a FD relay node and alternately serve as transmitter and receiver to continuously transmit data frames for every channel use. This scheme benefits from simplified curcit designs, and the optimal performance can be achieved if both HD relay nodes are able to perfectly cancel the inter-relay interference and decode their received messages\cite{Wicaksana2011,Fan2017}. Otherwise, error propagation effects degrade system performances.
In order to avoid error propagation effects, the authors in \cite{Fan2007,Wicaksana2011,Bobarshad2009,Zhang2015,Olivo2016} presented adaptive retransmission based relaying schemes to guarantee perfect decoding at relay nodes. However, this scheme has additional costs of time delay and signalling overhead \cite{Shikh-Bahaei2004,Olfat2005,Kobravi2007,Shadmand2009}. Alternatively, the authors in \cite{Al-Habian2011,Kwon2010,Yi2015} proposed symbol-level selective transmission schemes, where relay nodes can predict their correctly decoded symbols based on the pre-determined threshold and discard the the erroneously decoded symbols, resulting in fewer errors being forwarded to destination node. However, in order to guarantee the quality of decoding process at destination, the relay nodes need to forward the position of their discarded symbols to destination node, where the additional signalling overhead still exists. On the other hand, imperfect inter-relay interference cancellation may reduce the decoding capability at relay nodes \cite{Sun2011,Basar2014,Zarringhalam2009}. With the help of multiple receive-antenna, the detection techniques, such as zero-forcing (ZF) or minimum mean square error (MMSE), can be implemented to suppress the inter-relay interference. However, the majority drawback of ZF is noise power boost, and MMSE needs to know noise variance information.
Motivated by the above discussion, in this paper, we propose a multi-antenna assisted virtual FD relaying with reliability-aware iterative decoding at destination node in order to improve system spectral efficiency and reliability. Specifically, relay nodes first ultilize QR decodposition method to perfectly cancel the inter-relay interference and then forward their decoded data frames regardless the decoding errors. Subsequently, destination node performs the reliability-aware iterative detection/decoding method by taking the estimated probability of decoding error at relay nodes (e.g. $p_e$) into account. In the literature, multi-antenna assisted iterative decoding was proposed in \cite{Vikalo2004} for point-to-point communication network and in \cite{Lee2009} for network coding aided two-way relaying network. Comparison with the works in \cite{Vikalo2004,Lee2009}, our proposed scheme is for virtual FD relaying network and the additional challenge is how to jointly cancel inter-frame interference and enjoy spatial diversity gain at destination node. In addition, we also introduce an estimation method of $p_e$ for the iterative decoding process. Simulation results show that, without extra cost of time delay and signalling overhead , our proposed scheme exhibits up to 8 dB bit-error-rate (BER) performance gain by comparing with conventional cyclic redundancy check (CRC) based S-DF relaying and threshold based S-DF relaying schemes.
\section{System Model}
Fig.~\ref{F1} illustrates the multi-antenna assisted virtual FD relaying network with one single-antenna based source node (S), two $N$-antenna based HD relay nodes (R1/R2) and one $N$-antenna based destination node (D).
\begin{figure}[t]
\begin{center}
\epsfig{figure=F1.pdf,scale=0.43,angle=0}
\end{center}
\caption{Illustration of the virtrual FD relaying network.}\label{F1}
\end{figure}
All transmissions are using single antenna and all receptions are using $N$ antennas.\footnote{In order to keep consistent with the source node, two relay nodes transmit data frames with single antenna. The multi-antenna based transmission at relay nodes can be easily extended to enjoy additional antenna gain.} We also assume discrete-time block fading channels, which remain static over each time slot. All reception nodes can estimate and have their related channel state information. The source node first encodes the information bits $\mathbf{b}$ using the turbo-like encoders, e.g., $m$-rate serial concatenated convolutional codes, to generate the coded bits $\mathbf{c}$. Then, $\mathbf{c}$ are mapped into $\mathbf{x}$ transmission symbols based on $Q$-ary modulation scheme. Subsequently, $\mathbf{x}$ are divided into $L$ frames, and without loss of generality $L$ is assumed to be even.
As shown in Fig.~\ref{F1}, the source node continuously transmits the $l^{\mathrm{th}}$ frame in the $l^{\mathrm{th}}$ time slot, where $l=1,2,\ldots,L$. In the first time slot, only relay node one (R1) and destination node receive the first data frame from source node. Starting from the second time slot, two relay nodes alternatively transmit and receive data frames until the $L^{\mathrm{th}}$ time slot. Then, in the $(L+1)^{\mathrm{th}}$ time slot, only relay node two (R2) forwards its last decoded frame (i.e. the $L^{\mathrm{th}}$) to destination node. Let $\mathbf{Y}_{\mathrm{R}j}(l)\in\mathcal{C}^{N\times M}$ and $\mathbf{Y}_{\mathrm{D}}(l)\in\mathcal{C}^{N\times M}$ be the received signal matrices at the $j^{\mathrm{th}},j\in\{1,2\}$, relay and the destination node in the $l^{\mathrm{th}}$ time slot, respectively, where $M$ is the number of symbols per frame. Then, we have
\begin{equation}\label{eq01}
\mathbf{Y}_{\mathrm{R}j}(l)=\mathbf{H}_{\mathrm{R}j}(l)\mathbf{X}(l)+\mathbf{V}_{\mathrm{R}j}(l),
\end{equation}
\begin{equation}\label{eq02}
\mathbf{Y}_{\mathrm{D}}(l)=\mathbf{H}_{\mathrm{D}}(l)\mathbf{X}(l)+\mathbf{V}_{\mathrm{D}}(l),
\end{equation}
where $\mathbf{X}(l)\triangleq[\mathbf{x}_{\mathrm{R}\overline{j}}(l-1),\mathbf{x}_{\mathrm{S}}(l)]^{T}$ is the transmitted signal matrix in the $l^{\mathrm{th}}$ time slot composed by the $(l-1)^{\mathrm{th}}$ frame sent from the $\overline{j}^{\mathrm{th}},\overline{j}\neq j\in\{1,2\}$, relay, i.e., $\mathbf{x}_{\mathrm{R}}(l-1)\in\mathcal{C}^{M}$, and the $l^{\mathrm{th}}$ frame sent from the source node, i.e., $\mathbf{x}_{\mathrm{S}}(l)\in\mathcal{C}^{M}$; $(\cdot)^{T}$ is transpose of a matrix; $\mathbf{H}_{\mathrm{R}j}(l)\in\mathcal{C}^{N\times2}$ is the channel matrix which consists of the R-R link and S-R link in the $l^{\mathrm{th}}$ time slot; $\mathbf{H}_{\mathrm{D}}(l)\in\mathcal{C}^{N\times2}$ is the channel matrix which consists of R-D link and S-D link in the $l^{\mathrm{th}}$ time slot; $\mathbf{V}_{\mathrm{R}j}(l)\in\mathcal{C}^{N\times M}$ and $\mathbf{V}_{\mathrm{D}}(l)\in\mathcal{C}^{N\times M}$ are the additive white Gaussian noise (AWGN) matrices at the $j^{\mathrm{th}}$ relay and the destination nodes in the $l^{\mathrm{th}}$ time slot, respectively. It is worth noting that, in the first time slot, since relay node one receives data frame sent from the source node without inter-relay interference effect, the transmitted signal in this case should be $\mathbf{X}(1)=[\mathbf{0},\mathbf{x}_{\mathrm{S}}(1)]^{T}$, where $\mathbf{0}$ is a column vector with all zeros. Similarly, in the $(L+1)^{\mathrm{th}}$ time slot, since only relay node two (R2) sends data frame to destination node, the transmitted signal in this case should be $\mathbf{X}(L+1)=[\mathbf{x}_{\mathrm{R}2}(L),\mathbf{0}]^{T}$. Apart from that, in this paper, relay nodes decodes its received frame and then forwards its re-encoded version regardless of decoding errors, then the destination node will based on our proposed method to mitigate the error propagation effect.
\section{The Proposed Scheme for Interference and Error Propagation Mitigation}
According to the system model described above, there are three main factors that affect the system performances, which are the inter-relay interference generated at relay nodes, the inter-frame interference generated at destination node, and the error propagation effect due to decoding errors at relay nodes. In this section, we mainly introduce how to cancel and limit these impacts with our proposed methods.
\subsection{Inter-relay Interference Cancellation at Relay Nodes}
Due to alternative transmission and reception between two relay nodes, the presented inter-relay interference should be cancelled before they can decode the received data frames. In this case, by introducing QR decomposition, the inter-relay interference at relay nodes can be perfectly cancelled without boosting noise power. Specifically, we take the $j^{\mathrm{th}}$ relay node as an example and assume $\mathbf{Q}_{\mathrm{R}j}(l)\in\mathcal{C}^{N\times N}$ is the unitary matrix and $\mathbf{R}_{\mathrm{R}j}(l)\in\mathcal{C}^{N\times2}$ is the upper triangular matrix, where both matrices are generated from QR decomposition of the channel matrix $\mathbf{H}_{\mathrm{R}j}(l)$. Then, the received signal at the $j^{\mathrm{th}}$ relay node, after multiplying the unitary matrix $\mathbf{Q}_{\mathrm{R}j}(l)$, can be expressed as
\begin{eqnarray}\label{eq03}
\hat{\mathbf{Y}}_{\mathrm{R}j}(l)&=&\mathbf{Q}^{H}_{\mathrm{R}j}(l)\mathbf{Y}_{\mathrm{R}j}(l), \nonumber\\
&=&\mathbf{R}_{\mathrm{R}j}(l)\left[\begin{array}{cc}\mathbf{x}^{T}_{\mathrm{R}\overline{j}}(l-1) \\
\mathbf{x}^{T}_{\mathrm{S}}(l)
\end{array}\right]+\mathbf{Q}^{H}_{\mathrm{R}j}(l)\mathbf{V}_{\mathrm{R}j}(l),
\end{eqnarray}
where $(\cdot)^{H}$ is hermitian of a matrix. Due to the upper triangular structure of $\mathbf{R}_{\mathrm{R}j}(l)$, the desired signal sent from the source node without inter-relay interference effects is the second row of matrix $\hat{\mathbf{Y}}_{\mathrm{R}j}(l)$. In addition, due to the nice feature of $\mathbf{Q}_{\mathrm{R}j}(l)$, the multiplication of $\mathbf{V}_{\mathrm{R}j}(l)$ by $\mathbf{Q}_{\mathrm{R}j}(l)$ does not change the power of AWGN matrix. Thus, based on the above description, the $j^{\mathrm{th}}$ relay node can decode the second row of $\hat{\mathbf{Y}}_{\mathrm{R}j}(l)$ in order to regenerate data frame sent from source node, i.e., $\mathbf{x}_{\mathrm{R}j}(l)$. Then, it will be forwarded to destination node in the $(l+1)^{\mathrm{th}}$ time slot.
\subsection{Reliability-Aware Iterative Decoding at Destination Node}
In this paper, we allow relay nodes forward their received signals regardless decoding error. In this case, error propagation could degrade the system performance. In addition, due to simultaneously receiving signals sent from source and one of relay nodes, the inter-frame interference at destination node also affects system performance. In this subsection, we introduce a reliability-aware iterative detection/decoding method to mitigate the above effects.
It is shown in the literature that MAP based detection technique has been widely used to cancel inter-frame interference. However, in order to allow the MAP detector have the capacity to jointly cancel inter-frame interference and remove the error propagation effects, we need to modify its detection process. Specifically, we start from transforming \eqref{eq02} to the real-valued equivalent form. Such transformation allows the proposed method suit for different modulation schemes, e.g., generalized $Q$-ary modulation schemes. Let's assume that $\tilde{\mathbf{Y}}_{\mathrm{D}}(l)\in\mathcal{R}^{2N\times M}$, $\tilde{\mathbf{X}}(l)\in\mathcal{R}^{2N\times M}$, and $\tilde{\mathbf{V}}_{\mathrm{D}}(l)\in\mathcal{R}^{2N\times M}$ are the real matrices transformed from $\mathbf{Y}_{\mathrm{D}}(l)$, $\mathbf{X}(l)$, and $\mathbf{V}_{\mathrm{D}}(l)$, respectively, where we have
\begin{equation}\label{eq04}
\tilde{\mathbf{Y}}_{\mathrm{D}}(l)=[\mathcal{R}(\mathbf{Y}_{\mathrm{D}}(l))^{T},\mathcal{I}(\mathbf{Y}_{\mathrm{D}}(l))^{T}]^{T},
\end{equation}
\begin{equation}\label{eq05}
\tilde{\mathbf{X}}(l)=[\mathcal{R}(\mathbf{X}(l))^{T},\mathcal{I}(\mathbf{X}(l))^{T}]^{T},
\end{equation}
\begin{equation}\label{eq06}
\tilde{\mathbf{V}}_{\mathrm{D}}(l)=[\mathcal{R}(\mathbf{V}_{\mathrm{D}}(l))^{T},\mathcal{I}(\mathbf{V}_{\mathrm{D}}(l))^{T}]^{T}.
\end{equation}
Additionally, let $\tilde{\mathbf{H}}_{\mathrm{D}}(l)\in\mathcal{R}^{2N\times4}$ denote the real-valued channel matrix transformed from $\mathbf{H}_{\mathrm{D}}(l)$, as
\begin{eqnarray}\label{eq07}
\tilde{\mathbf{H}}_{\mathrm{D}}(l)=\left[\begin{array}{cc}
\mathcal{R}(\mathbf{H}_{\mathrm{D}}(l)) & -\mathcal{I}(\mathbf{H}_{\mathrm{D}}(l)) \\
\mathcal{I}(\mathbf{H}_{\mathrm{D}}(l)) & \mathcal{R}(\mathbf{H}_{\mathrm{D}}(l)) \end{array} \right].
\end{eqnarray}
Then, the real-valued equivalent form of \eqref{eq02} can be expressed as
\begin{equation}\label{eq08}
\tilde{\mathbf{Y}}_{\mathrm{D}}(l)=\tilde{\mathbf{H}}_{\mathrm{D}}(l)\tilde{\mathbf{X}}(l)+\tilde{\mathbf{V}}_{\mathrm{D}}(l).
\end{equation}
According to the work in \cite{Vikalo2004}, the conventional MAP detector is to solve the optimization problem, which is
\begin{equation}\label{eq09}
\min_{\tilde{\mathbf{x}}^{(m)}(l)}\Bigg[\underbrace{\|\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)-\tilde{\mathbf{H}}_{\mathrm{D}}(l)\tilde{\mathbf{x}}^{(m)}(l)\|^2}_{\triangleq\Delta^{(m)}(l)}-\sum^{2N}_{k=1}{\mathrm{log}p(\tilde{x}^{(l)}_{k,m})}\Bigg],
\end{equation}
where $\tilde{\mathbf{x}}^{(m)}(l)$ and $\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)$ are the $m^{\mathrm{th}}$ columns of $\tilde{\mathbf{X}}(l)$ and $\tilde{\mathbf{Y}}_{\mathrm{D}}(l)$, respectively;\footnote{It is worth noting that $\tilde{\mathbf{x}}^{(m)}(l)$ in \eqref{eq09} are the corresponding trial bits vector used in this hypothesis-detection problem.} $p(\tilde{x}^{(l)}_{k,m})$ is the \textit{a posteriori} probability for $\tilde{x}^{(l)}_{k,m}$, where $\tilde{x}^{(l)}_{k,m}$ is the $k^{\mathrm{th}}$ component of the symbol vector $\tilde{\mathbf{x}}^{(m)}(l)$. As each constellation symbol represents $\log_{2}Q$ modulated bits (e.g., for a $Q$-ary constellation), we assume $c^{(l)}_{i,m},\forall i\in[1,\ldots,2\log_{2}Q],$ are the modulated bits for the symbol vector $\tilde{\mathbf{x}}^{(m)}(l)$.\footnote{Two times of $\log_{2}Q$ represents two symbols included in $\tilde{\mathbf{x}}^{(m)}(l)$, where one is from the source node, and one is from the relay node.} Since the output of the MAP detector requires the soft information for the next step decoding process, the LLR of the $i^{\mathrm{th}}$ modulated bit for the symbol vector $\tilde{\mathbf{x}}^{(m)}(l)$ can be expressed as
\begin{eqnarray}\label{eq10}
L[c^{(l)}_{i,m}|\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)]=\log\frac{\mathrm{Pr}[c^{(l)}_{i,m}=+1|\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)]}{\mathrm{Pr}[c^{(l)}_{i,m}=-1|\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)]}~~~~~~~~~~~~~\nonumber\\
=\mathrm{log}\frac{\sum_{\mathbf{x}:c^{(l)}_{i,m}=+1}e^{-\Delta^{(m)}(l)+\sum_{j}{\mathrm{log}p(c^{(l)}_{j,m})}}}{\sum_{\mathbf{x}:c^{(l)}_{i,m}=-1}e^{-\Delta^{(m)}(l)+\sum_{j}{\mathrm{log}p(c^{(l)}_{j,m})}}},
\end{eqnarray}
where $\Delta^{(m)}(l)$ has been defined in \eqref{eq09}; $\mathbf{x}$ is the short-hand used to denote the symbol vector $\tilde{\mathbf{x}}^{(m)}(l)$; $c^{(l)}_{i,m}=+1$ represent logical 0 with amplitude level +1, and $c^{(l)}_{i,m}=-1$ represent logical 1 with amplitude level -1.
Different from the conventional MAP detector presented above, our proposed MAP detector needs to be aware of the error propagation effects. Specifically, we first define $p^{(l)}_e$ as the probability of decoding error at one of relay nodes in the $l^{\mathrm{th}}$ time slot. This probability can be estimated during the iterative decoding at destination node. More detailed analysis of its estimation method will be introduced in Section III-C. Based on the transformed $\tilde{\mathbf{x}}^{(m)}(l),\forall m,l$, in \eqref{eq05}, all the components from the odd rows of $\tilde{\mathbf{x}}^{(m)}(l)$ represent the symbol sent from one of relay nodes, and all the components from the even rows of $\tilde{\mathbf{x}}^{(m)}(l)$ represent the symbol sent from the source node. Let's define a modulated bits set $\mathcal{S}_{\mathrm{odd}}$ representing all the components from the odd rows of $\tilde{\mathbf{x}}^{(m)}(l)$, and $\overline{\mathcal{S}}_{\mathrm{odd}}$ is its complementary set representing all the components from the even rows of $\tilde{\mathbf{x}}^{(m)}(l)$. By taking $p^{(l)}_e$ into account, the \textit{a posteriori} probabilities of the modulated bits sent from the relay node can be modified as
\begin{IEEEeqnarray}{ll}
p(\hat{c}^{(l)}_{j,m}=+1)=(1-p^{(l)}_e)p(c^{(l)}_{j,m}=+1)+p^{(l)}_ep(c^{(l)}_{j,m}=-1),\nonumber\\
p(\hat{c}^{(l)}_{j,m}=-1)=(1-p^{(l)}_e)p(c^{(l)}_{j,m}=-1)+p^{(l)}_ep(c^{(l)}_{j,m}=+1),\nonumber
\end{IEEEeqnarray}
where $p(\hat{c}^{(l)}_{j,m})$ denotes the modified probability of the $j^{\mathrm{th}}$ bit sent from the relay node and $j\in\mathcal{S}_{\mathrm{odd}}$. Then, by replacing $p(c^{(l)}_{j,m})$ with $p(\hat{c}^{(l)}_{j,m})$ in \eqref{eq10} when $j\in\mathcal{S}_{\mathrm{odd}}$, the LLR value of the $i^{\mathrm{th}}$, $i\in\overline{\mathcal{S}}_{\mathrm{odd}}$, modulated bits sent from the source node can be expressed as
\begin{IEEEeqnarray}{ll}\label{eq11}
L[c^{(l)}_{i,m}|\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)]=\mathrm{log}\frac{\sum_{\mathbf{x}:c^{(l)}_{i,m}=+1}e^{-\Delta^{(m)}(l)+\Theta^{(m)}(l)}}{\sum_{\mathbf{x}:c^{(l)}_{k,m}=-1}e^{-\Delta^{(m)}(l)+\Theta^{(m)}(l)}},
\end{IEEEeqnarray}
where
\begin{IEEEeqnarray}{ll}\label{eq12}
\Theta^{(m)}(l)\triangleq\sum_{j,j\in\overline{\mathcal{S}}_{\mathrm{odd}}}{\mathrm{log}p(c^{(l)}_{j,m})}+\sum_{j,j\in\mathcal{S}_{\mathrm{odd}}}{\mathrm{log}p(\hat{c}^{(l)}_{j,m})}.
\end{IEEEeqnarray}
On the other hand, due to the error prorogation effects, the LLR value of the $i^{\mathrm{th}}$, $i\in\mathcal{S}_{\mathrm{odd}}$, modulated bits sent from the relay node can be expressed as \eqref{eq13}, shown on the top of next page,
\begin{figure*}
\normalsize \vspace*{-0pt}
\begin{IEEEeqnarray}{ll}\label{eq13}
L[c^{(l)}_{i,m}|\tilde{\mathbf{y}}^{(m)}_{\mathrm{D}}(l)]=\log\frac{p(c^{(l)}_{i,m}=+1)}{p(c^{(l)}_{i,m}=-1)}+\mathrm{log}\frac{(1-p^{(l)}_e)\sum_{\mathbf{x}:\hat{c}^{(l)}_{i,m}=+1}e^{-\Delta^{(m)}(l)+\tilde{\Theta}^{(m)}(l)}+p^{(l)}_e\sum_{\mathbf{x}:\hat{c}^{(l)}_{i,m}=-1}e^{-\Delta^{(m)}(l)+\tilde{\Theta}^{(m)}(l)}}{(1-p^{(l)}_e)\sum_{\mathbf{x}:\hat{c}^{(l)}_{i,m}=-1}e^{-\Delta^{(m)}(l)+\tilde{\Theta}^{(m)}(l)}+p^{(l)}_e\sum_{\mathbf{x}:\hat{c}^{(l)}_{i,m}=+1}e^{-\Delta^{(m)}(l)+\tilde{\Theta}^{(m)}(l)}},\nonumber\\
\end{IEEEeqnarray}
\vspace*{-0.2cm}\hrulefill
\end{figure*}
where
\begin{IEEEeqnarray}{ll}\label{eq14}
\tilde{\Theta}^{(m)}(l)\triangleq\sum_{j,j\in\overline{\mathcal{S}}_{\mathrm{odd}}}{\mathrm{log}p(c^{(l)}_{j,m})}+\sum_{j,j\neq i\in\mathcal{S}_{\mathrm{odd}}}{\mathrm{log}p(\hat{c}^{(l)}_{j,m})}.
\end{IEEEeqnarray}
Here, we omit the detailed derivation of \eqref{eq13} because of the limited space. On the other hand, the basic principle of deriving \eqref{eq13} is to solve $\sum_{j,j\in\mathcal{S}_{\mathrm{odd}}}{\mathrm{log}p(\hat{c}^{(l)}_{j,m})}$ on the condition that $c^{(l)}_{i,m}=+1$ for the numerator of \eqref{eq13} and on the condition that $c^{(l)}_{i,m}=-1$ for the denominator of \eqref{eq13}. Similar analisis method for the network coding aided two-way relaying system can be found in \cite{Lee2009}.
Fig.~\ref{F2} illustrates the iterative detection/decoding process at destination node.
\begin{figure}[t]
\begin{center}
\epsfig{figure=F2.pdf,scale=0.55,angle=0}
\end{center}
\caption{Reliability-aware iterative decoding model.}\label{F2}
\end{figure}
With the obtained LLR values from \eqref{eq11} and \eqref{eq13}, the modified MAP detector now is able to separate the data frames sent from the relay and the source nodes, meanwhile, the error propagation effect can be eliminated by taking $p^{(l)}_{e},\forall l,$ into account. Here, destination node needs to wait all $L$ frames being detected by the modified MAP detector before the iterative process started. Then, it will add up two versions of LLR values per frame, e.g., the $(l-1)^{\mathrm{th}}$ frame sent from the relay node in the $l^{\mathrm{th}}$ time slot and the $(l-1)^{\mathrm{th}}$ frame sent from the source node in the $(l-1)^{\mathrm{th}}$ time slot, for the turbo-like decoder. The LLR values of the coded bits per frame, after the turbo decoder, will then be duplicated and fed into the \textit{a posteriori} probability converter to generate the \textit{a posteriori} probabilities for the next iteration. In the \textit{a posteriori} probability converter, the \textit{a posteriori} probabilities are found from the LLR values using the modulator mapping function. In addition, in Fig.~\ref{F2}, the blocks ``-1'' denote backward-shifting the frame number, ``+1'' denote forward-shifting the frame number, and $p^{(l)}_e,\forall l,$ denote the error probability parameter estimation.
\subsection{Error Probability Parameter Estimation}
As discussed in Section III-B, in order to mitigate the error propagation effect, the probabilities of decoding error at relay nodes $p^{(l)}_e,\forall l,$ should be taken into account in the iterative decoding process. These probabilities can be estimated at relay nodes and then be forwarded to destination node at the cost of signalling overhead. On the other hand, they can be directly estimated at destination node, where the estimation method will be introduced in this subsection.
As shown in Fig.~\ref{F2}, the error probability estimation process should be carried out in each of detection/decoding iterations. In detail, let's define $B\triangleq\log_{2}Q$ is the number of modulated bits per each constellation symbol sent from either source node or one of relay nodes. Then, the estimated probabilities can be given by
\begin{IEEEeqnarray}{ll}\label{eq15}
p^{(l)}_{e} = \frac{1}{MB}\sum^{M}_{m=1}\sum^{B}_{j=1}\mathrm{Pr}[c^{(l)}_{j,m}=+1|S]\mathrm{Pr}[c^{(l)}_{j,m}=-1|R]\nonumber\\
~~~~~~~~+\mathrm{Pr}[c^{(l)}_{j,m}=-1|S]\mathrm{Pr}[c^{(l)}_{j,m}=+1|R],~\forall l,
\end{IEEEeqnarray}
where $\mathrm{Pr}[c^{(l)}_{j,m}=\pm1|S]$ are the probabilities for the bits sent from the source node; $\mathrm{Pr}[c^{(l)}_{j,m}=\pm1|R]$ are the probabilities for the bits sent from the relay node. Then, \eqref{eq15} can be further transformed to
\begin{IEEEeqnarray}{ll}\label{eq16}
p^{(l)}_{e} = \frac{1}{MB}\sum^{M}_{m=1}\sum^{B}_{j=1}\frac{e^{L[c^{(l)}_{j,m}|S]}+e^{L[c^{(l)}_{j,m}|R]}}{(1+e^{L[c^{(l)}_{j,m}|S]})(1+e^{L[c^{(l)}_{j,m}|R]})},~\forall l,\nonumber\\
\end{IEEEeqnarray}
where $L[c^{(l)}_{j,m}|S]\triangleq\log\frac{\mathrm{Pr}[c^{(l)}_{j,m}=+1|S]}{\mathrm{Pr}[c^{(l)}_{j,m}=-1|S]}$ represents the LLR value of the modulated bit sent from the source node, which is derived from \eqref{eq11}; $L[c^{(l)}_{j,m}|R]\triangleq\log\frac{\mathrm{Pr}[c^{(l)}_{j,m}=+1|R]}{\mathrm{Pr}[c^{(l)}_{j,m}=-1|R]}$ represents the LLR value of the modulated bit sent from one of relay node, which is derived from \eqref{eq13}. After the error probabilities for all $L$ frames being estimated in each iteration, they will be fed back and used by the modified MAP detector in the next iteration. It is worth noting that the accuracy of the estimated probabilities can be improved with the iteration number.
\subsection{Computational Complexity of Iterative Decoding}
Based on the above analysis, the proposed iterative decoding process can be summarized as
\begin{center}
{\normalsize
\begin{tabular}{cp{8cm}}
& \textit{Reliability-Aware Iterative Decoding}\\ \hline
&I.~ Initialize \textit{A posteriori} probabilities and set $i=0$;\\
&II.~\textbf{While} $i\leq Iter$:\\
&~~~~~$1.$ Increasing $i$ by 1;\\
&~~~~~$2.$ Calculate LLRs for all $L$ frames based on \eqref{eq11}\\
&~~~~~~~~ and \eqref{eq13};\\
&~~~~~$3.$ Estimate $p^{(l)}_e,\forall l,$ based on \eqref{eq16};\\
&~~~~~$4.$ Combine two version of LLR values per frame\\
&~~~~~~~~ and then send them to turbo decoder;\\
&~~~~~$5.$ Send the LLRs of coded bits to \textit{A posteriori}\\
&~~~~~~~~ probability converter;\\
&~~~~~$6.$ Feed back the output of \textit{A posteriori} probabili-\\
&~~~~~~~~ -ties to modified MAP detector;\\
&III.~ \textbf{End While}\\
\hline
\end{tabular}}
\end{center}
where $Iter=5$ is the number of outer iterations. The computational complexity mainly comes from Step II-2 and Step II-4. Specifically, for Step II-2, modified MAP detector is implemented and its complexity mainly depends on the number of receive antenna and modulation size; for Step II-4, the complexity of standard turbo-like decoder should be taken into account. For the rest steps, like Step II-3 and Step II-5, the complexity just involves certain element-wise multiplication and additions, which can be ignored by comparing with Step II-2 and Step II-4.
\section{Simulation Results}
Computer simulations are used to evaluate the BER performances of our proposed multi-antenna assisted virtual FD relaying with reliability-aware iterative decoding method. We assume all channels are generated as independent block Rayleigh fading, which remain static over each transmission time slot.
For transmitting, source and relay nodes are using one transmit antennas, and for receiving, relay and destination nodes are with $N=2$ receive antennas. In addition, there are $L=20$ frames being transmitted via $L+1$ time slots, and each frame consists of $M=512$ information bits. SNR denotes the transmission signal power normalized by noise power ratio. The quadrature phase-shift keying (QPSK) modulation and turbo-like channel coding are implemented, e.g., 1/2 rate serial concatenated convolutional code for both source and relay nodes, where the first encoder is the non-recursive non-systematic convolutional code with a generator polynomial $G=([3,2])_8$, and the second encoder is the doped-accumulator with a doping rate equalling two. Random interleavers are implemented, and the results are computed on an average over 1000 independent channel realizations.
Four baselines are used for comparisons: 1) \textit{Perfect decoding at relay nodes}: this is served as performance bound, where we assume both relay nodes can correctly decode their received messages all the time and the same iterative decoding at destination node as our proposed scheme; 2) \textit{Proposed iterative decoding (ID) without $p_e$}: relay nodes always forward their decoded symbols regardless decoding errors and the iterative decoding at destination node does not take $p_e$ into account; 3) \textit{CRC based S-DF}: relay nodes only forward if their received messages can be correctly decoded and the same iterative decoding at destination node as our proposed scheme; 4) \textit{Threshold based S-DF}: relay nodes only forwards if their decoded errors are less than 15\% and the same iterative decoding at destination node as our proposed scheme. In addition, the QR-decomposition based detection method at relay nodes is implemented for all the schemes.
Fig.\ref{F3} shows BER performances for different schemes with relay-location A.
\begin{figure}[t]
\begin{center}
\epsfig{figure=F3.pdf,scale=0.53,angle=0}
\end{center}
\caption{BER versus S-D link SNR with the relay-location A in block Rayleigh fading channels and block-length=1024.}\label{F3}
\end{figure}
Let's define $d_{i,j}$ as the distance between the $i^{\mathrm{th}}$ and $j^{\mathrm{th}}$ nodes, and $d_{\mathrm{S,D}}=d$. Thus, we have $d_{\mathrm{S,R1}}=d_{\mathrm{S,R2}}=d_{\mathrm{R1,D}}=d_{\mathrm{R2,D}}=d$ for the relay-location A. As a consequence, based on a simplified suburban area pathloss model with the pathloss exponent equalling to 3.52, SNR relationship in dB among different links can be approximated by $\gamma_{\mathrm{S,D}}=\gamma_{\mathrm{S,R1}}=\gamma_{\mathrm{S,R2}}=\gamma_{\mathrm{R1,D}}=\gamma_{\mathrm{R2,D}}$. As shown in Fig.\ref{F3}, \textit{Perfect decoding at relay nodes} scheme undoubtedly gives the best BER performance. \textit{Proposed ID with $p_e$} scheme gives the second best BER performance due to the iterative decoding process at destination node by taking the error propagation effects into account. In contrast, \textit{Proposed ID without $p_e$} scheme gives the worst BER performance. The BER performances of two S-DF based schemes are in the middle between \textit{Proposed ID with $p_e$} scheme and \textit{Proposed ID without $p_e$} scheme. This is because the S-DF based schemes prefer to remain silent rather than propagate decoding errors if the number of decoding errors above their limits.
Fig.\ref{F4} gives BER performances for different schemes with a different relay location (e.g. relay-location B).
\begin{figure}[t]
\begin{center}
\epsfig{figure=F4.pdf,scale=0.53,angle=0}
\end{center}
\caption{BER versus S-D link SNR with relay-location B in block Rayleigh fading channels and block-length=1024.}\label{F4}
\end{figure}
In this case, given $d_{\mathrm{S,D}}=d$, we have $d_{\mathrm{S,R1}}=d_{\mathrm{S,R2}}=\frac{1}{2}d$ and $d_{\mathrm{R1,D}}=d_{\mathrm{R2,D}}=\frac{3}{4}d$. Then, similar as relay-location A, the corresponding SNR relationship in dB among different links can be approximated by $\gamma_{\mathrm{S,R1}}=\gamma_{\mathrm{S,R2}}=\gamma_{\mathrm{S,D}}+10.6~\mathrm{dB}$ and $\gamma_{\mathrm{R1,D}}=\gamma_{\mathrm{R2,D}}=\gamma_{\mathrm{S,D}}+4.4~\mathrm{dB}$. Such location is the case that relay nodes are close to source node. As shown in Fig.\ref{F4}, the same BER performance trend as Fig.\ref{F3} is presented, where \textit{Proposed ID with $p_e$} scheme gives the second best BER performance. In comparison with Fig.\ref{F3}, the BER performance gap between \textit{Proposed ID with $p_e$} scheme and the conventional S-DF based schemes in Fig.\ref{F4} is larger than the ones in Fig.\ref{F3}, where the performance gap in Fig.\ref{F4} exhibits up to 8 dB gain. This is because, with relay-location B, the decoding capability of relay nodes increases due to high S-R link quality. On the other hand, for the conventional S-DF based schemes, one or a few bits decoding errors will stop them forwarding their received frames to enjoy the spatial diversity gain. This is also why there is error floor for \textit{CRC based S-DF} scheme.
\section{Conclusion}
In this paper, a multi-antenna assisted virtual FD relaying with reliability-aware iterative decoding has been proposed. Based on QR decomposition, our proposed scheme can cancel the inter-relay interference without boosting noise power at relay nodes. In addition, our proposed scheme allows two relay nodes forward the erroneously decoded symbols and jointly cancels inter-frame interference and error propagation effect at destination node. It has been shown that, without extra cost of time delay and signalling overhead, our proposed scheme exhibits up to 8 dB gain by comparing with the conventional S-DF based schemes especially when relay nodes are close to source node.
\section*{Acknowledgement}
This work was supported by the Engineering and Physical Science Research Council (EPSRC) through the Scalable Full Duplex Dense Wireless Networks (SENSE) grant EP/P003486/1.
|
\section{Introduction}\label{INTRODUCTION}
Magic and mystery, in the broad sense, have always been embraced by lots of people, including scientists. This also holds for the
so-called mysteries in quantum mechanics. The central mystery of quantum mechanics, wave-particle duality (interference), and
other mysteries like the Schr\"odinger's cat paradox (superposition), the Einstein-Podolsky-Rosen paradox (entanglement), and
various others, including their application in quantum teleportation, quantum cryptography and quantum computation, do not only
fascinate many scientists but also seem to capture the public imagination, resulting in many popular books and publications.
These mysteries are foreign to quantum theory itself as this theory describes features of a collective of outcomes only.
They are a consequence from attempts to explain what happens to single objects in thought or laboratory experiments
as they do not appear in statistical quantities.
Unfortunately, resolving mysteries is not as popular as the mysteries themselves.
One reason for this might be that we humans simply like mysteries.
Another reason for mystery cultivation may be found in the quote by Einstein ``We can't solve problems by using the same kind of thinking we used
when we created them''. In other words, demystifying these so-called quantum experiments requires some out of the box thinking,
that is one might have to leave one's comfort zone.
Recently, a new quantum mystery called quantum Cheshire Cat, has grabbed a lot of media attention.
The quantum Cheshire Cat displays mysterious behavior similar to that of the grinning Cheshire Cat
in the novel ``Alice's Adventures in Wonderland'' by Lewis Carroll~\cite{CARR65}.
In this children's story the Cheshire Cat is able to slowly vanish beginning with the end of its tail,
and ending with its grin, which remains for some time after the
rest of the cat has disappeared.
The scientific community has embraced the Cheshire Cat as a metaphor to explain several scientific phenomena.
The ``classical'' optical Cheshire Cat effect has been demonstrated in
an experiment with a mirror stereoscope allowing one eye of the viewer to see a person's face in front of a white
background and the other eye to see a solid white background~\cite{DUEN79}.
If the viewer looks at the smile of the face while a
hand makes a slow sweeping motion in the visual field of the other eye,
the motion causes the face to completely disappear leaving only the smile.
This ``non-quantum'' Cheshire Cat effect is an optical illusion caused by binocular rivalry~\cite{DUEN79},
a visual phenomenon in which perception alternates between different images presented to each eye, such that
motion in the field of one eye can trigger suppression of
the other visual field as a whole or in parts~\cite{GRIN65,DUEN79}.
The quantum Cheshire Cat was introduced by Aharonov {\sl et al.} in the form of a circularly polarized photon, whereby the photon
represents the cat and its polarization state the grin~\cite{AHAR13}.
Aharonov {\sl et al.} showed analytically that in a pre- and post-selected
experiment measuring weak values for the location of the photon and its polarization, the polarization can be disembodied from the
photon~\cite{AHAR13}.
Recently, Aharonov {\sl et al.} presented a dynamical analysis of weak values, thereby suggesting a dynamical
process through which the Quantum Cheshire Cat effect occurs~\cite{AHAR15}.
According to Aharonov {\sl et al.} the quantum Cheshire Cat effect is quite general, that is physical properties can
be disembodied from the objects they belong to in a pre- and postselected experiment~\cite{AHAR13}.
Soon after the introduction of the quantum Cheshire Cat by Aharonov {\sl et al.} proposals for more sorts of quantum Cheshire Cats
made their appearance in the literature~\cite{MATZ13,IBNO12,GURY12,YU14}.
Denkmayr {\sl et al.} performed weak measurements to probe the location of a neutron and its magnetic
moment ($z$-component only) in a neutron interferometry experiment to demonstrate the quantum Cheshire Cat effect~\cite{DENK14,HASE14}.
In Refs.~\onlinecite{DENK14,HASE15} Hasegawa and co-workers give various interpretations of their experimental observations and point
out that weak interactions between the probe and the neutron and its magnetic moment have observational effects on average so that it seems
as if the neutron and its magnetic moment are spatially separated.
These interpretations, not the outcome of the experiment itself, have been criticised on various grounds~\cite{DILO14,STUC14,ATHE15,CORR15}.
In this work, we provide
a mystery-free explanation for the experimentally observed facts in terms of a discrete-event simulation (DES) model
which accurately reproduces the data of the neutron experiments~\cite{DENK14}.
In other words, we offer a straightforward interpretation of the neutron interferometry experiment with no need to invoke
a quixotic ``quantum Cheshire Cat effect''.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\hsize]{figure1}
\caption{(color online)
Schematic picture of the single-neutron interferometry experiment for observing a quantum Cheshire Cat~\cite{DENK14}.
A source S emits a monochromatic neutron beam. A magnetic birefringent prism P produces from this beam two spatially separated beams of
polarized neutrons.
Polarized neutrons with their magnetic moments aligned antiparallel to a magnetic guide field $B^z$ (following the dash-dotted line)
are not considered in the experiment.
Polarized neutrons with their magnetic moments aligned parallel to $B^z$ first enter a spin turner
(ST$_1$), which rotates the magnetic moment by $\alpha=\pi/2$ about the $y$-axis before they enter a triple-Laue
interferometer~\cite{RAUC00}. BS0, $\ldots$, BS3: beam splitters; neutrons that are transmitted by BS1 or BS2 leave the
interferometer (following the dashed lines); SR$_1$ and SR$_2$: spin rotators for rotating the magnetic moment about the $z$-axis by
$\mu_1=0$ and $\mu_2=\pi$, respectively; phase shifter $\chi$: aluminum foil; ABS$_1$ and ABS$_2$: absorbers which can be placed
in path 1 and path 2 (indicated by the dotted lines), respectively; $B_1^z$ and $B_2^z$: weak additional magnetic fields which
can be applied in path 1 and path 2 (indicated by the dotted lines) for rotating the magnetic moment about the
$z$-axis by $\theta_1$ and $\theta_2$, respectively. For the purpose of postselection a spin turner ST$_2$ rotating the magnetic
moments of the neutrons by $\beta=\pi/2$ about the $y$-axis and a spin analyzer A is put in the O-beam~\cite{DENK14}.
For postselection in the H-beam (not performed in~\cite{DENK14}), a spin turner ST$_2$ and spin analyzer A could be put in the H-beam
(indicated by the dotted lines).
Detectors count the number of neutrons in the O- and H-beam.
}
\label{fig1}
\end{center}
\end{figure}
\section{Neutron experiment}\label{EXP}
Figure~\ref{fig1} shows a schematic picture of the single-neutron interferometry experiment~\cite{DENK14}.
A monochromatic neutron beam emitted by a source S passes through a magnetic birefringent prism P which produces two spatially separated beams
of polarized neutrons with their magnetic moments aligned parallel, respectively anti-parallel with respect to the magnetic axis
of the polarizer which is parallel to the magnetic guiding field ${\bf B}$ oriented along the $z$-axis and pointing in the $+z$-direction. The anti-parallel
polarized neutron beam (following the dash-dotted line) plays no role in the experiment. The parallel polarized neutron beam enters a spin turner (ST$_1$) which
rotates the magnetic moment of the neutrons by $\alpha=\pi/2$ about the $y$-axis such that they become aligned along the $x$-axis in the $+x$-direction.
On leaving the spin turner, the neutron beam impinges on a triple-Laue diffraction type silicon perfect single crystal interferometer~\cite{RAUC00}.
Laue diffraction on the first perfect crystal slab (beam splitter BS0) coherently splits the neutron beam
in a beam following path 1 (called the transmitted beam) and one following path 2 (called the reflected beam). Behind beam splitter BS0 and in front of beam
splitters BS1 and BS2, spin rotator SR$_1$ (SR$_2$) in path 1 (2) rotates the magnetic moment of the neutrons by $\mu_1=0$
($\mu_2=\pi$) about the $z$-axis so that they are aligned along the $x$-axis in the $+x$ ($-x$) direction.
This corresponds to the preselection process of the weak measurement procedure~\cite{DENK14}.
Neutrons that are transmitted by beam
splitters BS1 and BS2 (following the dashed lines) are not considered any further. Behind BS1 and BS2 absorbers ABS$_1$ and ABS$_2$ with
transmissivity $T_1=T_2=0.79$ can be inserted or additional magnetic fields $B_1^\mathrm{z}$ and $B_2^\mathrm{z}$ rotating the neutrons' magnetic
moments by $\theta_1=\theta_2=20^{\circ}$ can be applied for the weak measurement of the location of the neutrons or their magnetic
moments, respectively. These parameter choices fulfill the condition of a weak measurement, the idea being
that due to the weakness of the local coupling between the system and the measurement device, a probe, the subsequent
evolution of the system is not significantly altered~\cite{DENK14}. A rotatable-plate
phase shifter (e.~g. aluminum foil~\cite{RAUC00}) in front of beam splitter BS3 tunes the
relative phase $\chi$ between path 1 and path 2. BS3 takes as input the two neutron beams following path 1 and
path 2 and produces two output beams called the O-beam and H-beam. Neutrons in the O- and H-beam are immediately
detected or they undergo a postselection process depending on their magnetic moments. In the postselection process neutrons first
pass through spin turner ST$_2$, rotating the magnetic moment by $\beta=\pi/2$ about the $y$-axis, and a spin
analyzer A, selecting neutrons with their magnetic moments parallel to the guiding field, before being detected. In the
experiment by Denkmayr {\sl et al.} the neutrons in the H-beam are always detected without postselection and those in the O-beam
always with postselection. The neutron detectors in the O- and H-beam have a detection efficiency over $99\%$~\cite{RAUC00}. We refer to the
interferometer without absorbers ABS$_1$ and ABS$_2$ and extra magnetic fields $B_1^\mathrm{z}$ and $B_2^\mathrm{z}$ as the ``reference
interferometer''. Very important is that the neutron interferometry experiments are
performed under the condition that there is at most one neutron in the interferometer while producing,
after many single neutron passages through the interferometer, the same interference patterns as if a beam of neutrons would
have been used~\cite{RAUC00}.
\section{Quantum theory}\label{THEORY}
In Appendix A, we first give the quantum theoretical expressions for the probabilities
$P_\mathrm{H}(\chi,\theta_1,\theta_2,T_1,T_2)$ and $P_\mathrm{O}(\chi,\theta_1,\theta_2,T_1,T_2)$ for a neutron to trigger the
(ideal) H- or O-detector (see Eqs.~(\ref{app2a}) and (\ref{app2b})),
respectively, for the case that a spin turner and spin analyzer are present in the O-beam only (experimental setup~\cite{DENK14}).
Second, we give the corresponding expressions for the probabilities $\widetilde P_\mathrm{H}$ and $\widetilde P_\mathrm{O}$
(see Eqs.~(\ref{app2e}) and (\ref{app2f})) for the case that a spin turner
and spin analyzer are placed in the H-beam only. The probabilities for the two other cases, that is detection without spin turner and spin analyzer and
detection with spin turners and spin analyzers in both the O- and H-beam are then given by $\widetilde P_\mathrm{O}$, $P_\mathrm{H}$ and
$P_\mathrm{O}$, $\widetilde P_\mathrm{H}$, respectively,
The only parameter entering in this description is
the reflectivity $R$ of the beam splitter (the four beam splitters are assumed to be identical).
By fitting the quantum theoretical prediction for the
empty interferometer ($\mu_1=\mu_2=0$, $T_1=T_2=1$, $\theta_1=\theta_2=0$ and no postselection) to the experimental data, we
obtain $R=0.22$ (see Appendix B).
In the reference interferometer ($\mu_1=0,\mu_2=\pi$, $T_1=T_2=1$,
$\theta_1=\theta_2=0$ and postselection in the O-beam)
the neutron beams following path 1 and path 2 have orthogonal magnetic polarization
(the magnetic moments of the neutrons following path 1 are oriented in the $+x$-direction and those of the neutrons
following path 2 in the $-x$-direction)
and hence the probabilities $P_\mathrm{O}(\chi,0,0,1,1)$ and $P_\mathrm{H}(\chi,0,0,1,1)$ show no
dependence on $\chi$. In what follows, we use $P_\mathrm{O}(\chi,0,0,1,1)$ and $P_\mathrm{H}(\chi,0,0,1,1)$ as reference values for
comparison with the probabilities for interferometer configurations with absorbers or weak magnetic fields present.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\hsize]{figure2}
\caption{(color online)
Quantum theoretical results for the neutron Cheshire Cat experiment of Denkmayr {\sl et al.}~\cite{DENK14}
Solid lines: normalized probabilities
$P_\mathrm{H}(\chi)=P_\mathrm{H}(\chi,\theta_1,\theta_2,T_1,T_2)/P_\mathrm{H}(0,0,0,1,1)$ and
$P_\mathrm{O}(\chi)=P_\mathrm{O}(\chi,\theta_1,\theta_2,T_1,T_2)/P_\mathrm{O}(0,0,0,1,1)$
(see Appendix A, Eqs.~(\ref{app2a}) and (\ref{app2b}))
for observing a neutron in the H- or O-beam as a function of the phase shift $\chi$
and for the various experimental conditions with a spin turner and spin analyzer in the O-beam only~\cite{DENK14}.
Dashed lines: normalized probabilities
$P_\mathrm{O}(\chi)= P_\mathrm{O}(\chi,\theta_1,\theta_2,T_1,T_2)/ P_\mathrm{O}(0,0,0,1,1)$ and
$P_\mathrm{H}(\chi)=\widetilde P_\mathrm{H}(\chi,\theta_1,\theta_2,T_1,T_2)/\widetilde P_\mathrm{H}(0,0,0,1,1)$
(see Appendix A, Eqs.~(\ref{app2b}) and (\ref{app2e}))
for observing a neutron in the H- or O-beam as a function of the phase shift $\chi$
in an, as yet unperformed, experiment with a spin turner and spin analyzer in both the O- and the H-beam.
}
\label{fig2}
\end{center}
\end{figure}
The quantum theoretical predictions are presented in Fig.~\ref{fig2}.
The solid lines in the upper two rows of panels clearly show that placing an absorber in path 1 has no effect on the
probability of a neutron triggering the O-detector ($P_\mathrm{O}(\chi,0,,0.79,1)=P_\mathrm{O}(\chi,0,0,1,1)$), while placing the same absorber in
path 2 leads to a reduction of this probability compared to the one of the reference interferometer
($P_\mathrm{O}(\chi,0,0,1,0.79)<P_\mathrm{O}(\chi,0,0,1,1)$), in concert with the O-beam intensities measured in experiment~\cite{DENK14}.
Adopting the reasoning of Denkmayr {\sl et al.}, it seems as if the neutrons follow path 2 in the interferometer.
For the respective probabilities for a neutron to trigger the H-detector
we find that $P_\mathrm{H}(\chi,0,0,0.79,1)<P_\mathrm{H}(\chi,0,0,1,1)$ and $P_\mathrm{H}(\chi,0,0,1,0.79)\lesssim P_\mathrm{H}(\chi,0,0,1,1)$.
If we were to adopt the same reasoning to the H-detector data,
then the conclusion would be that most of the neutrons follow path 1 and only some follow path 2,
Obviously, this reasoning leads to a picture that is self-contradictory.
However, in the H-beam the neutrons are not postselected whereas successful
postselection is a necessity for the picture to hold~\cite{DENK14}.
Following Denkmayr {\sl et al.},
which-way information about the magnetic moment of the neutrons can be obtained by replacing the absorbers by magnetic fields
rotating the magnetic moment by a small angle about the $z$-axis. A magnetic
field in one of the paths ensures that the magnetic moment of the neutrons traveling path 1 and path 2 are no longer orthogonal.
implying that it is possible to observe interference.
As seen from the lower two rows of panels Fig.~\ref{fig2} (solid lines), a small magnetic field rotating the
magnetic moment by 20$^{\circ}$ in path 1 leads to a variation of both $P_\mathrm{O}(\chi,\pi/9,0,1,1)$ and
$P_\mathrm{H}(\chi,\pi/9,0,1,1)$ with $\chi$.
A small magnetic field in path 2 instead of path 1 leads to a variation of
$P_\mathrm{H}(\chi,0,\pi/9,1,1)$ with $\chi$ only. The probability $P_\mathrm{O}(\chi,0,\pi/9,1,1)<P_\mathrm{O}(\chi,0,0,1,1)$ shows no variation with $\chi$.
The experimental findings reported in Ref.~\cite{DENK14} show similar features.
Based on the significant changes in the intensity pattern recorded by the O-detector
Denkmayr {\sl et al.} argue that the magnetic moments of the neutrons follow path 1~\cite{DENK14}.
If we were to apply the same reasoning to the H-detector data, then this conclusion cannot be drawn since
a periodic variation of the intensity pattern is observed for both cases.
In other words, the H-detector data would suggest a picture in which the magnetic moments ($z$-components only) of the neutrons follow path 1 and/or path 2.
As in the case of the weak measurement of the neutron location, the picture that emerges is self-contradictory,
but again, in experiment no postselection is performed in the H-beam~\cite{DENK14}.
In Ref.~\cite{DENK14}, it is argued that only if the ensemble is successfully postselected, the magnetic
moments of the neutrons travel along path 1.
Therefore, we consider theoretically,
the experiment with postselection performed in both the O- and H-beam (see dashed lines in Fig.~\ref{fig2}).
With the absorbers in path 1 and path 2, we have (see Appendix A)
${P}_\mathrm{O}(\chi,0,0,0.79,1)={P}_\mathrm{O}(\chi,0,0,1,1)$, ${P}_\mathrm{O}(\chi,0,0,1,0.79)<{P}_\mathrm{O}(\chi,0,0,1,1)$,
${\widetilde P}_\mathrm{H}(\chi,0,0,0.79,1)={\widetilde P}_\mathrm{H}(\chi,0,0,1,1)$,
and ${\widetilde P}_\mathrm{H}(\chi,0,0,1,0.79)<{\widetilde P}_\mathrm{H}(\chi,0,0,1,1)$, leading to the consistent picture that neutrons follow path 2.
Placing the small magnetic field in path 2 does not lead to a variation of both ${P}_\mathrm{O}(\chi,0,\pi/9,1,1)$ and
${\widetilde P}_\mathrm{H}(\chi,0,\pi/9,1,1)$ with $\chi$, but placed in path 1
the same magnetic field leads to a variation of both ${P}_\mathrm{O}(\chi,\pi/9,0,1,1)$ and
${\widetilde P}_\mathrm{H}(\chi,\pi/9,0,1,1)$ with $\chi$. This supports the picture that the magnetic moments of the neutrons follow path 1.
Hence, if the experiment is performed with postselection in both the O- and H-beam, then
the picture that neutrons follow path 2 and that the $z$-components of their magnetic moments
follow path 1 in the interferometer is at least consistent.
The experimental observations (limited to the postselected O-beam data) and the rigorous quantum theoretical analysis given here
suggest that the neutrons behave as if they were Cheshire Cats. Note that this picture also holds if an absorber is
put in one path and the additional magnetic field in the other and if both an absorber and additional magnetic field
are put together in one of the paths, as can be seen from the formulas presented in Appendix A.
In other words, the picture inferred from separate measurements of the path taken by the neutrons and of the path taken
by their magnetic moments (five experiments including the one with the reference interferometer)
still hold when these measurements are performed at once, that is by placing an absorber in one path and
a weak magnetic field in the other or by placing both an absorber and additional magnetic field
together in one of the paths (three measurements including the one with the reference interferometer).
Therefore, the picture that the neutrons and their magnetic moments take different paths in the interferometer
is not a paradox of counterfactual reasoning~\cite{AHAR13}. Thus, following Aharonov {\sl et al.}, it may seem that in the interferometer
the $z$-component of the magnetic moment really becomes disembodied from the neutron. This by itself is quite
mysterious and requires a rational explanation. However, also the observation that the ensemble needs to be successfully
postselected in order to make the conclusion that neutrons and their magnetic moments take different paths in the interferometer
is mysterious. How can it be that an analysis performed on the neutron's magnetic moment {\bf after} the neutron and its
magnetic moment have gone through the interferometer has an influence on how the neutron and its magnetic moment travel through
the interferometer? In other words, how is it possible that the future influences the past? This phenomenon reminds of other
quantum mysteries like Wheeler's delayed choice and quantum erasure experiments~\cite{WHEE83,SCUL82}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\hsize]{figure3}
\caption{(color online)
Artistic impression of a quantum Cheshire Cat passing through a triple Laue interferometer.
In contrast to travelling through a Mach-Zehnder interferometer~\cite{DENK14}, the cat and its grin do not only take different paths but
the cat also loses its tail and half of its grin. As a result what arrives at the detectors are not Cheshire Cats anymore.
}
\label{splittedcat}
\end{center}
\end{figure}
Apart from these mysteries which require rational explanations there are various flaws in the quantum Cheshire Cat story.
The metaphor of a cat without grin and its disembodied grin taking different paths in the interferometer
is working nicely for a Mach-Zehnder interferometer with the phase shifter adjusted such that only one of the detectors click~\cite{DENK14}.
However, this metaphor is way too simple for the triple Laue interferometer used in the single neutron interferometry experiments.
A Cheshire Cat representing a neutron, enters the interferometer from the left and is ``split'' by beam splitter BS0
in two parts, e.g. in the cat without grin and its grin (see Fig.~\ref{splittedcat} for an artistic impression).
The grin follows path 1 and the cat without grin follows path 2.
Each of these two parts is split in two again at BS1 and BS2 giving four parts, e.g. two halves of a grin, a tail and the cat without grin and without tail.
At BS1 the left half of the grin leaves the interferometer and at BS2 the tail leaves the interferometer.
The right half of the grin and the cat without grin and without tail reunite at beam splitter BS3.
Beam splitter BS3 splits on its turn the cat with the right half of the grin and without tail in two parts, e.g. in its head and its body.
The intriguing question that arises is how the detectors that are designed to detect neutrons or their metaphors, the Cheshire Cats, can produce a
click if only part of the neutron or Cheshire Cat arrives at the detector.
The answer is that they simply cannot.
Hence, the story needs to be modified so that the detectors encounter complete Cheshire Cats.
One option could be to use the concept of self-interference whereby each quantum Cheshire Cat behaves as if it ``explores''
all the possible paths.
However, self-interference cannot act as the lifeguard of the quantum Cheshire Cat effect, and this for two reasons.
First, two parts of the quantum Cheshire Cat are split-off, namely (in the story above) the left part of its smile at BS1 and its tail at BS2,
and never reunite with the other remainings of the Cheshire Cat.
Hence, even with the hypothesis of self-interference, the detectors never encounter a Cheshire Cat and therefore they never click.
Second, if the Cheshire Cats behave as if they explored both paths to self-interfere, then it does not make sense at all to speak
about the Cheshire Cat effect since the Cheshire Cats never behave as if their smiles and the rest of their body
take different paths.
Alternatively, one could think of a description in terms of a collection (distribution) of quantum Cheshire Cats.
Adopting this view never leads to harmed cats since at the beam splitters parts of the distribution (a certain number of
quantum Cheshire Cats, not the cats themselves) split off.
Then it does not make sense to speak about the Cheshire Cat effect at all as the Cheshire Cats and their smiles never separate.
This ensemble/statistical interpretation of quantum theory is free of logical contradictions and mysterious elements~\cite{BALL70}.
However, it does not contain the elements to explain how, in the experiment of Denkmayr et al.~\cite{DENK14}, neutrons pass through the interferometer
and build up an interference pattern one neutron at a time (disregarding extremely rare events).
Hence, the picture (but not the description in terms) of a distribution of neutrons traveling through space seems mysterious as well.
\section{Weak measurement}\label{WEAK}
We now scrutinize the more quantitative analysis by Denkmayr {\sl et al.} of the conclusions drawn above,
which is based on the calculation of the so-called weak values. Weak values of quantum variables, obtained from weak
measurements, have been introduced in 1988 by Aharonov, Albert and Vaidman~\cite{AHAR88} in order to gain more information
about a quantum system than by performing ordinary measurements. The average outcome of a conventional
quantum measurement of any operator $O$ of a quantum system in the state $|\psi\rangle$ is given by its expectation
value, that is $\langle O \rangle = \langle\psi |O|\psi\rangle$. In a weak measurement scheme, wherein the measured system is
very weakly coupled to the measuring device, the probe, two states for a single system at a given time are involved, namely a preselected
state $|\phi\rangle$ and a postselected state $|\psi\rangle$. The underlying idea is that weak enough measurements do not
disturb these two states and that the outcomes of such measurements reflect properties of both states. The outcome of a
weak measurement of the operator $O$ is then defined as $\langle O \rangle_\mathrm{w} = \langle\phi |O|\psi\rangle / \langle\phi
|\psi\rangle$. For discussions of the application, the interpretation, the problems and paradoxes created by
the theory of weak measurement see Refs.~\cite{AHAR08,KOFM12,SVEN13,SVEN13a,TAMI13,DRES14,FERR14} and references therein.
The experiment performed by Denkmayr {\sl et al.} involves pre- and postselection and is a specific
implementation of a general measurement strategy known as weak measurements~\cite{AHAR88,LEGG89,PERE89}.
Denkmayr {\sl et al.} perform weak-value measurements of the paths taken by the neutron
(yielding weak values $\langle \Pi _1\rangle_\mathrm{w}$ and
$\langle \Pi _2\rangle_\mathrm{w}$ obtained for $T_1 =0.79$, $\theta_1=\theta_2=0$ and $T_2=0.79$, $\theta_1=\theta_2=0$, respectively)
and the paths taken by its magnetic moment (yielding weak values $\langle \sigma_z\Pi_1\rangle_\mathrm{w}$ and $\langle
\sigma_z\Pi_1\rangle_\mathrm{w}$ obtained for $\theta_1= 20^{\circ}$, $T_1=T_2=1$ and $\theta_2= 20^{\circ}$, $T_1=T_2=1$, respectively).
From the quantum theoretical description of the experiment in terms of the preselected state $|\phi\rangle$ and the postselected state
$|\psi\rangle$ (see Ref.~\cite{DENK14}), thereby ignoring the effect of the absorbers and additional magnetic fields on the
measurement outcome, it follows that $\langle \Pi _1\rangle_\mathrm{w}=0$, $\langle \Pi _2\rangle_\mathrm{w}=1$,
$\langle \sigma_z\Pi_1\rangle_\mathrm{w}=1$ and $\langle \sigma_z\Pi_2\rangle_\mathrm{w}=0$.
These weak values are then interpreted as if neutrons follow path 2 and their magnetic
moments follow path 1~\cite{DENK14}.
In Appendix C (Eqs.~(\ref{app4a})-(\ref{app4})), we derive expressions for these four weak values based on the standard quantum theoretical
description of the experimental setup including absorbers and additional magnetic fields and the definitions of the weak
values given in Ref.~\cite{DENK14}. Here we only discuss the case with postselection in the O-beam, corresponding
to the experimental configuration of Denkmayr {\sl et al.} We find that $\langle \Pi _1\rangle_\mathrm{w}=0$ and that $\langle \Pi
_2\rangle_\mathrm{w}=(\sqrt{T_2}+1)/2=0.94$. Note that these are the values with which the experimental outcomes $\langle \Pi
_1\rangle_\mathrm{w}=0.14$ and $\langle \Pi _2\rangle_\mathrm{w}=0.96$ should be compared, not with the ideal values 0 and 1 as derived from
the idealized weak measurement setup. In other words, the deviation of the weak values of 0 and 1 is not due to systematic
misalignments in the experiment, as argued in Ref.~\cite{DENK14}. The deviations are inherent to the measurement setup and would
also show up in an ideal experiment (which we assume is described by quantum theory).
However, based on this result one could still argue that in case of
an ideal weak measurement ($T_2\rightarrow 1$), the weak values for the neutron population along path 1 and 2 are $\langle
\Pi _1\rangle_\mathrm{w}=0$ and $\langle \Pi _2\rangle_\mathrm{w}=1$, respectively,
which correspond to the predictions of idealized weak-measurement theory.
For the weak value of the magnetic-moment population in path 2 we find $|\langle
\sigma_z\Pi_2\rangle_\mathrm{w}|^2=0$, to which the reported experimental value of 0.02 should be compared.
For path 1 we find that
$|\langle \sigma_z\Pi_1\rangle_\mathrm{w}|^2=1+2\sin\chi/\sin(\theta_1/2)=1+11.52\sin\chi$
whereas the quantum theoretical result for the idealized weak measurement setup is independent of $\chi$ and
reads $|\langle \sigma_z\Pi_1\rangle_\mathrm{w}|^2=1$~\cite{DENK14}.
In the analysis of the experimental data, without supporting argument, only intensity values for $\chi=0$ are
considered for calculating the weak values~\cite{DENK14}, in which case we find $|\langle\sigma_z\Pi_1\rangle_\mathrm{w}|^2=1$
in agreement with the idealized weak-measurement theory.
The experimentally obtained value is $|\langle \sigma_z\Pi_1\rangle_\mathrm{w}|^2=1.07$,
in good agreement with the theoretically derived value of 1 for $\chi=0$.
However, taking $\chi=-\pi/2$ theory predicts that $|\langle \sigma_z\Pi_1\rangle_\mathrm{w}|^2=-10.52$.
This large negative value indicates that this implementation of weak measurement theory cannot have any significance.
From what has been shown here it is clear that the quantitative interpretation of the experiment of Denkmayr {\sl et al.}
in terms of weak values cannot unambiguously contribute to ``explaining'' the mystery of the quantum Cheshire Cat effect.
In contrast to the qualitative analysis for which including postselection in the H-beam removed contradictory conclusions,
the extra postselection in the H-beam does not help to come to a clear (definite) conclusion based on weak values (see Appendix B).
\section{Discrete event simulation}\label{DES}
The analysis of the experimental data shows that neither quantum theory nor weak measurement theory are helpful in
unravelling the mystery of the quantum Cheshire Cat. Hence, a question that arises is: ``Does there exist another,
non-mysterious, interpretation of the observed interference patterns?'' In other words, is it possible to give another, but
rational, explanation for the interference patterns than the neutrons and the $z$-components of their magnetic moments
taking different paths in the interferometer? If so,
the quantum Cheshire Cat is nothing else than an illusion. Somehow, the situation is similar to being a viewer who watches a
magician performing a magic trick and instead of just experiencing the magic and being amazed starts wondering how the
trick was done.
Since the magician is unlikely to explain how the trick is done while it is being performed, the viewer, who has
limited information about the trick, can, in order to find an explanation, come up with any kind of moves as
long as the end result is creating the same illusion as the magic trick.
Even the magician may be unaware that the moves required to perform the trick are not unique.
In what follows, we employ discrete event simulation (DES)
to unravel the mystery of the quantum Cheshire Cat in the neutron interferometry experiment.
DES is a general form of computer-based modeling that provides a flexible approach to represent the behavior of complex systems in terms
of a sequence of well-defined events, that is operations
being performed on entities of certain types.
The entities are passive (in contrast to agents in agent based modeling),
but can have attributes that affect the way they are handled or may change as the entity flows through the process.
Typically, many details about the entities are ignored.
DES is used in a wide range of health care, manufacturing, logistics, science and engineering applications.
We use DES to construct an event-based model that reproduces the statistical distributions of quantum theory
by modeling physical phenomena as a chronological sequence of events whereby events can be the actions of an experimenter,
particle emissions by a source, signal generations by a detector, interactions of a particle with a material and so on.
The general idea is that simple rules define
discrete-event processes which may lead to the behavior that is observed in experiments, all this
without making use of the quantum theoretical prediction of the collective outcome of many events.
Evidently, mainly because of insufficient knowledge, the rules are not unique. Hence, the simplest
rules one could think of can be used until a new experiment indicates otherwise.
Reviews of the method and its application to single-photon experiments and
single-neutron interferometry experiments can be found in Refs.~\cite{RAED12a,RAED12b,MICH14a}.
A DES of the experiment of Denkmayr {\sl et al.} requires rules for the neutrons and for the various units in the
diagram (see Fig.~\ref{fig1}) representing the neutron interferometry experiment.
We regard a neutron as a messenger (called entity in DES) carrying a message (called attribute in DES). From experiments
we know that a neutron has a magnetic moment and that it moves
from one point in space to another within a certain time period, the time of flight.
Hence, we encode both the magnetic moment and the time of flight in the
message. For the technical details we refer to Ref.~\cite{RAED12b}.
The neutron source creates messengers one-by-one. The source waits until the messenger's message has
been processed by one of the detectors before creating the next messenger. Hence, the messengers cannot directly
communicate, but only indirectly through the units in the diagram (see Fig.~\ref{fig1}).
The messengers interact with the various units representing the beam splitters, the spin turners,
spin rotators, absorbers, magnetic fields, phase shifters and spin analyzers.
Each of these units interpret, and eventually process and change (part of) the message carried by the messengers.
The specific simple rules that each of these units use to emulate their real-world behavior for many neutrons passing
through the unit is given in Refs.~\cite{RAED12a,RAED12b,MICH14a}.
Finally, the messengers trigger one of the detectors in the O- or H-beam.
These detectors count all incoming messengers and hence have a detection efficiency of 100\%.
This is an idealization of real neutron detectors which can have a detection efficiency of 99\% and more~\cite{RAUC00}.
Upon detection the neutron is destroyed.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\hsize]{figure4}
\caption{(color online)
Comparison between the experimental data (open circles with error bars), kindly provided to us by T. Denkmayr,
and event-based simulation data (solid squares) of the neutron Cheshire Cat experiment~\cite{DENK14}.
The transmissivity for the absorbers is $T_1=T_2=0.79$, the reflectivity of the beam splitters is $R=0.22$.
More detailed information about the simulation parameters is given in Appendix D.
}
\label{fig3}
\end{center}
\end{figure}
In Fig.~\ref{fig3} we present a comparison of the experimental and event-based simulation data of the neutron Cheshire Cat experiment
(postselection in the O-beam only).
In the simulation we have taken into account that a fraction of neutrons is lost in the O-beam due to the spin analysis procedure
and that the transmissivity of the absorbers in path 1 and path 2 might be slightly different due to neutron scattering at the absorbers
(T. Denkmayr, private communication). Detailed information about the simulation parameters is given in Appendix D.
Taken these experimental details into account, the agreement between experimental and simulation data is excellent:
if one were to give the experimental and DES data sets to a third party for analysis
then it would be very hard, if not impossible, to distinguish between the two.
Hence, applying the same qualitative and quantitative analysis as described in Ref.~\cite{DENK14}
to the DES data obviously leads to the same mystery of neutrons behaving as quantum Cheshire Cats.
However, in the DES we know exactly what the messengers (the neutrons) do
in the interferometer: a neutron and its magnetic moment never separate and a neutron follows a definite trajectory, that is it follows
either path 1 or path 2 in the interferometer.
This is illustrated in Fig.~\ref{CAT} (see Appendix E) where we explicitly
show how many of the neutrons counted by the O- and H-beam detectors have been following path 1 and path 2.
From the O-beam data, it is clear that the contributions from path 1 and path 2 to the total count are roughly the same.
This is different for the H-beam data, due to the fact that in the H-beam no postselection on the basis of the magnetic moment
is performed.
Performing postselection in the H-beam as well leads to the same conclusion as for the O-beam, as can be seen from Fig.~\ref{CAT4}
(see Appendix E).
In other words, in the DES the neutron and its magnetic moment never separate and each neutron follows either path 1 or path 2.
Both figures also include the interference patterns obtained from quantum theory (solid lines), that
is the interference patterns for the ideal experiment. It is clearly seen that the neutron counts
obtained with the event-based simulation correspond very well to these quantum theoretical results.
\section{Conclusions}\label{CONCLUSIONS}
We have demonstrated that there exists an accurate description by discrete event simulation, free of paradoxes,
in which many neutrons (together with their magnetic moment)
travel one-by-one through the interferometer thereby taking only one path or the other, that yields the same interference patterns as those
observed in the experiment.
Since no real which-way information can be obtained from the experiment one has the choice to adopt one or the
other scenario.
One can adopt the mysterious description, quantified in terms of weak values, of the experiment with the neutrons acting as quantum Cheshire Cats whereby the neutrons
seem to travel different paths as the $z$-components of their magnetic moments
or one can adopt the rational description with the neutrons together with their magnetic moment simply taking one path or the other.
Hence, although the quantum Cheshire Cat is not a paradox of counterfactual reasoning, in contrast to the statement made by Aharonov {\sl et al.}
that there really is a quantum Cheshire Cat~\cite{AHAR13}, the Cheshire Cat is nothing else than an illusion.
It remains to be seen whether the alleged applications in
precision measurements~\cite{AHAR13, DENK14} and quantum information technology~\cite{DENK14} will be more than
an illusion.
\section*{Acknowledgement}
We thank T. Denkmayr and Y. Hasegawa for providing us with their neutron interferometry data
and for clarifying intricate aspects of their experiments.
|
\section{Introduction}
\emph{Path-pairability} is a graph theoretical notion that emerged from a practical networking problem introduced by Csaba, Faudree, Gy\'arf\'as, Lehel, and Schelp \cite{CS}, and further studied by Faudree, Gy\'arf\'as, and Lehel \cite{mpp,F,pp} and by Kubicka, Kubicki and Lehel \cite{grid}. Given a fixed integer $k$ and a simple undirected graph $G$ on at least $2k$ vertices, we say that $G$ is {\it $k$-path-pairable} if, for any pair of disjoint sets of distinct vertices $\{x_1,\dots,x_k\}$ and $\{y_1,\dots,y_k\}$ of $G$, there exist $k$ edge-disjoint paths $P_1,P_2,\dots,P_k$, such that $P_i$ is a path from $x_i$ to $y_i$, $1\leq i\leq k$. The path-pairability number of a graph $G$ is the largest positive integer $k$, for which $G$ is $k$-path-pairable, and it is denoted by $\pp(G)$. A $k$-path-pairable graph on $2k$ or $2k+1$ vertices is simply said to be {\it path-pairable}.
Path-pairability is related to the notion of \textit{linkedness}. A graph is $k$-\emph{linked} if for any choice of $2k$ vertices $\{s_1, \ldots ,
s_k, t_1, \ldots , t_k\}$ (not necessarily distinct), there are internally vertex disjoint paths $P_1, \ldots , P_k$ with $P_i$ joining $s_i$ to $t_i$ for $1 \le i \le k$. Bollob{\'a}s and Thomason~\cite{BollobasThomason} showed that any $2k$-connected graph with a lower bound on its edge density is $k$-linked. On the other hand, a graph being path-pairable imposes no constraint on the connectivity or edge-connectivity of the graph. The most illustrative examples of this phenomenon are the stars $K_{1, n-1}$. Indeed, it is easy to see that stars are path-pairable, while they are neither $2$-connected nor $2$-edge-connected. Note that, for any pairing of the vertices of $K_{1, n-1}$, joining two vertices in a pair is straightforward due to the presence of a vertex of high degree, and the fact that the diameter is small. This example motivates the study of two natural questions about path-pairable graphs: given a path-pairable graph $G$ on $n$ vertices, how small can its maximum degree $\Delta(G)$ be, and how large can its diameter $d(G)$ be? This note addresses some aspects of the second question. To be precise, for a family of graphs $\mathcal{G}$ let us define $d(n, \mathcal{G})$ as follows:
\[
d(n, \mathcal{G}) = \max\{d(G): G \in \mathcal{G} \text{ and } G \text{ is path-pairable on } n \text{ vertices}\}.
\]
When $\mathcal{G}$ is the family of path-pairable graphs, we shall simply write $d(n)$ instead of $d(n, \mathcal{G})$.
\begin{comment}
It was proved by Faudree et al. that $\Delta_\text{min}(n)$ has to grow with the size of the graph; in particular, if $G$ is a path-pairable graph on $n$ vertices with maximum degree $\Delta$, then $n\leq 2\Delta^\Delta$ holds. The result places sublogarithmic lower bound on $\Delta_\text{min}(n)$, that is, $\Delta_\text{min}(n) = \Omega\left(\frac{\log n}{\log\log n}\right)$. To date the best known asymptotic upper bound is $\Delta_\text{min}(n) = O(\log n)$ due to Gy\H ori et al \cite{ntp}.
\end{comment}
The maximum diameter of arbitrary path-pairable graphs was investigated by M\'esz\'aros \cite{me_diam} who proved that $d(n) \le 6 \sqrt{2} \sqrt{n}$.
Our aim in this note is to investigate the maximum diameter of path-pairable graphs when we impose restrictions on the number of edges and on how the edges are distributed.
To state our results, let us denote by $\mathcal{G}_m$ the family of graphs with at most $m$ edges.
The following result determines $d(n, \mathcal{G}_m)$ for a certain range of $m$.
\begin{thm}\label{diam_m}
If $2n \le m \le \frac{1}{4}n^{3/2}$ then
\[
\sqrt[3]{\frac{1}{2}m-n} \le d(n, \mathcal{G}_m) \le 16 \sqrt[3]{m}.
\]
\end{thm}
We remark that the upper bound in the Theorem~\ref{diam_m} holds for $m$ in any range, but when $m \ge \frac{1}{4}n^{3/2}$ the bound obtained by M\'esz\'aros \cite{me_diam} is sharper.
Determining the behaviour of the maximum diameter among path-pairable graphs on $n$ vertices with fewer than $2n$ edges remains an open problem.
In particular, we do not know if the maximum diameter must be bounded (see Section \ref{sec:final}).
Following this line of research, it is very natural to consider the problem of determining the maximum attainable diameter for other classes of graphs. For example, what is the behaviour of the maximum diameter of path-pairable \emph{planar} graphs? Although we could not give a satisfactory answer to this particular question, we were able to do so for graphs which are $c$-\emph{degenerate}.
As usual, we say that an $n$-vertex graph $G$ is $c$-\emph{degenerate} if there exists an ordering $v_1,\ldots,v_n$ of its vertices such that $|\{v_j: j > i, v_iv_j\in E(G) \}|\leq c$ holds for all $i=1,2,\ldots,n$. We let $\mathcal{G}_{c\text{-deg}}$ denote the family of $c$-degenerate graphs. Clearly all $c$-degenerate graphs have a linear number of edges, so Theorem~\ref{diam_m} implies that $d(n, \mathcal{G}_{c\text{-deg}}) = O(\sqrt[3]{n})$. However, as the next result shows, this bound is far from the truth.
\begin{thm}\label{diam_cdeg}
Let $c \ge 5$ be an integer. Then
\[
(2+o(1)) \frac{\log(n)}{\log({\frac{c}{c-2}})} \leq d(n, \mathcal{G}_{c\text{-deg}}) \leq (12+o(1)) \frac{\log(n)}{\log(\frac{c}{c-2})}
\]
as $n \rightarrow \infty$.
\end{thm}
We remark that we have not made an effort to optimize the constants appearing in the upper and lower bounds of Theorems~\ref{diam_m} and~\ref{diam_cdeg}.
\subsection{The Cut-Condition}
While path-pairable graphs need not be highly connected or edge-connected, they must satisfy certain `connectivity-like' conditions that we shall need in the remainder of the paper. We say a graph $G$ on $n$ vertices satisfies the \emph{cut-condition} if for every $X \subset V(G)$, $|X| \le n/2$, there are at least $|X|$ edges between $X$ and $V(G)\setminus X$. Clearly, a path-pairable graph has to satisfy the cut-condition. On the other hand, satisfying the cut-condition is not sufficient to guarantee path-pairability in a graph; see \cite{me_pp} for additional details.
\subsection{Organization and Notation}
The proofs of the lower bounds in Theorems~\ref{diam_m} and~\ref{diam_cdeg} require constructions of path-pairable graphs with large diameter. In Section~\ref{sec:blowup}, we
show how to obtain such graphs by proving that a more general class of graphs is path-pairable. In Sections~\ref{sec:proofdiam_m} and~\ref{sec:proofdiam_cdeg} we shall complete the proofs of
Theorems~\ref{diam_m} and~\ref{diam_cdeg}, respectively. Finally, we mention some open problems in Section~\ref{sec:final}.
Our notation is standard. Thus, for a (simple, undirected) graph $G$ we shall denote by $V(G)$ and $E(G)$ the vertex set and edge set of $G$, respectively. We also let $|G|$ and $d(G)$ denote the number of vertices and diameter of $G$, respectively. For a vertex $x \in V(G)$ we let $N_{G}(x)$ denote the neighbourhood of $x$ in $G$, and we shall omit the subscript `$G$' when no ambiguity arises.
\section{Path-pairable graphs from blowing up paths}\label{sec:blowup}
In this section, we will show how to construct a quite general class of graphs which have high diameter and are path-pairable.
Let $G$ be a graph with vertex set $V(G)=\{v_1,\ldots,v_k\}$, and let $G_1,\ldots, G_k$ be graphs. We define the \textit{blown-up graph} $G(G_1,\ldots, G_k)$ as follows: replace every vertex $v_i$ in $G$ by the corresponding graph $G_i$, and for every edge $v_iv_j \in E(G)$ insert a complete bipartite graph between the vertex sets of $G_i$ and $G_j$.
Let $P_k$ denote the path on $k$ vertices. The following lemma asserts that if we blow-up a path with graphs $G_1, \ldots , G_k$, such that $G_{i}$ is path-pairable for $i \le k-1$, and certain properties inherited from the cut-condition hold, then the resulting blow-up is path-pairable.
\begin{lemma}\label{blown-up}
Suppose that $G_1, \ldots ,G_k$ are graphs on $n_1, \ldots , n_k$ vertices, respectively, where $G_{i}$ is path-pairable for $i \le k-1$.
Let $n = \sum_{i =1}^k n_i$ and let $u_i = \sum_{j=1}^i n_j$ for $i=1, \dots , k-1$. Then $P_k(G_1,\ldots, G_k)$ is path-pairable if and only if
\begin{equation}\label{eq_1}
n_i\cdot n_{i+1}\geq \min(u_i,n-u_i)
\end{equation}
holds for $i=1,\ldots,k-1$.
\end{lemma}
\begin{proof}
For each $i = 1, \ldots , k$, let $U_i = \bigcup_{j=1}^i V(G_j)$ so that $u_i = |U_i|$. Now, if $P_k(G_1, \ldots , G_k)$ is path-pairable, then we may apply the cut-condition to the cut $\{U_i, V(G)\setminus U_i\}$. This implies $n_i\cdot n_{i+1}\geq \min(u_i,n-u_i)$ must hold for $i=1,\ldots , k-1$. In the remainder, we show that this simple condition is enough to yield the path-pairability of
$G := P_k(G_1, \ldots , G_k)$. Assume that a pairing $\mathcal{P}$ of the vertices of $G$ is given. If $\{u, v\} \in \mathcal{P}$ we shall say that $u$ is a \emph{sibling} of $v$ (and vice-versa).
We shall define an algorithm that sweeps through the classes $G_1,G_2,\ldots, G_k$ and joins each pair of siblings via edge-disjoint paths.
First we give an overview of the algorithm.
We proceed by first joining pairs $\{u, v\} \in \mathcal{P}$ via edge-disjoint paths such that $u$ and $v$ belong to different $G_i$'s, and then afterwards joining pairs that remain inside some $G_j$ (using the path-pairability of $G_j$).
Before round $1$ we use the path-pairability property of $G_{1}$ to join those siblings which belong to $G_{1}$.
In round $1$ we assign to every vertex $u$ of $G_1$ a vertex $v$ of $G_2$.
If $\{u, v\} \in \mathcal{P}$ are siblings, then we simply choose the edge $uv$.
Then we join the siblings which are in $G_{2}$ again using the path-pairability property of $G_{2}$.
For those paths $uv$ that have not ended (because $\{u, v\} \notin \mathcal{P}$) we shall continue by choosing a new vertex $w$ in $G_3$ and continue the path with edge $vw$, and so on.
Paths which have not finished joining a pair of siblings we shall call \emph{unfinished}; otherwise, we say the path is \emph{finished}.
The last edge which completes a finished path we shall call a \emph{path-ending edge}.
During round $i$ we shall first choose those vertices in $G_{i+1}$ which, together with some vertex of $G_i$, form path-ending edges.
At the end of round $i$, in $G_{i+1}$ we will have endpoints of unfinished paths and perhaps also some endpoints of finished paths.
Note that the vertices of $G_{i+1}$ might be endpoints of several unfinished paths.
For $x \in G_{i+1}$ let $w(x)$ denote the number of unfinished paths $P\cup \{x\}$ with $P \subset U_i$ at the end of round $i$ which are to be extended by a vertex of $G_{i+2}$ (including the single-vertex path $x$ in the case when $x$ was not joined to its sibling in the latest round).
Note that every such path corresponds to a yet not joined vertex in $U_{i+1}$ as well as to another vertex yet to be joined lying in $V(G)\setminus U_{i+1}$.
It follows that
\begin{equation}\label{eq:weights}
\sum_{x \in G_{i+1}}w(x) \le \min(u_{i+1}, n-u_{i+1}).
\end{equation}
Let us now be more explicit in how we make choices in each round.
We shall maintain the following two simple conditions throughout our procedure (the first of which has been mentioned above):
\begin{itemize}
\item[(a)] During round $i$ ($1\le i \le k-1$), if $w \in G_i$ is the current endpoint of the path which began at some vertex $u\in U_i$ (possibly $u=w$), and $\{u, v\} \in \mathcal{P}$ for $v \in G_{i+1}$, then we join $w$ to $v$. Informally, we choose path-ending edges when we can.
\item[(b)] $w(x) \leq n_{i+1}$ for all $x \in G_i$, for $i =1, \ldots , k-1$.
\end{itemize}
The second condition above is clearly necessary in order to proceed during round $i$, as ${|N(x)\cap G_{i+1}| = n_{i+1}}$ for every $x \in G_i$, and hence we cannot continue more than $n_{i+1}$ unfinished paths through $x$.
We claim that as long as both of the above conditions are maintained, the proposed algorithm finds a collection of edge-disjoint paths joining every pair in $\mathcal{P}$.
Both conditions are clearly satisfied for $i=1$ as $w(x) \le 1\leq n_2$ for all $x\in G_1$.
Let $i \ge 2$ and suppose both conditions hold for rounds $1, \ldots , i-1$.
Our aim is show that an appropriate selection of edges between $G_i$ and $G_{i+1}$ exists in round $i$ to maintain the conditions.
We start round $i$ by choosing all path-ending edges with endpoints in $G_i$ and $G_{i+1}$; this can be done since, by induction, $w(x) \le n_{i+1}$ for every $x \in G_i$.
Observe that if $i = k-1$ then the only remaining siblings are in $G_{k}$.
Then for every $\left\{ u, v \right\} \in \mathcal{P}$ such that $u,v \in G_{k}$ we can find a vertex $w$ in $G_{k-1}$ and join $u, v$ with the path $uwv$.
When $i < k-1$ then the remaining paths can be continued by assigning arbitrary vertices from $G_{i+1}$ (without using any edge multiple times).
We choose an assignment that balances the `weights' in $G_{i+1}$.
More precisely, let us choose an assignment of the vertices that minimizes
\[
\sum\limits_{a\in G_{i+1}}w(a)^2.
\]
If for every $x \in G_{i+1}$ we have that $w(x) \le n_{i+2}$ we are basically done.
It remains to find edge-disjoint paths inside $G_{i+1}$ for those pairs $\{x, y\} \in \mathcal{P}$ whose vertices belong to $G_{i+1}$.
But this is possible because of the assumption that $G_{i+1}$ is path-pairable.
Suppose then that in the above assignment there exists $x \in G_{i+1}$ with $w(x) \geq n_{i+2}+1$.
We first claim that, under this assignment, no other vertex of $G_{i+1}$ has small weight.
\begin{claim}\label{claim:y-big}
Every vertex $y\in G_{i+1}$ satisfies $w(y)\geq n_{i+2}-1$.
\end{claim}
\begin{proof}
Suppose there is $y \in G_{i+1}$ such that $w(y) \le n_{i+2}-2$.
Then, as $w(x)>w(y)+2$, there exist vertices $v_1,v_2\in G_i$ such that certain paths ending at $v_1$ and $v_2$ were joined in round $i$ to $x$ ($x$ was assigned as the next vertex of these paths) but no paths at $v_1$ or $v_2$ were assigned $y$ as their next vertex.
Observe that at least one of the edges $v_1x$ and $v_2x$ is not a path ending edge which could have been replaced by the appropriate $v_1y$ or $v_2y$ edge, respectively.
That operation would result in a new assignment with a smaller square sum $\sum_{a\in G_{i+1}}w(a)^2$, which is a contradiction.
\end{proof}
Therefore, we may assume $w(y)\geq n_{i+2} - 1$ for all $y\in G_{i+1}$.
In this case, partition the vertices of $G_{i+1}$ into three classes:
\begin{align*}
X &= \{v\in G_{i+1}: w(v) \geq n_{i+2} +1\}\\
Y &= \{v\in G_{i+1}: w(v) = n_{i+2} - 1\} \\
Z &= \{v\in G_{i+1}: w(v) = n_{i+2}\}.
\end{align*}
Observe first that $1 \le |X| \leq |Y|$, since otherwise using~(\ref{eq:weights}) we have \[n_{i+1}n_{i+2}+1\leq \sum\limits_{s\in G_{i+1}}w(s)\leq \min(u_{i+1},n-u_{i+1}),\] contradicting condition~(\ref{eq_1}).
Notice also that the same argument as in Claim~\ref{claim:y-big} shows that $w(v) \le n_{i+2}+1$ for every $v \in G_{i+1}$, hence we can actually write
\[
X = \left\{ v \in G_{i+1}: w(v) = n_{i+2}+1 \right\}.
\]
We will need the following claim which asserts that if there are siblings in $G_{i+1}$ then they must belong to $Z$.
\begin{claim}\label{claim:Z_pairs}
If $\{u,v\} \in \mathcal{P}$ and $u, v \in G_{i+1}$, then $u, v \in Z$.
\end{claim}
\begin{proof}
We first show that every $y \in Y$ is incident to a path-ending edge. Suppose, to the contrary, that there is $y \in Y$ such that there is no path-ending edge which ends at $y$.
It follows that there are at most $w(y)$ vertices in $G_{i}$ which had been joined to $y$.
Hence we can take any $x \in X$ and find $z \in G_{i}$ which was not joined to $y$, and such that $xz$ is not a path-ending edge.
Replacing $zx$ by $zy$ would result in a smaller square sum $\sum_{a \in G_{i+1}}w(a)^{2}$, which gives a contradiction.
Now, let $\{u, v\} \in \mathcal{P}$ such that $u, v \in G_{i+1}$.
Since every $y \in Y$ is incident to a path-ending edge, we have that $u, v \not\in Y$.
Suppose, for contradiction, that $u \in X$.
Then $u$ was joined to $w(u) = n_{i+2}+1$ vertices in $G_{i}$, and hence for every $y \in Y$, there is $z \in G_{i}$ which was joined to $u$ but not $y$.
Replacing $zu$ by $zy$ would result in a smaller square sum $\sum_{a \in G_{i+1}}w(a)^{2}$, which again gives a contradiction.
\end{proof}
Finally, we shall show that we can reduce the weights of the vertices in $X$ (and pair the siblings inside $G_{i+1}$) using the path-pairable property of $G_{i+1}$.
For every $x \in X$ pick a different vertex $y_{x} \in Y$ (which we can do, since $|Y| \ge |X|$) and let $\mathcal{P'} = \left\{ \{u, v\} \in \mathcal{P} : u, v \in G_{i+1} \right\} \cup \left\{ \{x, y_{x}\} : x \in X \right\}$.
Since $G_{i+1}$ is path-pairable, we can find edge-disjoint paths joining the siblings in $\mathcal{P'}$ (note that by Claim~\ref{claim:Z_pairs} none of the pairs $\{x, y_x\}$ interfere with any siblings $\{u, v\} \in \mathcal{P}$ with $u, v \in G_{i+1}$).
Observe now that for every $x \in X$ one path has been channeled to a vertex $y\in Y$, thus the number of unfinished path endpoints at $x$ has dropped to $n_{i+2}$ and so the condition is maintained.
\end{proof}
We close the section by pointing out that the condition that the graphs $G_{i}$ are path-pairable is necessary.
We do this by giving an example of a blown-up path $P_k(G_{n_1},\ldots,G_{n_k})$ that satisfies the cut-conditions of Lemma~\ref{blown-up} yet it is not path-pairable unless some of $G_i$'s are path-pairable as well.
For the sake of simplicity we set $k = 5$ and prove that $G_3$ has to be path-pairable.
Let $n = 2t^2 + t$ for some even $t \in \mathbb{N}$ and let $n_1 = n_5 = t^2-t$, $n_2=n_3 = n_4 = t$.
Clearly $P_5(G_{n_1},\ldots,G_{n_5})$ satisfies the Condition~\ref{eq_1} of Lemma~\ref{blown-up}.
Observe, that any pairing of the vertices in $G_{1} \cup G_{2}$ with the vertices in $G_{4} \cup G_{5}$ has to use all the edges between $G_{3}$ and $G_{2} \cup G_{4}$.
Therefore if we additionally pair the vertices inside $G_{3}$, then the paths joining those vertices can only use the edges in $G_{3}$, therefore $G_{3}$ has to be path-pairable.
\section{Proof of Theorem \ref{diam_m}}\label{sec:proofdiam_m}
Take $x,y\in V(G)$ such that $d(x,y) = d(G)$ and let $V_{i}$ be the set of vertices at distance exactly $i$ from $x$, for every $i$.
Observe that $V_0 = \{x\}$ and $y\in V_{d(G)}$.
For $i \in \left\{ 1, \dots, d(G) \right\}$ define $n_{i}$ to be the size of $V_{i}$ and let $u_{i} = \sum_{j=0}^{i} n_{j}$.
We need the following claim.
\begin{claim}
$u_{2k+1} \geq \binom{k+2}{2}$ as long as $u_{2k+1}\leq\frac{n}{2}$.
\end{claim}
\begin{proof}
We shall use induction on $k$.
For $k = 0$ it is clear.
Assume that $u_{2k-1} \ge \binom{k+1}{2}$.
By the cut-condition we have that the number of edges between $V_{2k}$ and $V_{2k+1}$ is at least $u_{2k-1}$, hence $n_{2k}\cdot n_{2k+1} \ge u_{2k-1} \ge \binom{k+1}{2}$.
By the arithmetic-geometric mean inequality, $n_{2k} + n_{2k+1} \ge 2\sqrt{\binom{k+1}{2}} \ge k+1$.
As $u_{2k+1} = u_{2k-1} + n_{2k} + n_{2k+1}$, we have $u_{2k+1} \ge \binom{k+2}{2}$.
\end{proof}
Now, let $A = \bigcup_{i=0}^{\lfloor d / 3 \rfloor}V_{i}$, $B = \bigcup_{i = \lfloor d / 3 \rfloor +1}^{2d/3} V_{i}$, $C = \bigcup_{i = \lfloor 2d/3 \rfloor + 1}^{d} V_{i}$.
Observe, that $|A|, |C| \ge \min\left\{\frac{n}{2}, \frac{d^{2}}{100}\right\}$, so joining vertices in $A$ with vertices in $C$ requires at least $ \min\left\{\frac{n}{2},\frac{d^2}{100}\right\}\cdot\frac{d}{3}$ edges.
Hence,
\[ \min\left\{\frac{n}{2},\frac{d^2}{100}\right\}\cdot\frac{d}{3}\leq m,\]
which implies
\[d\leq \max\left\{\frac{6m}{n},16\sqrt[3]{m}\right\}.\]
Notice that whenever $m \le 4n^{3/2}$ we have $d \le 16\sqrt[3]{m}$.
Let us remark that if $m \ge \frac{1}{4}n^{3/2}$ then the upper bound is trivially satisfied by the general upper bound obtained in \cite{me_diam}.
For the lower bound, let $n$ and $2n \le m \le \frac{1}{4}n^{3/2}$ be given.
For any natural number $\ell$ we shall denote by $S_\ell$ the star $K_{1, \ell-1}$ on $\ell$ vertices.
Consider the graph $G = P_{k}(G_{1},\dots,G_{k})$ on $n$ vertices,
where $k = \left\lfloor \sqrt[3]{\frac{m}{2}-n} \right\rfloor$ and $G_{1} = G_{2} = \dots = G_{k} = S_{k}$, $G_{k+1} = S_{k^{2}}$, $G_{k+2} = S_{2}$, and $G_{k+3}$ is an empty graph on $n-2k^{2}-2$ vertices.
Straightforward calculation shows that $u_i = i\cdot k$, for $i\leq k, u_{k+1}= 2k^2$, and $u_{k+2} = 2k^2 + 2$.
Also $n_1n_2=n_2n_3=\ldots=n_{k-1}n_{k}=k^2$, $n_{k}n_{k+1}=k^3$, $n_{k+1}n_{k+2}=2k^2$, and $n_{k+2}n_{k+3}=2n-4k^2-4$.
Therefore, for $i \in \left\{ 1, \dots, k+1 \right\}$ we have $n_{i} \cdot n_{i+1} \ge u_{i} \ge \min(u_{i}, n-u_{i})$ and $n_{k+2} \cdot n_{k+3} \ge n_{k+3} \ge \min(u_{k+2}, n-u_{k+2})$.
Hence it follows from Lemma~\ref{blown-up} that $G$ is path-pairable.
It is easy to check that the number of edges in $G$ is at most $2n + 2k^{3} \le m$.
On the other hand, the diameter of $G$ is $k+2 \ge \sqrt[3]{\frac{m}{2} - n}$.
\section{Proof of Theorem \ref{diam_cdeg}}\label{sec:proofdiam_cdeg}
In this section, we investigate the maximum diameter a path-pairable $c$-degenerate graph on $n$ vertices can have.
We shall assume that $c$ is an integer and $c\geq 5$.
Let $G$ be a $c$-degenerate graph on $n$ vertices with diameter $d$.
We shall show first that $d \le 4\log_{\frac{c+1}{c}}(n)+3$.
Let $x \in G$ be such that there is $y \in G$ with $d(x,y) = d$.
For $i \in \left\{ 0, \dots, d \right\}$, write $V_{i}$ for the set of vertices at distance $i$ from $x$.
Let $n_{i} = |V_{i}|$ and $u_{i} = \sum_{j = 0}^{i}n_{j}$.
Observe that $|V_{i}| \ge 1$ for every $i \in \left\{ 0, \dots, d \right\}$.
We can assume that $u_{\lfloor \frac{d}{2} \rfloor} \le \frac{n}{2}$ (otherwise we repeat the argument below with $V'_{i} = V_{d-i})$.
The result will easily follow from the following claim.
\begin{claim}
$u_{2k+1} \ge \left( \frac{c+1}{c} \right)^{k}$ as long as $u_{2k+1} \le \frac{n}{2}$.
\end{claim}
Let us assume the claim and prove the result.
Letting $k = \frac{\lfloor\frac{d}{2}\rfloor -1}{2}$, we have that $n/2 \ge u_{2k+1} \ge \left( \frac{c+1}{c} \right)^{\frac{\lfloor \frac{d}{2} \rfloor-1}{2}}$.
Hence $d \leq 4\log_{\frac{c+1}{c}}(n)+3 = 4 \frac{\log(n)}{\log(\frac{c+1}{c})} + 3 \le 4 \frac{\log(n)}{\log(\frac{c}{c-2})}\frac{\log(\frac{c}{c-2})}{\log(\frac{c+1}{c})} + 3\le 12 \frac{\log(n)}{\log(\frac{c}{c-2})}+3$, where the last inequality follows from the easy to check fact that $\frac{\log(\frac{c}{c-2})}{\log(\frac{c+1}{c})} \le 3$, for all $c \ge 5$.
\begin{proof}[Proof of the Claim]
We shall prove the claim by induction on $k$.
The base case when $k=0$ is trivial as $u_{1} \ge 2$.
Suppose the claim holds for every $l \le k-1$.
Since $G$ is $c$-degenerate we have that $e(V_{2k}, V_{2k+1}) \le c\left( n_{2k} + n_{2k+1} \right)$.
On the other hand, it follows from the cut-condition that $e(V_{2k}, V_{2k+1}) \ge u_{2k} = u_{2k-1}+ n_{2k}$.
Therefore, by the induction hypothesis, we have
$n_{2k} + n_{2k+1} \ge
\frac{1}{c}\left( u_{2k-1}+n_{2k} \right)
\ge \frac{1}{c} \left( \left( {\frac{c+1}{c}}\right)^{k-1} +n_{2k} \right)
\ge \frac{1}{c}\left( \frac{c+1}{c} \right)^{k-1}$. Hence, $u_{2k+1}=u_{2k-1}+n_{2k}+n_{2k+1}\geq \left( \frac{c+1}{c} \right)^{k-1}+ \frac{1}{c}\left( \frac{c+1}{c} \right)^{k-1}= \left( 1+\frac{1}{c} \right) \left( \frac{c+1}{c} \right)^{k-1} \ge \left( \frac{c+1}{c} \right)^{k}$, which proves the claim.
\end{proof}
We shall prove the lower bound assuming $c$ is an odd integer; when $c$ is even we apply the same argument for $c-1$.
To do so, consider the graph $G = P(G_{1},\ldots, G_{2m^{\prime}-1})$ for some $m^{\prime}\in \mathbb{N}$, which we specify later.
Firstly, we shall define the sizes of $G_i$ for $i \in \{1,\ldots, 2m'-1\}$.
To do so, let us define a sequence $\{n_i\}_{i\in\mathbb{N}}$ where $n_{2i} =\frac{c-1}{2} $ and $n_{2i+1}$ is defined recursively in the following way:
\begin{equation}
n_{2i+1}= \left \lceil \frac{2}{c-1} \cdot \sum_{j=1}^{2i} n_j \right \rceil \leq \frac{2}{c-1}\sum_{j=1}^{2i} n_j +1
\end{equation}
Let $m$ be the largest integer such that $\sum_{j=1}^{m} n_j \leq n/2$.
We let $m'=m$ when $m$ is odd and $m'=m-1$ when $m$ is even.
Moreover, let $|G_{m'}|=n-2\sum_{j=1}^{j=m'-1} n_j$ and let $|G_i|=n_i$ for $1\leq i < m'$ and $|G_{m'+j}|= |G_{m'-j}|$ for $ j \in \{1,\ldots, m'-1\}$.
For all $i\in \{1,\ldots 2m'-1\}$ let $G_i = S_{n_i}$ be a star on $n_i$ vertices.
It is easy to check that the graph $P_{2m'-1}(G_1,\ldots,G_{2m'-1}) $ is path-pairable by Lemma \ref{blown-up}.
It has diameter at least $2m-4$ and $m \geq \log_{\frac{c+1}{c-1}}(n)(1+o(1))$.
Again an easy verification shows that the graph $G$ is $c$-degenerate.
\section{Final remarks and open problems}\label{sec:final}
We obtained tight bounds on the parameter $d(n, \mathcal{G}_{m})$ when $(2+\epsilon)n \leq m \leq \frac{1}{4} n^{3/2}$, for any fixed $\epsilon >0$.
It is an interesting open problem to investigate what happens when the number of edges in a path-pairable graph on $n$ vertices is around $2n$. We ask the following:
\begin{question}
Is there a function $f$ such that for every $\epsilon>0$ and for every path-pairable graph $G$ on $n$ vertices with at most $(2-\epsilon)n$ edges, the diameter of $G$ is bounded by $f(\epsilon)$?
\end{question}
Another line of research concerns determining the behaviour of $d(n, \mathcal{P})$, where $\mathcal{P}$ is the family of planar graphs. Since planar graphs are $5$-degenerate, it follows from Theorem~\ref{diam_cdeg} that the diameter of a path-pairable planar graph on $n$ vertices cannot be larger than $c \log{n}$.
This fact makes us wonder whether there are path-pairable planar graphs with unbounded diameter.
\begin{question}
Is there a family of path-pairable planar graphs with arbitrarily large diameter?
\end{question}
The graph constructed in the proof of the lower bound in Theorem~\ref{diam_cdeg} when $c=5$ is not planar since it contains a copy of $K_{3,3}$. Therefore, it cannot be used to show that the diameter of a path-pairable planar graph can be arbitrarily large (note, however, that this graph does not contain a $K_7$-minor nor a $K_{6,6}$-minor). We end by remarking that we were able to construct an infinite family of path-pairable planar graphs with diameter $6$, but not larger.
\begin{comment}
\begin{thm} \label{diameter_thm}
For every even $n\geq 6$ there is a path-pairable planar graph on $n$ vertices and with diameter equal $6$.
\end{thm}
\begin{proof}
\textbf{There are few errors here.
First, the diameter is $5$.
When $k_{1} = 1$ the graph seems not to be path-pairable.
Also, when applying the induction we have to make sure that $A_{3}$ doesn't have too few vertices\dots}
Consider the following planar graph on $n=k_1+k_2+6$ vertices, where $k_1,k_2\geq 1$ and $k_1+k_2$ is even.
Partition the vertex set of $G$ into $7$ non-empty subsets $A_1, A_2,A_3,A_4,A_5,A_6$ and $A_7$, namely $A_1=\{a_1\}$,$A_2=\{a_2\}$, $A_3=\{v_1,v_2,...,v_{k_1}\}$, $A_4=\{a_3,a_4\}$ and $A_5=\{w_1,w_2,...,w_{k_2}\}$, $A_6=\{a_5\}$ and $A_7=\{a_6\}$.
Connect $a_1$ to $a_2$ and $a_2$ to every vertex in $A_3$. Let $a_3$ be joined to every vertex in $A_3\cup A_5$ and $a_4$ to the set $\{v_1,v_k,a_3,w_1,w_{k_2}\}$. Symmetrically, let $a_5$ be connected to every vertex in $A_4$ and finally let $a_6$ be connected to $a_5$. We also add the edges of respective paths inside $A_3, A_4,A_5$, that is, we join consecutive vertices of the sequences $(v_1, v_2,...,v_{k_1})$, $(a_3,a_4)$, and $(w_1,w_2,...,w_{k_2})$.
Clearly $d(a_1,a_6)=6$, so the diameter of $G$ is $6$. It is also easy to see that the graph is planar. We need to show it is path-pairable. The statement is fairly obvious for small values of $n$; we leave the verification of these cases to the reader. On the other hand, if $k_1+k_2>6$ (i.e. $n>12$), then every pairing $\mathcal{P}$ of the vertices contains at least one pair of terminals $(u,v)$ such that their corresponding vertices lie in $A_3\cup A_5$. We can join this pair by a path of length 2 through the vertex $a_3$ and complete the pairing process via induction that completes our proof.
\end{proof}
\end{comment}
\begin{comment}
-------------------------------------Need to re-write this proof------------------\\
For every pair between vertices in $\{a_3\}\cup A_2\cup A_3$ use the vertex $a_2$ to join them. If $(a_3,a_2) \in \mathcal{P}$ then just join them via the edge between them. Also if whenever $a_1$ is paired with some vertex $v_t$ in $A_2$, use the edge $(a_1,v_t)$ and similarly if $a_4$ is paired with a vertex in $A_4$.
We split our analysis into three cases, regarding how $\mathcal{P}$ pairs the vertices in ${a_1,a_2,a_3,a_4}$.
\begin{itemize}
\item[i)]$(a_1,a_4)\in \mathcal{P}$; \\
To join $(a_1,a_4)$ use the path $(a_1,v_1,a_3,w_1,a_4)$.
Then either $a_3$ is paired with $a_2$ (which we can solve) or $a_3$ is paired with something in $A_2\cup A_3$, say $v_i$, then use the path $a_3,v_k,v_{k-1},...,v_i$.
\item[ii)] $(a_1,a_2) \in \mathcal{P}$; (when $(a_4,a_2) \in \mathcal{P}$, it is symmetric)\\
Firstly assume $a_3$ and $a_4$ are paired with vertices $v_{j_1},v_{j_2} \in A_2$ with $j_1\leq j_2$ (the other case is symmetric), respectively. Then use the path $(a_1,v_j,a_2)$ to join $a_1$ to $a_2$ and the path $(a_3,v_1,v_2,...,v_{j_1})$ to join $a_3$ and the path $(a_4,w_1,a_3,v_k,v_{k-1},...,v_{j_2})$. The case when $a_3$ is paired with a vertex $w_j \in A_4$ the argument works. If $a_3$ is paired with $a_4$ then use the path $(a_1,v_1,a_3,a_2)$ to join $a_1$ to $a_2$ and the path $(a_3,w_1,a_4)$. When $a_3$ is paired with some vertex $w_l \in A_4$ and $a_4$ is paired with some vertex $v_j \in A_2$ then, as before, use the path $(a_1,v_j,a_2)$ to join $a_1$ to $a_2$, use the path $(a_3,w_1,w_2,...,w_l)$ to join $a_3$ to $w_l$ and the path $(a_4,w_k,a_3,v_1,v_2,...,v_j)$.
\item[iii)] $a_1$ and $a_4$ are paired with some vertex $w_j \in A_4$ and $ v_j \in A_2$, respectively.
Then use the paths $(a_1,v_1,a_3,w_1,a_2,w_j)$ and $(a_2,w_k,a_3,v_k,a_1,w_j)$ to join $a_1$ to $w_j$ and $a_4$ to $v_j$, respectively. Now if $a_3$ is paired $w_l$ (or $v_l$) then join them via $(a_3,a_2,w_l)$ (or $(a_3,a_2,v_l)$).
\end{itemize}
Our case analysis exhausted all possibilities, so we proved our graph is path-pairable.
\end{comment}
\bibliographystyle{acm}
|
\section{Introduction}
Entanglement is a basic concept in quantum information science. It
provides possibility to test quantum nonlocality against local
hidden theory \cite{BellPhysics1,Greenberger,DurPRA62}, and also
plays a key role in various quantum information tasks
\cite{KarlssonPRA58,DFGPRA72,EkertPRL67,DFGPRA68,BennettPRL69,LXSPRA65,SYBPRA81I}.
Therefore, preparing \cite{DZJPRA74,DLMPRL90}, transferring
\cite{WTJPRA85,HCYPRB83} and purifying \cite{RBCPRA90,PanNat410} all
kinds of entangled states in different physical systems become hot
topics in quantum information processing (QIP). As Bell states of
two qubits are easy to be obtained and manipulated, they have been
employed as the information carriers in quantum communications and
quantum computations \cite{EkertPRL67,BennettPRL69,BennettPRL68}.
Thus when using Bell states as information carriers, reading out
quantum information encoded on Bell states is an indispensable task,
which greatly motivated the researches on the Bell-state analysis.
At the beginning, researchers mainly paid their attentions on the
Bell-state analysis for polarized photons with liner optical
elements \cite{MattlePRL76,HouwelingenPRL96}. But unfortunately, it
have been proven by protocols \cite{VaidmanPRA59,CalsamigliaPRA65}
that the Bell-state analysis with only linear optical element have
optimal success probability of 0.5. Besides, the Bell-state analysis
usually destroys the entanglement which causes the waste of physical
resources. Therefore, to achieve complete and nondestructive
Bell-state analysis and to exploit the advantages of other physical
systems, researchers have turned their attentions on Bell states in
various systems by applying many new techniques, such as
nonlinearities and hyperentanglement. Until now, complete and
nondestructive Bell-state analysis for photons
\cite{SYBPRA81II,SYBPRA82,BarbieriPRA75,WTJPRA86,RBCOE20,BonatoPRL104,XYJOSAB31},
atoms \cite{HYCPB19}, spins inside quantum dots
\cite{WHRIJTP52,KYHAPB119} and nitrogen-vacancy centers
\cite{LJZIJTP56} have been reported.
In recent years, the superconducting system has been developed a
lot, and is now deemed as a very promising candidate to implement
quantum information tasks
\cite{MakhlinRMP73,XZLRMP85,VionSci296,YYSci296,YCPPRL92,YCPPRA67,YCPPRA74,YCPPRA82,ClarkeNature453,DevoretADP16,YCPPRA86,BlaisPRA69,WallraffNature431,YCPPRA87,ChiorescuNature431,SteinbachPRL87,FilippPRL102,BialczakNP6,YamamotoPRB82,ReedPRL105,MajerNature449,DiCarloNature460,SchmidtADP525,StrauchPRL105,KochPRA76},
as it possesses many advantages. Superconducting qubits, including
phase qubits, change qubits, flux qubits, etc., are outstanding with
their relatively long decoherence time \cite{ClarkeNature453} and
perfect scalability \cite{VionSci296,YYSci296,ChiorescuNature431}.
Among all kinds of superconducting qubits, the
superconducting-quantum-interference-device (SQUID) qubits in cavity
quantum electrodynamics (QED) have many advantages:
(1) The positions of SQUID qubits in a cavity are fixed. That makes
them holds superiority compared with neutral atoms, which requires
to be controlled the centers of mass motion in a cavity.
\cite{YCPPRL92,YCPPRA67}.
(2) When placing SQUID qubits into a superconducting cavity,
decoherence induced due to the external environment can be greatly
suppressed since the superconducting cavity could be considered as
the magnetic shield for SQUID qubits \cite{YCPPRA67}.
(3) The strong-coupling limit of the cavity QED can be easily
realized for SQUID qubits embedded in a cavity, while it is
difficult to be realized with atoms \cite{YCPPRL92}.
(4) The level structure of every individual SQUID qubit can be
adjusted easily \cite{YCPPRL92}.
The great advantages of SQUID qubits make them attractive choices to
implement quantum information tasks. So far, SQUID qubits have been
widely used in entanglement preparations
\cite{YCPPRL92,YCPPRA67,DZJPRA74,SKHPRA75}, information transfers
\cite{YCPPRL92,YCPPRA67}, logic gates \cite{YCPPRA67}. However,
Bell-state analysis for SQUID qubits still has plenty room for
researches.
On the other hand, when choosing superconducting system as the
platform for QIP, an ineluctable question is to design microwave
pulses driving superconducting qubits to complete various
operations. Interestingly, a new technique called by shortcut to
adiabaticity (STA)
\cite{TorronteguiAAMOP62,BerryJPA42,CXPRL105,CYHSR6,CampoPRL111,CampoPRL109,CXPRA83,MugaJPB42,CXPRL104,CampoSR2,BaksicPRL116,GaraotPRA89,SaberiPRA90,TorronteguiPRA89,TorosovPRA87,TorosovPRA89,IbanezPRA87,IbanezPRL109,SXKPRA93,TorronteguiPRA83,MugaJPB43,TorronteguiPRA85,MasudaPRA84,MasudaJPCA119,CXPRA82,CXPRA84,CampoPRA84,CampoEPL96,SchaffPRA82,CXPRA86,SantosPRA93,HenPRA91,DeffnerPRX4,SXKNJP6}
has been developing recently to control quantum evolutions. Rather
than confining quantum evolutions along one eigenstate or
superpositions of several eigenstates of the Hamiltonian under the
adiabatic condition, STA provides a lot of evolution paths by means
of various methods including transitionless tracking algorithm
\cite{BerryJPA42,CXPRL105,CXPRA83,CampoPRL111,CYHSR6},
Lewis-Riesenfeld invariants theory \cite{CXPRA83,CXPRL104}, Lie
algebra \cite{GaraotPRA89,TorronteguiPRA89}, picture transformations
\cite{BaksicPRL116,IbanezPRA87,IbanezPRL109,SXKPRA93}, fast-forward
scales \cite{MasudaPRA84,MasudaJPCA119}, etc.. These protocols
\cite{TorronteguiAAMOP62,BerryJPA42,CXPRL105,CYHSR6,CampoPRL111,CampoPRL109,CXPRA83,MugaJPB42,CXPRL104,CampoSR2,BaksicPRL116,GaraotPRA89,SaberiPRA90,TorronteguiPRA89,TorosovPRA87,TorosovPRA89,IbanezPRA87,IbanezPRL109,SXKPRA93,TorronteguiPRA83,MugaJPB43,TorronteguiPRA85,MasudaPRA84,MasudaJPCA119,CXPRA82,CXPRA84,CampoPRA84,CampoEPL96,SchaffPRA82,CXPRA86,SantosPRA93,HenPRA91,DeffnerPRX4,SXKNJP6}
have demonstrated that STA not only inherits the robustness of the
adiabatic passage, but also greatly accelerates adiabatic processes.
Moreover, constructing STA by using different methods produces
excellent feasibility to handle all kinds of quantum information
tasks. Thus, it may be a good idea applying STA in pulse design to
manipulate superconducting systems.
In this paper, motivated by (1) the importance of Bell-state
analysis in quantum information tasks, (2) the advantages of SQUID
qubits, (3) the requirement of Bell-state analysis from quantum
communications and computations within superconducting quantum
networks, (4) the advantages of STA in designing pulses to control
physical systems, we proposed a protocol for complete and
nondestructive Bell state analysis for two SQUID qubits. By using
transitionless tracking algorithm, a useful method of STA, a
sequence of microwave pulses are designed to complete the Bell-state
analysis. The information for distinguishing four Bell states would
be encoded in two auxiliary SQUID qubits, and could be read out with
current technology \cite{BennettSST20,TakayanagiSM32}. Therefore,
the operations of the Bell-state analysis are not difficult in real
experiments. Besides, the protocol combines the robustness of SQUID
qubits and the speediness of STA. Thus, we can see in numerical
simulation that high success probability to distinguish each Bell
state are still available when decoherence is considered. By
substituting experimentally realizable parameters, good performance
of the Bell-state analysis is shown.
The article is organized as follows. In Sec. II, we briefly review
the physical model of a SQUID qubit. In Sec. III, we amply
illuminate the procedures of the Bell-state analysis. In Sec. IV,
the transitionless tracking algorithm is utilized to design a
microwave pulses for realizing the Bell-state analysis. In Sec. V,
numerical simulations are performed to select suitable control
parameters and demonstrate the robustness of the Bell-state analysis
against decoherence. Finally, conclusions are given in Sec. VI.
\section{Physical model of a SQUID qubit}
Considering a single SQUID qubit with junction capacitance $C$ and
loop inductance $L$, its Hamiltonian reads \cite{YCPPRA67,YCPPRL92}
\begin{eqnarray}\label{e1}
H_{s}(t)=\frac{Q^2}{2C}+\frac{(\Phi-\Phi_x)^2}{2L}-E_J\cos(2\pi\frac{\Phi}{\Phi_0}),
\end{eqnarray}
where, $Q$ is the total charge on the capacitor; $\Phi$ is the
magnetic flux threading the loop, and $\Phi_x$ is the external flux
applied to the ring; $E_J=I_c\Phi_0/2\pi$ is the Josephson energy
with $I_c$ and $\Phi_0=h/2e$ being the critical current of the
junction and the flux quantum. By quantizing the Hamiltonian of the
SQUID qubit, the SQUID qubit can be described by level diagram with
a serial of energy levels $\{|k\rangle\}$ ($k=0,1,2,...$) shown in
Fig.~\ref{fig1} \cite{YCPPRA67,YCPPRL92}.
\begin{figure}
\scalebox{0.6}{\includegraphics[scale=1]{Fig1.eps}} \caption{The
level configuration of a single SQUID qubit.}\label{fig1}
\end{figure}
When a transition between two different levels $|k\rangle$ and
$|k'\rangle$ is driving by an classical microwave field, in the
frame of the rotating-wave approximation, the Rabi frequency of the
driving field could be written by \cite{YCPPRA67,YCPPRL92}
\begin{eqnarray}\label{e2}
\Omega_{kk'}(t)=\frac{1}{2L\hbar}\langle
k|\Phi|k'\rangle\int_{S}\tilde{\mathbf{B}}_{\mu
w}(\mathbf{r},t)\cdot d\mathbf{S},
\end{eqnarray}
where, $S$ is surface bounded by the loop of the SQUID qubit;
$\mathbf{B}_{\mu w}(\mathbf{r},t)=\tilde{\mathbf{B}}_{\mu
w}(\mathbf{r},t)\cos(2\pi\nu_{\mu w}t)$ is the magnetic components
of the classical microwave in the superconducting loop of the SQUID
qubit with frequency $\nu_{\mu w}$. Considering that the SQUID qubit
is placed in a microwave cavity, when a transition between two
different levels $|k\rangle$ and $|k'\rangle$ is coupled to a
quantized cavity field with frequency $\omega_c$, after the
rotating-wave approximation, the coupling constant reads
\cite{YCPPRA67,YCPPRL92}
\begin{eqnarray}\label{e3}
g_{kk'}=\frac{1}{L}\sqrt{\frac{\omega_c}{2\mu_0\hbar}}\langle
k|\Phi|k'\rangle\int_{S}\mathbf{B}_c(\mathbf{r})\cdot d\mathbf{S},
\end{eqnarray}
where, $\mathbf{B}_c(\mathbf{r})$ is the magnetic components of the
cavity mode in the superconducting loop of the SQUID qubit.
\section{Complete Bell-state analysis}
Consider a system that contains four SQUID qubits $A_1$, $A_2$,
$B_1$, $B_2$ placing inside a microwave cavity, which is shown in
Fig.~\ref{fig2}~(a). SQUID qubit $A_j$ ($j=1,2$) is employed as an
auxiliary qubit, whose level diagram is shown in
Fig.~\ref{fig2}~(b). We consider the lowest five levels
$|0\rangle_{A_j}$, $|1\rangle_{A_j}$, $|2\rangle_{A_j}$,
$|3\rangle_{A_j}$, and $|4\rangle_{A_j}$ of SQUID qubit $A_j$. The
transition between $|0\rangle_{A_j}$ and $|3\rangle_{A_j}$
($|4\rangle_{A_j}$) is resonantly driven by a classical microwave
field with Rabi frequency $\Omega_{03A_j}(t)$ ($\Omega_{04A_j}(t)$).
The transition between $|1\rangle_{A_j}$ and $|3\rangle_{A_j}$
($|4\rangle_{A_j}$) is resonantly driven by a classical microwave
field with Rabi frequency $\Omega_{13A_j}(t)$ ($\Omega_{14A_j}(t)$).
SQUID qubits $B_1$ and $B_2$ are information carriers, whose level
diagrams are shown in Fig.~\ref{fig2}~(c). We consider the lowest
three levels $|0\rangle_{B_j}$, $|1\rangle_{B_j}$, $|2\rangle_{B_j}$
and among them, information is encoded on $|0\rangle_{B_j}$ and
$|1\rangle_{B_j}$. Thus, the four Bell states to be distinguished
can be described as
\begin{eqnarray}\label{e4}
&&|\Psi_{\pm}\rangle_{B_1B_2}=\frac{1}{2}(|0\rangle_{B_1}|0\rangle_{B_2}\pm|1\rangle_{B_1}|1\rangle_{B_2}),\cr\cr&&
|\Phi_{\pm}\rangle_{B_1B_2}=\frac{1}{2}(|0\rangle_{B_1}|1\rangle_{B_2}\pm|1\rangle_{B_1}|0\rangle_{B_2}).
\end{eqnarray}
A classical microwave field with Rabi frequency $\Omega_{B_j}(t)$ is
applied on SQUID qubit $B_j$ to drive the transition between levels
$|0\rangle_{B_j}$ and $|1\rangle_{B_j}$.
\begin{figure}
\scalebox{0.6}{\includegraphics[scale=0.9]{Fig2.eps}} \caption{(a)
The auxiliary qubits $A_1$, $A_2$ and the information carriers $B_1$
and $B_2$ placed in a microwave cavity. (b) The level configuration
of SQUID qubit $A_j$ ($j=1,2$). (c) The level configurations of
SQUID qubits $B_1$ and $B_2$.}\label{fig2}
\end{figure}
Assuming the microwave cavity is a double-mode cavity, where two
cavity fields $a_1$ and $a_2$ (denote by their annihilation
operators) may exist. The cavity field $a_1$ ($a_2$) could
resonantly coupled with the transition between levels
$|2\rangle_{A_j}$ ($j=1,2$) and $|3\rangle_{A_j}$ ($|2\rangle_{A_j}$
and $|4\rangle_{A_j}$) of SQUID qubit $A_j$ with coupling constant
$g_{23A_j}$ ($g_{24A_j}$), and the transition between levels
$|0\rangle_{B_1}$ and $|2\rangle_{B_1}$ ($|0\rangle_{B_2}$ and
$|2\rangle_{B_2}$) with coupling constant $g_{B_1}$ ($g_{B_2}$).
Assuming the frequencies of cavity modes $a_1$ and $a_2$ are
$\omega_1$ and $\omega_2$, respectively. The transition frequency
between $|0\rangle_{B_1}$ and $|2\rangle_{B_1}$ ($|0\rangle_{B_2}$
and $|2\rangle_{B_2}$) of SQUID qubit $B_1$ ($B_2$) should be equal
to $\omega_1$ ($\omega_2$). According to Ref.~\cite{YCPPRL92}, level
structure of each individual SQUID qubit can be adjusted by either
design variations and/or changing local bias field. Thus, coupling
between microwave pulses (cavity fields) and any particular SQUID
qubits can be obtained selectively via frequency matching. When
$g_{B_1},g_{B_2}\ll|\omega_1-\omega_2|$, the interaction between
SQUID qubit $B_1$ and cavity mode $a_2$ (SQUID qubit $B_2$ and
cavity mode $a_1$) could be discarded \cite{YCPPRA67}. Therefore, in
the interaction picture, the total Hamiltonian of system for
Bell-state analysis could be written as
\begin{eqnarray}\label{e5}
&&H_{I}(t)=H_{mA}(t)+H_{mB}+H_{c},\cr\cr&&
H_{mA}(t)=\sum\limits_{j=1,2}\Omega_{03A_j}(t)|0\rangle\langle3|+\Omega_{04A_j}(t)|0\rangle\langle4|
+\Omega_{13A_j}(t)|1\rangle\langle3|+\Omega_{14A_j}(t)|1\rangle\langle4|+H.c.,\cr\cr&&
H_{mB}(t)=\sum\limits_{j=1,2}\Omega_{B_j}(t)e^{-i\epsilon_j}|0\rangle\langle1|+H.c.,\cr\cr&&
H_{c}=\sum\limits_{j=1,2}g_{23A_j}|3\rangle_{A_j}\langle2|a_1+g_{24A_j}|4\rangle_{A_j}\langle2|a_2+g_{B_j}|2\rangle_{B_j}\langle0|a_j+H.c.,
\end{eqnarray}
where, $\epsilon_j$ is the phase shift of $\Omega_{B_j}$. Here, we
take $\epsilon_j=\pi/2$ for the convenience of calculations and
descriptions.
Now, let us describe the procedures for Bell-state analysis. The
Bell-state can be divided into six steps. We do not discuss the
pulse design here, but leave it later in Sec. IV. Besides, for the
convenience of descriptions, we assume the operation time of each
step is $T$.
\textbf{Step 1:} Assume SQUID qubit $A_j$ is initially in state
$|0\rangle_{A_j}$, cavity field $a_j$ is initially in vacuum state
$|0\rangle_{a_j}$. We turn on $\Omega_{03A_1}(t)$ and
$\Omega_{13A_1}(t)$, but turn off other classical microwave fields.
In this case, SQUID qubit $A_2$ is decoupled to the system. Besides,
cavity fields $a_2$ keeps in vacuum state. Thus, whether SQUID qubit
$B_2$ is in state $|0\rangle_{B_2}$ or $|1\rangle_{B_2}$, it does
not evolve as well. Without considering cavity field $a_2$ and the
decoupled SQUID qubits, the system would evolve in a subspace
spanned by
\begin{eqnarray}\label{e6}
&&|\bar{\psi}_1\rangle=|0\rangle_{A_1}|0\rangle_{B_1}|0\rangle_{a_1},\
\
|\bar{\psi}_2\rangle=|3\rangle_{A_1}|0\rangle_{B_1}|0\rangle_{a_1},\
\
|\bar{\psi}_3\rangle=|2\rangle_{A_1}|0\rangle_{B_1}|1\rangle_{a_1},\cr\cr
&&|\bar{\psi}_4\rangle=|2\rangle_{A_1}|2\rangle_{B_1}|0\rangle_{a_1},\
\
|\bar{\psi}_5\rangle=|1\rangle_{A_1}|0\rangle_{B_1}|0\rangle_{a_1},\
\
|\bar{\psi}_6\rangle=|0\rangle_{A_1}|1\rangle_{B_1}|0\rangle_{a_1},\cr\cr
&&|\bar{\psi}_7\rangle=|3\rangle_{A_1}|1\rangle_{B_1}|0\rangle_{a_1},\
\
|\bar{\psi}_8\rangle=|2\rangle_{A_1}|1\rangle_{B_1}|1\rangle_{a_1},\
\
|\bar{\psi}_9\rangle=|1\rangle_{A_1}|1\rangle_{B_1}|0\rangle_{a_1},
\end{eqnarray}
where $|1\rangle_{a_1}$ denotes the one-photon state of cavity field
$a_1$. Rewriting the Hamiltonian of the system within the subspace,
we obtain
\begin{eqnarray}\label{e7}
&&H_{step1}(t)=H_{m1}(t)+H_{c1},\cr\cr&&
H_{m1}(t)=\Omega_{03A_1}(t)(|\bar{\psi}_1\rangle\langle\bar{\psi}_2|+|\bar{\psi}_6\rangle\langle\bar{\psi}_7|)+\Omega_{13A_1}(t)(|\bar{\psi}_5\rangle\langle\bar{\psi}_2|+|\bar{\psi}_9\rangle\langle\bar{\psi}_7|)+H.c.,\cr\cr&&
H_{c1}=g_{23A_1}(|\bar{\psi}_2\rangle\langle\bar{\psi}_3|+|\bar{\psi}_7\rangle\langle\bar{\psi}_8|)+g_{B_1}|\bar{\psi}_4\rangle_{B_1}\langle\bar{\psi}_3|+H.c..
\end{eqnarray}
The eigenstates of $H_{c1}$ are
\begin{eqnarray}\label{e8}
&&|\bar{\phi}_{0}\rangle=\frac{1}{\sqrt{g_{23A_1}^2+g_{B_1}^2}}(g_{B_1}|\bar{\psi}_2\rangle-g_{23A_1}|\bar{\psi}_4\rangle),\cr\cr
&&|\bar{\phi}_{1}\rangle=\frac{1}{\sqrt{2}}(|\bar{\psi}_7\rangle+|\bar{\psi}_8\rangle),\cr\cr
&&|\bar{\phi}_{2}\rangle=\frac{1}{\sqrt{2}}(|\bar{\psi}_7\rangle-|\bar{\psi}_8\rangle),\cr\cr
&&|\bar{\phi}_{3}\rangle=\frac{1}{\sqrt{2(g_{23A_1}^2+g_{B_1}^2)}}(g_{23A_1}|\bar{\psi}_2\rangle+\sqrt{g_{23A_1}^2+g_{B_1}^2}|\bar{\psi}_3\rangle+g_{B_1}|\bar{\psi}_4\rangle),\cr\cr
&&|\bar{\phi}_{4}\rangle=\frac{1}{\sqrt{2(g_{23A_1}^2+g_{B_1}^2)}}(g_{23A_1}|\bar{\psi}_2\rangle-\sqrt{g_{23A_1}^2+g_{B_1}^2}|\bar{\psi}_3\rangle+g_{B_1}|\bar{\psi}_4\rangle),
\end{eqnarray}
with corresponding eigenvalues 0, $g_{23A_1}$, $-g_{23A_1}$,
$\sqrt{g_{23A_1}^2+g_{B_1}^2}$, $-\sqrt{g_{23A_1}^2+g_{B_1}^2}$,
respectively. With the condition
$\Omega_{03A_1}(t),\Omega_{13A_1}(t)\ll g_{23A_1},g_{B_1}$, we can
derive the effective Hamiltonian of the system as
\begin{eqnarray}\label{e9}
H_{eff1}(t)=\frac{g_{B_1}}{\sqrt{g_{23A_1}^2+g_{B_1}^2}}[\Omega_{03A_1}(t)|\bar{\psi}_1\rangle\langle\bar{\phi}_0|+\Omega_{13A_1}(t)|\bar{\psi}_5\rangle\langle\bar{\phi}_0|]+H.c..
\end{eqnarray}
The details of the derivation of $H_{eff1}(t)$ are given in the
appendix. To realize
$|\bar{\psi}_1\rangle\rightarrow|\bar{\psi}_5\rangle$, we should
suitably design $\Omega_{03A_1}(t)$ and $\Omega_{13A_1}(t)$. The
design of $\Omega_{03A_1}(t)$ and $\Omega_{13A_1}(t)$ is amply
discussed in Sec. IV. Thus, if SQUID qubit $B_1$ is initially in
state $|0\rangle_{B_1}$, after Step 1, its state keeps unchanged,
while SQUID qubit $A_1$ evolves from $|0\rangle_{A_1}$ to
$|1\rangle_{A_1}$. But if qubit $B_1$ is initially in state
$|1\rangle_{B_1}$, both SQUID qubits $A_1$ and $B_1$ keep in their
initial states after Step 1.
\textbf{Step 2:} In this step, we turn on $\Omega_{04A_1}(t)$ and
$\Omega_{14A_1}(t)$, but turn off other classical microwave fields.
Similar to Step 1, SQUID qubit $A_2$ is decoupled to the system,
while cavity fields $a_1$ keeps in vacuum state. Thus, SQUID qubit
$B_1$ is decoupled to the system in this step. The system would
evolve in a subspace spanned by
\begin{eqnarray}\label{e10}
&&|\tilde{\psi}_1\rangle=|0\rangle_{A_1}|0\rangle_{B_2}|0\rangle_{a_2},\
\
|\tilde{\psi}_2\rangle=|4\rangle_{A_1}|0\rangle_{B_2}|0\rangle_{a_2},\
\
|\tilde{\psi}_3\rangle=|2\rangle_{A_1}|0\rangle_{B_2}|1\rangle_{a_2},\cr\cr
&&|\tilde{\psi}_4\rangle=|2\rangle_{A_1}|2\rangle_{B_2}|0\rangle_{a_2},\
\
|\tilde{\psi}_5\rangle=|1\rangle_{A_1}|0\rangle_{B_2}|0\rangle_{a_2},\
\
|\tilde{\psi}_6\rangle=|0\rangle_{A_1}|1\rangle_{B_2}|0\rangle_{a_2},\cr\cr
&&|\tilde{\psi}_7\rangle=|4\rangle_{A_1}|1\rangle_{B_2}|0\rangle_{a_2},\
\
|\tilde{\psi}_8\rangle=|2\rangle_{A_1}|1\rangle_{B_2}|1\rangle_{a_2},\
\
|\tilde{\psi}_9\rangle=|1\rangle_{A_1}|1\rangle_{B_2}|0\rangle_{a_2}.
\end{eqnarray}
Similar way, under the condition
$\Omega_{04A_1}(t),\Omega_{14A_1}(t)\ll g_{24A_1},g_{B_2}$, the
effective Hamiltonian of the system can be derived as
\begin{eqnarray}\label{e11}
H_{eff2}(t)=\frac{g_{B_2}}{\sqrt{g_{24A_1}^2+g_{B_2}^2}}[\Omega_{04A_1}(t)|\tilde{\psi}_1\rangle\langle\tilde{\phi}_0|+\Omega_{14A_1}(t)|\tilde{\psi}_5\rangle\langle\tilde{\phi}_0|]+H.c.,
\end{eqnarray}
with
$|\tilde{\phi}_{0}\rangle=\frac{1}{\sqrt{g_{24A_1}^2+g_{B_2}^2}}(g_{B_2}|\tilde{\psi}_2\rangle-g_{24A_1}|\tilde{\psi}_4\rangle)$.
The design of $\Omega_{04A_1}(t)$ and $\Omega_{14A_1}(t)$ for
achieving
$|\tilde{\psi}_1\rangle\leftrightarrow|\tilde{\psi}_5\rangle$ is
also discussed in Sec. IV. Therefore, if SQUID qubit $B_2$ is
initially in state $|0\rangle_{B_2}$, its state keeps unchanged,
while the evolution of SQUID qubit $A_1$ would be
$|0\rangle_{A_1}\rightarrow|1\rangle_{A_1}$ or
$|1\rangle_{A_1}\rightarrow|0\rangle_{A_1}$, where the initial state
of SQUID qubit $A_1$ in this step is decided by the result of Step
1. Otherwise, the initial state of SQUID qubit $B_2$ is
$|1\rangle_{B_2}$, both SQUID qubits $A_1$ and $B_2$ stay in their
initial states.
Step 1 and Step 2 can be summarized as Table I.
\begin{center}
\centering{\bf Table I. The evolution of $A_1$ in Step 1 and Step
2.} {\small\begin{tabular}{ccc} \hline\hline
\ \ \ \ The state of $B_1$ and $B_2$\ \ \ \ &\ \ \ \ Step 1\ \ \ \ &\ \ \ \ Step 2\ \ \ \ \\
\hline
\ \ \ \ $|0\rangle_{B_1}|0\rangle_{B_2}$\ \ \ \ &\ \ \ \ $|0\rangle_{A_1}\rightarrow|1\rangle_{A_1}$\ \ \ \ &\ \ \ \ $|1\rangle_{A_1}\rightarrow|0\rangle_{A_1}$\ \ \ \ \\
\ \ \ \ $|0\rangle_{B_1}|1\rangle_{B_2}$\ \ \ \ &\ \ \ \ $|0\rangle_{A_1}\rightarrow|1\rangle_{A_1}$\ \ \ \ &\ \ \ \ $|1\rangle_{A_1}\rightarrow|1\rangle_{A_1}$\ \ \ \ \\
\ \ \ \ $|1\rangle_{B_1}|0\rangle_{B_2}$\ \ \ \ &\ \ \ \ $|0\rangle_{A_1}\rightarrow|0\rangle_{A_1}$\ \ \ \ &\ \ \ \ $|0\rangle_{A_1}\rightarrow|1\rangle_{A_1}$\ \ \ \ \\
\ \ \ \ $|1\rangle_{B_1}|1\rangle_{B_2}$\ \ \ \ &\ \ \ \ $|0\rangle_{A_1}\rightarrow|0\rangle_{A_1}$\ \ \ \ &\ \ \ \ $|0\rangle_{A_1}\rightarrow|0\rangle_{A_1}$\ \ \ \ \\
\hline \hline
\end{tabular}}
\end{center}
Thus, after Step 1 and Step 2, information for distinguishing
$|\Psi_\pm\rangle_{B_1B_2}$ from $|\Phi_\pm\rangle_{B_1B_2}$ is
encoded on the SQUID qubit $A_1$.
\textbf{Step 3:} In this step, we only turn on classical microwave
fields $\Omega_{B_1}(t)$ and $\Omega_{B_2}(t)$ to perform
single-qubit operation on SQUID qubits $B_1$ and $B_2$. When
\begin{eqnarray}\label{e12}
\int_{2T}^{3T}\Omega_{B_j}(t')dt'=\pi/4,
\end{eqnarray}
the single-qubit operation on SQUID qubit $B_j$ is
\begin{eqnarray}\label{e13}
|0\rangle_{B_j}\rightarrow\frac{1}{\sqrt{2}}(|0\rangle_{B_j}+|1\rangle_{B_j}),\
\ \
|1\rangle_{B_j}\rightarrow-\frac{1}{\sqrt{2}}(|0\rangle_{B_j}-|1\rangle_{B_j}).
\end{eqnarray}
Therefore, the four Bell states change as follows
\begin{eqnarray}\label{e14}
&&|\Psi_+\rangle_{B_1B_2}\rightarrow|\Psi_+\rangle_{B_1B_2},\ \ \
|\Psi_-\rangle_{B_1B_2}\rightarrow|\Phi_+\rangle_{B_1B_2},\cr\cr
&&|\Phi_+\rangle_{B_1B_2}\rightarrow-|\Psi_-\rangle_{B_1B_2},\ \
|\Phi_-\rangle_{B_1B_2}\rightarrow|\Phi_-\rangle_{B_1B_2}.
\end{eqnarray}
Thus, we can transform the question of distinguishing
$|\Psi_+\rangle_{B_1B_2}$ ($|\Phi_+\rangle_{B_1B_2}$) from
$|\Psi_-\rangle_{B_1B_2}$ ($|\Phi_-\rangle_{B_1B_2}$) to the
question of distinguishing $|\Psi_\pm\rangle_{B_1B_2}$ from
$|\Phi_\pm\rangle_{B_1B_2}$.
\textbf{Step 4:} We turn on $\Omega_{03A_2}(t)$ and
$\Omega_{13A_2}(t)$, but turn off other classical microwave fields.
In this step, SQUID qubit $A_2$ takes the place of SQUID qubit $A_1$
in the Step 1. Thus, by suitably designing $\Omega_{03A_2}(t)$ and
$\Omega_{13A_2}(t)$, we can make SQUID qubit $A_2$ evolve from
$|0\rangle_{A_2}$ to $|1\rangle_{A_2}$ when SQUID qubit $B_1$ is in
state $|0\rangle_{B_1}$, while keep in $|0\rangle_{A_2}$ when qubit
$B_1$ is in state $|1\rangle_{B_1}$.
\textbf{Step 5:} We only turn on $\Omega_{04A_2}(t)$ and
$\Omega_{14A_2}(t)$. In this step, SQUID qubit $A_2$ takes the place
of SQUID qubit $A_1$ in the Step 2. Therefore, with suitable
$\Omega_{04A_2}(t)$ and $\Omega_{14A_2}(t)$, we can also make SQUID
qubit $A_2$ evolves from $|0\rangle_{A_2}$ to $|1\rangle_{A_2}$
($|1\rangle_{A_2}$ to $|0\rangle_{A_2}$) when SQUID qubit $B_2$ is
in state $|0\rangle_{B_2}$, while keep in $|0\rangle_{A_2}$ when
qubit $B_2$ is in state $|1\rangle_{B_2}$. After, Step 4 and Step 5,
the information for distinguishing $|\Psi_+\rangle_{B_1B_2}$
($|\Phi_+\rangle_{B_1B_2}$) from $|\Psi_-\rangle_{B_1B_2}$
($|\Phi_-\rangle_{B_1B_2}$) is encoded on SQUID qubit $A_2$.
\textbf{Step 6:} In this step, we only turn on classical microwave
fields $\Omega_{B_1}(t)$ and $\Omega_{B_2}(t)$ to perform inverse
transformations of Step 3 on SQUID qubits $B_1$ and $B_2$. When
\begin{eqnarray}\label{e15}
\int_{5T}^{6T}\Omega_{B_j}(t')dt'=-\pi/4,
\end{eqnarray}
the transformations for SQUID qubit $B_j$ are
\begin{eqnarray}\label{e16}
|0\rangle_{B_j}\rightarrow\frac{1}{\sqrt{2}}(|0\rangle_{B_j}-|1\rangle_{B_j}),\
\ \
|1\rangle_{B_j}\rightarrow\frac{1}{\sqrt{2}}(|0\rangle_{B_j}+|1\rangle_{B_j}).
\end{eqnarray}
After that, the states of SQUID qubits $B_1$ and $B_2$ recover to
their initial forms.
At the end of the Bell-state analysis, we detecting the states of
SQUID qubits $A_1$ and $A_2$, thus reading out the information for
distinguishing the four Bell states of SQUID qubits $B_1$ and $B_2$.
The measurement results of the states of SQUID qubits $A_1$ and
$A_2$ with corresponding Bell state of SQUID qubits $B_1$ and $B_2$
are shown in Table II. According to measurement results of the
states of SQUID qubits $A_1$ and $A_2$, a complete and
nondestructive Bell-state analysis can be realized.
\begin{center}
\centering{\bf Table II. Measurement results of the states of SQUID
qubits $A_1$ and $A_2$ with the corresponding Bell states of $B_1$
and $B_2$.} {\small\begin{tabular}{ccc} \hline\hline \ \ \
\ Measurement result\ \ \ \ &\ \ \ \ Corresponding Bell state\ \ \ \ \\
\hline
\ \ \ \ $|0\rangle_{A_1}|0\rangle_{A_2}$\ \ \ \ &\ \ \ \ $|\Psi_{+}\rangle_{B_1B_2}$\ \ \ \ \\
\ \ \ \ $|0\rangle_{A_1}|1\rangle_{A_2}$\ \ \ \ &\ \ \ \ $|\Psi_{-}\rangle_{B_1B_2}$\ \ \ \ \\
\ \ \ \ $|1\rangle_{A_1}|0\rangle_{A_2}$\ \ \ \ &\ \ \ \ $|\Phi_{+}\rangle_{B_1B_2}$\ \ \ \ \\
\ \ \ \ $|1\rangle_{A_1}|1\rangle_{A_2}$\ \ \ \ &\ \ \ \ $|\Phi_{-}\rangle_{B_1B_2}$\ \ \ \ \\
\hline \hline
\end{tabular}}
\end{center}
\section{Pulse design via STA}
In this section, let us design the microwave pulse via STA. As we
can see from Sec. III, the effective Hamiltonians of Step 1, Step 2,
Step 4 and Step 5 of the Bell-state analysis have the form
\begin{eqnarray}\label{e17}
H(t)=\Omega_1(t)|\psi_1\rangle\langle\psi_2|+\Omega_2(t)|\psi_2\rangle\langle\psi_3|+H.c..
\end{eqnarray}
The required transformation is
$|\psi_1\rangle\leftrightarrow|\psi_3\rangle$. Here, to build up
evolution paths, we consider the transitionless tracking algorithm.
As pointed out by Ref. \cite{CYHSR6} that, while utilizing
transitionless tracking algorithm, not only the eigenstates of the
Hamiltonian but also a set of orthonormalized time-dependent
vectors, can be selected as the evolution paths. For the current
protocol, we select
\begin{eqnarray}\label{e18}
&|\xi_1(t)\rangle&=(\cos\theta\sin\varphi\cos\vartheta+\sin\theta\sin\vartheta)|\psi_1\rangle+i\cos\varphi\cos\vartheta|\psi_2\rangle+(\sin\theta\sin\varphi\cos\vartheta-\cos\theta\sin\vartheta)|\psi_3\rangle,\cr\cr
&|\xi_2(t)\rangle&=(\cos\theta\sin\varphi\sin\vartheta-\sin\theta\cos\vartheta)|\psi_1\rangle+i\cos\varphi\sin\vartheta|\psi_2\rangle+(\sin\theta\sin\varphi\sin\vartheta+\cos\theta\cos\vartheta)|\psi_3\rangle,\cr\cr
&|\xi_3(t)\rangle&=\cos\theta\cos\varphi|\psi_1\rangle-i\sin\varphi|\psi_2\rangle+\sin\theta\cos\varphi|\psi_3\rangle,
\end{eqnarray}
where, $\theta$, $\varphi$ and $\vartheta$ are three time-dependent
parameters. According to transitionless tracking algorithm, the
evolution operator and the Hamiltonian for the evolution paths shown
in Eq.~(\ref{e18}) could be derived by
\begin{eqnarray}\label{e19}
&U'(t)&=\sum\limits_{n=1}^{3}|\xi_n(t)\rangle\langle\xi_n(0)|,
\end{eqnarray}
and
\begin{eqnarray}\label{e20}
&H'(t)&=i\sum\limits_{n=1}^{3}|\dot{\xi}_n(t)\rangle\langle\xi_n(t)|\cr\cr&&
=(\dot{\varphi}\cos\theta+\dot{\vartheta}\sin\theta\cos\varphi)|\psi_1\rangle\langle\psi_2|+(\dot{\varphi}\sin\theta-\dot{\vartheta}\cos\theta\cos\varphi)|\psi_2\rangle\langle\psi_3|\cr\cr&&
+i(\dot{\theta}-\dot{\vartheta}\sin\varphi)|\psi_3\rangle\langle\psi_1|+H.c..
\end{eqnarray}
To make $H'(t)=H(t)$, it requires
\begin{eqnarray}\label{e21}
&\Omega_1(t)&=\dot{\varphi}\cos\theta+\dot{\vartheta}\sin\theta\cos\varphi,\cr\cr
&\Omega_2(t)&=\dot{\varphi}\sin\theta-\dot{\vartheta}\cos\theta\cos\varphi,\cr\cr
&\dot{\theta}&=\dot{\vartheta}\sin\varphi.
\end{eqnarray}
Considering the time interval $[0,T]$ and the boundary condition
$\varphi(0)=-\pi/2$, $\varphi(T)=\pi/2$, $\vartheta(0)+\theta(0)=0$,
$\vartheta(T)-\theta(T)=\pi/2$, we have $U'(0)=I$ ($I$ is the
identical operator) and
$U'(T)=|\psi_1\rangle\langle\psi_3|+|\psi_3\rangle\langle\psi_1|+|\psi_2\rangle\langle\psi_2|$.
Thus, we can complete the transformation
$|\psi_1\rangle\leftrightarrow|\psi_3\rangle$. With the boundary
condition and Eq.~(\ref{e21}), parameters $\theta$, $\vartheta$ and
$\varphi$ can be selected to be
\begin{eqnarray}\label{e22}
\varphi=-\frac{\pi}{2}\cos(\frac{\pi t}{T}),\ \ \ \vartheta=\pi/4,\
\ \ \theta=-\pi/4.
\end{eqnarray}
Then, $\Omega_1(t)$ and $\Omega_2(t)$ could be derived as
\begin{eqnarray}\label{e23}
\Omega_1(t)=\frac{\pi^2}{2\sqrt{2}T}\sin(\pi t/T),\ \ \
\Omega_2(t)=-\frac{\pi^2}{2\sqrt{2}T}\sin(\pi t/T).
\end{eqnarray}
Therefore, if we set
\begin{eqnarray}\label{e24}
&&\Omega_{03A_1}(t)=\frac{\sqrt{g_{23A_1}^2+g_{B_1}^2}}{g_{B_1}}\Omega_1(t),\
\ \
\Omega_{13A_1}(t)=\frac{\sqrt{g_{23A_1}^2+g_{B_1}^2}}{g_{B_1}}\Omega_2(t),\cr\cr
&&\Omega_{04A_1}(t-T)=\frac{\sqrt{g_{24A_1}^2+g_{B_2}^2}}{g_{B_2}}\Omega_1(t),\
\
\Omega_{14A_1}(t-T)=\frac{\sqrt{g_{24A_1}^2+g_{B_2}^2}}{g_{B_2}}\Omega_2(t),\cr\cr
&&\Omega_{03A_2}(t-3T)=\frac{\sqrt{g_{23A_2}^2+g_{B_1}^2}}{g_{B_1}}\Omega_1(t),\
\
\Omega_{13A_2}(t-3T)=\frac{\sqrt{g_{23A_2}^2+g_{B_1}^2}}{g_{B_1}}\Omega_2(t),\cr\cr
&&\Omega_{04A_2}(t-4T)=\frac{\sqrt{g_{24A_2}^2+g_{B_2}^2}}{g_{B_2}}\Omega_1(t),\
\
\Omega_{14A_2}(t-4T)=\frac{\sqrt{g_{24A_2}^2+g_{B_2}^2}}{g_{B_2}}\Omega_2(t),
\end{eqnarray}
the microwave pulses could be used in the Bell-state analysis.
As for the $\Omega_{B_j}(t)$, which used to perform a single-qubit
operation on SQUID qubit $B_j$ in Step 3, according to
Eq.~(\ref{e12}), it could be chosen as
\begin{eqnarray}\label{e25}
\Omega_{B_j}(t-2T)=\frac{\pi^2}{8T}\sin(\pi t/T).
\end{eqnarray}
Similarly, for $\Omega_{B_j}(t)$, which used to perform a
single-qubit operation on SQUID qubit $B_j$ in Step 6, according to
Eq.~(\ref{e15}), it could be chosen as
\begin{eqnarray}\label{e26}
\Omega_{B_j}(t-5T)=-\frac{\pi^2}{8T}\sin(\pi t/T).
\end{eqnarray}
For simplicity, we consider $g_{23A_j}=g_{24A_j}=g_{B_j}=g$. The
microwave pulses for the Bell-state analysis is shown in
Fig.~\ref{fig3}. The maximal value of the amplitudes of the pulses
$\Omega_{\max}=\pi^2/2T\simeq4.9348/T$.
\begin{figure}
\scalebox{0.6}{\includegraphics[scale=1]{Fig3.eps}}
\caption{Microwave pulses for the Bell-state analysis.}\label{fig3}
\end{figure}
\section{Numerical simulations}
In this section, the robustness of the protocol is checked via
numerical simulations. Before the numerical simulations, we first
define the success probability for distinguishing each Bell state.
Assume that the density operator of the system is $\rho(t)$.
According to Table. II, the success probabilities could be defined
as
\begin{eqnarray}\label{e27}
&&P(\Psi_+)=\langle00\Psi_+|\rho(6T)|00\Psi_+\rangle,\ \
P(\Psi_-)=\langle01\Psi_-|\rho(6T)|01\Psi_-\rangle,\cr\cr
&&P(\Phi_+)=\langle10\Phi_+|\rho(6T)|10\Phi_+\rangle,\ \
P(\Phi_-)=\langle11\Phi_-|\rho(6T)|11\Phi_-\rangle,
\end{eqnarray}
where,
\begin{eqnarray}\label{e28}
&&|00\Psi_+\rangle=|0\rangle_{A_1}|0\rangle_{A_2}|\Psi_+\rangle_{B_1B_2},\
\
|01\Psi_-\rangle=|0\rangle_{A_1}|1\rangle_{A_2}|\Psi_-\rangle_{B_1B_2},\cr\cr
&&|10\Phi_+\rangle=|1\rangle_{A_1}|0\rangle_{A_2}|\Phi_+\rangle_{B_1B_2},\
\
|11\Phi_-\rangle=|1\rangle_{A_1}|1\rangle_{A_2}|\Phi_-\rangle_{B_1B_2}.
\end{eqnarray}
Firstly, as we perform the numerical simulation based on the
original Hamiltonian shown in Eq.~(\ref{e5}), a suitable coupling
constant $g$ should be chosen. Thus, we plot $P(\Psi_+)$,
$P(\Psi_-)$, $P(\Phi_+)$ and $P(\Phi_-)$ versus $g$ in
Fig.~\ref{fig4}. As shown in Fig.~\ref{fig4}, the success
probability are very low when $g$ is small due to bit flip errors,
while they are near 1 when $g\geq30/T$. According to
Fig.~\ref{fig4}, we choose $g=66/T$, which gives
$\max\{|1-P(\Psi_+)|,|1-P(\Psi_-)|,|1-P(\Phi_+)|,|1-P(\Phi_-)|\}\leq1.1847\times10^{-5}$.
\begin{figure}
\scalebox{0.6}{\includegraphics[scale=1]{Fig4.eps}} \caption{Suceess
probabilities versus $g$. The solid-blue line: $P(\Psi_+)$; The
dotted-green line: $P(\Psi_-),P(\Phi_+)$; The dashed-red line
$P(\Phi_-)$.}\label{fig4}
\end{figure}
Secondly, we investigate the robustness of the protocol against
decoherence. The main decoherent factors for the protocol are: (i)
cavity dissipations for cavity fields $a_1$ and $a_2$ with decay
rates $\kappa_1$ and $\kappa_2$, respectively; (2) spontaneous
emissions from $|3\rangle_{A_j}$ ($|4\rangle_{A_j}$) to
$|0\rangle_{A_j}$, $|1\rangle_{A_j}$ and $|2\rangle_{A_j}$ with
spontaneous emission rates $\gamma_{30A_j}$ ($\gamma_{40A_j}$),
$\gamma_{31A_j}$ ($\gamma_{41A_j}$) and $\gamma_{32A_j}$
($\gamma_{42A_j}$), respectively; (3) spontaneous emissions from
$|2\rangle_{B_j}$ to $|0\rangle_{B_j}$ and $|1\rangle_{B_j}$ with
spontaneous emission rates $\gamma_{20B_j}$ and $\gamma_{21B_j}$,
respectively; (4) dephasings between $|3\rangle_{A_j}$
($|4\rangle_{A_j}$) and $|0\rangle_{A_j}$, $|1\rangle_{A_j}$,
$|2\rangle_{A_j}$ with dephasing rates $\gamma_{\phi30A_j}$
($\gamma_{\phi40A_j}$), $\gamma_{\phi31A_j}$ ($\gamma_{\phi41A_j}$),
$\gamma_{\phi32A_j}$ ($\gamma_{\phi42A_j}$), respectively; (5)
dephasings between $|2\rangle_{B_j}$ and $|0\rangle_{B_j}$,
$|1\rangle_{B_j}$ with spontaneous emission rates
$\gamma_{\phi20B_j}$, $\gamma_{\phi21B_j}$, respectively. Here, we
neglect the spontaneous emission and dephsing between
$|4\rangle_{A_j}$ and $|3\rangle_{A_j}$ and that between
$|2\rangle_{A_j}$ and $|0\rangle_{A_j}$ ($|1\rangle_{A_j}$), because
they are much weaker than spontaneous emissions and dephasings
between other levels of SQUID qubit $A_j$. Thus, the evolution of
the system could be described using a master equation
\begin{eqnarray}\label{e29}
\dot{\rho}(t)=i[\rho(t),H(t)]+\sum\limits_{p=1}^{34}[L_p\rho
L_p^{\dag}-\frac{1}{2}(L_p^{\dag}L_p\rho+\rho L_p^{\dag}L_p)],
\end{eqnarray}
where, $L_p$ ($p=1,2,3,...,34$) is the Lindblad operator as
\begin{eqnarray}\label{e30}
&&L_1=\sqrt{\gamma_{30A_1}}|0\rangle_{A_1}\langle3|,\ \
L_2=\sqrt{\gamma_{31A_1}}|1\rangle_{A_1}\langle3|,\ \
L_3=\sqrt{\gamma_{32A_1}}|2\rangle_{A_1}\langle3|,\cr\cr
&&L_4=\sqrt{\gamma_{40A_1}}|0\rangle_{A_1}\langle4|,\ \
L_5=\sqrt{\gamma_{41A_1}}|1\rangle_{A_1}\langle4|,\ \
L_6=\sqrt{\gamma_{42A_1}}|2\rangle_{A_1}\langle4|,\cr\cr
&&L_7=\sqrt{\gamma_{30A_2}}|0\rangle_{A_2}\langle3|,\ \
L_8=\sqrt{\gamma_{31A_2}}|1\rangle_{A_2}\langle3|,\ \
L_9=\sqrt{\gamma_{32A_2}}|2\rangle_{A_2}\langle3|,\cr\cr
&&L_{10}=\sqrt{\gamma_{40A_2}}|0\rangle_{A_2}\langle4|,\ \
L_{11}=\sqrt{\gamma_{41A_2}}|1\rangle_{A_2}\langle4|,\ \
L_{12}=\sqrt{\gamma_{42A_2}}|2\rangle_{A_2}\langle4|,\cr\cr
&&L_{13}=\sqrt{\gamma_{20B_1}}|0\rangle_{B_1}\langle2|,\ \
L_{14}=\sqrt{\gamma_{21B_1}}|1\rangle_{B_1}\langle2|,\ \
L_{15}=\sqrt{\gamma_{20B_2}}|0\rangle_{B_2}\langle2|,\cr\cr
&&L_{16}=\sqrt{\gamma_{21B_2}}|1\rangle_{B_2}\langle2|,\ \,
L_{17}=\sqrt{\kappa_1}a_1,\ \ L_{18}=\sqrt{\kappa_2}a_2\cr\cr
&&L_{19}=\sqrt{\gamma_{\phi30A_1}}(|3\rangle_{A_1}\langle3|-|0\rangle_{A_1}\langle0|),\
\
L_{20}=\sqrt{\gamma_{\phi31A_1}}(|3\rangle_{A_1}\langle3|-|1\rangle_{A_1}\langle1|),\cr\cr
&&L_{21}=\sqrt{\gamma_{\phi32A_1}}(|3\rangle_{A_1}\langle3|-|2\rangle_{A_1}\langle2|),\
\
L_{22}=\sqrt{\gamma_{\phi40A_1}}(|4\rangle_{A_1}\langle4|-|0\rangle_{A_1}\langle0|),\cr\cr
&&L_{23}=\sqrt{\gamma_{\phi41A_1}}(|4\rangle_{A_1}\langle4|-|1\rangle_{A_1}\langle1|),\
\
L_{24}=\sqrt{\gamma_{\phi42A_1}}(|4\rangle_{A_1}\langle4|-|2\rangle_{A_1}\langle2|),\cr\cr
&&L_{25}=\sqrt{\gamma_{\phi30A_2}}(|3\rangle_{A_2}\langle3|-|0\rangle_{A_2}\langle0|),\
\
L_{26}=\sqrt{\gamma_{\phi31A_2}}(|3\rangle_{A_2}\langle3|-|1\rangle_{A_2}\langle1|),\cr\cr
&&L_{27}=\sqrt{\gamma_{\phi32A_2}}(|3\rangle_{A_2}\langle3|-|2\rangle_{A_2}\langle2|),\
\
L_{28}=\sqrt{\gamma_{\phi40A_2}}(|4\rangle_{A_2}\langle4|-|0\rangle_{A_2}\langle0|),\cr\cr
&&L_{29}=\sqrt{\gamma_{\phi41A_2}}(|4\rangle_{A_2}\langle4|-|1\rangle_{A_2}\langle1|),\
\
L_{30}=\sqrt{\gamma_{\phi42A_2}}(|4\rangle_{A_2}\langle4|-|2\rangle_{A_2}\langle2|),\cr\cr
&&L_{31}=\sqrt{\gamma_{\phi20B_1}}(|2\rangle_{B_1}\langle2|-|0\rangle_{B_1}\langle0|),\
\
L_{32}=\sqrt{\gamma_{\phi21B_1}}(|2\rangle_{B_1}\langle2|-|1\rangle_{B_1}\langle1|),\cr\cr
&&L_{33}=\sqrt{\gamma_{\phi20B_2}}(|2\rangle_{B_2}\langle2|-|0\rangle_{B_2}\langle0|),\
\
L_{34}=\sqrt{\gamma_{\phi21B_2}}(|2\rangle_{B_2}\langle2|-|1\rangle_{B_2}\langle1|).
\end{eqnarray}
For brief discussions, we assume
$\gamma_{30A_j}=\gamma_{31A_j}=\gamma_{32A_j}=\gamma_{40A_j}=\gamma_{41A_j}=\gamma_{42A_j}=\gamma_{20B_j}=\gamma_{21B_j}=\gamma$,
$\gamma_{\phi30A_j}=\gamma_{\phi31A_j}=\gamma_{\phi32A_j}=\gamma_{\phi40A_j}=\gamma_{\phi41A_j}=\gamma_{\phi42A_j}=\gamma_{\phi20B_j}=\gamma_{\phi21B_j}=\gamma_\phi$,
$\kappa_1=\kappa_2=\kappa$. $P(\Psi_+)$, $P(\Psi_-)$, $P(\Phi_+)$
and $P(\Phi_-)$ versus $\gamma/\Omega_{\max}$,
$10\gamma_\phi/\Omega_{\max}$ and $\kappa/\Omega_{\max}$ are plotted
in Figs.~\ref{fig5}~(a-d), respectively.
\begin{figure}
\scalebox{0.6}{\includegraphics[scale=1]{Fig5.eps}} \caption{(a)
$P(\Psi_+)$, (b) $P(\Psi_-)$, (c) $P(\Phi_+)$ and (d) $P(\Phi_-)$
versus $\kappa/\Omega_{\max}$ (the solid-red line),
$\gamma/\Omega_{\max}$ (the dashed-green line),
$10\gamma_\phi/\Omega_{\max}$ (the dotted-blue line).}\label{fig5}
\end{figure}
As shown by Fig.~\ref{fig5}, the Bell-state analysis is quite robust
against the cavity decays. The success probabilities of
distinguishing each Bell state are still higher than 0.9996 when
$\kappa/\Omega_{\max}=0.01$. We can also find that, the Bell-state
analysis is more sensitive to the spontaneous emissions of SQUID
qubits. When $\gamma/\Omega_{\max}=0.01$, the success probabilities
of distinguishing each Bell state are all a little less than 0.97.
As for the dephasings, they are the most troublesome decoherent
factors. When $\gamma_\phi/\Omega_{\max}=0.001$, the success
probabilities of distinguishing each Bell state are all higher than
0.96. We considered experimentally realizable parameters
$\Omega_{\max}=2\pi\times6.37$MHz, $g=2\pi\times85.14$MHz,
$\kappa=1.32$MHz, $\gamma=0.40$MHz, $\gamma_\phi=0.20$MHz
\cite{YCPPRA67,YCPPRL92,XZLRMP85}, we have $P(\Psi_+)=0.9555$,
$P(\Psi_-)=0.9554$, $P(\Phi_+)=0.9554$ and $P(\Phi_-)=0.9554$.
Therefore, the protocol possesses high success probability to
distinguish each one of four Bell states.
\section{Conclusion}
In conclusion, we have proposed a protocol for complete Bell-state
analysis for superconducting-quantum-interference-device qubits.
Although the Bell-state analysis is composed of six steps, one just
need to use a sequence of sinusoidal microwave pulses as shown in
Fig.~(\ref{fig3}), without any additional operations. Therefore, the
operations are not difficult in experiments. After the six steps,
the information for distinguishing four Bell states are encoding on
two auxiliary qubits $A_1$ and $A_2$. Thus, we can read out the
information by detecting the states of $A_1$ and $A_2$. The
detections of SQUID qubits have been reported in protocols
\cite{BennettSST20,TakayanagiSM32}. In this paper, we have not
considered the detection efficiency, while for a real experiment,
they should be taken into account. Moreover, apart from the
advantages of SQUID qubits, the protocol holds some other
advantages:
(1) The Bell-state analysis is complete and nondestructive. In other
words, we can completely distinguish four Bell states without
destroying them. Thus the physical resource could be saved.
(2) As shown by the numerical simulations, under the current
experimental conditions, the protocol still possess high success
probability of distinguishing each Bell state when decoherence is
taken into account.
(3) Since STA is used in the pulse design, the operation speed of
the protocol may be faster than that using the adiabatic passages.
Thus, we hope the protocol could contribute to information readout
for quantum communications and quantum computations in
superconducting quantum networks.
\section*{Acknowledgement}
This work was supported by the National Natural Science Foundation
of China under Grants No. 11575045, No. 11374054 and No. 11674060,
and the Major State Basic Research Development Program of China
under Grant No. 2012CB921601.
\section*{Appendix: The derivation of the effective Hamiltonian}
Here, we would like to amply described the derivation of the
effective Hamiltonian $H_{eff1}(t)$ shown in Eq.~(\ref{e9}) from
Eqs. (\ref{e7}) and (\ref{e8}). According to the eigenstates of
$H_{c1}$ shown in Eq.~(\ref{e8}), we can rewrite Eq.~(\ref{e7}) as
\begin{eqnarray}\label{e31}
&H_{step1}(t)&=H_{m1}(t)+H_{c1},\cr\cr
&H_{m1}(t)&=\frac{\Omega_{03A_1}(t)}{\sqrt{2(g_{23A_1}^2+g_{B_1}^2})}|\bar{\psi}_1\rangle(\sqrt{2}g_{B1}\langle\bar{\phi}_0|+g_{23A_1}\langle\bar{\phi}_3|+g_{23A_1}\langle\bar{\phi}_4|)\cr\cr&&
+\frac{\Omega_{03A_1}(t)}{\sqrt{2}}|\bar{\psi}_6\rangle(\langle\bar{\phi}_1|+\langle\bar{\phi}_2|)+\frac{\Omega_{13A_1}(t)}{\sqrt{2}}|\bar{\psi}_9\rangle(\langle\bar{\phi}_1|+\langle\bar{\phi}_2|)\cr\cr&&
+\frac{\Omega_{13A_1}(t)}{\sqrt{2(g_{23A_1}^2+g_{B_1}^2})}|\bar{\psi}_5\rangle(\sqrt{2}g_{B1}\langle\bar{\phi}_0|+g_{23A_1}\langle\bar{\phi}_3|+g_{23A_1}\langle\bar{\phi}_4|)+H.c.,\cr\cr
&H_{c1}&=g_{23A_1}(|\bar{\phi}_1\rangle\langle\bar{\phi}_1|-|\bar{\phi}_2\rangle\langle\bar{\phi}_2|)+\sqrt{g_{23A_1}^2+g_{B_1}^2}(|\bar{\phi}_3\rangle\langle\bar{\phi}_3|-|\bar{\phi}_4\rangle\langle\bar{\phi}_4|).
\end{eqnarray}
Considering $H_{c1}$ as a free Hamiltonian, we perform the picture
transformation $\bar{U}=e^{-iH_{c1}t}$ on Eq.~(\ref{e31}). The
Hamiltonian after the transformation is
\begin{eqnarray}\label{e32}
&H'_{step1}(t)&=\frac{\Omega_{03A_1}(t)}{\sqrt{2(g_{23A_1}^2+g_{B_1}^2})}|\bar{\psi}_1\rangle(\sqrt{2}g_{B1}\langle\bar{\phi}_0|+g_{23A_1}\langle\bar{\phi}_3|e^{-i\sqrt{g_{23A_1}^2+g_{B_1}^2}t}+g_{23A_1}\langle\bar{\phi}_4|e^{i\sqrt{g_{23A_1}^2+g_{B_1}^2}t})\cr\cr&&
+\frac{\Omega_{13A_1}(t)}{\sqrt{2(g_{23A_1}^2+g_{B_1}^2})}|\bar{\psi}_5\rangle(\sqrt{2}g_{B1}\langle\bar{\phi}_0|+g_{23A_1}\langle\bar{\phi}_3|e^{-i\sqrt{g_{23A_1}^2+g_{B_1}^2}t}+g_{23A_1}\langle\bar{\phi}_4|e^{i\sqrt{g_{23A_1}^2+g_{B_1}^2}t})\cr\cr&&
+\frac{\Omega_{03A_1}(t)}{\sqrt{2}}|\bar{\psi}_6\rangle(\langle\bar{\phi}_1|e^{-ig_{23A_1}t}+\langle\bar{\phi}_2|e^{ig_{23A_1}t})+\frac{\Omega_{13A_1}(t)}{\sqrt{2}}|\bar{\psi}_9\rangle(\langle\bar{\phi}_1|e^{-ig_{23A_1}t}+\langle\bar{\phi}_2|e^{ig_{23A_1}t})\cr\cr&&+H.c..
\end{eqnarray}
Under the condition $\Omega_{03A_1}(t),\Omega_{13A_1}(t)\ll
g_{23A_1},g_{B_1}$, we neglect the terms of high frequency
oscillations and obtain the effective Hamiltonian
\begin{eqnarray}\label{e33}
H_{eff1}(t)=\frac{g_{B_1}}{\sqrt{g_{23A_1}^2+g_{B_1}^2}}[\Omega_{03A_1}(t)|\bar{\psi}_1\rangle\langle\bar{\phi}_0|+\Omega_{13A_1}(t)|\bar{\psi}_5\rangle\langle\bar{\phi}_0|]+H.c..
\end{eqnarray}
In similar way, we can derive the effective Hamiltonian
$H_{eff2}(t)$ shown in Eq.~(\ref{e11}) from Eqs. (\ref{e9}) and
(\ref{e10}).
|
\section{Generalized Tribonacci sequence}
We consider the generalized Tribonacci Sequence, $\{V_{n}(V_{0},V_{1},V_{2};r,s,t)\}_{n\geq0}$, or briefly $\{V_{n}\}_{n\geq0}$, defined as follows:
\begin{equation}\label{eq:1}
V_{n}=rV_{n-1}+sV_{n-2}+tV_{n-3},\ n\geq3,
\end{equation}
where $V_{0}=a$, $V_{1}=b$, $V_{2}=c$ are arbitrary integers and $r$, $s$, $t$, are real numbers.
This sequence has been studied by Shannon and Horadam \cite{Sha}, Yalavigi \cite{Ya} and Pethe \cite{Pe}. If we set $r=s=t=1$ and $V_{0}=0=V_{1}$, $V_{2}=1$, then $\{V_{n}\}$ is the well-known Tribonacci sequence which has been considered extensively (see, for example, \cite{Fe}).
As the elements of this Tribonacci-type number sequence provide third order iterative relation, its characteristic equation is $x^{3}-rx^{2}-sx-t=0$, whose roots are $\alpha=\alpha(r,s,t)=\frac{r}{3}+A+B$, $\omega_{1}=\frac{r}{3}+\epsilon A+\epsilon^{2} B$ and $\omega_{2}=\frac{r}{3}+\epsilon^{2}A+\epsilon B$, where $$A=\left(\frac{r^{3}}{27}+\frac{rs}{6}+\frac{t}{2}+\sqrt{\Delta}\right)^{\frac{1}{3}},\ B=\left(\frac{r^{3}}{27}+\frac{rs}{6}+\frac{t}{2}-\sqrt{\Delta}\right)^{\frac{1}{3}},$$ with $\Delta=\Delta(r,s,t)=\frac{r^{3}t}{27}-\frac{r^{2}s^{2}}{108}+\frac{rst}{6}-\frac{s^{3}}{27}+\frac{t^{2}}{4}$ and $\epsilon=-\frac{1}{2}+\frac{i\sqrt{3}}{2}$.
In this paper, $\Delta(r,s,t)>0$, then the equation (\ref{eq:1}) has one real and two nonreal solutions, the latter being conjugate complex. Thus, the Binet formula for the generalized Tribonacci numbers can be expressed as:
\begin{equation}\label{eq:2}
V_{n}=\frac{P\alpha^{n}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{Q\omega_{1}^{n}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{R\omega_{2}^{n}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})},
\end{equation}
where $P=c-(\omega_{1}+\omega_{2})b+\omega_{1}\omega_{2}a$, $Q=c-(\alpha+\omega_{2})b+\alpha\omega_{2}a$ and $R=c-(\alpha+\omega_{1})b+\alpha\omega_{1}a$.
In fact, the generalized Tribonacci sequence is the generalization of the well-known sequences like Tribonacci, Padovan, Narayana and third order Jacobsthal. For example, $\{V_{n}(0,0,1;1,1,1)\}_{n\geq0}$, $\{V_{n}(0,1,0;0,1,1)\}_{n\geq0}$, are Tribonacci and Padovan sequences, respectively. The Binet formula for the generalized Tribonacci sequence is expressed as follows:
\begin{lemma}
The Binet formula for the generalized Tribonacci sequence is:
\begin{equation}\label{eq:3}
V_{n}=cU_{n}+(bs+at)U_{n-1}+btU_{n-2},
\end{equation}
where $a$, $b$ and $c$ are initial values and
\begin{equation}\label{eq:4}
U_{n}=\frac{\alpha^{n}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{\omega_{1}^{n}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{\omega_{2}^{n}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})}.
\end{equation}
\end{lemma}
\begin{proof}
The validity of this formula can be confirmed using the recurrence relation. Furthermore, $\{U_{n}\}_{n\geq0}=\{V_{n}(0,0,1;r,s,t)\}_{n\geq0}$.
\end{proof}
On the other hand, Horadam \cite{Ho} introduced the $n$-th Fibonacci and the $n$-th Lucas quaternion as follow:
\begin{equation}\label{equa:1}
Q_{n}=F_{n}+iF_{n+1}+jF_{n+2}+kF_{n+3},
\end{equation}
\begin{equation}\label{equa:2}
K_{n}=L_{n}+iL_{n+1}+jL_{n+2}+kL_{n+3},
\end{equation}
respectively. Here $F_{n}$ and $L_{n}$ are the $n$-th Fibonacci and Lucas numbers, respectively. Furthermore, the basis $i,j,k$ satisfies the following rules:
\begin{equation}\label{equa:3}
i^{2}=j^{2}=k^{2}=-1,\ ijk=-1.
\end{equation}
Note that the rules (\ref{equa:3}) imply $ij=-ji=k$, $jk=-kj=i$ and $ki=-ik=j$. In general, a quaternion is a hyper-complex number and is defined by $q=q_{0}+iq_{1}+jq_{2}+kq_{3}$, where $i,j,k$ are as in (\ref{equa:3}) and $q_{0},q_{1},q_{2}$ and $q_{3}$ are real numbers. Note that we can write $q=q_{0}+u$ where $u= iq_{1}+jq_{2}+kq_{3}$. The conjugate of the quaternion $q$ is denoted by $q^{*}=q_{0}-u$. The norm of a quaternion $q$ is defined by $Nr^{2}(q)=q_{0}^{2}+q_{1}^{2}+q_{2}^{2}+q_{3}^{2}$.
In this paper we introduce and study the generalized Tribonacci quaternions. We show that the new quaternion sequence that we introduced includes the previously introduced Tribonacci, Padovan, Narayana and Third order Jacobsthal quaternion sequences. We obtained the Binet formula and calculated a quadratic identity, summation formula and the norm value for this new quaternion sequence. Furthermore, we describe their properties also using a matrix representation.
\section{Generalized Tribonacci quaternions}
In this section, we define a new relation which generalizes all the third order recurrence relations which have been studied so far. We present some quadratic identities and some results corresponding to the results of previously studied sequences.
First, for $n\geq 2$ and $p$, $q$ integers, using $W_{n}=pW_{n-1}+qW_{n-2}$, $W_{0}=a$ and $W_{1}=b$, Halici and Karata\c{s} in \cite{Ha2} showed the presence of the following iterative relation $W_{n}=bT_{n}+aqT_{n-1}$, where $T_{n}=pT_{n-1}+qT_{n-2}$, $T_{0}=0$ and $T_{1}=1$. The above formula is a type of generalization for Fibonacci numbers. Furthermore, the authors made a new generalization for the Fibonacci quaternions with the help of this relation, which is expressed below:
\begin{equation}\label{eq:5}
Q_{w,n}=W_{n}+W_{n+1}i+W_{n+2}j+W_{n+3}k,\ n\geq0,
\end{equation}
where $W_{n}$ is the $n$-th Horadam number.
Now, let us propose a new type of quaternion similar to generalized Tribonacci:
\begin{equation}\label{eq:6}
Q_{v,n}=V_{n}+V_{n+1}i+V_{n+2}j+V_{n+3}k,\ n\geq0.
\end{equation}
Here, $V_{n}$ is the $n$-th generalized Tribonacci number defined by Eq. (\ref{eq:1}). This new quaternion can be called generalized Tribonacci quaternion due to its coefficients. The following iterative relation is obtained after some basic calculations:
\begin{equation}\label{eq:7}
Q_{v,n}=rQ_{v,n-1}+sQ_{v,n-2}+tQ_{v,n-3},\ n\geq3.
\end{equation}
The Eq. (\ref{eq:6}) generalizes different quaternionic sequences, which are known as quaternions similar to Tribonacci in the literature. Two of the most well-known of these are Narayana and Third order Jacobsthal quaternions. When we calculate the initial values of the sequence in Eq. (\ref{eq:7}), we find:
\begin{equation}\label{eq:8}
\left\{
\begin{array}{c}
Q_{v,0}=a+bi+cj+(rc+sb+ta)k, \\
Q_{v,1}=b+ci+(rc+sb+ta)j+((r^{2}+s)c+(t+rs)b+rta)k, \\
Q_{v,2}=c+(rc+sb+ta)i+((r^{2}+s)c+(t+rs)b+rta)j+V_{5}k.
\end{array}
\right.
\end{equation}
where $V_{5}=(r^{3}+2rs+t)c+(r^{2}s+s^{2}+rt)b+(r^{2}t+st)a$.
It must be noted that when $V_{n}(0,1,1;1,1,2)$ are taken instead of $V_{n}(r,s,t)$ in Eqs. (\ref{eq:8}), $Q_{v,0}=i+j+2k$, $Q_{v,0}=1+i+2j+5k$ and $Q_{v,0}=1+2i+5j+9k$ is obtained. Here $Q_{v,n}=QJ_{n}^{(3)}$ are Third order Jacobsthal quaternions found by Cerda-Morales in \cite{Ce}. For this case, the roots of the characteristic equation of the sequence are $x=2$ and $x_{1,2}=\frac{-1\pm i\sqrt{3}}{2}$. Note that the latter two are the complex conjugate cube roots of unity. In \cite{Ce}, with the help of the roots above, Binet formula for Third order Jacobsthal quaternions was given as follows:
\begin{equation}\label{eq:9}
JQ_{n}^{(3)}=\frac{1}{7}\left[2^{n+1}\underline{2}-\left(1+\frac{2i\sqrt{3}}{3}\right)x_{1}^{n}\underline{x_{1}}-\left(1-\frac{2i\sqrt{3}}{3}\right)x_{2}^{n}\underline{x_{2}}\right],
\end{equation}
where $x_{1,2}$ are the solutions of the equation $t^{2}+t+1=0$, and $\underline{2}=1+2i+4j+8k$, $\underline{x_{1}}=1+x_{1}i+x_{1}^{2}j+k$ and $\underline{x_{2}}=1+x_{2}i+x_{2}^{2}j+k$.
In the following theorem, we derive the Binet formula for generalized Tribonacci quaternions.
\begin{theorem}
Binet formula for generalized Tribonacci quaternions is
\begin{equation}\label{eq:10}
Q_{v,n}=\frac{P\underline{\alpha}\alpha^{n}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{Q\underline{\omega_{1}}\omega_{1}^{n}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{R\underline{\omega_{2}}\omega_{2}^{n}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})},
\end{equation}
where $P$, $Q$ and $R$ as in Eq. (\ref{eq:2}), and $\underline{\alpha}=1+\alpha i+\alpha^{2} i+\alpha^{3} k$, $\underline{\omega_{1}}=1+\omega_{1} i+\omega_{1}^{2} i+\omega_{1}^{3} k$ and $\underline{\omega_{2}}=1+\omega_{2} i+\omega_{2}^{2} i+\omega_{2}^{3} k$.
\end{theorem}
\begin{proof}
Using the definition of generalized Tribonacci quaternions $\{Q_{v,n}\}_{n\geq0}$, the following equation can be written:
\begin{align*}
Q_{v,n}&=\frac{P\alpha^{n}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{Q\omega_{1}^{n}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{R\omega_{2}^{n}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})}\\
&\ \ + \left(\frac{P\alpha^{n+1}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{Q\omega_{1}^{n+1}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{R\omega_{2}^{n+1}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})}\right)i\\
&\ \ + \left(\frac{P\alpha^{n+2}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{Q\omega_{1}^{n+2}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{R\omega_{2}^{n+2}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})}\right)j\\
&\ \ + \left(\frac{P\alpha^{n+3}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}-\frac{Q\omega_{1}^{n+3}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}+\frac{R\omega_{2}^{n+3}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})}\right)k.
\end{align*}
Then, some basic calculations are made and Eq. (\ref{eq:10}) is found.
\end{proof}
In the following table, we give the formula $\alpha(r,s,t)$ for some special third order sequences.
\begin{table}[ht]
\caption{Formulas for $\alpha(r,s,t)$ according to different values of $r,s,t$.}
\centering
\begin{tabular}{c c c}
\hline
Sequences & $(V_{0},V_{1},V_{2};r,s,t)$ & $\alpha(r,s,t)$\\ [0.5ex]
\hline
Narayana & $\left(0,1,1;1,0,1\right)$ & $\frac{1}{3}+\sqrt[3]{\frac{29}{54}+\frac{\sqrt{93}}{18}}+\sqrt[3]{\frac{29}{54}-\frac{\sqrt{93}}{18}}$\\
Tribonacci-Lucas & $\left(0,0,1;1,1,1\right)$ & $\frac{1}{3}+\sqrt[3]{\frac{19}{27}+\frac{\sqrt{33}}{9}}+\sqrt[3]{\frac{19}{27}-\frac{\sqrt{33}}{9}}$\\
Padovan-Perrin & $\left(0,1,0;0,1,1\right)$ & $\sqrt[3]{\frac{1}{2}+\frac{\sqrt{69}}{18}}+\sqrt[3]{\frac{1}{2}-\frac{\sqrt{69}}{18}}$\\
Third order Jacobsthal & $\left(0,1,1;1,1,2\right)$ & $\frac{4}{3}$ \\ [1ex]
\hline
\end{tabular}
\label{table:1}
\end{table}
The values given in the table are the same with the values given in some previous studies (see, for example, \cite{Di,Fe,Pe,Sha}). In other words, the formula we give is a generalization. For example, the Fibonacci-Narayana quaternions given in the first row of the table were also defined by Flaut and Shpakivskyi \cite{Fla1} by giving suitable initial values and by listing its properties. Also, Third order Jacobsthal quaternions in the fourth row of the table were defined by Cerda-Morales \cite{Ce}, who examined its properties. As a result, here we generalized different Tribonacci-like quaternion sequences in one formula.
In the following theorem we present the generating function for generalized Tribonacci quaternions.
\begin{theorem}
The generating function for the generalized Tribonacci quaternion $Q_{v,n}$ is
\begin{equation}\label{eq:11}
g(x)=\frac{Q_{v,0}+(Q_{v,1}-rQ_{v,0})x+(Q_{v,2}-rQ_{v,1}-sQ_{v,0})x^{2}}{1-rx-sx^{2}-tx^{3}}
\end{equation}
\end{theorem}
\begin{proof}
Assuming that the generating function of the generalized Tribonacci quaternion sequence $\{Q_{v,n}\}_{n\geq 0}$ has the form $g(x)=\sum_{n\geq0}Q_{v,n}x^{n}$, we obtain that
\begin{align*}
(1-rx-&sx^{2}-tx^{3})\sum_{n\geq0}Q_{v,n}x^{n}\\
&=Q_{v,0}+Q_{v,1}x+Q_{v,2}x^{2}+Q_{v,3}x^{3}+\cdots\\
&\ \ -rQ_{v,0}x-rQ_{v,1}x^{2}-rQ_{v,2}x^{3}-rQ_{v,3}x^{4}-\cdots\\
&\ \ -sQ_{v,0}x^{2}-sQ_{v,1}x^{3}-sQ_{v,2}x^{4}-sQ_{v,3}x^{5}-\cdots\\
&\ \ -tQ_{v,0}x^{3}-tQ_{v,1}x^{4}-tQ_{v,2}x^{5}-tQ_{v,3}x^{6}-\cdots\\
&=Q_{v,0}+(Q_{v,1}-rQ_{v,0})x+(Q_{v,2}-rQ_{v,1}-sQ_{v,0})x^{2},
\end{align*}
since $Q_{v,n}=rQ_{v,n-1}+sQ_{v,n-2}+tQ_{v,n-3}$, $n\geq3$ and the coefficients of $x^{n}$ for $n\geq 3$ are equal with zero. Then, we get $$\sum_{n\geq0}Q_{v,n}x^{n}=\frac{Q_{v,0}+(Q_{v,1}-rQ_{v,0})x+(Q_{v,2}-rQ_{v,1}-sQ_{v,0})x^{2}}{1-rx-sx^{2}-tx^{3}}.$$ The theorem is proved.
\end{proof}
In the following table, we examine some special cases of generating functions given in Eq. (\ref{eq:11}).
\begin{table}[ht]
\caption{Generating functions according to initial values.}
\centering
\begin{tabular}{ c c}
\hline
Narayana quaternions & $\frac{x+i+(1+x^{2})j+(1+x+x^{2})k}{1-x-x^{3}}$\\
Tribonacci quaternions & $\frac{x^{2}+xi+j+(1+x+x^{2})k}{1-x-x^{2}-x^{3}}$ \\
Padovan-Perrin quaternions & $\frac{x+i+(x+x^{2})j+(1+x)k}{1-x^{2}-x^{3}}$ \\
Third order Jacobsthal quaternions & $\frac{x+i+(1+x+2x^{2})j+(2+3x+2x^{2})k}{1-x-x^{2}-2x^{3}}$ \\ [1ex]
\hline
\end{tabular}
\label{table:2}
\end{table}
It must be restated that the calculations in the table above were made according to the initial condition of each sequence. The calculation in the first row was made by Flaut and Shpakivskyi \cite{Fla1}. The sequence in the fourth row of the table was studied by Cerda-Morales in \cite{Ce}.
Now, let us write the formula which gives the summation of the first $n$ generalized Tribonacci numbers and quaternions.
\begin{lemma}
For every integer $n\geq 0$, we have:
\begin{equation}\label{eq:12}
\sum_{l=0}^{n}V_{l}=\frac{1}{\delta(r,s,t)}\left(V_{n+2}+(1-r)V_{n+1}+tV_{n}+(r+s-1)a+(r-1)b-c\right),
\end{equation}
where $\delta=\delta(r,s,t)=r+s+t-1$ and $V_{n}$ denote the $n$-th term of the generalized Tribonacci numbers.
\end{lemma}
\begin{proof}
By the principal of mathematics induction since $\delta(r,s,t)V_{0}$ is equal to $V_{2}+(1-r)V_{1}+tV_{0}+(r+s-1)a+(r-1)b-c=\delta(r,s,t)a$, then the statement is true when $n=0$. Assume the given statement is true for all integer $n$. Then,
\begin{align*}
\delta(r,s,t)&\sum_{l=0}^{n+1}V_{l}=\delta(r,s,t)\sum_{l=0}^{n}V_{l}+\delta(r,s,t)V_{n+1}\\
&=V_{n+2}+(1-r)V_{n+1}+tV_{n}+(r+s-1)a+(r-1)b-c\\
&\ \ +(r+s+t-1)V_{n+1}\\
&=V_{n+3}+(1-r)V_{n+2}+tV_{n+1}+(r+s-1)a+(r-1)b-c.
\end{align*}
Thus, by the principal of mathematics induction the formula holds for every integer $n\geq0$.
\end{proof}
\begin{theorem}
The summation formula for generalized Tribonacci quaternions is as follows:
\begin{equation}\label{eq:13}
\sum_{l=0}^{n}Q_{v,l}=\frac{1}{\delta(r,s,t)}\left(Q_{v,n+2}+(1-r)Q_{v,n+1}+tQ_{v,n}+\omega(r,s,t)\right),
\end{equation}
where $\omega(r,s,t)=\lambda+i(\lambda-\delta a)+j(\lambda-\delta(a+b)+k(\lambda-\delta(a+b+c)$ and $\lambda=\lambda(r,s,t)$ is equal to $(r+s-1)a+(r-1)b-c$.
\end{theorem}
\begin{proof}
Using Eq. (\ref{eq:6}), we have
\begin{align*}
\sum_{l=0}^{n}Q_{v,l}&=\sum_{l=0}^{n}V_{l}+i\sum_{l=0}^{n}V_{l+1}+j\sum_{l=0}^{n}V_{l+2}+k\sum_{l=0}^{n}V_{l+3}\\
&=(V_{0}+V_{1}+V_{2}+\cdots+V_{n})+i(V_{1}+V_{2}+V_{3}+\cdots+V_{n+1})\\
&\ \ +j(V_{2}+V_{3}+V_{4}+\cdots+V_{n+2})+k(V_{3}+V_{4}+V_{5}+\cdots+V_{n+3}).
\end{align*}
Since from Eq. (\ref{eq:12}) and using the notation $\lambda(r,s,t)=(r+s-1)a+(r-1)b-c$, we can write
\begin{align*}
\delta(r,s,t)\sum_{l=0}^{n}Q_{v,l}&=V_{n+2}+(1-r)V_{n+1}+tV_{n}+\lambda(r,s,t)\\
&\ \ +i\left(V_{n+3}+(1-r)V_{n+2}+tV_{n+1}+\lambda(r,s,t)-\delta a\right)\\
&\ \ +j\left(V_{n+4}+(1-r)V_{n+3}+tV_{n+2}+\lambda(r,s,t)-\delta(a+b)\right)\\
&\ \ +k\left(V_{n+5}+(1-r)V_{n+4}+tV_{n+3}+\lambda(r,s,t)-\delta(a+b+c)\right)\\
&=Q_{v,n+2}+(1-r)Q_{v,n+1}+tQ_{v,n}+\omega(r,s,t),
\end{align*}
where $\omega(r,s,t)=\lambda+i(\lambda-\delta a)+j(\lambda-\delta(a+b)+k(\lambda-\delta(a+b+c)$. Finally, $$\sum_{l=0}^{n}Q_{v,l}=\frac{1}{\delta(r,s,t)}\left(Q_{v,n+2}+(1-r)Q_{v,n+1}+tQ_{v,n}+\omega(r,s,t)\right).$$
The theorem is proved.
\end{proof}
The summation formula in Eq. (\ref{eq:13}) gives the sum of the elements in the quaternion sequences which have been found in the studies conducted so far. This can be seen in the following table.
\begin{table}[ht]
\caption{Summation formulas according to initial values.}
\centering
\begin{tabular}{ c c }
\hline
Narayana quaternions & $Q_{v,n+3}-(1+i+2j+3k)$\\
Tribonacci quaternions & $\frac{1}{2}\left(Q_{v,n+2}+Q_{v,n}-(1+i+j+3k)\right)$ \\
Padovan-Perrin quaternions & $Q_{v,n+5}-(1+i+2j+2k)$ \\
Third order Jacobsthal quaternions & $\frac{1}{3}\left(Q_{v,n+2}+2Q_{v,n}-(1+i+4j+7k)\right)$ \\ [1ex]
\hline
\end{tabular}
\label{table:3}
\end{table}
Now, we present the formula which gives the norms for generalized Tribonacci quaternions. If we use the definition norm, then we obtain $Nr^{2}(Q_{v,n})=\sum_{l=0}^{3}V_{n+l}^{2}$. Moreover, by the Binet formula (\ref{eq:2}) we have $$\phi V_{n}=(\omega_{1}-\omega_{2})P\alpha^{n}-(\alpha-\omega_{2})Q\omega_{1}^{n}+(\alpha-\omega_{1})R\omega_{2}^{n},$$ where $\phi=\phi(\alpha,\omega_{1},\omega_{2})=(\alpha-\omega_{1})(\alpha-\omega_{2})(\omega_{1}-\omega_{2})$. Then,
\begin{align*}
\phi^{2} V_{n}^{2}&=(\omega_{1}-\omega_{2})^{2}P^{2}\alpha^{2n}+(\alpha-\omega_{2})^{2}Q^{2}\omega_{1}^{2n}+(\alpha-\omega_{1})^{2}R^{2}\omega_{2}^{2n}\\
&\ \ -2(\omega_{1}-\omega_{2})(\alpha-\omega_{2})PQ(\alpha\omega_{1})^{n}+2(\omega_{1}-\omega_{2})(\alpha-\omega_{1})PR(\alpha\omega_{2})^{n}\\
&\ \ -2(\alpha-\omega_{1})(\alpha-\omega_{2})QR(\omega_{1}\omega_{2})^{n}
\end{align*}
and
\begin{align*}
\phi^{2} Nr^{2}&(Q_{v,n})=\phi^{2} (V_{n}^{2}+V_{n+1}^{2}+V_{n+2}^{2}+V_{n+3}^{2})\\
&=(\omega_{1}-\omega_{2})^{2}P^{2}\overline{\alpha}\alpha^{2n}+(\alpha-\omega_{2})^{2}Q^{2}\overline{\omega_{1}}\omega_{1}^{2n}+(\alpha-\omega_{1})^{2}R^{2}\overline{\omega_{2}}\omega_{2}^{2n}\\
&\ \ -2(\omega_{1}-\omega_{2})(\alpha-\omega_{2})PQ \underline{\alpha\omega_{1}}(\alpha\omega_{1})^{n}-2(\omega_{1}-\omega_{2})(\omega_{1}-\alpha)PR\underline{\alpha\omega_{2}}(\alpha\omega_{2})^{n}\\
&\ \ -2(\alpha-\omega_{1})(\alpha-\omega_{2})QR\underline{\omega_{1}\omega_{2}}(\omega_{1}\omega_{2})^{n},
\end{align*}
where $\overline{\alpha}=1+\alpha^{2}+\alpha^{4}+\alpha^{6}$, $\overline{\omega_{1,2}}=1+\omega_{1,2}^{2}+\omega_{1,2}^{4}+\omega_{1,2}^{6}$, $\underline{\alpha \omega_{1,2}}=1+ \alpha \omega_{1,2}+ (\alpha \omega_{1,2})^{2}+ (\alpha \omega_{1,2})^{3}$ and $\underline{\omega_{1}\omega_{2}}=1+\omega_{1}\omega_{2}+(\omega_{1}\omega_{2})^{2}+(\omega_{1}\omega_{2})^{3}$.
Then, we obtain
\begin{theorem}
The norm value for generalized Tribonacci quaternions is given with the following formula:
\begin{equation}\label{eq:18}
Nr^{2}(Q_{v,n})=\frac{1}{\phi^{2}}\left(
\begin{array}{c}
(\omega_{1}-\omega_{2})^{2}P^{2}\overline{\alpha}\alpha^{2n}+(\alpha-\omega_{2})^{2}Q^{2}\overline{\omega_{1}}\omega_{1}^{2n} \\
+(\alpha-\omega_{1})^{2}R^{2}\overline{\omega_{2}}\omega_{2}^{2n}-2K%
\end{array}%
\right)
\end{equation}
where $K=(\omega_{1}-\omega_{2})(\alpha-\omega_{2})PQ \underline{\alpha\omega_{1}}(\alpha\omega_{1})^{n}+(\omega_{1}-\omega_{2})(\omega_{1}-\alpha)PR\underline{\alpha\omega_{2}}(\alpha\omega_{2})^{n}+(\alpha-\omega_{1})(\alpha-\omega_{2})QR\underline{\omega_{1}\omega_{2}}(\omega_{1}\omega_{2})^{n}$.
\end{theorem}
Now, we will give binomial summation of generalized Tribonacci quaternions as follows:
\begin{theorem}
For $n\geq 0$, we have the equality
\begin{equation}\label{eq:19}
Q_{v,3n}=\sum_{l=0}^{n}\sum_{m=0}^{l}\binom{n}{l}\binom{l}{m}r^{m}s^{l-m}t^{n-l}Q_{v,l+m}.
\end{equation}
\end{theorem}
\begin{proof}
Let $\alpha$ stand for a root of the characteristic equation of Eq. (\ref{eq:1}). Then, we have $\alpha^{3}=r\alpha^{2}+s\alpha+t$ and we can write by considering binomial expansion with $t\neq0$:
\begin{align*}
\left(\frac{\alpha^{3}}{t}\right)^{n}&=\sum_{l=0}^{n}\binom{n}{l}\left(\frac{\alpha^{3}}{t}-1\right)^{l}=\sum_{l=0}^{n}\binom{n}{l}\left(\frac{r}{t}\alpha^{2}+\frac{s}{t}\alpha\right)^{l}\\
&=\sum_{l=0}^{n}\binom{n}{l}\sum_{m=0}{l}\binom{l}{m}\left(\frac{r}{t}\alpha^{2}\right)^{m}\left(\frac{s}{t}\alpha\right)^{l-m}\\
&=\sum_{l=0}^{n}\sum_{m=0}^{l}\binom{n}{l}\binom{l}{m}\left(\frac{r}{s}\right)^{m}\left(\frac{s}{t}\right)^{l}\alpha^{l+m}.
\end{align*}
If we replace to $\omega_{1}$ and $\omega_{2}$ by $\alpha$ and rearrange, then we obtain
\begin{align*}
\frac{Q_{v,3n}}{t^{n}}&=\frac{P\underline{\alpha}}{(\alpha-\omega_{1})(\alpha-\omega_{2})}\frac{\alpha^{3n}}{t^{n}}-\frac{Q\underline{\omega_{1}}}{(\alpha-\omega_{1})(\omega_{1}-\omega_{2})}\frac{\omega_{1}^{3n}}{t^{n}}+\frac{R\underline{\omega_{2}}}{(\alpha-\omega_{2})(\omega_{1}-\omega_{2})}\frac{\omega_{2}^{3n}}{t^{n}}\\
&=\sum_{l=0}^{n}\sum_{m=0}^{l}\binom{n}{l}\binom{l}{m}\left(\frac{r}{s}\right)^{m}\left(\frac{s}{t}\right)^{l}Q_{v,l+m}.
\end{align*}
where $P$, $Q$ and $R$ as in Eq. (\ref{eq:2}), and $\underline{\alpha}$, $\underline{\omega_{1}}$ and $\underline{\omega_{2}}$ as in Eq. (\ref{eq:10}).
\end{proof}
\section{Matrix Representation of Generalized Tribonacci Quaternions}
For our purposes, the most useful technique for generating $\{V_{n}\}$ is by means of what we call the $S$-matrix which has been defined and used in \cite{Sha} and is a generalization of the $R$-matrix defined in \cite{Wa}. The $S$-matrix is defined as
\begin{equation}\label{eq:14}
\left[
\begin{array}{c}
V_{n+2} \\
V_{n+1} \\
V_{n}%
\end{array}%
\right]=\left[
\begin{array}{ccc}
r & s & t \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}%
\right]^{n}\left[
\begin{array}{c}
V_{2} \\
V_{1} \\
V_{0}%
\end{array}%
\right]
\end{equation}
and
\begin{equation}\label{eq:15}
S^{n}=\left[
\begin{array}{ccc}
r & s & t \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}%
\right]^{n}=\left[
\begin{array}{ccc}
U_{n+2} & sU_{n+1}+tU_{n} & tU_{n+1} \\
U_{n+1} & sU_{n}+tU_{n-1} & tU_{n} \\
U_{n} & sU_{n-1}+tU_{n-2} & tU_{n-1}
\end{array}%
\right],
\end{equation}
where $U_{-1}=\frac{1}{t}$ and $U_{-2}=-\frac{s}{t^{2}}$.
Now, let us define the following matrix as
\begin{equation}\label{eq:16}
Q_{S}=\left[
\begin{array}{ccc}
Q_{v,4} & sQ_{v,3}+tQ_{v,2} & tQ_{v,3} \\
Q_{v,3} & sQ_{v,2}+tQ_{v,1} & tQ_{v,2} \\
Q_{v,2} & sQ_{v,1}+tQ_{v,0} & tQ_{v,1}
\end{array}%
\right].
\end{equation}
This matrix can be called as the generalized Tribonacci quaternion matrix. Then, we can give the next theorem to the $Q_{S}$-matrix.
\begin{theorem}
If $Q_{v,n}$ be the $n$-th generalized Tribonacci quaternion. Then, for $n\geq0$:
\begin{equation}\label{eq:17}
Q_{S}\cdot\left[
\begin{array}{ccc}
r & s & t \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}%
\right]^{n}=\left[
\begin{array}{ccc}
Q_{v,n+4} & sQ_{v,n+3}+tQ_{v,n+2} & tQ_{v,n+3} \\
Q_{v,n+3} & sQ_{v,n+2}+tQ_{v,n+1} & tQ_{v,n+2} \\
Q_{v,n+2} & sQ_{v,n+1}+tQ_{v,n} & tQ_{v,n+1}%
\end{array}%
\right].
\end{equation}
\end{theorem}
\begin{proof}
(By induction on $n$) If $n=0$, then the result is obvious. Now, we suppose it is true for $n=m$, that is
$$Q_{S}\cdot S^{m}=\left[
\begin{array}{ccc}
Q_{v,m+4} & sQ_{v,m+3}+tQ_{v,m+2} & tQ_{v,m+3} \\
Q_{v,m+3} & sQ_{v,m+2}+tQ_{v,m+1} & tQ_{v,m+2} \\
Q_{v,m+2} & sQ_{v,m+1}+tQ_{v,m} & tQ_{v,m+1}%
\end{array}%
\right].$$
Using the Eq. (\ref{eq:7}), for $m\geq 0$, $Q_{v,m+3}=rQ_{v,m+2}+sQ_{v,m+1}+tQ_{v,m}$. Then, by induction hypothesis
\begin{align*}
Q_{S}\cdot S^{m+1}&=\left(Q_{S}\cdot S^{m}\right)\cdot S\\
&=\left[
\begin{array}{ccc}
Q_{v,m+4} & sQ_{v,m+3}+tQ_{v,m+2} & tQ_{v,m+3} \\
Q_{v,m+3} & sQ_{v,m+2}+tQ_{v,m+1} & tQ_{v,m+2} \\
Q_{v,m+2} & sQ_{v,m+1}+tQ_{v,m} & tQ_{v,m+1}%
\end{array}%
\right]\left[
\begin{array}{ccc}
r & s & t \\
1 & 0 & 0 \\
0 & 1 & 0%
\end{array}%
\right]\\
&=\left[
\begin{array}{ccc}
Q_{v,m+5} & sQ_{v,m+4}+tQ_{v,m+3} & tQ_{v,m+4} \\
Q_{v,m+4} & sQ_{v,m+3}+tQ_{v,m+2} & tQ_{v,m+3} \\
Q_{v,m+3} & sQ_{v,m+2}+tQ_{v,m+1} & tQ_{v,m+2}%
\end{array}%
\right].
\end{align*}
Hence, the Eq. (\ref{eq:17}) holds for all $n\geq0$.
\end{proof}
\begin{corollary}
For $n\geq0$,
\begin{equation}
Q_{w,n+2}=Q_{w,2}U_{n+2}+(sQ_{v,1}+tQ_{v,0})U_{n}+tQ_{v,1}U_{n}.
\end{equation}
\end{corollary}
\begin{proof}
The proof can be easily seen by the coefficient (3,1) of the matrix $Q_{S}\cdot S^{n}$ and the Eq. (\ref{eq:15}).
\end{proof}
\section{Conclusion}
This study examines and studied Tribonacci-type quaternion sequences with the help of a simple and general formula. For this purpose, generalized Tribonacci sequence $\{V_{n}\}_{n\geq0}$ was used. Generalized Tribonacci sequence was examined in detail particularly in section 1, and it was shown that this sequence is used to generalize all the third order linear recurrence relations on quaternions. In this study, Binet formulas, generating functions, some summation formulas and norm values of all quaternion sequences were obtained. As a result, all the formulas in the literature were given with the help of only one formula. Quaternions have great importance as they are used in quantum physics, applied mathematics, graph theory and differential equations. Thus, in our future studies we plan to examine generalized Tribonacci octonions and their key features.
|
\section{Introduction}
The term `coronal bright point' (CBP) describes a phenomenon in the solar
atmosphere that appears in extreme-ultraviolet (EUV) and X-ray images as a small-scale multi-loop system of enhanced coronal emission that is associated with magnetic bipolar features \citep[e.g.,][]{1977SoPh...53..111G, 1993SoPh..144...15W, 2001SoPh..201..305B, 2016ApJ...818....9M}. CBPs were detected for the first time in soft X-ray photographs taken during rocket missions in 1968--1973 \citep{1973SoPh...32...81V} and later were analysed in great detail during the \textit{Skylab} mission \citep[][and the references therein]{1976SoPh...49...79G, 1976SoPh...50..311G, 1992AnGeo..10...34H}. \citet{1990ApJ...352..333H} found that CBPs show no difference of their properties in coronal holes (CH) and the quiet Sun (QS), which led the authors to conclude that they do not depend on the structure of the surrounding background corona.
CBPs have sizes in the range 10\arcsec--50\arcsec. Their average lifetime derived from observations taken with the Fe~{\sc xii}~195~\AA\ filter of the Extreme-ultraviolet Imaging Telescope (EIT) on board the Solar and Heliospheric Observatory (SoHO) is 20 hours \citep{2001SoPh..198..347Z}, while recently \citet{2016ApJ...818....9M} reported a lifetime ranging from 2.7 to 58.8 hours from a study of 70 CBPs observed with Fe~{\sc xii}~193~\AA\ filter of the Atmospheric Imaging Assembly (AIA) on board the Solar Dynamic Observatory (SDO). In X-rays CBPs have lifetimes of only 8 hrs \citep{1974ApJ...189L..93G}. \citet{2001SoPh..198..347Z} concluded that the temperatures of BPs are generally below 2$\times$10$^6$ K, which also explains their smaller size and shorter lifetime in X-rays compared to lower temperature observations. \citet{1990ApJ...352..333H} found that simultaneously measured peaks of emission in six different lines (with a large range of formation temperatures from chromospheric to coronal) were not always co-spatial, implying that the BPs may consist of a complex of small-scale loops at different temperatures and heights. \citet{2012ApJ...757..167K} investigated the multi-thermal nature of
EUV BPs using the coronal 171, 193 and 284~\AA\, and the 304~\AA\ (chromosphere -- transition region) filter images from observations with the Extreme-UltraViolet Imager (EUVI) on board the twin STEREO satellites. The correlation coefficient between the different channels made the authors to conclude that
BPs at 171, 195, and 284~\AA\ belong to the same loop system while the BP emission in 304~\AA\ can be interpreted as coming from cool legs of the loops.
The coronal emission evolution of CBPs strongly correlates with the variation of the total unsigned photospheric magnetic flux of the associated magnetic bipolar features (MBFs) \citep[e.g.,][]{1999ApJ...510L..73P, 2003A&A...398..775M, 2004A&A...418..313U}. \citet{2016ApJ...818....9M} concluded that the formation of the MBFs associated with CBPs involves the processes of flux emergence, convergence, and local coalescence of the magnetic fluxes, while the formation of MBFs may involve more than one of these processes. Observational magnetic cancellation, that is defined as an empirical description of the disappearance and/or decrease of magnetic features of opposite polarity while they visibly interact with each other as observed in magnetograms, was found in all 70 analysed BP cases. The MBFs were found to evolve in three typical manners, namely, ``between a MBF and small weak magnetic features, within a MBF with the two polarities moving toward each other from a large distance, and within a MBF whose two main polarities emerge in the same place simultaneously''.
\citet{1987ARA&A..25...83Z} suggested two possible scenarios that can explain the observed magnetic cancellation. One is a simple submergence, when a pre-existing loop descends into the convection zone. Another scenario requires magnetic reconnection either above or below the photosphere, which is named reconnection submergence. Reconnection submergence is discussed in detail by \citet{1989ApJ...343..971V}. From simultaneous measurements of the magnetic field in the photosphere and chromosphere \citet{1999SoPh..190...35H} concluded that the magnetic flux ``is retracting below the surface for most, if not all, of the cancellation sites studied''. Furthermore, \citet{1994ApJ...427..459P} developed a model of BPs based on converging motions of magnetic features which can trigger magnetic reconnection and thereby energize a BP. The model was further developed in three-dimensions by \citet{1994SoPh..153..217P} and tested by two-dimensional numerical experiments by \citet{2006MNRAS.366..125V, 2006MNRAS.369...43V}.
Reconstructing the three dimensional (3D) magnetic topology of CBPs is essential in modelling these or any other solar phenomena. \citet{1994SoPh..151...57P} were the first to derive the magnetic topology of two observed CBP. They used a low number of point sources to capture the 3D structure of the magnetic field. This showed that the BP structure aligned with the separator line connecting two 3D null points located in the photosphere. \citet{2008A&A...492..575P} used the MPOLE topology code of \citet{1996SoPh..169...91L}, that is based on reducing a full magnetogram to a distribution of monopoles in the photospheric surface, to calculate the 3D potential magnetic field of one BP and analyse the skeleton of the magnetic field. The authors found a very close agreement between the calculated magnetic field configuration and some of the X-ray loops composing the studied bright point. This made them suggest that ``a large fraction of the magnetic field in the bright point is close to potential''. \citet{2011A&A...526A.134A} applied the same method to another BP with the same main conclusion. \citet{2012ApJ...746...19Z} extrapolated a potential field model to derive the magnetic topology of 13 CBPs using Global Oscillation Network Group (GONG) Big Bear and Learmonth magnetograms. Only two of their BPs were found to be associated with the so-called ``embedded bipolar region'' where one polarity is surrounded by another polarity. This resulted, for both cases, in a coronal magnetic null point defining a dome-shaped separatrix surface. The remaining 11 BPs were found to be associated with simple bipolar regions creating loop systems, as found in \citet{2008A&A...492..575P} and \cite{2011A&A...526A.134A}.
The aim of the present paper is to investigate the magnetic skeleton of CBPs
in both CH and QS regions using a long time-series of high resolution longitudinal magnetograms from the Solar Optical Telescope (SOT) on board the Hinode satellite combined with X-ray images taken with the X-ray Telescope (XRT) on board the same satellite. Knowing the magnetic skeleton of the BP regions is important for determining the possible mechanisms that are responsible for heating the plasma to coronal temperatures. On a longer time scale this allows us to perform time dependent 3D modellings to further enhance our understanding of the dynamical evolution of basic BP magnetic field structures.
In \Sec{obs.sec} we describe the observations. The results are reported in
\Sec{res.sec}. Discussion and conclusions are given in \Sec{disc.sec} and \Sec{conc.sec}.
\section{Observations, modelling and analysis methodologies}
\label{obs.sec}
\subsection{Observational material}
The observations analysed in the present study were taken by XRT \citep{2007SoPh..243...63G} and SOT \citep{2008SoPh..249..167T} on board the Hinode satellite. Five datasets obtained in QS and equatorial CH regions were selected and 10 bright points were identified for further analysis from which 7 were located in the CH regions and 3 in QS regions. Details on the XRT and SOT data can be found in \Tab{tab:bp_obs}. The CH and one of the QS (on 2007 November 27) XRT observations were taken with the Al$\_$Poly filter. The QS data on 2007 October 10 were obtained with both the C$\_$poly and Al$\_$poly filters, but only the C$\_$poly filter data were used for further analysis because of their better cadence rate. The temperature response of both filters peaks around 8$\times$10$^6$~K with the Al$\_$poly filter having a stronger response at lower temperatures. The pixel size of the X-ray images is 1\arcsec $\times$ 1\arcsec. The SOT observations represent series of Stokes V and I images taken with the Narrowband Filter Imager (NFI) in the Na~{\sc i} D 5896~\AA\ spectral line which provide measurements of the magnetic field in the chromosphere \citep{2008SoPh..249..233I}. The magnetograms have a pixel size of 0.16\arcsec $\times$ 0.16\arcsec. The data calibration and alignment are performed as described in \citet{2012A&A...548A..62H}. The QS SOT data on 2011 June 18 were taken after the launch of the Solar Dynamic Observatory (SDO) and, therefore, full-disk solar observations were available in the extreme-ultraviolet (EUV) obtained with the Fe~{\sc xii}~193~\AA\ filter of the Atmospheric Imaging Assembly \citep[AIA,][]{2012SoPh..275...17L} at 12~s cadence and 0.6\arcsec\ $\times$ 0.6\arcsec\ per pixel spatial resolution. To complete the SOT dataset that only had 9 out of 19 good quality magnetograms, we also used data from the Helioseismic and Magnetic Imager (HMI) on board SDO that have 45~s cadence and 0.5\arcsec $\times$ 0.5\arcsec\ pixel size resolution. From the observed regions given in \Tab{tab:bp_obs}, we have chosen the bright points that are isolated showing a simple structure and are located sufficiently far from the domain boundaries to allow for a meaningful magnetic field extrapolation. The typical error in the alignment between the magnetogram and XRT data is on the order of 2\arcsec\ and this error dominates over the projection effect due to the height difference in the two types of observations. This effect is most pronounced for the first dataset (see Table~\ref{tab:bp_obs}). For the 3D visualisations shown below, the aligned low resolution X-ray data have been interpolated to the associated magnetogram resolution using the idl procedure {\it interpolate} including the keyword {\it cubic=-0.5}.
\begin{table*}[!ht]
\centering
\caption{\label{tab:bp_obs} \textbf{Data used in the present study.}}
\begin{tabular}{l c c c c c c c }
\hline
\hline
Date &Time & {Cadence (s)} & \multicolumn{2}{c} {X-cen, Y-cen} & SOT FOV & XRT FOV & BP No. \\
& (UT) & XRT/SOT & XRT & SOT & &\\
\hline
2007-Nov-09 & 06:38--14:59 & 20/30 & -540\arcsec,-222\arcsec& -499\arcsec, -200\arcsec&261\arcsec\ $\times$ 148\arcsec & 366\arcsec $\times$ 366\arcsec & 1, 2, 3\\
2007-Nov-12 & 01:21--10:58 & 40/90 & -47\arcsec,-228\arcsec&-7\arcsec,-206\arcsec &276\arcsec\ $\times$ 163\arcsec & 394\arcsec $\times$ 394\arcsec & 4, 5, 6, 7\\
2007-Oct-10 & 18:47--23:58 & 30/60 & -214\arcsec,-70\arcsec & -174\arcsec, -39\arcsec & 276\arcsec\ $\times$ 164\arcsec&394\arcsec $\times$ 263\arcsec& 8 \\
2007-Nov-27 & 14:00--17:57 & 30/60 & 263\arcsec,-42\arcsec & 112\arcsec,-59\arcsec& 225\arcsec\ $\times$ 113\arcsec & 394\arcsec $\times$ 263\arcsec& 9 \\
2011-June-18 & 14:16--15:13 & --/180 &--& 7\arcsec, 46\arcsec & 225\arcsec\ $\times$ 113\arcsec & -- &10 \\
\hline
\end{tabular}
\end{table*}
\begin{table}[!ht]
\centering
\caption{\label{bp_relation.tab} \textbf{Bright point references relative \citet{2012A&A...548A..62H}}}
\begin{tabular}{c c}
\hline
\hline
Label in Huang et al. 2012 & Label in this paper \\
\hline
Huang 1, 9 Nov.& BP1 \\
Huang 4, 9 Nov.& BP2 \\
Huang 3, 9 Nov.& BP3 \\
Huang 2, 12 Nov.& BP4 \\
Huang 1, 12 Nov.& BP5 \\
Huang 8, 12 Nov.& BP6 \\
Huang 10, 12 Nov.& BP7 \\
\hline
\end{tabular}
\end{table}
\subsection{Potential and Magneto-static modelling}
\label{extrapolation.sec}
The field of view of the SOT data given in \Tab{tab:bp_obs} was used, at full observational data resolution, as the photospheric boundary condition for a potential field extrapolation. This approach uses a Fast-Fourier Transform (FFT) algorithm to solve for the scalar potential
and a 6th-order centred finite difference operator to derive the 3D magnetic field. The FFT approach requires the magnetogram data to be 2D periodic. This is achieved by smoothing the magnetogram data to zero along the domain edges using a hyperbolic tan function centred 10 pixels inside the domain, and with a halfwidth of 3 pixels. This alteration to the observational data is not critical for the extrapolated magnetic field as most of the investigated BP regions are located with a sufficient distance to the boundaries not to be significantly influenced by this change. The extrapolations are solved to a height equal to one quarter of the shortest magnetograms' sizes (except for BP9 that is much higher, see the discussion of BP9 below).
\begin{figure}[!ht]
{
\includegraphics[scale=.5]{BP1_skirt_MHS_tt.eps}
\hfill
\includegraphics[scale=.5]{BP1_skirt_potential_tt.eps}
\hfill
\includegraphics[scale=.5]{BP1_pot_skirt_tt.eps}
}
\caption[]{Two extrapolations of the magnetic field using a skirt of 10 pixels around the original magnetogram to enforce flux balance and one extrapolation without the skirt. Left -- the MS model extrapolation, middle -- the potential extrapolation and right -- potential extrapolation of the original data with the boundary smoothing to remove the signal to enforce a periodic setup. The field lines represent the magnetic field line structure around the null point associated with the BP1 case (see the discussion below).
\label{extrapolation_compare.fig}}
\end{figure}
A magneto-static (MS) field extrapolation \citep{2015ApJ...815...10W} was also used to test the reliability of the potential solution. The MS approach is based on a special class of linear MS equilibria as proposed by \citet{low91}. The electric current density is described as
\begin{equation}
\nabla \times {\bf B } = \alpha_0 {\bf B } + f(z) \nabla B_z \times {\bf e_z},
\label{lin_mhs}
\end{equation}
where $\alpha_0$ is the linear force-free parameter controlling field aligned currents and $f(z)$ is a function which controls horizontal currents. {In the following $f(z)=a exp(-k z]$ is used and the two constants have the following values; $\alpha_0=1$ and $a=1$.} The equilibria are solved with the help of a FFT method.
The MS modelling requires the input data to be in magnetic flux balance. The first dataset of the CH regions obtained on 2007 November 9 was used for the comparative test (see \Tab{tab:bp_obs}). The region is dominated by open magnetic field, and therefore, has a clear flux imbalance. To fulfil the flux balance requirement of the MS approach, a band of 10 pixels of uniform flux is placed around the observational data. These adjustments of the new magnetogram make clear alterations to the derived field line connectivity inside the domain, while still allowing the locally imbalanced flux to reach into the atmosphere before connecting down to the fictive flux region ``around'' the original data.
Both a MS and a potential extrapolation were derived for the modified magnetogram. The field line structure of the BP1 region found in the MS (left frame) and the potential (middle frame) extrapolations are shown in \Fig{extrapolation_compare.fig}. This demonstrates that the two extrapolation methods provide magnetic field structures that, for this region, are comparable. {This indicates, as also stated by \cite{}, that the magnetic field is close to being potential for the BP regions. For a closer comparison of the two solutions the characteristics of the central null point in the figure has been investigated.} The position and eigenvalues of the null point for the different extrapolation approaches are given in \Tab{null_compare.tab}. This shows that the two extrapolation methods produce comparable magnetic field structure. For a further comparison, a field line trace (right frame) in \Fig{extrapolation_compare.fig}, and the null point information is also provided for a potential extrapolation of the original data with the smoothing described above to provide a periodic dataset.
A significant deviation is only found for the height of the null point that is clearly higher in the original data. The differences are caused by the deviations from flux balance, which for the original data allow the magnetic field to expand higher into the atmosphere. In the following investigation, only the potential field extrapolation on the original datasets is used.
\begin{table*}[!ht]
\centering
\caption{\label{null_compare.tab} Parameters of the BP1 associated null point using two different extrapolations methods with imposed and without flux balance. The horizontal positions are given in gridpoints relative to the lower left corner of the SOT data set, where the pixel size is equivalent to the SOT magnetogram resolution of 0.16\arcsec. The height is measured in the same units.}
\begin{tabular}{c c c}
\hline
\hline
Extrapolation method and field & Position (grid points) & Eigenvalues \\
\hline
Mag-static-skirt & (1004.7, 839.3, 16.4) & (-0.39, 0.18, 0.21) \\
Potential-skirt & (1004.7, 839.4 , 15.8) & (-0.43, 0.23, 0.20) \\
Potential & (1002.2, 812.9, 36.5) & (-0.33, 0.15, 0.18) \\
\hline
\end{tabular}
\end{table*}
\subsection{Magnetic topology analysis methods}\label{top.sec}
The extrapolated 3D potential magnetic fields were investigated for magnetic null points using the method of \citet{2007PhPl...14h2107H}. This approach employs a trilinear method to identify null points in a 3D gritted magnetic field dataset. The method works successfully when the typical length scales of the magnetic field are clearly larger than the grid resolution. As the high frequency information in the FFT extrapolation decays exponentially with wave number and height, this condition is typically fulfilled above a given height in the atmosphere (5-10 horizontal grid points). This may further be tested by calculating the $\nabla \cdot B$ through summing up the eigenvalues of the null point. If this requirement is satisfied numerically, then the linear criterion is typically sufficiently fulfilled.
To identify the magnetic field structures at the null locations, their characteristics are derived solving for the eigenvalues of the associated Jacobian matrix of the magnetic field \citep{1997GApFD..84..245P}. The nullfinder approach provides a very high number of identified 3D null points inside the domain. Most of these are located very low in the atmosphere, as also found in \citet{2009SoPh..254...51L}, and a fair number of these are associated with the imposed boundary smoothing of the magnetic field. Only null points above a given minimum height (10 gridpoints) have been considered. To qualify as the null associated with a given BP region, they have to be located sufficiently close to the BP position. Typically, only one null point fits this description and its time evolution is easy to follow and is investigated below.
The 3D magnetic field is visualised using the VAPOR software to further investigate the magnetic field topology around the locations of the identified BPs. This analysis reveals that 9 of the 10 BPs in the sample contain a dominating magnetic null point located in the coronal domain. The null points have only real eigenvalues \citep{1997GApFD..84..245P}. The general structure of all BP null points is such that the spine axis (associated with the dominating eigenvalue and its eigenvector), in one direction, connects to the minority polarity below the null point and, in the opposite direction, connects towards the top boundary of the domain. In comparison, the fan planes are asymmetric with the main axis (dominating fan eigenvalue and vector) pointing towards the nearby dominating flux concentration. The presence of the null above the minority polarity flux concentration implies that all flux from the minority polarity closes back to the photosphere, while the ambient excess flux connects to the upper boundary.
\section{Results: data analysis and modelling}
\label{res.sec}
The seven selected CH bright points (BP1--BP7) were previously discussed in detail in \citet[][hereafter HBPs]{2012A&A...548A..62H} (see also \url{http://star.arm.ac.uk/highlights/2012/603/bpmag}). \Tab{bp_relation.tab} relates the bright point numbering in \citet{2012A&A...548A..62H} with the numbering used in the rest of this paper. The temporal evolution of the BPs' X-ray emission and their associated magnetic polarities including magnetic flux decrease rate of BP1, BP3 and BP4 are given by \citet{2012A&A...548A..62H}. The BPs were chosen as they have well isolated magnetic field structures only related to the bright point phenomena. Also, all examples are sufficiently long lived to allow for a qualitative investigation of the magnetic field evolution during their lifetime.
\begin{table*}[!ht]
\centering
\caption{\label{tab:bp_an} General bright point properties.}
\begin{tabular}{l l l l}
\hline
\hline
BP No. & MBF evolution & Null point & Projected X-ray/EUV location \\
\hline
1 & Convergence \& decrease & Decreasing height & Above one of the main polarity\\
& & & flux concentrations\\
2 & Convergence \& divergence. Diffuse & Moves up and down & Spread around mainly inside \\
& flux emergence & & the dome, concentrated at the \\
& & & strongest minority flux region \\
3 & Convergence & Low lying null point & Inside the fan-dome, \\
& & & appears during a jet\\
4 & Flux emergence and convergence of & Rise in the corona & Inside the fan-dome, moves \\
& emerged positive flux with pre-existing & & around\\
& negative & & \\
5 & Convergence of opposite flux & Decreases in height & Inside the fan-dome\\
6 & Flux convergence & Decreases in height & Located above a dominant flux \\
& & & concentration that includes a \\
& & & fraction of the fan-dome\\
7 & Granular size motions & Moves up and down & Inside the fan-dome \\
8 & Convergence of negative flux with & Increases in height and & In the area between the \\
& one of the strong positive fluxes & expands in volume & approaching flux patches \\
9 & Convective motion and possible & Located far above the & A low lying bipolar loop system \\
& small scale emergence & BP region & located inside the dome\\
10 & Increased flux concentrations with time & No null identified & A bipolar loop system \\
& and a slow convergence of opposite polarities & & \\
\hline
\end{tabular}
\end{table*}
All BPs reveal a dynamical evolution observed in X-rays/EUV (see for example Figures~6, 9 and 13 in HBPs) on time scales as short as the data cadence (see Table~\ref{tab:bp_obs}). The analysed BPs were observed during different stages of their lifetime but the long duration of the observations provides a sufficient coverage for the requirements of the present study. It is important to mention that the choice of data was entirely dictated by the quality of the magnetic field data. Presently, SOT provides the highest quality of uninterrupted long time series of magnetic field data that are crucial for modelling of small scale solar phenomena. Therefore, these data fulfil the requirements of our investigation on the magnetic field structure of BPs and the changes of their footpoint properties related to their dynamical coronal evolution.
This investigation focuses on the normal component magnetic data from NFI/SOT and HMI/SDO (one case) to provide the boundary condition for the potential field extrapolation. There are no available vector magnetogram that could have been used for these investigations.
This section is focused on the general changes in the photospheric magnetic field of the BPs, the specific characteristics of the magnetic field topology and the associated X-ray/EUV emission that identifies a feature as a coronal bright point. Here it has to be stressed that coronal bright points are only seen in X-rays
for a short period of time during their entire lifetime.
This implies that they are active to this temperature level only when their magnetic structure is sufficiently stressed. BPs have a longer lifetime when observed in the lower temperature passband imaging data \citep{2001SoPh..198..347Z}, i.e. EUV. It may, therefore, be possible that some of the general dynamical changes seen in the magnetic observations do provide energy releases, but that these are not sufficient to reach high coronal temperatures. This makes it difficult to see which specific changes in the photospheric footpoint positions of the magnetic field could result in important large energy releases. For this we need much more detailed dynamical investigations using at least the same quality of magnetic field data (SOT) together with high resolution and temperature coverage imaging data, e.g. AIA/SDO and IRIS, to obtain a better understanding on this relationship. An ongoing dedicated observing campaign should provide suitable data for further analysis.
\begin{figure*}[!ht]
{
\includegraphics[scale=.8]{BP1_mag_0000_tt.eps}
\hfill
\includegraphics[scale=.8]{BP1_mag_0008_tt.eps}
\hfill
\includegraphics[scale=.8]{BP1_mag_0016_tt.eps}
\hfill
\includegraphics[scale=.8]{BP1_mag_0021_tt.eps}
}
\vspace{.1cm}
{
\includegraphics[scale=.8]{BP1_xrt_0000_tt.eps}
\hfill
\includegraphics[scale=.8]{BP1_xrt_0008_tt.eps}
\hfill
\includegraphics[scale=.8]{BP1_xrt_0016_tt.eps}
\hfill
\includegraphics[scale=.8]{BP1_xrt_0021_tt.eps}
}
\caption{Time evolution of the BP1 case. Top row shows the evolution of the magnetogram and the corresponding potential magnetic field extrapolation. The magnetogram is scaled to $\pm$200~G, where blue represent negative and red positive polarity. The grey isosurface indicates the location of the magnetic null point, with the field lines being traced from the vicinity of the null region. The field line colour represents the magnetic field strength, with red being strong, and green-blue represent the weak field. The lower row, shows the same magnetic fields, while the bottom boundary represents the XRT observations shown on a linear scale, with the red being low values and purple representing the peak values. The images correspond to the times 06:38:13~UT, 08:38:05~UT, 10:38:08~UT and 11:59:35~UT (left to right) on 2007 November 9. See the included animation for the time evolution.
\label{BP1.fig}}
\end{figure*}
{\it BP1} (\Fig{BP1.fig}) is present from the start of the observations (at 06:38~UT) and it disappears at $\sim$12:00~UT. It was found in an equatorial coronal hole observed on 2007 November 9 with dominant negative polarity flux (hereafter dominant) and parasite or minority positive flux (hereafter minority). The XRT images show a relatively round shaped BP with a diameter of $\sim$10\arcsec\ that is preserved for most of the observing period. Impulsive variations of the X-ray emission are observed for the whole duration of the BP (see Figure~8 in HBPs), two of which at around 09:44~UT and 10:10~UT exceed more than twice the BP average brightness and last for several minutes. The BPs projected X-ray emission is located above a strong flux concentration. Small dominating magnetic fragments are constantly moving towards the minority polarity and slowly erode it away while a remote dominating polarity patch strengthens and slowly approaches the minority flux. When the BP vanishes in X-rays, the magnetic structure changes now avoiding the clear connection to the strong dominant flux concentration that previously ``hosted'' the X-ray emission (see the attached animation for the time evolution).
\begin{figure*}[!ht]
\center
\includegraphics[scale=.45]{null_info_9_1_1000_800.eps}\\
\includegraphics[scale=.45]{null_info_9_4_1150_600.eps}\\
\includegraphics[scale=.45]{null_info_9_3_770_613.eps}
\caption{{\bf Left column:} Change in the horizontal null point position relative to a fictive null position for BP1 (top row), BP2 (middle) and BP3 (bottom). The grey shaded stars mark the evolution in time going from black to grey. The first point is marked with a large triangle and the last point with a large diamond. {\bf Middle column:} Height change of the null point. {\bf Right column:} Time dependent change of the null point's eigenvalues.}
\label{BP_1_Nulls.fig}
\end{figure*}
The temporal evolution of the BP1 null point is given in \Fig{BP_1_Nulls.fig} (top row), where the left panel shows the change in the null point horizontal position as a function of time. It reveals a jittering of the horizontal position that is mostly imposed by the spatial changes in the photospheric flux distribution and only weakly by the instrument pointing. The middle panels present the null point height as a function of time. This shows that the null point has a near monotonic decrease in height as a function of time. This motion is coupled to the general convergence of the dominating flux patches over time. It leads to a constant decrease in the total volume under the fan dome, eventually reaching a point where most of the volume is too low in the atmosphere to allow further heating of the plasma inside the dome to X-ray temperatures. Finally, the right panel shows how the three eigenvalues change in time, where it is seen that their magnitudes increase only slightly with time. This may be expected as the null point decreases in height, while the field strength and magnetic gradients in the null region increase.
\begin{figure*}[!ht]
{
\includegraphics[scale=.8]{BP2_mag_0029_tt.eps}
\hfill
\includegraphics[scale=.8]{BP2_xrt_0029_tt.eps}
\hfill
\includegraphics[scale=.8]{BP3_mag_0015_tt.eps}
\hfill
\includegraphics[scale=.8]{BP3_xrt_0015_tt.eps}
}
\vspace{.1cm}
{
\includegraphics[scale=.8]{BP4_local_mag_0020_tt.eps}
\hfill
\includegraphics[scale=.8]{BP4_local_xrt_0021_tt.eps}
\hfill
\includegraphics[scale=.75]{BP5_mag_0018_tt.eps}
\hfill
\includegraphics[scale=.75]{BP5_xrt_0018_tt.eps}
}
\vspace{.1cm}
{
\includegraphics[scale=.75]{BP6_mag_0016_tt.eps}
\hfill
\includegraphics[scale=.75]{BP6_xrt_0016_tt.eps}
\hfill
\includegraphics[scale=.8]{BP7_local_3_mag_0028_tt.eps}
\hfill
\includegraphics[scale=.8]{BP7_local_3_xrt_0028_tt.eps}
}
\caption{Snapshots of CH bright points BP2--BP7 in pairs of two showing first the surface magnetic field followed by the XRT observations. In both cases the field lines represent the magnetic structure around the null region as in \Fig{BP1.fig}. The times for the frames are BP2 -- 14:14:36~UT, BP3 -- 10:23:07~UT, BP4 -- 06:51:03~UT, BP5 -- 06:06:03~UT, BP6 -- 05:21:35~UT and BP7 -- 08:06:05~UT on their respective dates, \Tab{tab:bp_obs}. In BP4 and BP7 additional single coloured field lines are included to show the local field line structure near the the projected X-ray BP regions. See the included corresponding animations for their time evolution.
\label{BPrest_CH.fig}}
\end{figure*}
{\it BP2} (\Fig{BPrest_CH.fig}) becomes visible in the same CH as BP1 around 07:38~UT and it lasts for the remainder of the observations. Magnetically, this case is very different from BP1: only the minority flux is concentrated, while for most of the time the dominating polarity consists of a weak background field. The minority flux concentration increases with time following the emergence and convergence of small scale (pixel size) magnetic flux elements. During this time the dominant flux has only minor flux concentrations scattered around the minority flux.
The X-ray intensity is lower than BP1's and it has a much more diffuse structure, with the emission ``moving'' around over a larger elongated area (around 20\arcsec $\times$ 30\arcsec), that is mainly distributed over one half of the projected dome region. It is unclear why the X-ray emission is located in this part of the dome from just looking at the time evolution of the underlying magnetogram. Around 14:14~UT the XRT data show the highest intensity in the BP. This is associated with a period where a strong dominant polarity becomes part of the fan dome. The middle row of \Fig{BP_1_Nulls.fig} shows a near systematic advection of the null point's horizontal position in time, while at the same time the height of the null point rises in the atmosphere. This is consistent with the continued accumulation of minority flux.
{\it BP3} is also formed in the CH observed on 2007 November 9 (\Fig{BPrest_CH.fig}), and it appears in X-rays around 11:10~UT. At $\sim$11:57~UT, the BP produces an X-ray jet that lasts for around 10~min (see HBPs for more details). The BP vanishes at around 12:29~UT. The magnetic structure is dominated by a concentrated minority polarity flux pattern and a ``wall'' like distribution of dominating polarity flux on ``one'' side of the minority flux. These flux concentrations advect towards each other, which makes the magnetic structure that host the BP shrink in size. The BP vanishes as the minority polarity flux disappears, leading to a unipolar open field region. In the bottom row panels of \Fig{BP_1_Nulls.fig} it can be seen how the null point is moving around in the horizontal direction and its height decreases as a function of time. At the beginning of the observations the X-ray emission appears more spread out under the entire dome area, which may be explained by the divergence of the flux concentrations. As some of the flux patches form stronger concentrations, the X-ray emission concentrates above the approaching opposite polarities, i.e. in only a minor part of the dome. Before the BP appearance in X-rays, the potential field already contains a null point that is located much higher than when the BP initiates. As the BP vanishes the dome has diminished in size and shortly after it disappears.
{\it BP4} has been observed in the same CH but 3 days later on November 12 (\Fig{BPrest_CH.fig}). It becomes visible in the X-rays around 04:40~UT. Its shape changes quite dynamically but its size remains around 20\arcsec $\times$ 20\arcsec. Before the BP appears, the magnetic structure of the region is dominated by the dominant polarity, with an almost open field structure. In the period just before the BP appearance at $\sim$04:00~UT, a null point forms and starts rising in the corona (see \Fig{BP_2_Nulls.fig} (top row)). This is driven by the emergence of magnetic flux. As the amount of minority flux increases the height of the null increases. The X-ray emission has a complex behaviour spreading over a large area which is clearly related to the complex magnetic field structure and its evolution. During the later stronger X-ray phase, the BP is located above a weak field region of opposite polarity to the stronger dominant flux concentrations. The dominant flux concentrations are enhanced during the emergence, relative to the spine footpoint. At the time of the most active X-ray phase, the amount of minority flux increases and moves in the direction of the strong X-ray emission, before the X-ray signature is finally located around the spine footpoint.
\begin{figure*}[!ht]
\center
\includegraphics[scale=.45]{null_info_12_2_730_670.eps}\\
\includegraphics[scale=.45]{null_info_12_1_310_750.eps}\\
\includegraphics[scale=.45]{null_info_12_8_275_220.eps}
\caption[]{The same as \Fig{BP_1_Nulls.fig} for BP4 (top row), BP5 (middle), and BP6 (bottom).}
\label{BP_2_Nulls.fig}
\end{figure*}
{\it BP5} is also found in the CH on November 12 (\Fig{BPrest_CH.fig}). It is present almost from the start of the observations (after 02:00~UT) and lasts until around 07:10~UT which coincides with the disappearance of the null point (\Fig{BP_2_Nulls.fig}, middle row). The BP X-ray signal shows quite some modulation with time. It has a round shape with a diameter smaller than 10\arcsec. From the start the magnetic null point is located high in the atmosphere. The height decreases with time as the minority polarity slowly approaches the dominant flux concentration and a new dominant patch is found to appear on its ``back side'' (with respect to the observer viewing direction in \Fig{BPrest_CH.fig}). Towards the end of the X-ray emission, the magnetic field complexity increases significantly, which is seen in the form of several magnetic null-point regions appearing in close proximity of the original null point at very low heights.
{\it BP6} is again a November 12 CH BP (\Fig{BPrest_CH.fig}). The BP is present throughout the whole observing period. It has almost a round shape (around 20\arcsec\ diameter) that changes to elliptic as two of the minority and one of the dominant polarities converge in time. Its intensity shows variations in time with a few major increases. The strong X-ray signal is found above the dominant flux region that occupies a small partition of the fan dome. The size of the region embedded by the fan dome is large, covering several granular cells. This is clearly seen in the time animation of the normal component of the magnetic field, that shows a jittering motion at different inter-granular lanes. Towards the end of the observations the BP decreases in size as the minority flux converges towards the largest dominating flux patch that hosted the X-ray signal. This motion leads to a decrease of the null points height, from $\sim$8~Mm to finally reaching a height of $\sim$3~Mm. This change in the magnetic character is nicely reflected in the properties of the null point, \Fig{BP_2_Nulls.fig} (bottom row). This shows both a systematic horizontal motion and the significant change
{\it BP7} is the last case belonging to the CH observed on November 12 (\Fig{BPrest_CH.fig}). It has a magnetic null point located high in the atmosphere for the whole time of the observations (\Fig{BP_3_Nulls.fig}, top row). The dominant flux concentrations seem to be rooted in different supergranular cell corners and are with time advected around. This results in a systematic advection of the null position with, first, a systematic increase in height, followed by an equivalent decrease in height. The X-ray emission is weak for most of the time with clear fluctuations in time and a significant emission increase around 07:50~UT that lasts for approximately 20 min. The BP shape changes from almost round (10\arcsec\ across) to slightly elongated with a size of less than 10\arcsec. The maximum X-ray emission is associated with a location of generally weak magnetic flux, which for a short period of time contains a lot of small scale mixed polarity flux. During this time there is a general advection of the magnetic spine region towards the X-ray region.
\begin{figure*}[!ht]
{
\includegraphics[scale=.8]{BP8_mag_comb_0007_tt.eps}
\hfill
\includegraphics[scale=.8]{BP8_xrt_comb_0007_tt.eps}
\hfill
\includegraphics[scale=.8]{BP9_mag_0005_tt.eps}
\hfill
\includegraphics[scale=.8]{BP9_xrt_0005_tt.eps}
\vspace{.1cm}
{
\includegraphics[scale=.8]{BP10_mag_0000_tt.eps}
\hspace{.1 cm}
\includegraphics[scale=.8]{BP10_AIA_0000_tt.eps}
}
\caption[]{Snapshots of QS BP8--BP10 as in \Fig{BPrest_CH.fig}. The times for the frames are BP8 -- 19:56:34~UT, BP9 -- 14:50:36~UT and BP10 -- 13:09:42~UT on their respective dates, \Tab{tab:bp_obs}. In BP8 and BP9 additional single coloured field lines are included to show the local field line structure near the projected X-ray regions Notice that for BP10 the base image in the right frame is the AIA 193. See the included animations for their time evolution.
\label{BPrest_QS.fig}
}}
\end{figure*}
\begin{figure*}[!t]
\center
\includegraphics[scale=.45]{null_info_12_10_600_300.eps}\\
\includegraphics[scale=.45]{null_info_10_1000_700.eps} \\
\includegraphics[scale=.45]{null_info_27_460_500.eps}
\caption[]{The same as \Fig{BP_1_Nulls.fig} for BP7 (top row), BP8 (middle), and BP9 (bottom).}
\label{BP_3_Nulls.fig}
\end{figure*}
{\it BP8} (\Fig{BPrest_QS.fig}) is found in a QS region on 2007 October 10 and it is present throughout the whole observational period. This BP was extensively investigated by \citet{2011A&A...526A.134A} including detailed important information of the BP region that need not be repeated here. This BP has a double structure: it can be said to be composed of two BPs that are contained inside the fan dome boundaries of a null point located relatively high in the corona. They appear as two sets of loops that connect the minority polarity, with two dominating polarities on either side of the spine axis. Both BPs have a size of approximately 20\arcsec $\times$ 20\arcsec\ that slightly vary in time. One of the BPs is brighter in X-rays and its X-ray intensity increases with time as the two corresponding polarities drift towards each other. Towards the end of the time series the null point changes character, both being more extended in horizontal direction -- implying a near 2D magnetic field structure in this region -- while the null point region increases significantly in height, like being ejected upwards. There do not seem to be any clear changes in the magnetogram that can explain this behaviour.
This evolution is only partly seen in \Fig{BP_3_Nulls.fig}, when the identification of the null point failed over this time period where the fast and significant changes took place.
\citet{2011A&A...526A.134A} use the MPOLE code to obtain information about the 3D magnetic field structure and find that the main structure in the X-ray emission is well represented by their potential model, that is shown by the blue field line region seen in \Fig{BPrest_QS.fig}. The two investigations are in good agreement on the low lying field line structure of this region.
\begin{figure}
{
\includegraphics[scale=.8]{BP9_mag_0005_tt.eps}
\hfill
\includegraphics[scale=.8]{BP9_xrt_0005_tt.eps}
}
\caption{Enlarged view of the BP9 case presented in \Fig{BPrest_QS.fig} showing the local field line structure around the concentrated X-ray emission. The left frame displays the surface magnetic field, while the right frame shows the associated X-ray emission at time 14:50:36~UT. See the included animation for the time evolution.
\label{BP9_local.fig}}
\end{figure}
{\it BP9} (\Fig{BPrest_CH.fig}) is found in a QS region on 2007 November 27 and is active for the full time of the observations. The BP has a single clear loop structure that shows minor variations over time. This BP is the largest of all studied cases with a size of $\approx$50\arcsec\ as measured from one end to the other of the loop structure. A null point is associated with this BP, but it is located very high above the BP region (\Fig{BP_3_Nulls.fig}). The null is located so high above the photosphere that the fan dome connects to the photosphere over a distance that is comparable to the smallest of the two dimensions of the observed SOT domain.
The connectivity of the null is, therefore, defined by the periodicity of the extrapolation. The presence of the null point is due to a weak flux imbalance of the observed data, which may change if one cuts the magnetogram differently. The location and structure of the null point, therefore, cannot be determined with full confidence for this BP.
In this case the BP appears to be embedded in a bipolar field region located clearly inside the fan dome structure for the entire observing period (blue field lines in \Fig{BPrest_QS.fig}). This loop system at one end connects the structurally complicated minority flux concentration that anchors the spine axis and a number of dominant flux regions next to it (see \Fig{BP9_local.fig} for a zoom of the local field line structure). This causes the connectivity to have a significant divergence across this split in the dominant fluxes and thereby creates local quasi-separatrix layer \citep{1995JGR...10023443P} structures that may provide locations for local energy release when appropriately stressed. An inspection of the magnetogram time series shows the region between the strong dominant flux concentrations to contain many smaller flux concentrations that constantly move around adding to the general stressing of the local magnetic field.
{\it BP10} (\Fig{BPrest_CH.fig}) is found in a QS region on 2011 July 18 and is active for the full time of the SOT observations which unfortunately last only 57~min. The BP is clearly associated with a bipolar loop system with a size of $~$13\arcsec\ estimated from one visible footpoint to the other. There are no identified null points associated with the BP. This BP is, therefore, clearly different from the other identified BPs in the sample.
\section{Discussion}
\label{disc.sec}
Knowledge on the magnetic topology of coronal bright points throughout its formation, evolution and decay, is benchmarking for any further understanding on the physical mechanism that sustains this main building segment of the high temperature low corona. We explored the magnetic topology of 10 coronal BPs and found that 9 of these contain a magnetic null point located in the coronal domain. Seven of these are CH BPs, and two are QS BPs. The general structure of all null points is such that the spine axis, in one direction, connects to the minority polarity below the null point and, in the opposite direction, connects towards the top boundary of the domain. For the CH BPs the null point is located relative low in the atmosphere, and the X-ray emission is located either inside the fan-dome or in a fraction of the dome, mostly hosted in a dominant polarity flux concentration. Due to the low height of the null point, this strongly suggests that the null point is an important feature in the magnetic field for hosting the BP structure. For the QS BPs we see that the two identified null points sit at much higher in the atmosphere with the X-ray emission being located clearly inside the fan-dome region and associated with bipolar magnetic field regions showing a clear loop structure. Finally, BP10 does not show any relation to a coronal null point, and seems to be residing in a bipolar loop system.
\citet{2010ApJ...714..130K} determined the heights of 210 BPs using a triangulation method on EUV images from the two STEREO spacecrafts in the 171~\AA, 195~\AA, 284~\AA\ and 304~\AA\ passbands. This provided them with heights for the BPs at different temperatures. For the 195~\AA\ filter they determine the BP mean height to be 6.7$\pm$2~Mm and determined lower heights for observations of BPs in passbands a with lower temperature response (171, 284 and 304~\AA). In comparison to this, \citet{2007AdSpR..39.1853T} used a linear field extrapolation of Kitt Peak magnetograms to investigate the alignment of the horizontal components of the extrapolated magnetic field at a given height with the emission contours of 10 BPs observed by EIT in the 195~\AA\ passband. They assumed a high correlation index to indicate the alignment of the emission with the loop summit and, therefore, can provide independent information about the height of the underlying BP loops. Their results show a scattering of BP heights within the 1--20 Mm range.
For all our CH cases, the null points are located relatively low in the atmosphere, 1--7 Mm, compared to the identified QS null points, 12--40 Mm. The CH nulls are therefore inside the typical hight range of the observations, while the QS nulls are on the high side in comparison to the two investigations mentioned above. With the differences in our null point heights for the two solar regions,
the CH nulls will have a much higher likelihood for being directly involved in the energy release process than for the QS BPs. This difference in height of the magnetic structures between the two regions is supported by the work of \citet{2004SoPh..225..227W}, where it is found that closed field lines are clearly lower in CH regions compared to QS regions. To investigate the importance of the null points in the different regions in relation to the BP regions, one has to conduct numerical 3D MHD modelling. To further check on the differences in magnetic structure between the two regions, more cases from QS regions must be studied in a comparable way. As mentioned above a dedicated observing campaign with SOT/EIS/Hinode and IRIS coupled with data from the AIA/SDO will be pursued.
In relation to this, \citet{2012ApJ...746...19Z} found that only two of the
13 bright points they selected in their region hosted a null point above the
BP, whereas the 11 remaining ones were located in general bipolar loop
systems. By inspection of full disk images from the day of their
observations, it is clear that these data correspond to a QS region and can
therefore be best compared with our BP8 to BP10 cases. Given the larger
number of QS BPs in their study we have tried to understand the possible
reasons for this discrepancy, in particular concerning the differences in the
underlying data and the treatment thereof. We do this in two parts:
\noindent a) a parameter that may influence the results is the spatial
resolution of the observational data that serve as a basis for the
extrapolation. \cite{2012ApJ...746...19Z} used longitudinal GONG magnetograms
starting at midnight on 2007 March 16.
From the header information of the corresponding data files (as
retrieved from the GONG data base) we derived a pixel size of
those observations of 2.55\arcsec, to be compared with the pixel size of
our SOT/NFI magnetograms (0.16\arcsec): SOT has $16$ times higher pixel resolution.
Would the results presented in earlier sections of the present paper change, with null
point structures possibly vanishing, if we took as a basis lower-resolution
magnetograms? To test for this, we have conducted an experiment where we
have degraded the SOT data to the GONG-magnetogram level by first convolving
with a Gaussian Point Spread Function (PSF) corresponding to the pixel size
ratio, followed by rebining to the pixel size of the GONG data. We have used
for that our 2007 Nov 9 initial data set, and carried out the degradation
process in four steps, i.e., degrading by pixel-size factors 2, 4, 8, and 16
to create intermediate-resolution cases. For each of the new resolutions
new potential extrapolations were done, and a visual comparison of the different
resolutions were carried out.
\begin{figure*}[!ht]
\hfill \,
\includegraphics[scale=2.7]{field_extrapolated_CH_BP1-2_QS_BP8_nep_collage_full_to_degraded_16_2_tt.eps}
\hfill \,
\caption[]{The frames show what happens when the pixel resolution of the
magnetogram is degraded in a number of steps using a PSF convolution
method. The left block of images show the extrapolations based on the
original SOT data, and with a factor of 2, 4, 8 and 16 times degradation in
spatial resolution for the BP1 and BP2 bright points. The right block of
images shows the original SOT data and a factor of 16 times lower resolution
for the BP8 bright point. The photospheric magnetograms are shown using a
$\pm$200 Gauss range; following the standard biasing technique of the VAPOR
package, the seeds for the field line tracing are planted at locations with
very low values of the total field strength in a box stretching from the
photosphere to several Mm height in the BP domain.
\label{resolution_compare.fig}
}
\end{figure*}
\Fig{resolution_compare.fig} shows the five different resolutions with their
associated extrapolated field line configurations for two of the coronal hole
bright points, namely BP1 and BP2 (left block of frames), together with the
extreme (i.e., SOT original and fully degraded) resolutions for one of the
QS BPs, namely BP8 (right block of frames). In both cases we see that
the null point structure is maintained in spite of the severe degradation of
the magnetograms, as apparent in the surface distributions in the
figure. This is reassuring in terms of the robustness of the extrapolation
technique we are using to locate null points above photospheric regions with
mixed (in particular, embedded) polarities, especially those linked to
coronal bright points in as far as the latter have sizes larger than the pixel
resolution of the magnetograms.
On the other hand, the puzzle of the discrepancy between the global
conclusion of \citet{2012ApJ...746...19Z} and ours must be explored further,
for instance as follows:
\noindent b) we have obtained GONG magnetogram data in the time range studied by
\citet{2012ApJ...746...19Z}, specifically, 2007 March 16, 00:01:16~UT, (as in
their Fig.~7) and carried out an extrapolation for a region of $900 \arcsec
\times 900 \arcsec$ that includes the domain with 13 bright points shown in
their Fig.~1. The magnetic structure in the neighbourhood of those BPs was
then re-investigated. Doing this we found that 8 BPs were clearly associated
with magnetic null points, 2 were structures of a different kind (akin to
bald patches), a further case was a simple bipolar loop configuration, and
there were two additional marginal cases where the identification with null
points was ambiguous. These results are significantly different to the
conclusions of \citet{2012ApJ...746...19Z}, who only found two null points:
we can only guess that the reason for the discrepancy lies with the specific
methods and techniques used to obtain the field lines, like, e.g., the region
chosen to carry out the extrapolation. We find that the average height above
the photosphere of the 8 clearly identified null points is $6$ Mm, with all
cases located at heights between $4$ and $10$ Mm.
To further explore this issue, additionally to the GONG data, we downloaded
the full-disk SOHO/MDI magnetogram at about the same time (01:39:30 ~UT) and
date, which has a pixel size of $1.99$\arcsec. We
selected from it a domain basically coinciding with the one used above for
the GONG magnetogram and carried out a field extrapolation. The MDI
magnetogram has an excellent level of flux balance (around $2\%$ ratio of
total signed flux to total unsigned flux for our region, or about $3.5 \%$
for \citeauthor{2012ApJ...746...19Z}'s original region), much better than the
corresponding numbers for the GONG data. The peak field strength in the MDI
region, about $500$ G, is also much higher than that in the GONG magnetogram
(close to $200$G). Our region being roughly centered on disk center, we do
not expect the deviation between line-of-sight and vertical field components
to play any relevant role in reaching the conclusions below. We could
clearly identify 8 null points associated with BPs in the list, with 6 of
them coinciding with positive identifications in our GONG extrapolation.
Two of the negative identifications were found in both
extrapolations. Two cases were marginal cases in either extrapolation; for
the remaining three cases there is no clear explanation for the discrepancy,
except perhaps through the large differences in the two magnetograms
concerning field strength of the flux concentrations. The null points
identified in the MDI extrapolation are located at slightly higher positions
above the photosphere, namely between $6$ and $12$ Mm, with an average of $8$
Mm.
All in all, we find that for a significant number of cases in the list of
\citet{2012ApJ...746...19Z} there is an association between a bright point and a null
point structure both in the GONG and the MDI datacheck.
From the analysis of the CH BP regions, we can conclude that the presence of a magnetic null point above the BP region appears to be very important. In all cases the BP regions are associated with some fraction of the fan-dome that encloses the minority polarity flux. Magnetic null points are known to be a location that can host magnetic reconnection processes, and through this process change the connectivity of the magnetic field globally while affecting all of the dome region due to the flipping of magnetic field lines through the diffusion region. The basic evolution of an initially potential magnetic null point has been investigated in detail by \citet{2007PhPl...14e2106P}, \citet{2011A&A...529A..20G}, and \citet{2011A&A...534A...2G}, which shows how the null point structure is perturbed strongly and generates systematic outflow regions that push the reconnected flux out of the current sheet region. A more realistic configuration is investigated by \citet{2013ApJ...774..154P}. It has a configuration that is comparable to the one shown in \Fig{extrapolation_compare.fig} where a flux balance of the data set is enforced. The investigation in \citet{2013ApJ...774..154P} shows how flux is transported through the null point while changing the connectivity patterns of the involved flux, and how different regions around the fan-dome change connectivity in the process. A similar evolution may take place in the CH BP magnetic field structures where the regions of change in connectivity is strongly altered by the inhomogeneity of the photospheric flux concentrations.
The skeleton defined by the null point located in the corona is very resilient and this structure may only be simplified by moving the null point out through the photospheric boundary. For the BPs discussed here the null points typically exist longer than the BPs are observed in X-rays. Only the BP4 case is different, where the X-ray emission is activated together with the formation of the null point through the photosphere and its following motion up into the coronal domain. As the X-ray emission vanishes, the null point vanished down through the photosphere again. This is the case that most clearly shows the importance of the null point in the appearance of the X-ray BP.
Having this characteristic magnetic skeleton raises the question about its role in the energy release process needed for heating the plasma to coronal temperatures. Looking at the time evolution of the magnetic field and the changes of the dome structures discussed in this paper (see the attached animations), it is clear that the convective motions constantly batter the footpoints of the magnetic field.
Convection is a stochastic process, so how does this provide a long time heating of a localised region of the fan-dome? The motions in the previous mentioned numerical investigations have all been imposing a systematic stressing of the null point system. So we may not directly infer these results to the BP systems. From the extrapolated field models shown in \Fig{BPrest_CH.fig} and \Fig{BPrest_QS.fig} one sees the X-ray emission, in many cases, to be associated with a dominating flux concentration, while there also exist examples where the emission is located above a weak field region (BP7). How do we explain how these different magnetic structured regions can host X-ray emission?
The dynamical response of the complicated magnetic field region to convective motions cannot be predicted from the knowledge of the magnetic field structure alone. To get a qualified understanding of the effect of the imposed convective motions, we have to turn to dedicated numerical experiments. Such work is presently in progress \citep{Moreno-Insertis-Galsgaard-2017}, where investigations of large scale systematic stressing is being studied, but we also need to explore the stochastic stressing process to establish the effect of these on the magnetic structure.
The plasma cooling time due to radiation (and conduction) of the BP X-ray plasma is on the order of 5~min \citep[See their Table~2]{2011A&A...526A.134A}, which indicates that a continuous energy input to the region is required. This energy may arise in part from free magnetic energy in the magnetic structure and from a continues stressing of the system. One may argue, therefore, that a potential field extrapolation is not the best choice when describing the magnetic field structure of BP regions. Presently though no sufficiently high quality vector magnetograms exist outside active regions that can be used for more advanced extrapolation methods like non-linear force free field extrapolations.
In addition, the requirement of the more advance methods is magnetic flux balance over the area of the adopted magnetogram. A condition that is never fulfilled in CH regions. Please see also the discussion in Section~2.2 on the comparison of magneto-static and potential field extrapolation methods.
\citet{2008A&A...492..575P} and \citet{2011A&A...526A.134A} investigated a few BP regions using the MPOLE code.
This code uses a relative low number of point sources to represent a magnetogram and from this analytic description they identify magnetic null points in the photosphere and their connections through spine axis and fan planes.
This reduced information was used to investigate the structure of the 3D magnetic field and to relate the magnetic structures to the BPs' X-ray emission. Both studies find the BP magnetic field structures to be bipolar in nature from identifying field lines that align with the X-ray emission. During this process they do not seem to investigate the large scale structure of the magnetic field, which prevents them from identifying any potential null points above the BP region. A direct comparison between their and our results on these scales is therefore not possible.
\section{Summary and Conclusions}
\label{conc.sec}
In this paper we have used SOT/XRT/Hinode and HMI/AIA/SDO observations to investigate the topology of X-ray/EUV BPs. 3D magnetic fields have been derived for different times using a potential field extrapolation for the full size of the observed magnetograms. Within these data the location of the BP regions were identified and the structure of the associated magnetic field investigated. The topology of the local magnetic field was visualised using the information of the position of the null points associated with the identified BP regions. This was used to investigate the close relation between the magnetic field structure and the projected location of the X-ray emission for the 10 selected BP regions.
The topology investigation showed a close relation between the CH BPs X-ray emission and the structure of the magnetic field defined by 3D null points. This shows that for all CH cases analysed here, the null points fan-dome structure embraces the X-ray emission in slightly different ways for the observed BPs. This strongly suggests that the presence of the null-point region located in the coronal region is crucial for the energy release processes for these BP regions.
Two cases of QS BPs also shows the presence of 3D null points, but these are located much higher in the atmosphere compared to the CH null points, while the third case has no coronal null point. This height of the two null points above the photosphere and the lack of knowledge of the height of the X-ray emission region makes it less clear how important the null points are for these BP regions. Just making field line traces from the photosphere makes one to assume that it is the low lying field lines, that show a bipolar structure, that represents the important structure of the BP regions. A good indication of the height of the X-ray emitting region would be important for identifying which part of the 3D magnetic field that hosts it. This would then allow for a discrimination between the different scenarios for the local magnetic field line structures responsible for hosting the BP regions and provide a substantial clue towards the heating process.
Ignoring the problem with the height of the X-ray region, and simply assuming that the simple approach for the local field line structure provides a sufficient representation for the structure of the field, then the investigated data indicate that two possible mechanisms exist for powering the BP regions. One which mainly works for the CH BPs where the coronal null point is the key location for the energy conversion. A different mechanism may represent the QS BPs where the local bipolar loop structure is more important than the presence of the associated null-point region. More observational data with high spacial resolution for QS BP are needed to better understand what the correct magnetic structures for this type of BP regions are. When this is determined more conclusively a much better starting point for investigating the dynamical evolution of the BPs will be presented.
\begin{acknowledgements}
KG acknowledges support through a research grant (VKR023406) from
VILLUMFONDEN. FMI is grateful for the financial
support by the Spanish Ministry of Economy and Competitiveness (MINECO)
through project AYA2014-55078-P. Z.H. acknowledges NSFC for financial support
via grant 41404135. TW acknowledges DLR-grant 50 OC 1301 and
DFG-grant WI 3211/4-1. The authors are grateful for the use of the
softwareImagery produced by VAPOR (www.vapor.ucar.edu),
a product of the Computational Information Systems Laboratory at the National
Center for Atmospheric Research.
\end{acknowledgements}
|
\section{Introduction} \label{sec-intro}
BPS states in supersymmetric field theories and string theory have remarkable properties, which have been actively studied in last decades. In this paper we consider BPS states that arise in D-brane systems, which encode properties of knots. The counting of these states leads to invariants of knots, referred to as Labastida-Mari{\~n}o-Ooguri-Vafa (LMOV) invariants (or Ooguri-Vafa invariants). On the other hand, dimensional reduction of such brane systems is expected to lead to a description in terms of a supersymmetric quiver quantum mechanics. In this paper we argue that this is indeed the case, and in consequence properties of BPS states are encoded in the data of moduli spaces of quiver representations, which leads to intriguing relations between knots and quivers. We presented a general idea of this correspondence in \cite{Kucharski:2017poe}. Now, in this paper, we explain more details of the identification between knot invariants and quiver moduli spaces, which enables us to identify explicitly relevant quivers for many knots, including some infinite families of knots, all knots with up to 6 crossings, some knots with thick HOMFLY-PT homology, etc. Understanding structural properties of generating series of knot polynomials also enables us to derive previously unknown expressions for colored HOMFLY-PT polynomials and superpolynomials (and their quadruply-graded generalizations) for several knots. More importantly, our correspondence relates generating series of colored HOMFLY-PT polynomials to motivic Donaldson-Thomas (DT) invariants, which then leads to the proof of the famous integrality of LMOV invariants, conjectured in \cite{OoguriV,Labastida:2000zp,Labastida:2000yw}. We also discuss many other consequences of the relation between BPS states, knots and quivers.
Both types of invariants mentioned above, i.e. LMOV invariants of knots and motivic Donaldson-Thomas invariants of quivers, are defined through factorization of some generating series. Our results, in particular the proof of the LMOV conjecture, follow from the identification of these series, which physically amounts to the identification of the corresponding BPS states. In case of knots, the series in question is the generating series of colored HOMFLY-PT polynomials, and it arises as the expectation value of the Ooguri-Vafa operator. This operator characterizes the system of branes, which provide topological string theory realization of observables in Chern-Simons theory. This system consists of $N$ A-model lagrangian branes wrapping $S^3$ in the deformed conifold $T^*S^3$ Calabi-Yau geometry, and intersecting -- along a knotted curve -- an additional set of lagrangian branes \cite{OoguriV}. Topological string amplitudes on each set of branes reduce to amplitudes in Chern-Simons theory, and the Ooguri-Vafa operator captures contributions from the scalar field describing strings stretched between these two sets of branes, whose amplitudes are identified with expectation values of Wilson loops in Chern-Simons theory \cite{Witten:1992fb,OoguriV}. According to the seminal work of Witten, such expectation values are identified with colored HOMFLY-PT polynomials \cite{Witten_Jones}, which are then assembled into a generating series that arises as the expectation value of the Ooguri-Vafa operator. The LMOV invariants that we consider are defined through factorization of this series \cite{OoguriV,Labastida:2000zp,Labastida:2000yw}. Upon the geometric transition, $N$ branes in the deformed conifold geometry are replaced by the resolved conifold geometry in the presence of additional lagrangian branes, which encode the topology of the original knot. Embedding this system in M-theory enables to interpret LMOV invariants as counting open M2-branes ending on M5-branes. However, integrality of these invariants has been verified only in some specific cases e.g. in \cite{OoguriV,Labastida:2000zp,Labastida:2000yw,Ramadevi_Sarkar,Mironov:2017hde}, as well as for some infinite families of knots and representations \cite{Garoufalidis:2015ewa,Kucharski:2016rlb}. In particular, in \cite{Kucharski:2016rlb} the relation of the framed unknot invariants (equivalently extremal invariants of twist knots, as well as open topological string amplitudes for branes in $\mathbb{C}^3$ geometry) to motivic Donaldson-Thomas invariants of the $m$-loop quiver was found, which led to the proof of integrality of BPS numbers in those cases; this relation was then analyzed and discussed also in \cite{Luo:2016oza,Zhu:2017lsn}.
Reducing the above mentioned open M2-brane states to their worldvolume is expected to lead to a description in terms of $\mathcal{N}=4$ supersymmetric quiver quantum mechanics. We find this quantum mechanics description by postulating that the Ooguri-Vafa generating function should be identified with the motivic generating series assigned to a putative quiver. Factorization of such a series defines motivic Donaldson-Thomas invariants, which also have an interpretation as the counts of BPS states \cite{Kontsevich:2008fj,Kontsevich:2010px}. If a quiver in question indeed exists, it is natural to identify these BPS states as the effective description of M2-M5 bound states in the Ooguri-Vafa description. As our main result -- announced already in \cite{Kucharski:2017poe} -- we show that the Ooguri-Vafa generating series indeed takes the form of the motivic generating series for some quiver, and we identify such quivers explicitly in various cases. For example, the quiver corresponding to the trefoil knot is shown in figure \ref{fig-trefoil}.
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.6\textwidth]{trefoil-quiver.png}
\caption{Trefoil knot and the corresponding quiver.} \label{fig-trefoil}
\end{center}
\end{figure}
BPS states that arise in the quiver description can be interpreted as elements of Cohomological Hall Algebras \cite{Kontsevich:2010px,COM:8276935,Rei12}, which provide prototype examples of algebras of BPS states, whose existence was postulated in \cite{Harvey:1996gc}. These structures are intimately related to the theory of wall-crossing and associated phenomena, which led to important results both in physics and mathematics in recent years. In our work we take advantage of some of those results, as well as suggest new directions of studies. For example, it has been proved that motivic Donaldson-Thomas invariants assigned to a symmetric quiver are integer \cite{efimov2012}. Our results lead to the identification of LMOV invariants with motivic Donaldson-Thomas invariants for symmetric quivers, which thus proves integrality of these LMOV invariants. More precisely, for knots for which we identify the corresponding quiver, the LMOV conjecture for all symmetric representations is automatically proved.
This is already an important result, and we expect that such corresponding quivers exist for all knots, and a general proof of the LMOV conjecture could be conducted along these lines. Some other identifications between quantities associated to knots and to quivers are shown in table \ref{tab-duality}.
\begin{table}
\begin{center}
\begin{tabular}{c|c}
{\bf Knots} & {\bf Quivers} \\
\hline
Homological degrees, framing & Number of loops \\
Colored HOMFLY-PT & Motivic generating series \\
LMOV invariants & Motivic DT-invariants \\
Classical LMOV invariants & Numerical DT-invariants \\
Algebra of BPS states & Cohom. Hall Algebra \\
\end{tabular}
\caption{Identification of various quantities associated to knots and quivers.} \label{tab-duality}
\end{center}
\end{table}
There are many other consequences and new relations that follow from our work. First, motivic Donaldson-Thomas invariants that we consider have an interpretation as certain topological characteristics of quiver moduli spaces \cite{MR,FR}. This suggests that quiver moduli spaces themselves should be interpreted as knot invariants, which leads to a novel kind of categorification in knot theory.
Second, we find that all HOMFLY-PT polynomials, as well as superpolynomials and their quadruply-graded generalizations, colored by arbitrary symmetric representations, are determined by a finite number of parameters: the matrix $C$ encoding the structure of the quiver corresponding to a given knot, and homological degrees of generators of the uncolored HOMFLY-PT homology. There should be a deeper reason why such limited information gives rise to rich structure and intricate properties of various infinite families of knot invariants.
Third, colored HOMFLY-PT polynomials and LMOV invariants can be defined for arbitrary (not only symmetric) representations and labeled by arbitrary Young diagrams. It is desirable to understand how this information is encoded in the corresponding quiver, or some generalization thereof. On the other hand, colored HOMFLY-PT polynomials labelled by symmetric representations satisfy a difference equation (encoded in $\widehat{A}$ operator), and their asymptotics is encoded in algebraic curves generalizing the A-polynomial \cite{AVqdef,Garoufalidis:2015ewa,Kucharski:2016rlb}. Such objects should also have an interesting interpretation in the context of quivers. In fact, for the $m$-loop quiver analogous functional equations have been discussed in \cite{COM:8276935}, and we expect that such relations should more generally play a role in quiver representation theory.
Fourth, having expressed colored HOMFLY-PT polynomials in the form of the motivic generating function, it is natural to replace one generating parameter associated with symmetric representations, by several parameters that naturally appear in motivic generating functions. This leads to a refinement of colored HOMFLY-PT polynomials, as well as LMOV invariants, and among others even stronger integrality statements.
Furthermore, motivic generating functions associated to quivers, as well as -- after our rewriting -- the generating functions of colored HOMFLY-PT polynomials, take the form of Nahm sums (with additional generating parameters) \cite{Nahm,Garoufalidis2015}. The Nahm sums have very intriguing properties, in special cases they are modular functions and may arise as characters of rational conformal field theories. It appears that both quiver representation theory, as well as knot theory, are rich sources of sums of this type.
Our work should also be related to many other results in literature. For example, uncolored HOMFLY-PT polynomials were related -- from a different perspective -- to Donaldson-Thomas invariants in \cite{Diaconescu-HOMFLY-DT}. A class of functions encoding integrality properties analogous to our generating functions has been analyzed in \cite{Schwarz:2013zua,Schwarz:2017gnp}. A detailed analysis of the LMOV conjecture was conducted in \cite{Dedushenko:2014nya}, and a refined LMOV conjecture was considered in \cite{Garoufalidis:2015ewa,Kameyama:2017ryw}.
This paper is organized as follows. In section \ref{sec-knots} we present appropriate background in knot theory and its relations to physics, including issues such as colored HOMFLY-PT polynomials, LMOV invariants, and knot homologies. In section \ref{sec-quivers} we introduce motivic Donaldson-Thomas invariants and other relevant notions from quiver representation theory. In section \ref{sec-knots-quivers} we present our main conjectures, motivated by physical interpretation of knot invariants in terms of supersymmetric quiver quantum mechanics, and relating various knot invariants to invariants of quivers. In section \ref{sec-results} we illustrate these conjectures in many examples, including infinite families of torus and twist knots, all knots with up to 6 crossings, and examples of thick knots. Using our results we also determine previously unknown HOMFLY-PT polynomials and superpolynomials colored by arbitrary symmetric representations for $6_2$ and $6_3$ knots, as well as for $(3,7)$ torus knot.
\section{Knot theory and physics} \label{sec-knots}
Knot theory plays a prominent role in contemporary high energy and mathematical physics. As a branch of topology, it is not surprising that it is intimately related to topological field and string theories. It is perhaps more surprising, that through these links not only physics provides an interpretation of mathematical facts, but it is also a source of new ideas, which are subsequently formalized and (hoped to be) proved by mathematicians. Examples of such ideas, relevant in the context of our work, include Labastida-Mari{\~n}o-Ooguri-Vafa (LMOV) invariants, superplynomials and HOMFLY-PT homologies, quadruply-graded homologies, etc. In this section we recall and briefly summarize all these notions and introduce notation used in what follows.
\subsection{Knot invariants and the LMOV conjecture}
Polynomial knot invariants, including the Alexander polynomial known for almost 100 years, and the much younger Jones polynomial, form one important class of knot invariants. The Jones polynomial was subsequently generalized to the two-parameter HOMFLY-PT polynomial, and colored versions of these polynomials were introduced. Witten's interpretation of these polynomials as expectation values of Wilson loops in Chern-Simons theory \cite{Witten_Jones} played an important role in those developments. Furthermore, the Chern-Simons interpretation was also shown to be related to topological string theory \cite{Witten:1992fb}. This paved the way to subsequent formulation of LMOV invariants and the famous conjecture, stating that these invariants are integer \cite{OoguriV,Labastida:2000zp,Labastida:2000yw,Labastida:2001ts}. While this conjecture was verified in various specific situations
\cite{OoguriV,Labastida:2000zp,Labastida:2000yw,Ramadevi_Sarkar,Mironov:2017hde,Garoufalidis:2015ewa,Kucharski:2016rlb}, and some attempts of its general proof were undertaken \cite{Liu:2007kv}, it still appears to be an open problem. One aim of our work is to provide a proof of this conjecture, at least for a large class of knots and representations. However, let us first introduce relevant notation.
As follows from \cite{Witten_Jones}, colored HOMFLY-PT polynomials can be interpreted as expectation values of Wilson loops in representation $R$ in Chern-Simons theory
\begin{equation}
\overline{P}_{R}(a,q) = \langle \textrm{Tr}_R U \rangle,
\end{equation}
where $U=P\,\exp\oint_K A$ is the holonomy of $U(N)$ Chern-Simons gauge field along a knot $K$. Here the HOMFLY-PT polynomial is unreduced, i.e. it is normalized as
\begin{equation}
\overline{P}_{R}(a,q)=\overline{P}_R^{\bf 0_1} P_R(a,q),
\end{equation}
where $P_R(a,q)$ is the corresponding reduced colored HOMFLY-PT polynomial (equal to 1 for the unknot), and $\overline{P}_R^{\bf 0_1}$ is the normalization factor of the unknot. As we will explain in what follows, our results depend in a crucial way on the choice of this normalization.
After embedding Chern-Simons theory in string theory, as we sketched in the introduction, it was shown in \cite{OoguriV} that the following generating function -- often referred to as the Ooguri-Vafa operator -- is natural to consider
\begin{equation}
Z(U,V) = \sum_R \textrm{Tr}_R U \, \textrm{Tr}_R V = \exp\Big( \sum_{n=1}^{\infty} \frac{1}{n} \textrm{Tr} U^n \textrm{Tr} V^n \Big),
\end{equation}
where $V$ is interpreted as a source, and the sum runs over all representations $R$, i.e. all two-dimensional partitions. The expectation value of this expression is the generating function of colored HOMFLY-PT polynomials. It was postulated in \cite{OoguriV,Labastida:2000zp,Labastida:2000yw,Labastida:2001ts} that this expectation value has the following structure
\begin{equation}
\big\langle Z(U,V) \big\rangle = \sum_R \overline{P}_{R}(a,q) \textrm{Tr}_R V =
\exp \Big( \sum_{n=1}^\infty \sum_R \frac{1}{n} f_{R}(a^n,q^n) \textrm{Tr}_R V^n \Big), \label{ZUV}
\end{equation}
where the functions $f_R(a,q)$ take the form
\begin{equation}
f_{R}(a,q) = \sum_{i,j} \frac{N_{R,i,j} a^i q^j}{q-q^{-1}} \label{fR}
\end{equation}
and encode conjecturally integer $N_{R,i,j}$ numbers. The functions $f_R(a,q)$ can be expressed as universal polynomials in colored HOMFLY-PT polynomials. The above statements, concerning the structure of $\big\langle Z(U,V) \big\rangle$ and integrality of $N_{R,i,j}$, are referred to as the LMOV conjecture, and $N_{R,i,j}$ are called LMOV invariants (or Ooguri-Vafa invariants). As indicated in the introduction, in the physics interpretation they count bound states of M2-branes ending on M5-branes.
Of our main interest in this work is the generating function of $S^r$-colored HOMFLY-PT polynomials. It can be obtained by considering a one-dimensional source $V=x$. In this case $\textrm{Tr}_R V \neq 0$ only for symmetric representations $R=S^r$, and then $\textrm{Tr}_{S^r}(x) = x^r$. Then (\ref{ZUV}) reduces to the generating function of $S^r$-colored HOMFLY-PT polynomials, and denoting the $S^r$-colored HOMFLY-PT polynomial by $\overline{P}_{r}(a,q)$ we get
\begin{equation}
P(x) = \langle Z(U,x) \rangle = \sum_{r=0}^\infty \overline{P}_{r}(a,q) x^r =
\exp\Big({\sum_{r,n\geq 1} \frac{1}{n} f_{r}(a^n,q^n)x^{n r}}\Big),
\label{Pz2}
\end{equation}
with
\begin{equation}
f_r(a,q)\equiv f_{S^r}(a,q) = \sum_{i,j} \frac{N_{r,i,j} a^i q^j}{q-q^{-1}},
\end{equation}
where LMOV invariants are denoted by $N_{r,i,j}\equiv N_{S^r,i,j}$. These functions are polynomials, with rational coefficients, of $\overline{P}_{d_1}(a^{d_2},q^{d_2})$ for some $d_1$ and $d_2$ (with $d_1 d_2\le r$):
\begin{align}
f_1(a,q) &= \overline{P}_1(a,q), \nonumber\\
f_2(a,q) &= \overline{P}_2(a,q) - \frac{1}{2} \overline{P}_1(a,q)^2 -\frac{1}{2} \overline{P}_1(a^2,q^2), \nonumber \\
f_3(a,q) &= \overline{P}_3(a,q) - \overline{P}_1(a,q)\overline{P}_2(a,q) + \frac{1}{3}\overline{P}_1(a,q)^3 - \frac{1}{3} \overline{P}_1(a^3,q^3), \nonumber
\end{align}
etc. One can also rewrite (\ref{Pz2}) in the product form
\begin{equation}
P(x) = \prod_{r\geq 1;i,j;k\geq 0} \Big(1 - x^r a^i q^{j+2k+1} \Big)^{N_{r,i,j}}.
\label{Pr-LMOV}
\end{equation}
One of our aims is to show integrality of BPS degeneracies $N_{r,i,j}$ encoded in this product.
In the (classical) limit $q\to 1$, a special role is played by a subset of LMOV invariants, referred to as classical LMOV invariants. To define them it is useful to consider the ratio
\begin{equation}
y(x,a) = \lim_{q\to 1} \frac{P(qx)}{P(x)} =
\lim_{q\to 1} \prod_{r\geq 1;i,j;k\geq 0} \Big(\frac{1 - x^r a^i q^{r+j+2k+1} }{1 - x^r a^i q^{j+2k+1}}\Big)^{N_{r,i,j}}
= \prod_{r\geq 1;i} (1 - x^r a^i)^{-r b_{r,i} / 2}, \label{yxa}
\end{equation}
with classical LMOV invariants defined as
\begin{equation}
b_{r,i} = \sum_j N_{r,i,j}. \label{bri}
\end{equation}
It turns out that $y=y(x,a)$ defined above satisfy algebraic equations
\begin{equation}
A(x,y)=0
\end{equation}
of A-polynomial type \cite{AVqdef,Garoufalidis:2015ewa}.
\subsection{Knot homologies} \label{ssec-knothom}
Another important class of knot invariants are knot homologies. First well understood examples of such structures are Khovanov homology \cite{Khovanov} and Khovanov-Rozansky homology \cite{KhR1,KhR2}. In our work an important role is played by their putative, highly nontrivial generalization, namely colored HOMFLY-PT homology $\mathcal{H}^{S^r}_{i,j,k}$, which categorifies the HOMFLY-PT polynomial colored by symmetric representations $S^r$. It has been defined rigorously by mathematicians only recently \cite{Cautis}, yet only for the unreduced version, and it is still not suitable for explicit computations (there also exist some constructions in the case of antisymmetric representations $\Lambda^r$, both reduced and unreduced versions, see e.g. \cite{Wedrich:2016smm}, which are conjecturally isomorphic to the homologies corresponding to the symmetric representations). Nonetheless, the conjectural Poincar\'{e} polynomial of (reduced) colored HOMFLY-PT homology, referred to as the superpolynomial
\begin{equation}
P_r (a,q,t) = \sum_{i,j,k} \, a^i q^j t^k \, \dim \mathcal{H}^{S^r}_{i,j,k}, \label{superpolynomial}
\end{equation}
can be determined for various families of knots, for example using the formalism of differentials and the structural properties of the (colored) homologies \cite{DGR,Rasmussen-differentials,Gukov:2011ry,Gukov:2015gmm} -- the formalism that we will exploit in the present paper. It has been postulated that HOMFLY-PT homology should be identified with the space of BPS states in relevant brane systems \cite{GSV,Gukov:2011ry}. Superpolynomials for the unknot or the Hopf-link can be also computed by techniques of refined topological string theory \cite{GIKV}, and superpolynomials for torus knots can be computed by means of refined Chern-Simons theory \cite{AS,FGS,DMMSS,superA,ItoyamaMMM}. Colored superpolynomials reduce to colored HOMFLY-PT polynomials upon the substitution $t=-1$. As we will see, one interesting result of our work is an explicit relation between colored HOMFLY-PT polynomials and the (uncolored) superpolynomial.
Let us briefly present structural properties of the reduced $S^r$-colored HOMFLY-PT homologies $\mathcal{H}^{S^r}$ of a given knot \cite{Gukov:2011ry,Gorsky:2013jxa}. First, for a given knot, for every $k=0,\ldots,r-1$, there exists a (positive, vertical) colored differential $d_{1-k}$ on $\mathcal{H}^{S^r}$, of $(a,q,t)$-degree $(-2,2-2k,-1)$, such that the homology of $\mathcal{H}^{S^r}$ with respect to $d_{1-k}$ is isomorphic to $\mathcal{H}^{S^k}$. Second, for every $k=0,\ldots,r-1$ there is another set of (negative, vertical) colored differentials $d_{-r-k}$ on $\mathcal{H}^{S^r}(K)$, of $(a,q,t)$-degree $(-2,-2r-2k,-3)$,
such that the homology of $\mathcal{H}^{S^r}(K)$ with respect to $d_{-r-k}$ is isomorphic to $\mathcal{H}^{S^k}$. Third, there is a universal colored differential $d_{2\to 1}$
of degree $(0,2,0)$ on the homology $\mathcal{H}^{S^2}(K)$, such that the homology of $\mathcal{H}^{S^2}$ with respect to $d_{2\to 1}$ is isomorphic to the uncolored homology $\mathcal{H}^{S}$. All these differentials relate homology theories with different values of $r$. The uncolored homology, corresponding to $r=1$, supposedly categorifies the reduced HOMFLY-PT polynomial, and its Poincar\'e polynomial is simply the original (uncolored) superpolynomial introduced in \cite{DGR}.
Furthermore, HOMFLY-PT homologies of a large class of knots satisfy the refined exponential growth, which implies the following relation for their colored superpolynomials
\begin{equation}
P_{S^r}(a,q=1,t)=\left( P_{\tableau{1}}(a,q=1,t) \right)^r, \label{exp-growth}
\end{equation}
see also \cite{Wedrich:2016smm}. Properties of colored differentials, together with the assumption of the exponential growth, enable to determine an explicit form of the colored superpolynomial $P_{S^r}(a,q,t)$ for various knots \cite{Gukov:2011ry,superA,FGSS}. For example, colored superpolynomials for the trefoil knot $3_1$ take the form \cite{superA,FGSS}
\begin{equation}
P_r(a,q,t)= \frac{a^{2r}}{q^{2r}}\sum_{k=0}^r {r \brack k} q^{2k(r+1)} t^{2k} \prod_{i=1}^k (1+a^2q^{2(i-2)}t), \label{Pr-31}
\end{equation}
where
\begin{equation}
{r \brack k} = \frac{(q^2;q^2)_r}{(q^2;q^2)_k(q^2;q^2)_{r-k}}, \label{q-binomial}
\end{equation}
and the $q$-Pochhammer symbol is defined as
\begin{equation}
(x;q)_r =\prod_{k=0}^{r-1} (1-x q^k),\qquad \qquad (x;q)_\infty =\prod_{k=0}^\infty (1-x q^k) \,. \label{qdilog}
\end{equation}
For $t=-1$, (\ref{Pr-31}) specializes to the reduced colored HOMFLY-PT polynomial for trefoil knot, while in the uncolored case (i.e. for $r=1$), (\ref{Pr-31}) reduces to
\begin{equation}
P_1(a,q,t) = \frac{a^2}{q^2} + a^2 q^2 t^2 + a^4 t^3. \label{P1-31}
\end{equation}
The monomials in this expression correspond to generators of the HOMFLY-PT homology, and powers of $t$ in each monomial -- in this example taking values $(0,2,3)$ -- are referred to as homological degrees.
\begin{figure}[b]
\begin{center}
\begin{tikzpicture}
\draw [help lines] (0,0) grid (3,2);
\draw [very thick] (0,2.2)--(0,0)--(3.2,0);
\draw [thick] (-1,-1)--(0,0);
\draw [thick] (0,-1)--(0,0);
\draw [thick] (-1,0)--(0,0);
\draw (-0.7,-0.3) node {{\textbf{\it a}}};
\draw (-0.3,-0.7) node {{\it q}};
\filldraw [black] (0.5,0.5) circle (2pt);
\filldraw [black] (1.5,1.5) circle (2pt);
\filldraw [black] (2.5,0.5) circle (2pt);
\draw (0.33,0.72) node {$0$};
\draw (1.3,1.7) node {$3$};
\draw (2.65,0.7) node {$2$};
\draw (-0.45,0.5) node {$2$};
\draw (-0.45,1.5) node {$4$};
\draw (0.48,-0.5) node {$-2$};
\draw (1.55,-0.5) node {$0$};
\draw (2.55,-0.5) node {$2$};
\draw [->,>=stealth,thin] (1.6,1.4)--(2.4,0.6);
\draw [->,>=stealth,thin] (1.4,1.4)--(0.6,0.6);
\draw [help lines] (7,-1) grid (10,2);
\draw [very thick] (7,2.2)--(7,-1)--(10.2,-1);
\draw [thick] (7,-1)--(6,-2);
\draw [thick] (7,-1)--(6,-1);
\draw [thick] (7,-1)--(7,-2);
\draw (6.3,-1.3) node {{\textbf{\it a}}};
\draw (6.7,-1.7) node {{\it q}};
\filldraw [black] (7.5,0.5) circle (2pt);
\filldraw [black] (8.5,0.5) circle (2pt);
\filldraw [black] (9.5,0.5) circle (2pt);
\filldraw [black] (8.5,1.5) circle (2pt);
\filldraw [black] (8.5,-0.5) circle (2pt);
\draw (7.38,-1.5) node {$-2$};
\draw (8.45,-1.5) node {$0$};
\draw (9.45,-1.5) node {$2$};
\draw (6.45,1.5) node {$2$};
\draw (6.45,0.5) node {$0$};
\draw (6.39,-0.5) node {$-2$};
\draw (7.32,0.73) node {$-1$};
\draw (8.5,0.73) node {$0$};
\draw (9.63,0.73) node {$1$};
\draw (8.4,1.75) node {$2$};
\draw (8.3,-0.77) node {$-2$};
\draw [->,>=stealth,thin] (8.6,1.4)--(9.4,0.6);
\draw [->,>=stealth,thin] (8.4,1.4)--(7.6,0.6);
\draw [->,>=stealth,thin] (7.6,0.4)--(8.4,-0.4);
\draw [->,>=stealth,thin] (9.4,0.4)--(8.6,-0.4);
\end{tikzpicture}
\end{center}
\caption{Diagrams of the reduced uncolored HOMFLY-PT homology of the trefoil (left) and the figure-eight knot (right).}
\label{diagrams}
\end{figure}
It is convenient to present the structure of (colored) HOMFLY-PT homology in terms of diagrams on a two-dimensional lattice. Each homology generator is represented by a dot at position $(i,j)$ in such a lattice, with $i$ and $j$ representing respectively $q$-degree and $a$-degree of this generator, and whose $t$-degree can in addition be written explicitly in the diagram as label of a corresponding dot. In addition, differentials acting between pairs of generators can be represented by arrows in the diagram. In figure \ref{diagrams} are presented the diagrams of the reduced uncolored HOMFLY-PT homologies of the trefoil and the figure-eight knot.
The structure of differentials implies that in the case of the uncolored homologies (the ones that are of our main interest in the paper) the generators must form two larger structures, which we call a zig-zag and a diamond. A zig-zag is a string of an odd number of generators, and a diamond consists of four generators (as the name indicates, distributed in the form of a diamond). For each knot, its HOMFLY-PT homology must contain precisely one zig-zag (possibly of length one, i.e. consisting of a single generator), and an arbitrary number of diamonds. For example, in the case of the diagrams in figure \ref{diagrams}, the diagram of the trefoil knot consists of a single zig-zag of length three, while the diagram for the figure-eight knot consists of a zig-zag of length one (only the generator in the middle of the diagram with the label 0), and one diamond formed by the remaining four generators. The (finite-dimensional) homology of a link has as many zig-zags as the number of its components.
The structure of colored HOMFLY-PT homology was further generalized to quadruply-graded homology, which has a richer structure of differentials \cite{Gorsky:2013jxa}. The Poincar\'e polynomial of this quadruply-graded homology $P_r(a,Q,t_r,t_c)$ depends on four parameters $a,Q,t_r$ and $t_c$, and specializes to the colored superpolynomial upon the identification
\begin{equation}
P_r(a,q,t) = P_r(a,Q=q,t_r=t q^{-1},t_c=q), \label{quad-super}
\end{equation}
and to the colored HOMFLY-PT polynomial upon
\begin{equation}
P_r(a,q) = P_r(a,Q=q,t_r=-q^{-1},t_c=q). \label{super-homfly}
\end{equation}
Quadruply-graded homologies for a large class of knots satisfy the refined exponential growth, which implies the following relation for the corresponding Poincar\'e polynomials
\begin{equation}
P_r(a,Q,t_r,t_c=1)=\big(P_1(a,Q,t_r,t_c=1)\big)^r. \label{exp-growth-quad}
\end{equation}
\section{Quiver moduli and Donaldson-Thomas invariants} \label{sec-quivers}
In this section we present basic properties of quivers and moduli spaces of their representations, which will be crucial in the rest of the paper. Moduli spaces of quiver representations have a rich structure, which among others provides a natural playground for the theory of (motivic) Donaldson-Thomas invariants, Cohomological Hall Algebras, etc.
A quiver $Q$ is an oriented graph with a finite set of vertices $Q_0$ and finitely many arrows between vertices $\alpha:i\rightarrow j$, for $i,j \in Q_0$. On $\mathbb{Z}Q_0$, we define the Euler form of $Q$ by
\begin{equation}
\langle {\bm d},{\bm e}\rangle_Q=\sum_{i\in Q_0}d_ie_i-\sum_{\alpha:i\rightarrow j}d_ie_j.
\end{equation}
A quiver representation assigns to each vertex $i\in Q_0$ a vector space of dimension $d_i$, and a linear map between two such spaces to each arrow. The vector ${\bf d}=(d_1,\ldots,d_m)$ is referred to as the dimension vector.
As we will see, quivers which appear in relation to knot invariants are symmetric, meaning that for any pair of their vertices $i$ and $j$, the number of arrows from $i$ to $j$ is equal to the number of arrows from $j$ to $i$. While explicit expressions for invariants describing moduli spaces of quiver representations are hard to find in general, they are quite well understood in the case of symmetric quivers \cite{Kontsevich:2010px,efimov2012,FR,MR}. An important information about the moduli space of representations of a symmetric quiver is encoded in the following generating series
\begin{equation}
P_Q({\bm x})=\sum_{{\bm d}\in\mathbb{N}Q_0}(-q)^{-\langle {\bm d},{\bm d}\rangle_Q} {\bm x}^{\bm d} \prod_{i\in Q_0}\prod_{j=1}^{d_i}\frac{1}{1-q^{-2j}} \label{PQx}
\end{equation}
where ${\bm x}^{\bm d}=\prod_{i\in Q_0}x_i^{d_i}$. In particular, motivic Donaldson-Thomas invariants $\Omega_{{\bm d},j}\equiv \Omega_{d_1,\ldots,d_m;j}$, assembled into
\begin{equation}\label{OmegaPQx}
\Omega_{\bm d}(q) = \sum_j \Omega_{{\bm d},j} (-q)^j,
\end{equation}
are defined once (\ref{PQx}) is rewritten as
\begin{equation}
P_Q({\bm x})={\rm Exp}(\frac{1}{q^{-1}-q}\sum_{{\bm d}\neq0}\Omega_{\bm d}(q){\bm x}^{\bm d}),
\end{equation}
where ${\rm Exp}$ is the plethystic exponential defined by ${\rm Exp}(f+g)={\rm Exp}(f)\cdot{\rm Exp}(g)$ and ${\rm Exp}(q^ix^d)=\frac{1}{1-q^ix^d}$ for $i\in\mathbb{Z}$ and $d\in\mathbb{N}Q_0$. This plethystic exponential form can be written equivalently as a product decomposition
\begin{equation}
P_Q({\bm x})=
\prod_{{\bm d}\neq 0} \prod_{j\in\mathbb{Z}} \prod_{k\geq 0} \big(1 - (-1)^j {\bm x}^{\bm d} q^{j+2k+1} \big)^{-\Omega_{{\bm d},j}}. \label{PQx-Omega}
\end{equation}
Two geometric interpretations of invariants $\Omega_{\bm d}(q)$, either as the intersection Betti numbers of the moduli space of all semisimple representations of $Q$ of dimension vector ${\bm d}$, or as the Chow-Betti numbers of the moduli space of all simple representations of $Q$ of dimension vector ${\bm d}$, were provided in \cite{MR,FR}. It was also proved these invariants are positive integers \cite{efimov2012}. One of our aims in this paper is to relate these invariants to LMOV invariants of knots.
One can also introduce numerical Donaldson-Thomas invariants of a quiver. To this end, for a vector ${\bm n}\in\mathbb{N}Q_0$, we denote by $P_Q(q^{\bm n}{\bm x})$ the series arising from $P_Q({\bm x})$ by replacing every ${\bm x}^{\bm d}$ by $q^{{\bm n}\cdot {\bm d}} {\bm x}^{\bm d}$, where ${\bm n}\cdot {\bm d}=\sum_{i\in Q_0}n_id_i$. Then
\begin{equation}
\frac{P_Q((-q)^{\bm n}{\bm x})}{P_Q((-q)^{-\bm n}{\bm x})}={\rm Exp}(\sum_{d\not=0}\frac{(-q)^{{\bm n}\cdot {\bm d}}-(-q)^{-{\bm n}\cdot {\bm d}}}{(-q)-(-q)^{-1}}\Omega_{\bm d}(q){\bm x}^{\bm d}),
\end{equation}
and in this equation we can specialize $q$ to $1$. By \cite{Engel2009}, the left hand side specializes then to the generating series of the Euler characteristic of certain Hilbert schemes ${\rm Hilb}_{d,n}(Q)$ attached to the quiver (these numbers admit a combinatorial interpretation by counting certain kinds of trees), so that we get
\begin{equation}
\sum_{d\in\mathbb{N}Q_0}\chi({\rm Hilb}_{{\bm d},{\bm n}}(Q)){\bm x}^{\bm d}={\rm Exp}(\sum_{{\bm d}\not=0}({\bm n}\cdot {\bm d})\Omega_d(1) {\bm x}^{\bm d})=\prod_{{\bm d}\not=0}(1-{\bm x}^{\bm d})^{-({\bm n}\cdot {\bm d})\Omega_{\bm d}(1)}. \label{Omega-classical}
\end{equation}
These $\Omega_{\bm d}(1)$ are the (numerical) Donaldson-Thomas invariants of the quiver.
\section{Knot invariants from quivers} \label{sec-knots-quivers}
In this section we first present our main claim, relating various knot invariants to quivers. We also discuss its various implications, and develop a formalism facilitating computations and enabling to determine quivers associated to knots. Our claim takes form of the following conjectures. We show that these conjectures are correct in many explicit and nontrivial examples in section \ref{sec-results}.
\subsection{Main conjectures}
\begin{conjecture}\label{PxCConj}
For a given knot, the generating function of its (appropriately normalized, as explained in detail in section \ref{ssec-unreduced}) colored HOMFLY-PT polynomials (\ref{Pz2}) can be written in the form
\begin{equation}
P(x) = \sum_{r=0}^{\infty} \overline{P}_r(a,q) x^r = \sum_{d_1,\ldots,d_m\geq 0} x^{d_1+\ldots + d_m} q^{\sum_{i,j} C_{i,j} d_i d_j} \frac{\prod_{i=1}^m q^{l_i d_i} a^{a_i d_i}(-1)^{t_i d_i}}{\prod_{i=1}^m(q^2;q^2)_{d_i}} \label{PxC}
\end{equation}
where $C$ is a (symmetric) $m\times m$ matrix, and $l_i,a_i$ and $t_i$ are fixed integers. Note that terms proportional to $x^r$, with fixed $r$, arise from sets of $\{d_i \}$ such that $r=d_1+\ldots +d_m$. Remarkably, (\ref{PxC}) has the same form as the motivic generating function (\ref{PQx}) of a symmetric quiver determined by the matrix $C$, up to the identification $q \mapsto -q$ and the specialization of variables
\begin{equation}
x_i = x a^{a_i} q^{l_i-1} (-1)^{t_i}. \label{specialize}
\end{equation}
The number of vertices $m$ of such a quiver is given by the size of $C$, and the number of arrows between vertices $i$ and $j$ is given by the matrix element $C_{i,j}$ (in particular $C_{i,i}$ denotes the number of loops at vertex $i$).
\end{conjecture}
It follows that to a given knot one can assign a quiver, so that various invariants of this knot are encoded in the data of moduli spaces of quiver representations of this corresponding quiver. Moreover, all this information is encoded in a finite set of parameters that determine (\ref{PxC}): the matrix $C$, as well as integers $l_i,a_i,t_i$ that are encoded in the (uncolored, reduced) superpolynomial of the knot in question. Recall that the uncolored, reduced superpolynomial for a given knot is a sum of monomials of the form $a^{a_i} q^{q_i} t^{t_i}$, which correspond to generators of the HOMFLY-PT homology.
\begin{conjecture} \label{conj2}
The size of the matrix $C$ (the number of vertices in the corresponding quiver) is equal to the number of generators of uncolored HOMFLY-PT homology. Furthermore, with appropriate ordering of vertices, $t_i$ in (\ref{PxC}) agree with homological degrees of generators of HOMFLY-PT homology, diagonal elements of $C$ are also equal to homological degrees, i.e. $C_{i,i}=t_i$, coefficients of linear powers of $q$ take the form $l_i=q_i-t_i$, and $a_i$ are equal to $a$-degrees of generators of uncolored HOMFLY-PT homology. An additional minus sign in (\ref{PxC}) comes with the power determined by $t_i$, so that it is relevant only for generators with odd $t$-grading.
\end{conjecture}
Note that it follows that homological degrees $t_i$ can be identified (as diagonal elements of matrix $C$) after rewriting the generating series (\ref{Pz2}) in the quiver-like form, and they are given by the number of loops in the corresponding quiver. This means, that the uncolored superpolynomial is encoded in the form of colored HOMFLY-PT polynomials, which is quite a surprising observation.
For knots that satisfy the refined exponential growth (\ref{exp-growth}) it is not hard to see where the coefficients of linear terms in $d_i$, in powers of $q,a$ and $(-1)$ in (\ref{PxC}), come from. In general, in expressions for quadruply-graded superpolynomials, $t_c$ does not appear in powers which are linear in summation variables, so -- if (\ref{exp-growth-quad}) holds -- linear powers of other parameters can be identified upon specialization $t_c=1$. Furthermore, recall that colored superpolynomials arise upon the substitution (\ref{quad-super}), and colored HOMFLY-PT polynomials upon (\ref{super-homfly}). It follows that for arbitrary color $r=d_1+\ldots+d_m$, a linear term in the exponent of $q$ takes the form
\begin{equation}
\sum_i (q_i-t_i) d_i, \label{linear}
\end{equation}
where the sum is over all generators $i$ of the uncolored homology, $q_i$ and $t_i$ are their $q$- and $t$-degrees, and $d_i$ is the corresponding summation index. Analogously, linear powers of parameters $a$ and $(-1)$ in (\ref{PxC}) must, respectively, take the form $\sum_i a_i d_i$ and $\sum_i t_i d_i$. The same formulas are valid for the unreduced homology, since the refined exponential growth holds for the unreduced homology of the unknot.
Note that one can also focus on those parts of colored HOMFLY-PT polynomials or superpolynomials which are proportional to the highest or lowest power of the variable $a$ \cite{Gorsky:2013jxa,Garoufalidis:2015ewa}. The corresponding generators of HOMFLY-PT homology lie respectively in the top or bottom row of the homology diagram, so such invariants are often referred to as top/bottom row invariants, or extremal invariants. For a large class of knots satisfying the exponential growth property (\ref{exp-growth}), the generating function of their colored extremal reduced HOMFLY-PT polynomials also takes a universal form (\ref{PxC}), however with the dependence on $a$ suppressed
\begin{equation}
P^{bottom/top}(x)= \sum_{d_1+d_2+\cdots+d_m\geq 0} x^{d_1+\ldots +d_m} q^{\sum_{i,j} C_{i,j} d_i d_j}\frac{q^{\sum_i (q_i-t_i) d_i} (-1)^{\sum_i {t_i d_i} }}{\prod_{i=1}^m (q^2;q^2)_{d_i}}. \label{quivgen2}
\end{equation}
Here $m$ is the dimension of the fundamental homology corresponding to the bottom/top row, and $q_i$ and $t_i$, $i=1,\ldots,m$, are $q$-degrees and $t$-degrees of these $m$ generators. In this case the matrix $C$ encodes a quiver which is a subquiver (capturing only extremal $a$-dependence) of the full quiver associated to a given knot.
The conjecture \ref{conj2} relates various quantities associated to knots to those of quiver moduli. Note that other relations of this type also follow -- one another example of such a relation is the dependence on framing. The operation of framing by $f\in \mathbb{Z}$ changes the colored HOMFLY-PT polynomial by a factor, which for the symmetric representation $S^r$ takes the form
\begin{equation}
a^{2fr}q^{f \, r(r-1)}. \label{framing}
\end{equation}
From the viewpoint of the quiver generating function (\ref{PxC}), the term with quadratic (in $r$) power of $q$
\begin{equation}
q^{fr^2}=q^{f(\sum_i d_i)^2} = q^{f\sum_{i,j} d_i d_j}
\end{equation}
shifts all elements of the matrix $C$ by $f$
\begin{equation}
C\ \mapsto \ C+f\left[\begin{array}{cccc} 1&1&\cdots&1\\
1&1&\cdots&1\\
\vdots&\vdots&\ddots&\vdots\\
1&1&\cdots&1
\end{array}\right] \label{C-frame}
\end{equation}
which in the quiver interpretation corresponds to adding $f$ loops at each vertex and $f$ pairs of oppositely-oriented arrows between all pairs of vertices. Note that, while in the context of quivers all entries of a matrix $C$ should be nonnegative, in some examples coming from knots we find matrices $C$ with negative entries. In this case a change of framing can be used to shift such negative values, and make all entries of $C$ nonnegative; such a modified matrix still describes the same knot.
Furthermore, we can also characterize the structure of the matrix $C$ in more detail.
\begin{conjecture} \label{conj3}
For a given knot, the matrix $C$ has a block structure
$$
C = \left[
\begin{array}{c|c|c|c}
b_{1,1} & \cdots & b_{1,k} & \cdots \\ \hline
\vdots & \ddots & \vdots & \\ \hline
b_{1,k}^T & \cdots & b_{k,k} & \\ \hline
\vdots& & & \ddots \\
\end{array}
\right]
$$
Diagonal blocks $b_{k,k}$ correspond to structural elements of the HOMFLY-PT homology, introduced in section \ref{ssec-knothom}. One of those blocks (in case of knots; or as many blocks as the number of components of a link) corresponds to the zig-zag element of length $2p+1$, and it has the same form (up to some permutation of homology generators, and up to an overall shift by a constant matrix with integer coefficients as in (\ref{C-frame}), corresponding to framing) as the matrix $C$ for the $(2,2p+1)$ torus knot (\ref{C-torus22p1}). All other diagonal block elements correspond to diamonds and (up to a permutation of homology generators) take the form
\begin{equation}
\left[\begin{array}{cccc} k&k&k+1&k+1\\
k&k+1&k+2&k+2\\
k+1&k+2&k+3&k+3\\
k+1&k+2&k+3&k+4
\end{array}\right] \label{diamond}
\end{equation}
for a fixed $k$ (which may be different for each block). The structure of other blocks $b_{l,k}$ for $l\neq k$ depends only on the structural elements corresponding to diagonal blocks $b_{l,l}$ and $b_{k,k}$.
\end{conjecture}
\subsection{Superpolynomials and quadruply-graded homology of knots from quivers}
In the above conjectures we related generating functions of colored HOMFLY-PT polynomials to the motivic generating series of some quiver. However, in addition we postulate that the same quiver encodes also generating functions of colored superpolynomials, as well as Poincar{\'e} polynomials of quadruply-graded HOMFLY-PT homology.
\begin{conjecture}\label{PxCConj-quadruply}
Consider a knot satisfying the exponential growth property (\ref{exp-growth-quad}), with the corresponding quiver -- determined as explained above -- represented by a matrix $C$, the size of the reduced colored homology denoted by $m$, and $(a,q,t)$-degrees of its generators denoted by $(a_i,q_i,t_i)$. Then, the Poincar{\'e} polynomial of the reduced quadruply-graded $S^r$-colored homology is also determined by the quiver matrix $C$, and it takes a universal form
\begin{align}
P_r(a,Q,t_r,t_c)=& \sum_{d_1+d_2+\ldots+d_m=r} \frac{(t_c^2;t_c^2)_r}{(t_c^2;t_c^2)_{d_1}(t_c^2;t_c^2)_{d_2} \cdots (t_c^2;t_c^2)_{d_m}} \times \nonumber \\
& \qquad \times a^{\sum_{i=1}^m a_i d_i}Q^{\sum_{i=1}^m q_i d_i} t_r^{\sum_{i=1}^m t_i d_i} t_c^{\sum_{i,j=1}^m {C_{i,j} d_i d_j} } . \label{quad-C}
\end{align}
\end{conjecture}
The generating function of such Poincar{\'e} polynomials, normalized by $(t_c^2;t_c^2)_r$,
\begin{equation}
P(x) = \sum_{r=0}^{\infty} P_r(a,Q,t_r,t_c) \frac{x^r}{(t_c^2;t_c^2)_r},
\end{equation}
can also be obtained as a specialization of (\ref{PQx}), with appropriate choice of $x_i$, which then gives rise to linear (in $d_i$) powers of $a,Q$ and $t_r$. Therefore the product decomposition (\ref{PQx-Omega}) leads to refined (quadruply-graded) LMOV invariants. Furthermore, (\ref{quad-C}) can be reduced to the generating function of colored superpolynomials upon the identification of variables given in (\ref{quad-super}), and the corresponding refined LMOV invariants can be identified (note that LMOV invariants, refined in the sense of including $t$-dependence, were also discussed in \cite{Garoufalidis:2015ewa,Kameyama:2017ryw}).
The expression (\ref{quad-C}) can also be reduced to the case of extremal powers of $a$ (i.e. top/bottom row). Let now $m$ denote the size of such a bottom or top row uncolored reduced homology, and denote $(q,t)$-degrees of its generators by $(q_i,t_i$). The Poincar{\'e} polynomial of the bottom (or top) row of the reduced quadruply-graded $S^r$-colored homology is then given by
\begin{equation}
P^{bottom/top}_r(Q,t_r,t_c)=\sum_{d_1+d_2+\ldots+d_m=r} Q^{\sum_{i=1}^m q_i d_i} t_r^{\sum_{i=1}^m t_i d_i} t_c^{\sum_{i,j=1}^m {C_{i,j} d_i d_j} } \frac{(t_c^2;t_c^2)_r}{(t_c^2;t_c^2)_{d_1} \cdots (t_c^2;t_c^2)_{d_m}}. \label{quad-extreme}
\end{equation}
In this case we suppressed the $a$-dependence, since the entire bottom (or top) row homology is characterized by the same $a$-degree. The matrix $C$ in (\ref{quad-extreme}) coincides with the one in (\ref{quivgen2}) and it encodes a subquiver (representing only the extremal $a$-dependence) of the full quiver associated to a given knot. Specializing the product decomposition (\ref{PQx-Omega}) results in refined (quadruply-graded), extremal LMOV invariants. Moreover (\ref{quad-extreme}) can be reduced to the generating function of (extremal) colored superpolynomials upon the identification of variables given in (\ref{quad-super}), which then encodes refined extremal LMOV invariants.
We stress that integrality of various refined (or quadruply-graded) LMOV invariants mentioned above follows automatically from the fact that the corresponding generating series arise as specializations of (\ref{PQx-Omega}), whose product decomposition is proved to give rise to integer invariants in general.
\subsection{Consequences: proof of the LMOV conjecture, new categorification, etc.}
Our conjectures imply that various knot invariants can be expressed in terms of invariants characterizing quiver moduli spaces. More generally, these conjectures imply that there are various -- unexpected, and highly nontrivial -- relations between knot theory and quiver representation theory; some of those relations are listed in table \ref{tab-duality}. We now briefly discuss some of these consequences, and we will illustrate them in various examples in the next section.
First, the fact that -- under appropriate specialization -- the motivic generating series of a quiver agrees with the generating function of colored HOMFLY-PT polynomials also means, that the product decomposition (\ref{PQx-Omega}) is identified with the product decomposition (\ref{Pr-LMOV}). This implies that the LMOV invariants $N_{r,i,j}$ take the form of linear combinations (with integer coefficients) of motivic Donaldson-Thomas invariants $\Omega_{{\bm d},j} = \Omega_{d_1,\ldots,d_m;j}$. The motivic Donaldson-Thomas invariants for symmetric quivers are proved to be integer \cite{efimov2012}, which therefore implies that the corresponding LMOV invariants are also integer -- which then proves the LMOV conjecture. Therefore, once a quiver corresponding to a given knot is identified (which we will do in many examples in the rest of the paper), it automatically follows that LMOV invariants for this knot, labeled by symmetric representations, are integer.
Second, quiver invariants automatically provide a refinement of knot invariants -- once a quiver is identified, its motivic generating series (\ref{PQx}) involves several generating parameters $x_1,\ldots,x_m$, encoding ``refined'' invariants $\Omega_{d_1,\ldots,d_m;j}$, and ``refined'' HOMFLY-PT polynomials. It is desirable to understand the meaning of those refined invariants from the knot theory perspective.
Third, we find that in some cases to a given knot one may assign several quivers, which give rise to the same generating function of HOMFLY-PT polynomials -- even though their original motivic generating series, without imposing the specialization (\ref{specialize}), are different. Such quivers differ by some permutation of their elements, as we will illustrate in various examples.
Moreover, the limit $q\to 1$ of the motivic generating series immediately implies integrality of classical LMOV invariants (\ref{bri}), which are expressed in terms of (integer) numerical Donaldson-Thomas invariants defined in (\ref{Omega-classical}).
Furthermore, the fact that (generating functions of) colored HOMFLY-PT polynomials and LMOV invariants are expressed in terms of motivic Donaldson-Thomas invariants -- which arise as certain Betti numbers of quiver moduli spaces -- provides a novel categorification of these knot invariants. Namely, quiver moduli spaces themselves can be regarded as new invariants of knots.
While, on one hand, knot invariants appear as specializations of invariants of quiver moduli spaces, on the other hand knot invariants (HOMFLY-PT polynomials, or LMOV invariants) can be defined in more general families, labeled by arbitrary Young diagrams (not just symmetric Young diagrams, which appear in (\ref{PxC})). It is desirable to understand how such more general invariants are related to, or could be extracted from, the data of quiver moduli spaces.
We also note that, as argued in \cite{Kontsevich:2010px}, the Cohomological Hall Algebra associated to a quiver should be identified with the algebra of BPS states \cite{Harvey:1996gc}. Furthermore, the generating functions (\ref{PxC}) take the form of products of $q$-series that appear in Nahm conjectures \cite{Nahm}, which suggests their relation to conformal field theories and integrability. All these issues are worth thorough further investigation.
\subsection{The strategy and $q$-identities}
In order to determine a quiver corresponding to a given knot, we have to rewrite the generating function of colored HOMFLY-PT polynomials of this knot in the form (\ref{PxC}). Colored HOMFLY-PT polynomials, which are known for various knots, can be written in terms of sums involving $q$-Pochhammer and $q$-binomial symbols \cite{superA,FGSS}, as e.g. in the expression (\ref{eq:Torus knots from FGS12}). However, in general in such expressions the number of summations is smaller than the number of terms in the superpolynomial, and it is not obvious that such sums can be rewritten in the form (\ref{PxC}), which involves as many summations as the number of terms in the superpolynomial. Therefore some algebraic manipulations are necessary in order to rewrite such formulas in the form that involves an appropriate number of additional summations, and in addition includes appropriate $q$-Pochhammer symbols in the denominator. To achieve this we take advantage of the following lemmas.
\begin{lemma}\label{lem1}
For any $d_1,\ldots,d_k\ge 0$, we have:
\begin{align}
\frac{(x;q)_{d_1+\ldots+d_k}}{(q;q)_{d_1}\cdots (q;q)_{d_k}}
=& \sum\limits_{\alpha_1+\beta_1=d_1}\sum\limits_{\alpha_2+\beta_2=d_2}\cdots\sum\limits_{\alpha_k+\beta_k=d_k}
\frac{1}{(q;q)_{\alpha_1}\cdots(q;q)_{\alpha_k}(q;q)_{\beta_1}\cdots(q;q)_{\beta_k}} \times \nonumber \\
& \times (-x)^{{\alpha_1+\ldots+\alpha_k}} q^{\frac{1}{2}(\alpha_1^2+\ldots+\alpha_k^2)}q^{\sum_{i=1}^{k-1} \alpha_{i+1} (d_1+\ldots+d_i)}q^{-\frac{1}{2}(\alpha_1+\ldots+\alpha_k)}. \label{lem1-eq}
\end{align}
\end{lemma}
\textbf{Proof:}\\
First, note that
\begin{equation}
(x;q)_{d_1+\ldots+d_k}=(x;q)_{d_1}(xq^{d_1};q)_{d_2}(xq^{d_1+d_2};q)_{d_3}\ldots(xq^{d_1+\ldots+d_{k-1}};q)_{d_k}.
\end{equation}
By expanding each of $k$ $q$-Pochhammers on the right hand side by using the quantum binomial identity
\begin{equation}
(x;q)_n= \sum\limits_{\alpha=0}^n (-x)^{\alpha} q^{\frac{1}{2}\alpha(\alpha-1)} \frac{(q;q)_{n}}{(q;q)_{\alpha} (q;q)_{n-\alpha}}
= \sum\limits_{\alpha+\beta=n} (-x)^{\alpha} q^{\frac{1}{2}\alpha(\alpha-1)}\frac{(q;q)_{n}}{(q;q)_{\alpha}(q;q)_{\beta}}, \label{qpoch-sum}
\end{equation}
we obtain
\begin{align}(x;q)_{d_1+\ldots+d_k}=
&\sum\limits_{\alpha_1+\beta_1=d_1}
\sum\limits_{\alpha_2+\beta_2=d_2}
\cdots
\sum\limits_{\alpha_k+\beta_k=d_k}
\frac{(q;q)_{d_1}}{(q;q)_{\alpha_1}(q;q)_{\beta_1}}\frac{(q;q)_{d_2}}{(q;q)_{\alpha_2}(q;q)_{\beta_2}}\cdots\frac{(q;q)_{d_k}}{(q;q)_{\alpha_k}(q;q)_{\beta_k}}
\nonumber \\
& \times (-x)^{\alpha_1+\ldots+\alpha_k}
q^{{\frac{1}{2}}(\alpha_1^2+\ldots+\alpha_k^2-\alpha_1-\ldots-\alpha_k)}
q^{\sum_{i=1}^{k-1}{ \alpha_{i+1} (d_1+\ldots+d_i)}}
\end{align}
which proves the lemma. \hfill\rule{6pt}{6pt}
Lemma \ref{lem1} enables rewriting the expression of the form
\begin{equation}
\frac{(x;q)_{d_1+\ldots+d_k}}{(q;q)_{d_1}\cdots (q;q)_{d_k}}
\end{equation}
as a sum of terms
\begin{equation}
\frac{1}{(q;q)_{\alpha_1}\cdots(q;q)_{\alpha_k}(q;q)_{\beta_1}\cdots(q;q)_{\beta_k}}
\end{equation}
weighted simply by linear and quadratic powers of $q$.
In this way (at least in some cases) we can introduce additional summations in expressions for colored HOMFLY-PT polynomials, in order to bring them into the form of the quiver generating series (\ref{PxC}). In more complicated situations, generalizing the relation
\begin{equation}\label{pomoc1}
(x;q)_{a+b}=(x;q)_a(xq^a;q)_b,
\end{equation}
we can take advantage of the following lemma which enables rewriting certain $q$-binomial coefficients. Recall that throughout the paper we are using the convention ${n\brack k}=\frac{(q^2;q^2)_n}{(q^2;q^2)_k(q^2;q^2)_{n-k}}$, cf. (\ref{q-binomial}).
\begin{lemma}\label{lem2}
For nonnegative integers $a,b$ and $k$ we have
\begin{equation}\label{prva}
{ a+b \brack k}=\sum_{i+j=k} q^{2(a-i)(k-i)} {a\brack i} {b\brack j}.
\end{equation}
More generally, let $a_1,\ldots,a_m$, $m\ge 1$, and $k_1,\ldots,k_p$, $p\ge 1$, be nonnegative integers. Then
\begin{align}
{a_1+a_2\brack k_1} {k_1\brack k_2} \cdots {k_{p-1}\brack k_p} =& \sum_{i_1+j_1=k_1}\sum_{i_2+j_2=k_2}\cdots
\sum_{i_p+j_p=k_p} {a_1\brack i_1} {i_1\brack i_2}\cdots { i_{p-1}\brack i_p} \, {a_2\brack j_1 }{ j_1\brack j_2 } \cdots { j_{p-1}\brack j_p} \nonumber \\
& \times q^{2\left((a_1-i_1)(k_1-i_1)+(i_1-i_2)(k_2-i_2)+\ldots+(i_{p-1}-i_p)(k_p-i_p)\right)} ,
\end{align}
\begin{align}
{ a_1+a_2+\ldots+a_m\brack k}=& \sum_{i_1+i_2+\ldots+i_m=k} { a_1\brack i_1 }{ a_2\brack i_2 }\cdots { a_{m}\brack i_m } \nonumber \\
& \times q^{2\left((a_1-i_1)(k-i_1)+(a_2-i_2)(k-i_1-i_2)+\ldots+(a_m-i_m)(k-i_1-i_2-\ldots-i_m)\right)}.
\end{align}
Furthermore
\begin{align}
&{ a_1+a_2+\ldots+a_m\brack k_1 } {k_1\brack k_2} \cdots { k_{p-1}\brack k_p }= \nonumber \\
& = \sum_{i^1_1+\ldots+i^1_m=k_1} \sum_{i^2_1+\ldots+i^2_m=k_2}\cdots \sum_{i^p_1+\ldots+i^p_m=k_p}
q^{X(\underline{a},\underline{i^1},k_1)+X(\underline{i^1},\underline{i^2},k_2)+\ldots+X(\underline{i^{p-1}},\underline{i^p},k_p)} \\
&\qquad\quad\times { a_1\brack i^1_1 }{ i^1_1\brack i^1_2 } \cdots { i^1_{p-1}\brack i^1_p } \ { a_2\brack i^2_1 }{ i^2_1\brack i^2_2} \cdots { i^2_{p-1}\brack i^2_p }\cdots {a_m\brack i^m_1 }{ i^m_1\brack i^m_2 }\cdots
{ i^m_{p-1}\brack i^m_p } \nonumber
\end{align}
where
$$
X(\underline{a},\underline{i},k)={2\left((a_1-i_1)(k-i_1)+(a_2-i_2)(k-i_1-i_2)+\ldots+(a_m-i_m)(k-i_1-i_2-\ldots-i_m)\right)},
$$
for sequences $\underline{a}=(a_1,\ldots,a_m)$ and $\underline{i}=(i_1,\ldots,i_m)$.
\end{lemma}
\textbf{Proof:}
These relations follow after straightforward calculations. For example, from the $q$-Pochhammer relation (\ref{pomoc1}) we get
\begin{equation}
\sum_{k=0}^{a+b}(-x)^k q^{k^2-k} { a+b\brack k } = \sum_{i=0}^a (-x)^i q^{i^2-i} { a\brack i } \sum_{j=0}^b (-x)^j q^{2aj} q^{j^2-j} { b\brack j }.
\end{equation}
By matching the powers of $x$ on both sides of this equation we get the relation (\ref{prva}). The remaining three equalities can now be obtained by induction, using (\ref{prva}). \hfill\rule{6pt}{6pt}
One can use the above lemma for example after rewriting the following product of binomial coefficients in terms of the $q$-Pochhammer symbols
\begin{equation}
{ a\brack k_1}{ k_1\brack k_2} \cdots { k_{p-1}\brack k_p}=\frac{(q^2;q^2)_a}{(q^2;q^2)_{a-k_1}(q^2;q^2)_{k_1-k_2}\cdots(q^2;q^2)_{k_{p-1}-k_p}(q^2;q^2)_{k_p}}.
\end{equation}
The right hand side is the quotient of the $q$-Pochhammer of length $a$ by the product of $p$ $q$-Pochhammers, whose lengths sum up also to $a$. As we explained above, this can be further transformed into a sum involving $q$-Pochhammers only in the denominator, weighted only by linear and quadratic powers of $q$, which is then of the required quiver form (\ref{PxC}).
\subsection{Reduced vs. unreduced invariants} \label{ssec-unreduced}
The values of parameters $l_i, a_i$ and $C_{i,i}$ in (\ref{PxC}) depend on the choice of normalization of $\overline{P}_r(a,q)$. The values mentioned in Conjecture \ref{conj2} arise when the normalization includes only the denominator $(q^2;q^2)_r$ of the colored HOMFLY-PT polynomial of the unknot, i.e.
\begin{equation}
\overline{P}_r(a,q)=\frac{P_r(a,q) }{ (q^2;q^2)_r}. \label{barPr-qq}
\end{equation}
In this case the values of $l_i, a_i$ and $C_{i,i}$ are related to reduced and uncolored HOMFLY-PT homology and superpolynomial. On the other hand, a more familiar normalization that involves the full unknot polynomial (\ref{Pbar-unknot})
\begin{equation}
\overline{P}_r(a,q)=a^{-r}q^r \frac{(a^2;q^2)_r}{(q^2;q^2)_r} P_r(a,q) \label{barPr-unknot}
\end{equation}
leads to a twice larger quiver, which encodes information about unreduced HOMFLY-PT homology, whose Poincar{\'e} polynomial is obtained by multiplying the (reduced) superpolynomial by $a^{-1}q(1+a^2 t)$. Suppose that colored polynomials normalized as in (\ref{barPr-qq}) lead to a quiver encoded in a matrix $C$. Multiplying (\ref{barPr-qq}) by an additional factor $a^{-r}q^r (a^2;q^2)_r$, we can use (\ref{lem1-eq}) to deal with the additional $q$-Pochhammer $(a^2;q^2)_r$. Introducing new summation variables $\alpha_i$ and $\beta_i$, such that $d_i = \alpha_i + \beta_i$, the expression (\ref{PxC}) is replaced by another summation which is also of the required form (\ref{PxC}), which however involves summations over $\alpha_i$ and $\beta_i$ with summands involving the following factors of $q$ in quadratic powers of summation variables
\begin{equation}
q^{\sum_{i,j} C_{i,j}(\alpha_i+\beta_i)(\alpha_j+\beta_j)} q^{\alpha_1^2 +\ldots \alpha_m^2} q^{2\sum_{i=1}^{m-1} \alpha_{i+1}(d_1 + \ldots d_i) } .
\end{equation}
The exponent of $q$ in this expression can be rewritten as
\begin{equation}
\sum_{i,j} C_{i,j} \beta_i \beta_j + \sum_{i,j} (C_{i,j}+1) \alpha_i \alpha_j + 2\sum_{i\leq j} C_{i,j} \alpha_i \beta_j + 2 \sum_{i>j} (C_{i,j}+1) \alpha_i \beta_j. \label{C-twice}
\end{equation}
This expression also encodes a quiver, which is however twice larger than $C$, and which decomposes into two parts: one which looks like the original quiver encoded in $C$ (determined by the first term $\sum_{i,j} C_{i,j} \beta_i \beta_j$), and another one which looks like the original quiver framed by 1 (as determined by the second term $\sum_{i,j} (C_{i,j}+1) \alpha_i \alpha_j$). These two subquivers are connected by arrows, whose structure is given by the last two summations in (\ref{C-twice}).
\section{Case studies} \label{sec-results}
In this section we illustrate our claims and conjectures in various examples. We show how to rewrite generating functions of known colored HOMFLY-PT polynomials in the form (\ref{PxC}) and identify corresponding quivers. In particular this automatically proves the LMOV conjecture (for symmetric representations) for the knots under consideration. Furthermore, assuming that generating functions of colored HOMFLY-PT polynomials should be of the form (\ref{PxC}), we derive previously unknown formulas for such polynomials for $6_2$ and $6_3$ knots, as well as for $(3,7)$ torus knot.
\subsection{Unknot}
The (unreduced) colored HOMFLY-PT polynomial for the unknot takes the form
\begin{equation}
\overline{P}_r(a,q) = a^{-r}q^r \frac{(a^2;q^2)_r}{(q^2;q^2)_r}. \label{Pbar-unknot}
\end{equation}
First, consider just the denominator of this expression, which includes a single $q$-Pochhammer. This is equivalent to the simpler (reduced) normalization discussed in section \ref{ssec-unreduced}, and up to the $q^r$ factor it agrees with the extremal (bottom row) unknot HOMFLY-PT polynomial. More generally, the generating series of the reduced colored HOMFLY-PT polynomials of the $f$-framed unknot takes the form
\begin{equation}
P(x) = \sum_{r=0}^\infty x^r \frac{q^{f(r^2-r)}}{(q^2;q^2)_r}, \label{Px-unknot}
\end{equation}
which essentially agrees with the motivic generating series associated to a quiver consisting of one vertex and $f$ loops, shown in figure \ref{fig-m-loop}. These are prototype and important examples of quivers, and properties of their moduli spaces were discussed in \cite{COM:8276935,Rei12}. The relation of this family of quivers to LMOV invariants of framed unknot (equivalently extremal invariants of twist knots, or open topological string amplitudes for branes in $\mathbb{C}^3$ geometry) was presented in \cite{Kucharski:2016rlb}, and discussed also in \cite{Luo:2016oza,Zhu:2017lsn}.
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.35\textwidth]{quiver-m-loop.png}
\caption{Quiver with one vertex and $f$ loops, encoding extremal framed unknot invariants (equivalently open topological string amplitudes for branes in $\mathbb{C}^3$ geometry).} \label{fig-m-loop}
\end{center}
\end{figure}
Consider now the generating function of the full unknot invariants (\ref{Pbar-unknot}) -- or equivalently open topological string amplitudes for branes in the resolved conifold geometry. Using (\ref{qpoch-sum}), this generating function can be rewritten as
\begin{align}
P(x) &= \sum_{r=0}^{\infty} x^r a^{-r}q^r \frac{(a^2;q^2)_r}{(q^2;q^2)_r} = \sum_{d_1,d_2=0}^{\infty} x^{d_1+d_2} \frac{(-1)^{d_1} a^{d_1-d_2} q^{d_1^2+d_2} }{(q^2;q^2)_{d_1} (q^2;q^2)_{d_2}} = \nonumber \\
& = \Big( \sum_{d_1=0}^{\infty} x^{d_1} \frac{(-1)^{d_1} a^{d_1} q^{d_1^2} }{(q^2;q^2)_{d_1}} \Big)
\Big( \sum_{d_2=0}^{\infty} x^{d_2} \frac{a^{-d_2} q^{d_2} }{(q^2;q^2)_{d_2}} \Big) = \frac{(xaq;q^2)_{\infty}}{(xq/a;q^2)_{\infty}}. \label{Px-unknot-a}
\end{align}
From the expression in the first line, or simply taking advantage of (\ref{C-twice}), we find that the corresponding quiver can be interpreted as a twice larger quiver associated to (\ref{Px-unknot}); this larger quiver consists of two disconnected vertices, with a single loop associated to one vertex (labeled by $d_1$). The final factorization into the ratio of two quantum dilogarithms means that there are only two non-zero LMOV (or motivic Donaldson-Thomas) invariants, which is a well known statement for the unknot \cite{OoguriV,Kucharski:2016rlb}.
More generally, including the framing dependence (\ref{framing}) in (\ref{Px-unknot-a}) results in a quiver with additional loops and arrows, as in (\ref{C-frame}). Contrary to the unframed case (\ref{Px-unknot-a}), such expressions do not factorize into a finite number of quantum dilogarithms, and they would encode an infinite number of LMOV invariants.
\subsection{Trefoil and cinquefoil knots} \label{ssec-3151}
We now illustrate how to identify a quiver corresponding to a knot in the example of the trefoil knot, i.e. the $(2,3)$ torus knot (also denoted $T_{2,3}$ or $3_1$), whose reduced colored HOMFLY-PT polynomials arise by setting $t=-1$ in (\ref{Pr-31})
\begin{equation}
P_r(a,q)= \frac{a^{2r}}{q^{2r}}\sum_{k=0}^r {r \brack k} q^{2k(r+1)} \prod_{i=1}^k (1-a^2q^{2(i-2)}), \label{Pr-31-homfly}
\end{equation}
with the $q$-binomial ${r\brack k}$ given in (\ref{q-binomial}). Using (\ref{qpoch-sum}), the $q$-binomial together with the last product in (\ref{Pr-31-homfly}) take the form
$$
{r \brack k} \big(\frac{a^2}{q^2};q^2\big)_k = \sum_{i=0}^k \frac{(q^2;q^2)_r \big( - \frac{a^2}{q^2} \big)^i q^{i(i-1)} }{(q^2;q^2)_{r-k} (q^2;q^2)_i (q^2;q^2)_{k-i}}.
$$
Introducing
\begin{equation}
r=d_1+d_2+d_3, \quad k=d_2+d_3, \quad i=d_3,
\end{equation}
with $d_i\geq 0$, and normalizing $P_r(a,q)$ by $(q^2;q^2)_r$, the generating function (\ref{Pz2}) takes the form
\begin{equation}
P(x) = \sum_{r=0}^{\infty} \frac{P_r(a,q)}{(q^2;q^2)_r} x^r = \sum_{d_1,d_2,d_3\geq 0} \frac{q^{\sum_{i,j} C^{T_{2,3}}_{i,j} d_i d_j -2d_1-3d_3} (-1)^{d_3} a^{2d_1+2d_2+4d_3} }{(q^2;q^2)_{d_1} (q^2;q^2)_{d_2} (q^2;q^2)_{d_3}} x^{d_1+d_2+d_3}, \label{PxC-31}
\end{equation}
where
\arraycolsep 3pt
\begin{equation}
C^{T_{2,3}} = \left[
\begin{array}{ccc}
0 & 1 & 1 \\
1 & 2 & 2 \\
1 & 2 & 3 \\
\end{array}
\right] \label{trefoil-quiver}
\end{equation}
The expression (\ref{PxC-31}) is indeed of the form (\ref{PxC}), with the corresponding quiver shown in figure \ref{fig-trefoil}. Vertices of this quiver, as stated in the previous section, correspond to generators of HOMFLY-PT homology. The diagonal elements $(0,2,3)$ of the matrix $C$ (representing numbers of loops at vertices of the quiver) indeed agree with homological degrees encoded in the uncolored superpolynomial (\ref{P1-31}), coefficients $l_i=-2,0,3$ of linear terms in $d_i$ in the power of $q$ in (\ref{PxC-31}) are given by $l_i=q_i-t_i$, coefficients $a_i=2,2,4$ in the power of $a$ agree with $a$-degrees of generators of HOMFLY-PT homology, and the additional minus sign $(-1)^{d_3}$ is determined by just one generator with odd $t$-degree $t_3=3$ (which is manifest in figure \ref{diagrams}).
Let us also discuss the normalization of the colored HOMFLY-PT polynomials by the full unknot invariant, following section \ref{ssec-unreduced}. Multiplying each summand in (\ref{PxC-31}) proportional to $x^r$ by $a^{-r}q^r (a^2;q^2)_r$ and taking advantage of (\ref{lem1-eq}), we get the generating function
\begin{align}
P(x) = & \sum_{r=0}^{\infty} x^r \sum\limits_{\alpha_1+\alpha_2+\alpha_3+\beta_1+\beta_2+\beta_3=r}
\frac{ q^{\alpha_1^2+\alpha_2^2+\alpha_3^2}q^{\alpha_2 d_1+\alpha_{3} (d_1+d_2)}q^{-\alpha_1-\alpha_2-\alpha_3} }{(q^2;q^2)_{\alpha_1}
(q^2;q^2)_{\alpha_2}(q^2;q^2)_{\alpha_3}(q^2;q^2)_{\beta_1}(q^2;q^2)_{\beta_2}(q^2;q^2)_{\beta_3}} \nonumber \\
&\quad \times (-1)^{d_3}a^{d_1+d_2+3d_3} q^{-d_1+d_2-2d_3}q^{2d_2^2+3d_3^2+2(d_1d_2+d_1d_3+2d_2d_3)} (-a^2 )^{\alpha_1+\alpha_2+\alpha_3}, \label{PxC-31-bis}
\end{align}
where $d_i=\alpha_i+\beta_i$, $i=1,2,3$. The form of the corresponding quiver can be read off from powers of $q$ in this generating function, or simply from the transformation (\ref{C-twice}) applied to the quiver (\ref{trefoil-quiver}). Ultimately we find a quiver with 6 nodes, whose structure, in the basis ordered as $(\beta_1,\alpha_1,\beta_2,\alpha_2,\beta_3,\alpha_3)$, is encoded in the matrix of the form
\begin{equation}
C^{T_{2,3}}_{unreduced} =
\left[
\begin{array}{cccccc}
0\ & 0\ & 1\ & 2\ & 1\ & 2 \\
0\ & 1\ & 1\ & 2\ & 1\ & 2 \\
1\ & 1\ & 2\ & 2\ & 2\ & 3 \\
2\ & 2\ & 2\ & 3\ & 2\ & 3 \\
1\ & 1\ & 2\ & 2\ & 3\ & 3 \\
2\ & 2\ & 3\ & 3\ & 3\ & 4 \\
\end{array} \label{C-trefoil-bis}
\right]
\end{equation}
As expected, the information about unreduced HOMFLY-PT homology is encoded in this quiver and the expression (\ref{PxC-31-bis}). For the trefoil this homology has 6 generators (obtained by multiplying the reduced superpolynomial by $a^{-1}q(1+a^2 t)$), with $t$-degrees $0, 1, 2, 3, 3, 4$, which indeed appear as diagonal elements in (\ref{C-trefoil-bis}). More generally, $(q,t)$-degrees of these six generators are $(-1,0),(-1,1),(3,2),(3,3),(1,3),(1,4)$, and the differences $q_i-t_i$ in (\ref{linear}) also match the coefficients of the linear term in the power of $q$ in (\ref{PxC-31-bis}), which are of the form $-\beta_1-2\alpha_1+\beta_2-2\beta_3-3\alpha_3$.
Let us consider a more involved example of the cinquefoil $(2,5)$ torus knot (also denoted $T_{2,5}$ or $5_1$). Its colored HOMFLY-PT polynomials are obtained as the $p=2$ case of (\ref{eq:Torus knots from FGS12}) and their generating function, normalized by $(q^2;q^2)_r$, takes the form
\begin{equation}
P(x)= \sum_{r=0}^{\infty} \frac{x^{r}a^{4r}q^{-4r}}{(q^2;q^2)_r} \sum_{0\leq k_{2}\leq k_{1}\leq r}
{r \brack k_1} {k_1 \brack k_2} q^{2(2r+1)(k_{1}+k_{2})-2rk_{1}-2k_{1}k_{2}}(a^{2}q^{-2};q^{2})_{k_{1}}. \label{eq:5_1 general formula}
\end{equation}
Rewriting the last $q$-Pochhammer symbol in this expression using (\ref{qpoch-sum}), and then taking advantage of (\ref{prva}), we get
\begin{align}
P(x)= & \sum_{r=0}^{\infty} \frac{x^{r} }{(q^2;q^2)_r} \sum_{0\leq k_{2}\leq k_{1}\leq r}\sum_{0\leq\alpha_{1}\leq k_{1}}
{r \brack k_1}{k_1 \brack k_2} {k_1 \brack \alpha_1}
\times\nonumber \\
& \times (-1)^{\alpha_1}a^{2\alpha_1+4r} q^{2(2r+1)(k_{1}+k_{2})-2rk_{1}-2k_{1}k_{2} + \alpha_{1}^{2}-3\alpha_{1} -4r} = \nonumber \\
= & \sum_{r=0}^{\infty} \frac{x^{r} }{(q^2;q^2)_r} \sum_{0\leq k_{2}\leq k_{1}\leq r}\sum_{0\leq\alpha_{1}\leq k_{1}}\sum_{0\leq\alpha_{2}\leq k_{2}}{r \brack k_1} {k_1 \brack \alpha_1} {k_{1}-\alpha_{1} \brack k_{2}-\alpha_{2}} {\alpha_{1} \brack \alpha_{2}}
\times\nonumber \\
& \times(-1)^{\alpha_{1}} a^{4r+2\alpha_{1}}q^{-4r+2(2r+1)(k_{1}+k_{2})-2rk_{1}-2k_{1}k_{2}+\alpha_{1}^{2}-3\alpha_{1}+2(\alpha_{1}-\alpha_{2})(k_{2}-\alpha_{2})}\label{eq:5_1 generating function}
\end{align}
(with the condition $\alpha_{2}\leq\alpha_{1}, k_{2}-\alpha_{2}\leq k_{1}-\alpha_{1}$). After the change of variables
\begin{equation}
d_{1} =r-k_{1}, \quad d_{2} =k_{1}-\alpha_{1}-(k_{2}-\alpha_{2}), \quad d_{3} =\alpha_{1}-\alpha_{2},
\quad d_{4} =k_{2}-\alpha_{2},\quad d_{5} =\alpha_{2},
\end{equation}
we finally get
\begin{align}
P(x)= & \sum_{d_{1},d_{2},\ldots,d_{5}\geq0}q^{\sum_{i,j}C_{i,j}^{T^{2,5}}d_{i}d_{j}} \frac{x^{d_{1}+d_{2}+\cdots+d_{5}}}{(q^{2};q^{2})_{d_{1}}(q^{2};q^{2})_{d_{2}}\cdots(q^{2};q^{2})_{d_{5}}}\times\label{eq:5_1 quiver version}\\
& \times(-1)^{d_{3}+d_{5}}a^{4d_{1}+4d_{2}+6d_{3}+4d_{4}+6d_{5}}q^{-4d_{1}-2d_{2}-5d_{3}-3d_{5}},\nonumber
\end{align}
where the matrix
\begin{equation}
C^{T_{2,5}}=\left[\begin{array}{ccccc}
0\ & 1\ & 1\ & 3\ & 3 \\
1\ & 2\ & 2\ & 3\ & 3 \\
1\ & 2\ & 3\ & 4\ & 4 \\
3\ & 3\ & 4\ & 4\ & 4 \\
3\ & 3\ & 4\ & 4\ & 5
\end{array}\right] \label{51-quiver}
\end{equation}
represents the quiver corresponding to the $(2,5)$ torus knot.
\subsection{$(2,2p+1)$ torus knots} \label{ssec-torus22p1}
Colored HOMFLY-PT polynomials for $(2,2p+1)$ torus knots (also denoted $T_{2,2p+1}$) can be obtained as the $t=-1$ specialization of the following colored superpolynomials, determined in \cite{FGS,FGSS}
\begin{align}
P_{S^{r}}^{T_{2,2p+1}}(a,q,t)= & a^{2pr}q^{-2pr}\sum_{0\leq k_{p}\leq\ldots\leq k_{2}\leq k_{1}\leq r} {r \brack k_1} {k_1 \brack k_2} \cdots { k_{p-1}\brack k_{p}} \times\nonumber \\
& \times q^{2\sum_{i=1}^{p}((2r+1)k_{i}-k_{i-1}k_{i})}t^{2(k_{1}+k_{2}+\ldots+k_{p})}\prod_{i=1}^{k_{1}}\big(1+a^{2}q^{2(i-2)}t\big). \label{eq:Torus knots from FGS12}
\end{align}
These expressions can be transformed to the form (\ref{PxC}) recursively, generalizing the step between trefoil and cinquefoil knots presented in the previous section. In order to determine the form of the generating function (\ref{PxC}) and the quiver for arbitrary $(2,2p+1)$ torus knot, we analyze first the following three modifications in the expression for colored HOMFLY-PT polynomials, when $p$ is changed to $p+1$
\begin{align}
a^{2pr}q^{-2pr}\ &\mapsto\ a^{2(p+1)r}q^{-2(p+1)r} \label{mod1} \\
\sum_{0\leq k_{p}\leq\ldots\leq k_{1}\leq r} {r\brack k_{1}}
\cdots {k_{p-1}\brack k_{p}}\ &\mapsto\ \sum_{0\leq k_{p+1}\leq\ldots\leq k_{1}\leq r}{ r\brack k_{1}}
\cdots {k_{p}\brack k_{p+1}} \label{mod2} \\
q^{2\sum_{i=1}^{p}\left[(2r+1)k_{i}-k_{i-1}k_{i}\right]}\ &\mapsto\ q^{2\sum_{i=1}^{p}\left[(2r+1)k_{i}-k_{i-1}k_{i}\right]+2(2r+1)k_{p+1}-2k_{p}k_{p+1}} \label{mod3}
\end{align}
These transformations generalize the relation between trefoil and cinquefoil knots, which we discussed in section \ref{ssec-3151}, and which corresponds to changing $p=1$ to $p=2$.
The first modification (\ref{mod1}) only affects the change of variables leading to the generating function of the quiver, but not the form of the quiver.
In the second transformation (\ref{mod2}) a new variable $k_{p+1}$ and an additional $q$-binomial ${k_{p}\brack k_{p+1}}$ are introduced. Let us discuss first the special $p=1$ case of trefoil and cinquefoil knots. As already analyzed above, in this case, in the generating series for the cinquefoil knot, we split ${k_{p}\brack k_{p+1}}\equiv {k_{1}\brack k_{2}}$ into ${k_{p}-\alpha_{p}\brack k_{p+1}-\alpha_{p+1}} {\alpha_{p}\brack \alpha_{p+1}}$ and changed variables accordingly
\begin{align}
d_{1} & =r-k_{1} & & & d_{1} & =r-k_{1}\nonumber \\
d_{2} & =k_{1}-\alpha_{1} & \longmapsto & & d_{2} & =k_{1}-\alpha_{1}-(k_{2}-\alpha_{2})\nonumber \\
d_{3} & =\alpha_{1} & & & d_{3} & =\alpha_{1}-\alpha_{2}\\
& & & & d_{4} & =k_{2}-\alpha_{2}\nonumber \\
& & & & d_{5} & =\alpha_{2}\nonumber
\end{align}
This is equivalent to the following modification of summation variables in the quiver generating series
\begin{align}
r & =d_{1}+d_{2}+d_{3} & & & r & =d_{1}+(d_{2}+d_{4})+(d_{3}+d_{5})\nonumber \\
k_{1} & =d_{2}+d_{3} & \longmapsto & & k_{1} & =(d_{2}+d_{4})+(d_{3}+d_{5})\label{eq:change of vars 5_1}\\
\alpha_{1} & =d_{3} & & & \alpha_{1} & =(d_{3}+d_{5})\nonumber
\end{align}
which means that the matrix representing the cinquefoil quiver is obtained from the one for the trefoil quiver by copying the first and the second column and row, respectively, into the third and the fourth column and row. In addition, changing ${k_{1}\brack k_{2}}$ into ${k_{p}-\alpha_{p}\brack k_{p+1}-\alpha_{p+1}} {\alpha_{p}\brack \alpha_{p+1}}$ introduces a new term in $\sum_{i,j}C_{i,j}d_{i}d_{j}$ of the form $2(\alpha_{1}-\alpha_{2})(k_{2}-\alpha_{2})=2d_{3}d_{4}$, which means that the matrix elements $C_{3,4}$ and $C_{4,3}$ are increased by $1$.
Generalizing the above transformation and splitting ${k_{p}\brack k_{p+1}}$ into ${k_{p}-\alpha_{p}\brack k_{p+1}-\alpha_{p+1}} {\alpha_{p}\brack \alpha_{p+1}}$ for arbitrary $p$, the relation (\ref{eq:change of vars 5_1}) is replaced by
\begin{align}
r & =d_{1}+d_{2}\ldots+d_{2p}+d_{2p+1} & & & r & =d_{1}+d_{2}\ldots+(d_{2p}+d_{2p+2})+(d_{2p+1}+d_{2p+3})\nonumber \\
k_{1} & =d_{2}+\ldots+d_{2p}+d_{2p+1} & & & k_{1} & =d_{2}+\ldots+(d_{2p}+d_{2p+2})+(d_{2p+1}+d_{2p+3})\nonumber \\
k_{2} & =d_{4}+\ldots+d_{2p}+d_{2p+1} & & & k_{2} & =d_{4}+\ldots+(d_{2p}+d_{2p+2})+(d_{2p+1}+d_{2p+3})\nonumber \\
& \vdots & & & & \vdots\nonumber \\
k_{p} & =d_{2p}+d_{2p+1} & \longmapsto & & k_{p} & =(d_{2p}+d_{2p+2})+(d_{2p+1}+d_{2p+3})\nonumber \\
\alpha_{1} & =d_{3}+d_{5}+\ldots+d_{2p+1} & & & \alpha_{1} & =d_{3}+d_{5}+\ldots+(d_{2p+1}+d_{2p+3})\nonumber \\
\alpha_{2} & =d_{5}+\ldots+d_{2p+1} & & & \alpha_{2} & =d_{5}+\ldots+(d_{2p+1}+d_{2p+3})\nonumber \\
& \vdots & & & & \vdots\label{eq:change of vars general}\\
\alpha_{p} & =d_{2p+1} & & & \alpha_{p} & =(d_{2p+1}+d_{2p+3})\nonumber
\end{align}
so that columns and rows of number $2p$ and $2p+1$ are copied respectively to those of number $2p+2$ and $2p+3$, and matrix elements $C_{2p+1,2p+2}$ and $C_{2p+2,2p+1}$ are increased by $1$.
Finally, the third transformation (\ref{mod3}) modifies the change of variables and adds $4rk_{p+1}-2k_{p}k_{p+1}$
to the sum $\sum{}_{i,j}C_{i,j}d_{i}d_{j}$. In the special case of $p=1$ we have
\begin{align}
4rk_{2}-2k_{1}k_{2}&= 4(r-k_{1})k_{2}+2(k_{1}-k_{2})k_{2}+2k_{2}^{2} = \nonumber \\
&= 4d_{1}\left(d_{4}+d_{5}\right)+2\left(d_{2}+d_{3}\right)\left(d_{4}+d_{5}\right)+2\left(d_{4}+d_{5}\right)^{2},
\end{align}
which means that $C_{1,4}, C_{1,5}, C_{4,5}$ (and transposed matrix elements)
and $C_{4,4}$, $C_{5,5}$ increase by $2$, and $C_{2,4}, C_{2,5}, C_{3,4}$ and $C_{3,5}$ (and transposed elements) increase by $1$. For general $p$
\begin{align}
4rk_{p+1}-2k_{p}k_{p+1} &= 4(r-k_{p})k_{p+1}+2(k_{p}-k_{p+1})k_{p+1}+2k_{p+1}^{2} =\nonumber \\
&= 4\left(d_{1}+\ldots+d_{2p-1}\right)\left(d_{2p+2}+d_{2p+3}\right) + \\
&\quad +2\left(d_{2p}+d_{2p+1}\right)\left(d_{2p+2}+d_{2p+3}\right)+2\left(d_{2p+2}+d_{2p+3}\right)^{2},\nonumber
\end{align}
which means increasing
$C_{1,2p+2}, \ldots, C_{2p-1,2p+2}$, $C_{1,2p+3}, \ldots,
C_{2p-1,2p+3}$, $C_{2p+2,2p+3}$ (and transposed elements) and $C_{2p+2,2p+2}$,
$C_{2p+3,2p+3}$ by $2$, as well as increasing $C_{2p,2p+2}$, $C_{2p+1,2p+2}$,
$C_{2p,2p+3}$, $C_{2p+1,2p+3}$ (and transposed elements) by $1$.
To sum up, once we know a matrix $C^{T_{2,2p+1}}$ representing a quiver for the $(2,2p+1)$ torus knot, the matrix $C^{T_{2,2p+3}}$ for a quiver associated to the $(2,2p+3)$ torus knot is obtained by copying columns and rows of the number $2p$ and $2p+1$ to $2p+2$ and $2p+3$ respectively, and increasing elements in the last two columns (and rows) by $2$, except for $C_{2p,2p+2}$, $C_{2p,2p+3}$, and $C_{2p+1,2p+3}$ (and transposed elements) that are increased by $1$. The solution of this recursion, for an arbitrary $(2,2p+1)$ torus knot, takes the form
\begin{equation}
C^{T_{2,2p+1}}=\left[\begin{array}{ccccccc}
F_{0} & F_{1} & F_{2} & F_{3} & \cdots & F_{p-1} & F_{p}\\
F_{1}^{T} & D_{1} & U_{2} & U_{3} & \cdots & U_{p-1} & U_{p}\\
F_{2}^{T} & U_{2}^{T} & D_{2} & U_{3} & \cdots & U_{p-1} & U_{p}\\
F_{3}^{T} & U_{3}^{T} & U_{3}^{T} & D_{3} & \cdots & U_{p-1} & U_{p}\\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots\\
F_{p-1}^{T} & U_{p-1}^{T} & U_{p-1}^{T} & U_{p-1}^{T} & \cdots & D_{p-1} & U_{p}\\
F_{p}^{T} & U_{p}^{T} & U_{p}^{T} & U_{p}^{T} & \cdots & U_{p}^{T} & D_{p}
\end{array}\right] \label{C-torus22p1}
\end{equation}
with the following block entries
\begin{equation}
F_{0}=\left[0\right],\qquad
F_{k}=\left[\begin{array}{cc}
2k-1 & 2k-1\end{array}\right],
\qquad
D_{k}=\left[\begin{array}{cc}
2k & 2k\\
2k & 2k+1
\end{array}\right], \qquad
U_{k}=\left[\begin{array}{cc}
2k-1 & 2k-1\\
2k & 2k
\end{array}\right] \nonumber
\end{equation}
The homological diagram for the $(2,2p+1)$ torus knot consists of a single zig-zag, which is a building block of homologies for more complicated knots (as stated in Conjecture \ref{conj3}), and the above matrix represents the corresponding quiver. It is also interesting that, while increasing $p$, all previously determined entries of the matrix (\ref{C-torus22p1}) remain unchanged, so that it makes sense to consider the limit $p\to\infty$ of an infinite quiver.
Furthermore, from (\ref{eq:change of vars general}) we find parameters that determine (\ref{PxC}), which are indeed consistent with our conjectures. In particular ${\alpha_{1}}=d_{3}+d_{5}+\ldots+d_{2p+1}$ gives rise to the minus sign $(-1)^{\alpha_1}$ in (\ref{PxC}), which is consistent with the sign $(-1)^{\sum_i t_i d_i}$ determined by homological degrees $t_i$, encoded in the diagonal of (\ref{C-torus22p1})
\begin{equation}
(t_i) = (0,2,3,4,5,\ldots,2p,2p+1).
\end{equation}
In addition, the parameters $a_i$ and $l_i$ (and so $q_i$) in (\ref{PxC}) are determined by
\begin{equation}
\sum_i a_{i} d_i = 2pr+2\alpha_{1}=2p(d_{1}+d_{2}+d_{4}+\ldots+d_{2p}) +2(p+1)(d_{3}+d_{5}+\ldots+d_{2p+1}),
\end{equation}
\begin{align}
\sum_{i}l_{i} d_{i} &= -2pr+2(k_{1}+k_{2}+\ldots+k_{p})-3\alpha_{1} = \nonumber \\
& = -2pd_{1}+2(1-p)d_{2}+\big(2(1-p)-3\big)d_{3} +\nonumber \\
&\quad +2(2-p)d_{4}+\big(2(2-p)-3\big)d_{5} +\nonumber \\
&\qquad \quad \vdots\\
&\quad +2(p-1-p)d_{2(p-1)}+\big(2(p-1-p)-3\big)d_{2(p-1)+1} + \nonumber \\
&\quad +2(p-p)d_{2p}+\big(2(p-p)-3\big)d_{2p+1}.\nonumber
\end{align}
As a confirmation, for trefoil and cinquefoil knots, restricting (\ref{C-torus22p1}) to $p=1$ and $p=2$, we reproduce respectively (\ref{trefoil-quiver}) and (\ref{51-quiver})
\begin{equation}
C^{T_{2,3}}=\left[\begin{array}{cc}
F_{0} & F_{1}\\
F_{1}^{T} & D_{1}
\end{array}\right]=\left[\begin{array}{ccc}
0 & 1 & 1\\
1 & 2 & 2\\
1 & 2 & 3
\end{array}\right] \qquad \qquad
C^{T_{2,5}}=\left[\begin{array}{ccc}
F_{0} & F_{1} & F_{2}\\
F_{1}^{T} & D_{1} & U_{2}\\
F_{2}^{T} & U_{2}^{T} & D_{2}
\end{array}\right]=\left[\begin{array}{ccccc}
0 & 1 & 1 & 3 & 3\\
1 & 2 & 2 & 3 & 3\\
1 & 2 & 3 & 4 & 4\\
3 & 3 & 4 & 4 & 4\\
3 & 3 & 4 & 4 & 5
\end{array}\right]
\end{equation}
\subsection{$(2,2p)$ torus links}
In general, the analysis of HOMFLY-PT homology of links is more involved. However if all components of a link are colored by the same representation, they have properties analogous to knots. In particular colored HOMFLY-PT polynomials for $(2,2p)$ torus links, with all components colored by the same symmetric representation $S^r$, take the form \cite{Gukov:2015gmm}
\begin{align}
P_{[r]}^{T_{2,2p}}(a,q) &= a^{2pr}q^{-2pr}\sum_{0\leq s_{1}\leq\ldots\leq s_{p}\leq s_{p+1}=r}(a^{2}q^{-2};q^{2})_{s_{p}}(q^{2};q^{2})_{r-s_{1}}\times\nonumber \\
& \quad \times \prod_{i=1}^{p}q^{4s_{i}}(-q)^{-2s_{i}}q^{2rs_{i}-s_{i}s_{i+1}}{ s_{i+1}\brack s_{i}} . \label{eq:Torus link formula from GNSSS}
\end{align}
This expression corresponds to the so-called "finite-dimensional" version, which is a suitably normalized reduced colored HOMFLY-PT polynomial, that is actually a polynomial. It can be also rewritten as
\begin{align}
P_{[r]}^{T_{2,2p}}(a,q) &= a^{2pr}q^{-2pr}\sum_{0\leq k_{p}\leq\ldots\leq k_{1}\leq k_{0}=r} {r\brack k_{1}}
{k_{1}\brack k_{2}} \cdots {k_{p-1}\brack k_{p}} \times\nonumber \\
& \quad \times q^{2\sum_{i=1}^{p}((2r+1)k_{i}-k_{i-1}k_{i})}(a^{2}q^{-2};q^{2})_{k_{1}}(q^{2};q^{2})_{r-k_{p}}.\label{eq:P_r for general torus link}
\end{align}
Following analogous manipulations as in section \ref{ssec-torus22p1} we find that this expression can be further rewritten in the form (\ref{PxC}), with the corresponding quiver encoded in the matrix
\begin{equation}
C^{T_{2,2p}}=\left[\begin{array}{ccccccc}
F_{0} & F_{1} & F_{2} & F_{3} & \cdots & F_{p-1} & F_{p}^{e}\\
F_{1}^{T} & D_{1} & U_{2} & U_{3} & \cdots & U_{p-1} & U_{p}^{e}\\
F_{2}^{T} & U_{2}^{T} & D_{2} & U_{3} & \cdots & U_{p-1} & U_{p}^{e}\\
F_{3}^{T} & U_{3}^{T} & U_{3}^{T} & D_{3} & \cdots & U_{p-1} & U_{p}^{e}\\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots\\
F_{p-1}^{T} & U_{p-1}^{T} & U_{p-1}^{T} & U_{p-1}^{T} & \cdots & D_{p-1} & U_{p}^{e}\\
F_{p}^{eT} & U_{p}^{eT} & U_{p}^{eT} & U_{p}^{eT} & \cdots & U_{p}^{eT} & D_{p}^{e}
\end{array}\right] \label{C-quiver22p}
\end{equation}
Apart from the last column and row, the block entries take the form
\begin{equation}
F_{0}=\left[\begin{array}{cc}
0 & 0\\
0 & 1
\end{array}\right], \qquad \qquad
F_{k}=\left[\begin{array}{cccc}
2k-1 & 2k-1 & 2k-1 & 2k-1\\
2k & 2k & 2k & 2k
\end{array}\right],
\end{equation}
and
\begin{equation}
D_{k}=\left[\begin{array}{cccc}
2k+1 & 2k+1 & 2k & 2k+1\\
2k+1 & 2k+2 & 2k & 2k+1\\
2k & 2k & 2k & 2k \\
2k+1 & 2k+1 & 2k & 2k+1
\end{array}\right],\qquad
U_{k}=\left[\begin{array}{cccc}
2k & 2k & 2k & 2k \\
2k+1 & 2k+1 & 2k+1 & 2k+1\\
2k-1 & 2k-1 & 2k-1 & 2k-1\\
2k & 2k & 2k & 2k
\end{array}\right]
\end{equation}
In addition, the terms in the last column and row take the form
\begin{equation}
F_{p}^{e}=\left[\begin{array}{cc}
2p-1 & 2p-1\\
2p-1 & 2p-1
\end{array}\right], \qquad
D_{p}^{e}=\left[\begin{array}{cc}
2p+1 & 2p\\
2p & 2p
\end{array}\right], \qquad
U_{p}^{e}=\left[\begin{array}{cc}
2p & 2p\\
2p & 2p\\
2p-1 & 2p-1\\
2p-1 & 2p-1
\end{array}\right]
\end{equation}
The matrix (\ref{C-quiver22p}), being assigned to a link with two components, in fact represents a combination of two (appropriately shifted) zig-zags (\ref{C-torus22p1}).
Furthermore, the linear terms that determine (\ref{PxC}) take the form
\begin{align}
(-1)^{\sum_{i} t_{i} d_{i}}&= (-1)^{ (d_{3}+d_{4})+(d_{7}+d_{8})+\ldots+(d_{4p-1}+d_{4p}) +2 (d_{5}+d_{9}+\ldots+d_{4p+1} )}, \nonumber \\
\sum_{i}a_{i} d_{i}& = 2p\big(d_{1}+(d_{2}+d_{3})+(d_{6}+d_{7})+\ldots+(d_{4(p-1)-2}+d_{4(p-1)-1})+d_{4p+1}\big)+ \nonumber \\
&\quad +2(p+1)\big( (d_{4}+d_{5})+(d_{8}+d_{9})+\ldots+(d_{4(p-1)}+d_{4(p-1)+1})\big), \\
\sum_{i}l_{i} d_{i} &= (-2p)d_{1}+(-2p)d_{2}+(-1-2p)d_{3}+(-1-2p)d_{4}+(-2-2p)d_{5} + \nonumber \\
&\quad +(2-2p)d_{6}+(1-2p)d_{7}+(1-2p)d_{8}+(-2p)d_{9} + \nonumber \\
&\quad +(4-2p)d_{10}+(3-2p)d_{11}+(3-2p)d_{12}+(2-2p)d_{13} + \nonumber \\
&\qquad \quad \vdots\\
&\quad -4d_{4(p-1)-2}-5d_{4(p-1)-1}-5d_{4(p-1)}-6d_{4(p-1)+1} + \nonumber \\
&\quad -2d_{4p-2}-3d_{4p-1}-3d_{4p}.\nonumber
\end{align}
Specializing (\ref{C-quiver22p}) we find, for example, that quivers for the Hopf link ($p=1$) and the $T_{2,4}$ link ($p=2$) are represented by matrices
\begin{equation}
C^{T_{2,2}}=\left[\begin{array}{cc}
F_{0} & F_{1}^{e}\\
F_{1}^{eT} & D_{1}^{e}
\end{array}\right]=\left[\begin{array}{cccc}
0 & 0 & 1 & 1\\
0 & 1 & 1 & 1\\
1 & 1 & 3 & 2\\
1 & 1 & 2 & 2
\end{array}\right] \qquad
C^{T_{2,4}}=\left[\begin{array}{ccc}
F_{0} & F_{1} & F_{2}^{e}\\
F_{1}^{T} & D_{1} & U_{2}^{e}\\
F_{2}^{eT} & U_{2}^{eT} & D_{2}^{e}
\end{array}\right]=\left[\begin{array}{cccccccc}
0 & 0 & 1 & 1 & 1 & 1 & 3 & 3\\
0 & 1 & 2 & 2 & 2 & 2 & 3 & 3\\
1 & 2 & 3 & 3 & 2 & 3 & 4 & 4\\
1 & 2 & 3 & 4 & 2 & 3 & 4 & 4\\
1 & 2 & 2 & 2 & 2 & 2 & 3 & 3\\
1 & 2 & 3 & 3 & 2 & 3 & 3 & 3\\
3 & 3 & 4 & 4 & 3 & 3 & 5 & 4\\
3 & 3 & 4 & 4 & 3 & 3 & 4 & 4
\end{array}\right]
\end{equation}
The matrix $C^{T_{2,2}}$ for the Hopf-link represents two zig-zags, respectively of length 3 (which is identical to a matrix for the trefoil knot (\ref{trefoil-quiver})) and of length 1 (representing a single homology generator of $t$-degree 1). The matrix $C^{T_{2,4}}$ for the $T_{2,4}$ link consists of one zig-zag of length 5 (identical to a matrix for the $5_1$ knot (\ref{51-quiver})), and another zig-zag of length 3 (identical to a matrix for the trefoil knot, but with all elements shifted by 1, with $(1,3,4)$ on the diagonal).
\subsection{$(3,p)$ torus knots}
We discuss now torus knots from the $(3,p)$ family, which enables us to present other interesting aspects of the duality with quivers. Properties of these knots are much more involved than for $(2,2p+1)$ torus knots, in particular their homology is thick. General formulas for colored superpolynomials for arbitrary $(3,p)$ torus knot are unknown. Although there are explicit expressions for the colored HOMFLY-PT polynomials of arbitrary colors for torus knots, via Rosso-Jones formula \cite{RossoJones:1993}, the formulas involve different plethysm coefficients, that are changing with the colors. In such a way they are not suitable for obtaining explicit expressions for arbitrary symmetric color and consequently for obtaining the explicit generating function of colored HOMFLY-PT polynomials for general torus knots that we need for our main Conjecture \ref{PxCConj}.
However colored superpolynomials for the special cases of $(3,4)$ and $(3,5)$ torus knots (equivalently, respectively, $8_{19}$ and $10_{124}$ knots) were determined in \cite{Gukov:2015gmm}. In what follows we show, first, that these formulas can be rewritten in the general quiver form, and we identify corresponding quivers. Second, we show that such quivers are determined not uniquely, but only up to a permutation of some of their entries, which indicates some symmetry of the corresponding quiver moduli spaces. Furthermore, by simply assuming that there should exist a corresponding quiver, we find explicit formulas for colored HOMFLY-PT polynomials of $(3,7)$ torus knot, which have not been known before, and which nicely illustrate the power of our formalism.
Here we focus only on $(3,p)$ knots, rather than links, so $p$ cannot be a multiple of 3. The case of $p=1$ is the framed unknot, and $p=2$ represents the trefoil, already analyzed in section \ref{ssec-3151}. Therefore the first nontrivial examples involve $p=3,4$ and 7. Moreover, as computations become more involved and technical, in this section we only consider extremal (bottom row) invariants (\ref{quivgen2}); with some patience, and taking advantage of structural properties presented in Conjecture \ref{conj3}, these results can be generalized to the full $a$-dependence.
Let us consider the $(3,4)$ torus knot first. Its quadruply-graded Poincar{\'e} polynomial determined in \cite{Gukov:2015gmm} reads
\begin{align}
P_r(a,Q,t_r,t_c) &= a^{6 r} Q^{6 r} t_c^{6 r^2} t_r^{6 r} \sum _{\alpha=0}^r \sum _{\beta=\alpha}^r \sum _{\gamma=\beta}^r \sum _{j=\gamma}^r a^{2 (j-\gamma)} {\beta \brack \alpha}_{t_c^{-2}} {\gamma \brack \beta}_{t_c^{-2}} {j \brack \gamma}_{t_c^{-2}} {r \brack j}_{t_c^{-2}} \times \nonumber \\
& \times Q^{-4 \alpha-4 \beta+4 \gamma-8 j} t_c^{-2 (\alpha^2+\beta^2+\gamma^2 )-2 (j-\gamma) (\alpha+\beta+\gamma)-(j-\gamma)^2}
t_r^{-2 (\alpha+\beta+\gamma)-(j-\gamma)} \times \\
& \times \Big(-\frac{Q^2}{a^2 t_r t_c}; {t_c}^{-2}\Big)_{j-\gamma} \big(-a^2 Q^2 t_r^3 t_c^{2 r+1};t_c^2\big)_j, \nonumber
\end{align}
where now we denote ${\beta \brack \alpha}_{t_c^{-2}}=\frac{(t_c^{-2};t_c^{-2})_{\beta}}{ (t_c^{-2};t_c^{-2})_{\alpha} (t_c^{-2};t_c^{-2})_{\beta-\alpha} }$. Upon the identification of variables (\ref{super-homfly}), and extracting the terms at the lowest powers of $a$, we find that that extremal (bottom row) colored HOMFLY-PT polynomials for the $(3,4)$ torus knot take form
\begin{align}
P_r^{bottom}(q) = q^{6 r^2} \sum _{\alpha=0}^r \sum_{\beta=\alpha}^r \sum _{\gamma=\beta}^r \sum _{j=\gamma}^r \frac{(q^2;q^2)_r \,q^{-2\alpha (\beta-\gamma+j+1)-2\beta (j+1)+2\gamma- 2j (r+2)}}{(q^2;q^2)_{\alpha} (q^2;q^2)_{\beta-\alpha} (q^2;q^2)_{\gamma-\beta} \label{P34-bottom} (q^2;q^2)_{j-\gamma} (q^2;q^2)_{r-j}},
\end{align}
while its uncolored HOMFLY-PT homology has 5 generators in the bottom row, whose $q$-degrees and $t$-degrees are
\begin{equation}
\begin{split}
(q_1,q_2,q_3,q_4,q_5)&=(-6,-2,0,2,6), \\
(t_1,t_2,t_3,t_4,t_5)&=(0,2,4,6,8). \label{T34-qt}
\end{split}
\end{equation}
Manipulating the expression (\ref{P34-bottom}) we find that the corresponding quiver is represented by the following matrix
\begin{equation}
C^{T_{3,4}}=\left[\begin{array}{ccccc}
0&1&2&3&5\\
1&2&3&3&5\\
2&3&4&4&5\\
3&3&4&4&5\\
5&5&5&5&6
\end{array}\right]
\end{equation}
This quiver, together with $q$-degrees and $t$-degrees of 5 bottom row generators in (\ref{T34-qt}), encode all extremal (bottom row) colored HOMFLY-PT polynomials for the $(3,4)$ torus knot, which can be reconstructed from (\ref{quivgen2}). Moreover, simply the fact that we are able to identify this quiver proves the LMOV conjecture for all symmetric representations for this knot. Furthermore, the matrix $C^{T_{3,4}}$ captures the structure of the bottom row generators of HOMFLY-PT homology for the $(3,4)$ torus knot, which consists of one zig-zag (the same as for the $(2,7)$ torus knot) and one diamond. The part of the matrix with $(0,2,4,6)$ on the diagonal represents the bottom row of the zig-zag, and an additional 4 on the diagonal is the bottom row of the diamond (\ref{diamond}).
The next knot in this series is the $(3,5)$ torus knot. Its HOMFLY-PT homology and colored superpolynomials were also considered in \cite{Gukov:2015gmm}. For brevity, we just recall that colored superpolynomials for this knot take form
\begin{align}
P_r&(a,q,t) = \sum_{j=0}^r \sum_{k_1 = 0}^j \sum_{k_2 = 0}^{k_1} \sum_{k_3 = 0}^{k_2} \sum_{k_4 = 0}^{k_3} \sum_{i=0}^{r-j} a^{8r} \bigg(\frac{t}{q}\bigg)^{2(i+2j - k_1 - k_2 -k_3 -k_4 + 2r)} \times \\
& \times q^{-2(k_1 k_2 + k_2 k_3 + k_3 k_4 + 2 (k_1+k_2+k_3+k_4)+r+(k_1+k_2+k_3+k_4)r-2r^2-i(2+k_1+r) -j(5+k_2+k_3+k_4+2r)} \times \nonumber \\
& \times {j \brack k_1} {k_1 \brack k_2} {k_2\brack k_3} {k_3\brack k_4} {r-j \brack i}
(-a^2 t q^{-2};q^2)_{r-j} (-a^2 q^{-2-2j+2r};q^2)_{k_4 - j} (-a^2 q^{2r} t^3;q^2)_{r-j}. \nonumber
\end{align}
Setting $t=-1$ and extracting coefficients of minimal powers of $a$ reveals the form of extremal (bottom row) colored HOMFLY-PT polynomials
\begin{align}
P&^{bottom}_r(q) = \sum_{j=0}^r \sum_{k_1 = 0}^j \sum_{k_2 = 0}^{k_1} \sum_{k_3 = 0}^{k_2} \sum_{k_4 = 0}^{k_3} \sum_{i=0}^{r-j}\ (q^2;q^2)_r \times \\
&\times \frac{q^{-2 (k_2+k_3+k_4)-2 (k_1+k_1 k_2+k_2 k_3+k_3 k_4)-6 r-2 (k_1+k_2+k_3+k_4) r+4 r^2+2 i (1+k_1+r)+2 j (3+k_2+k_3+k_4+2 r)}}{(q^2;q^2)_i (q^2;q^2)_{j-k_1} (q^2;q^2)_{k_1-k_2} (q^2;q^2)_{k_2-k_3} (q^2;q^2)_{k_3-k_4} (q^2;q^2)_{k_4} (q^2;q^2)_{r-j-i}}. \nonumber
\end{align}
Furthermore, the HOMFLY-PT homology for $(3,5)$ torus knot has 7 generators in the bottom row, with the following $q$-degrees and $t$-degrees
\begin{equation}
\begin{split}
(q_1,q_2,\ldots,q_7)&=(-8,-4,-2,0,2,4,8), \\
(t_1,t_2,\ldots,t_7)&=(0,2,4,4,6,6,8). \label{P35-qt}
\end{split}
\end{equation}
Manipulating the above expressions we find, surprisingly, that there are two quivers which reproduce the same extremal (bottom row) colored HOMFLY-PT polynomials via (\ref{quivgen2}). These two quivers are very similar, and their matrices differ only by a permutation of a few entries. They take the following form; the permuted entries are underlined in the second matrix
\begin{equation}
C^{T_{3,5}}=\left[\begin{array}{ccccccc}
0&1&2&3&4&5&7\\
1&2&3&3&5&5&7\\
2&3&4&4&5&5&7\\
3&3&4&4&6&5&7\\
4&5&5&6&6&6&7\\
5&5&5&5&6&6&7\\
7&7&7&7&7&7&8
\end{array}\right] \qquad \qquad
\widetilde{C}^{T_{3,5}}=\left[\begin{array}{ccccccc}
0&1&2&3&4&5&7\\
1&2&3&3&5&5&7\\
2&3&4&4&5&5&\underline{6}\\
3&3&4&4&6&5&7\\
4&5&5&6&6&\underline{7}&7\\
5&5&5&5&\underline{7}&6&7\\
7&7&\underline{6}&7&7&7&8
\end{array}\right]
\end{equation}
Again let us stress, that the quiver represented by either of the above matrcies, together with $q$-degrees and $t$-degrees in (\ref{P35-qt}), encode all extremal colored HOMFLY-PT polynomials for $(3,5)$ torus knot, which can be reconstructed from (\ref{quivgen2}). The fact that we are able to identify these quivers also proves the LMOV conjecture for all symmetric representations for this knot.
The next knot we consider is the $(3,7)$ torus knot. Its colored superpolynomials, or even explicit, closed-form expressions of HOMFLY-PT polynomials colored by arbitrary symmetric representations, have not been known before. However, based on the structure (\ref{quivgen2}), and by comparing the results with the Rosso-Jones formula for the first several symmetric representations, we are able to reconstruct the corresponding quiver, which then encodes HOMFLY-PT polynomials colored by arbitrary symmetric representations, and corresponding integral LMOV invariants. The homology of the bottom row of the $(3,7)$ torus knot has 12 generators with the following $q$-degrees and $t$-degrees
\begin{equation}
\begin{split}
(q_1,q_2,\ldots,q_{12})&=(-12,-8,-6,-4,-2,0,0,2,4,6,8,12), \\
(t_1,t_2,\ldots,t_{12})&=(0,2,4,4,6,6,8,8,8,10,10,12),
\end{split}
\end{equation}
and we find that the corresponding quiver is encoded in a matrix
\begin{equation}
C^{T_{3,7}}=\left[\begin{array}{cccccccccccc}
0&1&2&3&4&5&5&6&7&8&9&11\\
1&2&3&3&5&5&6&7&7&9&9&11\\
2&3&4&4&5&6&6&7&8&9&10&11\\
3&3&4&4&5&5&7&7&7&9&9&11\\
4&5&5&5&6&6&7&7&8&9&10&11\\
5&5&6&5&6&6&8&7&7&9&9&11\\
5&6&6&7&7&8&8&8&9&9&10&11\\
6&7&7&7&7&7&8&8&8&9&10&11\\
7&7&8&7&8&7&9&8&8&9&9&11\\
8&9&9&9&9&9&9&9&9&10&10&11\\
9&9&10&9&10&9&10&10&9&10&10&11\\
11&11&11&11&11&11&11&11&11&11&11&12
\end{array}\right].
\end{equation}
In this case we also find other quivers, whose matrices differ from the above one by permutation of several entries, and yet encode the same generating series (\ref{quivgen2}).
\subsection{Twist knots $4_1, 6_1, 8_1, \ldots$}
Another infinite family of knots that we consider are twist knots, which are labelled by an integer $p$. Negative values of this parameter, i.e. $p=-1,-2,-3,\ldots,$ correspond to $4_1, 6_1,8_1,\ldots$ knots; these are simply $(2|p|+2)_1$ knots, which are also denoted $TK_{2|p|+2}$. The colored HOMFLY-PT polynomials for these knots are determined in \cite{FGSS,Nawata} and they take the form
\begin{align}
P_{r}^{TK_{2|p|+2}}(a,q)= & \sum_{0\leq k_{|p|}\leq\ldots\leq k_{2}\leq k_{1}\leq r}{r \brack k_1}{k_1 \brack k_2}\cdots{k_{|p|-1}\brack k_{|p|}}\times\nonumber \\
& \qquad \times a^{2\sum_{i=1}^{|p|}k_{i}}q^{2\sum_{i=1}^{|p|}\left(k_{i}^{2}-k_{i}\right)}(a^{-2}q^{2};q^{-2})_{k_{1}}(a^{-2}q^{-2r};q^{-2})_{k_{1}}. \label{eq:P_r for general twist knot with p<0}
\end{align}
Following manipulations similar to the previous examples, we find that the quiver corresponding to a given $p<0$, i.e. to a given $TK_{2|p|+2}$ knot, is encoded in the matrix
\begin{equation}
C^{TK_{2|p|+2}}=\left[\begin{array}{ccccccc}
F_{0} & F & F & F & \cdots & F & F\\
F^{T} & D_{1} & R_{1} & R_{1} & \cdots & R_{1} & R_{1}\\
F^{T} & R_{1}^{T} & D_{2} & R_{2} & \cdots & R_{2} & R_{2}\\
F^{T} & R_{1}^{T} & R_{2}^{T} & D_{3} & \cdots & R_{3} & R_{3}\\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots\\
F^{T} & R_{1}^{T} & R_{2}^{T} & R_{3}^{T} & \cdots & D_{|p|-1} & R_{|p|-1}\\
F^{T} & R_{1}^{T} & R_{2}^{T} & R_{3}^{T} & \cdots & R_{|p|-1}^{T} & D_{|p|}
\end{array}\right]
\end{equation}
where
\begin{equation}
F_{0}=\left[0\right]\qquad\qquad F=\left[\begin{array}{cccc}
0 & -1 & 0 & -1\end{array}\right]
\end{equation}
and
\begin{equation}
D_{k}=\left[\begin{array}{cccc}
2k & 2k-2 & 2k-1 & 2k-3\\
2k-2 & 2k-3 & 2k-2 & 2k-4\\
2k-1 & 2k-2 & 2k-1 & 2k-3\\
2k-3 & 2k-4 & 2k-3 & 2k-4
\end{array}\right] \qquad
R_{k}=\left[\begin{array}{cccc}
2k & 2k-2 & 2k-1 & 2k-3\\
2k-1 & 2k-3 & 2k-2 & 2k-4\\
2k & 2k-1 & 2k-1 & 2k-3\\
2k-2 & 2k-3 & 2k-2 & 2k-4
\end{array}\right]
\end{equation}
The element $F_0$ represents a zig-zag of length 1, corresponding to a single homology generator, while the diagonal blocks $D_k$ represent (up to a permutation of homology generators, and an overall shift) diamonds (\ref{diamond}).
Note that in this case it also makes sense to consider the $-p\to\infty$ limit, and the corresponding infinite quiver.
The other parameters that determine the form of the generating series (\ref{PxC}) for the $TK_{2|p|+2}$ knot take the form
\begin{align}
(-1)^{\sum_{i}p_{i} d_{i}} &= (-1)^{ (d_{3}+d_{4})+(d_{7}+d_{8})+\ldots+(d_{4|p|-1}+d_{4|p|}) +2(d_{5}+d_{9}+\ldots+d_{4|p|+1}) } , \nonumber \\
\sum_{i}a_{i} d_{i}& = 2d_{2}+0d_{3}+0d_{4}-2d_{5}+\nonumber \\
& \qquad +4d_{6}+2d_{7}+2d_{8}+0d_{9}+\\
& \qquad \quad \vdots\nonumber \\
& \qquad +2|p|d_{4|p|-2}+\left(2|p|-2\right)d_{4|p|-1}+\left(2|p|-2\right)d_{4|p|}+\left(2|p|-4\right)d_{4|p|+1}, \nonumber \\
\sum_{i}l_{i} d_{i} &= -2d_{2}-d_{3}+d_{4}+2d_{5} + \nonumber \\
& \qquad -4d_{6}-3d_{7}-1d_{8}+0d_{9}+\nonumber \\
& \qquad \quad \vdots\\
& \qquad -2|p|d_{4|p|-2}+\left(1-2|p|\right)d_{4|p|-1}+\left(3-2|p|\right)d_{4|p|}+\left(4-2|p|\right)d_{4|p|+1}.\nonumber
\end{align}
For example, the quiver for the $p=-1$ case, i.e. the figure-eight knot $4_1$, whose homology diagram is shown in figure \ref{diagrams}, is represented by the matrix
\begin{equation}
C^{TK_{4}}=\left[\begin{array}{cc}
F_{0} & F\\
F^{T} & D_{1}
\end{array}\right]=\left[\begin{array}{ccccc}
0 & 0 & -1 & 0 & -1\\
0 & 2 & 0 & 1 & -1\\
-1 & 0 & -1 & 0 & -2\\
0 & 1 & 0 & 1 & -1\\
-1 & -1 & -2 & -1 & -2
\end{array}\right]
\end{equation}
This matrix is consistent with Conjecture \ref{conj3}: the top left entry $0$ represents the zig-zag of length 1, and the remaining diagonal block of size $4\times 4$ represents a diamond and agrees (up to a permutation of homology generators, and corresponding rows and columns) with (\ref{diamond}) for $k=-2$.
For $p=-2$, i.e. the $6_{1}$ knot, the quiver is represented by the following matrix
\begin{equation}
C^{TK_{6}}=\left[\begin{array}{ccc}
F_{0} & F & F\\
F^{T} & D_{1} & R_{1}\\
F^{T} & R_{1}^{T} & D_{2}
\end{array}\right]=\left[\begin{array}{ccccccccc}
0 & 0 & -1 & 0 & -1 & 0 & -1 & 0 & -1\\
0 & 2 & 0 & 1 & -1 & 2 & 0 & 1 & -1\\
-1 & 0 & -1 & 0 & -2 & 1 & -1 & 0 & -2\\
0 & 1 & 0 & 1 & -1 & 2 & 1 & 1 & -1\\
-1 & -1 & -2 & -1 & -2 & 0 & -1 & 0 & -2\\
0 & 2 & 1 & 2 & 0 & 4 & 2 & 3 & 1\\
-1 & 0 & -1 & 1 & -1 & 2 & 1 & 2 & 0\\
0 & 1 & 0 & 1 & 0 & 3 & 2 & 3 & 1\\
-1 & -1 & -2 & -1 & -2 & 1 & 0 & 1 & 0
\end{array}\right] \label{C-61}
\end{equation}
Note that in these examples some entries of matrices $C$ are negative. In order to have a proper quiver representation theory interpretation, we can change the framing (\ref{C-frame}) to shift all values of these matrices by a constant and make them nonnegative.
\subsection{Twist knots $3_1, 5_2, 7_2, 9_2, \ldots$}
Another class of twist knots is characterized by $p>0$, which are respectively $3_1, 5_2,7_2,9_2,\ldots$ knots, which have $2p+1$ crossings, and are also denoted $TK_{2p+1}$. Their superpolynomials take the form \cite{Nawata}
\begin{align}
P_{S^{n-1}}^{TK_{2p+1}}(a,q,t)= & \sum_{0\leq s_{1}\leq\ldots\leq s_{p}<\infty}(-t)^{-n+1}q^{2s_{p}}\frac{(-a^{2}tq^{-2};q^{2})_{s_{p}}}{(q^{2};q^{2})_{s_{p}}}(q^{2-2n};q^{2})_{s_{p}}(-a^{2}t^{3}q^{2n-2};q^{2})_{s_{p}}\times\nonumber \\
& \qquad\times\prod_{i=1}^{p-1}q^{4s_{i}}(a^{2}t^{2})^{s_{i}}q^{2s_{i}(s_{i}-1)}{s_{i+1}\brack s_{i}}. \label{eq:Twist knot with p>0 formula from NRZ}
\end{align}
Here we can illustrate another subtlety, which is the fact that sometimes -- in particular for $TK_{2p+1}$ knots -- more general quivers can be assigned to a given knot, which are however not consistent with our conjectures. For example, setting $p=1$ and $t=-1$ in (\ref{eq:Twist knot with p>0 formula from NRZ}), we find the following representation of the colored HOMFLY-PT polynomials for the trefoil knot
\begin{equation}
P_{S^{r}}^{TK_{3}}(a,q)=\sum_{0\leq s_{1}\leq r} {r\brack s_{1}} (-1)^{s_{1}}q^{-2s_{1}r+s_{1}^{2}+s_{1}}(a^{2}q^{-2};q^{2})_{s_{1}}(a^{2}q^{2r};q^{2})_{s_{1}}. \label{eq:Trefoil knot treated as twist formula}
\end{equation}
This expression is equal to (\ref{Pr-31-homfly}), however its naive rewriting in the form (\ref{PxC}) leads to the quiver represented by the following matrix
\begin{equation}
\left[\begin{array}{ccccc}
0 & -1 & 0 & -1 & 0\\
-1 & -1 & 0 & -1 & 0\\
0 & 0 & 2 & 1 & 2\\
-1 & -1 & 1 & 0 & 1\\
0 & 0 & 2 & 1 & 3
\end{array}\right]
\end{equation}
The last three rows and columns of this matrix contain the trefoil quiver matrix (\ref{trefoil-quiver}) that we found earlier, however now we find two additional rows and columns. In fact the same issue arises for all twist knots $TK_{2p+1}$ in this series. Moreover, the structure of terms with linear powers of $d_i$ in the generating series that determines such enlarged quivers, is also not quite consistent with the structure of the parameters in (\ref{PxC}) and their relation to homological degrees. Nonetheless, we can get rid of these additional rows and columns, at the same time fixing terms with linear powers of $d_i$, by taking advantage of the following lemma.
\begin{lemma}
Consider a generating function (not necessarily related to a knot) of the form (\ref{PxC}), determined by a quiver $C$ of size $n\times n$. Up to appropriate adjustment of terms with with linear powers of $d_i$, the same generating function is assigned to the modified quiver
\begin{equation}
C^{+}=\left[\begin{array}{ccccccc}
1+\alpha_{0} & \alpha_{0} & \alpha_{1} & \alpha_{2} & \cdots & \alpha_{n-1} & \alpha_{n}\\
\alpha_{0} & \alpha_{0} & \alpha_{1} & \alpha_{2} & \cdots & \alpha_{n-1} & \alpha_{n}\\
\alpha_{1} & \alpha_{1}\\
\alpha_{2} & \alpha_{2}\\
\vdots & \vdots & & & C\\
\alpha_{n-1} & \alpha_{n-1}\\
\alpha_{n} & \alpha_{n}
\end{array}\right]\label{eq:C+}
\end{equation}
for every $\alpha_{0},\alpha_{1},\ldots,\alpha_{n}\in\mathbb{Z}$.
\end{lemma}
\textbf{Proof:}\\
Note that, for $m\geq 1$, using (\ref{qpoch-sum}), we have
\begin{equation}
0=(1;q^{-2})_{m}=\sum\limits _{a+b=m}(-1)^{b}q^{-b^{2}+b-2ab}\frac{(q^2;q^2)_m}{(q^2;q^2)_a (q^2;q^2)_b}.
\end{equation}
It follows that, for arbitrary $\alpha_{0},\alpha_{1},\ldots,\alpha_{n},d_{1},\ldots,d_{n}\in\mathbb{Z}$,
\begin{align}
1 & = \sum\limits _{m\geq0}q^{(\alpha_{0}+1)m^{2}+2(\alpha_{1}d_{1}+\alpha_{2}d_{2}+\ldots+\alpha_{n}d_{n})m} x^m\frac{(1;q^{-2})_{m}}{(q^{2};q^{2})_{m}} = \nonumber \\
&= \sum\limits _{a,b\geq0}(-1)^{b}q^{-b^{2}+b-2ab+(\alpha_{0}+1)(a+b)^{2}+2(\alpha_{1}d_{1}+\alpha_{2}d_{2}+\ldots+\alpha_{n}d_{n})(a+b)}\frac{x^{a+b}}{(q^{2};q^{2})_{a}(q^{2};q^{2})_{b}}. \label{1ab}
\end{align}
Therefore, if $P_{C}$ is a generating function determined by a quiver $C$, then
\begin{equation}
P_{C^{+}}=\sum\limits _{a,b\geq0}q^{-b^{2}+b-2ab+(\alpha_{0}+1)(a+b)^{2}+2(\alpha_{1}d_{1}+\alpha_{2}d_{2}+\ldots+\alpha_{n}d_{n})(a+b)}\frac{(-1)^{b} x^{a+b}}{(q^{2};q^{2})_{a}(q^{2};q^{2})_{b}}P_{C}\label{eq:auxiliarity of the first two variables}
\end{equation}
is a generating function determined by a quiver $C^{+}$ in (\ref{eq:C+}), and from (\ref{1ab}) we clearly see that $P_{C^{+}}=P_{C}$, which completes the proof. \hfill\rule{6pt}{6pt}
Having in mind the above subtlety, in order to find a quiver representation consistent with our conjectures, we rewrite the colored HOMFLY-PT polynomials (\ref{eq:Twist knot with p>0 formula from NRZ}) in the form
\begin{align}
P_{r}^{TK_{2p+1}}(a,q)= & \sum_{0\leq k_{p}\leq\ldots\leq k_{2}\leq k_{1}\leq r} {r\brack k_{1}} {k_{1}\brack k_{2}}\cdots {k_{p-1}\brack k_{p} } \times\nonumber \\
& \qquad \times a^{2\sum_{i=2}^{p}k_{i}}q^{-2k_{1}r+k_{1}^{2}+k_{1}+2\sum_{i=2}^{p}\left(k_{i}^{2}-k_{i}\right)}(a^{2}q^{-2};q^{2})_{k_{1}}(a^{2}q^{2r};q^{2})_{k_{1}}.\label{eq:P_r for general twist knot with p>0}
\end{align}
Following manipulations analogous to the previous sections, we now find that the generating series (\ref{PxC}) for $TK_{2p+1}$ knot is determined by a quiver whose matrix takes the form
\begin{equation}
C^{TK_{2p+1}}=\left[\begin{array}{ccccccc}
D_{1} & R_{1} & R_{1} & R_{1} & \cdots & R_{1} & R_{1}\\
R_{1}^{T} & D_{2} & R_{2} & R_{2} & \cdots & R_{2} & R_{2}\\
R_{1}^{T} & R_{2}^{T} & D_{3} & R_{3} & \cdots & R_{3} & R_{3}\\
R_{1}^{T} & R_{2}^{T} & R_{3}^{T} & D_{4} & \cdots & R_{4} & R_{4}\\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots\\
R_{1}^{T} & R_{2}^{T} & R_{3}^{T} & R_{4}^{T} & \cdots & D_{p-1} & R_{p-1}\\
R_{1}^{T} & R_{2}^{T} & R_{3}^{T} & R_{4}^{T} & \cdots & R_{p-1}^{T} & D_{p}
\end{array}\right] \label{C-TK2p1}
\end{equation}
where the block elements in the first row and column are
\begin{equation}
D_{1}=\left[\begin{array}{ccc}
2 & 1 & 2\\
1 & 0 & 1\\
2 & 1 & 3
\end{array}\right]\qquad \qquad
R_{1}=\left[\begin{array}{cccc}
1 & 2 & 1 & 2\\
0 & 2 & 0 & 1\\
1 & 3 & 2 & 3
\end{array}\right] \label{D1R1-TK2p1}
\end{equation}
and all other elements, for $k>1$, take the form
\begin{equation}
D_{k}=\left[\begin{array}{cccc}
2k-3 & 2k-2 & 2k-3 & 2k-2\\
2k-2 & 2k & 2k-1 & 2k\\
2k-3 & 2k-1 & 2k-2 & 2k-1\\
2k-2 & 2k & 2k-1 & 2k+1
\end{array}\right] \qquad \quad
R_{k}=\left[\begin{array}{cccc}
2k-3 & 2k-2 & 2k-3 & 2k-2\\
2k-1 & 2k & 2k-1 & 2k \\
2k-2 & 2k & 2k-2 & 2k-1\\
2k-1 & 2k+1 & 2k & 2k+1
\end{array}\right]
\end{equation}
In this case $D_1$ represents a zig-zag of the same form as for the trefoil knot (\ref{trefoil-quiver}), and the $D_k$ (for $k>1$) represent (up to a permutation of homology generators and an overall constant shift) the diamonds (\ref{diamond}).
The other parameters that determine (\ref{PxC}), now with vertices of a quiver (or homology generators), and thus also the summation variables $d_i$ numbered from $3$ to $4p+1$ (after removing $d_1$ and $d_2$ using the above lemma), take the form
\begin{align}
(-1)^{\sum_{i=3}^{4p+1}t_{i} d_{i}} &= (-1)^{ (d_{3}+d_{4})+(d_{7}+d_{8})+\ldots+(d_{4p-1}+d_{4p}) +2(d_{5}+d_{9}+\ldots+d_{4p+1})}\nonumber \\
\sum_{i=3}^{4p+1}a_{i} d_{i} &= 2(d_{3}+d_{4})+4d_{5} + \nonumber \\
& \qquad +2d_{6}+4d_{7}+4d_{8}+6d_{9} + \\
& \qquad +4d_{10}+6d_{11}+6d_{12}+8d_{13} + \nonumber \\
& \qquad \quad \vdots\nonumber \\
& \qquad +2(p-1)d_{4p-2}+2pd_{4p-1}+2pd_{4p}+2(p+1)d_{4p+1}\nonumber \\
\sum_{i=2}^{4p}l_{i} d_{i} &= -2d_{4}-3d_{5} +\nonumber \\
&\qquad -d_{6}-2d_{7}-4d_{8}-5d_{9}+\nonumber \\
&\qquad -3d_{10}-4d_{11}-6d_{12}-7d_{13}+\nonumber \\
&\qquad \quad \vdots\\
&\qquad +(1-2p)d_{4p-2}+(2-2p)d_{4p-1}+(-2p)d_{4p}+(-1-2p)d_{4p+1}.\nonumber
\end{align}
For example, for $p=1$, which represents simply the trefoil knot, the quiver matrix (\ref{C-TK2p1}) consists only of the element $D_1$ in (\ref{D1R1-TK2p1}), and up to permutation of vertices it is equivalent to (\ref{trefoil-quiver}). For $p=2$, i.e. for the knot $5_{2}$, from (\ref{C-TK2p1}) we obtain a quiver represented by the matrix
\begin{equation}
C^{TK_{5}}=\left[\begin{array}{ccc}
D_{1} & R_{1} & R_{1}\\
R_{1}^{T} & D_{2} & R_{2}\\
R_{1}^{T} & R_{2}^{T} & D_{3}
\end{array}\right]=\left[\begin{array}{ccccccc}
2 & 1 & 2 & 1 & 2 & 1 & 2\\
1 & 0 & 1 & 0 & 2 & 0 & 1\\
2 & 1 & 3 & 1 & 3 & 2 & 3\\
1 & 0 & 1 & 1 & 2 & 1 & 2\\
2 & 2 & 3 & 2 & 4 & 3 & 4\\
1 & 0 & 2 & 1 & 3 & 2 & 3\\
2 & 1 & 3 & 2 & 4 & 3 & 5
\end{array}\right]
\end{equation}
\subsection{$6_2$ and $6_3$ knots}
Finally we discuss knots with six crossings, $6_2$ and $6_3$ (the third prime knot with six crossings is the twist knot $6_1$, whose quiver we already identified in (\ref{C-61})). Explicit expressions for colored polynomials for those knots have not been known before. Assuming that they are consistent with our conjectures, we are able to determine such expressions, as being encoded in the corresponding quivers. This again shows the power of our formalism.
Let us consider $6_2$ knot first. Its uncolored HOMFLY-PT homology has 11 generators, which have the following degrees
\begin{equation}
\begin{split}
(a_1,\ldots,a_{11})&=(0,2,2,0,2,2,2,4,4,2,4) , \\
(q_1,\ldots,q_{11})&=(-2,-4,-2,2,0,0,2,-2,0,4,2), \\
(t_1,\ldots,t_{11})&=(-2,-1,0,0,1,1,2,2,3,3,4).
\end{split}
\end{equation}
Assuming that the generating series of colored HOMFLY-PT polynomials takes the form (\ref{PxC}), and comparing that generating series with several first such polynomials obtained using the Rosso-Jones formula, we find that the corresponding quiver is encoded in the following matrix
\begin{equation}
C^{6_2}=\left[
\begin{array}{ccccccccccc}
-2\ & -2\ & -1\ & -1\ & -1\ & -1\ & 0\ & -1\ & 1\ & 1\ & 1 \\
-2\ & -1\ & -1\ & 0\ & 0\ & 0\ & 1\ & 0\ & 1\ & 2\ & 2 \\
-1\ & -1\ & 0\ & 1\ & 0\ & 0\ & 1\ & 0\ & 1\ & 2\ & 2 \\
-1\ & 0\ & 1\ & 0\ & 0\ & 0\ & 1\ & 0\ & 2\ & 1\ & 1 \\
-1\ & 0\ & 0\ & 0\ & 1\ & 1\ & 1\ & 1\ & 2\ & 2\ & 2 \\
-1\ & 0\ & 0\ & 0\ & 1\ & 1\ & 1\ & 1\ & 2\ & 2\ & 2 \\
0\ & 1\ & 1\ & 1\ & 1\ & 1\ & 2\ & 1\ & 2\ & 2\ & 2 \\
-1\ & 0\ & 0\ & 0\ & 1\ & 1\ & 1\ & 2\ & 2\ & 3\ & 3 \\
1\ & 1\ & 1\ & 2\ & 2\ & 2\ & 2\ & 2\ & 3\ & 3\ & 3 \\
1\ & 2\ & 2\ & 1\ & 2\ & 2\ & 2\ & 3\ & 3\ & 3\ & 3 \\
1\ & 2\ & 2\ & 1\ & 2\ & 2\ & 2\ & 3\ & 3\ & 3\ & 4 \\
\end{array}
\right]
\end{equation}
In this matrix one can identify diagonal blocks, one corresponding to a zig-zag of the same form as in the trefoil quiver (\ref{trefoil-quiver}), and two diamonds of the form (\ref{diamond}).
Analogously we analyze the generating series for the $6_3$ knot, whose uncolored HOMFLY-PT homology has 13 generators of the following degrees
\begin{equation}
\begin{split}
(a_1,\ldots,a_{13})&=(0,2,0,0,-2,2,0,0,-2,2,0,0,-2), \\
(q_1,\ldots,q_{13})&=(0,-2,0,-4,-2,0,2,-2,0,2,4,0,2), \\
(t_1,\ldots,t_{13})&=(0,1,0,-2,-3,2,1,-1,-2,3,2,0,-1).
\end{split}
\end{equation}
Similarly, comparing first few colored HOMFLY-PT polynomials from (\ref{PxC}) with the Rosso-Jones formula, we find the corresponding quiver
\begin{equation}
C^{6_3}=\left[
\begin{array}{ccccccccccccc}
0\ & 0\ & 0 & -1\ & -1\ & 0\ & 0\ & -1\ & -1\ & 0\ & 0\ & -1\ & -1 \\
0\ & 1\ & 0 & -1\ & -2\ & 1\ & 0\ & -1\ & -2\ & 1\ & 1\ & 0\ & -1 \\
0\ & 0\ & 0 & -1\ & -2\ & 1\ & 0\ & 0\ & -2\ & 1\ & 1\ & 0\ & 0 \\
-1\ & -1\ & -1\ & -2\ & -3\ & 0\ & -1\ & -2\ & -3\ & -1\ & 0\ & -2\ & -2 \\
-1\ & -2\ & -2\ & -3\ & -3\ & -1\ & -1\ & -2\ & -3\ & -1\ & -1\ & -2\ & -2 \\
0\ & 1\ & 1\ & 0 & -1\ & 2\ & 1\ & 0\ & -1\ & 2\ & 1\ & 1\ & -1 \\
0\ & 0\ & 0\ & -1 & -1\ & 1\ & 1\ & 0\ & -1\ & 2\ & 1\ & 1\ & 0 \\
-1\ & -1\ & 0\ & -2 & -2\ & 0\ & 0\ & -1\ & -2\ & 0\ & 0\ & -1\ & -2 \\
-1\ & -2\ & -2\ & -3 & -3\ & -1\ & -1\ & -2\ & -2\ & 0\ & -1\ & -1\ & -2 \\
0\ & 1 & 1\ & -1\ & -1\ & 2\ & 2 & 0\ & 0\ & 3\ & 2\ & 1\ & 0 \\
0\ & 1 & 1\ & 0\ & -1\ & 1\ & 1 & 0\ & -1\ & 2\ & 2\ & 1\ & 0 \\
-1\ & 0 & 0\ & -2\ & -2\ & 1\ & 1 & -1\ & -1\ & 1\ & 1\ & 0\ & -1 \\
-1\ & -1 & 0\ & -2\ & -2\ & -1\ & 0 & -2\ & -2\ & 0\ & 0\ & -1\ & -1 \\
\end{array}
\right]
\end{equation}
In this matrix one can identify diagonal blocks, one corresponding to a zig-zag of length 1 (representing a homology generator with $t$-degree 0), and three diamonds of the form (\ref{diamond}).
We checked that the above results agree with those obtained using the formalism of differentials, presented in \cite{Nawata:2015wya}.
\acknowledgments{We thank Sergei Gukov, Satoshi Nawata, Mi{\l}osz Panfil, Yan Soibelman, Richard Thomas, Cumrun Vafa, and Paul Wedrich for comments and discussions. We are grateful to the American Institute of Mathematics (San Jose), the Isaac Newton Institute for Mathematical Sciences (Cambridge University), and Institut Henri Poincar{\'e} (Paris) for hospitality. The last author thanks the California Institute of Technology, the Matrix Institute (University of Melbourne, Creswick), University of California Davis, and the Isaac Newton Institute for Mathematical Sciences (Cambridge University) for hospitality and opportunity to present the results of this work during seminars and conferences within the last year. This work is supported by the ERC Starting Grant no. 335739 \emph{``Quantum fields and knot homologies''} funded by the European Research Council under the European Union’s Seventh Framework Programme, and the Foundation for Polish Science. M.S. is partially supported by the Ministry of Science of Serbia, project no. 174012.}
\newpage
\bibliographystyle{JHEP}
|
\section{Introduction}
In \cite{linear-open-arxiv}, the widely known linear time algorithm for computing the maximum area triangle in a given convex polygon $P$ (given in \cite{linear-wrong-DS}) was found incorrect.
An alternative algorithm given in \cite{kgon82} is also wrong for the same reason.
Also, another algorithm (see \cite{linear-wrong-india}) not mentioned in \cite{linear-open-arxiv} -- perhaps because published by an informal journal -- is also wrong.
Whether this problem can be solved in linear time was proposed as an open problem in \cite{linear-open-arxiv},
and is of significance because follow-up results in computational geometry (e.g. polygon collision detection) may depend on finding the maximum area triangle as a preprocessing step.
To study this problem, \cite{linear-open-arxiv} introduced the 3-stable triangles, i.e.\ those triangles whose corners lie at $P$'s vertices and which cannot be improved to a larger one by adjusting the position of one corner.
They proved that there are $O(n)$ such triangles since they are pairwise interleaving (see Definition~\ref{def:interleave} below), where $n$ is the number of vertices in $P$.
But it was left as an open problem whether we can compute all the 3-stable triangles in $O(n)$ time.
However, via personal communication, the authors of \cite{linear-open-arxiv} later pointed us to an algorithm of Chandran and Mount \cite{linear-correct},
which correctly computes a superset of the 3-stable triangles in linear time.
This algorithm is a typical and clever application of the well-known rotating-caliper technique \cite{rotatingcaliper}.
It in fact computes the maximum triangles in $P$ and the minimum triangle enclosing $P$ simultaneously (see more details in Subsection~\ref{subsect:CM}).
Moreover, Chandran and Mount showed in detail how to convert their algorithm into a parallel version.
\smallskip In this paper, by applying a new technique for solving extremal polygon inclusion problems which we call \emph{Rotate-and-Kill},
we present a different linear time algorithm for computing the 3-stable triangles.
This algorithm (Alg-J) has advantages over the previous algorithm (Alg-CM) in almost every aspect, except for the sole aspect from parallel computation.\smallskip
\begin{description}
\item[Conciseness.] Comparing the description of Alg-J (mainly in Algorithm~\ref{alg:r-and-k}) with that of Alg-CM (mainly in pages 9--10 of \cite{linear-correct}),
the new algorithm is conceptually simpler. The latter contains complicated subroutines for handling three different cases in each iteration.\smallskip
\item[Easier proof.] Correspondingly, our proof is much shorter and easier to understand.
Our proof is based on few prerequisite results -- only the aforementioned pairwise interleaving property proved in \cite{linear-open-arxiv},
whereas the proof of Alg-CM applies nontrivial intermediate results of \cite{Tri-Enclose-Area-kleeMid,Tri-Enclose-Area-nlogn2,Tri-Enclose-Area}. In fact, \cite{linear-correct} has not given a rigourous proof for its correctness.
\item[Easier implementation.] Alg-CM has not been implemented before our work.
In order to justify the superiority of our algorithm which was doubted by previous reviewers,
we implement both algorithms by \emph{C++ programs} (uploaded to \cite{implementation}).
Alg-J is 4 times shorter in the code length.
In this place, we point out that an implementation of Alg-CM requires much more attentions because many degenerated cases are not discussed in \cite{linear-correct}.
For example, in Figure~5 of their paper, $c$ could be coincided with $A$ or $B$, and then angles $\theta_a=\angle axa',\theta_b=\angle bxb'$ could be undefined;
and in Figure~6 of their paper, $x$ may not lie in the ray originating from $a$ towards $a'$, which makes the definition of $\theta_b$ problematic.
\item[Efficiency.] Our experiments (demonstrated in Figure~\ref{fig:experiments}) indicate that Alg-J is approximately 30 times faster, which matches our expectation.
Although both run in $O(n)$ time, Alg-J has a smaller constant behind the asymptotic complexity, and it only executes basic operations $+,-,\times,/$, whereas Alg-CM seems unable to avoid trigonometric functions.
\item[Robust.] The \emph{C++ program} for Alg-J easily solves the large data cases where $n=10^7$, whereas the \emph{C++ program} for Alg-CM fails even if $n=1000$ due to accuracy problems of floating point operations.
This shows the fact that Alg-J is much more robust.
\end{description}
More importantly, the Rotate-and-Kill technique developed in this paper (summarized in Subsection~\ref{sect:techover}) is powerful in that it also solves other polygon inclusion or circumscribing problems.
For one example, the following dual problem only admits $O(n\log n)$ time solutions in literature (\cite{kgon94,kgon95soda}) but can be solved in linear time by our new approach:
Given $n$ half-planes which constitute a convex polygon, find three of them whose intersecting area is minimum. This is reported in a subsequent paper \cite{Jin17b}.
For another example, in the second part of this paper (Appendix~\ref{sect:extended}), we use the new technique to compute all the generally 3-stable triangles, i.e.\ those triangles
whose corners lie in $P$'s boundary (rather than $P$'s vertices) and whose area cannot be improved through adjusting one corner (which include all the 3-stable triangles).
This further leads to a new linear time solution for computing the minimum area triangle enclosing $P$.
Thus the algorithm in this paper and the one in \cite{linear-correct} can both compute the maximum triangle in $P$ and the minimum triangle enclosing $P$ simultaneously.
These examples suggest that more efficient algorithms may exist for other related problems potentially, as discussed in Subsection~\ref{subsect:CM} below,
and we hope that this paper has brought incentives for revisiting these problems.
Note that, as elucidated in \cite{Jin17b}, some extra efforts seem necessary in applying our technique to any specific problem.
\subsection{Previous known algorithm given in \cite{linear-correct} and other related work}\label{subsect:CM}
We now briefly introduce Chandran and Mount's method. They first define the \emph{$P$-stable} triangles (outside $P$).
All sides of a $P$-stable triangle must be touched by $P$.
In particular, two of them must have their midpoints touched by $P$; these two are called the \emph{legs} whereas the remaining one is called the \emph{base}.
Moreover, one of the following holds. (1) The base is flushed with $P$ (namely, it contains an edge of $P$).
(2) one of the legs is flushed with an edge of $P$ and has as its midpoint a vertex of this edge.
Note that if (1) holds, it is said \emph{$P$-anchored}.
The $P$-anchored triangles are in fact introduced in \cite{Tri-Enclose-Area-nlogn2},
where it is proved that they contain all the local minimums of enclosing triangles of $P$.
It is then proved in \cite{Tri-Enclose-Area} that all the $P$-anchored triangles are ``interspersing'' -- if we move the base to the next edge, the midpoints of two legs will both move in clockwise.
Then, by using the rotating-caliper technique \cite{rotatingcaliper} together with some clever algorithmic tricks,
\cite{Tri-Enclose-Area} enumerates all the $P$-anchored triangles in linear time and thus find the minimum enclosing triangles.
The algorithm is somewhat involved and description is only highlighted, so details on implementation is given in \cite{Tri-Enclose-Area-Imp}.
Later, by using some ``more involved'' (as said in the paper) observations and techniques,
\cite{linear-correct} managed to enumerate all the $P$-stable triangles in linear time. Their algorithm is actually a terrific example of the rotating-caliper technique.
Moreover, it was proved that by enumerating all the $P$-stable triangles, one can easily obtain all local maximums of triangles enclosed in $P$.
To sum up, the known algorithm in \cite{linear-correct} indeed computes all the maximum area triangles enclosed in $P$ and all the minimum area triangles enclosing $P$ simultaneously.
\smallskip The problems of searching for extremal figures with special properties inside or outside a polygon were initiated in \cite{linear-wrong-DS, kgon82, kgonE-area-FOCS84},
and have since been studied extensively.
The minimum area triangle enclosing a convex polygon can be found in $O(n\log^2n)$ time \cite{Tri-Enclose-Area-nlogn2} or even in $O(n)$ time \cite{Tri-Enclose-Area}.
The minimum perimeter triangle enclosing a convex polygon is solved in $O(n)$ time \cite{Tri-Enclose-Peri}.
The maximum perimeter enclosed triangle can be solved in $O(n\log n)$ time \cite{kgon82}.
It is interesting to know whether this can be optimized to linear time.
At the same time when this paper was written, \cite{Kallus17b} presented an alternative linear time algorithm for computing the maximum area triangles (which looks more complicated than ours).
Moreover, \cite{kgon82, kgon87, kgon94, kgonE-area-FOCS84, kgonE-area-VC85, kgonE-peri-IPL08} studied extremal area / perimeter $k$-gon inside or outside a convex polygon. The maximum $k$-gon can be computed in $O(n\log n)$ time when $k$ is a constant and it is interesting to know whether this can be optimized to linear time (at least for $k=4$).
Furthermore, \cite{3d-Suri-JC02, 3d-Vivien-CGTA04} asked and solved the extremal polytope problems in three dimensional space.
Besides, \cite{shortestpathforGO} considered maximum triangles inside a simple polygon.
Instead of $k$-gon, people care about extremal figures that admit more geometric properties, such as equilateral triangle and squares, rectangles and parallelograms, disks and ellipses.
In particular, the largest-area parallelogram is studied by the same author of this paper in \cite{JinM11,Jin15},
who gave a nontrivial algorithm for computing all locally maximal parallelograms in a convex polygon in $O(n\log^2n)$ time. See \cite{Jin15} for more introductions in this area.
De Pano, Ke and O'Rourke \cite{YanKe87square} pointed out that finding extremal polygons inside a given polygon is usually more difficult than finding extremal polygons outside a given polygon.
The latter problems can usually be solved by the rotating-caliper technique.
\smallskip To be more complete, some properties of figures inscribed or circumscribed in convex polygons or curves can be found in \cite{Math-number-tri-DCG01, Math-number-tri-CTA09, Math-MMD-AMS47, Math-Levi-AMM02, Math-Miernowski-JG07, Math-Lassak-AG} and the references within.
\subsection{Preliminary}
Let $v_1,\ldots,v_n$ be a clockwise enumeration of the vertices of $P$. Assume that no three vertices lie in the same line.
Let $e_1,\ldots,e_n$ denote the $n$ edges of $P$, so that $e_i$ is the directed line segment $\overrightarrow{v_iv_{i+1}}$.
Denote the boundary of $P$ by $\partial P$.
For convenience, when $A$ denotes a vertex of $P$, we use notation $A-1$ and $A+1$ to denote the clockwise previous and next vertex of $A$ respectively.
Given two distinct points $A,B$, let $\overleftrightarrow{AB}$ denote the line connecting them.
\textbf{For any triangle that forms by points in $\partial P$, we list its corners in clockwise order;
when $\triangle ABC$ is written, we always assume $A,B,C$ lie in clockwise order}.
Consider any triangle $\triangle ABC$ in which all corners lie in $\partial P$.
Note that corner $A$ lies on the right of $\overrightarrow{BC}$. We say $A$ is \emph{stable} if it has the largest distance to $\overleftrightarrow{BC}$ among all points of $P$ that lie on the right of $\overrightarrow{BC}$.
Similarly, corner $B$ is \emph{stable} if it has the largest distance to $\overleftrightarrow{CA}$ among all points of $P$ that lie on the right of $\overrightarrow{CA}$.
Corner $C$ is \emph{stable} if it has the largest distance to $\overleftrightarrow{AB}$ among all points of $P$ that lie on the right of $\overrightarrow{AB}$.
The triangle $\triangle ABC$ is said \emph{3-stable}, if all its corners are stable and lie on the vertices of $P$.
\begin{definition}[\cite{kgon82,linear-open-arxiv}]\label{def:interleave}
Let $[A\circlearrowright B]$ denote the portion of $\partial P$ starting from point $A$ and clockwise to $B$.
Assume $\triangle A_1A_2A_3,\triangle B_1B_2B_3$ have all their corners lying on $P$'s boundary.
They are \emph{interleaving} if there exists at least one corner of $B$ lying in $[A_i\circlearrowright A_{i+1}]$ for each $i\in \{1,2,3\}$ and
there exists at least one corner of $A$ lying in $[B_i\circlearrowright B_{i+1}]$ for each $i\in \{1,2,3\}$.
\end{definition}
\begin{lemma}[\cite{linear-open-arxiv}]\label{lemma:interleaving}
Any two 3-stable triangles in a convex polygon are interleaving.
\end{lemma}
\begin{corollary}[\cite{linear-open-arxiv}]
There are $O(n)$ 3-stable triangles.
\end{corollary}
\subsection{Technique overview}\label{sect:techover}
We start by computing one 3-stable triangle using an ordinary method (see Section~\ref{sect:alg-one}); denote the resulting 3-stable triangle by $\triangle v_rv_sv_t$.
By Lemma~\ref{lemma:interleaving}, each 3-stable triangle interleaves $\triangle v_rv_sv_t$,
and hence has an edge $BC$ such that $B\in \{v_s,\ldots,v_t\}$ whereas $C\in \{v_t,\ldots,v_r\}$.
We then compute all the 3-stable triangles in a process called \textbf{Rotate-and-Kill}.
Initially, set two pointers $(B,C)=(v_s,v_t)$.
In each iteration, we first compute $A$ so that $A$ has the largest distance to $\overleftrightarrow{BC}$ among vertices of $P$ on the right of $\overrightarrow{BC}$.
This only costs amortized $O(1)$ time because the slope of $BC$ keeps decreasing and so $A$ goes only in clockwise direction.
Next, check whether $\triangle ABC$ is 3-stable and report it if so,
and go to the next iteration by either killing $B$ (i.e.\ moving pointer $B$ to its next vertex) or killing $C$ (i.e.\ moving pointer $C$ to its next vertex).
The \textbf{criterion} for killing $B$ or $C$ obeys two important rules:
$B$ is killed only when (1) its related pairs $(B,{C+1}),(B,{C+2}),...,(B,v_r)$ cannot form an edge of any 3-stable triangle; and
$C$ is killed only when (2) its related pairs $(B+1,C),(B+2,C),...(v_t,C)$ cannot form an edge of any 3-stable triangle.
In this way, noticing that $(v_t,v_r)$ forms an edge of the 3-stable triangle $\triangle v_sv_tv_r$,
we make sure that $(B,C)$ eventually reaches $(v_t,v_r)$ and that all the 3-stable triangles are going to be reported before that moment.
The core of our technique lies in designing an efficient criterion for killing.
We may attempt either of the following criterions at first: 1. Determine (1) directly and kill $B$ when (1) holds and $C$ otherwise. 2. Determine (2) directly and kill $C$ when (2) holds and $B$ otherwise.
Notice that at least one of (1) and (2) holds in every iteration of $(B,C)$ (this is an easy corollary of Lemma~\ref{lemma:interleaving}; see the details in Observation~\ref{obs:criterion_exists} and its proof below),
so both criterions satisfy the aforementioned rules.
However, determining (1) or (2) directly would cost $O(n)$ time by trivial methods or $O(\log n)$ time by binary searches.
For a linear running time of the Rotate-and-Kill process, we must find a better criterion that can be computed in $O(1)$ time.
Such a criterion is presented in Section~\ref{sect:Rotate-and-Kill}, which is surprisingly simple.
We see the Rotate-and-Kill technique is distinct from the rotating-caliper technique \cite{rotatingcaliper}.
Instead, it can be regarded as a corrective version of Dobkin and Snyder's technique \cite{linear-wrong-DS}.
\section{Compute one 3-stable triangle}\label{sect:alg-one}
In this section, we show how to compute one 3-stable triangle.\footnote{Not only one previous reviewers mistakenly claimed that our algorithm given in this section is the same as the algorithm given in \cite{linear-wrong-DS}. In fact, Step~2 of our algorithm is the same as the kernel step of the previous algorithm (by coincidence), but Step~1 will set the initial value differently -- we will start from $(B_0,C_0)$ so that $A_0B_0C_0$ is the largest triangle rooted at $A_0$ but their algorithm starts from $(B',C')$ where $(B',C')$ is the first pair of 2-stable corners with respect to $A_0$. So the two algorithms are different.}
We say a triangle is \emph{rooted} at a vertex if one of its corner lies on that vertex.
\smallskip \noindent \textbf{Step~1.}
Choose $A$ to be an arbitrary vertex of $P$; say $A=v_1$.
Find $B,C$ so that $ABC$ is the largest triangle rooted at $A$.
This can be done in $O(n)$ time.
We can enumerate a vertex $B$ and maintain $C_B$ in amortized $O(1)$ time so that $C_B$ is the vertex with the largest distance to $\overleftrightarrow{AB}$ among all vertices on the right of $\overrightarrow{AB}$.
Then, select $B$ so that $ABC_B$ is maximum.
\medskip Since $ABC$ is the largest triangle rooted at $A$, corners $B,C$ are stable.
Moreover, if corner $A$ is also stable, $\triangle ABC$ is 3-stable and we can go to the next section.
Now, assume that $A$ is not stable.
Moreover, assume that $A+1$ is further than $A$ in the distance to $\overleftrightarrow{BC}$.
Otherwise $A-1$ is further than $A$ and it is symmetric.
\medskip \noindent \textbf{Step~2.}
This step is presented in Algorithm~\ref{alg:I2}.
\begin{algorithm}[h]
\caption{Algorithm for Step 2}\label{alg:I2}
$A\leftarrow A+1$;\\
\Repeat{$(B+1)\leq B$ (to $\overleftrightarrow{AC}$) and $(C+1)\leq C$ (to $\overleftrightarrow{AB}$)}{
\While{($(B+1)>B$ in the distance to $\overleftrightarrow{AC}$)}{$B\leftarrow B+1$;}
\While{($(C+1)>C$ in the distance to $\overleftrightarrow{AB}$)}{$C\leftarrow C+1$;}
}
\end{algorithm}
To distinguish, we denote the value of the three pointers $(A,B,C)$ at the end phase of this algorithm by $(A_1,B_1,C_1)$;
and the value at the beginning phase by $(A_0,B_0,C_0)$.
\begin{figure}[h]
\centering \includegraphics[width=.9\textwidth]{first3s}\\
\caption{Illustration of Step 2 and Step 3}\label{fig:I2}
\end{figure}
\begin{observation}\label{obs:BC-stable}
\begin{enumerate}
\item Corners $B_1,C_1$ are stable in $\triangle A_1B_1C_1$.
\item $Area(\triangle A_1B_1C_1)>Area(\triangle A_0B_1C_1)$.
\end{enumerate}
\end{observation}
\begin{proof}
1. It reduces to prove the following arguments.
\begin{enumerate}
\item[(1)] $B_1$ is at least as far as $B_1-1$ in the distance to $\overleftrightarrow{A_1C_1}$.
\item[(2)] $B_1$ is at least as far as $B_1+1$ in the distance to $\overleftrightarrow{A_1C_1}$.
\item[(3)] $C_1$ is at least as far as $C_1-1$ in the distance to $\overleftrightarrow{A_1B_1}$.
\item[(4)] $C_1$ is at least as far as $C_1+1$ in the distance to $\overleftrightarrow{A_1B_1}$.
\end{enumerate}
Due to the termination condition of the algorithm, (2) and (4) hold.
In the following we point out two facts that hold throughout the algorithm, which respectively imply (1) and (3).
\begin{enumerate}
\item[(i)] $B$ is at least as far as $B-1$ in the distance to $\overleftrightarrow{AC}$.
\item[(ii)] $C$ is at least as far as $C-1$ in the distance to $\overleftrightarrow{AB}$.
\end{enumerate}
These facts hold at the beginning, because $B_0,C_0$ are stable in $A_0B_0C_0$.
When $A,B$ or $C$ are increased by 1 (at Line 1,4,7), these facts still hold as illustrated in Figure~\ref{fig:I2}~(b),(c),(d).
\smallskip \noindent 2. See Figure~\ref{fig:I2}~(a). By the assumption right above ``Step~2'', $A_1=A_0+1$ is further than $A_0$ in the distance to $\overleftrightarrow{B_0C_0}$.
Comparing the slopes of $\overleftrightarrow{B_1C_1}$ and $\overleftrightarrow{B_0C_0}$,
it implies that $A_1$ is further than $A_0$ in the distance to $\overleftrightarrow{B_1C_1}$, which is another way of saying Claim~2.
\end{proof}
\noindent \textbf{Step~3.}
So far, we obtain another triangle $\triangle A_1B_1C_1$ where $B_1,C_1$ are stable (according to Observation~\ref{obs:BC-stable}.1).
If $A_1$ is also stable, $\triangle A_1B_1C_1$ is 3-stable and we proceed to next section.
Now, consider the case where $A_1$ is not stable.
Since $A_1$ is not stable, applying Observation~\ref{obs:BC-stable}.2, vertex $A_1+1$, rather than $A_1-1$, must be further than $A_1$ in the distance to $\overleftrightarrow{B_1C_1}$.
We call Algorithm~\ref{alg:I2} once again with initial value $(A_1,B_1,C_1)$ and terminal value $(A_2,B_2,C_2)$,
and repeat such a process until $\triangle A_iB_iC_i$ is 3-stable for some integer $i$.
\subparagraph*{Correctness and running time.}
At every change of $A,B,C$ in Algorithm~\ref{alg:I2}, the area of $\triangle ABC$ increases.
So the above algorithm terminates eventually.
Because pointers $A,B,C$ can only move in the clockwise direction and pointer $A$ cannot return to $A_0$ (otherwise we obtain a triangle rooted at $A_0$ with a larger area than $A_0B_0C_0$, which contradicts the assumption of $A_0B_0C_0$), and because $A,B,C$ always lie in clockwise order, the total number of changes of $A,B,C$ is bounded by $O(n)$ and hence the algorithm runs in $O(n)$ time.
\section{Compute all the 3-stable triangles}
In this section, we assume $\triangle v_rv_sv_t$ is 3-stable and we compute all 3-stable triangles by a Rotate-and-Kill process as mentioned in Subsection~\ref{sect:techover}.
We start by an observation.
A vertex pair $(v_j,v_k)$ is \textbf{\emph{dead}} if it cannot form an edge of any 3-stable triangle;
precisely, if there is no vertex $v_i$ such that $\triangle v_iv_jv_k$ is 3-stable and $v_i,v_j,v_k$ lie in clockwise order.
\begin{observation}\label{obs:criterion_exists}
Assume $B\in \{v_s,\ldots, v_t\}$, $C\in \{v_t, \ldots, v_r\}$, and $B\neq C$.
Then,
\begin{enumerate}
\item[(1)] $(B,C+1),(B,C+2),\ldots,(B,v_r)$ are all dead; or
\item[(2)] $(B+1,C),(B+2,C),\ldots,(v_t,C)$ are all dead.
\end{enumerate}
\end{observation}
\begin{proof}
Suppose that (1) and (2) are both false.
Without loss of generality, assume that $(B,C')$ and $(B',C)$ are not dead,
where $C'\in \{C+1,\ldots, v_r\}$ and $B'\in \{B+1,\ldots,v_t\}$.
This implies that there exist $A_1,A_2$ so that $\triangle A_1BC'$ and $\triangle A_2B'C$ are 3-stable.
Because $C'\in \{C+1,\ldots, v_r\}$ whereas $B'\in \{B+1,\ldots,v_t\}$, triangle $\triangle A_1BC'$ does not interleave $\triangle A_2B'C$.
So at least one of them is not 3-stable due to Lemma~\ref{lemma:interleaving}. Contradictory.
\end{proof}
\begin{figure}[b]
\centering \includegraphics[width=.95\textwidth]{QIJKDef}\\
\caption{Definition of the area $Q_{j,k}$.}\label{fig:QIJKDef}
\end{figure}
\begin{definition}\label{def:Q}
Assume $v_j,v_k$ are distinct vertices of $P$. See Figure~\ref{fig:QIJKDef}.
Let $r_1$ (respectively, $r_2$) denote the ray originating from $v_j$ with the same direction as $e_{k-1}$ (respectively, $e_k$).
Let $r'_1$ (respectively, $r'_2$) denote the ray originating from $v_k$ with the opposite direction to $e_{j-1}$ (respectively, $e_j$).
(Here, recall that each edge is a directed line segment; specifically, the direction of $e_i$ is from $v_i$ to $v_{i+1}$.)
These four rays $r_1,r_2,r'_1,r'_2$ together define a region, denoted by $Q_{j,k}$.
The intersecting point between $r_1,r'_1$ is denoted by $H_{j,k}$.
The intersecting point between $r_2,r'_2$ is denoted by $I_{j,k}$.
The intersecting point between $r_1,r'_2$ is denoted by $J_{j,k}$.
The intersecting point between $r_2,r'_1$ is denoted by $K_{j,k}$.
\textbf{Note}: When $I_{j,k}$ (or $H_{j,k},J_{j,k},K_{j,k}$) is undefined in the above (when the corresponding two rays do not intersect), we define $I_{j,k}=\infty$ (or $H_{j,k}=\infty,J_{j,k}=\infty,K_{j,k}=\infty$, respectively) and assume that it has an infinite distance to $\overline{v_jv_k}$.
We consider that the region $Q_{j,k}$ contains its boundary.
We consider that polygon $P$ also contains its boundary.
\end{definition}
Roughly speaking, $Q_{j,k}$ indicates the potential region of a corner of a 3-stable triangle when the other two corners are fixed in $v_j$ and $v_k$. This is made precisely as follows.
\begin{observation}\label{obs:Q}
Suppose $v_j,v_k$ are stable in $\triangle v_iv_jv_k$. Then, $v_i$ lies in $Q_{j,k}$.\\
(Recall that whenever we write $\triangle v_iv_jv_k$, we assume that $v_i,v_j,v_k$ lie in clockwise order.)
\end{observation}
\begin{proof}
See Figure~\ref{fig:QIJKDef}.
Because $v_j$ is stable, $v_i$ can only lie in the sector area bounded by $r'_1$ and $r'_2$.
Because $v_k$ is stable, $v_i$ can only lie in the sector area bounded by $r_1$ and $r_2$.
Together, $v_i$ lies in the intersection of these two sector areas, which is defined as $Q_{j,k}$.
\end{proof}
\begin{definition}\label{def:legal}
We regard vertex pair $(v_j,v_k)$ \textbf{\emph{illegal}} if $v_{k+1}$ is closer than $v_k$ in the distance to $\overleftrightarrow{v_jv_{j+1}}$ (as shown in Figure~\ref{fig:legalDef}.a)
and regard $(v_j,v_k)$ \textbf{\emph{legal}} otherwise (see Figure~\ref{fig:legalDef}.b).
Note: For each vertex $v_j$ of $P$, the sequence of vertex pairs $(v_j,v_{j+1}),(v_j,v_{j+2}),\ldots,(v_j,v_{j-1})$ starts with several legal pairs and then is followed by several illegal pairs.
In particular, $(v_j,v_{j+1})$ is legal and $(v_j,v_{j-1})$ is illegal. (All subscripts taken modulo $n$.)
\end{definition}
\begin{figure}[h]
\centering \includegraphics[width=\textwidth]{illegalDef}\\
\caption{Illustration of the notion of ``legal pair''.}\label{fig:legalDef}
\end{figure}
\begin{observation}[Sufficient conditions for determining that $(v_j,v_k)$ is dead]\label{obs:VV-dead}~
\begin{enumerate}
\item If $Q_{j,k}$ does not intersect $P$, then $(v_j,v_k)$ is dead.
\item When $(v_j,v_k)$ is illegal, $Q_{j,k}$ does not intersect $P$ and hence is dead.
\item If there exists some point $A$ in $P$ which lies on the right of $\overrightarrow{v_jv_k}$ and is further than $K_{j,k}$ in the distance to $\overleftrightarrow{v_jv_k}$,
then we can determine that $(v_j,v_k)$ is dead.
\end{enumerate}
\end{observation}
\begin{proof}
1. Suppose to the opposite $(v_j,v_k)$ is not dead, then there exists $v_i$ such that $\triangle v_iv_jv_k$ is 3-stable.
By Observation~\ref{obs:Q}, $Q_{j,k}$ contains $v_i$ and hence intersects $P$.
2. Recall the two rays $r'_1,r'_2$ in Definition~\ref{def:Q}.
When $(v_j,v_k)$ is illegal, the area bounded by $r'_1,r'_2$ (except apex $v_k$) lies outside $P$ (see Figure~\ref{fig:legalDef}~(c)).
So, the subregion $Q_{j,k}$ of this area cannot intersect $P$ on any position other than $v_k$.
Moreover, since $(v_j,v_{j+1})$ is always legal, we know $k\neq j+1$ when $(v_j,v_k)$ is illegal, and thus $v_k$ does not lie in $Q_{j,k}$ (see Figure~\ref{fig:QIJKDef}).
Together, $Q_{j,k}$ has no intersection with $P$, hence $(v_j,v_k)$ is dead according to Claim~1.
3. Suppose to the opposite that $(v_j,v_k)$ is not dead but $\triangle v_iv_jv_k$ is 3-stable for some vertex $v_i$.
By Observation~\ref{obs:Q}, $v_i$ is contained in $Q_{j,k}$.
Further applying the definition of $Q_{j,k}$ (see Figure~\ref{fig:QIJKDef}~(a)), this means $v_i$ is not further than $K_{j,k}$ in the distance to $\overleftrightarrow{v_jv_k}$.
Therefore, $A$ is further than $v_i$ in the distance to $\overleftrightarrow{v_jv_k}$, thus $v_i$ is not stable in $\triangle v_iv_jv_k$. Contradictory.
\end{proof}
\subsection{Rotate-and-Kill process}\label{sect:Rotate-and-Kill}
Recall the Rotate-and-Kill process briefly discussed in Section~\ref{sect:techover}.
We present its details below in Algorithm~\ref{alg:r-and-k} and demonstrate it by a concrete example in Figure~\ref{fig:example}.
Its running time is already analyzed in Section~\ref{sect:techover} and its correctness is assured by the next lemma.
\begin{algorithm}[h]
\caption{The Rotate-and-Kill process}\label{alg:r-and-k}
\KwIn{$r,s,t$ such that $\triangle v_rv_sv_t$ is 3-stable.}
$(A,B,C)\leftarrow (v_r,v_s,v_t)$;\\
\Repeat{$(B,C)=(v_t,v_r)$}
{
\While{$(A+1)>A$ in the distance to $\overleftrightarrow{BC}$ \label{code:A++}}
{$A\leftarrow A+1$;}
Output $ABC$ if it is 3-stable;\qquad Output $(A+1)BC$ if it is 3-stable; \label{code:output}\\
\eIf{$A>I_{B,C}$ in the distance to $\overleftrightarrow{BC}$}
{$C\leftarrow C+1$; \label{code:C++Key} }
{$B\leftarrow B+1$; \label{code:B++Key} }
}
\end{algorithm}
\begin{figure}[h]
\centering \includegraphics[width=\textwidth]{example}\\ \smallskip
\centering \includegraphics[width=\textwidth]{example-table}\\
\caption{Example of the Rotate-and-Kill process.}\label{fig:example}
\end{figure}
\begin{lemma}\label{lemma:correct}
\begin{enumerate}
\item When vertex $B$ is about to be killed in Algorithm~\ref{alg:r-and-k} (that means we are about to run Line~\ref{code:B++Key}), $(B,C+1),(B,C+2),\ldots,(B,v_r)$ are dead.
\item When vertex $C$ is about to be killed in Algorithm~\ref{alg:r-and-k} (that means we are about to run Line~\ref{code:C++Key}), $(B+1,C),(B+2,C),\ldots,(v_t,C)$ are dead.
\item Throughout the ``repeat'' statement, $B\in \{v_s,\ldots, v_t\}$, $C\in \{v_t, \ldots, v_r\}$, and $B\neq C$.
Moreover, the ``repeat'' statement will terminate successfully at $(B,C)=(v_t,v_r)$.
\item Altogether, every 3-stable triangle will be computed by Algorithm~\ref{alg:r-and-k}.
\end{enumerate}
\end{lemma}
\begin{proof}
3. For ease of understanding, we first prove Claim~3, assuming that Claim~1,2 are correct.
Notice that $(v_t,v_r)$ forms an edge of a 3-stable triangle (that is $\triangle v_rv_sv_t$).
Therefore, Claim~1 and 2 imply that $B$ would \textbf{not} be killed when $(B=v_t,C<v_r)$, and $C$ would \textbf{not} be killed when $(B<v_t,C=v_r)$.
So $(B,C)$ eventually become $(v_t,v_r)$, at which moment the algorithm terminates, and we have $B\in \{v_s,\ldots, v_t\}$ and $C\in \{v_t, \ldots, v_r\}$ throughout.
In addition, we argue that $B\neq C$ throughout. Suppose $B=C$ in some iteration. It must be $B=C=v_t$.
This means that we have killed $B$ at the previous iteration $(B,C)=(v_{t-1},v_t)$.
But this is impossible by the following analysis.
When $(B,C)=(v_{t-1},v_t)$, we know $C=B+1$, and hence $I_{B,C}=B$, which further implies that ``$A>I_{B,C}$ in the distance to $\overleftrightarrow{BC}$''.
So we would enter Line~\ref{code:C++Key} and kill $C$ rather than enter Line~\ref{code:B++Key} and kill $B$ at that iteration.
\medskip \noindent 1. Assume we are at Line~\ref{code:B++Key}, and so $A\leq I_{B,C}$ in the distance to $\overleftrightarrow{BC}$.
Further assume that \emph{$C$ is closer than $C+1$ in the distance to $\overleftrightarrow{B(B+1)}$}, as shown in Figure~\ref{fig:march-VV1}.
Otherwise, $(B,C+1),...,(B,v_r)$ are all illegal and hence are dead (by Observation~\ref{obs:VV-dead}).
Let $l$ denote the unique line at $I_{B,C}$ that is parallel to $\overleftrightarrow{BC}$.
Let $C^*$ denote the last vertex in the sequence $C+1,\ldots,v_r$ so that $(B,C^*)$ is legal. We first state two facts.
\begin{enumerate}
\item[a.] Polygon $P$ lies in the (closed) half-plane delimited by $l$ and containing $BC$.
\item[b.] $Q_{B,C+1},\ldots,Q_{B,C^*}$ lie in the (open) half-plane delimited by $l$ and not containing $BC$.
\end{enumerate}
\begin{figure}[h]
\centering \includegraphics[width=.6\textwidth]{march-VV1}\\
\caption{Illustration for the correctness on Line~\ref{code:B++Key}.}\label{fig:march-VV1}
\end{figure}
Combining Fact a and b, no region in $\{Q_{B,C+1},\ldots,Q_{B,C^*}\}$ intersects $P$.
So $(B,C+1),\ldots,(B,C^*)$ are dead according to Observation~\ref{obs:VV-dead}.1.
Moreover, $\{(B,C^*+1),\ldots,(B,v_r)\}$ are dead because of illegal (by Observation~\ref{obs:VV-dead}.2).
Together, Claim~1 holds.
\smallskip \noindent \emph{Proof of a}.
By the assumption, we know $A\leq I_{B,C}$ in the distance to $\overleftrightarrow{BC}$.
Moreover, $A$ has the largest distance among all vertices of $P$ that lie on the right of $\overrightarrow{BC}$ (this is is guaranteed by the ``while-do'' sentence in Algorithm~\ref{alg:r-and-k}).
Together, we get claim~a.
\smallskip \noindent \emph{Proof of b}.
Here, we define a region $R_{j,k}$ for each legal pair $(v_j,v_k)$ so that $k\neq j+1$.
Recall $H_{j,k},I_{j,k},J_{j,k}$ in Definition~\ref{def:Q}. See Figure~\ref{fig:QIJKDef}~(a) and (b).
The unique quadrant bounded by $\overrightarrow{J_{j,k}I_{j,k}},\overrightarrow{J_{j,k}H_{j,k}}$ and containing $Q_{j,k}$ is defined as $R_{j,k}$.
Observe Figure~\ref{fig:march-VV1}~(b), we have
$$R_{B,C+1}\supseteq R_{B,C+2}\supseteq \ldots \supseteq R_{B,C^*}.$$
It implies that $Q_{B,C+1},Q_{B,C+2},\ldots,Q_{B,C^*}$ lie in $R_{B,C+1}$.
Now, it reduces to show that $R_{B,C+1}$ lies entirely in the (open) half-plane delimited by $l$ and not containing $BC$ (see Figure~\ref{fig:march-VV1}~(a)).
This holds because apex $J_{B,C+1}$ of $R_{B,C+1}$ lies in that half-plane, which is because $\overrightarrow{I_{B,C}J_{B,C+1}}$ is a translate of $\overrightarrow{C(C+1)}$
and $C$ is closer than $C+1$ to $\overleftrightarrow{B(B+1)}$.
\medskip \noindent 2. Assume we are at Line~\ref{code:C++Key}, and so $A>I_{B,C}$ in the distance to $\overleftrightarrow{BC}$.
Since $A>I_{B,C}$, we have $I_{B,C}\neq \infty$, so \emph{$C$ is closer than $C+1$ in the distance to $\overleftrightarrow{BB+1}$}, as shown in Figure~\ref{fig:march-VV2}.
Let $D$ denote the intersection between $\overleftrightarrow{C(C+1)}$ and $l$.
Let $d_{B,C}(X)$ denote the distance from $X$ to $\overleftrightarrow{BC}$ for any vertex pair $(B,C)$.
See Figure~\ref{fig:march-VV2}~(a). We claim two inequalities:
\begin{enumerate}
\item[I] $d_{B+1,C}(D)<d_{B+1,C}(A)$.
\item[II] $d_{B+1,C}(K_{B+1,C})=d_{B+1,C}(D)$.
\end{enumerate}
\begin{figure}[h]
\centering \includegraphics[width=.8\textwidth]{march-VV2}\\
\caption{Illustration for the correctness on Line~\ref{code:C++Key}.}\label{fig:march-VV2}
\end{figure}
Together, $d_{B+1,C}(A)>d_{B+1,C}(K_{B+1,C})$, thus $(B+1,C)$ is dead by Observation~\ref{obs:VV-dead}.3.
\smallskip \noindent \emph{Proof of I}. Since $A>I_{B,C}$ in the distance to $\overleftrightarrow{BC}$, it lies in the (open) half-plane delimited by $l$ and not containing $BC$.
Moreover, since $P$ is convex, $A$ lies in the (closed) half-plane delimited by $\overleftrightarrow{CC+1}$ and containing $P$. Together, we obtain (I).
\noindent \emph{Proof of II}. Obviously, both segment $(B+1)K_{B+1,C}$ and $CD$ are translates of $BI_{B,C}$. So, $(B+1)K_{B+1,C}DC$ is a parallelogram. This implies (II).
\smallskip Next, we argue that $(B+2,C)$ is dead. This reduces to prove the following:
\begin{enumerate}
\item[(III)] $d_{B+2,C}(A)>d_{B+2,C}(K_{B+2,C})$.
\end{enumerate}
(III) is implied by the above inequality $d_{B+1,C}(A)>d_{B+1,C}(K_{B+1,C})$. This is simply illustrated in Figure~\ref{fig:march-VV2}~(b). The proof is similar to the proofs of (I) and (II) and is omitted.
Furthermore, by induction, $(B+1,C),(B+2,C),\ldots (v_t,C)$ are all dead.
\medskip \noindent 4. Assume triangle $T$ is 3-stable. Since $T$ and $\triangle v_rv_sv_t$ are both 3-stable, they interleave by Lemma~\ref{lemma:interleaving}.
Without loss of generality, we can assume that $T=\triangle v_iv_jv_k$ where $v_i,v_j,v_k$ respectively lies in $\{v_r,\ldots,v_s\}$, $\{v_s,\ldots,v_t\}$, $\{v_t,\ldots,v_r\}$.
Since $\triangle v_iv_jv_k$ is 3-stable, the vertex pair $(v_j,v_k)$ is not dead.
Applying Claim~1, 2 and 3 altogether, this means we will enter an iteration in which $(B,C)=(v_j,v_k)$.
Let $A^*$ denote the value of $A$ after the first ``while-do'' statement (which updates $A$ to be the vertex with the largest distance to $\overleftrightarrow{BC}$) in this iteration.
Because $v_i$ is stable in $T$, we know $v_i$ has the largest distance to $\overleftrightarrow{v_jv_k}=\overleftrightarrow{BC}$,
which implies that $v_i\in \{A^*,A^*+1\}$. Therefore, $\triangle v_iv_jv_k$ will be outputted at Line~\ref{code:output}.
\end{proof}
\subparagraph*{Experiments.} The implementations of our algorithm and the algorithm of \cite{linear-correct} are given in a zip file at \cite{implementation}.
The experiment results are demonstrated in Appendix~\ref{sect:experiments}.
\subparagraph*{Extension.} We introduce generally 3-stable triangles and design a linear time algorithm for computing them in Appendix~\ref{sect:extended}.
This further algorithms leads to a linear time algorithm for computing the minimum area enclosing triangle as shown in Appendix~\ref{sect:application}.
\paragraph*{Acknowledgement} The author thank Professor Zhiyi Huang for discussion.
We also thank the authors of \cite{linear-open-arxiv} for giving us the opportunity to revisit such an interesting and classic problem.
\bibliographystyle{plainurl}
|
\section*{Figure Captions\markboth
{FIGURECAPTIONS}{FIGURECAPTIONS}}\list
{Figure \arabic{enumi}:\hfill}{\settowidth\labelwidth{Figure
999:}
\leftmargin\labelwidth
\advance\leftmargin\labelsep\usecounter{enumi}}}
\let\endfigcap\endlist \relax
\def\tablecap{\section*{Table Captions\markboth
{TABLECAPTIONS}{TABLECAPTIONS}}\list
{Table \arabic{enumi}:\hfill}{\settowidth\labelwidth{Table
999:}
\leftmargin\labelwidth
\advance\leftmargin\labelsep\usecounter{enumi}}}
\let\endtablecap\endlist \relax
\def\reflist{\section*{References\markboth
{REFLIST}{REFLIST}}\list
{[\arabic{enumi}]\hfill}{\settowidth\labelwidth{[999]}
\leftmargin\labelwidth
\advance\leftmargin\labelsep\usecounter{enumi}}}
\let\endreflist\endlist \relax
\def\list{}{\rightmargin\leftmargin}\item[]{\list{}{\rightmargin\leftmargin}\item[]}
\let\endquote=\endlist
\makeatletter
\newcounter{pubctr}
\def\@ifnextchar[{\@publist}{\@@publist}{\@ifnextchar[{\@publist}{\@@publist}}
\def\@publist[#1]{\list
{[\arabic{pubctr}]\hfill}{\settowidth\labelwidth{[999]}
\leftmargin\labelwidth
\advance\leftmargin\labelsep
\@nmbrlisttrue\def\@listctr{pubctr}
\setcounter{pubctr}{#1}\addtocounter{pubctr}{-1}}}
\def\@@publist{\list
{[\arabic{pubctr}]\hfill}{\settowidth\labelwidth{[999]}
\leftmargin\labelwidth
\advance\leftmargin\labelsep
\@nmbrlisttrue\def\@listctr{pubctr}}}
\let\endpublist\endlist \relax
\makeatother
\newskip\humongous \humongous=0pt plus 1000pt minus 1000pt
\def\mathsurround=0pt{\mathsurround=0pt}
\def\eqalign#1{\,\vcenter{\openup1\jot \mathsurround=0pt
\ialign{\strut \hfil$\displaystyle{##}$&$
\displaystyle{{}##}$\hfil\crcr#1\crcr}}\,}
\newif\ifdtup
\def\panorama{\global\dtuptrue \openup1\jot \mathsurround=0pt
\everycr{\noalign{\ifdtup \global\dtupfalse
\vskip-\lineskiplimit \vskip\normallineskiplimit
\else \penalty\interdisplaylinepenalty \fi}}}
\def\eqalignno#1{\panorama \tabskip=\humongous
\halign to\displaywidth{\hfil$\displaystyle{##}$
\tabskip=0pt&$\displaystyle{{}##}$\hfil
\tabskip=\humongous&\llap{$##$}\tabskip=0pt
\crcr#1\crcr}}
\relax
\def\begin{equation}{\begin{equation}}
\def\end{equation}{\end{equation}}
\def\begin{eqnarray}{\begin{eqnarray}}
\def\end{eqnarray}{\end{eqnarray}}
\def\dot{\beta}{\bar{\partial}}
\def\bar{J}{\bar{J}}
\def\partial{\partial}
\def\bar{\partial}{\bar{\partial}}
\def f_{,i} { f_{,i} }
\def F_{,i} { F_{,i} }
\def f_{,u} { f_{,u} }
\def f_{,v} { f_{,v} }
\def F_{,u} { F_{,u} }
\def F_{,v} { F_{,v} }
\def A_{,u} { A_{,u} }
\def A_{,v} { A_{,v} }
\def g_{,u} { g_{,u} }
\def g_{,v} { g_{,v} }
\def\kappa{\kappa}
\def\rho{\rho}
\def\alpha} \def\dot{\alpha}{{\dot\alpha}{\alpha}
\def {\bar A} {\Alpha}
\def\beta} \def\db{{\dot\beta}{\beta}
\def\Beta{\Beta}
\def\gamma} \def\Gamma{\Gamma} \def\dot{\gamma}{{\dot\gamma}{\gamma}
\def\Gamma{\Gamma}
\def\delta} \def\Delta{\Delta} \def\dot{\delta}{\dot\delta{\delta}
\def\Delta{\Delta}
\def\epsilon} \def\vare{\varepsilon{\epsilon}
\def\Epsilon{\Epsilon}
\def\pi} \def\Pi{\Pi{\pi}
\def\Pi{\Pi}
\def\chi{\chi}
\def\Chi{\Chi}
\def\theta} \def\Theta{\Theta} \def\vth{\vartheta{\theta}
\def\Theta{\Theta}
\def\mu} \def\nu{\nu{\mu}
\def\nu{\nu}
\def\omega{\omega}
\def\Omega{\Omega}
\def\lambda} \def\Lambda{\Lambda{\lambda}
\def\Lambda{\Lambda}
\def\sigma} \def\Sigma{\Sigma{\sigma}
\def\Sigma{\Sigma}
\def\varphi{\varphi}
\def{\cal N}{{\cal N}}
\def{\cal M}{{\cal M}}
\def\tilde V{\tilde V}
\def{\cal V}{{\cal V}}
\def\tilde{\cal V}{\tilde{\cal V}}
\def{\cal L}{{\cal L}}
\def{\cal R}{{\cal R}}
\def{\cal A}{{\cal A}}
\def{\cal{G} }{{\cal{G} }}
\def{\cal D}} \def{\cal E}{{\cal E}} \def\cF{{\cal F}{{\cal{D} } }
\def\tilde{m}{\tilde{m}}
\def{\cal A}{{\cal A}} \def\cB{{\cal B}} \def\cC{{\cal C}}
\def{\cal D}} \def{\cal E}{{\cal E}} \def\cF{{\cal F}{{\cal D}} \def{\cal E}{{\cal E}} \def\cF{{\cal F}}
\def{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}}
\def{\cal J}} \def\cK{{\cal K}} \def\cL{{\cal L}{{\cal J}} \def\cK{{\cal K}} \def{\cal L}{{\cal L}}
\def{\cal M}{{\cal M}} \def{\cal N}{{\cal N}} \def\cO{{\cal O}}
\def{\cal P}} \def\cQ{{\cal Q}} \def\cR{{\cal R}{{\cal P}} \def\cQ{{\cal Q}} \def{\cal R}{{\cal R}}
\def{\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}{{\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}}
\def{\cal V}{{\cal V}} \def\cW{{\cal W}} \def\cX{{\cal X}}
\def{\cal Y}} \def\cZ{{\cal Z}}\def\cf{{\cal f}{{\cal Y}} \def\cZ{{\cal Z}}\def\cf{{\cal f}}
\def{\cal b}{{\cal b}}
\newcommand{\vev}[1]{{\left< {#1} \right>}}
\newcommand{\bra}[1]{{\left< {#1} \right|}}
\newcommand{\ket}[1]{{\left| {#1} \right>}}
\newcommand{\prt}[1]{{\left( {#1} \right)}}
\newcommand{\prtt}[1]{{\left[ {#1} \right]}}
\defSchwarzschild {Schwarzschild}
\defReissner-Nordstr\"om {Reissner-Nordstr\"om}
\defChristoffel {Christoffel}
\defMinkowski {Minkowski}
\def\bigskip{\bigskip}
\def\noindent{\noindent}
\def\hfill\break{\hfill\break}
\def\qquad{\qquad}
\def\bigl{\bigl}
\def\bigr{\bigr}
\def\overline\del{\overline\partial}
\def\relax{\rm I\kern-.18em R}{\relax{\rm I\kern-.18em R}}
\def$SL(2,\IR)_{-k'}\otimes SU(2)_k/(\IR \otimes \tilde \IR)${$SL(2,\relax{\rm I\kern-.18em R})_{-k'}\otimes SU(2)_k/(\relax{\rm I\kern-.18em R} \otimes \tilde \relax{\rm I\kern-.18em R})$}
\def Nucl. Phys. { Nucl. Phys. }
\def Phys. Lett. { Phys. Lett. }
\def Mod. Phys. Lett. { Mod. Phys. Lett. }
\def Phys. Rev. Lett. { Phys. Rev. Lett. }
\def Phys. Rev. { Phys. Rev. }
\def Ann. Phys. { Ann. Phys. }
\def Commun. Math. Phys. { Commun. Math. Phys. }
\def Int. J. Mod. Phys. { Int. J. Mod. Phys. }
\def\partial_+{\partial_+}
\def\partial_-{\partial_-}
\def\partial_{\pm}{\partial_{\pm}}
\def\partial_{\mp}{\partial_{\mp}}
\def\partial_{\tau}{\partial_{\tau}}
\def \bar \del {\bar \partial}
\def {\bar h} { {\bar h} }
\def {\bar \phi} { {\bar \phi} }
\def \zeta { {\bar z} }
\def {\bar A} { {\bar A} }
\def {\tilde {A }} { {\tilde {A }}}
\def {\tilde {\A }} { {\tilde { {\bar A} }}}
\def {\bar J} {{\bar J} }
\def {\tilde J} { {\tilde {J }}}
\def {1\over 2} {{1\over 2}}
\def {1\over 3} {{1\over 3}}
\def \over {\over}
\def\int_{\Sigma} d^2 z{\int_{\Sigma} d^2 z}
\def{\rm diag}{{\rm diag}}
\def{\rm const.}{{\rm const.}}
\def\relax{\rm I\kern-.18em R}{\relax{\rm I\kern-.18em R}}
\def\relax{\rm I\kern-.18em L}{\relax{\rm I\kern-.18em L}}
\def^{\raise.15ex\hbox{${\scriptscriptstyle -}$}\kern-.05em 1}{^{\raise.15ex\hbox{${\scriptscriptstyle -}$}\kern-.05em 1}}
\def$SL(2,\IR)\otimes SO(1,1)^{d-2}/SO(1,1)${$SL(2,\relax{\rm I\kern-.18em R})\otimes SO(1,1)^{d-2}/SO(1,1)$}
\def$SL(2,\IR)_{-k'}\otimes SU(2)_k/(\IR \otimes \tilde \IR)${$SL(2,\relax{\rm I\kern-.18em R})_{-k'}\otimes SU(2)_k/(\relax{\rm I\kern-.18em R} \otimes \tilde \relax{\rm I\kern-.18em R})$}
\def$SO(d-1,2)_{-k}/ SO(d-1,1)_{-k}${$SO(d-1,2)_{-k}/ SO(d-1,1)_{-k}$}
\def$SO(d-1,2)/ SO(d-1,1)${$SO(d-1,2)/ SO(d-1,1)$}
\def{\cal M}{{\cal M}}
\def\tilde V{\tilde V}
\def{\cal V}{{\cal V}}
\def\tilde{\cal V}{\tilde{\cal V}}
\def{\cal L}{{\cal L}}
\def{\cal R}{{\cal R}}
\def{\cal A}{{\cal A}}
\def{\cal E}{{\cal E}}
\def{\tilde X}{{\tilde X}}
\def{\tilde J}{{\tilde J}}
\def{\tilde P}{{\tilde P}}
\def{\tilde L}{{\tilde L}}
\def{\rm tr}{{\rm tr}}
\def{\rm Tr}{{\rm Tr}}
\def\begin{eqnarray}{\begin{eqnarray}}
\def\end{eqnarray}{\end{eqnarray}}
\newcommand{\eq}[1]{(\ref{#1})}
\def\nonumber{\nonumber}
\def\partial{\partial}
\newcommand{\la}[1]{\label{#1}}
\def\alpha} \def\dot{\alpha}{{\dot\alpha}{\alpha} \def\dot{\alpha}{{\dot\alpha}}
\def\beta} \def\db{{\dot\beta}{\beta} \def\dot{\beta}{{\dot\beta}}
\def\gamma} \def\Gamma{\Gamma} \def\dot{\gamma}{{\dot\gamma}{\gamma} \def\Gamma{\Gamma} \def\dot{\gamma}{{\dot\gamma}}
\def\delta} \def\Delta{\Delta} \def\dot{\delta}{\dot\delta{\delta} \def\Delta{\Delta} \def\dot{\delta}{\dot\delta}
\def\epsilon} \def\vare{\varepsilon{\epsilon} \def\vare{\varepsilon}
\def\phi} \def\F{\Phi} \def\vvf{\f{\phi} \def\F{\Phi} \def\vvf{\phi} \def\F{\Phi} \def\vvf{\f}
\def\eta{\eta}
\def\kappa{\kappa}
\def\lambda} \def\Lambda{\Lambda{\lambda} \def\Lambda{\Lambda}
\def\mu} \def\nu{\nu{\mu} \def\nu{\nu}
\def\omega{\omega}
\def\pi} \def\Pi{\Pi{\pi} \def\Pi{\Pi}
\def\rho{\rho}
\def\sigma} \def\Sigma{\Sigma{\sigma} \def\Sigma{\Sigma}
\def\tau{\tau}
\def\theta} \def\Theta{\Theta} \def\vth{\vartheta{\theta} \def\Theta{\Theta} \def\vth{\vartheta}
\def\Xeta{\Xeta}
\def\zeta{\zeta}
\def \dot{\alpha} {\dot{\alpha}}
\def \dot{\beta} {\dot{\beta}}
\def \dot{\gamma} {\dot{\gamma}}
\def \dot{\delta} {\dot{\delta}}
\def \dot{\xi} {\dot{\xi}}
\def \dot{t} {\dot{t}}
\def \dot{\rho} {\dot{\rho}}
\def \dot{r} {\dot{r}}
\def \ddot{r} {\ddot{r}}
\def \dot{R} {\dot{R}}
\def \ddot{R} {\ddot{R}}
\def \dot{\theta} {\dot{\theta}}
\def \dot{\theta_1} {\dot{\theta_1}}
\def \dot{\phi} {\dot{\phi}}
\def \dot{\phi_1} {\dot{\phi_1}}
\def \dot{\psi} {\dot{\psi}}
\def \dot{\psi_1} {\dot{\psi_1}}
\def \dot{\varphi} {\dot{\varphi}}
\def \dot{\omega} {\dot{\omega}}
\definecolor{markcolor2}{rgb}{1,0,0}
\newcommand{\emr}[1]{{\textcolor{markcolor2}{#1}}}
\definecolor{markcolor3}{rgb}{0,1,0}
\newcommand{\emg}[1]{{\textcolor{markcolor3}{#1}}}
\newcommand{\frac{1}{2}}{\frac{1}{2}}
\newcommand{\mathbb{C} \mathrm{P}^{1}}{\mathbb{C} \mathrm{P}^{1}}
\newcommand{\mathbb{C} \mathrm{P}^{2}}{\mathbb{C} \mathrm{P}^{2}}
\newcommand{\mathbb{C} \mathrm{P}^{3}}{\mathbb{C} \mathrm{P}^{3}}
\newcommand{\mathbb{R} \mathrm{P}^{3}}{\mathbb{R} \mathrm{P}^{3}}
\newcommand{\mathrm{S}}{\mathrm{S}}
\newcommand{\mathrm{AdS}}{\mathrm{AdS}}
\newcommand{\mathrm{d}}{\mathrm{d}}
\newcommand{{\bar{\jmath}}}{{\bar{\jmath}}}
\newcommand{\bar{z}}{\bar{z}}
\newcommand{\bar{w}}{\bar{w}}
\newcommand{\bar{u}}{\bar{u}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{F}}{\mathcal{F}}
\newcommand{\mathcal{H}}{\mathcal{H}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathcal{Y}}{\mathcal{Y}}
\newcommand{\mathcal{W}}{\mathcal{W}}
\newcommand{\mathrm{su}}{\mathrm{su}}
\newcommand{\mathrm{GL}}{\mathrm{GL}}
\newcommand{\mathrm{SU}}{\mathrm{SU}}
\newcommand{\mathrm{SO}}{\mathrm{SO}}
\newcommand{\mathrm{SL}}{\mathrm{SL}}
\newcommand{\mathrm{Sp}}{\mathrm{Sp}}
\newcommand{\mathrm{U}}{\mathrm{U}}
\newcommand{\mparagraph}[1]{\paragraph{#1}\mbox{}}
\newcommand{\mathrm{CP}}{\mathrm{CP}}
\newcommand{{\dot{\alpha}}}{{\dot{\alpha}}}
\newcommand{{\dot{\theta}}}{{\dot{\theta}}}
\newcommand{{\ddot{\alpha}}}{{\ddot{\alpha}}}
\newcommand{{\ddot{\theta}}}{{\ddot{\theta}}}
\renewcommand{\headrulewidth}{0.4pt}
\begin{document}
\renewcommand{\thesection.\arabic{equation}}}{\thesection.\arabic{equation}}
\csname @addtoreset\endcsname{equation}{section}
\newcommand{\begin{equation}}{\begin{equation}}
\newcommand{\eeq}[1]{\label{#1}\end{equation}}
\newcommand{\begin{eqnarray}}{\begin{eqnarray}}
\newcommand{\eer}[1]{\label{#1}\end{eqnarray}}
\newcommand{\eqn}[1]{(\ref{#1})}
\begin{titlepage}
\begin{center}
~
\vskip 1 cm
{\Large
\bf Non-integrability and Chaos with Unquenched Flavor}
\vskip 0.5in
{\bf Dimitrios Giataganas$^{1}$ \phantom{x}and\phantom{x}Konstantinos Zoubos}$^{2,3}$
\vskip 0.1in
{\em
${}^1$ Physics Division, National Center for Theoretical
Sciences, \\
National Tsing-Hua University, Hsinchu, 30013, Taiwan
\vskip .15in
${}^2$
Department of Physics, University of Pretoria\\
Private Bag X20, Hatfield 0028, South Africa
\vskip .15in
${}^3$
National Institute for Theoretical Physics (NITheP) \\
Gauteng, South Africa
\\\vskip .1in
{\tt <EMAIL> , <EMAIL>}\\
}
\vskip .2in
\end{center}
\vskip .4in
\centerline{\bf Abstract}
We study (non-)integrability and the presence of chaos in gravity dual backgrounds of strongly coupled gauge theories with
unquenched flavor, specifically of the four-dimensional ${\cal N}=2$ super Yang-Mills theory and the three-dimensional ABJM theory. By
examining string motion on the geometries corresponding to backreacted D3/D7 and D2/D6 systems, we show that integrable
theories with quenched flavor become non-integrable when the virtual quark loops are taken into account. For the string
solutions in the backreacted D3/D7 system, we compute the leading Lyapunov exponent which turns out to saturate to a
positive value as the number of flavors increases. The exponent depends very weakly on the number of flavors when they
approach the number of colors. This suggests that once a particular flavor number in the theory is reached, a further
increase does not lead to more severe chaotic phenomena, implying certain saturation effects on chaos.
\noindent
\end{titlepage}
\vfill
\eject
\noindent
\def1.2{1.2}
\baselineskip 19 pt
\noindent
\setcounter{equation}{0}
\tableofcontents
\section{Introduction}
The study of integrability (see e.g. \cite{Beisert:2010jr, Bombardelli:2016rwb} for reviews) has proved to be a
very useful direction in obtaining exact results in AdS/CFT. An important question concerns which gauge theory
systems admit integrability or not. A fruitful approach to this problem from the dual gravity side has been to
apply techniques of analytic non-integrability. The technique, introduced in \cite{Basu:2011fw}, consists of
finding solutions for classical string motion on the gravitational background dual to a given gauge theory
which can be reduced to a Hamiltonian system on which one can apply the variational techniques of analytic
non-integrability outlined in \cite{MoralesRuiz99}.
Moreover recently there has been an extensive discussion of chaos in quantum field theories, which has further
potential applications relating the physics of black holes and quantum information. The Lyapunov exponent $\lambda} \def\Lambda{\Lambda$ of
out-of-time-ordered correlators has been proved to be bounded by the temperature $T$ of the theory as
$\lambda} \def\Lambda{\Lambda\le 2 \pi T/\hbar$ \cite{Maldacena:2015waa} using shock waves near black hole horizons
\cite{Shenker:2013pqa,Shenker:2013yza}, or by looking at the particle motion near the horizon \cite{Hashimoto:2016dfz}.
In this work we aim to study, from the dual gravity side, the integrability of theories with unquenched flavor
(see for example \cite{Erdmenger:2007cm,Nunez:2010sf} and references therein), where the virtual quark loops are
taken into account contributing to the gauge propagators. Moreover we discuss in detail chaotic string motion on
these backgrounds and, at least in the limit we are working with, we find a convergent Lyapunov exponent to a
non-zero value, although no black hole horizon is present.
We initiate our study in the case of four-dimensional theories with $\mathcal{N}=2$ supersymmetry, obtained by adding
D7-branes to a D3-brane system as in \cite{Karch:2002sh}, and taking into account their backreaction.
The background for this system was constructed in \cite{Burrington:2004id,Kirsch:2005uy}. We will work in the ``near-core'' regime of \cite{Kirsch:2005uy} where the solution is known analytically. In the unquenched case the effects of creation and annihilation of virtual quark-anti-quark pairs on the gauge degrees of freedom are taken into account. On the gravity side, this is translated as having the Dp-branes backreacting on the ``gluonic'' pure AdS space. In principle this happens when the number $N_f$ of flavor branes approaches the number $N_c$ of the color branes. In this limit the theory located at the D3/D7-brane intersection has positive beta function and possesses a chiral anomaly. Therefore the gravity dual background ceases to be conformal as can be naturally found by taking into account the effects of backreaction, while the axion and the dilaton become non-trivial. The closed string solutions we find are exactly in the region where the approximation closely follows the exact numerical solution.
We perturb the closed string solutions and apply variational methods to obtain the Normal Variational Equation (NVE) and show that it does not have Liouvillian solutions. Therefore we show that string motion is non-integrable on this background. The integrability of our solutions is restored for particle motion or when we switch off the backreaction of the D7-flavor branes. We moreover examine chaotic string motion in the non-integrable theory. We find, for the string solutions we examine, that the Lyapunov exponent in the Veneziano regime is surprisingly saturated to a value and depends weakly on the number of flavors. The value depends on the energy density of the system and its other parameters. By pushing the backreacted solution to the limits of its validity by reducing the number of flavors, we see that the maximum value of the Lyapunov exponent is obtained very quickly as we increase the flavors. We note that our analysis is done at zero temperature and the gravity dual theory has no black hole horizon. Furthermore, looking at the more generic features of the backreacted Dp-Dq brane geometries, we expect that integrability is broken in the unquenched flavor limit for a wider class of field theories.
Then we move on to examine the case of adding unquenched flavor to the ABJM model \cite{Aharony:2008ug}. This solution is obtained by introducing D6-branes and taking into account their backreaction. These backgrounds were constructed in \cite{Conde:2011sw} for the case with massless flavors, and extended to the massive case in \cite{Bea:2013jxa}. Unlike the D3/D7 case, these backgrounds are exact, while the D6 branes are smeared along the internal space. After reviewing the solution of \cite{Conde:2011sw}, we consider the general classical string equations of motion on this background and truncate them to a consistent ansatz which is suitable for applying analytic non-integrability techniques. We find analytically closed string solutions and we derive the NVE to study its solutions using the Kovacic algorithm. We conclude that, as expected and in agreement with the above D3/D7 analysis, the backreacted flavor deformation breaks the integrability present in the ABJM case. However we also point out a puzzling case where for a single string ansatz we do not find Liouvillian solutions of the corresponding NVE even when the backreaction is switched off -- conflicting our expectations of integrability in that limit. By investigating in depth the geometry we argue that it should be an issue of the string solution in the particular coordinate system used to obtain the backreaction.
As a byproduct of our study of the unquenched ABJM case, we study integrability in a range of other theories. There are special values of the parameter that controls the squashing of the space and the number of flavors, where the background is dual to other theories as well. For example, these backgrounds also appear as the IR fixed points of D2-D6 brane system flows. By computing the backreaction of the $N_f$ flavor D6-branes smeared over a six-dimensional nearly K\"{a}hler manifold, to the $N_c$ color branes, it has been found that the solutions flow to an $\mathrm{AdS}_4$ fixed point dual to Chern-Simons matter theory and which is the special case of the metric we present in section 4 \cite{Faedo:2015ula}. The family of squashed $\mathbb{C} \mathrm{P}^{3}$ metrics is relevant also for the construction of squashed seven-sphere metrics, which are $\mathrm{S}^1$ bundles over a squashed $\mathbb{C} \mathrm{P}^{3}$ base. The special case $q=5$ corresponds to an ${\cal N}=1$ supergravity background, being a gravity dual of superconformal Chern-Simons matter theory with $SO\prt{5}\times U\prt{1}$ global symmetry \cite{Ooguri:2008dk}. Therefore, our work applies in certain cases for different types of theories.
There are several works where non-integrability and/or chaos has been studied in an AdS/CFT context. In \cite{Hashimoto:2016wme} the time evolution of the homogeneous quark condensate in supersymmetric ${\cal N}=2$ QCD was studied and it was found, by looking at probe D-branes, that there exists an energy density where the phase space is dominated by chaos. Methods similar to analytic non-integrability have been applied to theories such as $\mathrm{AdS}_5\times T^{pq}$ \cite{Basu:2011fw}, Dp-brane backgrounds \cite{Stepanchuk:2012xi}, the Lunin-Maldacena background dual to the $\beta$-deformations of $\mathcal{N}=4$ SYM \cite{Giataganas:2013dha}, confining backgrounds \cite{Basu:2012ae,Ishii:2016rlk}, in theories beyond the planar limit \cite{Chervonyi:2013eja} although at special large N limits some integrability does appear \cite{Koch:2011hb}, and in non-relativistic theories \cite{Giataganas:2014hma}. Other works along these directions include \cite{ Farahi:2014lta,Ma:2014aha,Asano:2015eha, Asano:2015qwa,Panigrahi:2016zny, Asano:2016qsv, Basu:2016zkr}.
Our paper is organised as follows. In section 2, we introduce a very compact notation to analyse the string equations of motion for a generic metric. In section 3, we study the integrability of the ${\cal N}=2$ super Yang-Mills theory with $N_f$ hypermultiplets in the fundamental representation and one in the adjoint representation of the $SU(N_c)$ gauge group. We present the string solutions, the variational analysis and the proof of non-integrability. Moreover we examine the chaotic motion by solving the string equations numerically and we extract the Lyapunov exponent. In the next section 4, we study the geometry of the ABJM background with backreaction. We present several sting solutions and the NVE showing non-integrability. Moreover we mention a puzzling string solution with non-integrable motion even in the zero backreaction limit. The section is supported by two appendices. In appendix A we present more details of the $\mathbb{C} \mathrm{P}^{3}$ geometry, while in Appendix B we show how it is possible that the non-integrable $T^{11}$ space can arise as a foliation of $\mathbb{C} \mathrm{P}^{3}$ without conflicting with integrability.
\section{String equations of motion for a generic metric} \label{section:gen}
In this section we generate the string equations of motion in a generic framework and study their simplifications
for a particular string parametrisation. The metric may be written as
\begin{eqnarray}\nonumber
ds^2=g_{tt} \mathrm{d} t^2 +g_{ii} \mathrm{d} x^i \mathrm{d} x^i+2 g_{ij} \mathrm{d} x^i \mathrm{d} x^j~,
\end{eqnarray}
where in the last term $i<j$ and the indices $i,j=1,\ldots,d$ with $d$ being the number of space dimensions and $t$ being the coordinate time. Let's assume that we have $n$ cyclic coordinates, such that the labels $i\ge n$ label the cyclic angles, on which the metric elements functions are not dependent. The Nambu-Goto action is (using $\frac{\partial}{\partial \tau}:=\dot{},\frac{\partial}{\partial\sigma}:={}^\prime$)
\begin{equation}\nonumber
S=\int d\sigma} \def\Sigma{\Sigma d\tau \prt{g_{tt} \dot{t}^2+ g_{ii} \prt{ x^{i\prime}{~}^2-\dot{x}^{i}{~}^2 }+2 g_{ij} \prt{x^{i\prime} x^{j\prime}- \dot{x}^i\dot{x}^j}}:=\int d\sigma} \def\Sigma{\Sigma d\tau \prt{g_{tt} \dot{t}^2+ {\cal L}_s}~.
\end{equation}
We parametrise the string in such a way that all the non-cyclic coordinates $\alpha} \def\dot{\alpha}{{\dot\alpha}^i:=x^i$ with $i<d-n$ depend on the world-sheet time $\tau$ and therefore the string is localised along those directions. We also take the string to extend along all the cyclic coordinates $\phi^k:=x^k$ with $k\ge n$ in a way that each of them is linear in the world-sheet space parameter $\sigma} \def\Sigma{\Sigma$ with proportionality constant $m_k$.
The equation of motion for the time $t$, for a static space reads
\begin{equation}\la{gentau}
\dot{t}= \frac{\kappa }{g_{tt}}~,
\end{equation}
where $g_{tt}$ is a function of $\tau$. The equations of motion for the non-cyclic angles are
\begin{equation}\label{eomnocycle}
\partial_{\alpha} \def\dot{\alpha}{{\dot\alpha}^i} {\cal L}_s +2 \partial_\tau \prt{g_{ij}\dot{\alpha} \def\dot{\alpha}{{\dot\alpha}}^j}=0~.
\end{equation}
Notice the simplification in the absence of the $\sigma} \def\Sigma{\Sigma$ derivatives as a result of the string parametrisation.
The simplest equations for the system are the ones for the $U\prt{1}$ angles reading
\begin{equation}\label{eomcycle}
\partial_0 \prt{g_{\phi_i\phi_j}\dot{\phi_j}}- \partial_1 \prt{g_{\phi_i\phi_j}\phi_j'}=0~.
\end{equation}
For the parametrisation we have considered these are satisfied trivially since the metric elements depend on the non-cyclic coordinates.
The Virasoro constraints become
\begin{eqnarray} \label{vcg1}
g_{ij} \dot{x}^i x'^{j}= 0~,\\ \label{vcg2}
g_{tt} \dot{t}^2+{\cal L}_{s+}=0~,
\end{eqnarray}
where ${\cal L}_{s+}$ is derived from the Lagrangian density expression by flipping the negative signs in front of the kinetic terms
to positive ones.
So far the non-trivial equations are \eq{gentau}, \eq{eomnocycle} and the Virasoro constraints \eq{vcg1} and \eq{vcg2}. A further simplification happens when the metric has no non-diagonal terms between the non-cyclic $\alpha} \def\dot{\alpha}{{\dot\alpha}_i$ and the cyclic angles $\phi_i$. Then the Virasoro constraint \eq{vcg1} is satisfied trivially and the equations \eq{eomnocycle} are summed only in the non-cyclic directions.
Below we will apply the general formalism developed in this section to the gravity dual theories under examination.
\section{Backreacted flavors in four dimensions}
\subsection{Dual geometry to $\mathcal{N}=2$ SYM with unquenched flavor}
In this section we examine the $\mathcal{N}=2$ super Yang-Mills theory with $N_f$ hypermultiplets in the fundamental representation and one in the adjoint representation of the $SU(N_c)$ gauge group. The gravity dual of this background is the near horizon limit of $N_c$ coincident branes with $N_f$ number of D7-branes sharing the four spacetime directions with the D3-branes and extending along four of the six transverse directions. In the quenched approximation, where the D7-branes do not backreact on the geometry, the theory was discussed in \cite{Karch:2002sh}. As the geometry in this limit is the same as the case of no flavors, integrability for closed strings is unaffected. Integrability
for the open string sector was established in \cite{Mann:2006rh}. Here we are interested in going beyond the quenched approximation and study integrability of string motion
for the geometry dual to the unquenched theory, where the effects of backreaction of a large number $N_f$ of flavor branes are included.\footnote{One way to
obtain an exact background with flavors is through an orientifolding procedure, which leads to an $\mathrm{AdS}_5\times \mathrm{S}^5/\mathbb{Z}_2$ geometry. Integrability of
open strings in this setup was studied in \cite{Stefanski:2003qr,Chen:2004yf}. In this case the number of flavors is low, far from the $N_f\sim N_c$ regime that we are interested in.} Building on \cite{Aharony:1998xz, Grana:2001xn}, this geometry was constructed in \cite{Burrington:2004id,Kirsch:2005uy}.
The background has the metric
\begin{equation}\la{metricd3d7}
\mathrm{d} s^2=h^{-1/2}(r,\rho) \mathrm{d} x_\mu} \def\nu{\nu \mathrm{d} x^\mu} \def\nu{\nu+ h^{1/2}(r,\rho)\prt{\mathrm{d} r^2+r^2 \mathrm{d} \Omega_3^2+e^{\Psi(\rho)}\prt{\mathrm{d}\rho^2+\rho^2 ~ \mathrm{d}\psi^2}}~,
\end{equation}
where an $SO(4)\times SO(2)$ symmetry is present. The warp factor $h$ has been found in terms of a convergent series\footnote{In \cite{Aharony:1998xz} it was found to a first order approximation in the transverse 2-plane around a fixed point above the D7-branes.} in \cite{Kirsch:2005uy} following the methodology developed for the Schr\"{o}dinger equation of electrons in a logarithmic potential \cite{Gesztesy:1977vd}. Around the vicinity of the D7-branes and in the near horizon limit the warp factor becomes
\begin{equation}\la{warp}
h(r,\rho)=\frac{Q_{D3}}{(r^2+\rho^2 e^{\Psi(\rho)})^2}~,
\end{equation}
with
\begin{equation}
Q_{D3}=4\pi g_s N_c l_s^4 ,\quad \Psi(\rho)=\log\prt{b_f \log\frac{\rho_L}{\rho}}~,\quad \rho_L:=e^{\frac{2\pi N_c}{\lambda} \def\Lambda{\Lambda N_f}}~,\quad b_f:=\frac{N_f}{2\pi}~.
\end{equation}
The constant $\rho_L$ is chosen such that in the absence of the flavors $e^\phi=g_s$. The warp factor expression is only valid for values away from $\rho_L$, as $\Psi(\rho)$ diverges around this value.
In the following sections we will parametrise the $\mathrm{S}^3$ of the metric \eq{metricd3d7} as
\begin{equation}\la{param1}
\mathrm{d}\Omega_3^2=\mathrm{d}\theta} \def\Theta{\Theta} \def\vth{\vartheta^2+\sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta\prt{\mathrm{d}\alpha} \def\dot{\alpha}{{\dot\alpha}^2+\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} \mathrm{d}\chi^2}~,
\end{equation}
while the four dimensional spacetime is parametrised as
\begin{equation}\la{param2}
\mathrm{d} x_\mu} \def\nu{\nu \mathrm{d} x^\mu} \def\nu{\nu= -\mathrm{d} t^2+\mathrm{d} R^2+R^2 \mathrm{d} \phi^2+\mathrm{d} z^2~,
\end{equation}
which is a convenient choice to describe circular strings with radius $R$.\footnote{In the following, as in \cite{Kirsch:2005uy}
and much of the literature on e.g. pulsating string solutions, we will allow $R$ and $r$ to take negative values, signifying a
string of opposite orientation.}
\subsection{String solutions and non-integrability}
In order to study (non)-integrability of the theory, we examine classical string solutions on the above background.
We are interested in string solutions that extend along the cyclic angles $(\phi,\chi,\psi)$ linearly with the $\sigma} \def\Sigma{\Sigma$ string
worldsheet parameter, where the linearity constants are set to be $(m_\phi,~m_\chi,~m_\psi)$. In other words, the string is
wrapping these angles an amount of times given by the corresponding $m$. All the other angles, as well as the coordinate
time itself, are only allowed to be functions of the time world-sheet parameter $\tau$.
The equations of motion are derived by the Nambu-Goto action in the geometry of \eq{metricd3d7} with the parametrisations
\eq{param1} and \eq{param2}. Our parametrisation ensures that all the cyclic equations of motion are trivially
satisfied, while the time equation of motion gives
\begin{equation}\la{timed7}
\dot{t}(\tau)=\frac{\kappa}{2}\frac{\sqrt{Q_{D3}}}{|r(t)^2-b_f \log(\frac{\rho(\tau)}{\rho_L})\rho(\tau)^2|}~,
\end{equation}
where $\kappa$ is the integration constant.
By placing the string on the equators $(\theta} \def\Theta{\Theta} \def\vth{\vartheta=\pi/2,~ \alpha} \def\dot{\alpha}{{\dot\alpha}=\pi/2)$ of the $\mathrm{S}^3$ internal sphere the two relevant equations are
satisfied. Moreover, the $z$ equation of motion is satisfied by switching off this direction for the string. We are now
left with the three equations of motion corresponding to $(R(\tau),r(\tau),\rho(\tau))$ functions and one non-trivial Virasoro constraint.
It is important to observe that the harmonic function $h$ contains the term $\log(\rho(\tau)) \rho(\tau)^2$ where its
derivative with respect to $\tau$ or $\rho(\tau)$ gives the same factor $1+2 \log(\rho(\tau))$. Due to this fact the $\rho$
equation of motion is satisfied for $\rho= \rho_L/\sqrt{e}$. Note that this value is within the regime where
the background approximation is valid. The remaining equations of motion are those for $R(\tau)$ and $r(\tau)$, which read
\begin{eqnarray}\label{eaf1}
&&m_\phi^2\prt{b_f\rho_L^2+2 e r(\tau)^2}R(\tau)+4 e r(\tau)\dot{r}(\tau) \dot{R}(\tau)+\prt{b_f \rho_L^2+2 e r(\tau)^2}\ddot{R}(\tau)=0~,\\\nonumber
&&\frac{2}{\sqrt{Q_{D3}}(b_f \rho_L^2+2 e r(\tau)^2)^2}\bigg( r(t)\Big(-m_\phi^2\prt{b_f \rho_L^2+2 e r(\tau)^2}^2 R(t)^2+e Q_{D3} (-\kappa^2 e+ 4 e m_\chi^2 +2 b_f m_{\psi}^2 \rho_L^2\\&&+4 e r'(t)^2) +\prt{b_f \rho_L^2 +2 e r(t)^2}^2 R'(t)^2\Big)-2 e Q_{D3} \prt{ b_f \rho_L^2+2 e r(t)^2}r''(t)\bigg)=0~,\label{eaf2}
\end{eqnarray}
while the non-trivial Virasoro equation reads
\begin{eqnarray}\nonumber \label{virf1}
&&\frac{1}{2 e \sqrt{Q_{D3}}(b_f \rho_L^2+2 e r(\tau)^2)} \bigg(m_\phi^2 (b_f \rho_L^2+2 e r(\tau)^2)^2 R(t)^2 + e Q_{D3} (-\kappa^2 e+ 4 e m_\chi^2\\
&&\hspace{4cm} +2 b_f m_{\psi}^2 \rho_L^2 +4 e r'(t)^2)+(b_f \rho_L^2+2 e r(\tau)^2)^2 R'(t)^2\bigg)=0~.
\end{eqnarray}
A string solution of the above system and its NVE does exist. To simplify the system significantly without losing generality, we localise the string on the internal sphere by setting $m_\chi=0$. This allows an analytic $r(\tau)$ solution if the
string is taken to move in the $R(\tau)=0$ plane
\begin{equation}
r(\tau)=\frac{\sqrt{\kappa^2 e-2 b_f m_\psi^2 \rho_L^2}}{2 \sqrt{e}} \tau+c_1~,
\end{equation}
where $c_1$ is an integration constant. By applying variations along the $R(\tau)=0+\eta(\tau)$ direction, the corresponding equation of motion gives an NVE with rational coefficients
\begin{equation}\label{nve01
\ddot{\eta}(\tau)+\frac{2 \prt{\kappa^2 e -2 b_f m_\psi^2 \rho_L^2}\tau}{(\kappa^2 e -2 b_f \rho_L^2 m_\psi^2 )\tau^2+2 b_f \rho_L^2}\dot{\eta}(\tau)+m_\phi^2 \eta(\tau)=0~,
\end{equation}
where we set $c_1=0$. The above equation does not have Liouvillian solutions and the Kovacic algorithm fails. Therefore this means that the flavor background in the unquenched limit is non-integrable and the quark loops destroy the integrability.
In fact one may even further simplify the string solution and the corresponding NVE and still show non-integrability.
The string may be localised consistently in the transverse $SO(2)$ plane, parametrised by $(\rho,\psi)$, by setting
$m_\psi=0$. Choosing the appropriate initial conditions we get an invariant plane solution $R(\tau)=0$ with
$r(\tau)=\kappa~ \tau$, and the variation of $R(\tau)=0+\eta(\tau)$ leads to
\begin{equation}\la{nvesimple}
\ddot{\eta}(\tau)+\frac{2 \kappa^2 e \tau}{2 b_f \rho_L^2+e \kappa^2 \tau^2}\dot{\eta}(\tau)+m_\phi^2 \eta(\tau)=0~,
\end{equation}
which can be also obtained directly from \eq{nve01}.
The minimal string NVE \eq{nvesimple} does not have a Liouvillian solution proving non-integrability in the strong coupling limit flavor backreacted backgrounds.
Integrability is recovered for particle motion $(m_\phi=0)$ where the NVE admits Liouvillian solutions
\begin{equation}\label{etaparticle}
\eta(\tau)=c_3+c_4 \frac{\arctan(\frac{\kappa \sqrt{e}\tau}{\sqrt{2 b_f} \rho_L})}{\kappa \sqrt{2 e b_f}\rho_L}~,
\end{equation}
where $c_3$ and $c_4$ are the integration constants.
A comment is in order for the quenched limit of our solution. The expression
\begin{equation}\label{parameter1}
b_f \rho_L^2=\frac{N_f}{2\pi} e^{\frac{2\pi N_c}{\lambda} \def\Lambda{\Lambda N_f}}~,
\end{equation}
is infinite both for small and for large number of flavors, while for $N_f=2\pi N_c/\lambda} \def\Lambda{\Lambda$ the function has a minimum at $ e N_c/\lambda} \def\Lambda{\Lambda$, depending on the number of colors (and flavors) as well as the choice of the t'Hooft coupling. Moreover the string position in this limit goes deeply in the IR as $\rho\to \infty$ when considering the t'Hooft limit. As a result the NVE \eq{nve01} for $N_f=0$ does not have a smooth limit to the unquenched approximation. This limit may be recovered only together with the point-like string motion giving the integrable motion \eq{etaparticle}.
To summarise this section, we showed the existence of non-integrable string solutions in the flavor-backreacted D3/D7 background of \cite{Kirsch:2005uy}. Integrability is recovered for particle motion, a situation that has been observed in several other non-integrable theories. In the next section we study the existence of chaos in the flavor backreacted system.
\begin{figure}
\begin{minipage}[ht]{0.5\textwidth}
\begin{flushleft}
\centerline{\includegraphics[width=60mm]{3dmotion2.pdf}}
\caption{\small{The evolution of string motion along the holographic direction and its radius. The two strings have close initial conditions while the final endpoints begin to differ significantly. We simulate the motion until the initial moments of the chaos appearance. The axes have been rescaled for better optical results. }}
\label{figure:a1}
\end{flushleft}
\end{minipage}
\hspace{0.3cm}
\begin{minipage}[ht]{0.5\textwidth}
\begin{flushleft}
\centerline{\includegraphics[width=75mm ]{3dprojrb.pdf}}
\caption{\small{The evolution of the strings along the direction $r$ for the same initial conditions as in Figure \ref{figure:a1}. Notice that the negative values of $r(\tau)$ and $R(\tau)$ correspond to radii of $|r(\tau)|$ and $|R(\tau)|$ with changed string orientation comparing to the positive values. Both plots have $N_f/N_c\sim 0.3,~\lambda=12$, where the values have chosen for best optical results. Higher values of $N_f/N_c$ give a similar pattern and make the chaotic oscillations more severe and observed at earlier times.} }
\label{figure:a2}\vspace{.0cm}
\end{flushleft}
\end{minipage}
\end{figure}
\subsection{Chaotic string motion}
Having shown analytically the non-integrability of string motion on the backreacted D3/D7 background, we examine the
presence of chaos in this background. To do this, we consider the same string ansatz as above but without choosing any specific solution for $R(\tau)$. So the fields that will have a nontrivial motion are $r(\tau)$ and $R(\tau)$, corresponding to a time
evolution of a boundary operator. In terms of the Hamiltonian the system is reduced to a particle motion with four
dynamical variables $(r,R,\dot{r},\dot{R})$ and certain constant parameters. The conserved quantity of the system is the total energy. We look for chaos by varying the initial conditions and the theory parameters
resulting in modifying the energy density, and each time we integrate the equations of motion numerically.
Let us fix the value of the 't Hooft coupling $\lambda$ and examine the appearance of chaos as a function of $N_c/N_f$. We solve the system of differential equations \eq{eaf1} and \eq{eaf2}, with the energy constrained by \eq{virf1}. The 3-dimensional string motion with two neighbouring initial conditions is presented in Figure \ref{figure:a1}, where the chaotic nature starts to become apparent as the time evolves. In Figure \ref{figure:a2} we project the string motion to one dimension $r(\tau)$ where the chaotic nature also begins to be apparent.
\begin{figure*}[!ht]
\centerline{\includegraphics[width=140mm]{poincare.pdf}}
\caption{\small{The Poincar\'e sections for the flavor backreacted theory. The section is chosen such that $r(\tau)=0$. The constant $\kappa=140$, the winding number $m=1$ and $\lambda/N_c$ is around unity. The bottom right plot is the Poincar\'e section for $N_f/N_c\sim 1$ where the chaos is obvious. The other three plots from top left to right correspond to $N_f/N_c=0.094,~0.096,~0.098$. Note that at this limit the range of the validity of the gravity solution is already challenged. However we allow ourselves to push towards this limit to illustrate the formation of chaos. }}
\label{figure:a3}
\end{figure*}
The chaotic motion can be quantified in terms of $N_c, ~N_f$ and $\lambda} \def\Lambda{\Lambda$, parameters that our system and the equations of motion depends on through $b_f \rho_L^2$ and is equal to \eq{parameter1}. The integrability of our equation is recovered in the limit $N_c/(\lambda} \def\Lambda{\Lambda N_f)=1/(2\pi)$ with $N_f\sim N_c/\lambda} \def\Lambda{\Lambda\to 0$, which is beyond the regime of validity of our background.
A way to observe how chaos depends on the parameters of the system is by the construction of the Poincar\'e sections which are presented in Figure \ref{figure:a3} for an increasing number of flavors.
The choice of the constant $\kappa$ plays a crucial role in the energy density and has to be chosen in such a
way as to allow consistent initial conditions. We are already in the chaotic region and therefore we cannot
clearly observe the destruction of tori. However by slightly extending our analysis outside the
appropriate limit for $N_f/N_c$ and considering a low number of flavors we get a picture of how chaos starts
to form.
To quantify the sensitivity of our system to the initial conditions we compute the leading Lyapunov exponent. We obtain the exponent by a long-time numerical calculation of two initially neighboring trajectories, where each time we measure the exponent, we rescale the distance in order to keep the nearby trajectory separation within the linearised flow range. Our main finding is that when $N_f\sim N_c$
the Lyapunov coefficient depends weakly on the increase of the number of flavors with a slow increasing trend. It almost converges to a value depending on the energy of the system. In the computation we vary $N_f$ by pushing again
the limits of our background validity. This is tempting since we find that for low values of flavor numbers we see
lower Lyapunov coefficient, which nevertheless quickly converge to the particular value. The Lyapunov exponent in terms of
the ratio $N_f/N_c$ can be well approximated with shifted sigmoid functions like the $\arctan$.
\begin{figure}
\centerline{\includegraphics[width=100mm]{lyapd3d7.pdf}}
\caption{\small{The leading Lyapunov exponent in terms of number of flavors over number of colors. The main result is that in the region $N_f\sim N_c$ the exponent quantifying the chaotic motion is weakly depending on the number of flavors. It is tempting to extend our computation for lower values of flavor pushing the limits of the validity of our background. Starting from small values
of $N_f/N_c$ and increasing, we see that the exponent quickly converges to the constant value already at $N_f\sim 0.4 N_c$. }}
\label{figure:a4}
\end{figure}
\section{Backreacted flavors in three dimensions}
In this section we consider the integrability of string motion on backgrounds dual to three-dimensional
gauge theories with unquenched flavor. These backgrounds arise by adding branes to the $\mathcal{N}=6$ ABJM
Chern-Simons-matter theory, which is known to be integrable
(see \cite{Arutyunov:2008if,Stefanski:2008ik} for
the Lax-pair construction for the string sigma model),
and allowing them to backreact. In the following we review the geometry of the ABJM background and discuss how to
obtain the deformation corresponding to backreacted flavor, before turning to the study of string solutions
relevant for the study of integrability.
\subsection{The ABJM background} \label{Section:ABJM}
The ABJM background \cite{Aharony:2008ug} is given by
\begin{equation} \label{metric01}
\mathrm{d} s^2=L^2 \mathrm{d} s^2_{\mathrm{AdS}_4}+4 L^2 \mathrm{d} s^2_{\mathbb{C} \mathrm{P}^{3}}~,
\end{equation}
where $L^4=2 \pi^2 \frac{N_c}{k}$, and the dilaton, RR 2- and 4-form field strengths are given by
\begin{equation}
e^\phi=\frac{2 L}{k}=2 \sqrt{\pi}\prt{\frac{2 N}{k^5}}^\frac{1}{4}~,
\quad F_2=2 k J~,\quad F_4=\frac{3}{2} k L^2 \Omega_{\mathrm{AdS}_4}=\frac{3 \pi}{\sqrt{2}}\sqrt{k N} \Omega_{\mathrm{AdS}_4}~,
\end{equation}
where $J$ is the K\"{a}hler form of $\mathbb{C} \mathrm{P}^{3}$ and $\lambda} \def\Lambda{\Lambda=N_c/k$ is the t'Hooft coupling. This is dual to a Chern-Simons-matter theory,
of rank $N_c$ and at level $k$, in the limit $N_c^\frac{1}{5}\ll k\ll N_c$, corresponding to
large $\mathrm{AdS}$ radius and small string coupling.
We will mostly be concerned with the internal $\mathbb{C} \mathrm{P}^{3}$ space, so we will provide some details
on its geometry. The form of $\mathbb{C} \mathrm{P}^{3}$ which is convenient for introducing the backreaction is as an $\mathrm{S}^2$-bundle over $\mathrm{S}^4$, using the self-dual $\mathrm{SU}\prt{2}$ instanton on the four-sphere
\begin{equation}\label{metrica1}
\mathrm{d} s^2_{\mathbb{C} \mathrm{P}^{3}}=\frac{1}{4}\prt{\mathrm{d} s^2_{\mathrm{S}^4}+\prt{\mathrm{d} x^i+\epsilon} \def\vare{\varepsilon^{ijk}A^j x^k}^2}~.
\end{equation}
Here the $x^i$ are the Cartesian coordinates of the $\mathrm{S}^2$ satisfying $x_i^2=1$ and $A^i$ is the non-abelian one-form
connection of the $\mathrm{SU}\prt{2}$ instanton. The metric of the four-sphere reads\footnote{In the parametrisation of
\cite{Conde:2011sw} a coordinate $\xi$ is used instead of $\alpha$, with the relation between these coordinates
being $\sin\alpha} \def\dot{\alpha}{{\dot\alpha}=\frac{2 \xi}{1+\xi^2}$, or $\xi=\tan\frac{\alpha}{2}$.
}
\begin{equation}
\mathrm{d} s^2_{\mathbb{S}^4} =
\Big[\mathrm{d} \alpha} \def\dot{\alpha}{{\dot\alpha}^2+\frac{1}{4}\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} \,\sum_{i=1}^3 ( \omega^i)^2\,\Big]\,\,,
\end{equation}
where the range of $\alpha} \def\dot{\alpha}{{\dot\alpha}$ is $0\leq\alpha} \def\dot{\alpha}{{\dot\alpha} \le \pi$
and the $\omega_i$ are $\mathrm{SU}\prt{2}$ left-invariant forms satisfying
$\mathrm{d} \omega^i=\frac{1}{2} \epsilon} \def\vare{\varepsilon_{ijk} \omega^j\wedge \omega^k$. We will write them explicitly in terms of angular coordinates as
\begin{eqnarray} \label{leftinv1}
\omega^1 & = & \cos\psi_1\,\mathrm{d}\theta_1+\sin\psi_1\,\sin\theta_1\,\mathrm{d}\varphi_1~,\\ \label{leftinv2}
\omega^2 & = & \sin\psi_1\,\mathrm{d}\theta_1-\cos\psi_1\,\sin\theta_1\,\mathrm{d}\varphi_1~, \\ \label{leftinv3}
\omega^3 & = & \mathrm{d}\psi_1+\cos\theta_1 \,\mathrm{d}\varphi_1\,\, .
\label{omegas}
\end{eqnarray}
Then the one-forms $A^i$ are expressed as
\begin{equation}
A^i=-\sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\omega^i~,
\end{equation}
while the $x^i$ are parametrised with the spherical angles $\theta} \def\Theta{\Theta} \def\vth{\vartheta$ and $\varphi$ giving
$\prt{\mathrm{d} x^i+\epsilon} \def\vare{\varepsilon^{ijk}A^j x^k}^2= \prt{E^1}^2+\prt{E^2}^2$.
The $E^1$ and $E^2$ are two one-forms that can be written as
\begin{eqnarray}
E^1 & = & \mathrm{d}\theta+\sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\prt{\sin\varphi~\omega^1 -\cos\varphi~\omega^2}~, \\
E^2 & = & \sin\theta\prt{\mathrm{d}\varphi-\sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\omega^3}
+\sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\cos\theta\prt{ \cos\varphi\omega^1+\sin\varphi\omega^2} ~ .
\end{eqnarray}
The RR two-form can be written as
\begin{equation}
F_2=\frac{k}{2}\prt{E^1\wedge E^2-\prt{{\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}^\alpha} \def\dot{\alpha}{{\dot\alpha}\wedge {\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}^3+{\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}^1\wedge {\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}^2}}~,
\end{equation}
wher
\begin{eqnarray}\nonumber
&&{\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}^\alpha} \def\dot{\alpha}{{\dot\alpha}= \mathrm{d}\alpha} \def\dot{\alpha}{{\dot\alpha}~,\qquad {\cal S}} \def\cT{{\cal T}} \def\cU{{\cal U}^i=\frac{\sin{\alpha} \def\dot{\alpha}{{\dot\alpha}}}{2}S^i~,\\\nonumber
&&S^1=\sin\varphi \omega^1-\cos\varphi \omega^2~,\quad
S^2=\sin\theta \omega^3-\cos\theta\prt{\cos\varphi\,\omega^1+
\sin\varphi\,\omega^2}~, \\
&&
S^3=-\cos\theta \omega^3-\sin\theta\prt{\cos\varphi\,\omega^1+
\sin\varphi\,\omega^2}~.
\la{si}
\end{eqnarray}
The addition of flavor D6 branes to the ABJM background, in the probe approximation, was
considered in \cite{Hohenegger:2009as,Gaiotto:2009tk,Hikida:2009tp,Jensen:2010vx,Ammon:2009wc,Zafrir:2012yg}. The
D6-branes are wrapped around an $\mathbb{R} \mathrm{P}^{3}$ subspace of $\mathbb{C} \mathrm{P}^{3}$, with the configuration preserving $\mathcal{N}=3$ supersymmetry. The integrability of the corresponding gauge theory (taking into account open spin chains corresponding
to operators with fundamental fields) was recently discussed in \cite{Bai:2017jpe}.
In this work we are interested in the geometry arising when the backreaction of the flavor branes is taken into account,
corresponding to unquenched flavors in the ABJM gauge theory. We will discuss the resulting geometry in the next section.
\subsection{The flavor-backreacted background}
The deformed $\mathrm{AdS}_4\times \mathbb{C} \mathrm{P}^{3}$
metric derived in \cite{Conde:2011sw} is
\begin{equation}\label{metric1}
\mathrm{d} s_{10}^2 =
L^2\prt{-\cosh^2\rho ~\mathrm{d} t^2+\mathrm{d} \rho^2+\sinh^2\rho~ \mathrm{d} \Omega_2^2} + \frac{L^2}{b^2}\left(q \mathrm{d} s^2_{\mathbb{S}^4}+(E^1)^2+(E^2)^2\right) \ ,
\end{equation}
where the $\mathrm{AdS}_4$ metric is written in global coordinates and $q$ and $b$ are constants measuring the degree of the deformation from the original metric, by changing the size of the $\mathrm{S}^4$ compared with $\mathrm{S}^2$ of the $\mathbb{C} \mathrm{P}^{3}$ manifold. The special case of $q=1$ corresponds to the undeformed ${\cal N}=6$ supersymmetric ABJM background, while for all other values of $q$ the space is squashed. The other special case $q=5$ corresponds to an ${\cal N}=1$ supergravity background, dual to superconformal Chern-Simons matter theory with $\mathrm{SO}\prt{5}\times \mathrm{U}\prt{1}$ global symmetry \cite{Ooguri:2008dk}. The parameter $b$ can be thought as the scaling of the energy of a localised observable in the AdS space.
The flavored $\mathrm{AdS}$ solutions for constant parameters are given by the solution of the BPS conditions as
\begin{eqnarray} \label{qvalues}
q = 3+\frac{3}{2}\hat\epsilon \pm 2\sqrt{1+\hat\epsilon+\frac{9}{16}\hat\epsilon^2}~,\qquad
b = \frac{2q}{q+1}\ ,\label{constants}
\end{eqnarray}
where $\hat\epsilon$ is the usual Veneziano parameter reading
\begin{equation}
\hat \epsilon =\frac{3}{4}\frac{N_f}{N_c}\lambda \ .
\label{epsilon}
\end{equation}
The branch with the minus sign extends from $q=1$ to $5/3$ and is the flavored backreacted ABJM, since $q=1$ is included. The branch with positive sign extends from $q=5$ upwards and corresponds to the flavored version of the deformed ${\cal N}=1$ CS matter theory, since the value $q=5$ is included.
The mentioned backgrounds appear also as the IR fixed points of D2-D6 brane system flows. By computing the backreaction
of the $N_f$ flavor D6-branes smeared over a six-dimensional nearly K\"{a}hler manifold, to the $N_c$ color branes, it has
been found that the solutions flow to an $\mathrm{AdS}_4$ fixed point dual to Chern-Simons-matter theory and which is the special
case of the metric \eq{metric1} for a specific $q$-value and a squashing parameter appearing in the two-form \cite{Faedo:2015ula}.
It is important to note that the flavor-backreacted geometry contains an $\mathrm{AdS}_4$ factor, as the addition of
fundamental flavors to ABJM does not generically break conformal invariance (see
\cite{Bianchi:2009ja,Bianchi:2009rf} for perturbative gauge theory considerations). This is unlike the
four-dimensional case discussed above and of course results in a much simpler geometric
background.
Since the construction of the above background involved smearing D6-branes with different orientations,
the resulting supersymmetry is $\mathcal{N}=1$ in three dimensions,
instead of $\mathcal{N}=3$, which would only be expected for the supergravity solution with
localised flavor branes \cite{Conde:2011sw}.
\subsection{String Solutions on the backreacted metric}
In this section we explore the integrability of string motion on the flavor backreacted ABJM background above.
We will consider three ans\"atze, which are distinguished by whether we set $\theta=0$ ($\theta$ being
the angle of the $\mathrm{S}^2$), $\theta=\pi/2$, or $\theta_1=0$ ($\theta_1$ being the angle of the $\mathrm{S}^4$).
\subsubsection{Static String on $\mathrm{S}^2$}
We will now consider an ansatz for string motion which will turn out to be suitable for the application of analytic non-integrability techniques. We choose:
\begin{equation} \label{Theta1Ansatz1}
t=\kappa\tau~,\quad \alpha} \def\dot{\alpha}{{\dot\alpha}=\alpha} \def\dot{\alpha}{{\dot\alpha}\prt{\tau}~,\quad \theta=0~, \quad \theta_1=\theta_1\prt{\tau}~,\quad \varphi_1= m\sigma} \def\Sigma{\Sigma~,\quad \varphi=\frac{\pi}{2}~,\quad\psi_1=\frac{\pi}{2}\;.
\end{equation}
As this parametrisation is contained within the general analysis of section \ref{section:gen}, it is immediate to analyse
the equations of motion. Doing this we find that the only non-trivial equations of motion are those for $\theta_1$ and $\alpha$
\begin{equation}
\begin{split}
\partial_\alpha} \def\dot{\alpha}{{\dot\alpha}{\cal L}_s+2\partial_0\prt{g_{\alpha} \def\dot{\alpha}{{\dot\alpha}\a}\dot{\alpha} \def\dot{\alpha}{{\dot\alpha}}}=0~,\\
\partial_{\theta_1} {\cal L}_s+2\partial_0\prt{g_{\theta_1\theta_1}\dot{\theta_1}}=0~,\\
g_{tt} \dot{t}^2+{\cal L}_{s+}=0~.
\end{split}
\end{equation}
Writing these out more explicitly, we obtain
\begin{equation}
q \ddot \alpha-\frac14{\dot\theta_1}^2\sin\alpha(1+\cos\alpha(q-1))
+\frac14m^2\sin\alpha\left(q \cos\alpha+(1-\cos\alpha)\sin^2\theta_1\right)=0
\end{equation}
and
\begin{equation}
\ddot\theta_1(1+q\sin^2\alpha+\cos^2\alpha-2\cos\alpha)+2\dot\theta_1\dot\alpha\sin\alpha (1+(q-1)\cos\alpha)
+ m^2\sin\theta_1\cos\theta_1(\cos\alpha-1)^2=0~.
\end{equation}
The two equations of motion may be thought as coming from an effective particle lagrangian
\begin{equation}
\begin{split}
\mathcal{L}_{\text{eff}}&=b^2\kappa^2+\frac{q}{4}{\dot{\alpha}}^2+\frac{1}{16}\left(4 \sin^4\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}+q \sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} \right){\dot{\theta}}_1^2 -\frac{m^2}{16}\prt{4 \sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta_1 \sin^4\frac{\alpha}{2}+q\sin^2\alpha}~.
\end{split}
\end{equation}
Converting to the Hamiltonian using $p_\alpha=\partial \mathcal{L}_{\text{eff}}/\partial {\dot{\alpha}}$, $p_{\theta_1}=\partial \mathcal{L}_{\text{eff}}/\partial {\dot{\theta}}_1$,
we obtain
\begin{equation}
\begin{split}
\mathcal{H}&=\frac{p_\alpha^2}{q}+\frac{4 p_{\theta_1}^2}{4 \sin^4\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}+q \sin^2\alpha} \def\dot{\alpha}{{\dot\alpha}}-b^2\kappa^2+\frac{ m^2}{16}\prt{4 \sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta_1 \sin^4\frac{\alpha}{2}+q\sin^2\alpha}~.
\end{split}
\end{equation}
The Virasoro constraint sets the energy to zero, which must be imposed on the dynamics of this Hamiltonian. Using these formulas we can treat the system as one of particle motion and construct the Poincar\'e maps as in Figure \ref{Theta0ansatz}, which presents different values of $q$. We find that, at least for the initial conditions we have considered, there is only weak evidence of chaotic behaviour in this parameter space. This is confirmed by a computation of the Lyapunov exponent,
which, although positive, turns out to be small and not significant enough to make a conclusive statement about chaos in this system.
\begin{figure}[h]
\begin{center}
\resizebox{2.0in}{!}{\includegraphics{PSansatz1qis1.png}} \,\,
\resizebox{2.0in}{!}{\includegraphics{PSansatz1qis1point1.png}}\,\,
\resizebox{2.0in}{!}{\includegraphics{PSansatz1qis1point65.png}}\vspace{-.2cm}
\caption{ Sample Poincar\'e plots for the ansatz (\ref{Theta1Ansatz1}), here plotted for $m=1,\kappa=0.6$, and varying values of the deformation parameter
$q$. (a) $q=1$, (b) $q=1.1$ and (c) $q=1.65$. The total energy is restricted to $H=0$ due to the Virasoro constraint.} \label{Theta0ansatz}
\end{center}
\end{figure}
In the following we will turn to the techniques of analytic non-integrability, for which we need to choose an
invariant plane. We will consider two cases.
\textbf{ Case I: The $\theta_1=0$ plane}
On this plane, the only non-trivial equation of motion is for $\alpha(\tau)$:
\begin{equation}
\ddot\alpha+\frac 14 m^2 \sin\alpha\cos\alpha=0
\end{equation}
which, using the Virasoro constraint, gives
\begin{equation}
\dot\alpha^2=\frac{1}{4q}\left(16b^2\kappa^2-m^2q\sin^2\alpha\right)~.
\end{equation}
To obtain the NVE, we expand $\theta_1(\tau)=0+\eta(\tau)$ and arrive at
\begin{equation}
\frac12\ddot \eta \sin^2\frac{\alpha}{2}(1+q+(q-1) \cos\alpha} \def\dot{\alpha}{{\dot\alpha})
+\frac12\dot\eta\dot\alpha\sin\alpha (1+(q-1)\cos\alpha)
+ m^2\sin^4\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\eta=0~.
\end{equation}
We need to bring this equation into a form suitable for applying the Kovacic algorithm. In particular it needs to be
an ODE with rational coefficients. We can achieve this by making the substitution $z=\cos\alpha$, after which we find the
NVE\footnote{Primes denote $\frac{\partial}{\partial z}$.}
\begin{equation}
\begin{split}\nonumber
&\eta''(z-1)^2(z+1)\prt{(q-1)z+q+1}\prt{m^2 q( z^2-1)+16 b^2 \kappa^2}+2\eta'(z-1)\big( m^2 q(z^2-1)\\
&\cdot(1+z+2 q z+2(q-1)z^2)+8b^2k^2(2+z+3 q z +3(q-1)z^2)\big)
+4\eta q m^2 (z-1)^2=0~.
\end{split}
\end{equation}
Applying the Kovacic algorithm we find that this NVE does not admit integrable solutions for generic $q,m$.
However, specialising to $m=0$ (the point-like limit) or $q=1$ (the undeformed $\mathbb{C} \mathrm{P}^{3}$ limit) we find
an integrable NVE. This is consistent with the integrability of geodesic motion on the backreacted backgrounds
as well as the integrability of the ABJM theory.
\textbf{Case II: The $\theta_1=\frac{\pi}2$ plane}
The other invariant plane in our problem is $\theta_1=\frac{\pi}{2}$. On this plane the $\alpha$ equations
of motion become
\begin{equation}
\ddot \alpha=-\frac{m^2\sin\alpha}{4q}\left(1+(q-1)\cos\alpha\right)~.
\end{equation}
Expanding $\theta_1=\frac{\pi}2+\eta(\tau)$ and changing variables to $z=\cos\alpha$, we find the NVE
\begin{equation}
\begin{split}
&\eta''(z+1)(z-1)^2\prt{1+q +z(q-1)}(16 b^2 \kappa^2+m^2(z-1)(1+q +z(q-1))\\
&+\eta' (z-1)\cdot \bigg(16\kappa^2 b^2\prt{2+z+3 q z+3(q-1)z^2}+m^2(z-1)\prt{1+q+(q-1)z}\\
&\cdot \prt{3+z+4 q z+4 (q-1) z^2}\bigg)-4 m^2 q (z-1)^2 \eta=0~.
\end{split}
\end{equation}
As above, we find that the Kovacic algorithm concludes for $m=0$ and for $q=1$, while it does not conclude with Liouvillian solutions in the generic case. We see that the backreacted background ($q>1$) of \cite{Conde:2011sw} is non-integrable, while for $q=1$ integrability appears to be recovered, for this string solution at least. In the next section we present the analysis of a string moving in a $\mathbb{C} \mathrm{P}^{2}$ subspace of $\mathbb{C} \mathrm{P}^{3}$.
\subsubsection{The string on $\mathbb{C} \mathrm{P}^{2}$} \label{CP2section}
Let us now consider an ansatz corresponding to string motion purely on the $\mathbb{C} \mathrm{P}^{2}$ part of the
deformed $\mathbb{C} \mathrm{P}^{3}$ spacetime. As discussed in appendix \ref{CP2restriction}, the restriction to the subspace defined by $\theta=\phi=\pi/2$, of the undeformed $\mathbb{C} \mathrm{P}^{3}$ metric (in the $\mathrm{S}^2\rightarrow \mathbb{C} \mathrm{P}^{3}\rightarrow \mathrm{S}^4$ fibration picture, where $\theta,\phi$ are the coordinates of the $\mathrm{S}^2$) leaves us with the Fubini-Study metric on $\mathbb{C} \mathrm{P}^{2}$. It is thus interesting to consider strings moving in this subspace both in the quenched and unquenched limit.
The string is parametrised as
\begin{equation} \label{CP2ansatz}
t=\kappa\tau\;,\quad \theta=\phi =\frac{\pi}2,\quad \alpha =\alpha(\tau),\quad\theta_1 =\theta_1(\tau),\quad\phi_1 =m\sigma, \quad\psi_1 =\frac{\pi}{2}~,
\end{equation}
giving non-trivial equations of motion for $\alpha$ and $\theta_1$ which are
\begin{equation}
-\frac{q}{2}{\ddot{\alpha}}+\frac{q}{8}\sin\alpha\cos\alpha{\dot{\theta}}_1^2-\frac{q m^2}{8}\sin\alpha\cos\alpha
-\frac{m^2}{8}(1-\cos\alpha)\sin\alpha=0~,
\end{equation}
while the $\theta_1$ EOM:
\begin{equation}
-\frac{1}{8}\sin^2\alpha {\ddot{\theta}}_1-\frac{1}{4}\sin\alpha\cos\alpha {\dot{\alpha}} {\dot{\theta}}_1=0~.
\end{equation}
We note that $q$ does not enter into this equation.
The system can be reduced to a particle one with effective lagrangian reading
\begin{equation}
\mathcal{L}_{\text{eff}}=\frac{q}{4}{\dot{\alpha}}^2+\frac{q}{16}\sin^2\alpha{\dot{\theta}}_1^2+b^2\kappa^2-\frac{q m^2}{16}\sin^2\alpha
-\frac{m^2}{16}(1-\cos\alpha)^2
\end{equation}
and the corresponding Hamiltonian
\begin{equation}
\mathcal{H}=\frac{1}{q}p_\alpha^2+\frac{4}{q\sin^2\alpha}p^2_{\theta_1}-b^2\kappa^2+\frac{q m^2}{16}\sin^2\alpha
+\frac{m^2}{16}(1-\cos\alpha)^2~,
\end{equation}
where we note that $\theta_1$ does not appear, only its conjugate momentum. Having these we can look at two nearby trajectories.
\begin{figure}[h]
\begin{center}
\resizebox{2.0in}{!}{\includegraphics{PSplotCP2qis1.png}} \,
\resizebox{2.0in}{!}{\includegraphics{PSplotCP2qis3over2.png}}\,
\resizebox{2.0in}{!}{\includegraphics{PSplotCP2qis5over3.png}}\vspace{-.2cm}
\caption{ Phase space trajectories with nearby initial conditions for the $\mathbb{C} \mathrm{P}^{2}$ ansatz, here plotted for $m=1,\kappa=5$, and varying values of the deformation parameter
$q$. (a) $q=1$, (b) $q=3/2$ and (c) $q=5/3$. The total energy is restricted to $H=0$ due to the Virasoro constraint.} \label{CP2ansatzplots}
\end{center}
\end{figure}
From Fig. \ref{CP2ansatzplots} we see that motion on $\mathbb{C} \mathrm{P}^{2}$ is consistent with integrability, as expected. This is true for both $q=1$ and $q\neq 1$ in this simple string solution.
We now proceed to choose an invariant plane and check the non-integrability analytically. We take $\theta_1=0$. The $\theta_1$ equation of
motion is automatically satisfied, while the $\alpha$ equation of motion becomes:
\begin{equation}
{\ddot{\alpha}}=-\frac{m^2}{4q}\sin\alpha((q-1)\cos\alpha+1)
\end{equation}
and from the Virasoro constraints we find
\begin{equation}
{\dot{\alpha}}^2=-\frac{1}{4q}\left(m^2(q\sin^2\alpha+\cos^2\alpha-2\cos\alpha+1)-16b^2\kappa^2\right)~.
\end{equation}
Expanding $\theta_1(\tau)=0+\eta(\tau)$, keeping the first order in $\eta$, changing
variables from $\eta(\tau)$ to $\eta(z)$, where $z=\cos\alpha$, and substituting the above, we finally
find the NVE:
\begin{equation}
\begin{split}
(z-1)(z+1)&\left(m^2(z-1)(qz+q-z+1)+16b\kappa^2\right)\ddot{\eta}(z)\\
&+\left((z-1)(4qz^2+4qz-4z^2+3z+1)m^2+48b^2\kappa^2z\right)\dot{\eta}(z)=0~.
\end{split}
\end{equation}
Applying the Kovacic algorithm to this NVE, we find that it completes, consistent with string motion on
the $\theta=\phi=\pi/2$ subspace being integrable for any $q$, as the numerical results above also indicate.
For $q=1$ this agrees with our expectations
of integrability, but of course for $q\neq 1$ this could just be an artifact of our string ansatz being too simple
to show any non-integrability. To actually prove integrability on this subspace one would need to construct a
Lax pair for the metric (\ref{CP2qnot1}), similarly to \cite{Arutyunov:2008if,Stefanski:2008ik}. It would be
interesting to study string motion on this deformed $\mathbb{C} \mathrm{P}^{2}$ further in order to prove or disprove integrability.
In the next section we present the analysis of a different ansatz for string motion on the backreacted background,
which exhibits puzzling behavior in the undeformed $\mathbb{C} \mathrm{P}^{3}$ limit.
\subsubsection{The localised string on $\mathrm{S}^2$} \label{S2localised}
In this section we consider a generic parametrisation of string motion such that the "induced" $\mathrm{SU}(2)$ left-invariant forms are $\omega_2=\omega_3=0~$, while $\omega_3=\mathrm{d} \varphi_1$.
The metric effectively becomes
\begin{equation} \label{effectiveansatz1}
\begin{split}
\mathrm{d} s^2=-\cosh^2\rho ~\mathrm{d} t^2+&\frac{1}{b^2}\Big(\mathrm{d} \theta} \def\Theta{\Theta} \def\vth{\vartheta^2 + q \mathrm{d} \alpha^2+ \sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta \mathrm{d} \varphi^2
+\prt{ \frac{q}{4} \sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} +\sin^4\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta}\mathrm{d} \varphi_1^2\\
&\quad-2 \sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta \mathrm{d}\varphi \mathrm{d}\varphi_1\Big)~.
\end{split}
\end{equation}
To restrict the string motion consistently to the above space we localise the string in AdS at $\rho=0$ and set $\theta_1= \psi_1=0~$ and for further convenience $\varphi=0$. By applying the formalism of section \ref{section:gen} to the full system of equations using the above parametrisation we may specify for the trivial part of the solution
\begin{equation}
t=\kappa \tau~,\quad \alpha} \def\dot{\alpha}{{\dot\alpha}=\alpha} \def\dot{\alpha}{{\dot\alpha}(\tau)~,\quad \theta} \def\Theta{\Theta} \def\vth{\vartheta=\theta} \def\Theta{\Theta} \def\vth{\vartheta(\tau)~,\quad \varphi_1 =m\sigma} \def\Sigma{\Sigma ~.
\end{equation}
To specify the remaining functions $\alpha} \def\dot{\alpha}{{\dot\alpha}$ and $\theta} \def\Theta{\Theta} \def\vth{\vartheta$ we look at their equations of motion \eq{eomnocycle}
\begin{eqnarray}\la{xi1a}
&&\ddot \alpha} \def\dot{\alpha}{{\dot\alpha} +\frac{m^2}{4 q} \prt{q \cos \alpha} \def\dot{\alpha}{{\dot\alpha}+2 \sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2} \sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta}=0~,\\
&&\ddot\theta} \def\Theta{\Theta} \def\vth{\vartheta+ \frac{m^2}{2}\sin^4\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2} \sin2\theta} \def\Theta{\Theta} \def\vth{\vartheta =0~,
\end{eqnarray}
where the first equation can be integrated to obtain the Virasoro constraint
\begin{equation}\la{xi2a}
q \dot{\alpha}^2+\dot{\theta}^2+ m^2 \prt{\frac{ q}{4}\sin^2 \alpha} \def\dot{\alpha}{{\dot\alpha}+\sin^4\frac{\alpha}{2} \sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta} -4 \kappa^2 b^2 =0~.
\end{equation}
\textbf{Case I: The $\theta=0$ plane}
We may choose to localise the string solution in the $\theta} \def\Theta{\Theta} \def\vth{\vartheta=0$ plane. The motion can be mapped to a solution of the undeformed theory with a different energy. The deformation parameters are absorbed in the energy of the solution by redefining it as
\begin{equation}
\tilde{\kappa}=\frac{\kappa b}{\sqrt{q}}~
\end{equation}
and the motion along the angle $\alpha$ is given by the inverse of the elliptic integral of first kind. Nevertheless, the solution we obtain by departing from $\theta} \def\Theta{\Theta} \def\vth{\vartheta=0$ can not be related to the undeformed system. By varying the solution in the invariant plane $\theta} \def\Theta{\Theta} \def\vth{\vartheta=p_\theta} \def\Theta{\Theta} \def\vth{\vartheta=0$ of the effective Hamiltonian
we get
\begin{equation}\la{puzzle1}
\eta''(z)\prt{16\kappa^2+m^2 q( z^2-1)}(z^2-1)+\eta'(z) 2 z\prt{8\kappa^2+m^2 q(z^2-1)}-m^2 q(1-z)^2 \eta(z)=0~,
\end{equation}
where we have expressed everything in terms of $z=\cos(\alpha} \def\dot{\alpha}{{\dot\alpha}(\tau))$ to get a differential equation with rational coefficients.
It turns out that the NVE does not have Liouvillian solutions for any value of $q$, even for $q=1$. We expect that string motion on $\mathbb{C} \mathrm{P}^{3}$ is integrable, and indeed we have performed the analogous analysis for several string solutions on round $\mathbb{C} \mathrm{P}^{3}$ (expressed, for instance, using the metrics (\ref{FSmetricangles}) or (\ref{roundmetric1})). As expected, no signs of non-integrability were found.
To understand this puzzling behaviour, it is useful to re-express the specific string ansatz
above in one of these other coordinate systems, which of course can only be done for $q=1$. In appendix
\ref{FBtoinst} it is shown how to convert to the coordinate
system (\ref{roundmetric1}). Examining the equations of motion for this ansatz in the new coordinate system, we find that there
are three non-trivial equations of motion instead of two in our original coordinate system, for our two functions
$(\theta(\tau),\alpha(\tau))$.\footnote{This is unlike, for instance, the $\mathbb{C} \mathrm{P}^{2}$ ansatz of the previous section which, when converted to the coordinate
system (\ref{roundmetric1}) using (\ref{CP2convert}) leads to just two non-trivial equations, whose NVE analysis is consistent with integrability.} It thus appears that, after converting to the round-type
coordinate system (\ref{roundmetric1}), the restriction of string motion to the space (\ref{effectiveansatz1})
has certain peculiarities, at least for $q=1$. It would be very interesting to study this further and understand,
directly in the original coordinate system, whether there is indeed an inconsistency or instability, and whether it
affects only the case $q=1$ (where we are able to map the string to the round $\mathbb{C} \mathrm{P}^{3}$ metric) or the whole solution.
We point out that, as usual, integrability is recovered for point-like particle motion.
\textbf{Case II: The $\theta=\frac{\pi}{2}$ plane}
To elaborate further on the puzzling behavior above we can also localise the string at the equator
of the sphere $\theta} \def\Theta{\Theta} \def\vth{\vartheta=\frac{\pi}{2}.$
The equation of motion for $\theta} \def\Theta{\Theta} \def\vth{\vartheta$ is satisfied and we remain with the equation of $\alpha} \def\dot{\alpha}{{\dot\alpha}$ \eq{xi1a} or its equivalent \eq{xi2a}. To find the NVE we perturb along the solution $\theta} \def\Theta{\Theta} \def\vth{\vartheta=\frac{\pi}{2}+\eta\prt{t}$ to get
\begin{eqnarray}
&&\eta''(z)\prt{16\kappa^2+m^2(z-1)\prt{1+q+(q-1)z}(z^2-1)} \\ \nonumber
&&+\eta'(z)\prt{16 \kappa^2 z +m^2(z-1)\prt{1+z+2q z+2(q-1)z^2}}+m^2 q (z-1)^2\eta(z)=0~.
\end{eqnarray}
This equation again does not admit Liouvillian solutions for any $q$ as in \eq{puzzle1} giving the same puzzling behavior for the same reasons discussed there.
\section{Conclusions and Outlook}
In this work we considered two classes of AdS/CFT backgrounds dual to gauge theories with backreacted
flavors. We have shown that the corresponding theories at zero temperature are not integrable, at least in the sector dual to semiclassical strings, and we have studied the existence of chaos for string motion on these backgrounds.
In the case of four-dimensional theories with $\mathcal{N}=2$ supersymmetry, corresponding to the D3/D7 gravity backreacted gravity background, we have found evidence that shows the non-integrability of the theory. We find analytically closed string solutions that are not integrable, with the integrability restored in the point-like limit for particle motion or when we switch off the backreaction of the D7-flavor branes. This is also the case in the backreacted flavor solution of the D6-branes in the ABJM theory. We conclude that the backreacted flavor deformation breaks the integrability present in the three and four dimensional theories. We have also discussed a puzzling behavior where for a certain string ansatz we did not find Liouvillian solutions of the corresponding NVE even when switching off the backreaction, where the background is just $\mathbb{C} \mathrm{P}^{3}$. As discussed in section \ref{S2localised}, although this puzzling string solution appears valid in our original coordinate system (which is adapted to introducing backreaction), it has some unwelcome features after mapping it to a different metric on $\mathbb{C} \mathrm{P}^{3}$. So we interpret the non-integrable behaviour at $q=1$ as an inconsistency of our string ansatz, and not as a sign of non-integrability of the background. It would be interesting to further study this issue, as it could provide insight into additional (and not previously noticed in the literature) requirements on string solutions, in order for the analytic non-integrability method to be applicable.
In this work we considered two backgrounds with backreacted flavor, one with localised flavor branes and one with the flavor branes smeared along the internal direction. Our belief is that we should expect to see similar behaviours for an even wider class of theories where the virtual quark loops are taken into account and contribute to the gauge propagators. Integrability for these theories should be lost. A strong argument, is that the Hamiltonian of the strings in our methods written in the phase space coordinates capturing the information of non-integrability and chaos, is expected always to be involved enough in the case of backreacting flavor geometries. Given the sensibility of non-integrability and chaos to the Hamiltonian, and in direct analogy to mechanical systems, we can expect that integrability is lost in the presence of unquenched flavors for generic theories.
An important portion of our work consists of analyzing the presence of chaos in the flavor backreacted theories. We have quantified the chaotic motion by computing the leading Lyapunov exponent and observing its dependence on the $N_c, N_f$ and 't Hooft parameters of the theory. Interestingly we find that the Lyapunov exponent in the Veneziano regime converges to a certain value which depends on the energy density of the system and the parameters of the theory, but depends weakly on the number of flavors. We then reduce the number of flavors and, being a bit optimistic
regarding the range of validity of the backreacted solution, we see that the maximum value of the Lyapunov exponent is obtained very quickly as we increase the flavors. Independent phenomena related to properties of the bound states of heavy quarks have been found to be weakly dependent on the numbers of flavors or colors beyond a certain low number,
exactly like the strength of chaos we studied here. It would be interesting if one could establish any connection or common explanation for those phenomena and our observations regarding the chaos. We also note that our analysis is done in zero temperature and the gravity dual theory has no black hole horizon.
We should emphasise that although both of the cases we consider are very good models for backreacted flavor backgrounds and have been used widely in the literature, neither of them is exactly dual to a gauge theory with unquenched flavor: In the D3/D7 case we work in the ``near-core'' region, where the background admits a tractable analytic solution. In the D2/D6 case we work with the background of \cite{Conde:2011sw} which was obtained by considering a smeared distribution of D6 branes instead of a localised one. This leads to a simple squashed $\mathbb{C} \mathrm{P}^{3}$ internal space instead of the much more involved tri-Sasaki space of the localised case, and an associated reduction of supersymmetry from $\mathcal{N}=3$ to $\mathcal{N}=1$. In both cases the full, localised backgrounds are significantly more complicated and it is natural to expect that the loss of integrability seen in our simpler cases
will be a feature of those models as well. In general it is interesting to ask to what extent the analytic non-integrability method, applied to non-exactly known dual backgrounds, can reliably capture the physics of the full gauge theory. Our work can be considered as providing two such examples, which can possibly be validated by further future work going beyond the approximations discussed above.
Our study of the unquenched ABJM theory can act as a starting point for several other studies. For special values of the parameter that controls the squashing of the space and the number of flavors, the background appears in other gauge/gravity dualities. The backreaction of the $N_f$ flavor D6-branes, smeared over a six-dimensional nearly K\"{a}hler manifold on the $N_c$ color branes, flows to an $\mathrm{AdS}_4$ fixed point dual to Chern-Simons matter theory and which is the special case of our metric \cite{Faedo:2015ula}. The special case $q=5$ corresponds to an ${\cal N}=1$ supergravity background, being a gravity dual of superconformal Chern-Simons matter theory with $\mathrm{SO}\prt{5}\times \mathrm{U}\prt{1}$ global symmetry \cite{Ooguri:2008dk}.
It should also be mentioned that a full study of analytic non-integrability in flavor backgrounds (whether backreacted or not) should also include the open-string sector, dual to gauge theory operators with fundamental fields at the endpoints instead of traces of adjoint fields. The analytic non-integrability studies performed so far in the literature have been based on closed string solutions, and extending the method to open strings would open up several new avenues for the study of non-integrability. Beyond theories with flavors, this might also include the sector of open strings attached to giant graviton operators, as well as defect CFTs, where the study of integrability was recently initiated in \cite{deLeeuw:2015hxa}. It would be interesting to apply our methods to this sector.
\subsection*{Acknowledgments}
We are thankful to C.S. Chu, K. Hashimoto, C. Kristjansen, K. Sfetsos and K. Yoshida for useful
conversations and correspondence.
This work of DG is supported in part by the National Center of Theoretical Science
(NCTS) and the grants 101-2112-M-007-021-MY3 and 104-2112-M-007 -001 -MY3 of the
Ministry of Science and Technology of Taiwan. KZ acknowledges support from the National Research Foundation
of South Africa under grant CSUR 93735.
We are also grateful to the South African National Institute for Theoretical Physics (NITheP) for supporting a visit by DG to the University of Pretoria and to NCTS for supporting a visit of KZ to NCTS, Hsinchu. DG would like to thank the Albert Einstein Institute for Fundamental Physics,
Bern University, for hospitality during the final stages of this project. KZ would like to thank the Niels Bohr Institute, Copenhagen, for hospitality and the opportunity to present a preliminary version of this work.
\begin{appendices}
\section{The geometry of $\mathbb{C} \mathrm{P}^{3}$}
The internal part of the metric of \cite{Conde:2011sw} for generic values of the deformation parameter $q$, is that
of a squashed $\mathbb{C} \mathrm{P}^{3}$. This is a coset $\mathrm{Sp}(2)/\mathrm{U}(2)$. In this appendix we will review some aspects of this
geometry, focusing also on the limit $q=1$ where the metric becomes just $\mathbb{C} \mathrm{P}^{3}$.
\subsection{The squashed $\mathbb{C} \mathrm{P}^{3}$ metric}
Let us start from the metric describing an $\mathrm{SU}(2)$ instanton on $\mathrm{S}^4$:
\begin{equation} \label{S2S4fibremetric}
\mathrm{d} s^2=q \frac{\mathrm{d} w_a \mathrm{d} \bar{w}^a}{(1+w_a\bar{w}^a)^2}+\frac{(\mathrm{d} u+\mathcal{A})(\mathrm{d} \bar{u}+\bar{\mathcal{A}})}{(1+u\bar{u})^2}~,
\end{equation}
where the $w_a,\bar{w}^a$ are coordinates on the $\mathrm{S}^4$ and $u$ is the projective coordinate on $\mathrm{S}^2$ (thought of as
a $\mathbb{C} \mathrm{P}^{1}$). $\mathcal{A}$ is an $\mathrm{SU}(2)$ instanton
given by
\begin{equation}
\mathcal{A}=\frac{(\bar{w}_1+w_2u)(u\mathrm{d} w_1-\mathrm{d} \bar{w}_2)+(\bar{w}_2-w_1 u)(u\mathrm{d} w_2+\mathrm{d} \bar{w}_1)}{1+w_a\bar{w}^a}\;,
\end{equation}
while $q$ is the squashing parameter which controls the length of the fibre.\footnote{We note that in most of the
literature the squashing parameter is placed
in front of the second term in (\ref{S2S4fibremetric}) and denoted $\lambda^2$. Therefore, effectively our $q$ corresponds
to $1/\lambda^2$.} It is known (e.g. \cite{Aldazabal:2007sn}) that there are two special
values of $q$, which are distinguished by the metric being Einstein: At $q=1$ the metric is equivalent to the
standard Fubini-Study metric on $\mathbb{C} \mathrm{P}^{3}$, which is of course K\"ahler. At $q=2$, the metric is instead
\emph{nearly-K\"ahler}. For all other values of $q$ the metric is not Einstein.
To bring (\ref{S2S4fibremetric}) into the form of the flavor backreacted metric of \cite{Conde:2011sw}, we can start
by choosing coordinates for the $\mathrm{S}^4$ part as
\begin{equation} \label{S4coords}
w_1=\tan\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2} e^{i\frac{\psi_1+\varphi_1}2}\cos\frac{\theta_1}{2}\;,\quad
w_2=-\tan\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2} e^{i\frac{\psi_1-\varphi_1}2}\sin\frac{\theta_1}{2}~.
\end{equation}
Then we find that
\begin{equation}
\mathrm{d}_{\mathrm{S}^4}^2=\frac{1}{4}\prt{\mathrm{d}\alpha} \def\dot{\alpha}{{\dot\alpha}^2+\frac{1}{4}\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} \sum_i\omega_i^2}~,
\end{equation}
where the $\mathrm{SU}(2)$ left-invariant forms we use were given in (\ref{leftinv1}-\ref{leftinv3}).
Turning to the $\mathrm{S}^2$ part, we express
\begin{equation} \label{S2coords}
u=-i\cot\frac{\theta}{2} e^{-i\varphi}\;
\end{equation}
such that
\begin{equation}
\frac{\mathrm{d} u\mathrm{d} \bar{u}}{(1+u\bar{u})^2}=\frac{1}{4}\left(\mathrm{d}\theta^2+\sin^2\theta\mathrm{d}\varphi^2\right)~.
\end{equation}
Combining these substitutions we can straightforwardly check that (\ref{S2S4fibremetric}) takes the form
\begin{equation}
\mathrm{d} s^2_{\mathbb{C} \mathrm{P}^{3}}=\frac{1}{4}\prt{q\left(\mathrm{d} \alpha} \def\dot{\alpha}{{\dot\alpha}^2+\frac{1}{4}\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} \sum_i\omega_i^2\right)+E_1^2+E_2^2}~,
\end{equation}
with the $E_i$ defined in Section \ref{Section:ABJM}. We see that the flavor-backreacted metric of
\cite{Conde:2011sw} is equivalent to the squashed $\mathbb{C} \mathrm{P}^{3}$ metric. It should be noted,
however, that the physics of the backreacted flavor construction restricts $q$ to take the values discussed
below (\ref{qvalues}), which do not include the nearly-K\"ahler case $q=2$.
As mentioned, taking $q=1$ in (\ref{S2S4fibremetric}) one obtains undeformed $\mathbb{C} \mathrm{P}^{3}$, with the instanton-type
metric that appears, for instance, in the twistor construction. We will call this the ``twistor metric'' for $\mathbb{C} \mathrm{P}^{3}$.
Its explicit symmetries are those of the coset $\mathrm{Sp}(2)/\mathrm{U}(2)$. It will be useful to compare it with the ``round'' $\mathrm{SU}(4)/\mathrm{U}(3)$ metric on
$\mathbb{C} \mathrm{P}^{3}$, which we will do in the following.
\subsection{The Fubini-Study metric on $\mathbb{C} \mathrm{P}^{3}$}
We start by reviewing the Fubini-Study metric on $\mathbb{C} \mathrm{P}^{3}$, which is the one that has been
used in studies of the ABJM geometry in the literature.
Since $\mathbb{C} \mathrm{P}^{3}$ is a K\"ahler manifold, a metric on it can be found through a potential:
\begin{equation}
\mathrm{d} s^2 = \partial \overline\del K = \frac{1}{2} g_{i{\bar{\jmath}}} \left( \mathrm{d} z^i\otimes \mathrm{d} \bar{z}^{{\bar{\jmath}}}+\mathrm{d}\bar{z}^{{\bar{\jmath}}}\otimes \mathrm{d} z^i\right)~,
\end{equation}
with
\begin{equation}
g_{i{\bar{\jmath}}}=\frac{\partial^2}{\partial z^i \partial \bar{z}^{{\bar{\jmath}}}} K~,
\end{equation}
where $z, \bar{z}$ are complex coordinates on the manifold.
The K\"ahler potential can be expressed in terms of inhomogeneous coordinates $Z^I$, $I=1\ldots4$, as
\begin{equation}
K=\ln \bar{Z}^I Z^I~.
\end{equation}
This makes explicit the $\mathrm{SU}(4)$ symmetry of $\mathbb{C} \mathrm{P}^{3}$. However, it is usually convenient to rewrite the metric
in terms of homogeneous coordinates $z^i$, $i=1\ldots 3$, defined as $z^1=Z^2/Z^1,z^2=Z^3/Z^1,z^3=Z^4/Z^1$,
so that $K$ becomes:
\begin{equation}
K=\ln\left(1+\bar{z}^i z^i\right)~.
\end{equation}
Starting from this potential one obtains the Fubini-Study metric as
\begin{equation}\label{metricz}
\mathrm{d} s^2_{\mathbb{C} \mathrm{P}^{3}}=\frac{\mathrm{d} z^i\mathrm{d} \bar{z}^i}{1+z^k\bar{z}^k}-\frac{(\bar{z}^i\mathrm{d} z^i) (z^j\mathrm{d} \bar{z}^j)}{(1+z^k\bar{z}^k)^2}~,
\end{equation}
while the K\"ahler form is given by
\begin{equation} \label{Kahlerform}
\omega=\frac{i}{1+\bar{z}^k z^k}\mathrm{d} z^i\wedge \mathrm{d} \bar{z}^i-\frac{i}{(1+\bar{z}^k z^k)^2} \bar{z}^i \mathrm{d} z^i\wedge z^j\mathrm{d}\bar{z}^j~.
\end{equation}
For applications to string motion, it is convenient to parametrise these coordinates in terms of angles. One commonly
used parametrisation (see e.g. \cite{Pope:1984bd}) is
\begin{equation}
\begin{split}
z^1&=\tan \frac{\theta} \def\Theta{\Theta} \def\vth{\vartheta}{2}\sin\alpha} \def\dot{\alpha}{{\dot\alpha}\sin\frac{\theta} \def\Theta{\Theta} \def\vth{\vartheta_1}{2}e^{i \frac{\prt{\psi -\chi}}{2}}e^{i\phi}~,\\
z^2&=\tan \frac{\theta} \def\Theta{\Theta} \def\vth{\vartheta}{2}\sin\alpha} \def\dot{\alpha}{{\dot\alpha}\cos\frac{\theta} \def\Theta{\Theta} \def\vth{\vartheta_1}{2}e^{i \frac{\prt{\psi +\chi}}{2}}e^{i\phi}~,\\
z^3&=\tan \frac{\theta} \def\Theta{\Theta} \def\vth{\vartheta}{2}\cos\alpha} \def\dot{\alpha}{{\dot\alpha} e^{i\phi}~,
\end{split}
\end{equation}
where $0\leq\theta} \def\Theta{\Theta} \def\vth{\vartheta,\theta} \def\Theta{\Theta} \def\vth{\vartheta_1\leq \pi,~ 0\leq \phi,\chi\leq 2 \pi,~ 0\leq\alpha} \def\dot{\alpha}{{\dot\alpha}\leq\frac{\pi}{2},~ 0\leq\psi\leq 4 \pi$. The metric becomes
\begin{equation} \label{FSmetricangles}
\mathrm{d} s^2=\sin^2 \frac{\theta} \def\Theta{\Theta} \def\vth{\vartheta}{2}\prt{\mathrm{d} \alpha} \def\dot{\alpha}{{\dot\alpha}^2+\frac{1}{4}\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha}\prt{\omega_1^2+\omega_2^2+ \cos^2\alpha} \def\dot{\alpha}{{\dot\alpha} \omega_3^2}}+\frac{1}{4}\prt{\mathrm{d}\theta} \def\Theta{\Theta} \def\vth{\vartheta^2 +\sin^2\theta} \def\Theta{\Theta} \def\vth{\vartheta\prt{\mathrm{d}\phi+\frac{1}{2}\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha} \omega_3}^2}~,
\end{equation}
where the $\omega_i$ are as in (\ref{leftinv1}) with the substitutions $\psi_1\rightarrow\psi$, $\phi_1\rightarrow \chi$.
Another frequently used metric can be obtained from (\ref{metricz}) through the substitution (e.g. \cite{Abbott:2008qd})
\begin{equation} \label{FStoround1}
z^1=\cot\frac{\theta_2}{2} e^{i\phi_2}\;,
\;\;z^2=\frac{\cot\xi\cos\frac{\theta_1}{2}e^{-\frac{i}{2}(\psi-\phi_1-\phi_2)}}{\sin\frac{\theta_2}{2}}\;,
\;\;z^3=\frac{\cot\xi\sin\frac{\theta_1}{2}e^{-\frac{i}{2}(\psi+\phi_1-\phi_2)}}{\sin\frac{\theta_2}{2}}\;,\;\;
\end{equation}
leading to
\begin{equation} \label{roundmetric1}
\begin{split}
\mathrm{d} s^2&=\mathrm{d} \xi^2+\frac14\cos^2\xi(\mathrm{d}\theta_1^2+\sin^2\theta_1\mathrm{d}\phi_1^2)
+\frac14\sin^2\xi(\mathrm{d}\theta_2^2+\sin^2\theta_2\mathrm{d}\phi_2^2)\\
&+\frac14\sin^2\xi\cos^2\xi(\mathrm{d}\psi-\cos\theta_1\mathrm{d}\phi_1+\cos\theta_1\mathrm{d}\phi_2)^2~.
\end{split}
\end{equation}
This parametrisation is used e.g. in \cite{Cvetic:2000yp}. In the next section we show how some of our anz\"atze for string motion, written in the twistor-like coordinate
system which is appropriate for squashing, can be expressed in this metric, and we will also make use of it in Appendix \ref{T11restriction}, where we study
integrability on a $T^{11}$ subspace of $\mathbb{C} \mathrm{P}^{3}$.
\subsection{From the Fubini-Study metric to the twistor metric} \label{FBtoinst}
The relation between the Fubini-Study metric (\ref{metricz}) and the twistor metric ((\ref{S2S4fibremetric}) with $q=1$)
has been discussed in \cite{Bellucci:2003yx} in the context of relating the Hall mechanics of a particle on $\mathrm{S}^4$ and $\mathbb{C} \mathrm{P}^{3}$ (see also \cite{Krivonostalk} for more details and generalisations). The relation is simply the following:
\begin{equation} \label{redef}
z^1=w_1 u -\bar{w}_2\;,\quad z^2=w_2 u+\bar{w}_1\;,\quad z^3=u~.
\end{equation}
Performing this coordinate change and substituting (\ref{S4coords}) and (\ref{S2coords}), it can be checked that
one obtains the twistor metric, while, as also mentioned in \cite{Conde:2011sw}, the K\"ahler form
(\ref{Kahlerform}) can be expressed in terms of the $E^{1,2}, \mathcal{S}^\alpha$ and $\mathcal{S}^i$ forms defined in Section \ref{Section:ABJM} as
\begin{equation}
\omega=\frac{1}{2}\left(E^1\wedge E^2-\mathcal{S}^\alpha\wedge \mathcal{S}^3-\mathcal{S}^1\wedge \mathcal{S}^2\right)\;.
\end{equation}
Of course, in order to compare explicit string solutions we would like to go beyond the identification (\ref{redef})
and relate angle coordinates for the Fubini-Study metric to those for the twistor metric. Although this is not
straightforward to do in full generality, it can be done if we restrict to specific subspaces of
$\mathbb{C} \mathrm{P}^{3}$.
One such case concerns the ansatz discussed in section \ref{S2localised}, which we repeat
here for convenience:
\begin{equation}
\alpha=\alpha(\tau)\;,\quad\theta=0\;,\quad\phi_1=m\sigma\;,\quad\phi=\frac{\pi}2\;,\quad\theta_1=\theta_1(\tau)\;,\quad\psi_1=\frac{\pi}2 \;.
\end{equation}
Combining the expressions (\ref{S4coords}), (\ref{S2coords}) and (\ref{redef}) we find
that this corresponds to the following choice for the $z^i$ coordinates:
\begin{equation} \label{ziansatz3}
z^1=-i\,e^{{{im\sigma}\over{2}}}\,\tan\frac{\alpha(\tau)}{2}\cot\frac{\theta(\tau)}2\;,\quad
z^2=e^{-\frac{i m \sigma}{2}}\tan\frac{\alpha(\tau)}{2}\;,\quad z^3=-i \cos\frac{\theta(\tau)}{2}~.
\end{equation}
Let us now convert the above parametrisation to the coordinate system (\ref{roundmetric1}), which is commonly used in
studies of $\mathbb{C} \mathrm{P}^{3}$. To do so, we make the following identification of coordinates:
\begin{equation}
\xi(\tau)=\frac{\alpha(\tau)-\pi}{2},\quad\theta_1(\tau)=\theta(\tau),\quad \theta_2(\tau)=\theta(\tau),\quad \phi_1=m\sigma+\frac{3\pi}2,\quad
\phi_2=-\frac{\pi}{2},\quad\psi=0\;.
\end{equation}
With this choice, and also redefining (\ref{FStoround1}) as $z^1\rightarrow z^3, z^2\rightarrow z^2, z^3\rightarrow z^2$,
the inhomogeneous coordinates (\ref{FStoround1}) become equal to (\ref{ziansatz3}).
For completeness, let us also mention how the $\mathbb{C} \mathrm{P}^{2}$ ansatz (\ref{CP2ansatz}) can be written in the coordinates of (\ref{roundmetric1}). One redefines
the $z^i$ as above, this time with the following identifications:
\begin{equation} \label{CP2convert}
\xi(\tau)=\frac{\alpha(\tau)-\pi}{2}\;,\quad \theta_1=\theta_2=\frac{\pi}{2}\;,\quad \phi_1=m\sigma-\frac{\pi}2\;,\quad\phi_2=\pi,\quad \psi(\tau)=-\theta_1(\tau)\;.
\end{equation}
\subsection{Restriction to $\mathbb{C} \mathrm{P}^{2}$} \label{CP2restriction}
As noted in \cite{Conde:2011sw}, an interesting subspace of the $\mathrm{S}^2$-bundle
over $\mathrm{S}^4$ that we have been discussing arises when the two-sphere angles take the values
\begin{equation}
\theta} \def\Theta{\Theta} \def\vth{\vartheta=\frac{\pi}{2}~,\quad \varphi=\frac{\pi}{2}~,
\end{equation}
giving the four-dimensional metric
\begin{equation} \label{CP2qnot1}
ds^2= \frac{1}{4}\prt{q d\alpha} \def\dot{\alpha}{{\dot\alpha}^2+ \frac{1}{2}\prt{1+q+(1-q)\cos\alpha} \def\dot{\alpha}{{\dot\alpha}}\sin^2 \frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\prt{\omega_1^2+\omega_3^2}+\frac{q \sin^2\alpha} \def\dot{\alpha}{{\dot\alpha}}{4} \omega_2^2}\;.
\end{equation}
If we turn off the backreaction by setting $q=1$, this metric is
\begin{equation}
ds^2_{CP^2}= \frac{1}{4}\prt{d\alpha} \def\dot{\alpha}{{\dot\alpha}^2+ \sin^2 \frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\prt{\omega_1^2+\omega_3^2}+\frac{\sin^2\alpha} \def\dot{\alpha}{{\dot\alpha}}{4} \omega_2^2}~.
\end{equation}
This is the Fubini-Study metric on $\mathbb{C} \mathrm{P}^{2}$. As string motion on $\mathbb{C} \mathrm{P}^{2}$ is expected to be integrable, it is
worthwhile to perform an analysis of integrability on this subspace, for $q\neq 1$ as well.
This is addressed in section \ref{CP2section}.
\subsection{Seven-sphere lift}
The family of squashed $\mathbb{C} \mathrm{P}^{3}$ metrics is also relevant for the construction of squashed seven-sphere
metrics, which are $\mathrm{S}^1$ bundles over a squashed $\mathbb{C} \mathrm{P}^{3}$ base. The round metric on $\mathrm{S}^7$ is equivalent to
the metric obtained via an $\mathrm{S}^1$
bundle over the $q=1$ $\mathbb{C} \mathrm{P}^{3}$ metric. There is one other value of $q$ for which the squashed $\mathrm{S}^7$
space is Einstein, which is $q=5$: Even though the squashed $\mathbb{C} \mathrm{P}^{3}$ metric is not Einstein for this
value, the corresponding squashed seven-sphere is. The AdS/CFT dual theory to this squashed sphere
has been considered in \cite{Ooguri:2008dk}.
Let us briefly discuss the uplift of $\mathbb{C} \mathrm{P}^{3}$ (for $q=1$) to $\mathrm{S}^7$, following the discussion in
\cite{Conde:2011sw}. Calling the additional direction $\psi$, and
recalling that $\theta$ and $\varphi$ are the $\mathrm{S}^2$ angles, we introduce a new set of left-invariant forms,
\begin{equation}
\begin{split}
\tilde{\omega}^1&=-\sin\varphi\mathrm{d}\theta+\cos\varphi\sin\theta\mathrm{d}\psi~,\\
\tilde{\omega}^2&=\cos\varphi\mathrm{d} \theta+\sin\varphi\sin\theta\mathrm{d} \psi~,\\
\tilde{\omega}^3&=\mathrm{d} \varphi+\cos\theta\mathrm{d}\psi~,
\end{split}
\end{equation}
in terms of which the $\mathrm{S}^7$ metric is\footnote{For the actual uplift of ABJM to M-theory one also needs to perform a $\mathbb{Z}_k$
orbifold of the $\psi$ angle.}
\begin{equation} \label{Hopfmetric}
\mathrm{d} s^2 =\frac{1}{4}\left(\mathrm{d} s_{\mathrm{S}^4}^2+\sum_{i=1}^3 \left(\tilde{\omega}^i+A^i\right)^2\right)~,
\end{equation}
where we recall that $A^i=-\sin^2\frac{\alpha} \def\dot{\alpha}{{\dot\alpha}}{2}\omega^i=-\frac{1-\cos\alpha}{2}\omega^i$. We see that the uplift of the twistor metric of $\mathbb{C} \mathrm{P}^{3}$
naturally exhibits $\mathrm{S}^7$ as an $\mathrm{S}^3$ (Hopf) fibration over $\mathrm{S}^4$.
Similarly to the discussion above for $\mathbb{C} \mathrm{P}^{3}$, there is a question of how to show the equivalence of the above
$\mathrm{S}^7$ metric to the more usual round metric on $\mathrm{S}^7$. This is addressed e.g. in \cite{Hatsuda:2009vj}. For
completeness we sketch their construction below, while referring to their work for additional details. Let us start by
writing (\ref{Hopfmetric}) more explicitly (also suppressing the summation symbol) as
\begin{equation}
\mathrm{d} s^2=\frac{1}{4}\mathrm{d}\alpha^2+\frac{1}{16}\sin^2\alpha \omega_i^2+\frac{1}{4}(\tilde{\omega}_i+A_i)^2~,
\end{equation}
in order to better exhibit the roles of the two sets of left-invariant forms.
Let us now write
\begin{equation} \label{omegashift}
\omega_i=\omega_i-\tilde{\omega}_i+\tilde{\omega}_i=\nu_i+\tilde{\omega}_i~,
\end{equation}
i.e. we have defined a new one-form $\nu_i=\omega_i-\tilde{\omega}_i$. Then
\begin{equation}
\frac{\sin^2\alpha}{16}\omega_i^2=\frac{\sin^2\alpha}{16}(\nu_i^2+2 \nu_i \tilde{\omega}_i +\tilde{\omega}_i^2)~,
\end{equation}
while
\begin{equation}
(\tilde{\omega}_i-\frac{1-\cos\alpha}{2}\omega_i)^2=\cos\alpha \tilde{\omega}_i^2-(1-\cos\alpha)\tilde{\omega}_i\nu_i
+\frac{(1-\cos\alpha)^2}{4}(\nu_i^2+2\nu_i \tilde{\omega}_i +\tilde{\omega}_i^2)~.
\end{equation}
Adding these terms, and also substituting $\alpha=2\hat{\alpha}$, a short calculation leads to
\begin{equation}
\frac{1}{4}\mathrm{d}\alpha^2+\frac{1}{16}\sin^2\alpha \omega_i^2+\frac{1}{4}(\tilde{\omega}_i+A_i)^2
=\mathrm{d}\hat{\alpha}^2+\frac14\sin^2\hat{\alpha}~ \nu_i^2+\frac14\cos^2\hat{\alpha} ~\tilde{\omega}_i^2~.
\end{equation}
The right-hand side is what is commonly considered as the round metric on $\mathrm{S}^7$. For more details on how
this relation between the metrics is derived, we refer to \cite{Hatsuda:2009vj}. Of course, although (\ref{omegashift})
is suggestive, the actual identification of angles that implements this transformation is not straightforward.
\section{Strings on the $T^{11}$ restriction of $\mathbb{C} \mathrm{P}^{3}$} \label{T11restriction}
An interesting feature of the round $\mathbb{C} \mathrm{P}^{3}$ metric is that it admits a restriction to a 5-dimensional $T^{11}$
space. This can be most clearly seen starting from the metric (\ref{roundmetric1}) \cite{Cvetic:2000yp}. The constant $\xi$
surfaces are $\mathrm{U}(1)$ bundles on $\mathrm{S}^2\times \mathrm{S}^2$, and as such the restriction of (\ref{roundmetric1}) to constant $\xi$
has the geometric features of $T^{11}$. However, the metric one obtains by taking $\xi$ constant is not the Einstein metric on $T^{11}$,
which requires relations between the coefficients of the three factors that cannot be satisfied for any choice of $\xi$.
Integrability of string motion on $T^{11}$ (and its generalisations with arbitrary coefficients for
the three factors) was considered in \cite{Basu:2011fw}, where it was established that these backgrounds
are not integrable. So one can wonder how they can arise as foliations of the integrable $\mathbb{C} \mathrm{P}^{3}$ geometry.
We can examine this question by considering string motion on $\mathbb{C} \mathrm{P}^{3}$, parametrised as in (\ref{roundmetric1}), and
choosing an ansatz for a string that lives at constant $\xi$:
\begin{equation}
\begin{split}
t(\tau,\sigma)&=k\tau\;,\quad\xi(\tau,\sigma)=\xi\;,\quad\theta_1(\tau,\sigma)=\theta_1(\tau)\;,\quad
\phi_1(\tau,\sigma)=m\sigma\;,\;\\
&\theta_2(\tau,\sigma)=\theta_2(\tau)\;,\quad \phi_2(\tau,\sigma)=n\sigma\;,\quad
\psi(\tau,\sigma)=0~.
\end{split}
\end{equation}
While most equations of motion following from this ansatz are equivalent to those of \cite{Basu:2011fw},
there is an additional equation coming from the $\xi$ equation of motion:
\begin{equation}
\dot{\theta}_1^2-\dot{\theta}_2^2+(\cos^2\xi-\sin^2\xi)(m\cos\theta_1-n\cos\theta_2)^2 -m^2\sin^2\theta_1+n^2\sin\theta_2~.
\end{equation}
Following \cite{Basu:2011fw}, we take our straight line solution to be $\theta_2=0$. It turns out
that the additional equation of motion then restricts $\theta_1$ to take a constant value:
\begin{equation}
\bar{\theta}_1=\arccos\left(\frac23\frac{(2\cos^2\xi-1)n}{m\cos^2\xi}\right)~.
\end{equation}
Perturbing $\theta_2(\tau)=0+\eta(\tau)$, linearising and substituting the solution for $\theta_1$,
we find the NVE
\begin{equation}
\ddot{\eta}+\frac{n^2}{3}(1+\cos^2\xi)=0~.
\end{equation}
This is integrable, unlike the NVE found for $T^{11}$ in \cite{Basu:2011fw} (where $\theta_1$ was
a function of $\tau$). So it appears that the
additional EOM has led us to a simpler NVE which does not exhibit non-integrable behaviour. This is
of course consistent with our expectations from integrability of the $\mathbb{C} \mathrm{P}^{3}$ sigma model.
\end{appendices}
|
\section{Introduction}
Cloud Computing \cite{cc} has changed the way IT services are provisioned, managed, and delivered to users. Companies do not have to over-provision costly hardware infrastructure and personnel to offer large scale services. Cloud Computing allows services to scale as needed and companies only pay for the amount of resources used. It can be considered as a pool of computing resources designed to provide ``a computing function as a utility" \cite{green}. In addition, through statistical multiplexing, public clouds, such as Amazon EC2 \cite{ec2}, are able to provide such flexible and scalable services at minimum costs.
Datacenters are the infrastructure upon which cloud computing services are provided. To improve service provisioning, such infrastructures are usually spanned over multiple physical sites (multiple datacenters) closer to customers \cite{yahoo}. This also allows for higher reliability by keeping extra copies of user data on different datacenters. For example, Google's G-Scale network connects its 12 datacenters \cite{b4}. Also, Amazon EC2 platform runs on multiple datacenters located in different continents.
Many applications require traffic exchange between datacenters to synchronize data, access backup copies or perform geo-replication; examples of which include content delivery networks (CDNs), cloud storage and search and indexing services. Most of these transfers have to be completed prior to a deadline which is usually within the range of an hour to a couple of days and can also be as large as petabytes \cite{calendaring}. In addition, links that connect these datacenters are expensive to build and maintain. As a result, for any algorithm used to manage these resources, it is desired to have the following features:
\textbf{Efficiency:} We want to maximize the transfer goodput. This means maximizing both link utilization and number of successfully finished transfers.
\textbf{Speed:} For large scale applications that have millions of users, large number of transfers have to be processed and allocated. It is crucial that our algorithm is fast enough to allocate new requests in a short time.
In \cite{d2tcp}, authors propose Deadline-aware Datacenter TCP (D2TCP) which increases the number of transfers that complete prior to their assigned deadlines by adjusting the transmission rate of such transfers based on their deadlines. Also, multiple previous studies have focused on improving the efficiency and performance of inter-datacenter communications through proper scheduling of transfers. In \cite{b4, swan, calendaring}, authors propose B4, TEMPUS and SWAN. B4 and SWAN focus on maximizing link utilization through centralized traffic engineering and application of software defined networking. TEMPUS improves fairness by maximizing the minimum portion of transfers delivered to destination before the transfer deadlines. None of these schemes guarantees that admitted transfers are completed prior to their deadlines.
In \cite{amoeba}, authors propose Deadline-based Network Abstraction (DNA) which allows tenants to specify deadlines for transfers, and a system called Amoeba which guarantees that admitted transfers are completed prior to the specified deadlines. When a request is submitted, Amoeba performs admission control and decides whether the new request can be satisfied using available resources. If a transfer cannot be completed prior to its deadline, Amoeba tries to reschedule a subset of previously admitted requests to push traffic further away out of the new request's window. The admission process is performed on a first-come-first-served (FCFS) basis and requests are not preempted. In addition, each request is initially scheduled to finish as soon as possible.
In this paper, we propose RCD: a simple and fast traffic scheduler that minimizes the time required to perform the admission process. RCD does not have to move already admitted requests to decide whether a new request can be admitted. Also, it achieves high utilization by effectively using resources.
In the following sections, we first explain the problem of scheduling, present the rules based on which RCD operates, and compare RCD with Amoeba through simulations.
\section{The Scheduling Problem}
As mentioned in several previous studies \cite{swan,calendaring}, inter-datacenter traffic can be categorized into three different classes:
\textbf{Highpri:} This type of traffic is mainly initiated by users interacting with a website or online service. It is latency-sensitive as it directly affects user experience and has to be satisfied as soon as possible.
\textbf{Background:} Like previous traffic type, this one doesn't have a deadline either. It consists of throughput-oriented large transfers that are not critical, and have the least priority.
\textbf{Elastic:} Traffic requests of this type have to be fully satisfied before their deadline. Deadlines can be either hard or soft. Finishing an elastic request with a hard deadline after its deadline is pointless while in a soft deadline case, value of the transfer drops based off a value function depending on how late the transfer is finished.
Now imagine we have two datacenters $A$ and $B$ connected using one single link. New traffic requests of types mentioned above are generated in datacenter $A$ destined for datacenter $B$. Since highpri traffic is highly sensitive, it has to be sent on the link as soon as it is generated. The remaining bandwidth is first given to elastic requests and then background requests. The scheduling algorithm should decide how much bandwidth is given to each request at every time instance.
It is possible to estimate the volume of highpri traffic (it is usually between 5-15\% of the link capacity) \cite{amoeba}. As a result, we can set aside a percentage of the bandwidth on the link for this traffic and solve the scheduling problem for elastic and background traffic considering only the left over capacity. Elastic traffic is always allocated first as it has a higher priority. Background traffic goes through when there is no more elastic traffic to send. Throughout this paper, we focus on Elastic traffic and ways of allocating it.
Let us assume capacity $C$ is what is left after reserving the required bandwidth for highpri traffic. As in \cite{amoeba}, our aim here is to maximize the number of elastic transfers that finish before their deadlines as well as link utilization. In this paper, we will use linear programming to implement traffic scheduling.
In the following sections, we propose our method and compare its time complexity with the method proposed in \cite{amoeba}. Finally, we discuss how the ideas presented next can be further developed and applied to a network case where in addition to volume and deadline, each request is identified with a source and destination node.
\section{Rapid Close to Deadline Scheduling}
In order to flexibly allocate traffic with varying rates over time, we break the timeline into small timeslots. We do not assume an exact length for these timeslots as there are trade-offs involved. Having smaller timeslots can lead to large linear programs (LPs) which are time-consuming and inefficient to solve while having larger timeslots results in a less flexible allocation because the transmission rate is considered constant over a timeslot. In implementation of the previous study \cite{amoeba}, a length of 3 to 5 minutes is used.
We schedule traffic for future timeslots because focusing only on current timeslot provides poor performance as we cannot provide long term promises and cannot guarantee that important transfers finish before their deadlines \cite{calendaring}.
Assume we are allocating traffic for a timeline starting at $t_{now}$ representing current time and ending at $Td_{E}$ which corresponds to the latest deadline for all submitted requests. New requests may be submitted to the scheduler at any time. Each request is identified with three parameters, and is shown as $R(Q, Td_1, Td_2)$ in which $Q$ is the transfer volume, $Td_1$ is the first deadline where value of the transfer starts to drop if finished after this time, and $Td_2$ is the deadline after which finishing the transfer is pointless. For a request with a hard deadline $Td$, we have $Td = Td_1 = Td_2$.
In order to schedule traffic, we can create and solve a linear program (LP) involving all of submitted requests with demand and capacity constraints populated based off link capacities and request volumes. This LP has to be solved every time a new request is submitted and can result in changing the allocation of already scheduled requests. The problem with this approach is its high complexity (solving such a large LP over and over is computationally inefficient) as the frequency of arrivals increases.
In this section we discuss our proposal for bandwidth allocation. The useful property of elastic traffic is that it does not matter how soon or late we finish the transfer as long as we finish it before its deadline. We use this characteristic to further improve the allocation process based on the following rules:
\textbf{Rule 1:} Similar to previous schemes \cite{amoeba}, RCD does not support preemption. Preempting a request that is partly transmitted is wasteful. Also, it may result in thrashing if requests are consecutively preempted in favor of future requests.
\textbf{Rule 2:} To be fast, RCD does not change the allocation of already allocated traffic unless there is leftover bandwidth in current timeslot ($t_{now}$). In which case it fetches traffic from the earliest timeslot that is not empty and sends it. This is done until either we fully utilize the current timeslot or there is no more elastic traffic to send.
In addition, we make the following simplifying assumptions. Study of the algorithm without these assumptions is left for future work.
\begin{itemize}
\setlength{\itemsep}{0.5em}
\setlength{\labelsep}{1em}
\item Requests are instantly allocated upon arrival. They are allocated over timeslots for which $t > t_{now}$. If a request cannot be allocated upon arrival, it is rejected.
\item All requests have a hard deadline.
\item Bandwidth used by highpri traffic is constant and can be estimated with high accuracy.
\end{itemize}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{timeline}
\caption{Elastic traffic and different allocations}
\label{fig:timeline}
\end{figure*}
Now let's discuss an example considering our allocation rules. Fig. \ref{fig:timeline} shows three traffic allocation approaches and their behavior when a new traffic request is submitted:
\textbf{Allocation 1:} Tries to finish everything as soon as possible and does not change anything that is already allocated. This is the least complex approach; however, it fails to accommodate the new elastic request.
\textbf{Allocation 2:} Tries to finish everything as soon as possible but upon arrival of the new request, moves part of the already allocated request ahead to allocate the new request. Depending on how many already allocated requests are moved to accommodate the new request, this method has a variable chance of success. If it reallocates a large number of requests then it can get very slow as the frequency of arrivals increases. Conversely, it can have a small chance of success if only a few requests are reallocated.
\textbf{Allocation 3:} Based on our three rules, plans on finishing everything as late as possible, however, by sending traffic from future timeslots on $t_{now}$, it results in everything finished as soon as possible. Does not move allocated requests, only sends parts of already allocated requests to utilize the available capacity. Every time a new request is submitted, a small linear program involving only the new request is solved.
When a new elastic request $E$ is submitted, RCD creates a small LP to schedule it. The size of this LP is $(Td - t_{now})$. Assume the amount of bandwidth allocated for which at time $t$ is $E(t)$, the total demand is $Q$, and the deadline is $Td$. In addition, let's assume $C_{t}$ is the residual capacity on the link at timeslot $t$. We use the LP of equation \ref{eq:eq2} with the objective function of equation \ref{eq:eq1} to do the allocation. If following LP does not yield a feasible solution, we reject the request.
Now assume elastic requests $R_1, R_2, ~... ~ R_K$ are submitted to the scheduler in order. We want to show that upon arrival of $R_k, ~ 1 \le k \le K$, our allocation for previously admitted requests is in a way that we cannot increase the admission chance of $R_k$ by rearranging the allocation of already allocated requests. Let's show the deadline of $R_k$ as $Td_k$ and at any time $t$, show the latest deadline of all admitted requests as $Td_E$.
\begin{equation}
f(E) \triangleq \sum_{t=t_{now}+1}^{Td} E(t) \times (Td - t) \\
\label{eq:eq1}
\end{equation}
\begin{equation}
\begin{aligned}
Obj: ~~ & min\{f(E)\} \\
S.t: ~~ & \sum_{t = t_{now}+1}^{Td} E(t) = Q \\
& 0 \le E(t) \le C_{t}, ~~~ t_{now} < t \le Td
\label{eq:eq2}
\end{aligned}
\end{equation}
\textbf{Theorem 1:} \textit{If we draw a vertical line at time $Td_k \le Td_E$ in our traffic allocation, it is not possible to increase the free space before the line by moving traffic from left side of the line ($t \le Td_k$) to the right side ($Td_E \ge t > Td_k$).}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\textwidth]{proof1}
\caption{A general allocation used in theorem 1}
\label{fig:theorem1}
\end{figure*}
\textbf{Proof:} Let's assume we have the allocation shown in Fig. \ref{fig:theorem1}. To schedule all requests, we used the cost function of equation \ref{eq:eq1} which assigns a smaller cost to future timeslots. Assume we can move some traffic volume from left side to the right side. If so, this volume belongs to at least one of the admitted requests and that means we are able to decrease the allocation cost for that request even further. This is not possible because the LP in equation \ref{eq:eq2} gives the minimum cost solution. Therefore, we cannot move traffic from left side of the line to the right side. Doing so will either result in violation of link capacity or violation of a deadline.
Now let's assume a new elastic traffic arrives. If it can be allocated on the residual link capacity then we can accept it. If not, based on theorem 1, there is no way we can shift already allocated traffic so that we can accommodate the new elastic traffic. Therefore, we simply reject the request.
There may be cases when we want to reserve some bandwidth for an important traffic request while we don't have the traffic content yet. For example, at $6PM$ we find out an important traffic will arrive at $10PM$ and has to be fully transferred before midnight. We allocate such traffic the same as other elastic requests. This way we can be sure that when the traffic arrives, we have enough room to send it. If that important traffic is still not available by the time we finish sending all traffic requests that are allocated before it, we simply fetch traffic from the next closest request ahead of that transfer and if possible, push the important transfer to the next timeslot(s). Fig. \ref{fig:res} is an example of this case.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{reserve}
\caption{Reserving bandwidth for a future request}
\label{fig:res}
\end{figure}
If the reserved traffic is still not available when we are too close to its deadline that we cannot finish it anymore, we free the reserved bandwidth in favor of future elastic requests.
\section{Simulation Results}
We compare the performance and speed of RCD with Amoeba \cite{amoeba}. Other schemes, such as \cite{swan, calendaring}, are deadline-agnostic and have an effective link utilization of less than $50\%$ \cite{amoeba}. Amoeba, on the other hand, only accepts requests when it can guarantee that the deadline can be fully met.
\textbf{Simulation Setup:} As mentioned earlier, we assume highpri traffic takes a fixed amount of bandwidth and allocate the leftover among elastic requests. Simulation is performed for 576 timeslots each lasting 5 minutes which is equal to 2 days. We performed the simulations three times and calculated the average.
\textbf{Metrics:} Percentage of failed elastic requests, average link utilization, and average allocation time are the three metrics measured and presented.
\textbf{Workload:} We generate elastic requests based off a Poisson distribution of rate $1 \le \lambda \le 8$. The difference between the arrival time of requests and their deadlines follows an exponential distribution with an average of $12$ timeslots. In addition, the demand of each request also follows an exponential distribution with an average of $0.286$ (a maximum of $1$ unit of traffic can be sent in each timeslot on the link).
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{results_1_embedded}
\caption{Comparing Amoeba and RCD}
\label{fig:results_1}
\end{figure*}
Fig. \ref{fig:results_1} shows the aforementioned simulation metrics for both Amoeba and RCD. As can be seen, both algorithms result in similar failure rate and utilization. However, RCD achieves the same performance metrics with much less complexity: it is up to $15$ times faster than Amoeba. Also, the complexity of RCD grows very slowly as the frequency of arrivals increases: up to $1.6$ times while arrivals increase by a factor of $8$.
With regards to the trend for time complexity as shown in Fig. \ref{fig:results_1}, when the request arrival rate is small, most of the capacity is left unused. Therefore, Amoeba does not have to move already allocated requests to push in a new one. As the arrival rate increases, we see a higher utilization. Starting the arrival rate of $4$, utilization gets close to $1$ and we can see a huge jump in the time complexity of Amoeba (by a factor of $3.7$). That is because Amoeba has to move around multiple already allocated requests to push in the new request.
For an arrival rate of $8$ requests per timeslot, we see that both algorithms drop almost half of the requests. This can happen as a result of a failure in the network. For example, when a datacenter is connected using only two links and one of them fails for a few timeslots. While Amoeba can get really slow, RCD is able to handle such situations almost as fast as when there is low link utilization.
Up until now, the focus was only on one link. The network case will be discussed in the next section.
\section{RCD in Network}
For a single link, it was demonstrated how RCD allows for independent allocation of newly arriving requests: one only needs to know the residual link capacity to allocate a new request, and can be sure that there will be no performance degradation. For a network case however, it is not as simple. In this section we discuss one possible way of extending RCD for application in a network connecting multiple datacenters. Providing a detailed solution is beyond the scope of this paper and is identified as future work.
For a network case, each request is added a source and destination parameter and is shown as $R(Q, Td, S, D)$. We use the same rules as mentioned in section three. However, we have to augment the linear program in equation \ref{eq:eq2} to account for multiple paths and different capacity constraints of network links. Assume $L$ represents the set of network links $L_1, L_2, ~...~ L_m$ and $C_{t,l}$ is the residual capacity of link $l$ at time $t$. We want to allocate an elastic request with a source datacenter $A$ and a destination datacenter $B$. For a node $N$ in the network, $N_{out}$ refers to set of links going out of $N$ and $N_{in}$ refers to set of links going into $N$. We create the LP demonstrated in equation \ref{eq:eq4}, which corresponds to a flow network, with the objective function in equation \ref{eq:eq3} to allocate bandwidth upon arrival of a new request.
This LP has a size of $m(Td-t_{now})$ variables. For a network such as Google's G-Scale network \cite{b4} with 19 links, assuming 5-minute timeslots and a deadline of 24 hours, this LP will have 5500 variables. Using a laptop with Intel Core-i5 2.5GHz CPU, such an LP can be solved in less than a second. We can further optimize this LP using the k-shortest paths method: only links on those paths will be considered in routing.
Based on the objective function in equation \ref{eq:eq3}, similar to single link case, the resulting allocation will favor later timeslots. However, this time traffic is initially scheduled as late as possible on whole paths (not each single link). It is straight-forward to show that theorem 1 still holds: assuming no preemption, if a new request arrives it is not possible to move any previously allocated traffic further into the future to open up space before new request's deadline.
\begin{equation}
f(E) \triangleq \sum_{t=t_{now}+1}^{Td} (\sum_{l=L_1}^{l=L_m} E(t,l)) \times (Td - t) \\
\label{eq:eq3}
\end{equation}
\begin{equation}
\begin{aligned}
Obj: ~~ & min\{f(E)\} \\
S.t: ~~ & \sum_{t=t_{now}+1}^{Td} \sum_{l \in A_{out}}^{} E(t,l) = Q \\
& \sum_{t=t_{now}+1}^{Td} \sum_{l \in B_{in}}^{} E(t,l) = Q \\
& \sum_{l \in N_{in}}^{} E(t,l) = \sum_{l \in N_{out}}^{} E(t,l), ~ t_{now} < t \le Td \\
& 0 \le E(t,l) \le C_{t,l}, ~ t_{now} < t \le Td, \forall l~ \in~ L
\label{eq:eq4}
\end{aligned}
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{network}
\caption{Example of RCD allocation for network}
\label{fig:net}
\end{figure}
Fig. \ref{fig:net} shows an example of network case allocation. As can be seen, resources are not allocated as late as possible for $L_2, L_3$, however, the overall allocation cannot be pushed further into the future.
Utilizing unused resources in current timeslot by sending traffic from future timeslots is not as straight-forward anymore. A central network scheduler will have to decide how to do that. The idea is to look at the first traffic scheduled on timeline for all links, and if for some of the links, there is some unused bandwidth in current timeslot, pull some traffic from future timeslots and send it.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{network_rcd_pull_back}
\caption{Multiple solutions while pulling traffic}
\label{fig:net2}
\end{figure}
In this case, since there are multiple links and paths, it may not be possible to propose an optimal solution. In a single link case, all requests will have to use the same resource: bandwidth of one link. In a network case, however, requests may use resources from different links. In addition, we may not be able to pull the first scheduled request on all links and send them in current timeslot. That's because a request that is scheduled closest to current time on one link, may not be the closest to current time on next or previous link of the path.
Fig. \ref{fig:net2} is an example of such cases. Assume $R1$ and $R2$ are two possible requests that may be submitted in $t_{now}$, however, we don't know which one it will be. At the beginning of $t_{now}$, we can either choose to go with solution 1 or 2. We prefer the first solution if $R1$ comes in, because the second solution will reject the request. For the same reason, we prefer the second solution if $R2$ gets submitted. Our lack of knowledge prohibits us from making a definite optimal decision. We plan on studying this subject further in our future studies.
\section{Conclusions and Future Work}
In this paper, we proposed RCD, a technique that makes fast scheduling of bandwidth resources possible for large transfers with specific deadlines, provides deadline guarantees, and allows for maximum bandwidth utilization. Our simulations demonstrate that for high arrival rates, RCD is up to 15 times faster than Amoeba, a system with similar objectives. RCD makes it possible to decide on request admission knowing only the residual bandwidth. We also proposed a method to reserve bandwidth for important requests with deadlines whose content is not available for delivery at current time. Finally, we proposed an idea on how RCD can be further extended and applied to a network connecting multiple datacenters. Further study of RCD considering soft deadlines, variable highpri traffic rate, application and simulation of RCD for a network case, and comparison of RCD for a network case with other schemes are identified as future work.
|
\paragraph{\paragraph}
\setlength\overfullrule{5pt}
\usepackage{graphicx}
\usepackage[bookmarks=true,hyperfigures=true]{hyperref}
\usepackage[usenames, dvipsnames]{color}
\makeatletter
\let\o@a@f\@author@finish
\def\@author@finish{\o@a@f%
\let\@authors\empty\def\AF@opr##1{}\def\CO@opr##1##2##3{}%
\def\AU@opr##1##2##3{\ifx\@authors\@empty\toks@\expandafter{##2}%
\else\toks@\expandafter{\@authors, ##2}\fi\edef\@authors{\the\toks@}}
\@AAC@list}
\makeatother
\fi
\expandafter\def\expandafter\bfseries\expandafter{\bfseries\ifmmode\else\boldmath\fi}
\expandafter\def\expandafter\mdseries\expandafter{\mdseries\ifmmode\else\unboldmath\fi}
\expandafter\def\expandafter\normalfont\expandafter{\normalfont\ifmmode\else\unboldmath\fi}
\usepackage{fixmath}
\RequirePackage[extdef]{delimset}
\let\barefrac=\frac
\renewcommand{\frac}[2]{\mathinner{\barefrac{#1}{#2}}}
\let\baresqrt=\sqrt
\makeatletter
\renewcommand{\sqrt}{\@ifnextchar[\@sqrt@space@a\@sqrt@space@b}
\def\@sqrt@space@a[#1]#2{\mathinner{\mathchoice{\mkern-3mu}{\mkern-3mu}{}{}\baresqrt[#1]{#2}}}
\def\@sqrt@space@b#1{\mathinner{\mathchoice{\mkern-3mu}{\mkern-3mu}{}{}\baresqrt{#1}}}
\makeatother
\makeatlette
\let\per@dot@old=\.
\def\.{\ifmmode\def\per@dot@sel{\mkern3mu}\else\def\per@dot@sel{\per@dot@old}\fi\per@dot@sel}
\makeatother
\providecommand{\genfrac}[6]{\left#1{{#5}\above#3{#6}}\right#2}
\newcommand{\genfrac{}{}{0pt}{}}{\genfrac{}{}{0pt}{}}
\newcommand{\sfrac}[2]{{\textstyle\mathord{\frac{#1}{#2}}}}
\newcommand{\sfrac{1}{2}}{\sfrac{1}{2}}
\newcommand{\sfrac{i}{2}}{\sfrac{i}{2}}
\newcommand{\sfrac{1}{4}}{\sfrac{1}{4}}
\newcommand{\vfrac}[2]{\ifmmode\mathinner{\textstyle^{#1}\!/\!_{#2}}\else$^{#1}\!/\!_{#2}$\fi}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{C}}{\mathbb{C}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{\mathbb{V}}{\mathbb{V}}
\newcommand{\alg}[1]{\mathfrak{#1}}
\newcommand{\mathrm{Y}}{\mathrm{Y}}
\newcommand{\grp}[1]{\mathrm{#1}}
\newcommand{\gener}[1]{\mathbb{#1}}
\newcommand{\gener}{\gener}
\newcommand{\genY}[1]{\widehat{\gener{#1}}{}}
\makeatletter\let\Re\@undefined\let\Im\@undefined\makeatother
\DeclareMathOperator{\Re}{Re}
\DeclareMathOperator{\Im}{Im}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\diag}{diag}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{{\mathsf{T}}}{{\mathsf{T}}}
\newcommand{\nonumber}{\nonumber}
\newcommand{\nonumber\\}{\nonumber\\}
\newcommand{\alignbreak}[1][\nonumber]{#1\\&\mathrel{\phantom{=}}\mathord{}}
\def\[{\begin{equation}}
\def\]{\end{equation}}
\providecommand{\href}[2]{#2}
\newcommand{\arxivlink}[1]{\href{http://arxiv.org/abs/#1}{arxiv:#1}}
\makeatletter
\def\mr@ignsp#1 {\ifx\:#1\@empty\else #1\expandafter\mr@ignsp\fi}%
\newcommand{\multiref}[1]{\begingrou
\xdef\mr@no@sparg{\expandafter\mr@ignsp#1 \: }%
\def\mr@comma{}%
\@for\mr@refs:=\mr@no@sparg\do{\mr@comma\def\mr@comma{,}\ref{\mr@refs}}%
\endgroup}
\renewcommand{\eqref}[1]{(\multiref{#1})}
\makeatother
\makeatletter
\newcommand{\namedref}[2]{\hyperref[#2]{#1~\ref*{#2}}}
\newcommand{\@ifstar{\namedref{Section}}{\namedref{Sec.}}}{\@ifstar{\namedref{Section}}{\namedref{Sec.}}}
\newcommand{\@ifstar{\namedref{Appendix}}{\namedref{App.}}}{\@ifstar{\namedref{Appendix}}{\namedref{App.}}}
\newcommand{\@ifstar{\namedref{Table}}{\namedref{Tab.}}}{\@ifstar{\namedref{Table}}{\namedref{Tab.}}}
\newcommand{\@ifstar{\namedref{Figure}}{\namedref{Fig.}}}{\@ifstar{\namedref{Figure}}{\namedref{Fig.}}}
\makeatother
\let\oldbib=\phantomsection\addcontentsline{toc}{section}{\refname}\oldbib
\def\phantomsection\addcontentsline{toc}{section}{\refname}\oldbib{\phantomsection\addcontentsline{toc}{section}{\refname}\oldbib}
\let\oldtoc=\phantomsection\addcontentsline{toc}{section}{\contentsname}\oldtoc
\def\phantomsection\addcontentsline{toc}{section}{\contentsname}\oldtoc{\phantomsection\addcontentsline{toc}{section}{\contentsname}\oldtoc}
\providecommand{\hypersetup}[1]{}
\providecommand{\texorpdfstring}[2]{#1}
\hypersetup{plainpages=false}
\hypersetup{pdfpagemode=UseNone}
\hypersetup{bookmarksnumbered=true}
\hypersetup{pdfstartview=FitH}
\hypersetup{colorlinks=false}
\hypersetup{citebordercolor={.5 1 .5}}
\hypersetup{urlbordercolor={.5 1 1}}
\hypersetup{linkbordercolor={1 .7 .7}}
\makeatletter
\let\@keywords\@empty
\let\@subject\@empty
\providecommand{\keywords}[1]{\gdef\@keywords{#1}}
\providecommand{\subject}[1]{\gdef\@subject{#1}}
\def\@title{\@title}
\def\@author{\@author}
\def\@subject{\@subject}
\def\@date{\@date}
\def\@keywords{\@keywords}
\makeatother
\AtBeginDocument{
\hypersetup{pdftitle={\@title}}%
\hypersetup{pdfauthor={\@author}}%
\hypersetup{pdfsubject={\@subject}}%
\hypersetup{pdfkeywords={\@keywords}}%
}
\ifpublic
\newcommand{\remark}[2][]{\ignorespaces}
\else
\usepackage{ifpdf}
\ifpdf\else\RequirePackage[active]{srcltx}\fi
\RequirePackage{color}
\newcommand{\remark}[2][]{{\normalfont\sffamily\hspace{1ex}\def\tmparga{#1}%
\def\tmpargb{MR}\ifx\tmparga\tmpargb\color{magenta}\fi%
\def\tmpargb{NB}\ifx\tmparga\tmpargb\color{blue}\fi%
\def\tmpargb{AG}\ifx\tmparga\tmpargb\color{Orange}\fi%
\def\tmpargb{Refs}\ifx\tmparga\tmpargb\color{Orchid}\fi%
\def\tmpargb{}\ifx\tmparga\tmpargb\color{red}\fi%
\def\tmpargb{}\ifx\tmparga\tmpargb\else \textbf{#1:} \fi%
#2\hspace{1ex}}}
\fi
\newcommand{\remarkref}[1][]{{\def\tmparga{#1}\def\tmpargb{}%
\ifx\tmparga\tmpargb\remark[Refs]{needed}\else\remark[Refs]{#1}\fi}}
\begin{document}
\title{Untwisting the symmetries of $\beta$-deformed Super-Yang--Mills}
\long\def\theabstract{
We demonstrate that the planar real-$\beta$-deformed Super-Yang--Mills theory possesses an infinitely-dimensional Yangian symmetry algebra and thus is classically integrable. This is achieved by the introduction of the twisted coproduct which allows us to lift the apparent $\mathcal{N}=1$ supersymmetry first to the full $\mathcal{N}=4$ symmetry of the parent $\mathcal{N} = 4$ SYM theory, and subsequently to its Yangian.
}
\ifarxiv\else
\author{Aleksander Garus}
\email{<EMAIL>}
\affiliation{Institut f\"ur Theoretische Physik,
Eidgen\"ossische Technische Hochschule Z\"urich,
Wolfgang-Pauli-Strasse 27, 8093 Z\"urich, Switzerland}
\begin{abstract}
\theabstract
\end{abstract}
\maketitle
\fi
\ifarxiv
\pdfbookmark[1]{Title Page}{title}
\thispagestyle{empty}
\vspace*{2cm}
\begin{center}%
\begingroup\Large\bfseries\@title\par\endgroup
\vspace{1cm}
\begingroup\scshape
Aleksander Garus
\endgroup
\vspace{5mm}
\textit{Institut f\"ur Theoretische Physik,\\
Eidgen\"ossische Technische Hochschule Z\"urich,\\
Wolfgang-Pauli-Strasse 27, 8093 Z\"urich, Switzerland}
\vspace{0.1cm}
\begingroup\ttfamily\small
[email protected]\par
\endgroup
\vspace{5mm}
\vfill
\textbf{Abstract}\vspace{5mm}
\begin{minipage}{12.7cm}
\theabstract
\end{minipage}
\vspace*{4cm}
\end{center}
\newpage
\fi
\paragraph{Introduction}
\label{sec:intro}
The discovery of integrability (see \cite{Beisert:2010jr} for review) of the planar $\mathcal{N} = 4$ Super-Yang--Mills (SYM) theory and its gravity dual, the SUGRA limit of type IIB superstring theory on $AdS_5 \times S^5$, allowed for what is up to date the most convincing tests of the AdS/CFT correspondence \cite{Maldacena:1997re}. Indeed, the huge amount of usually not obvious symmetries in those theories made computations of non-trivial observables on both sides feasible. The integrability techniques enabled to find some observables at any value of coupling, thus being an indisposable tool for testing the weak/strong duality.
After the first convincing results stemming from $\mathcal{N} = 4$ SYM and ABJM theories (and their respective duals), a search for other examples of dualities began, ones that would allow to further verify Maldacena's proposal.
A way of controllably obtaining such theories which proved very successful was to deform the initial ones in a way which preserves conformal invariance. For $\mathcal{N} = 4$ SYM such deformations -- modifications of the superpotential -- were first discussed by Leigh and Strassler in \cite{Leigh:1995ep}. It was early observed that one of the deformations, the real-$\beta$-deformation, preserves integrability of $\mathcal{N} = 4$ SYM (see \cite{Zoubos:2010kh} and references therein) in the planar limit. Its gravity dual was found by Lunin and Maldacena \cite{Lunin:2005jy}, and later demonstrated to stem from TsT transformations of the original $AdS_5 \times S^5$ background \cite{Frolov:2005ty}, \cite{Frolov:2005dj}.
The real-$\beta$-deformed theory is manifestly an $\mathcal{N} =1$ SYM. In \cite{Mansson:2008xv}, \cite{Dlamini:2016aaa} it was demonstrated however, that the manifest $SU(3) \times U(1)$ R-symmetry of $\mathcal{N} = 4$ SYM expressed in the $\mathcal{N} = 1$ language (see Section \ref{sec:betadef} ) does survive the deformation, albeit in a twisted way. Thus the $\mathcal{N} = 4$ supersymmetry is not necessarily broken, but rather hidden. This result is backed up by the study of amplitudes in \cite{Khoze:2005nd}. The conclusion there was that the amplitudes in the twisted theory can be easily obtained from the ones of $\mathcal{N} = 4$ SYM by a procedure that depends only on the \emph{external} legs, irrespectively of the internal structure, even though the vertices get deformed too.
The superconformal symmetry by itself is not sufficient to account for integrability of a field theory. For the parent $\mathcal{N} = 4$ SYM the correct infinite-dimensional algebra has been identified as the Yangian $Y(\mathfrak{psu}(2,2|4))$ \cite{Dolan:2003uh}. It is natural to expect that the integrability of the real-$\beta$-deformed theory will be explained by a suitable deformation of this Yangian algebra. It was indeed shown in \cite{Ihry:2008gm} that some closed subsectors of it do enjoy a twisted Yangian symmetry, resonating well with the results of \cite{Beisert:2005if} on the twisted R-matrix.
In our recent work \cite{Beisert:2017pnr} we presented a novel framework which allows to establish a non-local symmetry of a given theory and used it to show the Yangian invariance of planar $\mathcal{N} = 4$ SYM. In this work we will push our formalism further and demonstrate the extended symmetries of the real-$\beta$-deformed SYM, where nonlocality appears already at the level of R-symmetry.
The article is organized as follows. In section \ref{sec:betadef} we introduce the real-$\beta$-deformed SYM theory. That is followed by the construction of the twisted coproduct for R-symmetry generators so that we obtain the 12 missing supercharges of $\mathcal{N} = 4$ SYM in section \ref{sec:untwist}. In section \ref{sec:yangian} we give a brief introduction to Yangian algebras, construct the twisted level-1 momentum generator $\widehat{P}_{\alpha \dot\alpha}$ and show that it is a symmetry of the theory. We comment on the results in section \ref{sec:conclusions}. The appendices contain all the formulae necessary to reproduce the results.
\paragraph{$\beta$-deformed $\mathcal{N}=4$ SYM}
\label{sec:betadef}
The action of the real-$\beta$-deformed $\mathcal{N} = 4$ SYM \cite{Leigh:1995ep} is most conveniently expressend in the $\mathcal{N} = 1$ language, the field content being three chiral and one vector superfield. Working with component fields, $\Phi_i$ are the three complex scalar fields, $\lambda_{i \alpha}$ their superpartners ($i=1,\ 2,\ 3$). The gauge field $A_{\alpha \dot{\alpha}}$ has the gluino $\lambda_{4 \alpha}$ as its superpartner and acts as a connection for the covariant derivative $D_{\alpha \dot{\alpha}}=\partial_{\alpha \dot{\alpha}}+i A_{\alpha \dot{\alpha}}$. All the fields are in the adjoint of the gauge group $U(N)$.
Written out explicitly, the Langrangian of the theory takes the form:
\begin{align}
\mathcal{L}=\tr \Bigg(& -\frac{1}{8}\epsilon^{\alpha \beta}\epsilon^{\gamma \kappa} \epsilon^{\dot{\alpha} \dot{\beta}} \epsilon^{\dot{\gamma}\dot{\kappa}} [D_{\alpha \dot{\beta}},D_{\gamma \dot{\kappa}}][D_{\beta \dot{\alpha}}, D_{\kappa \dot{\gamma}}]-\frac{1}{2}\epsilon^{\alpha \beta}\epsilon^{\dot{\alpha}\dot{\beta}}[D_{\alpha \dot{\alpha}},\bar{\Phi}^i][D_{\beta \dot{\beta}},\Phi_i]+ \nonumber \\
&-\frac{1}{2}[\Phi_i, \Phi_j]_{\beta_{ij}}[\bar{\Phi}^i,\bar{\Phi}^j]_{\beta_{ij}}+\frac{1}{4}[\Phi_i,\bar{\Phi}^i][\Phi_j,\bar{\Phi}^j]+\epsilon^{\alpha \beta} \epsilon^{\dot{\alpha} \dot{\beta}}\bar{\lambda}^4_{\dot{\beta}} [D_{\beta \dot{\alpha}},\lambda_{4 \alpha}]+\nonumber \\
&+ \epsilon^{\alpha \beta} \epsilon^{\dot{\alpha} \dot{\beta}}\bar{\lambda}^i_{\dot{\beta}} [D_{\beta \dot{\alpha}},\lambda_{i \alpha}]+i\left(\epsilon^{\alpha \beta}[\lambda_{4 \alpha},\lambda_{i \beta}]\bar{\Phi}^i+\epsilon^{\dot{\alpha} \dot{\beta}}[\bar{\lambda}^4_{\dot{\alpha}},\bar{\lambda}^i_{\dot{\beta}}]\Phi_i \right) + \nonumber \\
&+\frac{i}{2} \left(\epsilon^{ijk}\epsilon^{\alpha \beta} [\lambda_{i \alpha},\lambda_{j \beta}]_{\beta_{ij}}\Phi_k+ \epsilon_{ijk}\epsilon^{\dot{\alpha} \dot{\beta}}[\bar{\lambda}^i_{\dot{\alpha}},\bar{\lambda}^j_{\dot{\beta}}]_{\beta_{ij}}\bar{\Phi}^k \right)\Bigg) .
\label{eq:n4sym_act_beta}
\end{align}
In the above we have introduced a $\beta$-deformed graded commutator:
\begin{equation}
[f_i, g_j]_{\beta_{ij}}=e^{i \pi \beta_{ij}}f_i g_j - (-1)^{|f_i||g_j|}e^{i \pi \beta_{ji}} g_j f_i ,
\end{equation}
where $\beta_{ij}=-\beta_{ji}$, $\beta_{12}=\beta_{23}=\beta_{31}=\beta \in \mathbb{R}$.
The undotted and dotted spinor indices $\alpha$, $\dot{\alpha}$ take values $1,\ 2$. One can reintroduce the more common vector indices by contracting the spinor ones with Pauli matrices: $x_\mu = \frac{1}{ \sqrt{2} }\sigma_\mu^{\alpha \dot{\alpha}} x_{\alpha \dot{\alpha}}$.
Taking $\beta=0$ we recover the maximally symmetric $\mathcal{N}=4$ SYM theory. In the present form however, only $SU(3) \times U(1)$ subgroup of the $SU(4)$ R-symmetry is manifestly present, with the $U(1)$ factor corresponding to the vector superfield and the $SU(3)$ to 3 chirals.
For an aribtrary real $\beta$ the action \eqref{eq:n4sym_act_beta} is invariant under $\mathcal{N} = 1$ supersymmetry with the charges $Q_{\alpha}$, $\bar{Q}_{\dot{\alpha}}$, which in the undeformed theory correspond to $Q^4_{\alpha}$, $\bar{Q}_{4 \dot{\alpha}}$ (see Appendix B).
As was shown in \cite{Leigh:1995ep}, the theory remains conformal even on a quantum level -- a property shared with the full $\mathcal{N} =4$ SYM -- and thus yields another example of a superconformal quantum field theory, as already alluded to in the Introduction.
\paragraph{Untwisting the twist}
\label{sec:untwist}
As already discussed in the Introduction, the $\beta$-deformed theory shares a lot of similarities with its parent, the $\mathcal{N} = 4$ SYM. In this section we want to show that the deformed theory in fact possesses the full (albeit deformed and non-local) $\mathcal{N} = 4$ supersymmetry.
To this end it is enough to demonstrate that the theory is invariant under the full $SU(4)$ R-symmetry group. All the remaining supersymmetry generators can be then recovered from the $Q_\alpha=Q^4_{\alpha}$ and $\bar{Q}_{\dot\alpha}=\bar{Q}_{4 \dot\alpha}$ via:
\[
Q^i_{\alpha}=[R^i {}_4, Q_\alpha]
\label{eq:miss_sca}
\]
and similarly for the conjugate ones.
Corresponding to the $\mathcal{N}=1$ supersymmetry, the action is invariant under the $R^4 {}_4$ component of the R-symmetry (in this case the $U(1)$ generator). In the case of a planar limit of the real-$\beta$ deformation however, the symmetry is actually larger, as all the diagonal elements $R^c {}_c$ survive. The additional $U(1)$ symmetries enabled the construction of the background of the gravity dual \cite{Lunin:2005jy} (see also \cite{Mansson:2008xv} and \cite{Dlamini:2016aaa} for a discussion of extended symmetry of the theory).
We will now show that all the remaining generators $R^a {}_b$ (see Appendix A for the detailed action) can be promoted to the symmetries of the action \eqref{eq:n4sym_act_beta}.
This will be achieved by a suitable modification of the coproduct of the generators, i.e. the action on products of fields.
In the real-$\beta$-deformed SYM this is achieved by the Drinfeld-Reshetikhin twist of the comultiplication \cite{vanTongeren:2015soa}:
\begin{equation}
\Delta_\mathcal{F} = \mathcal{F} \Delta \mathcal{F}^{-1},
\end{equation}
\noindent with $\mathcal{F}$ given in terms of $SU(4)$ charges of the fields:
\begin{equation}
\mathcal{F}=e^{i \pi \beta (h_1 \wedge h_2 + h_2 \wedge h_3 - h_1 \wedge h_3)},
\end{equation}
where $h_i = R^i {}_{i}+R^4 {}_4$ (no sum over $i$).
We will omit the subscript $\mathcal{F}$ from now on.
As alluded to above, for the diagonal R-symmetry generators the modification acts as identity:
\begin{equation}
\Delta R^c {}_c = \mathbb{I} \otimes R^c {}_c + R^c {}_c \otimes \mathbb{I} .
\label{eq:nondef_copr}
\end{equation}
\noindent The coproduct of non-diagonal elements however changes to
\begin{equation}
\Delta R^a {}_b = \mathbb{K}_{ab} \otimes R^a {}_b + R^a {}_b \otimes \mathbb{K}_{ba}
\label{eq:def_copr}
\end{equation}
where $\mathbb{K}_{ba}=\mathbb{K}_{ab}^{-1}$. The element $\mathbb{K}_{ab}$ is group-like:
\begin{equation}
\Delta \mathbb{K}_{ab} = \mathbb{K}_{ab} \otimes \mathbb{K}_{ab}.
\label{eq:kr_cop}
\end{equation}
\noindent Its action on the fields of the theory amounts to a multiplication by a $\beta$-dependent phase (see Appendix A).
Observe that the coproduct \eqref{eq:kr_cop} does not cope well with cyclicity. It is a usual problem with non-trivial coproducts, since due to appearance of trace in \eqref{eq:n4sym_act_beta} (and any other Lagrangian) the action of a theory is cyclic. To circumvent this problem, in \cite{Beisert:2017pnr} we developed an equation-of-motion-based formalism which is equivalent to invariance of action. We will sketch it here. For a usual Lie-type symmetry with a trivial coproduct which leaves the action invariant:
\[
J S := (J Z_a) \frac{\delta S}{\delta Z_a} = 0,
\label{eq:jseq0}
\]
where $J$ is an algebra generator, we can differentiate \eqref{eq:jseq0} with respect to an arbitrary field $Z_c$ to obtain:
\[
J \frac{\delta S}{\delta Z_c} = - \frac{\delta (J Z_a)}{\delta Z_c} \frac{\delta S}{\delta Z_a}.
\label{eq:magic_lvl0}
\]
Observe that the equality \eqref{eq:magic_lvl0}, contrary to \eqref{eq:jseq0}, contains no cyclic objects. What \eqref{eq:magic_lvl0} states is that the result of $J$ acting on an equation of motion $\frac{\delta S}{\delta Z_c}$ is a particular combination of other equations of motion with coefficients determined purely by a representation of $J$ on fields $Z_a$ of the theory. We argued in \cite{Beisert:2017pnr} that \eqref{eq:magic_lvl0} is equivalent to the invariance of the action.
The question now is what is the equivalent of \eqref{eq:magic_lvl0} in case of a twisted coproduct \eqref{eq:kr_cop}.
Surprisingly, the answer is that \eqref{eq:magic_lvl0} holds in this case in an unchanged form, with no explicit appearance of generators $\mathbb{K}_{ab}$ (even though they do contribute while acting on an equation of motion) - a fact we confirmed by direct computation.
Indeed, it can actually be shown that their action amounts to an overall factor while acting on equations of motion and hence becomes unobservable. Heuristically this can be explained by noting that the kinetic term is always undeformed. All the transformations are thus fully determined by the single-field action of $R^a {}_b$. Having shown that all the R-symmetry generators are symmetries of the action, we now may use them to construct the missing supercharges according to \eqref{eq:miss_sca} and thus argue that indeed the real-$\beta$-deformed SYM possesses a $\mathcal{N} = 4$ supersymmetry. The formula \eqref{eq:magic_lvl0} holds then also for the generators $Q^a_{\alpha}$ and $\bar{Q}_{a \dot\alpha}$, as we verified again by explicit calculation.
Concluding this paragraph, let us mention that it may actually be directly shown that the action obtained by integrating the Lagrangian \eqref{eq:n4sym_act_beta} is invariant under all R-symmetry generators $R^a {}_b$, without resorting to the equations of motion. To this end, we cannot rely on picking an arbitrary cyclic representative, but rather need to consider the action as the averaged sum of all the possible ones . This of course is a trivial operation under the trace, but in order to act with a generator of transformation, we need to \emph{cut the trace open} at some point. A trivial coproduct respects cyclicity of the trace and hence the choice of the opening point is irrelevant, but that ceases to be true for coproducts like \eqref{eq:def_copr}. We hence rewrite the terms forming the action as:
\[
\text{tr} \left( Z_1 Z_2 ... Z_n \right) \rightarrow \frac{1}{n} \sum_{\sigma \in \mathbb{Z}_n} Z_{\sigma(1)} Z_{\sigma(2)} ... Z_{\sigma(n)},
\]
and then act on them using iterated coproducts. A heuristic picture to have in mind is cutting open the closed (due to trace) chain of fields at every possible point. Under so defined application of the generators, the action is invariant.
As this approach however does not easily generalise to higher level Yangian symmetries discussed in the next section, we will not discuss it here in more detail.
\paragraph{Twisted Yangian}
\label{sec:yangian}
Having established a twisted $\mathcal{N} = 4$ superconformal symmetry of the real-$\beta$-deformed theory, we are now in position to observe what happens to the Yangian symmetry. Indeed, for the $\mathcal{N} = 4$ SYM the Yangian of $\mathfrak{psu} (2,2|4)$ underlies its integrability. Twisted Yangian algebra has been identified in \cite{Ihry:2008gm} as the symmetry of some subsectors of the real-$\beta$-deformed SYM. We would now like to promote this discussion to a full theory, taking into account the nonlinearities of the symmetries.
The Yangian algebra is built on a Lie (super)-algebra as follows \cite{Drinfeld:1985rx}, \cite{Loebbert:2016cdm}. Starting from the usual commutation relations:
\[
[J^A,J^B \} = f^{AB} {}_C J^C,
\]
we introduce a new set of \emph{level-1} generators $\widehat{J}^A$ transforming in the adjoint of the original algebra:
\[
[J^A, \widehat{J}^B \} = f^{AB} {}_C \widehat{J}^C.
\]
An infinite tower of higher-level generators is obtained by the commutation of the lower level ones, i.e. level-2 generators can be obtained by commuting two level-1 ones. The generators are subject to Serre relations, which we will not discuss in the current work.
The crucial feature of the level-1 generators is their non-trivial coproduct:
\[
\Delta \widehat{J}^A = \mathbb{I} \otimes \widehat{J}^A + \widehat{J}^A \otimes \mathbb{I} + f^A {}_{BC} J^B \otimes J^C,
\]
which features two level-0 generators acting simultaneously. This coproduct renders the application of level-1 Yangian generators directly to the action cumbersome, which is why we work with the equations of motion of the theory (see \cite{Beisert:2017pnr} for details).
In the case of $\mathcal{N}=4$ SYM, the simplest level-1 Yangian generator is the level-1 momentum $\widehat{P}_{\alpha \dot{\alpha}}$, whose coproduct is given by:
\begin{align}
\Delta \widehat{P}_{\alpha \dot{\alpha}, \ \mathcal{N}=4} &= \mathbb{I}\otimes \widehat{P}_{\alpha \dot\alpha} + \widehat{P}_{\alpha \dot\alpha} \otimes \mathbb{I} + D \wedge P_{\alpha \dot\alpha} + P_{\beta \dot\alpha} \wedge L^\beta {}_\alpha + P_{\alpha \dot\beta} \wedge \bar{L}^{\dot\beta} {}_{\dot\alpha} \nonumber \\
& -\sfrac{1}{2} Q_{\alpha} \wedge \bar{Q}_{\dot\alpha} -\sfrac{1}{2} Q^i_{\alpha} \wedge \bar{Q}_{i \dot\alpha} \nonumber \\
&=\mathbb{I}\otimes \widehat{P}_{\alpha \dot\alpha} + \widehat{P}_{\alpha \dot\alpha} \otimes \mathbb{I} + h^A {}_{BC} J^B \otimes J^C,
\end{align}
\noindent where $h^A {}_{BC}$ are the $\mathfrak{psu}(2,2|4)$ structure constants.
Since the generators of conformal algebra and $Q_{\alpha}$ and $\bar{Q}_{\dot{\alpha}}$ are insensitive to the $\beta$-deformation, the only possible twists in the coproduct for $\widehat{P}_{\alpha \dot{\alpha}}$ can appear in the terms containing $Q^i_{\alpha}$ and $\bar{Q}_{i \dot{\alpha}}$. Indeed, following the methods developed in \cite{Beisert:2017pnr} we checked via explicit computations that the modified coproduct:
\begin{align}
\Delta \widehat{P}_{\alpha \dot{\alpha}, \ \beta \neq 0} &= \mathbb{I}\otimes \widehat{P}_{\alpha \dot\alpha} + \widehat{P}_{\alpha \dot\alpha} \otimes \mathbb{I} + D \wedge P_{\alpha \dot\alpha} + P_{\beta \dot\alpha} \wedge L^\beta {}_\alpha + P_{\alpha \dot\beta} \wedge \bar{L}^{\dot\beta} {}_{\dot\alpha} \nonumber \\
& -\sfrac{1}{2} Q_{\alpha} \wedge \bar{Q}_{\dot\alpha} -\sfrac{1}{2} Q^i_{\alpha} \mathbb{K}_{i \alpha}^{-1} \otimes \bar{Q}_{i \dot\alpha} \bar{\mathbb{K}}_{i \dot\alpha} - \sfrac{1}{2} \bar{Q}_{i \dot\alpha}\bar{\mathbb{K}}_{i \dot\alpha}^{-1} \otimes Q^i_{\alpha} \mathbb{K}_{i \alpha}
\label{eq:phat_def_copr}
\end{align}
\noindent maps equations of motion of the theory to each other, provided the single-field action of $\widehat{P}_{\alpha \dot\alpha}$ is given by eq. \eqref{eq:phat_single}
\begin{align}
&\widehat{P}_{\alpha \dot\alpha} \Phi_i = 0 \nonumber \\
&\widehat{P}_{\alpha \dot\alpha} \bar{\Phi}^i = 0 \nonumber \\
&\widehat{P}_{\alpha \dot\alpha} \lambda_{i \beta} = i \epsilon_{\alpha \beta} \epsilon_{ijk} \{\bar{\lambda}^j_{\dot\alpha}, \bar{\Phi}^k \}_{\beta_{jk}} -i \epsilon_{\alpha \beta} \{ \Phi_i, \bar{\lambda}_{4 \dot\alpha} \} \nonumber \\
&\widehat{P}_{\alpha \dot\alpha} \bar{\lambda}^i_{\dot\beta} = i \epsilon_{\dot\alpha \dot\beta} \epsilon^{ijk} \{\lambda_{j \alpha}, \Phi_k \}_{\beta_{jk}} -i \epsilon_{\dot\alpha \dot\beta} \{ \bar{\Phi}^i, \lambda_{4 \alpha} \} \nonumber \\
&\widehat{P}_{\alpha \dot\alpha} \lambda_{4 \beta} = i \epsilon_{\alpha \beta} \{\bar{\lambda}^i_{\dot\alpha}, \Phi_i \} \nonumber \\
&\widehat{P}_{\alpha \dot\alpha} \bar{\lambda}_{4 \dot\beta} = i \epsilon_{\dot\alpha \dot\beta} \{\lambda_{i \alpha}, \bar{\Phi}^i \} \nonumber \\
&\widehat{P}_{\alpha \dot\alpha} D_{\beta \dot\beta} = -\epsilon_{\alpha \beta} \epsilon_{\dot\alpha \dot\beta} \{\Phi_i, \bar{\Phi}^i \}.
\label{eq:phat_single}
\end{align}
Observe that the modified part is no longer antisymmetric. Furthermore, due to the group-like nature of twist generators $\mathbb{K}$ the coproduct is no-longer \emph{bilocal}, but actually may encompass all the fields from the product. This is in contrast with the undeformed theory, but already made its appearance at level-0.
In the case of undeformed Yangian algebras we showed that the following formula, an level-1 analogue of \eqref{eq:magic_lvl0}, is equivalent to the invariance of the action and utilises only equations of motion, thus circumventing the issue of the cyclicity:
\[
\widehat{J}^A \frac{\delta S}{\delta Z_m} = -\frac{\widehat{J}^A Z_n}{\delta Z_m} \frac{\delta S}{\delta Z_n} + h^A {}_{BC} \frac{\delta S}{\delta Z_n } \left(J^B \wedge \frac{\delta}{\delta Z_m} \right) (J^C Z_n).
\]
In case the of real-$\beta$-deformed theory, this equality gets a predictable deformation for the $\widehat{P}_{\alpha \dot\alpha}$ generator:
\begin{align}
&\widehat{P}_{\alpha \dot\alpha} \frac{\delta S}{\delta Z_m} = -\frac{\widehat{P}_{\alpha \dot\alpha} Z_n}{\delta Z_m} \frac{\delta S}{\delta Z_n} + f^{P_{\alpha \dot\alpha}} {}_{BC} \frac{\delta S}{\delta Z_n } \left(J^B \wedge \frac{\delta}{\delta Z_m} \right) (J^C Z_n)+ \nonumber \\
&\frac{\delta S}{\delta Z_n} \left(Q^i_{\alpha} \mathbb{K}_{i \alpha}^{-1} \wedge \frac{\delta }{\delta Z_m} \right) (\bar{Q}_{i \dot\alpha} \bar{\mathbb{K}}_{i \dot\alpha} \ Z_n) + \frac{\delta S}{\delta Z_n} \left(\bar{Q}_{i \dot\alpha}\bar{\mathbb{K}}_{i \dot\alpha}^{-1} \wedge \frac{\delta }{\delta Z_m}\right)(Q^i_{\alpha} \mathbb{K}_{i \alpha} \ Z_n) ,
\label{eq:magic_def}
\end{align}
\noindent where $f^A {}_{BC}$ are $\mathfrak{psu}(2,2|1)$ structure constants. Indeed, we checked that the equations of motion of the real-$\beta$-deformed SYM satisfy \eqref{eq:magic_def}.
Eventually, we see that the properly defined level-1 Yangian generator $\widehat{P}_{\alpha \dot\alpha}$ is a symmetry of real-$\beta$-deformed $\mathcal{N} =4$ SYM. We thus obtain an infinitely-dimensional symmetry algebra also for the case of the twisted theory, which explains its observed integrability. On the level of algebra, the symmetry is just the Yangian $Y\left(\mathfrak{psu}(2,2|4) \right)$, the difference with the $\mathcal{N} = 4$ SYM being only the Hopf structure.
\paragraph{Comments and conclusions}
\label{sec:conclusions}
In this work we demonstrated that the planar real-$\beta$-deformed SYM theory \eqref{eq:n4sym_act_beta} possesses a much richer set of symmetries than manifest on the level of the action. By the use of a twisted coproduct we applied all the $\mathcal{N} = 4$ SUSY generators and verified that they leave the action invariant and thus the planar real-$\beta$-deformed SYM shares the symmetry algebra $\mathfrak{psu}(2,2|4)$ with the $\mathcal{N} = 4$ SYM. We then showed that those results can be lifted to the full Yangian $Y\left(\mathfrak{psu}(2,2|4) \right)$ symmetry of the theory.
Our result not only generalizes earlier constructions of Yangian algebra in the case of a deformed theory \cite{Ihry:2008gm}, but moreover verifies that the formalism we introduced in \cite{Beisert:2017pnr} can also be used to study deformed theories, with the main result, \eqref{eq:magic_def} being in a clear correspondence with the coproduct \eqref{eq:phat_def_copr} of the generator under consideration.
Our formalism can thus serve as a classical integrability check for a wide class of theories.
The above results are clearly purely classical. Though very close on the classical level, the undeformed and deformed theory differ when loop effects are considered (i.e. wrapping corrections appear already at 1-loop for the deformed theory, whereas they are postponed to higher loops for the $\mathcal{N} = 4$ SYM). It would be interesting to see, what is the fate of Yangian symmetry once quantum corrections are introduced and whether it differs from the behavior of the parent theory. Even though the $\mathcal{N} = 1$ superconformal symmetry persists, it is not known what happens to twisted generators, and thus also to higher level Yangian ones.
\pdfbookmark[1]{Acknowledgements}{ack}
\paragraph*{Acknowledgements}
I would like to thank N. Beisert, R. Hecht, B. Hoare, M. Rosso and M. Sprenger for valuable discussions and comments.
The work is partially supported by grant no.\ 615203
from the European Research Council under the FP7
and by the Swiss National Science Foundation
through the NCCR SwissMAP.
\ifarxiv
\bibliographystyle{nb}
\fi
|
\section{Introduction}
When dealing with numerical optimization problems in engineering applications, one is often faced with the optimization of an expensive process that depends on a number of tuning parameters. Examples include the outcome of a biological experiment, the training of large scale machine learning algorithms or the outcome of exploratory drilling. We are concerned with problem instances wherein there is the capacity to perform $k$ experiments in \emph{parallel} and, if needed, repeatedly select further batches with cardinality $k$ as part of some sequential decision making process. Given the cost of the process, we wish to select the parameters at each stage of evaluations carefully so as to optimize the process using as few experimental evaluations as possible.
It is common to assume a surrogate model for the outcome $f: \mathbb{R}^n \mapsto \mathbb{R}$ of the process to be optimized. This model, which is based on both prior assumptions and past function evaluations, is used to determine a collection of $k$ input points for the next set of evaluations. Bayesian Optimization provides an elegant surrogate model approach and has been shown to outperform other state-of-the-art algorithms on a number of challenging benchmark functions \citep{Jones2001}. It models the unknown function $f$ with a Gaussian Process (GP) \citep{Rasmussen2005}, a probabilistic function approximator which can incorporate prior knowledge such as smoothness, trends, etc.
A comprehensive introduction to GPs can be found in \cite{Rasmussen2005}. In short, modeling a function with a GP amounts to modeling the function as a realization of a stochastic process. In particular, we assume that the outcomes of function evaluations are normally distributed random variables with known \emph{prior} mean function $m : \mathbb{R}^n \mapsto \mathbb{R}$ and \emph{prior} covariance function $\kappa : \mathbb{R}^n \times \mathbb{R}^{n} \mapsto \mathbb{R}$. Prior knowledge about $f$, such as smoothness and trends, can be incorporated through judicious choice of the covariance function $\kappa$, while the mean function $m$ is commonly assumed to be zero.
A training dataset $\mathcal{D} = (X^d, y^d)$ of $\ell$ past function evaluations $y^d_i = f(X^d_i)$ for $i = 1\dots \ell$, with $y^d \in \mathbb{R}^\ell, X^d \in \mathbb{R}^{\ell \times n}$ is then used to calculate the \emph{posterior} of $f$.
The celebrated GP regression equations \citep{Rasmussen2005} give the posterior
\begin{equation}
y | \mathcal{D} \sim \mathcal{N}(\mu(X),
\Sigma(X))
\label{eqn:GPPrediction}
\end{equation}
on a batch of $k$ test locations $X \in \mathbb{R}^{k \times n}$ as a multi-variate normal distribution in a closed form formula. The posterior mean value $\mu$ and variance $\Sigma$ depend also on the dataset $\mathcal{D}$, but we do not explicitly indicate this dependence in order to simplify the notation. Likewise, the posterior $y | \mathcal{D}$ is a normally distributed random variable whose mean $\mu(X)$ and covariance $\Sigma(X)$ depend on $X$, but we do not make this explicit.
Given the surrogate model, we wish to identify some selection criterion for choosing the next batch of points to be evaluated. Such a selection criterion is known as an \emph{acquisition function} in the terminology of Bayesian Optimization. Ideally, such an acquisition function would take into account the number of remaining evaluations that we can afford, e.g.\ by computing a solution via dynamic programming to construct an optimal sequence of policies for future batch selections. However, a probabilistic treatment of such a criterion is computationally intractable, involving multiple nested minimization-marginalization steps \citep{Gonzalez2015}.
To avoid this computational complexity, myopic acquisition functions that only consider the one-step return are typically used instead. For example, one could choose to minimize the one-step \emph{Expected Improvement} (described more fully in \S \ref{sec:expected_improvement}) over the best evaluation observed so far, or maximize the probability of having an improvement in the next batch over the best evaluation. Other criteria use ideas from the bandit \citep{Desautels2014} and information theory \citep{Shah2015} literature. In other words, the intractability of the multistep lookahead problem has spurred instead the introduction of a wide variety of myopic heuristics for batch selection.
\subsection{Expected improvement}
\label{sec:expected_improvement}
We will focus on the (one-step) Expected Improvement criterion, which is a standard choice and has been shown to achieve good results in practice \citep{Snoek2012}. In order to give a formal description we first require some definitions related to the optimization procedure of the original process. At each step of the optimization procedure, define $y^d \in \mathbb{R}^\ell$ as the vector of $\ell$ past function values evaluated at the points $X^d \in \mathbb{R}^{\ell \times n}$, and $X \in \mathbb{R}^{k \times n}$ as a candidate set of $k$ points for the next batch of evaluations. Then the classical expected improvement acquisition function is defined as
\begin{equation}
\begin{gathered}
\alpha(X) = \mathbb{E}[\min(y_{1}, \dots, y_{k}, \underline{y^d}) |
\mathcal{D}] - \underline{y^d} \\
\quad \text{with} \; y | \mathcal{D} \sim \mathcal{N}\bigl(\mu(X),
\Sigma(X)\bigr)
\label{eqn:ExpectedImprovement}
\end{gathered}
\end{equation}
where $\underline{y^d}$ is the element-wise minimum of $y^d$, i.e.\ the minimum value of the function $f$ achieved so far by any known function input. In the above definition we assume perfect knowledge of $\underline{y^d}$, which implies a noiseless objective. A noisy objective requires the introduction of heuristics discussed in detail in \cite{Picheny2013}. For the purposes of clarity, a noiseless objective is assumed for the rest of the document. This is not constraining, as most of the heuristics discussed in \citep{Picheny2013} are compatible with the theoretical analysis presented in the rest of the paper.
Selection of a batch of points to be evaluated with optimal expected improvement amounts to finding some $X \in \arg\min\left[\alpha(X)\right].$ Unfortunately, direct evaluation of the acquisition function $\alpha$ requires the $k$--dimensional integration of a piecewise affine function; this is potentially a computationally expensive operation. This is particularly problematic for gradient-based optimization methods, wherein $\alpha(X)$ may be evaluated many times when searching for a minimizer. Regardless of the optimization method used, such a minimizer must also be computed again for every step in the original optimization process, i.e.\ every time a new batch of points is selected for evaluation. Therefore a tractable acquisition function should be used. In contrast to \eqref{eqn:ExpectedImprovement}, the acquisition function we will introduce in Section \ref{sec:main_result} avoids expensive integrations and can be calculated efficiently with standard software tools.
\subsection{Related work}
Despite the intractability of \eqref{eqn:ExpectedImprovement}, \citet{Chevalier2013} presented an efficient way of approximating $\alpha$ and its derivative $d\alpha/dX$ \citep{Marmin2015} by decomposing it into a sum of $q$--dimensional Gaussian Cumulative Distributions, which can be calculated efficiently using the seminal work of \citet{Genz2009}. There are two issues with this approach: First the number of calls to the $q$--dimensional Gaussian Cumulative Distribution grows quadratically with respect to the batch size $q$, and secondly, there are no guarantees about the accuracy of the approximation or its gradient. Indeed, approximations of the multi-point expected improvement, as calculated with the \texttt{R} package \texttt{DiceOptim} \citep{Roustant2012} can exhibit points of failure in simple examples (see Figure \ref{fig:qei_problem} and Table \ref{table:Timings}). To avoid these issues, \citet{Gonzalez2016} and \citet{Ginsbourger2009} rely on heuristics to derive a multi-point criterion. Both methods choose the batch points in a greedy, sequential way, which restricts them from exploiting the interactions between the batch points in a probabilistic manner. Finally Parallel Predictive Entropy Search (PPES), a non-greedy information theoretic acquisition function was derived by \citet{Shah2015}.
\section{Distributionally ambiguous optimization for Bayesian optimization}
\label{sec:main_result}
We now proceed to the main contribution of the paper. We draw upon ideas from the Distributionally Ambiguous Optimization community to derive a novel, tractable, acquisition function that lower bounds the expectation in~\eqref{eqn:ExpectedImprovement}. Our acquisition function:
\begin{itemize}
\itemsep0.1em
\item is theoretically grounded;
\item is numerically reliable and consistent, unlike Expected Improvement-based alternatives (see \S\ref{sec:empirical_analysis});
\item is fast and scales well with the batch size; and
\item provides first and second order derivative information inexpensively.
\end{itemize}
In particular, we use the GP posterior \eqref{eqn:GPPrediction} derived from the GP to determine the mean $\mu(X)$ and variance $\Sigma(X)$ of $y | \mathcal{D}$ given a candidate batch selection $X$, but we thereafter ignore the Gaussian assumption and consider only that $y | D$ has a distribution embedded within a family of distributions $\mathcal{P}$ that share the mean $\mu(X)$ and covariance $\Sigma(X)$ calculated by the standard GP regression equations. In other words, we define
\[
\mathcal{P}(\mu,\Sigma) = \set{\mathbb P}{\mathbb{E}_\mathbb{P}[\xi] = \mu,
\mathbb{E}_\mathbb{P}[\xi\xi^T] = \Sigma}.
\]
We denote the set $\mathcal{P}(\mu(X),\Sigma(X))$ simply as $\mathcal{P}$ where the context is clear. Note in particular that $\mathcal{N}(\mu,\Sigma) \in \mathcal{P}(\mu,\Sigma)$ for any choice of mean $\mu$ or covariance $\Sigma$.
One can then construct lower and upper bound for the Expected Improvement by minimizing or maximizing over the set $\mathcal{P}$ respectively, i.e.\ by writing
\begin{equation}
\inf_{\mathbb{P} \in \mathcal{P}} \mathbb{E}_{\mathbb{P}}[g(\xi)]
\leq \mathbb{E}_{\mathcal{N}(\mu, \Sigma)}[g(\xi)]
\leq \sup_{\mathbb{P} \in \mathcal{P}} \mathbb{E}_{\mathbb{P}}[g(\xi)]
\label{eqn:ExpectedImprovementBounds}
\end{equation}
where the random vector $\xi \in \mathbb{R}^k$ and the function $g : \mathbb{R}^k \mapsto \mathbb{R}$ are chosen such that $\alpha(X) = \mathbb{E}_{\mathcal{N}(\mu, \Sigma)}[g(\xi)]$, i.e. $\xi = y|\mathcal{D}$ and
\begin{equation}\label{eqn:definition_g}
g(\xi) = \min(\xi_1, \dots, \xi_k, \underline{y^d}) - \underline{y^d}.
\end{equation}
Observe that the middle term in \eqref{eqn:ExpectedImprovementBounds} is equivalent to the expectation in~\eqref{eqn:ExpectedImprovement}.
Perhaps surprisingly, both of the bounds in~\eqref{eqn:ExpectedImprovementBounds} are computationally tractable even though they seemingly require optimization over the infinite-dimensional (but convex) set of distributions $\mathcal{P}$. For either case, these bounds can be computed exactly via transformation of the problem to a tractable, convex semidefinite optimization problem using distributionally ambiguous optimization techniques \citep{Zymler2013}.
We will focus on the lower bound $\inf_{\mathbb{P} \in \mathcal{P}} \mathbb{E}_{\mathbb{P}}[g(\xi)]$ in \eqref{eqn:ExpectedImprovementBounds}, hence adopting an \emph{optimistic} modeling approach. This is because the upper bound is of limited use as it can be shown to be trivial to evaluate and independent of $\Sigma$. Hence, we informally assume that the distribution of function values is such that it yields the lowest possible expectation compatible with the mean and covariance computed by the GP, which we put together in the second order moment matrix $\Omega$ of the posterior as
\begin{equation}
\Omega \eqdef
\begin{bmatrix}
\Sigma + \mu\mu^T & \mu \\
\mu^T & 1
\end{bmatrix}
\label{eqn:Omega}.
\end{equation}
We will occasionally write this explicitly as $\Omega(X)$ to highlight the dependency of the second order moment matrix on $X$.
The following result says that the lower (i.e.\ optimistic) bound in \eqref{eqn:ExpectedImprovementBounds} can be computed via the solution of a convex optimization problem whose objective function is linear in $\Omega$:
\begin{theorem}\label{thm:ValueTheorem}
For any $\Sigma \succ 0$ the optimal value of the semi-infinite optimization problem
\[
\inf_{\mathbb{P} \in \mathcal{P}(\mu, \Sigma)} \mathbb{E}_{\mathbb{P}}[g(\xi)]
\]
coincides with the optimal value of the following semidefinite program:
\begin{equation}
\begin{aligned}
p(\Omega) \eqdef \; &\sup
& & \langle \Omega, M \rangle\\
& \text{s.t.} & & M - C_i \preceq 0, \quad i = 0,\dots,k,
\end{aligned}
\label{eqn:PrimalOptProb}
\tag{$P$}
\end{equation}
where $M \in \mathbb{S}^{k+1}$ is the decision variable, and $C_0 \eqdef 0,$
\begin{equation}
C_i \eqdef
\left[
\begin{array}{cc}
0 & e_i/2 \\
e_i^T/2 & -\underline{y^d}
\end{array}
\right],
\qquad i = 1,\dots,k,
\qquad
\label{eqn:PrimalConstraints}
\end{equation}
are auxiliary matrices defined using
$\underline{y^d}$ and the standard basis vectors $e_i$ in $\mathbb{R}^k$.
\end{theorem}
\begin{proof}
See Appendix \ref{app:PrimalOptProb}.
\end{proof}
Problem \eqref{eqn:PrimalOptProb} is a semidefinite program (SDP). SDPs are convex optimization problems with a linear objective and convex conic constraints (i.e. constraints over the set of symmetric matrices $\mathbb{S}^{k}$, positive semidefinite/definite matrices $\mathbb{S}^{k}_{+}$/$\mathbb{S}^{k}_{++}$) and can be solved to global optimality with standard software tools~\citep{Sturm1999, SCS2016}. We therefore propose the computationally tractable acquisition function
\begin{equation*}
\bar \alpha(X) \eqdef p\bigl(\Omega(X)\bigr) \leq \alpha (X) \quad \forall X
\in
\mathbb{R}^{k\times n},
\end{equation*}
which we will call \emph{Optimistic Expected Improvement (OEI)}, as it is an optimistic variant of the Expected Improvement function in \eqref{eqn:ExpectedImprovement}.
This computational tractability comes at the cost of inexactness in the bounds \eqref{eqn:ExpectedImprovementBounds}, which is a consequence of minimizing over a set of distributions containing the Gaussian distribution as just one of its members. Indeed, it can be proved that the minimizing distribution is discrete with $k+1$ possible outcomes and can be constructed by the Lagrange multipliers of \eqref{eqn:PrimalOptProb} \citep{VanParysThesis2015}. We show in \S\ref{sec:empirical_analysis} that this inexactness is of limited consequence in practice and it mainly renders the acquisition function more explorative. Nevertheless, there remains significant scope for tightening the bounds in \eqref{eqn:ExpectedImprovementBounds} via imposition of additional convex constraints on the set $\mathcal{P}$, e.g.\ by restricting $\mathcal{P}$ to symmetric or unimodal distributions \citep{VanParys2015}. Most of the results in this work would still apply, \emph{mutatis mutandis}, if such structural constraints were to be included.
In contrast to the side-effect of inexactness, the distributional ambiguity is useful for integrating out the uncertainty of the GP's hyperparameters efficiently for our acquisition function. Given $q$ samples of the hyperparameters, resulting in $q$ second order moment matrices $\{\Omega_i\}_{i=1,\dots,q}$, we can estimate the resulting second moment matrix $\tilde \Omega$ of the marginalized, non-Gaussian, posterior as
\begin{equation*}
\tilde \Omega \approx \frac{1}{q}\sum_{i=1}^q \Omega_i.
\end{equation*}
Due to the distributional ambiguity of our approach, both bounds of \eqref{eqn:ExpectedImprovementBounds} can be calculated directly based on $\tilde \Omega$, hence avoiding multiple calls to the acquisition function.
Although the value of $p(\Omega)$ for any fixed $\Omega$ is computable via solution of an SDP, the non-convexity of the GP posterior \eqref{eqn:GPPrediction} that defines the mapping $X \mapsto \Omega(X)$ means that $\bar \alpha (X) = p \bigl(\Omega(X) \bigr)$ is still non-convex in $X$. This is unfortunate, since we ultimately wish to minimize $\bar \alpha(X)$ in order to identify the next batch of points to be evaluated experimentally.
However we can still optimize $\bar \alpha$ locally via non-linear programming. We will establish that a second order method is applicable by showing that $\bar \alpha (X)$ is twice differentiable under mild conditions. Such an approach would also be efficient as the hessian of $\bar \alpha$ can be calculated inexpensively. To show these results we will begin by considering the differentiability of $p$ as a function of $\Omega$.
\begin{theorem}\label{thm:gradientTheorem}
The optimal value function $p : \mathbb{S}^{k+1}_{++} \mapsto \mathbb{R}$ defined in problem \eqref{eqn:PrimalOptProb} is differentiable on
its domain with $\partial p(\Omega)/ \partial \Omega = \bar M(\Omega)$,
where $\bar M(\Omega)$ is the unique optimal solution of
\eqref{eqn:PrimalOptProb} at $\Omega$.
\end{theorem}
\begin{proof}
See Appendix \ref{app:DerivativeCalculations}.
\end{proof}
The preceding result shows that $\partial p(\Omega)/ \partial \Omega$ is produced as a byproduct of evaluation of $\inf_{\mathbb{P} \in \mathcal{P}} \mathbb{E}_{\mathbb{P}}[g(\xi)]$, since it is simply $\bar M(\Omega)$, the unique optimizer of \eqref{eqn:PrimalOptProb}. The simplicity of this result suggests consideration of second derivative information of $p(\Omega)$, i.e.\ derivatives of $\bar M(\Omega)$. The following result proves that this is well defined and tractable for any $\Omega \succ 0$:
\begin{theorem}\label{thm:dotM}
The optimal solution function $\bar M: \mathbb{S}_{++}^{k+1} \mapsto \mathbb{S}^{k+1}$ in problem \eqref{eqn:PrimalOptProb} is differentiable on $\mathbb{S}_{++}^{k+1}$. Every directional derivative of $\bar M (\Omega)$ is the unique solution to a sparse linear system with $\mathcal{O}(k^3)$ nonzeros.
\end{theorem}
\begin{proof}
See Appendix \ref{app:Mdot}.
\end{proof}
We can now consider the differentiability of $\bar \alpha = p\circ\Omega$. The following Corollary establishes this under certain conditions.
\begin{corollary}
$\bar \alpha: \mathbb{R}^{k \times n} \mapsto \mathbb{R}$ is twice differentiable on any $X$ for which $\Sigma(X) \succ 0$ and the mean and kernel functions of the underlying GP are twice differentiable.
\end{corollary}
\begin{proof}
By examining the GP Regression equations \citep{Rasmussen2005} and Equation \eqref{eqn:Omega}, we conclude that $\Omega(X)$ is twice differentiable on $\mathbb{R}^{k\times n}$ if the kernel and mean functions of the underlying Gaussian Process are twice differentiable. Hence, $\bar \alpha (X) = p(\Omega(X))$ is twice differentiable for any $\Omega(X) \succ 0$ as a composition of twice differentiable functions. Examining \eqref{eqn:Omega} reveals that $\Omega(X) \succ 0$ is equivalent to $\Sigma(X) \succ 0$, which concludes the proof.
\end{proof}
A rank deficient $\Sigma(X) \nsucc 0$ implies perfectly correlated outcomes. At these points both OEI and QEI are, in general, non-differentiable as shown in Proposition \ref{prop:nonDifferentiability}. This contradicts \cite{Marmin2015} which claims differentiability of QEI in the general setting. However, this is not constraining in practice as both QEI and OEI can be calculated by considering a smaller, equivalent problem. It is also not an issue for descent based methods for minimizing $\bar \alpha$, as a descent direction can be obtained by an appropriate perturbation of the perfectly correlated points.
We are now in a position to derive expressions for the gradient and the hessian of $\bar \alpha = p\circ \Omega$. For simplicity of notation we consider derivatives over $\bar x = \text{vec}(X)$. Application of the chain rule to $\bar \alpha(\bar x) = p(\Omega(\bar x))$ gives:
\begin{equation}
\frac{\partial \bar \alpha (\bar x)}{\partial \bar x_{(i)}}
=
\innerprod{
\frac{\partial p(\Omega)}
{\partial \Omega}
}{
\frac{\partial \Omega(\bar x )}
{\partial \bar x_{(i)}}
} =
\innerprod{
\bar M(\Omega)
}{
\frac{\partial \Omega(\bar x)}
{\partial \bar x_{(i)}}
}
\label{eqn:gradient}
\end{equation}
Note that the second term in the rightmost inner product above depends on the particular choice of covariance function $\kappa$ and mean function $m$. It is straightforward to compute \eqref{eqn:gradient} in modern graph-based \texttt{autodiff} frameworks, such as the \texttt{TensorFlow}-based \texttt{GPflow}. Differentiating again \eqref{eqn:gradient} gives the hessian of $\bar \alpha$:
\begin{equation}
\frac{\partial^2 \bar \alpha(\bar x)}{\partial \bar x_{(i)} \partial \bar x_{(j)}}
=
\frac{\partial}{\partial x_{(i)}}
\innerprod{
\bar M(\Omega)
}{
\frac{\partial \Omega(\bar x)}
{\partial \bar x_{(j)}}
}
=
\innerprod{
\bar M(\Omega)
}
{
\frac{\partial^2
\Omega(\bar x)}
{\partial \bar x_{(i)} \partial \bar x_{(j)}}
}
+
\innerprod{
\frac{\partial \bar M(\Omega(\bar x))}{\partial \bar x_{(i)}}
}{
\frac{\partial \Omega(\bar x)}{\partial \bar x_{(j)}}
}
\label{eqn:ChainRule}
\end{equation}
where $\partial \bar M / \partial \bar x_{(i)}$ is the directional derivative of $\bar M(\Omega)$ across the perturbation $\partial \Omega(\bar x ) / \partial \bar x_{(i)}$. According to Theorem \ref{thm:dotM}, each of these directional derivatives exists and can be computed via solution of a sparse linear system.
\section{Empirical analysis} \label{sec:empirical_analysis}
\begin{table}[ht]
\caption{List of acquisition functions} \label{sample-table}
\begin{center}
\begin{tabular}{lc}
\toprule
{Key} & {Description} \\
\hline \\
OEI &
Optimistic Expected Improvement
\emph{(Our novel algorithm)} \\
QEI &
Multi-point Expected Improvement
\citep{Marmin2015} \\
QEI-CL &
Constant Liar (\emph{``mix''} strategy)
\citep{Ginsbourger2010} \\
LP-EI &
Local Penalization Expected Improvement
\citep{Gonzalez2016} \\
BLCB &
Batch Lower Confidence Bound
\citep{Desautels2014} \\
\end{tabular}
\end{center}
\label{table:AlgList}
\end{table}
In this section we demonstrate the effectiveness of our acquisition function against a number of state-of-the-art alternatives. The acquisition functions we consider are listed in Table \ref{table:AlgList}. We do not compare against PPES as it is substantially more expensive and elaborate than our approach and there is no publicly available implementation of this method.
We show that our acquisition function OEI achieves better performance than alternatives and highlight simple failure cases exhibited by competing methods. In making the following comparisons, extra care should be taken in the setup used. This is because Bayesian Optimization is a multifaceted procedure that depends on a collection of disparate elements (e.g. kernel/mean function choice, normalization of data, acquisition function, optimization of the acquisition function) each of which can have a considerable effect on the resulting performance \citep{Snoek2012,Shahriari2016}. For this reason we test the different algorithms on a unified testing framework, based on \texttt{GPflow}, available online at \href{https://github.com/oxfordcontrol/Bayesian-Optimization}{https://github.com/oxfordcontrol/Bayesian-Optimization}.
Our acquisition function is evaluated via solution of a semidefinite program, and as such it benefits from the huge advances of the convex optimization field. A variety of standard tools exist for solving such problems, including \texttt{MOSEK} \citep{MosekPy}, \texttt{SCS} \citep{ODonoghue2016} and \texttt{CDCS} \citep{Zheng2017}. We chose the first-order \citep{Boyd2004}, freely-available solver \texttt{SCS}, which scales well with batch size and allows for solver warm-starting between acquisition function evaluations.
Warm starting allows for a significant speedup since the acquisition function is evaluated repeatedly at nearby points by the non-linear solver. This results in solving \eqref{eqn:PrimalOptProb} repeatedly for similar $\Omega$. Warm-starting the SDP solver with the previous solution reduces \texttt{SCS}'s iterations by $77\%$ when performing the experiments of Figure $\ref{fig:SyntheticBatch20}$. Moreover, Theorem \ref{thm:gradientTheorem} provides the means for a first-order warm starting. Indeed, the derivative of the solution across the change of the cost matrix $\Omega$ can be calculated, allowing us to take a step in the direction of the gradient and warm start from that point. This reduces \texttt{SCS}'s iterations by a further $43\%$.
Indicative timing results for the calculation of OEI, QEI and their derivatives are listed in Table \ref{table:Timings}. The dominant operation for calculating OEI and its gradient is solving \eqref{eqn:PrimalOptProb}.
This makes OEI much faster than QEI, which is in line with the complexity of the dominant operation in SDP solvers based on first-order operator splitting methods such as \texttt{SCS} or \texttt{CDCS} which, for our problem, is $\mathcal{O}\bigl(k^4 \bigr)$. Assume that, given the solution of \eqref{eqn:PrimalOptProb}, we want to also calculate the hessian of OEI. This would entail the following two operations:
\begin{description}
\item[Calculating $\partial \bar M / \partial X_{(i, j)}$ given $\partial \Omega(X) / \partial X_{(i, j)}.$] According to Lemma \ref{thm:dotM} this can be obtained as a solution to a sparse linear system. We used \texttt{Intel$^{\tiny{\textregistered}}$ MKL PARDISO} to solve efficiently these linear systems.
\item[Calculate $\partial \Omega(X) / \partial X_{(i, j)}$ and apply chain rules] of \eqref{eqn:ChainRule} to get the hessian of the acquisition function $\bar \alpha = p \circ \Omega$ given the gradient and hessian of $p$. We used \texttt{Tensorflow}'s Automatic Differentiation for this part, without any effort to optimize its performance. Considerable speedups can be brought by e.g. running this part on GPU, or automatically generating low-level code optimized specifically for this operation \citep{Vasilache2018}.
\end{description}
Note that the computational tractability of the hessian is only allowed due to the novelty of Theorem \ref{thm:gradientTheorem} which exploits the ``structure'' of \eqref{eqn:PrimalOptProb}'s optimizer.
\begin{table}[t]
\caption{Average execution time of the acquisition function, its gradient and hessian when running BO in the Eggholder function on an Intel$^{\tiny{\textregistered}}$ E5-2640v3 CPU. For batch size 40, QEI fails, i.e. it always returns 0 without any warning message. For the execution time of the hessian we assume knowledge of the solution of \eqref{eqn:PrimalOptProb}. Its timing is split into two parts as described in the main text. }
\vskip 0.15in
\begin{center}
\begin{tabular}{cc|c|cc}
\toprule
Batch Size & QEI: $\alpha(X), \nabla \alpha(X)$ & OEI: $\bar \alpha(X), \nabla \bar \alpha(X)$ & \multicolumn{2}{c}{$\nabla^2 \bar \alpha(X)$} \\
\midrule
& & Solve \eqref{eqn:PrimalOptProb} & $\partial \bar M$ & \texttt{Tensorflow} Part\\
2 & $5.6 \cdot 10^{-3}$ & $2.1 \cdot 10^{-4}$ & $5.3 \cdot 10^{-4}$ & $1.0 \cdot 10^{-2}$ \\
3 & $1.2 \cdot 10^{-2}$ & $3.8 \cdot 10^{-4}$ & $7.5 \cdot 10^{-4}$ & $1.4 \cdot 10^{-2}$ \\
6 & $1.1 \cdot 10^{-1}$ & $1.5 \cdot 10^{-3}$ & $1.7 \cdot 10^{-3}$ & $2.0 \cdot 10^{-2}$ \\
10 & $1.1$ & $8.2 \cdot 10^{-3}$ & $5.9 \cdot 10^{-3}$ & $3.2 \cdot 10^{-2}$\\
20 & $2.1 \cdot 10^{1}$ & $4.7 \cdot 10^{-2}$ & $2.3 \cdot 10^{-2}$ & $8.7 \cdot 10^{-2}$ \\
40 & $-$ & $3.4 \cdot 10^{-1}$ & $1.4 \cdot 10^{-1}$ & $3.5 \cdot 10^{-1}$ \\
\bottomrule
\end{tabular}
\end{center}
\vskip -0.1in
\label{table:Timings}
\end{table}
We chose the \texttt{KNITRO v10.3} \citep{Byrd2006} Sequential Quadratic Optimization (SQP) non-linear solver with the default parameters for the optimization of OEI. Explicitly providing the hessian on the experiments of Figure \ref{fig:SyntheticBatch20} reduces \texttt{KNITRO}'s iterations by $49\%$ as compared to estimating the hessian via the symmetric-rank-one (SR1) update method included in the \texttt{KNITRO} suite. Given the inexpensiveness of calculating the hessian and the fact that \texttt{KNITRO} requests the calculation of the hessian less than a third as often as it requires the objective function evaluation we conclude that including the hessian is beneficiary.
We will now present simulation results to demonstrate the performance of OEI in various scenarios.
\subsection{Perfect modeling assumptions}
We first demonstrate that the competing Expected Improvement based algorithms produce clearly suboptimal choices in a simple \mbox{1--dimensional} example. We consider a 1--d Gaussian Process on the interval $[-1, 1]$ with a squared exponential kernel \citep{Rasmussen2005} of lengthscale $1/10$, variance $10$, noise $10^{-6}$ and a mean function $m(x) = (5x)^2$. An example posterior of 10 observations is depicted in Figure \ref{fig:GP_draw}. Given the GP and the 10 observations, we depict the optimal 3-point batch chosen by minimizing each acquisition function. Note that in this case we assume perfect modeling assumptions -- the GP is completely known and representative of the actual process. We can observe in Figure~\ref{fig:GP_draw} that the OEI choice is very close to the one proposed by QEI while being slightly more explorative, as OEI allows for the possibility of more exotic distributions than the Gaussian.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{.45\textwidth}
\includegraphics[width=1\textwidth]{plots/130.png}
\caption{Suggested 3-point batches of different algorithms for a GP posterior given 10 observations. The thick blue line depicts the GP mean, the light blue shade the uncertainty intervals ($\pm$ sd) and the black dots the observations. The locations of the batch chosen by each algorithm are depicted with colored vertical lines at the bottom of the figure.}
\label{fig:GP_draw}
\end{subfigure}
\quad
\begin{subfigure}[t]{.45\textwidth}
\centering
\includegraphics[width=.9\textwidth]{plots/EI_vs_batch_size.png}
\caption{Averaged one-step Expected Improvement on 200 GP posteriors of sets of 10 observations with the same generative characteristics (kernel, mean, noise) as the one in Figure \ref{fig:GP_draw} for different algorithms across varying batch size.}
\label{fig:EI_vs_batch_size}
\end{subfigure}
\end{figure}
In contrast the LP-EI heuristic samples almost at the same point all three times. This can be explained as follows: LP-EI is based on a Lipschitz criterion to penalize areas around previous choices. However, the Lipschitz constant for this function is dominated by the end points of the function (due to the quadratic trend), which is clearly not suitable for the area of the minimizer (around zero), where the Lipschitz constant is approximately zero. On the other hand, QEI-CL favors suboptimal regions. This is because QEI-CL postulates outputs equal to the mean value of the observations which significantly alter the GP posterior.
Testing the algorithms on 200 different posteriors, generated by creating sets of 10 observations drawn from the previously defined GP, suggests OEI as the clear winner. For each of the 200 posteriors, each algorithm chooses a batch, the performance of which is evaluated by sampling the multipoint expected improvement \eqref{eqn:ExpectedImprovement}. The averaged results are depicted in Figure \ref{fig:EI_vs_batch_size}. For a batch size of 1 all of the algorithms perform the same, except for OEI which performs slightly worse due to the relaxation of Gaussianity. For batch sizes 2-3, QEI is the best strategy, while OEI is a very close second. For batch sizes 4-5 OEI performs significantly better. The deterioration of the performance for QEI in batch sizes 4 and 5 can be explained in Figure \ref{fig:qei_problem}. In particular, after batch size 3, the calculation of QEI via the \texttt{R} package \texttt{DiceOptim} exhibits some points of failure, leading to suboptimal choices. Figure \ref{fig:qei_problem} also explains the very good performance of OEI. Although always different from the sampled estimate, it is \emph{reliable} and closely approximates the actual expected improvement in the sense that their optimizers and level sets are in close agreement.
\begin{figure}[t]
\centering
\includegraphics[width=.5\textwidth]{plots/qei_problem.pdf}
\caption{Demonstration of a case of failure in the calculation of the Multi-Point Expected Improvement with the \texttt{R} package \texttt{\tt DiceOptim v2} \citep{Roustant2012}. The figure depicts how the expected improvement varies around a starting batch of 5 points, $X_0$, when perturbing the selection across a specific direction $X_t$ i.e. $f(t) = \mathrm{EI}(X_0 + tX_t)$. The \texttt{\tt DiceOptim}'s calculation is almost always very close to the sampled estimate, except for a single breaking point. In contrast our acquisition function OEI, although always different from the sampled estimate, is \emph{reliable} and closely approximates QEI in the sense that their optimizers and level sets nearly coincide.}
\label{fig:qei_problem}
\end{figure}
\subsection{Synthetic functions}
\begin{figure}[!ht]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/sixhumpcamel.pdf}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/cosines.pdf}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/eggholder.pdf}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/hart6.pdf}
\end{subfigure}
\caption{BO of \emph{batch size 5} on synthetic functions. Red, blue, green, yellow and black dots depict runs of OEI, LP-EI, BLCB, QEI and Random algorithms respectively. Diamonds depict the median regret for each algorithm.}
\label{fig:SyntheticBatch5}
\end{figure}
\begin{figure}[!ht]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/eggholder_batch_20.pdf}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/hart6_batch_20.pdf}
\end{subfigure}
\caption{BO of \emph{batch size 20} on the challenging Hartmann 6-d Eggholder 2-d functions where OEI exhibits clearly superior performance. Red, blue, green and black dots depict runs of OEI, LP-EI, BLCB and Random algorithms respectively. Diamonds depict the median regret for each algorithm. Compare the above results with Figure \ref{fig:SyntheticBatch5} for the same runs with a smaller batch size. Note that QEI is not included in this case, as it does not scale to large batch sizes (see Table \ref{table:Timings}).}
\label{fig:SyntheticBatch20}
\end{figure}
Next, we evaluate the performance of OEI in minimizing synthetic benchmark functions. The functions considered are: the Six-Hump Camel function defined on $[-2, 2] \times [-1, 1]$, the Hartmann 6 dimensional function defined on $[0, 1]^6$ and the Eggholder function, defined on $[-512, 512]^2$. We compare the performance of OEI against QEI, LP-EI and BLCB as well as random uniform sampling. The initial dataset consists of 10 random points for all the functions. A Matern 3/2 kernel is used for the GP modeling \citep{Rasmussen2005}. As all of the considered functions are noiseless, we set the likelihood variance to a fixed small number $10^{-6}$ for numerical reasons. For the purpose of generality, the input domain of every function is scaled to $[-0.5, 0.5]^n$ while the observation dataset $y_d$ is normalized at each iteration, such that $\mathrm{Var}[y^d] = 1$. The same transformations are applied to QEI, LP-EI and BLCB for reasons of consistency. All the acquisition functions except OEI are optimized with the quasi-Newton L-BFGS-B algorithm \citep{Fletcher1987} with 20 random restarts. We use point estimates for the kernel's hyperparameters obtained by maximizing the marginal likelihood via L-BFGS-B restarted on 20 random initial points.
First, we consider a small-scale scenario of \emph{batch size 5}. The results of 40 runs of Bayesian Optimization on a mixture of Cosines, the Six-Hump Camel, Eggholder, and 6-d Hartmann functions are depicted in Figure \ref{fig:SyntheticBatch5}. In these functions, OEI approaches the minimums faster than QEI and BLCB while considerably outperforming LP-EI, which exhibits outliers with bad behavior. The explorative nature of OEI can be observed in the optimization of the Hartmann function. OEI quickly reaches the vicinity of the minimum, but then decides not to refine the solution further but explore instead the rest of the 6-d space. Increasing the \emph{batch size} to \emph{20} for the challenging Eggholder and Hartmann functions shows a further advantage for OEI. Indeed, as we can observe in Figure \ref{fig:SyntheticBatch20}, OEI successfully exploits the increased batch size. BLCB also improves its performance though not to the extent of OEI. In contrast, LP-EI fails to manage the increased batch size. This is partially expected due to the heuristic based nature of LP-EI: the Lipschitz constant estimated by LP-EI is rarely suitable for all the 20 locations of the suggested batch. Even worse, LP-EI's performance is \emph{deteriorated} as compared to smaller batch sizes. LP-EI is plagued by numerical issues in the calculation of its gradient, and suggests multiple re-evaluations of the same points. This multiple re-samplings affects the GP modeling, resulting in an inferior overall BO performance.
\subsection{Real world example: Tuning a Reinforcement Learning Algorithm on various tasks}
\begin{figure}[!ht]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/InvertedDoublePendulum-v1.pdf}
\end{subfigure}
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/Hopper-v1.pdf}
\end{subfigure}
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/Reacher-v1.pdf}
\end{subfigure}
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[width=1\textwidth]{plots/InvertedPendulumSwingup-v1.pdf}
\end{subfigure}
\caption{BO of batch size 20 on tuning PPO on a variety of robotic tasks. Red, blue, green and black dots depict runs of OEI, LP-EI, BLCB and Random algorithms respectively. Diamonds depict the median loss of all the runs for each algorithm.}
\label{fig:PPOResults}
\end{figure}
Finally we perform Bayesian Optimization to tune Proximal Policy Optimization (PPO), a state-of-the-art Deep Reinforcement Learning algorithm that has been shown to outperform several policy gradient reinforcement learning algorithms \citep{Schulman2017}. The problem is particularly challenging, as deep reinforcement learning can be notoriously hard to tune, without any guarantees about convergence or performance. We use the implementation \citet{Dhariwal2017} published by the authors of PPO and tune the reinforcement algorithm on 4 \texttt{OpenAI~Gym} tasks (Hopper, InvertedDoublePendulum, Reacher and InvertedPendulumSwingup) using the \texttt{Roboschool} robot simulator. We tune a set of 5 hyper-parameters which are listed in Table \ref{table:hyperparameters}. We define as objective function the negative average reward per episode over the entire training period ($4\cdot 10^5$ timesteps), which favors fast learning \citep{Schulman2017}. All of the other parameters are the same as the original implementation \citep{Schulman2017}.
We run Bayesian Optimization with batch size of 20, with the same modeling, preprocessing and optimization choices as the ones used in the benchmark functions. The results of 20 runs are depicted in Figure \ref{fig:PPOResults}. OEI outperforms BLCB, which performs comparably to Random search, and LP-EI, which exhibits high variance with occasional outliers stuck in inferior solutions.
\begin{table}[t]
\caption{List of PPO's Hyperparameters used for tuning. Items with asterisk are tuned in the log-space.}
\label{table:hyperparameters}
\vskip 0.15in
\begin{center}
\begin{tabular}{cc}
\toprule
{Hyperparameter} & {Range} \\
\midrule
Adam step-size & $[10^{-5}, 10^{-3}]^*$ \\
Clipping parameter & $[0.05, 0.5]$ \\
Batch size & $24, \dots, 256$ \\
Discount Factor $(\gamma)$ & $1 - [10^{-3}, 10^{-3/2}]^*$ \\
GAE parameter $(\lambda)$ & $1 - [10^{-2}, 10^{-1}]^*$ \\
\bottomrule
\end{tabular}
\end{center}
\vskip -0.1in
\end{table}
\section{Conclusions}
We have introduced a new acquisition function that is a tractable, probabilistic relaxation of the multi-point Expected Improvement, drawing ideas from the Distributionally Ambiguous Optimization community. Novel theoretical results allowed inexpensive calculation of first and second derivative information resulting in efficient Bayesian Optimization on large batch sizes.
\acks{This work was supported by the EPSRC AIMS CDT grant EP/L015987/1 and Schlumberger. The authors would like to acknowledge the use of the University of Oxford Advanced Research Computing (ARC) facility in carrying out this work. \url{http://dx.doi.org/10.5281/zenodo.22558}. Many thanks to Leonard Berrada for various useful discussions.}
|
\section{Introduction}
\label{sec:introduction}
The Stern-Gerlach apparatus, used in the last century to obtain an experimental evidence of angular momentum quantization~\cite{Stern21}, has also interesting features for atom optics~\cite{Robert91}. In particular, this device has been used successfully in the first observation of a geometric phase~\cite{Berry84} in atom interferometry~\cite{Miniatura92PRL}. This system entangles the external atomic motion with the total angular atomic momentum, and produces a spatial separation between atomic wave-packets corresponding to different angular momenta. The Stern-Gerlach device can be used to perform transformations on the atomic spins, mapping initial angular momentum states to determined final states.
These transformations are usually achieved with a magnetic field presenting an helicoidal profile~\cite{Miniatura92APB}. A simple way to map reliably initial spin states to well-defined target spin states is to design the magnetic field profile in such a way that the atomic spin follows the locally rotating magnetic field adiabatically in the course of its propagation through the device. The bottleneck of this approach is that
the adiabatic regime imposes a minimum length over which the magnetic field may change its direction. This length is proportional to the atomic velocity and inversely proportional to the magnetic field modulus. Since strong magnetic fields may be undesirable, there is a trade-off between the speed of the adiabatic evolution and the magnetic field strength.
The purpose of this article is to show that this trade-off can be greatly improved by using the shortcut to adiabaticity(STA) technique~\cite{STAReview13}. More precisely, we design a suitable magnetic field profile by using the reverse engineering approach based on the Lewis-Riesenfeld invariants~\cite{Lewis69,PRLSTA1_2010,NJP}. These methods enable one to guarantee a transitionless evolution faster than the time scale imposed by the adiabatic regime. The STA approach has been shown experimentally to efficently speed up the transport or manipulation of wave functions \cite{David08b,Bowler12,Walther12,Schaff10,Schaff11,Rohringer15,David13,David16} and even thermodynamical transformations \cite{Boltzmann,NaturePhys,APL}. Concerning the transfer of quantum states, recent impressive implementations have been reported in cold atoms experiments \cite{Bason12}, solid-state architectures~\cite{STASolidZhou17} or in opto-mechanical systems~\cite{STAOptoZhang17}. The transposition of those ideas to integrated optics devices has been recently explored \cite{Tseng12,Martinez14,Valle16}.
Our proposal of STA-engineered Stern-Gerlach device outperforms the traditional rotating field in terms of speed of the quantum evolution for a given amount of resources - here the magnetic field -. The STA-engineered Stern-Gerlach is also robust toward toward a dispersion in the atomic velocities, and may achieve very high fidelities in an atomic spin state transfer. We will illustrate our arguments by considering specifically the case of spin-one particles such as the hydrogen fragments issued from $H_2$ dissociation~\cite{Medina11,Medina12,Robert2013}. This example is relevant, since an experiment based on an arrangement of Stern-Gerlach devices has been recently proposed to evaluate the spin-coherence of this dissociation~\cite{Carvalho15}.
The paper is organized as follows. In Section~\ref{sec:Standard Stern Gerlach}, we investigate the efficiency of a Stern-Gerlach device using a plain helicoidal configuration of the magnetic field. In Section~\ref{sec:Shortcut Adiabaticity}, we provide the general framework to determine a suitable magnetic field in a Stern-Gerlach device using STA technique. In Section~\ref{sec:examples STA}, we study an example of such a magnetic field profile suitable to realize fast and robust angular momentum evolution. In particular, we estimate the quantum fidelity and the speed-up of the spin transfer enabled by this configuration. We also study the resilience toward a dispersion in the atomic propagation times, and compare the performance of this configuration with respect to a standard Stern-Gerlach apparatus using an equivalent magnetic field.\\\
\section{Efficiency of a Stern-Gerlach device with an helicoidal magnetic field}
\label{sec:Standard Stern Gerlach}
In this Section, we review the propagation of a spin-one particle in a standard Stern-Gerlach device. This apparatus involves an helicoidal magnetic field which rotates of an angle $\pi/2$ over a certain length, corresponding to a propagation time $T$ for a given class of atomic velocities. An atom propagating at constant velocity along the helicoid axis experiences locally a uniform rotation of the magnetic field. We investigate the transfer of the angular momentum from the initial state $| J=1, m_z=1 \rangle$ to the target state $| J=1, m_x=1 \rangle$ as a function of the atomic propagation time $T$.
For this purpose, the particle is subjected to the time-dependent Hamiltonian $\hat{H}(\mathbf{B}(t)) = - \gamma \: \mathbf{B}(t) \cdot \hat{\mathbf{J}}$ between the initial time $t=0$ and final time $t=T$. $\vec{\hat{J}}$ is the vectorial angular momentum operator of a spin-one particle and $\gamma$ accounts for the strength of the coupling and includes the atomic Land\'e factor. The magnetic field, $\mathbf{B}(t)$, in the comoving frame seen at the central atomic position~\cite{Footnote1}, reads $\mathbf{B}_{\rm st.}(t)= \Re \left[ B_0 (\hat{\mathbf{z}}+ i \hat{\mathbf{x}}) e^{i \frac {\pi} {2} \frac {t} {T}} \right] .$ As there is only a single characteristic time scale, namely the Larmor time $T_L=2 \pi / \gamma B_0 $ associated with the spin precession in the magnetic field, the efficiency of the quantum transfer may only depend on the ratio $T / T_L.$ We have used the Qutip package~\cite{Qutip1,Qutip2} to simulate the evolution of the initial spin state $| m_z=1 \rangle$ in a time-dependent rotating magnetic field. The efficiency of the spin transfer is quantified by the fidelity $\mathcal{F}$~\cite{Jozsa94} of the final state $| \psi( T ) \rangle$ to the target state $\mathcal{F}= | \langle m_x=1 | \psi( T ) \rangle |^2$.
The results are shown in Figure~\ref{fig:FidelityStandardSternGerlach}. For short durations $T \ll T_L,$ the fidelity of the final state is barely higher than that of the initial state. As expected, in this limit, the atomic angular momentum state is almost unaffected by the Stern-Gerlach device: the rotation of the magnetic field is too fast to drive the atomic spin. On the other hand, for times $T \gg T_L,$ the atomic spin follows adiabatically the field, yielding a final state very close to the target state. A quantum fidelity equal to unity is achieved when $T=T_L.$ For total times $T$ {greater or equal} than a few Larmor times $T_L$, the fidelity stays higher than $99\%$. The Larmor time $T_L$, inversely proportional to the strength of the magnetic field, thus corresponds to a reliable spin transfer in a standard Stern-Gerlach apparatus. We shall use this duration to determine the figure of merit of a Stern-Gerlach device enhanced by the STA technique. Using typical experimental parameters for Stern-Gerlach atom interferometry with hydrogen fragments~\cite{Miniatura92PRL}, we consider atomic velocities $v \: \sim \: 10 \: \: {\rm km/s}$, a gyromagnetic ratio $\gamma \simeq \mu_b / \hbar$ and a magnetic field $B_0 \: = \: 0.1 \: {\rm G}$. a Larmor time
$\nu_L \: = \: 0.14 \: {\rm MHz}.$ The corresponding Larmor time $T_L \: \simeq \: 7 \: \: {\rm \mu s}$ then yields the minimum length of a few $L\: \simeq \: 7 \: \: {\rm cm}$ for an efficient spin transfer in a Stern-Gerlach device with an helicoidal magnetic field.
Finally, we note that the fidelity is a non-monotonic function of the ratio $T/T_L$. The small oscillations shown in Figure~~\ref{fig:FidelityStandardSternGerlach} are consistent with the predictions for the spin transition amplitudes of atomic spins experiencing a non-adiabatic evolution in an inhomogeneous magnetic field~\cite{Hight77}, evidenced experimentally in~for metastable hydrogen atoms~\cite{Hight78,Robert92}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=8 cm]{Fig1new}
\end{center} \caption{Fidelity of the final state $\mathcal{F}= | \langle m_x=1 | \psi( T ) \rangle |^2$ as a function of the ratio $T/T_L$ between the total time and the Larmor time. This profile is independent of the considered Larmor pulsation, which has been set to $\omega_L \: = \:10^6 \: \: {\rm rad/s}$ in the numerical simulation.}
\label{fig:FidelityStandardSternGerlach}
\end{figure}
\section{Determination of the magnetic field from Lewis Riesenfeld invariants}
\label{sec:Shortcut Adiabaticity}
In this Section, we search for a magnetic field profile suitable to realize fast spin evolution of a spin-one particle in a Stern-Gerlach device, with an initial angular momentum along the $Oz$ axis and a target angular momentum state along the $Ox$ axis. We find the general equations for the field and determine suitable boundary conditions to be satisfied by the Lewis-Riesenfeld invariants~\cite{Lewis69}.
\subsection{General equations for the dynamical invariant}
For a given time-dependent Hamiltonian, $ \hat{H}(t)$, a dynamical invariant, $\hat I(t)$, fulfills the equation \cite{Lewis69}:
\begin{equation}
i \hbar \frac {\partial \hat{I}(t)} {\partial t} = [ \hat{H}(t),\hat{I}(t)].
\label{invariant}
\end{equation}
A natural choice for the rotation of a spin 1 is to search for a dynamical invariant in the form $\hat{I}(t) = \mathbf{u}(t) \cdot \mathbf{J}(t)$ where the spin operators $\{ \hat{J}_{k} \}$ with $k=x,y,z$ form a closed Lie algebra~\cite{LieAlgebra1,LieAlgebra2} as the Pauli matrices for SU(2) and $ \mathbf{u}(t)$ is a vector that needs to be determined. Interestingly, the dynamical operatorial equation (\ref{invariant}) can be translated in a simple linear differential equation
describing a clockwise precession of the vector $\mathbf{u}(t)$ around the magnetic field $\mathbf{B}(t)$
\begin{equation}
\frac {d \mathbf{u}(t)} {dt} = \gamma \mathbf{u}(t) \times \mathbf{B}(t).
\label{precession}
\end{equation}
In the following, we set the evolution of $\mathbf{u}(t)$ and infer from Eq.~(\ref{precession}) the explicit expression for $\mathbf{B}(t)$ \cite{DavidarXiv17}. This amounts to reverse engineer Eq.~(\ref{precession}).
Actually, we have a lot of freedom to choose the function $\mathbf{u}(t)$ {and this choice does not fully constrains the function $\mathbf{B}(t)$. In what follows, we fix $B_y=0$ and proceed to determine $\mathbf{B}(t)$ from Eq.~(\ref{precession})~\cite{Footnote2}. Finally, to connect the eigenstates at initial and final time of $\hat I (t)$ and $\hat H (t)$, we impose the following commutation relations $[ \hat{H}(0),\hat{I}(0)]=[ \hat{H}(T),\hat{I}(T)]=0$. \\
\subsection{Resilience of the atomic spin transfer towards an atomic velocity dispersion}
\label{subsec:resilience}
Before proceeding, we highlight an interesting property arising from the commutation between the invariant and Hamiltonian operators at the final time, i.e. $[ \hat{H}(T),\hat{I}(T)]=0$. From the equation of motion~(\ref{invariant}), this condition implies that $\left. \frac {\partial \hat{I}} {\partial t} \right|_{T} =0.$ Consequently, to leading order,
the eigenstates of the dynamical invariant $\hat{I}(T)$ are unaffected by a small fluctuation of the interaction time $T$ with respect to a reference value $T_0.$ By construction, the atomic spin is an eigenstate of the invariant operator at all times. One thus expects that the final atomic spin state should also be unchanged to leading order by a small fluctuation of the atomic time of flight. This property enables one to perform an efficient spin rotation over a broad range of particle velocities, turning the apparatus robust against an atomic velocity dispersion. In the next Section, we will verify numerically this resilience of the spin transfer for a concrete example of magnetic field.
\subsection{Relation between the magnetic field and dynamical invariant}
The vector $\mathbf{u}(t)$ is conveniently parametrized in spherical coordinates as a unit vector, $\mathbf{u}(t)= (\sin \theta \cos \varphi, \sin \theta \sin \varphi, \cos \theta)$. From the precession equation (\ref{precession}), we get the components of the magnetic field as a function of the spherical angles $\theta(t)$ and $\varphi(t)$
\begin{equation}
\gamma B_x (t) = \frac {\dot{\theta}} {\sin \varphi}, \quad
\gamma B_z (t) = -\dot{\varphi} + \frac {\dot{\theta} \cos \theta \cos \varphi} {\sin \theta \sin \varphi} \, . \label{eq:BfieldEqs}
\end{equation}
We shall thus search for acceptable angular functions avoiding divergences in the magnetic field. Typically, as seen from the equations above, such divergence may occur when the invariant pointer crosses the equatorial line defined by $\theta=0$ or the meridians defined by $\varphi=0,\pi$. This geometric constraint on the pointer trajectory sets a limit on the shortest spin transfer time achievable with the STA method when using polynomial angles of a certain degree.
\subsection{Boundary conditions on the spherical coordinates of the invariant.}
We derive here the boundary conditions to be fulfilled by the angular functions $(\theta(t),\varphi(t)).$ These functions must enable the commutation between the invariant pointer and the Hamiltonian at the initial and final times.
Since the initial and target states must be eigenvectors of the initial and final Hamiltonian respectively, the magnetic field direction at these times is fixed according to $\mathbf{B}(0)= B_z^I \: \hat{\mathbf{z}}$ and $\mathbf{B}(T)= B_x^F \hat{\mathbf{x}}$. The commutation between the invariant and Hamiltonian is then ensured by setting the pointer $\mathbf{u}(t)$ parallel to the magnetic field at these times. This is achieved by imposing the following conditions on the spherical coordinates
\begin{eqnarray}
\theta(0)=\pi, \quad \theta(T)=\pi/2, \nonumber \\
\varphi(0)=\pi/2, \quad \varphi(T)=0 \, . \label{eq:boundary1}
\end{eqnarray}
We now search for suitable expansions of the angular functions near the initial and final times that yield through Eq.~(\ref{eq:BfieldEqs}) the magnetic field $ B_x (0)=0$, $ B_z (0)=B_z^I$ and $B_x (T)=B_x^F$, $B_z (T)=0.$ Using the perturbative expansion $\theta(t)=\pi+ \alpha \: t^m + o(t^m) $ and $\varphi(t)=\pi/2+ \beta \: t^n + o(t^n) $ in the vicinity of $t=0$, one obtains that the condition on the initial magnetic field is equivalent to $m>1$, $n=1$, and $\beta = - \gamma B_z^I / (m+n).$ The lowest-order expansion compatible with this condition corresponds to $n=1$ and $m=2$, that is to say
\begin{equation}
\dot{\theta}(0)=0, \quad \ddot{\theta}(0)>0, \quad \dot{\varphi}(0)= - \gamma B_z^I /3 \label{eq:boundary2} \,.
\end{equation}
Using a similar expansion close to the final time with $\tau=t-T,$ $\theta(\tau)=\pi/2+ \alpha' \: \tau^{m'} + o(\tau^{m'}) $ and $\varphi(\tau)= \beta' \: \tau^{n'} + o(\tau^{n'}) $ with $\tau=T-t,$ one obtains that the second set of conditions is equivalent to $m' = n' + 1$, $n'>1$ and $m' \alpha'= \gamma \beta' B_x^F$. It can be fulfilled by choosing $n'=2$ and $m'=3$, yielding another set of conditions
\begin{eqnarray}
& \varphi(T)=0, & \quad \dot{\varphi}(T)=0, \nonumber \\
& \dot{\theta}(T)=0, & \quad \ddot{\theta}(T)=0, \quad \dddot{\theta}(T)= \gamma B_x^F \ddot{\varphi}(T) \label{eq:boundary3} \,.
\end{eqnarray}
\section{Example of fast magnetic spin driving with shortcut to adiabaticity.}
\label{sec:examples STA}
In this Section, we give an example of magnetic field profile realizing the STA, by finding suitable polynomial functions for the spherical coordinates of the invariant pointer. We show by numerical simulations that this magnetic field profile yields an extremely reliable transfer of a single quantum state. We also discuss quantitatively the enhancement brought by the STA in a Stern-Gerlach device. For this purpose, one must consider the trade-off between the speed-up brought by the STA and the amount of resources involved~\cite{STAOptoZhang17}. In contrast to the standard Stern-Gerlach device, the STA-engineered Stern-Gerlach involves a magnetic field with a time-dependent amplitude. To work out explicitly the comparison between the two devices, we shall either consider as a resource the average magnetic field seen by the atom in the STA device, i.e. $B_{\rm av}= (1/T) \int_0^T dt ||\mathbf{B}(t)||,$ or the maximum magnetic field $B_{\rm max}= {\rm max} \{ ||\mathbf{B}(t)||,t \in [0,T] \}.$ These two criteria will provide different figures of merit. As seen below, with both criteria the STA-engineered Stern-Gerlach outperforms the standard device.
\subsection{Example of suitable magnetic field profile.}
A simple way to match simultaneously the conditions~(\ref{eq:boundary1},\ref{eq:boundary2},\ref{eq:boundary3}) is to look for polynomials functions of the form $\theta(t)=P\left( \frac {t} {T} \right)$ and $\varphi(t)=Q \left( \frac {t} {T} \right)$. Lowest-order suitable polynomials can be obtained as:
\begin{eqnarray}
P(x) & = & \pi - 3 \pi x^2 + 4 \pi x^3 - \frac {3 \pi} {2} x^4 \nonumber \\
Q(x) & = & \frac {\pi} {2} - \frac {\mathcal{B}_z^I } {3} x + \left(- 3 \pi - \frac {6 \pi} {\mathcal{B}_x^F}+ \mathcal{B}_z^I \right) x^2 \label{eq:polynomials} \\
& + & \left( 4 \pi + \frac {12 \pi} {\mathcal{B}_x^F} - \mathcal{B}_z^I \right) x^3
+ \left(- \frac {3 \pi} {2} - \frac {6 \pi} {\mathcal{B}_x^F} +\frac {\mathcal{B}_z^I } {3} \right) x^4 \nonumber
\end{eqnarray}
where we have introduced the adimensional magnetic fields $\overrightarrow{\mathcal{B}}^I \: = \: \gamma \: T \: \mathbf{B}(0)$ and $\overrightarrow{\mathcal{B}}^F \: = \: \gamma \: T \: \mathbf{B}(T).$ Let us stress that this choice of lowest-order polynomials is by no means unique, since one of the relations only constrains the ratio of the time derivatives of the angular functions. With this choice, the azimuthal angle $\theta(t)$ of the field is completely independent of the initial and final values of the magnetic field, which affect the longitude angle $\varphi(t)$.\\
At this stage, one can determine the full magnetic field profile from Eq.~(\ref{eq:BfieldEqs}). The values of the angle $\varphi(t)$ should be kept in the interval $]0,\pi[$ in order to avoid a divergence in the magnetic field~\cite{Footnote3}. For this purpose, one must choose carefully the sign of the magnetic field component $B_x^F$ at the final time. Indeed, from Eq.(\ref{eq:polynomials}) one has $\dddot{\theta}(T)<0,$ so by virtue of Eq.(\ref{eq:boundary3}) one must have $B_x^F <0$ to ensure that $\varphi(T)=0$ is a local minimum. As a consequence of this choice, the considered device maps the initial state $| m_z=1\rangle$ to the target spin state $| m_x=-1\rangle$. The Figure below shows an example of magnetic field profile determined by Eq.(\ref{eq:polynomials}). For the considered parameters and with the propagation time $T \: = \: 2 \: {\mu s}$, an efficient state transfer is achieved in a STA-enhanced Stern-Gerlach with a maximum magnetic field $B_{\rm max} \:\simeq \: 0.22 \: {\rm G}$ to be compared with the magnetic field $B_{\rm st}^0= 2 \pi / \gamma T \simeq \: 0.36 \: {\rm G}$ required in a standard Stern-Gerlach.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=8.8 cm]{Fig2new}
\end{center} \caption{Time-dependent magnetic field determined by reverse engineering with the Lewis-Riesenfeld invariant method for a total atomic propagation time $T \: = \: 2 \: {\mu s}$ and with the initial and final magnetic field components fixed to $B_z^I \: = \: 0.1{\rm G}$ and $B_x^F \: = \: - 0.1{\rm G}$ respectively. (a) Trajectory of the pointer vector $\mathbf{u}(t)$ on the unit sphere. The initial point is on the pole and ending and the final point on the equator. (b) Parametric plot of the magnetic field seen at the atomic central position. (c) Plot of the norm $|| \mathbf{B}(t) ||$ (solid purple line), of the $-B_x(t)$ (dash-dotted purple line) and $B_z(t)$ (dotted purple line) components of the magnetic field as a function of time. The dashed red line represents the larger magnetic field $B^0_{\rm st}$ enabling a reliable spin transfer in a standard device during the time $T$.}
\label{fig:Magnetic Field Profile}
\end{figure}
\subsection{Fidelity and resilience of the spin transfer in a STA-engineered Stern-Gerlach device.}
Using the magnetic field defined in Eqs.~(\ref{eq:BfieldEqs}), we simulate~\cite{Qutip1,Qutip2} the temporal evolution of an atomic spin in a Stern-Gerlach device with a STA-designed magnetic field. Precisely, we keep track of the expectation values of the angular momentum projections $\langle \hat{J}_z \rangle (t)$ and $\langle \hat{J}_x \rangle (t)$, as well of the fidelity of the atomic spin state with respect to the target state $ | m_x=-1 \rangle$. The latter may be written as $\mathcal{F}(t) = | \langle m_x=-1 | \psi(t) \rangle|^2$, where the quantum state $|\psi(t) \rangle= \mathcal{T} \exp \left[ - \frac {i} {\hbar} \int_0^T dt \hat{H}(\mathbf{B}(t)) \right] | m_z=1 \rangle$ is the time-dependent atomic spin state evolved from the initial state $| m_z=1 \rangle$ through the interaction with the time-dependent magnetic field $\mathbf{B}(t).$
The corresponding results are shown on Figure~\ref{fig:SpinExpectation} and Figure~\ref{fig:FidelitySTASternGerlach} respectively. We first consider a STA Stern-Gerlach device and a standard Stern-Gerlach apparatus using equivalent magnetic fields. When the atomic spin propagates in a standard Stern-Gerlach device, whose magnetic field is $B_{\rm av}= (1/T) \int_0^T dt ||\mathbf{B}(t)||,$ the expectation value $\langle \hat{J}_x \rangle (T)> -1$ and the relatively low fidelity achieved ($\mathcal{F} \simeq 0.60$) reveal an imperfect transfer to the target state $ | m_x=-1 \rangle$. Even when the maximum field modulus $B_{\rm max}$ is used in the standard Stern-Gerlach device, the fidelity of the spin state to the target state saturates at the value $\mathcal{F} \simeq 0.80$. In contrast, when the atomic spin propagates in the STA-designed magnetic field, the atomic spin projections expectation values are very close to $\langle \hat{J}_z \rangle (T)=0$ and $\langle \hat{J}_x \rangle (T) =-1,$ showing that the final spin state is very close to the target state $| m_x=-1 \rangle$ at the final time $T$. With a standard Stern-Gerlach device using a larger magnetic field $B_{\rm st}^0$, one may achieve a good fidelity in the quantum state transfer over a large interval of propagation times $t$.
Nevertheless, the behaviour of the fidelity near the optimal time is different for the standard and for the STA Stern-Gerlach, enabling the latter to reach very high fidelities in a single state transfer. Indeed, the error committed in this transfer $\epsilon(t)=1-\mathcal{F}(t)$ decreases sharply for the STA device and can become arbitrarily low in the vicinity of the ideal propagation time $T$. Differently, for the standard SG device a finite error remains even at this ideal time. Its value depends on the magnetic field involved. In order to obtain a quantitative measure of the reliability of the spin transfer with a STA device, we analyse different fidelity thresholds for propagation times $t$ close to the ideal time $T$. Noting $\Delta t= t-T$ the time propagation mismatch, the inset of Figure~\ref{fig:FidelitySTASternGerlach} reveals that the error committed can be as low as $\epsilon(t) \leq 10^{-8}$ for $ | \Delta t / T | \leq 2 \times 10^{-3} $, or $\epsilon(t) \leq 10^{-5}$ for $ | \Delta t / T | \leq 1.5 \: \times 10^{-2},$ and $\epsilon(t) \leq 10^{-2}$ for $ |\Delta t / T| \: \leq 0.1 .$
Equivalently, these fidelities can be achieved for a certain velocity interval $\Delta v=v-v_0$ around a reference velocity $v_0$ for which the STA magnetic field has been designed. The resilience of the STA and standard Stern-Gerlach devices increases with the strength of the magnetic field involved. With a maximum magnetic field of $B_{\rm max} \: = \: 0.22 \: {\rm G}$, the STA Stern-Gerlach achieves a fidelity $\mathcal{F} \geq 99 \%$ for a class of velocities $\Delta v$ such that $| \Delta v /v_0| \leq 10 \%$. This corresponds to a velocity spread of $\Delta v \simeq 1 {\rm km/s}$ for the experiments~\cite{Hight78,Miniatura92APB} or for the slow beams of metastable hydrogen obtained from molecular dissociation in~\cite{Medina12,Robert2013}, or $\Delta v \simeq 4 {\rm km/s}$ for the fast hydrogen beams~\cite{Medina11,Robert2013}.
These measures can be compared with the reliability threshold for a universal set of quantum gates in order to obtain scalable quantum error correction~\cite{Shor96,Knill98,Preskill98}. As shown in these references, the availability of a finite set of gates enabling universal quantum computation and with a probability of failure below a certain threshold indeed enables one to implement large scale quantum error correction. The exact value of this threshold depends on a variety of factors such as the structure of the code employed and on the noise model~\cite{Gottesman09}. Earlier estimates of this threshold gave a maximum failure probability of $p \simeq 2.10^{-5}$~\cite{Preskill06} for seven-qubit codes. Using a different topology, surface code quantum computing~\cite{Fowler12,MartinisNature14} enables efficient quantum computation with gates fidelities of $\mathcal{F} \geq 99 \%$.
The Stern Gerlach device enhanced by the STA technique may thus obtain the transfer of a single quantum state with a reliability over this threshold for a significant range of propagation times, suggesting that this apparatus could be used within a quantum computing architecture.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=8 cm]{Fig3new}
\end{center} \caption{ Average values of the angular momentum projections (a) $\langle \hat{J}_z \rangle(t)$ and (b) $\langle \hat{J}_x \rangle(t)$ in units of $\hbar$ as a function of the propagation time $t$. We consider the propagation in a STA-enhanced Stern-Gerlach device (solid purple lines), in standard Stern-Gerlach devices using either a magnetic field equal to the temporal average of the STA magnetic field (dashed purple lines) either a larger magnetic field $B^0_{\rm st}$ enabling an efficient spin transfer during the time $T \: = \: 2 \: {\mu s}$ (dashed red lines). Purple lines represent the use of equivalent resources in terms of magnetic field. The numerical parameters and magnetic fields are identical to those used in Figure~\ref{fig:Magnetic Field Profile}.} \label{fig:SpinExpectation}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9 cm]{Fig4new}
\end{center} \caption{Fidelity of the atomic spin state during its propagation in the Stern-Gerlach device: STA device (solid purple line), in the standard Stern-Gerlach apparatus with a magnetic field of modulus $B_{\rm av}$ (dotted purple line), $B_{\rm max}$ (dash-dotted purple line), and $B_{\rm st}^0$ (dotted red line). Purple lines represent the use of equivalent resources in terms of magnetic field, while the dotted red line corresponds to a stronger magnetic field $B_{\rm st}^0.$ The inset gives a logarithmic plot of the error committed $\epsilon(t)=1-\mathcal{F}(t)$ in the quantum state transfer for times $t$ in the vicinity of the reference time $T= 2 \: {\mu s}$. The magnetic fields correspond to the parameters of Figure~\ref{fig:Magnetic Field Profile}.}
\label{fig:FidelitySTASternGerlach}
\end{figure}
\subsection{Speed-up offered by the STA-engineered magnetic field.}
In order to evaluate quantitatively the benefits offered by the STA-engineered magnetic field over the standard helicoidal configuration, we estimate the resources -- in terms of magnetic field -- required to achieve a perfect spin transfer in a given total propagation time. As previously, we consider either the average or the maximum magnetic fields involved in the STA configuration to determine the equivalent magnetic field in the standard STA device. For a range of times $T$, we derive the STA-engineered magnetic field from Eq.(\ref{eq:polynomials}), which gives readily the average and maximum magnetic fields involved. As seen previously, the magnetic field required in the standard device is inversely proportional to the propagation time $T$.
Figure \ref{fig:EfficiencySTAResources} compares the performances of the STA-engineered and the standard Stern-Gerlach apparatus from two equivalent points of views. Figure~\ref{fig:EfficiencySTAResources}(a) shows magnetic field required in both devices for a range of atomic spin transfer time $T$, while Figure~\ref{fig:EfficiencySTAResources}(b) reveals the speed-up offered by a STA Stern Gerlach in comparison with a standard Stern Gerlach loaded with a magnetic field of similar strength. These figures reveal that for a total transfer time $T \gtrsim 0.8 \: {\rm \mu s},$ the STA-engineered Stern Gerlach device is more performant. The maximum magnetic field involved in the STA device is smaller than the magnetic field required in a standard Stern-Gerlach device. Equivalently, the standard Stern-Gerlach requires a longer atomic propagation time when using a magnetic field of modulus equal to the maximum STA-magnetic field. Note that the curves associated with the maximum magnetic field in the STA device present a slope discontinuity~\cite{Footnote4}.
On the other hand, when the atomic spin transfer time is such that $T \lesssim 0.8 \: {\rm \mu s},$ the standard Stern-Gerlach becomes more efficient. Indeed, close to $T=0.5 \: {\rm \mu s}$, the magnetic field involved in the STA device diverges. This issue is related to the divergences generated by the roots of the angular functions $\theta(t)$ and $\phi(t)$, which appear in the interval $]0,T[$ when $T$ goes below a certain value. By using polynomial functions of higher order, it is possible to go to shorter times, while preserving a small enough magnetic field. When considering a family of polynomials of a given order, these divergences set a lower bound on the times achievable by the STA-engineered Stern Gerlach devices.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=7.5 cm]{Fig5new}
\end{center} \caption{ (a) Maximum (solid red line) and average (solid purple line) magnetic fields used in a STA-engineered Stern-Gerlach device, compared to the magnetic field in a standard Stern-Gerlach device (dotted black line), in order to perform an efficient spin transfer during the total time $T$. (b) Speed-up offered by the STA configuration. Ratio between the transfer time $T_{\rm st.}$ required in a standard Stern-Gerlach device to the time $T_{\rm STA}$ required in a STA device. The standard Stern-Gerlach uses either a magnetic field of modulus $B_{\rm max}$ corresponding to the maximum field in the STA device (red line) or a magnetic field of modulus $B_{\rm av}$ corresponding to the average field in the STA device (purple line). }
\label{fig:EfficiencySTAResources}
\end{figure}
To conclude, we have proposed to enhance the performances of a Stern-Gerlach device by using the technique of shortcut to adiabaticity. We have considered the propagation of spin-one particles in the device. Using the Lewis-Riesenfeld invariant approach, we have found a suitable magnetic field by reverse-engineering of the dynamical equation of motion for the invariant. The commutation between the invariant pointer and the Hamiltonian at initial and final times reduces the sensibility of the final state to the total propagation time. Using numerical simulations, we have demonstrated the validity of our approach, and provided a quantitative picture of the enhancement offered by this technique. The STA-engineered Stern-Gerlach apparatus appears to offer a better trade-off between time of propagation and magnetic field involved in the device. This conclusion is generic and valid for higher angular momentum. It provides the physical limits for the miniaturization of Stern and Gerlach device to entangle external and internal degrees of freedom. Finally, the STA-engineered Stern-Gerlach apparatus may achieve for the transfer of a single atomic spin extremely high fidelities for a narrow range of velocities, and fidelities above $99\%$ over a broad range of velocities. Such fidelities are below the error threshold for scalable quantum computation with surface codes. The fidelity enhancement provided by the STA method may open the possibility to use the Stern-Gerlach devices in the context of quantum information processing.
\acknowledgments
F.I. acknowledges a very fruitful collaboration and enlightening discussions on atom optics and Stern-Gerlach interferometry with Carlos Renato de Carvalho, Ginette Jalbert and
Nelson Velho de Castro Faria.
|
\section{Introduction}
\IEEEPARstart{W}{ith} the rapid growth of advanced mobile devices and social networking services, tremendous
amount of landmark images have been generated and disseminated in popular
social networks. Mobile landmark search (MLS)
is gaining its importance and increasingly becomes one of the most important
techniques to pervasively and intelligently access knowledge about the landmarks of interest
\cite{sigpro/ChengS16,DBLP:conf/mmm/ChengRSM13}.
However, MLS still remains unsolved due to two important challenges. 1) \emph{Low-bit query transmission}.
Mobile devices have limited computational power, relatively small memory storage, and short battery lifetime.
Consequently, a client-server architecture is the main stream search paradigms adopted in
existing landmark search systems, where query is captured and submitted by mobile devices, computation-intensive landmark search
is performed on the remote sever with strong computing capability and rich image resources.
Since wireless bandwidth from mobile devices to server
is limited, how to generate a compact signature for query to achieve low bit
data transmission becomes vital important. 2) \emph{Huge visual query variations}. Visual
splendour of a landmark can be photographed by multiple tourists under various circumstances, e.g. a wide sampling of positions, viewpoints, focal lengths, various weather conditions or illuminations. Therefore, the recorded images will have visual variations.
Besides, landmarks could be comprised of a wide range of attractive sub-spots. The images taken for sub-spots of the same landmark may appear with more visual diversity \cite{tmm/ZhuSJXZ15}. A typical example is presented in Fig. \ref{fig:diversity}. All the characteristics of landmark inevitably diversify the
visual appearances of the recorded query images, thus posing great challenges on MLS search system to robustly accommodate visual queries.
\begin{figure}
\centering
\includegraphics[width=85mm]{diversity.pdf}
\caption{The left sub-figure shows the images recoded for landmark \emph{Pantheon} from different viewpoints. The right sub-figure presents images recorded for different sub-pots of landmark \emph{Tiananmen Square}. All above landmark images demonstrate huge visual diversity, which presents great challenges on the mobile landmark search.}
\label{fig:diversity}
\vspace{-4mm}
\end{figure}
Hashing \cite{7915734,7047876,Song2017,leizhutcyb,leitkde2017} is a promising technique to achieve low-bit query transmission for
supporting efficient mobile landmark search. Its main objective is to learn compact binary codes from high-dimensional data, while
preserving the similarity of the original data. Thus, it can significantly reduce transmission cost with storage-efficient binary embedding, and moreover, speedup the search process with simple but efficient Hamming distance computations \cite{DCFS,rhcmass}. However, most existing hashing strategies developed for MLS are based on uni-modal visual-words based features \cite{IJCVMLS,AIJMLS,TCFSCH}. They generally suffer from
1) limited discriminative capability and 2) poor robustness against visual variations \cite{MLSTMtao,TCFSCH}.
Although general multi-modal hashing techniques \cite{MVAGH,HASHMFTWO,HASHMFONE,MVLH} can
improve discriminative capability of binary codes \cite{tcyb/ZhuSJZX15,Xie2016}, they are designed based on simple
multiple feature integration without specific query robustness accommodation. Hence, their performance may be limited
on searching mobile landmark images.
Motivated by the above considerations, in this paper, we propose a novel hashing scheme, named as canonical view based discrete
multi-modal hashing (CV-DMH), to facilitate efficient and robust MLS. We define canonical views as the
views which capture the key characteristics of visual landmark appearances with limited redundancy. Based on them, an arbitrary image
captured by tourists or users can be robustly represented using either a specific canonical view or the cross-scenery of
multiple canonical views. Accordingly, varied visual contents of landmark can
be effectively characterized using their visual correlations to certain
canonical views. Through encoding these relations into the binary codes, various visual queries can be robustly accommodated. Furthermore, the low-cost query transmission and fast search can be well supported.
Specifically, CV-DMH works with a three-stage learning procedure.
First, a submodular function is designed to measure visual representativeness and redundancy of a view set.
With it, the optimal canonical views are efficiently discovered
by an iterative mining strategy with theoretical guarantee. Second, multi-modal sparse
coding is applied to transform visual features from multiple
modalities into a unified intermediate representation. It can robustly
characterize visual contents of varied landmark images with certain canonical views. Finally, compact binary codes are
learned on intermediate representation within a tailored discrete
binary embedding model. The learning objective is to preserve visual relations of
images measured with canonical views and remove the involved noises. In
this part, we not only explicitly deal with
the discrete constraint of hashing codes, but also consider the bit-uncorrelated
constraint and balance constraint together. A novel augmented Lagrangian
multiplier (ALM) \cite{ALMone} based optimization method is developed to directly solve
the discrete binary codes. It can desirably avoid the accumulated quantization errors in conventional
hashing method which simply adopts ``relaxing+rounding" optimization framework.
The contributions of this paper are summarized as follows:
\begin{enumerate}[1.]
\item A submodular function is designed to measure visual representativeness
and redundancy of a view set. With it, an iterative mining strategy is proposed
to efficiently identify canonical views of landmarks using multiple modalities.
Theoretical analysis demonstrates that it can obtain near-optimal solutions.
\item A novel intermediate representation generated by multi-modal sparse coding
is proposed to robustly characterize the visual contents of varied landmark images.
It provides a natural and effective connection between the canonical views and binary
embedding model.
\item A binary embedding model tailored for canonical views is developed to
preserve visual relations of images into binary codes and thus support efficient
MLS with great robustness. We propose a direct discrete hashing optimization method based on ALM.
It effectively avoids the accumulated quantization errors in
conventional hashing methods which simply adopt ``relaxing+rounding" optimization framework.
\end{enumerate}
Compared with our previous work \cite{2CVR}, several enhancements have been made in this paper. They can be summarized as follows:
\begin{enumerate}[1.]
\item We conduct a comprehensive review of related work and introduce more details of the proposed approach.
\item We propose a novel discrete binary embedding model. The iterative computation for solving hashing codes enjoys an efficient optimization process. Moreover, direct discrete optimization can successfully avoid the accumulated quantization errors.
\item More experiments are conducted and the presented results validate the effects of the proposed approach.
\end{enumerate}
The rest of the study is structured as follows. Section \ref{sec:2} introduces related work.
System overview of CV-DMH based MLS system is illustrated in Section \ref{sec:so}.
Details about CV-DMH are introduced in Section \ref{sec:3}. Experimental configuration is presented
in Section \ref{sec:4}. In Section \ref{sec:5}, we give experimental results and analysis.
Section \ref{sec:6} concludes the study with a summary and future work.
\section{Related Work}
\label{sec:2}
Due to the constraints of space, only the work most related to this study is introduced in this section. In particular,
we present a short literature review on mobile landmark search, multi-modal hashing, and discrete hashing optimization.
\subsection{Mobile Landmark Search}
Most existing approaches developed for mobile landmark search focus on compressing high-dimensional visual-words based landmark feature into binary codes to achieve low bit rate data transmission. Ji \emph{et al}. \cite{IJCVMLS} present a location discriminative vocabulary coding (LDVC) to compress bag-of-visual-words (BoVW) with location awareness by combining both visual content and geographical context. The compact landmark descriptor is generated with an iterative optimization between geographical segmentation and descriptor learning. Duan \emph{et al}. \cite{AIJMLS} explore multiple information sources, such as landmark image, GPS, crowd-sourced hotspot
WLAN, and cell tower locations, to extract location discriminative compact landmark image
descriptor. Chen \emph{et al}. \cite{MLSTMtao} develop a soft bag-of-visual phrase (BoVP)
to learn category-dependent visual phrases, by capturing co-occurrence features of
neighbouring visual-words. The context location and direction information captured by
mobile devices are also integrated with the proposed BoVP. To alleviate the online memory cost,
Zhou \emph{et al}. \cite{TCFSCH} propose codebook-free scalable cascaded hashing (SCH) for mobile landmark search.
In their approach, matching recall rate is ensured first and false positive matches are then removed with a subsequent verification step.
All the aforementioned techniques learn compact binary codes from only visual-words based
features, without considering valuable information from other visual modalities. Therefore,
they will generate binary codes with limited discriminative information. To the best of our knowledge,
only \cite{MLS3D} is proposed to robustly accommodate landmark images under various conditions. However, it
sends compressed query images with low resolution for landmark search, Consequently, it still consumes considerable transmission bandwidth.
\begin{figure*}
\centering
\includegraphics[width=180mm]{framework.pdf}
\vspace{-3mm}
\caption{Framework of the CV-DMH based mobile landmark search system. The main aim of the offline learning is
to learn hashing codes of database images and hashing functions for online query. This part mainly consists of three steps: multi-modal canonical view mining, intermediate representation generation, and discrete binary embedding. The online searching part generates hashing codes for queries and performs efficient online similarity search in Hamming space.}
\label{fig:sys}
\vspace{-3mm}
\end{figure*}
\subsection{Multi-modal Hashing}
Multi-modal hashing has been emerging as a promising technique to generate compact binary
code based on multiple features. The earliest study on this topic is composite hashing with
multiple information sources (CHMIS) \cite{HASHMFFOUR}. It integrates discriminative information from multiple
sources into the hashing codes via weight adjustment on each individual source to maximize the hashing embedding performance.
However, CHMIS just post-integrates linear output of features and fails to fully exploit their correlations.
Kim \emph{et al}. \cite{MVAGH} present multi-view anchor graph hashing (MVAGH) to extend anchor graph hashing
(AGH) \cite{HASHGRAPH} in multiple views (visual modalities). The hashing codes are determined as the subset
of eigenvectors calculated from an averaged similarity matrix induced by multiple anchor graphs.
Song \emph{et al}. \cite{HASHMFTWO} develop multiple feature hashing (MFH). The learning process preserves
the local structure information of each individual feature, and simultaneously considers the global
alignment of local structures for all the features. By using the learned hashing hyper-plane,
MFH first concatenates all the features into a single vector and then projects it into binary codes.
Liu \emph{et al}. \cite{HASHMFONE} propose compact kernel hashing (CKH). It formulates
the similarity preserving hashing with linearly combined multiple kernels corresponding to different features.
More recently, multi-view latent hashing (MVLH) \cite{MVLH} is proposed to incorporate
multi-modal features in binary hashing learning by discovering the latent factors
shared by multiple modalities. Multi-view alignment hashing (MVAH) is presented in \cite{TIPMVAH} to combine
multiple image features for learning effective hashing functions based on the regularized kernel nonnegative matrix factorization.
Distinguished from the hashing methods presented above, CV-DMH learns multi-modal hashing codes on canonical views
by capturing the key characteristics of landmarks. With this design, the generated binary codes
can enjoy desirable robustness on query accommodation. Moreover, CV-DMH directly solves discrete hashing codes. It can effectively avoid the accumulated quantization errors in existing multi-modal hashing methods adopting simplified ``relaxing+rounding" optimization framework.
\subsection{Discrete Hashing Learning}
Most existing hashing methods simply apply two-step ``relaxing+rounding" optimization framework to solve hashing codes. However, as indicated by recent literature \cite{DGH,SDH,luo2017robust}, this simple relaxing will bring significant information quantization loss. To alleviate this problem, several approaches are proposed to directly deal with the discrete optimization challenge. Discrete graph hashing (DGH) \cite{DGH} reformulates the graph hashing with a discrete optimization framework and solves the problem with a tractable alternating maximization algorithm. Supervised discrete hashing (SDH) \cite{SDH} learns discrete hashing codes via supervised learning. Cyclic coordinate descent is applied to calculate discrete hashing bits in a closed form. Coordinate discrete hashing (CDH) \cite{CDOECVR} is designed for cross-modal hashing \cite{xu2017learning}, and the discrete optimization proceeds in a block coordinate descent manner. Column sampling based discrete supervised hashing (COSDISH) is proposed in \cite{CSBDSH} to learn discrete hashing codes from semantic information by column sampling. Discrete proximal linearized minimization (DPLM) is presented in \cite{TIP2016binary} to solve discrete hashing codes. It reformulates the hashing learning as minimizing the sum of a smooth loss term with a nonsmooth indicator function. The problem is finally solved by an iterative procedure with each iteration admitting an analytical discrete solution. Kernel-based supervised discrete hashing (KSDH) \cite{Shi2016} solves discrete hashing codes via asymmetric relaxation strategy. It relaxes the hashing function into a general binary code matrix which is calculated within an alternative strategy. Although these approaches achieve certain success, their proposed discrete optimization solutions are specially designed for uni-modal hashing and particular hashing types. Hence, they cannot be easily generalized to other hashing learning formulations.
Our work is an advocate of discrete hashing optimization but focuses on learning robust hashing codes for mobile landmark search. We integrate the discriminative information from multi-modal features directly into discrete hashing codes based on the discovered informative canonical views. Moreover, our proposed discrete optimization strategy can not only explicitly deal with the discrete constraint of binary codes, but also consider the bit-uncorrelated constraint and balance constraint together\footnote{DPLM can cope with uncorrelation and balance constraints. However, it simply transfers two constraints to the objective function and tries to optimize a relaxed equivalent problem.}. To the best of our knowledge, there is still no similar work.
\section{System Overview}
\label{sec:so}
This section briefly introduces system overview of the proposed CV-DMH based mobile landmark search system. As shown in the Fig. \ref{fig:sys}, the system is mainly comprised of two key components: offline learning and online searching.
\begin{itemize}
\itemsep=3pt
\item \emph{Offline Learning}: The aim of this part is to learn hashing functions which can project high-dimensional multi-modal features of both query and database images into binary hashing codes. Specifically, offline learning is further divided into four subsequent sub-processes: feature extraction, multi-modal canonical view mining, intermediate representation generation, and discrete binary embedding. In the system, multi-modal features are first extracted from heterogeneous visual modalities to represent images. Then, multi-modal canonical view mining is proposed in order to efficiently discover a compact but informative canonical view set from noisy landmark image collections to capture key visual appearances of landmarks. Next, in order to robustly model diverse visual contents, an intermediate representation is generated by computing multi-modal sparse reconstruction coefficients between image and canonical views. Finally, compact binary codes of database images and hashing functions for online queries are learned by formulating a unified discrete binary embedding model.
\item \emph{Online Searching}: Query image is first submitted by user from mobile devices. Multi-modal visual features of it are then extracted on image and transformed to an intermediate image representation with the same pipeline conducted on database images. Next, binary codes of query are generated with the hashing functions learned from offline learning. Finally, the Hamming distances are computed with simple bit operations and ranked in ascending order, and their corresponding landmark images are returned.
\end{itemize}
\section{The Proposed CV-DMH}
\label{sec:3}
This section provides the details of the proposed CV-DMH. First, we introduce multi-modal canonical view mining. Second, we give details of intermediate representation generation. Third, we formulate the discrete binary embedding model and give an efficient discrete optimization solution. Finally, we summarize the key steps of CV-DMH and give a computation complexity analysis.
\subsection{Multi-modal Canonical View Mining}
\label{sec:3:2}
Landmark images have an interesting characteristic that is different from general images. Landmark is located on specific venue or attraction. On the one hand, in real practice, only the spectacular and attractive views of a landmark will be photographed by various tourists spontaneously. On the other hand, many of them would like to share their recorded images in social websites (e.g. Flickr). These accumulated canonical views of landmarks in social websites coincidentally reflect the common preferences of tourists. On the perspective of technique, they capture the key visual characteristics of landmarks, and an arbitrary query image recorded by tourists can be characterized with several canonical views of landmark. Therefore, it is promising to leverage canonical views of landmark as hashing learning bases to robustly accommodate visual variations of query images captured by mobile devices. Principally, optimal canonical views should possess two important properties: 1) Representative. The canonical views should capture the key visual landmark components from large quantities of recorded landmark images. 2) Compact. Redundant views will bring noises and increase extra computation burden.
Motivated by the above analysis, we propose an efficient submodular function based mining algorithm, measuring the representativeness and compactness of view set, to iteratively discover canonical views. Specifically, we first quantitatively define the above two properties of canonical view set\footnote{Canonical view set is comprised of canonical views.}. Then, a submodular function is designed accordingly and an efficient iterative mining approach with theoretical guarantee is developed for canonical view discovery.
\begin{definition}
Let $\mathcal I$ denote image space, $\mathcal I= \{\mathcal I_n\}_{n=1}^N$, $N$ is the number of database images.
Let $\mathcal L$ denote landmark space, $\mathcal L= \{\mathcal L_m\}_{m=1}^M$, $M$ is the number of landmarks in database.
$\mathcal L_m$ is defined as a set of images which are recorded at the nearby positions of the $m_{th}$ landmark.
Let $\mathcal V$ denote a view set of $\mathcal L$. It is defined as a set of images $\{v_i\}_{i=1}^{|\mathcal V|}$ belonging to $\mathcal I$, $\mathcal V \subseteq \mathcal I$, $|\mathcal V|\ll |\mathcal I|$.
\end{definition}
\begin{definition}
Let $\texttt{Rep}(\mathcal V)$
denote the visual representativeness of view set $\mathcal V$ over $\mathcal L$.
It is defined as $\texttt{Rep}(\mathcal V)=\sum_{v_i\in \mathcal V}\texttt{Rep}(v_i)=\sum_{v_i\in \mathcal V} \sum_{v_j\in \mathcal I, i\neq j}g_{ij}$, $g$ is the function which measures the
feature similarity of two image views, $g_{ij}$ is short for $g(v_i,v_j)$.
Let $\texttt{Red}(\mathcal V)$ denote the visual redundancy of view set $\mathcal V$.
It is defined as $\texttt{Red}(\mathcal V)=\sum_{v_i\in V}
\texttt{Red}(v_i)=\sum_{v_i,v_j\in \mathcal V, i\neq j} g_{ij}$.
\end{definition}
\begin{definition}
Let $\mathcal C$ denote the canonical view set of $\mathcal L$.
The views involved in $\mathcal C$ can comprehensively represent
diverse visual contents of landmark, and meanwhile, have less visual redundancy. In this paper, it is defined as
$\mathcal C=\arg\max_{\mathcal V\subseteq \mathcal I, |\mathcal V|=T} h(\mathcal V), h(\mathcal V)=\texttt{Rep}(\mathcal V)- \texttt{Red}(\mathcal V)$, $T$ is cardinality of canonical view set.
\end{definition}
\begin{lemma}
$h(\mathcal V)$ is submodular function. That is, $\forall \mathcal V_1\subseteq \mathcal V_2\subseteq \mathcal V, \forall v_j\notin \mathcal V$, $h(\mathcal V_1\cup v_j)-h(\mathcal V_1)\geq h(\mathcal V_2\cup v_j)-h(\mathcal V_2)$.
\end{lemma}
\vspace{-4mm}
\begin{equation*}
\begin{aligned}
\small
& \emph{Proof} \quad \texttt{Rep}(\mathcal V_1\cup v_j)-\texttt{Rep}(\mathcal V_1) =\texttt{Rep}(\mathcal V_2\cup v_j)- \\
&\texttt{Rep}(\mathcal V_2)-(\texttt{Red}(\mathcal V_1\cup v_j)-\texttt{Red}(\mathcal V_1))=-2\sum_{v_i\in \mathcal V_1 \setminus v_j} g_{ij}\\
&\geq -2\sum_{v_i\in \mathcal V_2 \setminus v_j} g_{ij} = -(\texttt{Red}(\mathcal V_2\cup v_j)-\texttt{Red}(\mathcal V_2)) \\
&\Rightarrow \ h(\mathcal V_1\cup v_j)-h(\mathcal V_1)\geq h(\mathcal V_2\cup v_j)-h(\mathcal V_2)
\end{aligned}
\end{equation*}
\begin{lemma}
$h(\mathcal V)$ is monotonically nondecreasing function. That is, $\forall \mathcal V_1\subseteq \mathcal V_2\subseteq \mathcal V$, $h(\mathcal V_1)\leq h(\mathcal V_2)$.
\end{lemma}
\vspace{-4mm}
\begin{equation*}
\small
\begin{aligned}
&\emph{Proof} \quad h(\mathcal V)=\sum_{v_i\in \mathcal V}\sum_{v_j\in \mathcal I \setminus \mathcal V, i\neq j} g_{ij} \\
& \Rightarrow \ h(\mathcal V_1)=\sum_{v_i\in \mathcal V_1}\sum_{v_j\in \mathcal V_2 \setminus \mathcal V_1,i\neq j} g_{ij}+ \sum_{v_i\in \mathcal V_1}\sum_{v_j\in \mathcal I \setminus \mathcal V_2,i\neq j} g_{ij}\\
& \Rightarrow \ h(\mathcal V_2)=\sum_{v_i\in \mathcal V_2\setminus\mathcal V_1} \sum_{v_j\in \mathcal I \mathcal \setminus V_2, i\neq j} g_{ij} + \sum_{v_i\in \mathcal V_1}\sum_{v_j\in \mathcal I \setminus \mathcal V_2,i\neq j} g_{ij}\\
&Since \ in \ our \ case, \ |\mathcal V_1|, |\mathcal V_2| \ll \mathcal I \\
& \Rightarrow \ \sum_{v_i\in \mathcal V_1}\sum_{v_j\in \mathcal V_2 \setminus \mathcal V_1,i\neq j} g_{ij}\leq \sum_{v_i\in \mathcal V_2\setminus\mathcal V_1} \sum_{v_j\in \mathcal I \setminus \mathcal V_2, i\neq j} g_{ij}\\
& \Rightarrow \ h(\mathcal V_1)\leq h(\mathcal V_2)
\end{aligned}
\end{equation*}
As indicated in \textbf{Definition 3}, to discover optimal canonical
view set, the function $h(\mathcal V)$ should be maximized in theory. However,
since $h(\mathcal V)$ is submodular function, the maximization of it
is a NP-complete optimization problem. Fortunately, $h(\mathcal V)$ is monotonically nondecreasing
with a cardinality constraint as indicated by two above Lemmas. Canonical views can be discovered near optimally
by greedy strategy as following steps\vspace{-3mm}\newline
\begin{algorithm}
\caption{Canonical view mining}
\label{cvm}
\begin{algorithmic}[1]
\STATE Extract visual feature for all landmark images. \\
\STATE Set canonical view set as empty, $\mathcal C=\emptyset$.\\
\FOR{$t=1$ to $T$}
\STATE Compute $\texttt{diff}(\mathcal I_{n})=h(\mathcal C \cup \mathcal I_{n})-h(\mathcal I_{n})$
for each landmark image $\mathcal I_{n}\in \mathcal I$. \\
\STATE Select the image with the maximum $\texttt{diff}$ into $\mathcal C$, $\mathcal I^*=\arg\max_{I_n\in \mathcal I}\texttt{diff}(I_n)$, and
simultaneously remove it from $\mathcal I$, $\mathcal C \leftarrow \mathcal C \cup \mathcal I^*$, $\mathcal I \leftarrow \mathcal I \setminus \mathcal I^*$.\\
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\cite{AAAItheorem} Let $\mathcal S^*$ denote the global optimal solution that solves the combinatorial optimization problem $\arg\max_{\mathcal S\subseteq \mathcal L, |\mathcal S|=T} h(\mathcal S)$, $\mathcal S$ denote approximate solution found by the greedy algorithm. If $h(\mathcal S)$ is nondecreasing
submodular function with $h(\emptyset)=0$, we can have
\begin{equation*}
\begin{aligned}
\small
h(\mathcal S)\geq h(\mathcal S^*)\frac{\zeta-1}{\zeta}
\end{aligned}
\end{equation*}
where $\zeta$ refers to the natural exponential.
\end{theorem}
As validated by \textbf{Theorem 1}, this greedy algorithm can achieve a result that is no worse than a constant
fraction $\frac{\zeta-1}{\zeta}$ away from the optimal value. The time complexity of canonical view mining is reduced to $O(NT)$.
Hence, the canonical view discovery process can be completed efficiently. Different modalities generally include complete information \cite{7565615,DBLP:conf/cvpr/ChangYYX16,yang2014exploiting,Cheng2016,ChangMLYH17,7442559}. Canonical views in different modalities may be different.
To comprehensively cover visual appearances of landmarks,
canonical view mining is performed in multiple modalities,
obtaining canonical view set $\{\mathcal C^p\}_{p=1}^P$, $P$ is number of modalities.
We concatenate features of canonical views and
construct a matrix $E^p=[e_1^p, ..., e_{T}^p]\in \mathbb{R}^{d_p\times T}$
in modality $p$, $d_p$ is the corresponding feature dimension.
\begin{figure}
\centering
\includegraphics[width=85mm]{sample.pdf}
\caption{Visual contents of an arbitrary query landmark image can be sparsely represented with its relations to canonical views.}
\label{fig:sample}
\vspace{-4mm}
\end{figure}
\subsection{Intermediate Representation Generation}
\label{sec:3:3}
With the discovered canonical views, we generate an intermediate representation for subsequent binary embedding.
As illustrated above, an arbitrary recorded landmark image
either describes a certain canonical view
or the cross-scenery among several particular canonical views.
In both cases, visual contents of the image can be sparsely represented with
its relations to several particular canonical views (as shown in Fig. \ref{fig:sample}). Sparse coding performs well on
robust representation with fix visual bases. In this paper, we leverage it for intermediate representation generation. Specifically, we calculate multi-modal sparse reconstruction coefficients
between image and canonical views, and the auto-generated response coefficients
are determined as the dimensions of intermediate representation. Principally, the intermediate representation can effectively characterize diverse visual
contents by adaptively adjusting the response coefficients on canonical views. Hence, it can construct a robust
foundation for subsequent discrete binary embedding. Mathematically, the concrete computation form is
\begin{equation}
\small
\begin{aligned}
\label{eq:afm}
\min_{\{Y^p\}_{p=1}^P} \ & \sum_{p=1}^P ||X^p-E^pY^p||_F^2 + \sigma\sum_{p=1}^P\sum_{n=1}^N||d_n^p\otimes y_n^p||_F^2 \\
s.t. \ & 1_T^\texttt{T}y_n^p=1, d_n^p=\texttt{exp}(\frac{\texttt{dist}(x_n^p, E^p)}{\rho}), \forall p, n \\
\end{aligned}
\end{equation}
\noindent where the first term measures the sparse reconstruction errors from raw features to exemplars, the second term is locality adaptor that ensures the generated sparse codes are proportional to the similarities between their corresponding raw features and exemplars. $\sigma>0$ is a constant factor that adjusts the balance between terms, $\rho$ is set to be the mean of pairwise distances, $\otimes$ denotes the element-wise product, $1_T\in \mathbb{R}^\texttt{T}$ denotes a column vector with all ones. $X^p=[x_1^p,..., x_N^p]\in \mathbb{R}^{d_p\times N}$ denotes
features of database images in modality $p$. $Y^p=[y_1^p,...,y_N^p]\in \mathbb{R}^{T\times N}$
denotes modality-specific canonical view based intermediate representation. Each column
has $r$ non-zeros coding coefficients.
$d_n^p$ is a vector that measures the distances between raw feature and exemplars,
$\texttt{dist}(x_n^p, E^p)= [\texttt{dist}(x_n^p, e_1^p), ..., \texttt{dist}(x_n^p, e_{T}^p)]$, $\texttt{dist}(x_n^p, e_1^p)$ is Euclidean distance between
$x_n^p$ and $e_{T}^p$.
The problem in (\ref{eq:afm}) can be efficiently solved by using the alternating direction method of multipliers (ADMM) \cite{PAMISSC}. After solving it, we concatenate the calculated $Y^p$ and construct dimensions of intermediate representation
\begin{equation}
\label{calculatey}
\small
\begin{aligned}
Y=[Y^1; ... ; Y^P]\in \mathbb{R}^{TP\times N}
\end{aligned}
\end{equation}
\subsection{Discrete Binary Embedding Model}
\label{sec:3:4}
Based on the intermediate representation, we design a discrete binary embedding model to
learn final binary hashing codes. Let us define $V\in [-1, 1]^{c\times N}$ as the hashing codes of database images,
$c$ is the hashing code length. Due to approximate canonical view mining, the intermediate representation
generation inevitably brings about noises from inaccurate canonical views.
In addition, as a result of information integration
from multiple modalities, there exist information redundancies among dimensions of
intermediate representation calculated from different modalities. Therefore, it is very important to remove the involved noises and redundancies
during the binary embedding. To achieve this goal, we propose latent structure learning
with matrix factorization to extract orthogonal Hamming space. In this space, each dimension corresponds to one hashing bit. Besides, we construct a canonical view based graph \cite{7547296,JGLVS,GWC} to measure visual relationships among images captured by canonical views. That is, if two landmark images have similar visual distributions on intermediate representations, they are forced to be projected to close points in hamming space. Moreover, we learn linear projection based hashing functions to support queries that are out of the database.
By integrating the aforementioned considerations, we derive the overall discrete binary embedding formulation as
\begin{equation}
\small
\label{disobj}
\begin{aligned}
\min_{W, U, V} \ & ||Y-UV||_F^2 + \alpha Tr(VLV^{\texttt{T}}) + \beta(||V- W^{\texttt{T}}Y||_F^2+\gamma ||W||_F^2) \\
s.t. \ & V\in \{-1,1\}^{c\times N}, VV^{\texttt{T}}=NI_c, V1_N=0
\end{aligned}
\end{equation}
\noindent where $\alpha, \beta, \gamma>0$ adjust the balance of terms. $V\in [-1, 1]^{c\times N}$ is discrete constraint on hashing codes.
In addition, we also consider the bit-uncorrelated constraint $VV^{\texttt{T}}=NI_c$ which
guarantees that the learned Hamming space to be orthogonal for information redundancy removal, and bit balance constraint $V1_N=0$
which forces each bit to have equal chance to occur in the whole image database. $||Y-UV||_F^2$
is for latent structure learning, $U$ is mapping between latent structure and intermediate representation.
$Tr(VLV^{\texttt{T}})$ is graph regularizer which
preserves visual relations of landmark images measured by canonical views.
$Tr(\cdot)$ is trace operation, $L$ is Laplacian matrix calculated on canonical view based graph, it is
constructed based on intermediate representation, to measure the visual relations of images
on canonical views. $||V-W^{\texttt{T}}Y||_F^2+\gamma ||W||_F^2$ learns
linear projection matrix $W\in \mathbb{R}^{TP\times c}$ equipped in hashing functions,
$||V-W^{\texttt{T}}Y||_F^2$ is to reduce the loss between
binary codes and the projected values. It is worth noting that,
as linear projection is leveraged, the online mobile landmark search process is efficient.
The optimal $W$ that solves Eq.(\ref{disobj}) can be expressed in terms of $Y$. We can derive the following theorem.
\begin{theorem}
Let $W, U, V$ and $Y$ be defined as before. Then the optimal $W$ that solves learning problem in (\ref{disobj}) is given by
$W=(YY^{\texttt{T}}+\gamma I)^{-1}YV^{\texttt{T}}$.
The minimum problem in (\ref{disobj}) is equivalent to the following problem
\begin{equation}
\small
\label{obj:1}
\begin{aligned}
\min_{U, V} \ & ||Y-UV||_F^2+\alpha Tr(VAV^{\texttt{T}}) \\
s.t. \ & V\in \{-1,1\}^{c\times N}, VV^{\texttt{T}}=NI_c, V1_N=0
\end{aligned}
\end{equation}
\noindent where $A=L+\frac{\beta}{\alpha}(I-Y^{\texttt{T}}QY), Q=(YY^{\texttt{T}}+\gamma I)^{-1}$.
\end{theorem}
\begin{proof}
By calculating the derivation of
the objective function in Eq.(\ref{disobj}) \emph{w.r.t} $W$ and setting it to be zero, we can have
\begin{equation}
\small
\label{eq:3:1}
\begin{aligned}
W=(YY^{\texttt{T}}+\gamma I)^{-1}YV^{\texttt{T}} \\
\end{aligned}
\end{equation}
Let $Q=(YY^{\texttt{T}}+\gamma I)^{-1}$, then $W=QYV^{\texttt{T}}$.
By replacing $W$ into Eq.(\ref{disobj}), we can derive that
\begin{equation*}
\small
\label{formuM}
\begin{aligned}
||V-W^{\texttt{T}}Y||_F^2+\gamma ||W||_F^2= Tr(V(I_N-Y^{\texttt{T}}QY)V^{\texttt{T}})\\
\end{aligned}
\end{equation*}
By summing three terms together, we derive that
\begin{equation*}
\small
\begin{aligned}
&\ ||Y-UV||_F^2 + \alpha Tr(VLV^{\texttt{T}})+ \beta Tr(V(I-Y^{\texttt{T}} \\
&(YY^{\texttt{T}}+\gamma I)^{-1}Y)V^{\texttt{T}}) = ||Y-UV||_F^2+\alpha Tr(VAV^{\texttt{T}})\\
\end{aligned}
\end{equation*}
\noindent where $A=L+\frac{\beta}{\alpha}(I-Y^{\texttt{T}}QY), Q=(YY^{\texttt{T}}+\gamma I)^{-1}$. This completes the proof of the theorem.
\end{proof}
With \textbf{Theorem 2}, the problem in (\ref{disobj}) is transformed to
\begin{equation}
\small
\begin{aligned}
\label{disretobject}
&\min_{U, V} ||Y-UV||_F^2+\alpha Tr(VAV^{\texttt{T}})\\
s.t. \ & V\in \{-1,1\}^{c\times N}, VV^{\texttt{T}}=NI_c, V1_N=0
\end{aligned}
\end{equation}
The above objective function can be abstracted as graph-based matrix factorization \cite{TPAMINMF}. The main difference between our formulation and \cite{TPAMINMF} is the constraints imposed on hashing codes. These three constraints bring new challenges to the optimization process which have not yet been touched by existing graph-based matrix factorization methods.
\subsection{Discrete Solution}
\label{sec:3:5}
Essentially, solving problem (\ref{disretobject}) is a challenging combinatorial optimization problem due to three constraints. Most existing hashing approaches apply ``relaxing+rounding" optimization \cite{hashingsurvey}. They first relax three constraints to calculate continuous values, and then binarize them to hashing codes via rounding. This two-step learning can actually simplify the solving process, but it may cause significant information loss \cite{SDH,DGH}. In recent literature, several discrete hashing solutions are proposed. However, they are developed for particular hashing types. For example, graph hashing \cite{DGH}, supervised hashing \cite{SDH,CSBDSH,KSDH}, cross-modal hashing \cite{CDOECVR}. Therefore, their designed learning strategies cannot be directly applied to solve our problem.
In this paper, we propose an effective optimization algorithm based on augmented Lagrangian multiplier (ALM) \cite{ALMone} to calculate the discrete solution within one step. Our method can not only explicitly deal with the discrete constraint, but also consider the bit-uncorrelated constraint and balance constraint together. Our key idea is adding auxiliary variables $\Gamma$, $\Theta$ to separate three challenging constraints, and transform the objective function to an equivalent one. The substituted variables and corresponding auxiliary ones are forced to be close with each other. Specifically, in this paper, we set $\Gamma=Y-UV, \Theta=V$. Problem (\ref{disretobject}) is transformed as
\begin{equation}
\small
\begin{aligned}
\label{eq:tj}
& \min_{U, V, \Gamma, \Theta} \ ||\Gamma||_F+ \frac{\eta}{2} ||Y-UV-\Gamma+\frac{E_{\eta}}{\eta}||_F^2 \\
& \alpha Tr(VA\Theta^{\texttt{T}}) + \frac{\mu}{2} ||V-\Theta+\frac{E_{\mu}}{\mu}||_F^2 \\
& s.t. \quad V\in \{-1,1\}^{c\times N}, \Theta\Theta^\texttt{T}=NI_c, \Theta1_N=0
\end{aligned}
\end{equation}
\noindent where $E_{\eta}, E_{\mu}$ measure the difference between the target and auxiliary variables, $\eta, \mu>0$ adjust the balance between terms.
We adopt alternate optimization to iteratively solve problem (\ref{eq:tj}). We optimize the objective function with respective to one variable while fixing other remaining variables. The key steps for solving $Z$ are summarized in Algorithm \ref{calv}.
\begin{algorithm}
\caption{Obtaining $V$ via solving problem (\ref{eq:tj})}
\label{calv}
\begin{algorithmic}[1]
\STATE Initialize $E_{\eta}$, $E_{\mu}$, $V$; \\
\WHILE{not convergence}
\STATE Optimize $\Gamma$ while fixing the others with Eq.(\ref{eq:tl});
\STATE Optimize $U$ while fixing the others with Eq.(\ref{eq:tp});
\STATE Optimize $\Theta$ while fixing the others with Eq.(\ref{eq:optitheta});
\STATE Optimize $V$ while fixing the others with Eq.(\ref{eq:tz});
\STATE Update $E_{\eta}$, $E_{\mu}$, $\eta$, $\mu$ with Eq.(\ref{eq:tu}); \\
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\textbf{Step 3: Update $\Gamma$.} By fixing other variables, the optimization formulas for $\Gamma$ are
\begin{equation}
\small
\begin{aligned}
\label{eq:tk}
&\min_{\Gamma} \ ||\Gamma||_F+ \frac{\eta}{2} ||Y-UV-\Gamma+\frac{E_{\eta}}{\eta}||_F^2)
\end{aligned}
\end{equation}
\noindent By calculating the derivative of the objective function with respective to $\Gamma$, and setting it to 0, we can obtain that
\begin{equation}
\small
\begin{aligned}
\label{eq:tl}
\Gamma=\frac{\eta Y-\eta UV+E_{\eta}}{2+\eta}
\end{aligned}
\end{equation}
\textbf{Step 4: Update $U$}. The optimization formula for $U$ is
\begin{equation}
\small
\begin{aligned}
\label{eq:tn}
\min_{U} \ & ||Y-UV-\Gamma + \frac{E_{\eta}}{\eta}||_F^2 \\
\end{aligned}
\end{equation}
\noindent By calculating the derivative of the objective function with respective to $U$, and setting it to 0, we can obtain that
\begin{equation}
\small
\begin{aligned}
UV = Y-\Gamma+\frac{E_{\eta}}{\eta} \\
\end{aligned}
\end{equation}
\noindent Since $VV^\texttt{T}=NI_c$, we can further derive that
\begin{equation}
\small
\begin{aligned}
\label{eq:tp}
U = \frac{1}{N}(Y-\Gamma+\frac{E_{\eta}}{\eta})V^\texttt{T}
\end{aligned}
\end{equation}
\textbf{Step 5: Update $\Theta$}. The optimization formula for $\Theta$ is
\begin{equation}
\small
\begin{aligned}
\label{eq:tq}
\min_{\Theta} \ & \alpha Tr(VA\Theta^\texttt{T})+\frac{\mu}{2}||V-\Theta+\frac{E_{\mu}}{\mu}||_F^2 \\
& s.t. \ \Theta\Theta^\texttt{T}=NI_c, \Theta1_N=0
\end{aligned}
\end{equation}
\noindent The objective function in Eq.(\ref{eq:tq}) can be simplified as
\begin{equation}
\small
\begin{aligned}
\label{eq:tre}
\min_{\Theta} & \ ||\Theta-(V+\frac{E_{\mu}}{\mu}-\frac{\alpha}{\mu}VA)||_F^2\\
\end{aligned}
\end{equation}
\noindent where $C=V+\frac{E_{\mu}}{\mu}-\frac{\alpha}{\mu}VA$.
The above equation is equivalent to the following maximization problem
\begin{equation}
\small
\begin{aligned}
\label{eq:ts}
\max_{\Theta} \ & Tr(\Theta^\texttt{T}C) \quad s.t. \ \Theta\Theta^\texttt{T}=NI_c, \Theta1_N=0
\end{aligned}
\end{equation}
Mathematically, with singular value decomposition (SVD), $C$ can be decomposed as $C=P\Lambda Q^\texttt{T}$
, where the columns of $P$ and $Q$ are left-singular vectors and right-singular vectors of $C$ respectively, $\Lambda$ is rectangular diagonal matrix and its diagonal entries are singular values of $C$. Then, we can derive that $\max_{V} \ Tr(\Theta^\texttt{T}P\Lambda Q^\texttt{T}) \Leftrightarrow \max_{V} \ Tr(\Lambda Q^\texttt{T}\Theta^\texttt{T}P)$.
\begin{theorem}\label{calzthre}
Given any matrix $G$ which meets $GG^\texttt{T}=NI$ and diagonal matrix $\Lambda \ge 0$, the solution of $\max_{G} Tr(\Lambda G)$ is $\texttt{diag}(\sqrt{N})$.
\end{theorem}
\begin{proof}
Let us assume $\lambda_{ii}$ and $g_{ii}$ are the $i_{th}$ diagonal entry of $\Lambda$ and $G$ respectively, $Tr(\Lambda G)=\sum_{i}\lambda_{ii}g_{ii}$. Since $GG^\texttt{T}=NI$, $g_{ii}\leq \sqrt{N}$. $Tr(\Lambda G)=\sum_{i}\lambda_{ii}g_{ii}\leq \sqrt{N}\sum_{i}\lambda_{ii}$. The equality holds only when $g_{ii}=\sqrt{N}, g_{ij}=0, \forall i,j$. $Tr(\Lambda G)$ achieves its maximum when $G=\texttt{diag}(\sqrt{N})$.
\end{proof}
As $\Lambda$ is calculated by SVD, we can obtain that $\Lambda \ge 0$. On other hand, we can easily derive that $Q^\texttt{T}\Theta^\texttt{T}PP^\texttt{T}\Theta Q=NI$. Therefore, according to the \textbf{Theorem \ref{calzthre}}, the optimal $\Theta$ can only be obtained when $Q^\texttt{T}\Theta^\texttt{T}P=\texttt{diag}(\sqrt{N})$. Hence, the solution of $\Theta$ is
\begin{equation}
\small
\begin{aligned}
\label{eq:solutiony}
\Theta=\sqrt{N}PQ^\texttt{T}
\end{aligned}
\end{equation}
Moreover, in order to satisfy the balance constraint $\Theta1_N=0$, we apply Gram-Schmidt process as \cite{DGH} to construct matrices $\hat{P}$ and $\hat{Q}$, so that $\hat{P}^\texttt{T}\hat{P}=I_{L-R}$, $[P, 1]^\texttt{T}\hat{P}=0$, $\hat{Q}^\texttt{T}\hat{Q}=I_{L-R}$, $Q\hat{Q}^\texttt{T}=0$, $R$ is the rank of $C$. The close form solution for $\Theta$ is
\begin{equation}
\small
\begin{aligned}
\label{eq:optitheta}
\Theta=\sqrt{N}[P, \hat{P}][Q, \hat{Q}]^\texttt{T}
\end{aligned}
\end{equation}
\textbf{Step 6: Update $V$}. By fixing other variables, the optimization formula for $V$ is
\begin{equation}
\small
\begin{aligned}
\label{eq:tr}
\min_{V\in [-1, 1]^{c\times N}} \ & \frac{\eta}{2}||Y-UV-\Gamma+\frac{E_{\eta}}{\eta}||_F^2 +\alpha Tr(VA\Theta^{\texttt{T}})\\
& +\frac{\mu}{2} ||V-\Theta+\frac{E_{\mu}}{\mu}||_F^2
\end{aligned}
\end{equation}
\noindent The above problem can be transformed as
\begin{equation}
\small
\begin{aligned}
\label{eq:trrr}
\min_{V\in [-1, 1]^{c\times N}} & \ ||V-(\Theta-\frac{E_{\mu}}{\mu}-\frac{\alpha}{\mu}\Theta A + \frac{\eta}{\mu} U^\texttt{T}(Y-\Gamma+\frac{E_{\eta}}{\eta}))||_F^2
\end{aligned}
\end{equation}
The discrete solution of $V$ can be directly represented as
\begin{equation}
\small
\begin{aligned}
\label{eq:tz}
V=\texttt{Sgn}(\Theta-\frac{E_{\mu}}{\mu}-\frac{\alpha}{\mu}\Theta A + \frac{\eta}{\mu} U^\texttt{T}(Y-\Gamma+\frac{E_{\eta}}{\eta}))
\end{aligned}
\end{equation}
\noindent where $\texttt{Sgn}(\cdot)$ is signum function which returns -1 if $x<0$, and 1 if $x\geq 0$.
\textbf{Step 7: Update $E_{\eta}$, $E_{\mu}$, $\eta$, $\mu$}. The update rules are
\begin{equation}
\small
\begin{aligned}
\label{eq:tu}
&E_{\eta}=E_{\eta}+\eta(Y-UV-\Gamma)\\
&E_{\mu}=E_{\mu}+\mu(V-\Theta)\\
&\eta=\rho\eta, \mu=\rho\mu\\
\end{aligned}
\end{equation}
$\rho>1$ is learning rate that controls the convergence.
\textbf{Hashing Function Learning.} After obtaining $V$, we substitute the value into $W=(YY^{\texttt{T}}+\gamma I)^{-1}YV^{\texttt{T}}$ and get the linear projection matrix. The hashing functions can be represented as $H(x)=\texttt{sgn}(W^{\texttt{T}}x)$.
\textbf{Online Mobile Landmark Search.} Given a landmark query image $q$ from mobile device, we first extract intermediate representation $Y_q$ as Eq.(\ref{calculatey}). Its binary hashing codes are calculated as $V_q=\texttt{sgn}(W^{\texttt{T}}Y_q)$.
\begin{algorithm}
\caption{CV-DMH based mobile landmark search}
\label{alg:summary}
\begin{algorithmic}[1]
\REQUIRE ~~\\
Query image $q$. Database images $\mathcal I= \{\mathcal I_n\}_{n=1}^N$. \\
\ENSURE ~~\\
Discrete hashing codes of database images $V$. Hashing functions $H$. Image retrieval results for image query $q$.\\
\emph{\textbf{Offline Learning}}
\STATE Extract multi-modal canonical views as illustrated in \ref{sec:3:2}; \\
\STATE Generate intermediate landmark representation as \ref{sec:3:3}; \\
\STATE Compute hashing codes of database images $V$ by solving problem (\ref{disretobject}) with Algorithm \ref{calv}; \\
\STATE Construct hashing functions $H$ with the projection matrix; \\
\emph{\textbf{Online Searching}}
\STATE Extract multi-modal visual features of query image and transform them to a unified intermediate representation; \\
\STATE Project query into hashing codes with the learned $H$; \\
\STATE Perform searching in Hamming space and return results. \\
\end{algorithmic}
\end{algorithm}
\subsection{Summarization and Computational Complexity Analysis}
The key steps of CV-DMH based MLS are described in Algorithm \ref{alg:summary}.
It can be easily derived that the computation cost of multi-modal canonical view mining is $O(NPT)$, as there are $P$ visual modalities
and $T$ iterations for mining. The part of intermediate generation solves a constrained least square fitting problem which also consumes $O(NPT)$.
The computational complexity of discrete optimization is $O(\#iter(TPN+ TPc + cN))$, where $\#iter$ denotes the number of iterations in Algorithm \ref{calv}. Given $N\gg TP>c$, this process scales linearly with $N$. The computation of hashing functions solves a linear system, which consumes $O(N)$.
\section{Experimental Configuration}
\label{sec:4}
\subsection{Experimental Datasets and Setting}
In this paper, three real landmark datasets, \emph{Oxford5K} \cite{CVPRoxford}, \emph{Paris6K} \cite{parisdataset}, and \emph{Paris500K} \cite{iccvlandmark}\footnote{In this experiment, the maximum number of images in each category is limited to 2000 to avoid bias.}, are applied in empirical study. \emph{Oxford5K} is comprised of 5,062 images recorded for 17 landmarks in Oxford.
\emph{Paris6K} consists of 6,412 \emph{Paris} landmark images in 12 categories.
\emph{Paris500K} contains 41,673 images with clustering ground truth which describes
79 landmarks. For \emph{Oxford5K} and \emph{Paris6K}, 10\%, 20\%, and 70\% images
are used as query images, training images, and database images, respectively.
For \emph{Paris500K}, the corresponding ratios are 10\%, 10\%, and 80\%. For three datasets, both
query and database images appear with great visual diversity. Each image is represented by features in 5 heterogenous visual modalities: 81-D Color Moments (CM) \cite{CM}, 58-D Local Binary Pattern (LBP) \cite{LBP}, 80-D Edge Direction Histogram (EDH)
\cite{EDH}, 1,000-D BoVW\footnote{128-D SIFT is employed as local
descriptor.} \cite{bovw}, and 512-D GIST \cite{GIST}.
\subsection{Evaluation Metrics}
\label{sec:4:2}
In our experimental study, mean average precision (mAP) is adopted as the evaluation metric for effectiveness.
The metric has been widely used in literature \cite{surveyhashing}. For a given query, average precision (AP) is calculated as
\begin{equation}
\label{mAP}
\begin{aligned}
AP=\frac{1}{NR}\sum_{r=1}^{R} pre(r)rel(r)
\end{aligned}
\end{equation}
\noindent where $R$ is the total number of retrieved images, $NR$ is the number of relevant images in retrieved set, $pre(r)$ denotes the precision of top $r$ retrieval images, which is defined as the ration between the number of the relevant images and the number of retrieved images $r$, and $rel(r)$ is indicator function which equals to 1 if the $r_{th}$ image is relevant to query, and 0 vice versa. mAP is defined as the average of the AP of all queries. Larger mAP indicates the better retrieval performance. In experiments, we set $R$ as 100 to collect experimental results. Furthermore, \emph{Precision-Scope} curve is also reported to reflect the retrieval performance variations with respect to the number of retrieved images.
\begin{table*}
\small
\caption{mAP of all approaches on three datasets. The best performance in each column is marked with bold.}
\label{resulttable}
\centering
\begin{tabular}{|p{14mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}
|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}
|p{8mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{\emph{Oxford5K}} & \multicolumn{4}{c|}{\emph{Paris6K}} & \multicolumn{4}{c|}{\emph{Paris500K}}\\
\cline{2-13}
& 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 \\
\hline
AGH & \textbf{0.3164} & 0.2990 & 0.3098 & 0.3067 & 0.3301 & 0.3283 & 0.3524 & 0.3267 & 0.3427 & 0.3798 & 0.3872 & 0.3944 \\
\hline
ITQ & 0.2778 & 0.2800 & 0.2910 & 0.2917 & 0.2835 & 0.2942 & 0.3059 & 0.3275 & 0.2251 & 0.2765 & 0.3063 & 0.3621 \\
\hline
SGH & 0.2944 & 0.3000 & 0.3190 & 0.3315 & 0.3214 & 0.3497 & 0.3594 & 0.3793 & 0.3242 & 0.3718 & 0.4103 & 0.4620 \\
\hline
DPLM & 0.2985 & 0.3134 & 0.3112 & 0.3065 & 0.2706 & 0.2800 & 0.2962 & 0.3112 & 0.3044 & 0.3490 & 0.3663 & 0.3903 \\
\hline
CHMIS & 0.2959 & 0.3087 & 0.2991 & 0.3211 & 0.3256 & 0.3253 & 0.3265 & 0.3674 & 0.3937 & 0.4429 & 0.4681 & 0.5312 \\
\hline
MVAGH & 0.2947 & 0.3124 & 0.3105 & 0.3011 & 0.2646 & 0.2725 & 0.2882 & 0.3205 & 0.2907 & 0.3181 & 0.3403 & 0.3647 \\
\hline
MFH & 0.2718 & 0.2874 & 0.3027 & 0.3206 & 0.2897 & 0.3066 & 0.3108 & 0.3546 & 0.3537 & 0.4202 & 0.4461 & 0.5211 \\
\hline
CMKH & 0.2999 & 0.3034 & 0.2994 & 0.3026 & \textbf{0.3396} & 0.3351 & 0.3441 & 0.3481 & 0.4094 & 0.4680 & 0.5096 & 0.5493 \\
\hline
MVLH & 0.2918 & 0.3092 & 0.3051 & 0.3203 & 0.3152 & 0.3183 & 0.3531 & 0.3936 & 0.2965 & 0.3433 & 0.3570 & 0.4124 \\
\hline
MAVH & 0.2980 & 0.2856 & 0.2861 & 0.3021 & 0.2975 & 0.3188 & 0.3385 & 0.3565 & 0.2889 & 0.3177 & 0.3353 & 0.3876 \\
\hline
CV-DMH & 0.3028 & \textbf{0.3153} & \textbf{0.3275} & \textbf{0.3458} & 0.3350 & \textbf{0.3534} & \textbf{0.3799} & \textbf{0.4190} & \textbf{0.4533} & \textbf{0.5029} & \textbf{0.5293} & \textbf{0.6025}\\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\centering
\mbox{
\subfigure{\includegraphics[width=42mm]{topN32_oxford.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN48_oxford.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN64_oxford.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN128_oxford.pdf}}
}
\vspace{-3mm}
\caption{\emph{Precision-Scope} curves on \emph{Oxford5K} varying code length.}
\label{fig_oxford5k}
\vspace{-3mm}
\end{figure*}
\begin{figure*}
\centering
\mbox{
\subfigure{\includegraphics[width=42mm]{topN32_paris.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN48_paris.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN64_paris.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN128_paris.pdf}}
}
\vspace{-3mm}
\caption{\emph{Precision-Scope} curves on \emph{Paris6K} varying code length.}
\label{fig_paris6k}
\vspace{-3mm}
\end{figure*}
\begin{figure*}
\centering
\mbox{
\subfigure{\includegraphics[width=42mm]{topN32_mmparis.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN48_mmparis.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN64_mmparis.pdf}}
}\mbox{
\subfigure{\includegraphics[width=42mm]{topN128_mmparis.pdf}}
}
\vspace{-3mm}
\caption{\emph{Precision-Scope} curves on \emph{Paris500K} varying code length.}
\label{fig_mmaris}
\vspace{-3mm}
\end{figure*}
\subsection{Compared Approaches}
We compare CV-DMH with several state-of-the-art multi-modal hashing approaches.
They include\footnote{For CHMIS, MFH, and CMKH, implementation codes of
them are provided by the authors. For MVAGH, MVLH, and MVAH, we implement them according to the relevant
literature.}:
\begin{itemize}
\itemsep=3pt
\item Composite hashing with multiple information sources (CHMIS) \cite{HASHMFFOUR}. It is the first work to extend uni-modal spectral hashing to multiple modalities. CHMIS preserves visual similarity with graph and simultaneously integrates information from multiple modalities into the hashing codes with adjusted weights.
\item Multi-view anchor graph hashing (MVAGH) \cite{MVAGH}. It generates the nonlinearly integrated binary codes as a subset of eigenvectors calculated from an averaged similarity matrix.
\item Multiple feature hashing (MFH) \cite{HASHMFTWO}. MFH preserves the local structural information of each individual feature modality and also globally considers the local structures for all the features to learn hashing codes.
\item Compact multiple kernel hashing (CMKH) \cite{HASHMFONE}. CMKH formulates multi-modal hashing as a similarity preserving problem with linearly combined multiple kernels.
\item Multi-view latent hashing (MVLH) \cite{MVLH}. MVLH first transforms multi-modal features into a unified kernel feature space, then applies matrix factorization to learn the latent factors as the target binary codes.
\item Multi-view alignment hashing (MVAH) \cite{TIPMVAH}. MVAH first learns multi-modal fused projected vectors with a regularized kernel nonnegative matrix factorization. Then, hashing functions are learned via multivariable logistic regression.
\end{itemize}
Besides, we also compare CV-DMH with several state-of-the-art uni-modal hashing
approaches\footnote{Implementation codes of all these methods are downloaded directly from author websites}: AGH \cite{HASHGRAPH}, ITQ \cite{ITQ}, SGH \cite{SGH} and DPLM \cite{TIP2016binary}. For them, multi-modal features are concatenated into a unified vector
for subsequent learning. The involved parameters of the compared approaches are strictly
adjusted to report the maximum performance with the guidance of relevant literature.
\subsection{Implementation Details}
CV-DMH has three parameters: $\alpha, \beta$, and $\gamma$ in hashing objective function Eq.(\ref{disobj}). They are used to play the balance between the formulated regularization terms. In experiment, we choose the best parameters from $\{10^{-4}, 10^{-2}, 1, 10^2, 10^4\}$. The best performance of CV-DMH is achieved when $\{\alpha=10^{-2}, \beta=10^{-4}, \gamma=10^2\}$, $\{\alpha=10^{-2}, \beta=10^{-2}, \gamma=10^4\}$, $\{\alpha=10^{-2}, \beta=10^{-2}, \gamma=10^2\}$ on \emph{Oxford5K}, \emph{Paris6K}, and \emph{Paris500K} respectively. The parameters $\mu$ and $\eta$ in Eq.(\ref{eq:tj}) are set for ALM based discrete optimization. The best performance is obtained when $\{\mu=1, \eta=1\}$, $\{\mu=0.01, \eta=1\}$, $\{\mu=0.01, \eta=1\}$ on \emph{Oxford5K}, \emph{Paris6K}, and \emph{Paris500K} respectively. In addition, the best canonical view size $T$ is set to 100 on \emph{Oxford5K}
and \emph{Paris6K}, and 300 on \emph{Paris500K}.
The best number of nearest canonical views $r$ in Eq.(\ref{eq:afm}) is set to 70 on \emph{Oxford5K}
and \emph{Paris6K}, and 200 on \emph{Paris500K}. $\sigma$ in Eq.(\ref{eq:afm}) is set to $10^{-4}$ to maximize the performance.
In experiments, hashing code length on all datasets is
varied in the range of [32, 48, 64, 128] to observe the performance. Further,
The retrieval scope on three datasets is set from 100 to 1000 with step size 100. In the first step of Algorithm \ref{calv}, the initial values of $E_{\eta}, E_{\mu}$ are set to 0. The values of $V$ is calculated by PCAH \cite{PCAH}.
\begin{table*}
\small
\caption{Canonical views improve the robustness of CV-DMH. CV-DMH-I denotes direct multi-modal hashing without canonical views.}
\label{effcan}
\centering
\begin{tabular}{|p{16mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}
|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}
|p{8mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{\emph{Oxford5K}} & \multicolumn{4}{c|}{\emph{Paris6K}} & \multicolumn{4}{c|}{\emph{Paris500K}}\\
\cline{2-13}
& 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 \\
\hline
CV-DMH-I & 0.2855 & 0.3005 & 0.3053 & 0.3319 & 0.2905 & 0.3177 & 0.3225 & 0.3488 & 0.3558 & 0.4229 & 0.4509 & 0.5222\\
\hline
CV-DMH & \textbf{0.3028} & \textbf{0.3153} & \textbf{0.3275} & \textbf{0.3458} & \textbf{0.3350} & \textbf{0.3534} & \textbf{0.3799} & \textbf{0.4190} & \textbf{0.4533} & \textbf{0.5029} & \textbf{0.5293} & \textbf{0.6025}\\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\small
\caption{Effects of canonical view mining in multiple modalities.}
\label{feat}
\centering
\begin{tabular}{|p{14mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}
|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}
|p{8mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{\emph{Oxford5K}} & \multicolumn{4}{c|}{\emph{Paris6K}} & \multicolumn{4}{c|}{\emph{Paris500K}}\\
\cline{2-13}
& 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 \\
\hline
CM & 0.2131 & 0.2188 & 0.2192 & 0.2287 & 0.1954 & 0.1971 & 0.2043 & 0.2230 & 0.1511 & 0.1782 & 0.1945 & 0.2451\\
\hline
LBP & 0.2619 & 0.2695 & 0.2872 & 0.2950 & 0.2249 & 0.2284 & 0.2504 & 0.2582 & 0.2585 & 0.3107 & 0.3422 & 0.3899 \\
\hline
EDH & 0.2465 & 0.2658 & 0.2663 & 0.2645 & 0.2224 & 0.2325 & 0.2380 & 0.2620 & 0.2291 & 0.2821 & 0.3147 & 0.3965 \\
\hline
BOVW & 0.2932 & 0.3155 & 0.3022 & 0.3300 & 0.3060 & 0.3252 & 0.3380 & 0.3776 & 0.3255 & 0.3959 & 0.4381 & 0.5112 \\
\hline
GIST & 0.2589 & 0.2634 & 0.2675 & 0.2781 & 0.2517 & 0.2669 & 0.2797 & 0.3042 & 0.2528 & 0.3200 & 0.3745 & 0.4695 \\
\hline
CV-DMH & \textbf{0.3028} & \textbf{0.3153} & \textbf{0.3275} & \textbf{0.3458} & \textbf{0.3350} & \textbf{0.3534} & \textbf{0.3799} & \textbf{0.4190} & \textbf{0.4533} & \textbf{0.5029} & \textbf{0.5293} & \textbf{0.6025}\\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\small
\caption{Effects of submodular function based canonical view discovery.}
\label{effesumbo}
\centering
\begin{tabular}{|p{18mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}
|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}
|p{8mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{\emph{Oxford5K}} & \multicolumn{4}{c|}{\emph{Paris6K}} & \multicolumn{4}{c|}{\emph{Paris500K}}\\
\cline{2-13}
& 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 \\
\hline
\emph{Random} & 0.2740 & 0.2795 & 0.2727 & 0.3097 & 0.3254 & 0.3394 & 0.3529 & 0.3859 & 0.4028 & 0.4415 & 0.4659 & 0.5206 \\
\hline
\emph{K-means} & 0.2648 & 0.2877 & 0.2829 & 0.3199 & 0.3102 & 0.3267 & 0.3466 & 0.3879 & 0.3944 & 0.4418 & 0.4739 & 0.5363 \\
\hline
DL & 0.2523 & 0.2683 & 0.2699 & 0.2848 & 0.2615 & 0.2807 & 0.2822 & 0.3191 & 0.3306 & 0.3568 & 0.3770 & 0.4440 \\
\hline
R-CV-DMH & 0.2890 & 0.3036 & 0.3151 & 0.3434 & 0.3350 & \textbf{0.3627} & \textbf{0.3832 }& 0.4181 & 0.4443 & 0.4916 & 0.5198 & 0.5923\\
\hline
CV-DMH & \textbf{0.3028} & \textbf{0.3153} & \textbf{0.3275} & \textbf{0.3458} & \textbf{0.3350} & 0.3534 & 0.3799 & \textbf{0.4190} & \textbf{0.4533} & \textbf{0.5029} & \textbf{0.5293} & \textbf{0.6025}\\
\hline
\end{tabular}
\end{table*}
\section{Experimental Results and Discussions}
\label{sec:5}
\subsection{Performance Comparison Results}
We report mAP results and \emph{Precision-Scope}
curves of all approaches in Table \ref{resulttable} and Figure \ref{fig_oxford5k}, \ref{fig_paris6k}, \ref{fig_mmaris},
respectively. From the presented results, we can easily find that CV-DMH outperforms the
competitors on almost all cases. It is interesting to find that, even with less binary bits,
CV-DMH can still achieve higher mAP than many competitors with longer binary codes.
Further, Figure \ref{fig_mmaris} shows that, on \emph{Paris500K} and 128 bits, the precision gain of CV-DMH over the second best
approach is more than 10\%, and it becomes larger when more images are returned. Moreover,
we observe that performance improvement on \emph{Paris500K} is more than that obtained
on \emph{Oxford5K} and \emph{Paris6K}. Indeed, images in \emph{Paris500K} have more landmark images and
diverse visual appearances. This experimental phenomenon validates the desirable property of CV-DMH on
accommodating the visual variations. Finally, we observe that the retrieval performance of CV-DMH on \emph{Oxford5K}
is steadily improved when binary code length increases. However, we don't gain similar
observations for many approaches studied in this experimental study.
This is because CV-DMH adopts discrete optimization to solve hashing codes directly without any relaxing. The design can
successfully avoid the accumulated quantization errors brought in compared approaches. In this case, more binary bits will bring more discriminative information and thus enable CV-DMH to gain higher discriminative capability.
\begin{table*}
\small
\caption{Canonical views improve the robustness of CV-DMH. CV-DMH-II denotes direct multi-modal hashing with anchor transformation.}
\label{effinter}
\centering
\begin{tabular}{|p{16mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}
|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}
|p{8mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{\emph{Oxford5K}} & \multicolumn{4}{c|}{\emph{Paris6K}} & \multicolumn{4}{c|}{\emph{Paris500K}}\\
\cline{2-13}
& 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 \\
\hline
CV-DMH-II & 0.2829 & 0.2816 & 0.2903 & 0.2994 & 0.2699 & 0.2768 & 0.2852 & 0.2822 & 0.2877 & 0.3248 & 0.3450 & 0.3750 \\
\hline
CV-DMH & \textbf{0.3028} & \textbf{0.3153} & \textbf{0.3275} & \textbf{0.3458} & \textbf{0.3350} & \textbf{0.3534} & \textbf{0.3799} & \textbf{0.4190} & \textbf{0.4533} & \textbf{0.5029} & \textbf{0.5293} & \textbf{0.6025}\\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\small
\caption{Effects of discrete optimization. CV-DMH-II: it only considers bit-uncorrelated constraint by relaxing the discrete constraint and removing balance constraint in the Eq.(\ref{disobj}). CV-DMH-III: it considers bit-uncorrelated constraint and balance constraint together with discrete constraint relaxing. }
\label{effdsop}
\centering
\begin{tabular}{|p{18mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}
|p{8mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}|p{10mm}<{\centering}|p{8mm}<{\centering}|p{8mm}<{\centering}
|p{8mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{\emph{Oxford5K}} & \multicolumn{4}{c|}{\emph{Paris6K}} & \multicolumn{4}{c|}{\emph{Paris500K}}\\
\cline{2-13}
& 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 & 32 & 48 & 64 & 128 \\
\hline
CV-DMH-III & 0.2905 & 0.3179 & 0.3160 & 0.3403 & 0.3110 & 0.3194 & 0.3397 & 0.3887 & 0.4423 & 0.4918 & 0.5189 & 0.5917 \\
\hline
CV-DMH-IV & 0.2932 & 0.3177 & 0.3219 & 0.3373 & 0.3175 & 0.3510 & 0.3510 & 0.4003 & 0.4421 & 0.4935 & 0.5205 & 0.5915 \\
\hline
CV-DMH & \textbf{0.3028} & \textbf{0.3153} & \textbf{0.3275} & \textbf{0.3458} & \textbf{0.3350} & \textbf{0.3534} & \textbf{0.3799} & \textbf{0.4190} & \textbf{0.4533} & \textbf{0.5029} & \textbf{0.5293} & \textbf{0.6025}\\
\hline
\end{tabular}
\end{table*}
\subsection{Canonical View or Not?}
To see how the canonical view mining can benefit multi-modal hashing learning, we first compare the performance of CV-DMH with the one (denoted as CV-DMH-II) which performs hashing (Eq.(\ref{disobj})) directly on raw concatenated multiple low-level features. Table \ref{effcan} presents the detailed comparative results. From it, we easily find that CV-DMH can consistently yield better performance. On three datasets, the maximum search precision improvements reach about 2\%, 7\%, and 9\%, respectively. The performance improvement is attributed to the fact that, canonical views capture key visual contents of landmarks, diverse visual contents can be robustly accommodated with intermediate representation, and thus hashing codes learned on intermediate representation enjoy better robustness.
Then, we investigate the effects of canonical view mining in multiple modalities. We compare the performance of CV-DMH with the approaches that perform hashing (Eq.(\ref{disobj})) on only uni-modal canonical view set. We denote them directly with the corresponding modality names: CM, LBP, EDH, BOVW, and GIST respectively. Table \ref{feat} presents the main results. It demonstrates that CV-DMH can achieve the best performance. The reason is that, with multi-modal learning, canonical view set can cover more visual variations and thus CV-DMH can enjoy better robustness. All the above results clearly demonstrate that CV-DMH adopts a reasonable strategy by employing canonical views for MLS.
Finally, we validate the effects of the proposed submodular function based canonical view selection approach. We compare the performance of CV-DMH with four variants of our method. The first two competitors \emph{random} and \emph{K-means} discover canonical views by randomly sampling and K-means respectively. The third competitor DL learns canonical views by dictionary learning with K-SVD \cite{Aharon2006}. The last one R-CV-DMH selects canonical views by only considering representativeness of view set. The detailed comparison results are presented in Table \ref{effesumbo}. It can be easily observed that CV-DMH can achieve better performance in most cases. These results demonstrate the effectiveness of submodular function on discovering canonical views of landmarks. The reasons for the weakness of competitors can be explained as follows: the approach \emph{random} discovers canonical views without considering any underlying visual landmark distributions. \emph{K-means} and R-CV-DMH only consider representativeness of the selected view set, they unfortunately ignore the visual redundancy of views. In this case, with the fixed size of canonical view set, real representative views may be excluded by redundant ones. DL seeks the optimal ``dictionary" views that fit the visual distributions of training data. However, the learned dictionaries may not well adapt the query visual contents. In contrast, our approach comprehensively considers representativeness and redundancy of view set. The canonical views capture the key visual characteristics of landmarks and thus can characterize the visual contents of both query and database images.
\subsection{Effects of Intermediate Representation}
Intermediate representation bridges the canonical view mining and discrete binary embedding model. It is generated by calculating the multi-modal sparse reconstruction coefficients between image and canonical views. This subsection evaluates its effectiveness. Specifically, we compare CV-DMH with the variant approach CV-DMH-II that generates intermediate representation by calculating the similarities between images and canonical views. The calculating process is similar to anchor transformation in many existing hashing approaches \cite{SDH,DGH,HASHGRAPH}. In this experiment, we set the same number of nearest canonical views in two compared approaches for calculation. Table \ref{effinter} summarizes the main results. It clearly demonstrates that CV-DMH can consistently outperform the CV-DMH-II on all datasets and code lengths. The potential reason is that: the auto-generated sparse reconstruction coefficient can robustly accommodate the visual variations of landmark images with certain canonical views, they can provide more robust representation bases for subsequent binary embedding.
\subsection{Effects of Discrete Optimization}
To evaluate the effects of discrete optimization, we compare the performance between CV-DMH and its two variants. CV-DMH-III: it only considers bit-uncorrelated constraint by relaxing the discrete constraint and removing balance constraint in the Eq.(\ref{disobj}). CV-DMH-IV: it considers bit-uncorrelated constraint and balance constraint together with discrete constraint relaxing. Both two variants adopt conventional ``relaxing+rounding" optimization as many existing hashing approaches. The relaxed hashing values are also solved with ALM and the final binary hashing codes are generated by mean thresholding. Table \ref{effdsop} gives main mAP comparison results. We can clearly observe that CV-DMH can achieve better performance in all cases. These results validate the effects of direct discrete optimization and considering three constraints.
\begin{figure*}
\centering
\mbox{
\subfigure[$\alpha$ is fixed to $10^{-2}$]{\includegraphics[width=40mm]{para_alpha.pdf}}
}\mbox{
\subfigure[$\beta$ is fixed to $10^{-4}$]{\includegraphics[width=40mm]{para_beta.pdf}}
}\mbox{
\subfigure[$\gamma$ is fixed to $10^2$]{\includegraphics[width=40mm]{para_gamma.pdf}}
}\mbox{
\subfigure[]{\includegraphics[width=40mm]{iter_Paris500K.pdf}}
}
\caption{(a-c) CV-DMH performance variations with parameters in Eq.(\ref{disobj}) on \emph{Paris5K} when binary code length is 128. (d) Performance performance variations with iterations in Algorithm \ref{calv}.}
\label{fig_robustness}
\end{figure*}
\subsection{Convergency Analysis}
At each iteration in Algorithm \ref{calv}, the updating of variables will decrease the objective function value. As indicated by ALM optimization theory \cite{ALMone}, the iterations will make the optimization process converged. We also conduct empirical study on the convergence property using \emph{Paris500k}. Fig.\ref{fig_robustness}(d) presents the results. We observe that the objective function value first decreases with the number of iterations and then becomes steady after certain iterations (about 7 $\sim$ 8 iterations). This result demonstrates that the convergence of the proposed method.
\subsection{Parameter Study}
In this subsection, we conduct empirical experiments to study the performance variations with involved parameters in CV-DMH. Specifically, we observe the performance variations of CV-DMH with $\alpha$, $\beta$, and $\gamma$. They are used in Eq.(\ref{disobj}) to play the trade-off between regularization terms and empirical loss. Due to the limited space, we only report the results on \emph{Paris5K} when hashing code length is 128. Similar results can be found on other code lengths and datasets. We test the results when $\alpha$, $\beta$, $\gamma$ are varied from $[10^{-4}, 10^{-2}, 1, 10^2, 10^4]$. We observe performance variations with two of them by fixing the other parameter. Experimental results are presented in Fig.\ref{fig_robustness}(a-c). From these figures, we can clearly find that the performance is relatively stable to a wide range of $\alpha, \beta, \gamma$ variations.
\section{Conclusion and Future Work}
\label{sec:6}
In this paper, we propose a novel hashing scheme CV-DMH to learn compact hashing codes for supporting efficient and robust mobile landmark search. The design of CV-DMH has inspired by an interesting observation that: only canonical views of landmarks are frequently photographed and disseminated by different tourists, these views naturally provide effective visual representation basis of landmark hashing. We first develop submodular function based iterative mining to select canonical views that are discriminative and with limited redundancy. An intermediate representation is then generated on canonical views to robustly characterize the diverse visual landmark contents with sparse visual relations. At the final stage, we design a discrete binary embedding model to transform continuous intermediate representation to compact binary codes. To avoid relaxing quantization errors brought in conventional methods, we propose ALM-based optimization method to directly solve discrete solution. Experimental results on three real landmark datasets demonstrate that our proposed approach can achieve superior performance compared with several state-of-the-art approaches.
The current work will continue along several directions for further investigation:
\begin{enumerate}[1.]
\item Recent works \cite{sigpro/ChengS16} indicate that contextual modalities associated with landmark images enjoy better discriminative capability. They inspire us to integrate contextual modalities with our hashing model and thus embed more discriminative semantics into hashing codes.
\item Current three learning components of our approach are operated in three subsequent steps. In the further, we will try to develop a unified learning formulation to systemically integrate them together for further performance improvement.
\item The Laplacian matrix construction based on visual graph in Eq.(\ref{disobj}) consumes $O(N^2)$. In the future, we will explore strategies to further reduce the computation from algorithmic perspective.
\item We will apply our hashing model to other applications with the similar characteristics with landmarks.
\end{enumerate}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\section*{Acknowledgment}
The authors would like to thank the anonymous reviewers for their constructive and helpful suggestions.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
|
\section{Introduction}\label{sec:intro}
Let $X=(X_n: n\geq 0)$ be a Markov chain taking values in a state space $S$. For the purpose of this paper, the state space $S$ may be discrete or continuous. In many applications settings, it is natural to consider the behavior of $X$ as a function of a parameter $\theta$ that affects the transition dynamics of the process. In particular, suppose that for each $\theta$ in some open neighborhood of $\theta_0\in \mathbb{R}^d$, $P(\theta) = (P(\theta, x, dy): x,y\in S)$ defines the one-step transition kernel of $X$ associated with parameter choice $\theta$. In such a setting, computing the derivative of some application-specific expectation is often of interest.
Such derivatives play a key role when one is numerically optimizing an objective function, defined as a Markov chain's expected value, over the decision parameter $\theta$. In addition, such derivatives describe the sensitivity of the expected value under consideration to perturbations in $\theta$. Such sensitivities are valuable in statistical applications, and arise when one applies (for example) the ``delta method'' in conjunction with estimating equations involving some expectation of the observed Markov chain; see, for example, \cite{lehmann06}. More generally, sensitivity analysis is important when one is interested in understanding how robust the model is to uncertainties in the input parameters.
In particular, suppose that $\theta$ is a vector of statistical parameters, and that a data set of size $n$ has been collected to estimate the underlying true parameter $\theta^*$. In significant generality, the associated estiamtor $\hat \theta_n$ for $\theta^*$ will satisfy a central limit theorem (CLT) of the form
$$
n^{1/2} (\hat \theta_n - \theta^*) \Rightarrow N(0,C)
$$
as $n\to \infty$, where $\Rightarrow$ denotes weak convergence and $N(0,C)$ is a normally distributed random column vector with mean $0$ and covariance matrix $C$; see, for example, \cite{ibragimov81}.
In many applications, one wishes to understand how the uncertainty in our estimator $\hat \theta_n$ of $\theta^*$ propagates through the model associated with $X$ to produce uncertainty in output measures of interest.
Suppose, for example, that the decision-maker focuses her attention on a performance measure of the form $\alpha(\theta) = \mathbf{E}_\theta Z$, where $Z$ is some appropriately chosen random variable (rv) and $\mathbf{E}_\theta(\cdot)$ is the expectation operator under which $X$ evolvoes according to $\P(\theta)$.
If $\alpha(\cdot)$ is differentiable at $\theta^*$, then
$$
n^{1/2}(\alpha(\hat \theta_n)-\alpha(\theta^*)) \Rightarrow \nabla \alpha(\theta^*) N(0,C)
$$
as $n\to \infty$, where $\nabla\alpha(\theta)$ is the (row) gradient vector evaluated at $\theta$; see p.122 of \cite{serfling80}. If, in addition, $\nabla \alpha(\cdot)$ is continuous at $\theta^*$ and $C$ can be consistently estimated from the observed data via an estimator $C_n$, the interval
\begin{equation}\label{eqn_1_1}
\left[\alpha(\hat \theta_n) - z\frac{\sigma_n}{\sqrt{n}},\ \alpha(\hat \theta_n) + z\frac{\sigma_n}{\sqrt{n}}\right]
\end{equation}
is an asymptotic $100(1-\delta)$\% confidence interval for $\alpha(\theta^*)$ (provided $\nabla \alpha(\theta^*)C\nabla \alpha(\theta^*)^T > 0$), where $z$ is chosen so that $P(-z\leq N(0,1)\leq z) = 1-\delta$ and $\sigma_n = \sqrt{\nabla \alpha(\hat \theta_n)C_n \nabla \alpha(\hat \theta_n)^T}$. The confidence interval (\ref{eqn_1_1}) provides the modeler with the desired sensitivity and robustness of the model described by $X$ to the statistical uncertainties present in the estiamtion of $\theta^*$. Of course, this approach rests on the differentiability of $\alpha(\cdot)$ and on one's ability to compute the gradient. This paper provides conditions guaranteeing differentiability in the general state space Markov chian settings and provides representations for those derivatives suitable for computation.
The problem of determining such differentiability has a long history and has been addressed through various approaches including weak differentiation (\citealp{vazquez1992, pflug1992}), likelihood ratio \citep{glynn95}, measure-valued differentiation (\citealp{heidergott06}, \linebreak\citealp{heidergott2006B}), and derivative regenration (\citealp{glasserman1993}).
However, most of the previous approaches are limited to special classes of problems. For example, the results in \linebreak\cite{vazquez1992} and \cite{pflug1992} are limited to bounded performance functionals;
\cite{glasserman1993} impose special structures in the the transition dynamics of the Markov chains and their parametrization;
\cite{glynn95} assumes for random horizon expectations that the associated stopping times have finite exponential moments, and for stationary expectations that the Markov chain is geometrically ergodic.
\cite{heidergott2006B} provide weaker conditions for random horizon performance measures based on measure-valued differentiation approach, but their sufficient conditions are difficult to verify in general and still require that the associated stopping times possess finite (at least) second moment.
Also based on measure-valued differentiation, \cite{heidergott06} study stationary expectations. However, the sufficient conditions verifiable based on the model building blocks in the paper require geometric ergodicity of the Markov chain.
In this paper, on the other hand, we provide easily verifiable sufficient conditions for random horizon expectations that do not require any moment conditions for the associated stopping times---hence, allowing even infinite horizon expectations.
For stationary expectations, we provide (again, easily verifiable) sufficient conditions that does not require geometric ergodicity.
We illustrate the sharpness of our differentiability criteria with the example of waiting times of G/G/1 queues with heavy tailed service times.
The rest of the paper is organized as follows.
Section 2 develops a preliminary theory for both random-horizon expectations and stationary expectations based on simple and clean operator theoretic arguments.
Section 3 provides more general criteria for differentiability of random horizon expectations based on the stochastic Lyapunov type inequalities arguments.
In Section 4, also taking Lyapunov inequalities approach, we establish the differentiability criteria for stationary expectations.
\section{Operator-theoretic Criteria for Differentiability}\label{sec:operator}
We start by studying differentiability in a setting in which one can use operator arguments to establish existance of derivatives. In this operator setting, the proofs and theorem statements are especially straightforward.
Consider a Markov chain $X=(X_n:n\geq 0)$ living on state space $S$, with one step transition kernel $P=(P(x,dy): x,y\in S)$, where
$$
P(x,dy) = P(X_{n+1}\in dy|X_n = x)
$$
for $x,y\in S$. We focus first on expectations of the form
\begin{equation}\label{def2:u_star}
u^*(x) = \mathbf{E}_x \sum_{j=0}^{T-1} \exp(\sum_{k=0}^{j-1} g(X_k)) f(X_j) + \exp(\sum_{k=0}^{T-1}g(X_k)) f(X_T),
\end{equation}
where $T = \inf\{n\geq 0: X_n \in C^c\}$ is the first hitting time of the ``target set'' $C^c\subseteq U$, $f:S\to\mathbb{R}_+$, $g:S\to \mathbb{R}$, and $\mathbf{E}_x(\cdot)\triangleq \mathbf{E}(\cdot|X_0=x)$.
In (\ref{def2:u_star}), we permit the possibility that $C^c = \phi$, in which case $T=\infty$ a.s., and $u^*$ is then to be interpreted as the ``infinite horizon discounted reward''
\begin{align*}
u^*(x) = \mathbf{E} _x \sum_{j=0}^\infty \exp\left(\sum_{k=0}^{j-1} g(X_k)\right) f(X_j).
\end{align*}
In addition to subsuming infinite horizon discounted rewards, (\ref{def2:u_star}) also includes expected hitting times ($g\equiv 0, f=1$ on $C$ and $f=0$ on $C^c$), exit probabilities ($g\equiv 0$, $f=0$ on $C$, and $f(x) = I(x\in B)$ for $x\in C^c$, when one is considering $P(X_T\in B | X_0 = x)$), and many other natural Markov chain expectations.
It is easy to verify that
\begin{equation}\label{eq:u_star}
u^* = \sum_{n=0}^\infty K^n \tilde f
\end{equation}
where $K = (K(x,dy): x, y\in C)$ is the non-negative kernel for which
\begin{equation}\label{eq2:2.3}
K(x,dy) = \exp(g(x)) P(x,dy)
\end{equation}
for $x,y\in S$, and
\begin{equation*}
\tilde f(x) = f(x) + \int_{C^c} \exp(g(x)) P(x,dy)f(y)
\end{equation*}
for $x \in C$. Here, we are taking advantage in (\ref{def2:u_star}) of the (common) notational convention that for a function $h:B \to \mathbb{R}$, a measure $\eta$ on $B$, and kernels $Q_1$ and $Q_2$ on $B$, the scalar $\eta h$, the function $Q_1 h$, the measure $\eta Q_1$, and the kernel $Q_1 Q_2$ are respectively defined via
$$\eta h = \int_B h(y) \eta(dy),$$
$$(Q_1 h)(x) = \int_B h(y) Q_1(x,dy),$$
$$(\eta Q_1)(A) = \int_B \eta(dx) Q_1 (x,A),$$
$$(Q_1 Q_2)(x,A) = \int_B Q_1(x,dy) Q_2(y,A),$$
whenever the right-hand sides are well-defined. Furthermore, we define the kernels $Q^n$ via $Q^0 (x,dy) = \delta_x(dy)$ (where $\delta_x( . )$ is a unit point mass at $x$), and $Q^n = Q Q^{(n-1)}$ for $n\geq1$.
Our goal is to use operator-theoretic tools to study the differentiability of (\ref{eq:u_star}). To this end, we start by defining the appropriate linear spaces that underlie this approach. Given a measurable space $(B, \mathcal B)$, measurable $w:B\to[1,\infty)$, and $h:B\to\mathbb{R}$, let $\|h\|_w = \sup\{ |h(x)|/w(x): x\in B\}$ and $L_w=\{h\in L:\|h\|_w < \infty\}$ where $L$ is the set of measurable functions. For a linear operator $Q:L_w\to L_w$ and a functional $\eta:L_w\to \mathbb{R}$, set
$$\vertiii{Q}_w = \sup_{h\in L_w: \|h\|_w \neq 0} \frac{\|Qh\|_w}{\|h\|_w}$$
and
$$\|\eta\|_{w} = \sup\{|\eta h|: h\in L_w, \|h\|_w \leq 1\}.$$
Then, let , $\mathcal L_w = \{Q\in \mathcal L: \vertiii{Q}_w < \infty\}$, and $\mathcal M_w = \{\eta\in \mathcal M: \|\eta\|_w < \infty\}$ where $\mathcal L$ and $\mathcal M$ are the sets of kernels, and measures, respectively.
Each of the spaces $L_w$, $\mathcal L_w$, and $\mathcal M_w$ are Banach spaces under their respective norms and addition / scalar multiplication operations; see Appendix~\ref{appendix:B}.
Furthermore, for $Q_1, Q_2 \in \mathcal L_w$, $h\in L_w$, and $\eta \in \mathcal M_w$, it is easy to show that
\begin{equation}\label{eq:2.4}
\vertiii{Q_1Q_2}_w \leq \vertiii{Q_1}_w\cdot\vertiii{Q_2}_w
\end{equation}
and
\begin{equation}\label{eq:2.5}
\begin{aligned}
\|Qh\|_w &\leq \vertiii{Q}_w \cdot\|h\|_w,\\
\|\eta Q\|_w &\leq \|\eta\|_w \cdot \vertiii{Q}_w,\\
|\eta h| &\leq \|\eta\|_w \cdot \|h \|_w;
\end{aligned}
\end{equation}
see, for example, \cite{Dunford71} for the special case $w\equiv 1$.
In view of (\ref{eq:2.4}), if $\vertiii{Q^m}_w < 1$ for some $m\geq 1$, then $(I-Q)$ is invertible on $\mathcal L_w$ and
$$ (I-Q)^{-1} = \sum_{n=0}^\infty Q^n.$$
Given a parametrized family of kernels $(Q(\theta)\in \mathcal L_w: \theta \in (a,b))$, we say that $Q(\cdot)$ is \emph{continuous} in $\mathcal L_w$ at $\theta_0 \in (a,b)$ if
$\vertiii{Q(\theta_0 + h)-Q(\theta_0) }_w \to 0$
as $h\to 0$,
and \emph{differentiable} in $\mathcal L_w$ at $\theta_0 \in (a,b)$ with derivative $Q'(\theta_0)$ if there exists a kernel $Q'(\theta_0) \in \mathcal L_w$ for which
$$\vertiii{\frac{Q(\theta_0 + h)-Q(\theta_0)}{h} - Q'(\theta_0)}_w \to 0$$
as $h\to 0$.
If $Q(\cdot)$ is differentiable in a neighborhood of $\theta_0$ with derivative $Q'(\cdot)$, and $Q'(\cdot)$ is continuous in $\mathcal L_w$, then we say that $Q(\cdot)$ is continuously differentiable at $\theta_0$.
Similarly, given families $(f(\theta)\in L_w: \theta\in (a,b))$ and $(\eta(\theta) \in \mathcal M_w: \theta\in (a,b))$, we say that $f(\cdot)$ is
\emph{continuous} in $ L_w$ at $\theta_0$ if
$\left\|f(\theta_0+h) - f(\theta_0)\right\|_w \to 0$
as $h\to 0$,
and \emph{differentiable} in $ L_w$ at $\theta_0$ if there exists $f'(\theta_0)\in L_w$ such that
$$\left\|\frac{f(\theta_0+h) - f(\theta_0)}{h} - f'(\theta_0)\right\|_w \to 0$$
as $h\to 0$; and $\eta(\cdot)$ is
\emph{continuous} in $\mathcal M_w$ at $\theta_0$ if
$ \left\|\eta(\theta_0 + h) - \eta(\theta_0)\right\|_w \to 0$
as $h\to0$,
and differentiable in $\mathcal M_w$ at $\theta_0$ if there exists $\eta'(\theta_0) \in \mathcal M_w$ such that
$$ \left\|\frac{\eta(\theta_0 + h) - \eta(\theta_0)}{h} - \eta'(\theta_0)\right\|_w \to 0$$
as $h\to0$. As in $\mathcal L_w$, if $f(\cdot)$ and $\eta(\cdot)$ are differentiable and their derivatives are continuous at $\theta_0$ in $L_w$ and $\mathcal M_w$ respectively, we say that they are continuously differentiable.
Assuming that $(Q(\theta): \theta \in (a,b))$ is $n$-times differentiable in some neighborhood $\mathcal N$ of $\theta_0$, with derivative $(Q^{(n)}(\theta):\theta\in\mathcal N)$, we say that $Q(\cdot)$ is $(n+1)$-times differentiable in $\mathcal L_w$ at $\theta_0$ if $(Q^{(n)}(\theta): \theta\in \mathcal N)$ is differentiable at $\theta_0$, with corresponding derivative $Q^{(n+1)}(\theta_0)$.
We can analogously define $f^{(n+1)}(\theta_0)$ and $\eta^{(n+1)}(\theta_0)$ in the spaces $L_w$ and $\mathcal M_w$, respectively.
(We restrict our discussion in this paper to scalar $\theta$, since the vector case introduces no new mathematical issues.)
We can now state our first result, pertaining to differentiability of $u^*$.
\begin{theorem}\label{thm:01}
Suppose there exists $w:C\to[1,\infty)$ and $\theta_0 \in (a,b)$ for which:
\begin{enumerate}
\item[(a)] $\vertiii{K^m(\theta_0)}_w < 1$ for some $m\geq 1$;
\item[(b)] $K(\cdot)$ is \what{(continuously)} differentiable in $\mathcal L_w$ at $\theta_0$, with derivative $K'(\theta_0)$;
\item[(c)] $\tilde f(\cdot)$ is \what{(continuously)} differentiable in $L_w$ at $\theta_0$, with derivative $\tilde f'(\theta_0)$.
\end{enumerate}
Then:
\begin{enumerate}
\item[(i)] $(I-K(\theta))$ is invertible on $L_w$ for $\theta$ in a neighborhood of $\theta_0$;
\item[(ii)] Setting $G(\theta) = (I-K(\theta))^{-1}$, $G(\cdot)$ is \what{(continuously)} differentiable in $\mathcal L_w$ at $\theta_0$, and
\begin{equation*}
G'(\theta_0) = G(\theta_0) K'(\theta_0) G(\theta_0);
\end{equation*}
\item[(iii)] $u^*(\theta) = \sum_{n=0}^\infty K^n(\theta) \tilde f(\theta)$ is \what{(continuously)} differentiable in $L_w$ at $\theta_0$, with
\begin{equation}\label{eq:4.3a}
(u^*)'(\theta_0) = G'(\theta_0) \tilde f(\theta_0) + G(\theta_0)\tilde f'(\theta_0).
\end{equation}
\end{enumerate}
If, in addition, $K(\cdot)$ and $\tilde f(\cdot)$ are $n$-times \what{(continuously)} differentiable in $\mathcal L_w$ and $L_w$, respectively, at $\theta_0$, then $Q(\cdot)$ and $u^*(\cdot)$ are $n$-times \what{(continuously)} differentiable at $\theta_0$ in $\mathcal L_w$ and $L_w$, respectively, and $Q^{(n)}(\theta_0)$ and $(u^*)^{(n)}(\theta_0)$ can be recursively computed via
\begin{align}\label{eq2:D}
G^{(n)}(\theta_0) = \sum_{j=0}^{n-1} {n\choose j}G^{(j)}(\theta_0)K^{(n-j)}(\theta_0) G(\theta_0)
\end{align}
and
\begin{align}\label{eq2:E}
(u^*)^{(n)}(\theta_0) = G(\theta_0) \bigg(\tilde f^{(n)}(\theta_0) + \sum_{j=0}^{n-1}{n\choose j} K^{(n-j)} (\theta_0) (u^*)^{(j)}(\theta_0)\bigg)
\end{align}
where, as usual, $K^{(0)}(\theta)\equiv K(\theta)$ and $\tilde f^{(0)}(\theta) = \tilde f(\theta)$.
\end{theorem}
\begin{proof}{Proof.
Part (i) is \rvout{well known: see for example, \what{[reference]}}\rvin{obvious}. For part (ii), note that assumptions (a) and (b) imply that there exists a neighborhood $\mathcal N$ of $\theta_0$ for which $\sup_{\theta\in \mathcal N}\vertiii{K^m(\theta)}_w < 1$ and $\sup_{\theta\in \mathcal N} \vertiii{K(\theta)}_w < \infty$, from which it follows that $\sup_{\theta\in \mathcal N}\vertiii{G(\theta)}_w < \infty$.
Furthermore, since $(I-K(\theta_0+h)) G(\theta_0+h) = G(\theta_0+h) (I-K(\theta_0+h)) = I$, evidently
\begin{equation*}
(G(\theta_0+h) - G(\theta_0)) (I-K(\theta_0)) = G(\theta_0+h) (K(\theta_0+h) - K(\theta_0)),
\end{equation*}
so that
\begin{equation}\label{eq2:2.9}
G(\theta_0 + h) - G(\theta_0) = G(\theta_0 + h) (K(\theta_0+h) - K(\theta_0)) G(\theta_0)
\end{equation}
Clearly, this implies that $\vertiii{G(\theta_0+h) - G(\theta_0)}_w \leq \vertiii{G(\theta_0+h)}_w\vertiii{K(\theta_0+h)- K(\theta_0)}_w\vertiii{G(\theta_0)}_w \to 0$ as $h \to 0$, so $G(\cdot)$ is continuous in $\mathcal L_w$ at $\theta_0$. Consequently, (\ref{eq2:2.9}) implies that $G(\cdot)$ is differentiable in $\mathcal L_w$ at $\theta_0$, with $G'(\theta_0) = G(\theta_0) K'(\theta_0)G(\theta_0)$. \what{In case $K'$ is continuous, continuity of $G'$ is also immediate from this expression.}
For part (iii), the result follows analogously from the identity
\begin{equation*}
u^*(\theta_0+h) - u^*(\theta_0) = G(\theta_0) (\tilde f(\theta_0+h) - \tilde f(\theta_0)) + (G(\theta_0+h) - G(\theta_0)) \tilde f(\theta_0+h).
\end{equation*}
The proof for the $n$-fold derivatives for $n\geq 2$ is very similar and therefore omitted.
\end{proof}
\begin{remark}
Suppose that $K(\cdot)$ posesses a density $(k(\cdot, x, y): x,y \in C)$ that is $n$-times differentiable (with (pointwise) derivative $(k^{(n)}(\cdot, x,y):x,y\in C))$. For $\epsilon>0$ and $0\leq j \leq n$, let $\tilde \omega_\epsilon^{(j)}(x,y) = \sup_{|\theta-\theta_0|<\epsilon} |k^{(j)}(x, y)|$. Then, the conditions
\begin{align}
&\sup_{x\in C} \int_C K^m(\theta_0, x, dy) \frac{w(y)}{w(x)} < 1\quad \text{for some } m\geq 1,\label{eq:02.10}\\
&\sup_{x\in C} \int_C \omega_\epsilon^{(j)}(x,y)\frac{w(y)}{w(x)} K(\theta_0, x, dy) < \infty,\label{eq:02.11}\\
&\sup_{x\in C} \int_{C^c} (1+\tilde \omega_\epsilon^{(j)}(x,y))\frac{|f(y)|}{w(x)}K(\theta_0, x, dy) < \infty,\label{eq:02.12}
\end{align}
for $j=0,\ldots,n$ imply (a), (b), and (c) of Theorem~\ref{thm:01} implying the validity of (\ref{eq2:D}) and (\ref{eq2:E}).
\end{remark}
There is an analgous differentiability results for measures. For a given initial distribution $\mu$ on $C$, let $\nu$ be the measure defined by
\begin{equation*}
\nu(dy) = \mathbf{E}_\mu \sum_{j=0}^{T-1} \exp \left(\sum_{k=0}^{j-1} g(X_k)\right) \mathbb{I}(X_j \in dy)
\end{equation*}
for $y\in S$, where $\mathbf{E}_\mu(\cdot) \triangleq \int_C \mu(dx) \mathbf{E}_x(\cdot)$. Then,
\begin{equation*}
\nu = \sum_{n=0}^\infty \mu K^n,
\end{equation*}
where $K$ is defined as in (\ref{eq2:2.3}). Assume that $\mu(\cdot)$ and $K(\cdot)$ now depend on the parameter $\theta$ (so that $\nu$ does as well). The following result has a proof identical to that of Theorem~\ref{thm:01}, and is therefore omitted.
\begin{theorem}\label{thm:02}
Suppose there exists $w:C\to[1,\infty)$ and $\theta_0 \in (a,b)$ for which:
\begin{itemize}
\item[(a)] $\vertiii{K^m(\theta_0)}_w < 1$ for some $m\geq 1$;
\item[(b)] $K(\cdot)$ is (continuously) differentiable in $\mathcal L_w$ at $\theta_0$, with derivative $K'(\theta_0)$
\item[(c)] $\mu(\cdot)$ is (continuously) differentiable in $\mathcal M_w$ at $\theta_0$, with derivative $\mu'(\theta_0)$.
\end{itemize}
Then, $\nu(\theta) = \sum_{n=0}^\infty \mu(\theta)K^m(\theta)$ is (continuously) differentiable in $\mathcal M_w$ in $\theta_0$, with
\begin{equation*}
\nu'(\theta_0) = \mu'(\theta_0)G(\theta_0) + \nu(\theta_0) G'(\theta_0).
\end{equation*}
If, in addition, $K(\cdot)$ and $\mu(\cdot)$ are $n$-times (continuously) differentiable in $\mathcal L_w$ and $\mathcal M_w$, respectively, at $\theta_0$, then $\nu(\cdot)$ is $n$-times (continuously) differentiable in $\mathcal M_w$, and $\nu^{(n)}(\theta_0)$ can be recursively computed via
\begin{equation*}
\nu^{(n)} (\theta_0) = \left(\mu^{(n)}(\theta_0) + \sum_{j=0}^{n-1}{n \choose j}\nu^{(j)}K^{(n-j)}(\theta_0)\right)G(\theta_0).
\end{equation*}
\end{theorem}
We finish this section with a short operator-theoretic argument establishing existence of a derivative for the stationary distribution under the assumption of geometric ergodicity (see condition (a) below, which is the key Lyapunov condition that implies geometric ergodicity in Chapter 15 of \cite{meyn09}).
\begin{theorem}\label{thm:4}
Suppose that there exists a subset $A\subseteq S$, $\epsilon,c > 0$, $\lambda, r \in (0,1)$, an integer $m\geq 1$, a probability measure $\varphi$ on $S$, and $w:S\to[1,\infty)$ such that:
\begin{enumerate}
\item[(a)] $(P(\theta_0) w)(x) \leq rw(x) + cI(x\in A)$ \quad for $x\in S$;
\item[(b)] $P^m(\theta,x,dy) \geq \lambda \varphi(dy)$ for $x\in A$, $y\in S$, and $|\theta-\theta_0|<\epsilon$;
\item[(c)] $P(\cdot)$ is \what{(continuously)} differentiable in $\mathcal L_w$ at $\theta_0$.
\end{enumerate}
Then, $X$ is positive Harris recurrent for $\theta$ in a neighborhood of $\theta_0$, and the stationary distributions $\pi(\theta)\in \mathcal M_w$ for $\theta$ in a neighborhood of $\theta_0$ are \what{(continuously)} differentiable in $\mathcal M_w$ at $\theta_0$.
Furthermore, if $\Pi(\theta_0)$ is the kernel defined by $\Pi(\theta_0, x, dy) = \pi(\theta_0, dy)$ for $x,y\in S$, $(I-P(\theta_0)+ \Pi(\theta_0))$ has an inverse on $\mathcal L_w$ and
\begin{equation}\label{eq2:2.13}
\pi'(\theta_0) = \pi(\theta_0) P'(\theta_0) (I-P(\theta_0)+\Pi(\theta_0))^{-1}.
\end{equation}
If, in addition, $P(\cdot)$ is $n$-times \what{(continuously)} differentiable in $\mathcal L_w$ at $\theta_0$, then $\pi(\cdot)$ is $n$-times \what{(continuously)} differentiable in $\mathcal M_w$ at $\theta_0$, and $\pi^{(n)}(\theta_0)$ can be recursively computed via
\begin{equation*}
\pi^{(n)}(\theta_0) = \sum_{j=0}^{n-1} {n \choose j} \pi^{(j)}(\theta_0) P^{(n-j)}(\theta_0) (I-P(\theta_0) + \Pi(\theta_0))^{-1}.
\end{equation*}
\end{theorem}
\begin{remark}
Note that Theorem 4 of \cite{glynn95} is closely related to the above theorem. See also Remark 11 and the Kendall set assumption in \cite{glynn95}. \cite{heidergott06} also imposes similar assumption to establish the measure-valued derivative of the stationary distribution.
\end{remark}
\begin{proof}{Proof.}
In view of (a) and (c), there exists $r'<1$ such that
\begin{equation}\label{eq2:2.14}
(P(\theta_0+h) w)(x) \leq r' w(x) + cI(x\in A)
\end{equation}
for $x\in S$ and $|h|$ sufficiently small. Assumptions (a) and (b), and the fact that $w\geq 1$ implies that $X$ is positive Harris recurrent for $\theta$ in a neighborhood of $\theta_0$. We can now appeal to Theorem 2.3 of \cite{glynn96} to establish that $(I-P(\theta_0) + \Pi(\theta_0))$ is invertible on $\mathcal L_w$, with $(I-P(\theta_0) + \Pi(\theta_0))^{-1} \in \mathcal L_w$.
Furthermore, according to \cite{glynn08}, (\ref{eq2:2.14}) implies that $\pi(\theta_0+h) w \leq c/(1-r')$, and hence $\|\pi(\theta_0+h)\|_w \leq c/(1-r').$ Also,
\begin{align*}
(\pi(\theta_0+h) - \pi(\theta_0)) (I-P(\theta_0))
&= \pi(\theta_0+h) (I-P(\theta_0))\\
&= \pi(\theta_0+h) (P(\theta_0+h) - P(\theta_0)).
\end{align*}
In addition, $\nu \Pi(\theta_0) = \pi(\theta_0) $ for any probability $\nu$ on $S$. So $(\pi(\theta_0+h) - \pi(\theta_0))\Pi(\theta_0) = 0$. Consequently,
\begin{equation*}
(\pi(\theta_0 + h) - \pi(\theta_0)) (I-P(\theta_0)+\Pi(\theta_0))
= \pi(\theta_0+h) (P(\theta_0+h)-P(\theta_0)),
\end{equation*}
from which it follows that
\begin{equation}\label{eq2:2.15}
\pi(\theta_0 + h) - \pi(\theta_0)
= \pi(\theta_0+h) (P(\theta_0+h)-P(\theta_0))(I-P(\theta_0)+\Pi(\theta_0))^{-1}.
\end{equation}
Thus,
\begin{equation}\label{continuity_of_pi}
\|\pi(\theta_0+h) - \pi(\theta_0)\|_w
\leq \frac{c}{1-r'}\vertiii{P(\theta_0+h) - P(\theta_0)}_w\cdot \vertiii{(I-P(\theta_0)+\Pi(\theta_0))^{-1}}_w
\end{equation}
Since $P(\cdot)$ is differentiable in $\mathcal L_w$, $\vertiii{P(\theta_0+h)-P(\theta_0)}_w \to 0$ as $h \to 0$, so $\pi(\theta_0+h) \to \pi(\theta_0)$ in $\mathcal M_w$ as $h\to 0$. Letting $h\to 0$ in $(\ref{eq2:2.15})$ then yields (\ref{eq2:2.13}).
\what{
For the continuity of the derivative in case $P(\cdot)$ is continuously differentiable, note first that (a) and (b)
imply that $\vertiii{(P(\theta_0)-\Pi(\theta_0))^m}_w<1$ for some $m\geq 1$; this along with the continuity of $P(\cdot)$ and $\pi(\cdot)$, in turn, implies that $\sup_{|h|\leq h_0}\vertiii{(P(\theta_0+h)-\Pi(\theta_0+h))^m}_w<1$ for a small enough $h_0$. Therefore, we conclude that $\vertiii{(I-P(\theta_0+h) + \Pi(\theta_0+h))^{-1}}_w$ is bounded (uniformly w.r.t.\ $h$).
From this, it is easy to see that the same argument as for (\ref{eq2:2.13}) works with $\theta = \theta_0+h$ instead of $\theta_0$ and proves that
\begin{equation}\label{pi_prime_general}
\pi'(\theta_0+h) = \pi(\theta_0+h) P'(\theta_0+h) (I-P(\theta_0+h)+\Pi(\theta_0+h))^{-1}.
\end{equation}
Now,
\begin{align*}
\pi'(\theta_0+h) - \pi'(\theta_0)
&=
\left(\pi'(\theta_0+h) - \frac{\pi(\theta_0)-\pi(\theta_0+h)}{-h}\right) - \left(\pi'(\theta_0)- \frac{\pi(\theta_0+h)-\pi(\theta_0)}{h} \right)
= \text{(I)} - \text{(II)}
\end{align*}
where we have already seen that (II) converges to 0.
To show that (I) also vanishes, note that
\begin{align*}
(\pi(\theta_0) - \pi(\theta_0+h))(I-P(\theta_0+h)+\Pi(\theta_0+h))
&
=(\pi(\theta_0) - \pi(\theta_0+h))(I-P(\theta_0+h))
\\
&
= \pi(\theta_0)(I-P(\theta_0+h))
= \pi(\theta_0)(P(\theta_0)-P(\theta_0+h)),
\end{align*}
and hence,
\begin{equation}\label{pi_finite_difference}
\frac{\pi(\theta_0) - \pi(\theta_0+h)}{-h}
=
\pi(\theta_0)\frac{P(\theta_0+h)-P(\theta_0)}{h}(I-P(\theta_0+h)+\Pi(\theta_0+h))^{-1}.
\end{equation}
From (\ref{pi_prime_general}), (\ref{pi_finite_difference}), the continuity of $\pi(\cdot)$, the continuous differentiability of $P(\cdot)$, and the uniform boundedness of the norm of $(I-P(\theta_0+h)+\Pi(\theta_0+h))^{-1}$, we conclude that (I) vanishes. Therefore, $\pi'(\cdot)$ is continuous at $\theta_0$.
}
Finally, as in Proposition 2, the proof for the $n$-fold derivatives for $n\geq 2$ follows similar lines, and is therefore omitted.
\end{proof}
This result establishes, in the presence of a single Lyapunov function $w$, the $n$-fold differentiability of the stationary distribution $\pi(\cdot)$ in $\mathcal M_w$. Of course, the simplicity of the result comes at the cost of assuming geometric ergodicity of $X$.
\section{Lyapunov Criteria for Differentiability of Random Horizon Expectations}\label{sec:random_horizon}
Let $\Lambda = (a,b)$ be an open interval containing $\theta_0$. For each $\theta \in \Lambda$, let $\mathbf{E}_x^\theta (\cdot) \triangleq \mathbf{E}^\theta(\cdot|X_0 = x)$ be the expectation operator associated with $X$, when $X$ is driven by the one-step transition kernel $P(\theta)$.
As in Section~\ref{sec:operator}, we consider
\begin{align}\label{def:u_star}
u^*(\theta, x) = \mathbf{E}_x^\theta \sum_{j=0}^{T-1} \exp\left(\sum_{k=0}^{j-1} g(X_k)\right)f(X_j) + \exp\left(\sum_{k=0}^{T-1} g(X_k)\right)f(X_T)
\end{align}
for each $x\in C$
given $f:S\to \mathbb{R}$, $g:S\to \mathbb{R}$, $\phi \neq C\subseteq S$, and $T=\inf \{n\geq 0: X_n \in C^c\}$.
Our goal, in this section, is to provide Lyapunov conditions under which $u^*(\theta) = (u^*(\theta,x): x \in C)$ is differentiable at $\theta_0$, and to provide an expression for the derivative ${u^*}'(\theta)$.
Note that if $f$ is non-negative, then $u^*(\theta)$ is always well-defined. Furthermore, by conditioning on $X_1$, it is easily seen that
\begin{align*}
u^*(\theta, x) = f(x) + \int_{C^c} \exp(g(x)) P(\theta,x,dy)f(y) + \int_C \exp(g(x))P(\theta, x,dy) u^*(\theta,y)
\end{align*}
for $x\in C$, and hence
\begin{equation}\label{eq:u_star}
u^*(\theta) = \tilde f(\theta) + K(\theta)u^*(\theta),
\end{equation}
where as in Section~\ref{sec:operator}
\begin{equation*}
\tilde f(\theta,x) = f(x) + \int_{C^c} \exp(g(x)) P(\theta,x,dy)f(y)
\end{equation*}
for $x\in C$, and $K(\theta) = (K(\theta, x, dy): x, y \in C)$ is the non-negative kernel on $C$ for which
\begin{equation*}
K(\theta, x, dy ) = \exp(g(x))P(\theta,x,dy).
\end{equation*}
Given (\ref{eq:u_star}), formal differentiation of both sides of the equation yields
\begin{equation}
{u^*}'(\theta_0) = \tilde f'(\theta_0) + K'(\theta_0){u^*}(\theta_0) + K(\theta_0) {u^*}'(\theta_0),
\end{equation}
so that ${u^*}'(\theta)$ should satisfy the linear system
\begin{equation}\label{eq:2.3a}
(I-K(\theta_0)){u^*}'(\theta_0) = \tilde f'(\theta_0) + K'(\theta_0)u^*(\theta_0).
\end{equation}
When $|C|$ is finite, it will frequently be the case that the matrix $K(\theta_0)$ has spectral radius less than $1$, in which case $I-K(\theta_0)$ is invertible and
\begin{equation}\label{eq:potential}
(I-K(\theta_0))^{-1} = \sum_{n=0}^\infty K^n(\theta_0)
\end{equation}
In this case,
\begin{equation*}
{u^*}'(\theta_0) = \sum_{n=0}^\infty K^n(\theta_0)\left(\tilde f'(\theta_0) + K'(\theta_0) u^*(\theta_0)\right).
\end{equation*}
But (\ref{eq:u_star}) and (\ref{eq:potential}) further imply that
\begin{equation}\label{rep:u_star}
u^*(\theta_0) = \sum_{n=0}^\infty K^n(\theta_0)\tilde f(\theta_0),
\end{equation}
and hence we arrive at the formula
\begin{equation}\label{eq:2.5}
{u^*}'(\theta_0) = \sum_{m=0}^\infty\sum_{n=0}^\infty K^m(\theta_0) K'(\theta_0) K^n(\theta_0) \tilde f(\theta_0) + \sum_{m=0}^\infty K^m(\theta_0) \tilde f'(\theta_0).
\end{equation}
The remainder of this section is largely concerned with rigorously extending the formula (\ref{eq:2.5}) to the general state space setting, under Lyapunov criteria that are close to minimal (and easily checkable from the model building blocks). We start by observing that when $f$ is non-negative, Fubini's theorem implies that
\begin{align}
u^*(\theta, x)
&= \sum_{j=0}^\infty \mathbf{E}_x^\theta \exp\left(\sum_{k=0}^{j-1} g(X_k)\right)f(X_j)I(T>j)\nonumber\\
&\qquad + \sum_{j=0}^\infty \mathbf{E}_x^{\theta} \exp\left( \sum_{k=0}^{j-1} g(X_k)\right) I(T\geq j) \hcancel{e^{g(X_j)}}f(X_j) \mathbb{I}(X_j \in C^c)\nonumber\\
&= \sum_{j=0}^\infty (K^{j}(\theta)\tilde f(\theta))(x),\label{eq:2.5a}
\end{align}
thereby rigorously verifying (\ref{rep:u_star}). To simplify the notation in the remainder of this paper, we set $K = K(\theta_0)$ and put
\begin{equation}\label{eq:2.6}
G = \sum_{n=0}^\infty K^n.
\end{equation}
Our path to providing rigorous conditions under which (\ref{eq:2.5}) holds involves the following key ``absolute continuity'' assumption:
\begin{itemize}
\item[A1.]
The kernels $(K(\theta): \theta \in \Lambda)$ are absolutely continuous with respect to $K$, in the sense that there exists a (measurable) density $(k(\theta, x,y): x,y\in C)$ such that
\begin{equation*}
K(\theta, x,dy) = k(\theta, x, y) K(x,dy)
\end{equation*}
for $\theta\in \Lambda$, $x,y\in C$.
\end{itemize}
Our absolute continuity condition is often a mild hypothesis. For example, when $X$ has a transition density with respect to a reference measure $\eta$, A1 is in force when the support of the density is independent of $\theta$.
We also need to assume that $K(\theta)$ is suitably differentiable at $\theta_0$.
\begin{itemize}
\item[A2.]
There exists $\epsilon>0$ such that for each $x,y\in C$, $k(\cdot, x,y)$ is continuously differentiable, with derivative $k'(\cdot, x,y)$, in $[\theta_0-\epsilon, \theta_0+\epsilon]$.
\end{itemize}
Set $\omega_\epsilon(x,y) = \sup\{|k'(\theta,x,y)|: |\theta-\theta_0|<\epsilon\}$, $k'(x,y) = k'(\theta_0, x,y)$, and $K'(x,dy) = k'(x,y) \allowbreak K(x,dy)$. (Note that $K'$ is a signed kernel, and not non-negative.)
Our hypotheses are stated in terms of $K(\theta)$, not $P(\theta)$, in order to offer the extra generality needed to cover settings in which derivatives involving parameters in the discount factor $\exp(g(\cdot))$ are of interest. Such derivatives are commonly considered in the finance literature when attempting to hedge uncertainty in the so-called ``short rate.'' (The resulting derivative is called \emph{rho} in the finance context.)
Finally, we also need to assume $\tilde f(\theta)$ is suitably differentiable at $\theta_0$. To permit derivatives in parameters that involve the discount factor, we write $\tilde f(\theta)$ in the form
\begin{equation}
\tilde f(\theta, x) = f(x) + \int_{C^c} K(\theta, x, dy) f(y).
\end{equation}
\begin{itemize}
\item[A3.]
The family of measures $(K(\theta, x, dy): \theta\in \Lambda, x \in C, y \in C^c)$ is absolutely continuous with respect to $(K(\theta_0, x, dy): x\in C, y \in C^c)$, in the sense that there exists a (measurable) density $(k(\theta, x,y): x\in C, y \in C^c)$ such that
\begin{equation*}
K(\theta, x, dy) = k(\theta, x, y) K(\theta_0, x, dy)
\end{equation*}
for $\theta \in \Lambda$, $x \in C$, $y\in C^c$. Furthermore, there exists $\epsilon>0$ such that for $x \in C$, $y\in C^c$, $k(\cdot, x,y)$ is continuously differentiable, with derivative $k'(x,y),$ in $[\theta_0-\epsilon, \theta_0 +\epsilon]$. Also, we assume that
\begin{equation*}
\tilde r_\epsilon(x) \triangleq \int_{C^c} \tilde \omega_\epsilon(x,y)|f(y)|K(\theta_0,x,dy)<\infty
\end{equation*}
for $x \in C$, where
\begin{equation*}
\tilde \omega_\epsilon(x,y) = \sup_{|\theta - \theta_0 | < \epsilon} |k'(\theta, x,y)|.
\end{equation*}
\end{itemize}
In many applications, $\tilde f(\theta)$ is independent of $\theta$ and A3 need not be verified (e.g. expected hitting times).
For $x\in C$, $y\in C^c$, set $K(x,dy)= K(\theta_0, x, dy)$ and $K'(x,dy)= k'(\theta_0, x,y) K(x,dy)$. We are now ready to state the main theorem of this section.
\begin{theorem}\label{thm:1}
Assume A1, A2, and A3. Suppose there exists $\epsilon > 0$ and two finite-valued non-negative functions $v_0$ and $v_1$ defined on $C$ for which
\begin{equation}\label{eq:2.7}
(K(\theta)v_0)(x) \leq v_0(x) - |\tilde f(\theta, x)|
\end{equation}
for $x\in C$ and $|\theta - \theta_0| < \epsilon$, and
\begin{equation}\label{eq:2.8}
(Kv_1)(x) \leq v_1(x) - \int_C \omega_\epsilon(x,y) v_0(y) K(x,dy) - \tilde r_\epsilon(x)
\end{equation}
for $x\in C$. Then, $u^*(\cdot,x)$ is differentiable at $\theta_0$ and
\begin{equation}\label{eq:2.9}
{u^*}'(\theta_0) = \int_C\int_C\int_C G(x,dy) K'(y,dz) G(z, dw) f(w) + \int_C \int_{C^c} G(x, dy) K'(y,dz) f(z).
\end{equation}
\what{%
If, in addition,
\begin{equation}\label{cond:random-horizon-C1-2}
\int_C \omega_\epsilon(x,y)v_1(y)K(x,dy)<\infty
\end{equation}
and (\ref{eq:2.8}) holds in a neighborhood of $\theta_0$, i.e., for $\theta\in[\theta_0-\epsilon, \theta_0+\epsilon]$
\begin{equation}\label{cond:random-horizon-C1-3}\tag{\ref*{eq:2.8}$'$}
(K(\theta)v_1)(x) \leq v_1(x) - \int_C \omega_\epsilon(x,y) v_0(y) K(\theta,x,dy) - \int_{C^c} \tilde \omega_\epsilon(x,y)|f(y)|K(\theta,x,dy)
\end{equation}
then ${u^*}'(\cdot,x)$ is continuous on $[\theta_0-\epsilon, \theta_0+\epsilon]$.
}
\end{theorem}
Recalling the definition of $G$, we see that (\ref{eq:2.9}) is indeed the general state space analog of (\ref{eq:2.5}). The functions $v_0$ and $v_1$ appearing in Theorem~\ref{thm:1} are often called (stochastic) Lyapunov functions. A standard means of guessing good choices for $v_0$ and $v_1$ is to recognize that $u^*(\theta_0)$ satisfies (\ref{eq:2.7}) with equality,
\rvin{if $\tilde f$ is non-negative}
while
\begin{equation*}
\int_C \left[\int_C K(y, dz) \omega_\epsilon(y,z) v_0(z) + r_\epsilon(y)\right] G(x,dy)
\end{equation*}
satisfies (\ref{eq:2.8}) with equality. When $C\subseteq \mathbb{R}^m$ is unbounded, one can often approximate the large $x$ behavior of these functions, and use these approximations as choices for $v_1$ and $v_2$, respectively.
The proof of Theorem~\ref{thm:1} rests on the following easy bound.
\begin{proposition}\label{prop:1}
Suppose that $Q=(Q(x,dy): x,y\in C)$ is a non-negative kernel and that $f: C\to \mathbb{R}_+$. If $v:C\to \mathbb{R}_+$ is a finite-valued function for which
\begin{equation}\label{eq:2.10}
Qv \leq v - f,
\end{equation}
then
\begin{equation}\label{eq:2.11}
\sum_{n=0}^\infty Q^n f \leq v.
\end{equation}
\end{proposition}
\begin{proof}{Proof.}
Note that (\ref{eq:2.10}) implies that $Qv \leq v$, and hence $Q^nv \leq v$ for $n\geq 0$. It follows that $Q^nv$ is finite-valued for $n\geq 0$. Inequality (\ref{eq:2.10}) can be re-written as
\begin{equation}\label{eq:2.12}
f \leq v - Qv.
\end{equation}
Applying $Q^j$ to both sides of (\ref{eq:2.12}), we get
\begin{equation}\label{eq:2.13}
Q^jf\leq Q^j v - Q^{j+1}v
\end{equation}
Summing both sides of (\ref{eq:2.13}) over $j = 0, 1, \ldots, n$, we find that
\begin{equation*}
\sum_{j=0}^n Q^j f \leq v - Q^{n+1} v \leq v.
\end{equation*}
Sending $n \to \infty$ yields (\ref{eq:2.11}).
\end{proof}
\begin{proof}{Proof of Theorem~\ref{thm:1}.}
For the purposes of this proof, $\epsilon$ is taken as the smallest of the $\epsilon$'s appearing in A2, A3, and the statement of the theorem. We start by observing that Proposition~\ref{prop:1}, applied to the Lyapunov bound (\ref{eq:2.7}), guarantees that
\begin{equation*}
\sum_{n=0}^\infty K^n(\theta) |\tilde f(\theta) | \leq v_0
\end{equation*}
and hence \rvout{(\ref{rep:u_star})} \rvin{Fubini's theorem} implies that $u^*(\theta)$ is finite-valued, $u^*(\theta) = \sum_{n=0}^\infty K^n(\theta)\tilde f(\theta)$, and $|u^*(\theta)|\leq v_0$. Since $u^*(\theta)$ is finite-valued (as is $K(\theta)u^*(\theta)$), we can write
\begin{equation*}
u^*(\theta_0+h) - u^*(\theta_0) = K(\theta_0+h) u^*(\theta_0+h) - K(\theta_0) u^*(\theta_0) + \tilde f(\theta_0 + h) - \tilde f(\theta_0)
\end{equation*}
and hence
\begin{equation}\label{eq:2.14}
(I-K) \big(u^*(\theta_0+h)- u^*(\theta_0)\big) = \big(K(\theta_0 + h)- K(\theta_0) \big) u^*(\theta_0 + h) + \big(\tilde f(\theta_0 + h) - \tilde f(\theta_0)\big).
\end{equation}
For $|h| < \epsilon$,
\begin{align*}
&\left| \int_C (K(\theta_0 + h, x, dy) - K(\theta_0, x,dy))u^*(\theta_0+h,y) \right|\\
&\leq \int_C | k(\theta_0 + h, x, y) - k(\theta_0 , x , y)| K(x,dy) v_0(y)\\
&\leq |h| \int_C \sup_{|\theta - \theta_0| < \epsilon} |k'(\theta, x,y) | K(x,dy) v_0(y)\\
&= |h| \int_C \omega_\epsilon(x,y) K(x,dy) v_0(y).
\end{align*}
Similarly, for $|h|<\epsilon$,
\begin{align*}
&|\tilde f(\theta_0 + h, x) - \tilde f(\theta_0, x)|\\
&\leq |h| \int_{C^c} \tilde \omega_\epsilon (x,y)K(x,dy)|f(y)|\\
&\leq |h| \tilde r_\epsilon (x).
\end{align*}
Consequently, Proposition~\ref{prop:1}, together with the Lyapunov bound (\ref{eq:2.8}), ensures that
\begin{equation*}
\int_C G(x,dy)\bigg( \left| \int_C (K(\theta_0 + h, y, dz) - K(\theta_0, y, dz) ) u^*(\theta_0+h, z) \right| +\left|\tilde f(\theta_0+h,y) - \tilde f(\theta_0, y)\right|\bigg) \leq |h|v_1(x).
\end{equation*}
It follows from (\ref{eq:2.14}) that $u^*(\theta,x)$ is continuous at $\theta_0$ and
\cf{
Is there a short justification? Referees may ask to justify.
}
\begin{align*}
\frac{u^*(\theta_0+h,x) - u^*(\theta_0,x)}{h}
&= \int_C G(x,dy) \left[\int_C \frac{k(\theta_0+h,y,z) - k(\theta_0, y,z)}{h} u^*(\theta_0+h,z)K(y,dz) \right.\\
&\qquad\qquad\quad\qquad+ \left.\int_{C^c} \frac{k(\theta_0 + h, y,z) - k(\theta_0, y,z)}{h} f(z) K(y,dz)
\right].
\end{align*}
But
\begin{equation}\label{eq:2.15}
\frac{k(\theta_0+h,y,z)-k(\theta_0, y, z)}{h} \to k'(y,z)
\end{equation}
and
\begin{equation}\label{eq:2.16}
u^*(\theta_0+h,z) \to u^*(\theta_0, z)
\end{equation}
as $h\to 0$. Also,
\begin{equation}\label{eq:2.17}
\left|\frac{k(\theta_0+h,y,z) - k(\theta_0,y,z)}{h}u^*(\theta_0+h,z)\right| \leq \omega_\epsilon(y,z) v_0(z)
\end{equation}
for $y,z\in C$, and
\begin{equation}\label{eq:2.18}
\frac{|k(\theta_0+h,y,z) - k(\theta_0,y,z)|}{h} \leq \tilde \omega_\epsilon(y,z)
\end{equation}
for $y\in C$, $z\in C^c$. The Lyapunov bound (\ref{eq:2.8}), together with Proposition~\ref{prop:1}, guarantees that
\begin{equation}\label{eq:2.19}
\int_C G(x,dy)\left(\int_C \omega_\epsilon(y,z) v_0(z) K(y,dz) + \int_{C^c} \tilde \omega_\epsilon (y,z) |f(z)| K(y,dz) \right) < \infty.
\end{equation}
In view of (\ref{eq:2.15}) through (\ref{eq:2.19}), the Dominated Convergence Theorem therefore establishes that $u^*(\theta, x)$ is differentiable at $\theta_0$, and
\begin{equation}\label{eq:sec3:u-star-prime-in-the-proof}
{u^*}'(\theta_0, x) = \int_C G(x,dy) \int_C k'(y,z) u^*(\theta_0, z) K(y,dz) + \int_C G(x,dy)\int_{C^c} k'(y,z)f(z) K(y,dz),
\end{equation}
which is equivalent to (\ref{eq:2.9}).
\what{%
Turning to the continuity of ${u^*}'(\cdot, x)$, note that one can easily check that
$$
{u^*}'(\theta) = \tilde f'(\theta) + K'(\theta)u^*(\theta) + K(\theta){u^*}'(\theta)
$$
for $\theta\in[\theta_0-\epsilon, \theta_0+\epsilon]$ where $K'(\theta)u^*(\theta,x) = \int_C k'(\theta, x,y) u^*(\theta,y)K(x,dy)$, and hence,
\begin{align*}
{u^*}'(\theta+h) - {u^*}'(\theta)
&=
G(\theta)
\big(
\tilde f'(\theta+h) - \tilde f'(\theta)
\big)
+
G(\theta)
\big(
(K'(\theta+h)-K'(\theta))u^*(\theta+h)
\big)\\
&\hspace{20pt}
+
G(\theta)
\big(
K'(\theta)(u^*(\theta+h)-u^*(\theta))
\big)
+
G(\theta)
\big(
(K(\theta+h)-K(\theta)){u^*}'(\theta+h)
\big).
\end{align*}
Now, a similar argument (via dominated convergence and the Lyapunov conditions) as the one that leads to (\ref{eq:sec3:u-star-prime-in-the-proof})---along with (\ref{cond:random-horizon-C1-2}), and (\ref{cond:random-horizon-C1-3})---shows that ${u^*}'(\theta+h) - {u^*}'(\theta) \to 0$ for $\theta\in[\theta_0-\epsilon, \theta_0+\epsilon]$.
}
\end{proof}
Our proof also yields the following (computable) bound on ${u^*}'(\theta_0)$, namely,
\begin{equation}\label{eq:2.20}
|{u^*}'(\theta_0, x) | \leq v_1(x)
\end{equation}
for $x \in C$.
In many applications, the parameter $\theta$ enters the dynamics in a very specific way, which allows further simplification of the result.
In particular, whenever $S$ is a separable metric space, we can always express $X$ as the solution to a stochastic recursion; see, for example,
\cite{kifer1986}. Namely, we can find a mapping $r:S\times S' \to S$ and a sequence $(Z_n:n\geq 1)$ of independent and identically distributed (iid) $S'$-valued random elements such that
\begin{equation}\label{eqA}
X_{n+1} = r(X_n, Z_{n+1})
\end{equation}
for $n\geq 0$. Suppose that $\theta$ affects the dynamics of $X$ only through the distribution of the $Z_n$'s. Assume that for $z\in S'$,
\begin{equation}\label{eqAB}
P^\theta(Z_1 \in dz) = p(\theta, z) P^{\theta_0} (Z_1\in dz),
\end{equation}
where $p(\cdot,z)$ is continuously differentiable for $z\in S'$. If $u^*(\theta, x)$ is defined as in (\ref{def:u_star}), then $u^*(\cdot, x)$ is differentiable at $\theta_0$ and ${u^*}'(\theta_0,x)$ is given by (\ref{eq:2.9}) (where $K'(x,dy) = \mathbf{E}^{\theta_0}\mathbb{I}(r(x,Z_1)\in dy)p'(\theta_0, Z_1)$), provided that there exists $\epsilon>0$ and finite-valued non-negative function $v_0$ and $v_1$ defined on $C\subseteq S$ for which
\begin{equation}\label{eqC}
\mathbf{E}^{\theta_0} v_0(r(x, Z_1)) p(\theta, Z_1) \leq v_0(x) - |\tilde f(\theta, x)|
\end{equation}
for $x \in C$ and $|\theta - \theta_0| < \epsilon$, and
\begin{align*}\label{eqD}
\mathbf{E}^{\theta_0} v_1 (r(x,Z_1))
\leq v_1 (x) &- \mathbf{E} ^{\theta_0}v_0(r(x,Z_1)) \sup_{|\theta - \theta_0|<\epsilon} |p'(\theta,Z_1)| \mathbb{I}(r(x,Z_1) \in C) \\
&- \mathbf{E}^{\theta_0} |f(r(x,Z_1)) |\sup_{|\theta - \theta_0|< \epsilon} |p'(\theta,Z_1)| \mathbb{I}(r(x,Z_1)\in C^c).
\end{align*}
for $x\in C$; the proof is essentially identical to that of Theorem~\ref{thm:1} and is omitted.
According to Theorem~\ref{thm:1}, for functions $f$ satisfying the Lyapunov bound,
\begin{equation*}
{u^*}'(\theta_0, x) = \int_S \nu'(x,dy)f(y)
\end{equation*}
where
\begin{align*}
\nu'(w,dz) = \begin{cases}
\int_C G(w,dx)\int_C K'(x,dy) \int_C G(y,dz), & w,z\in C\\
\int_C G(w,dx)\int_{C^c}K'(x,dz), & w\in C, z \in C^c.
\end{cases}
\end{align*}
Hence, our derivative can be represented in terms of a signed measure. (In general, $\nu'(x,S)$ is non-zero in this setting.)
The above approach also extends, in a straightforward way, to higher-order derivatives. Formal differentiation of (\ref{eq:u_star}) $n$ times yields the identity
\begin{equation*}
u^{*(n)}(\theta) = \tilde f^{(n)} (\theta) + \sum_{j=0}^n \binom{n}{j} K^{(n-j)}(\theta) u^{*(j)}(\theta),
\end{equation*}
which suggests that the $n$\textsuperscript{th} order derivative $u^{*(n)}(\theta)$ can then be recursively computed from $u^{*(0)}(\theta)$, \ldots, $u^{*(n-1)}(\theta)$ by solving the linear (integral) equation
\begin{equation}
(I-K(\theta))u^{*(n)} (\theta)
= \rvin{\tilde f^{(n)}(\theta)} + \sum_{j=0}^{n-1}\binom{n}{j} K^{(n-j)} (\theta) u^{*(j)}(\theta).
\end{equation}
In particular, it should follow that
\begin{equation}\label{eq:2.21}
u^{*(n)}(\theta) = G\left(\tilde f^{(n)}(\theta)+\sum_{j=0}^{n-1}\binom{n}{j} K^{(n-j)}(\theta) u^{*(j)}(\theta)\right).
\end{equation}
Rigorous verification of (\ref{eq:2.21}) can be implemented with a family $v_0, v_1, \ldots, v_n$ of Lyapunov functions. Specifically assume that the densities $k(\cdot, x, y)$ (for $x\in S, y \in S$) are $n$-times continuously differentiable in some neighborhood $[\theta_0-\epsilon, \theta_0 + \epsilon]$ of $\theta_0$, and set
\begin{equation*}
\omega_\epsilon^{(j)}(x,y) = \sup_{|\theta-\theta_0|<\epsilon} |k^{(j)}(\theta, x, y)|
\end{equation*}
for $x,y\in C$ and
\begin{equation*}
\tilde \omega_\epsilon^{(j)}(x,y) = \sup_{|\theta-\theta_0|<\epsilon} |k^{(j)}(\theta, x, y)|
\end{equation*}
for $x\in C$, $y\in C^c$.
\begin{theorem}\label{thm:2}
Suppose that there exists $\epsilon > 0$ and a family of finite-valued non-negative functions $v_0, v_1, \ldots, v_n$ defined on $C$ for which
\begin{equation*}
(K(\theta)v_0)(x) \leq v_0(x) - |\tilde f(\theta,x)|
\end{equation*}
for $x\in C$ and $|\theta - \theta_0| < \epsilon$;
\begin{align*}
(K(\theta)v_l)(x) \leq v_l(x) - \sum_{j=0}^{l-1} \binom{l}{j} \int_C \omega_\epsilon^{(l-j)} (x,y) v_j(y)K(\theta,x,dy) - \int_{C^c} \tilde \omega_\epsilon^{(l)}(x,y) |f(y)| K(\theta,x,dy)
\end{align*}
for $x\in C$, $|\theta - \theta_0| < \epsilon$, and $1\leq l\leq n$; \what{and
$$
\int_C \omega_\epsilon^{(n)}v_n(y)K(x,dy) < \infty
$$
for $x\in C.$
}
Then, $u^*(\cdot, x)$ is $n$-times continuously differentiable at $\theta_0$, and the derivative can be recursively computed from the equations
\begin{align*}
u^{*(l)}(\theta_0,x)
&= \int_C G(x,dy)
\int_C\sum_{j=0}^{l-1}\binom{l}{j} k^{(l-j)} (\theta_0, x, y) u^{*(j)}(y)K(x,dy)\\
&\qquad\quad+\int_C G(x,dy)\int_{C^c} k^{(l)}(y,z)f(z) K(y,dz)
\end{align*}
\end{theorem}
The proof of Theorem~\ref{thm:2} mirrors that of Theorem~\ref{thm:1}, and is therefore omitted. As in the proof of Theorem~\ref{thm:1}, the argument establishes the bound $|u^{*(n)}(\theta_0, x)| \leq v_n(x)$ for $x\in C$ on the $n$\textsuperscript{th} order derivative.
\section{Lyapunov Criteria for Differentiability of Stationary Expectations}\label{sec:equilibrium}
Perhaps the most commonly occurring expectations that arise in applications are those associated with steady-state behavior. Our Lyapunov approach is also well-suited to establishing differentiability in this context. As in Section~\ref{sec:random_horizon}, it is informative to first study the problem non-rigorously.
A stationary distribution $\pi(\theta) = (\pi(\theta, dx): x \in S)$ of the Markov chain $X$ associated with one-step transition kernel $P(\theta)$ will satisfy
\begin{equation}\label{eq:3.1a}
\pi(\theta) = \pi(\theta)P(\theta).
\end{equation}
Differentiating both sides of (\ref{eq:3.1a}) with respect to $\theta$, we obtain
\begin{equation*}
\pi'(\theta) = \pi'(\theta)P(\theta) + \pi(\theta)P'(\theta),
\end{equation*}
which leads to the equation
\begin{equation*}
\pi'(\theta) (I-P(\theta))=\pi(\theta)P'(\theta).
\end{equation*}
This equation is similar to (\ref{eq:2.3a}). However, unlike (\ref{eq:2.3a}), the operator $I-P(\theta)$ appearing here will never be invertible, even when $|S|<\infty$. In addition, $I - P(\theta)$ is acting on a measure rather than a function in this setting. Thus, a different approach is needed here.
For a given function $f: S\to \mathbb{R}$, set $\alpha(\theta) = \pi(\theta) f$. Thus,
\begin{align}
\alpha(\theta_0 + h) - \alpha (\theta_0)
&= \pi(\theta_0+h)f-\pi(\theta_0)f\nonumber\\
&= \pi(\theta_0+h)f_c,\label{eq:3.1}
\end{align}
where $f_c(x) = f(x) - \pi(\theta_0) f$. While $I-P(\theta_0)$ is singular, the \emph{Poisson's equation}
\begin{equation}\label{eq:3.2}
(I-P(\theta_0))g = f_c
\end{equation}
is, under suitable technical conditions, generally solvable for $g$ (because of the special structure of the right-hand side, namely $\pi(\theta_0)f_c= 0$). Substituting (\ref{eq:3.2}) into (\ref{eq:3.1}), we get
\begin{align}
\alpha(\theta_0 + h) - \alpha(\theta_0)
&= \pi(\theta_0 + h) (I-P(\theta_0))g\nonumber\\
&= \pi(\theta_0 + h) (P(\theta_0+h) - P(\theta_0))g.\label{eq:3.3a}
\end{align}
This suggests that
\begin{equation}\label{eq:3.3}
\alpha'(\theta_0) = \pi(\theta_0) P'(\theta_0)g.
\end{equation}
We now turn to making this argument rigorous.
We start by assuming that $(P(\theta): \theta \in \Lambda)$ itself satisfies the absolute continuity condition:
\begin{itemize}
\item[A4.]
The family of one-step transition kernels $(P(\theta): \theta \in \Lambda)$ is absolutely continuous with respect to $P(\theta_0)$, in the sense that there exists a density $(p(\theta, x, y): \theta\in \Lambda, x,y \in S)$ for which
\begin{equation*}
P(\theta, x, dy) = p(\theta, x, y) P(\theta_0, x, dy)
\end{equation*}
for $x, y \in S$, and $\theta\in \Lambda$. Furthermore, there exists $\epsilon > 0$ for which $p(\cdot, x, y)$ is continuously differentiable on $[\theta_0-\epsilon, \theta_0 + \epsilon]$ for each $x,y\in S$.
\end{itemize}
Set $\omega_\epsilon(x,y) = \sup_{|\theta-\theta_0|<\epsilon}|p'(\theta,x,y)|.$ Our next assumption involves a (uniform) minorization condition over the set $A$, which is standard in the theory of Harris recurrent Markov chains; see, for example, p.102 of \cite{meyn09}
\begin{itemize}
\item[A5.]
There exists $\epsilon > 0$, a subset $A\subseteq S$, an integer $n\geq 1$, $\lambda>0$, and a probability $\varphi$ for which
\begin{equation*}
P^n(\theta, x, dy) \geq \lambda \varphi(dy)
\end{equation*}
for $x\in A$, $y\in S$, and $|\theta-\theta_0| < \epsilon$.
\end{itemize}
For $a, b\in \mathbb{R}$, let $a\vee b \triangleq \max(a,b)$. We can now state our main theorem on differentiability of stationary expectations.
\begin{theorem}\label{thm:3}
Assume that A4 and A5 hold.
Let $\kappa: \mathbb{R}_+ \to \mathbb{R}_+$ be a function for which $\kappa(x)\geq x$ and $\kappa(x)/x \to\infty$ as $x \to \infty$.
Suppose that there exist positive constants $\epsilon$, $c_0$, and $c_1$, and non-negative finite-valued functions $q$, $v_0$, and $v_1$ for which
\begin{align}
(P(\theta)v_0)(x) &\leq v_0(x) - (\rvin{q(x)}\vee 1) + c_0 \mathbb{I}(x\in A), \label{eq:3.4}\\
(P(\theta)v_1)(x)
&\leq
v_1(x) -
\kappa\left( \int_S ( 1 \vee \omega_\epsilon(x,y) ) (v_0(y) + 1) P(\theta, x, dy)\right)
+ c_1 \mathbb{I}(x \in A),
\label{eq:3.5}
\end{align}
for $x\in S$, $|\theta-\theta_0|<\epsilon$, and
\begin{equation}\label{bound:sup_v_0}
\sup_{x\in A} v_0(x) < \infty
\end{equation}
Then:
\begin{itemize}
\item[(i)]
There exists an open interval $\mathcal N$ containing $\theta_0$ for which $X$ is a positive recurrent Harris chain under $P(\theta)$ for each $\theta\in \mathcal N$;
\item[(ii)]
There exists a unique stationary distribution $\pi(\theta)$ satisfying $\pi(\theta) = \pi(\theta)P(\theta)$ for each $\theta\in \mathcal N$ and $\pi(\theta)\rvin{q} \leq c_0$ for $\theta\in \mathcal N$;
\item[(iii)]
For each $f$ such that $|f(x)|\leq q(x) \vee 1$ for $x\in S$, there exists a solution $g$ (denoted $g=\Gamma f$) of Poisson's equation satisfying
\begin{equation*}
((I-P(\theta_0))g)(x) = f(x) - \pi(\theta_0) f
\end{equation*}
for $x\in S$, and $|g(x)| = |(\Gamma f)(x)| \leq a(v_0(x) + 1)$ for $x \in S$, where $a$ is a finite constant;
\item[(iv)]
For each $f$ such that $|f(x)| \leq q(x) \rvin{\vee} 1$, $\alpha(\theta) = \pi(\theta)f$ is \what{continuously} differentiable at $\theta_0$, and
\begin{equation}\label{C}
\alpha'(\theta_0) = \int_S \pi(\theta_0, dx) \int_S p'(\theta_0, x,y) (\Gamma f) (y) P(\theta_0, x, dy).
\end{equation}
\end{itemize}
\end{theorem}
\begin{proof}{Proof.}
It is a standard fact that A5, (\ref{eq:3.4}), and (\ref{bound:sup_v_0}) imply that $X$ is a positive recurrent Harris chain under $P(\theta)$ for $\theta \in \mathcal N$ (where $\mathcal N$ is selected so that A5, (\ref{eq:3.4}) and (\ref{bound:sup_v_0}) are all in force); see, for example, \citet[p.313]{meyn09}. As a consequence, there exists a unique stationary distribution $\pi(\theta)$ for each $\theta \in \mathcal N$. Furthermore, (\ref{eq:3.4}) implies that the bound $\pi(\theta)\rvin{q}\leq c_0$ holds for $\theta\in \mathcal N$; see, for example, Corollary 4 of \cite{glynn08}. Because $X$ is Harris recurrent (and (\ref{eq:3.4}) holds), one can now invoke Theorem 2.3 of \cite{glynn96} to obtain (iii).
Turning to (iv), note that (\ref{eq:3.5}) guarantees that $\pi(\theta)\rvin{v_0} < \infty $ for $\theta\in \mathcal N$, so that $\pi(\theta)|\Gamma f| < \infty$.
\cf{
Note that $\pi|f|<\infty$ is equivalent to $(\pi P)|f|<\infty$ and hence conditional Fubini applies: $\pi (P f) = (\pi P) f = \pi f$. Therefore,
$
\pi (Pf-f) = 0.
$
This validates (\ref{eq:3.3a}).}%
With the above conclusions having been verified, we can now appeal to (\ref{eq:3.3a}) to write
\begin{align}
\pi(\theta_0+h) f - \pi(\theta_0) f
&= \pi(\theta_0 + h) \big( P(\theta_0+h) - P(\theta_0)\big) \Gamma f\nonumber\\
&=\int_S \pi(\theta_0 + h, dx) \int_S \big(P(\theta_0+h, x, dy) - P(\theta_0, x, dy)\big)(\Gamma f)(y).\label{eq:3.6}
\end{align}
Set $s(x) = \int_S \omega_\epsilon(x,y) (v_0(y) + 1) P(\theta_0, x, dy)$ and put $\mathbb{I}_m(x) = \mathbb{I}(s(x) \geq m)$, $\mathbb{I}_m^c(x) = \mathbb{I}(s(x) < m)$.
\cf{ Then $s(x)$ bound the finite difference
$$\frac{P_h-P}{h}\Gamma f \leq a s$$
and from Glynn and Zeevi (2008) and (\ref{eq:3.5})
\begin{align*}
\pi(\theta) s
&\leq \int\pi(\theta,dx) \int(\omega_\epsilon(x,y)\vee 1)(v_0(y) + 1) P(\theta_0, x, dy)\\
&\leq \int\pi(\theta,dx) \int\kappa\big((\omega_\epsilon(x,y)\vee 1)(v_0(y) + 1)\big) P(\theta_0, x, dy)\\
&\leq c_1
\end{align*}
}
Observe that since $|p(\theta_0+h, x, y) - p(\theta_0, x, y) |/h \leq \omega_\epsilon(x,y)$, and $\rvin{|}(\Gamma f)(y)\rvin{|} \leq a(v_0(y) +1 )$,
\begin{align}
&\int_S \pi(\theta_0 + h, dx) \mathbb{I}_m(x) \rvin{\bigg|}\left(\frac{P(\theta_0+h)-P(\theta_0)}{h} (\Gamma f)\right)(x)\rvin{\bigg|}\nonumber\\
&\leq \int_S \pi(\theta_0 + h, dx) \mathbb{I}_m(x) \int_S \omega_\epsilon(x,y) \,a(v_0(y)+1) P(\theta_0, x, dy)\nonumber\\
&\leq a\int_S \pi(\theta_0 + h, dx) \mathbb{I}_m(x) s(x)\nonumber\\
&\leq \frac{a}{\inf\{ \frac{\kappa(s(y))}{s(y)}: s(y)\geq m)\}} \int_S \pi(\theta_0+h, dx) \frac{\kappa(s(x))}{s(x)}s(x)\nonumber\\
&\leq \frac{a}{\inf\{ \frac{\kappa(s(y))}{s(y)}: s(y)\geq m)\}}\int_S \pi(\theta_0+h,dx) \kappa\left(\int_S (1\vee\omega_\epsilon(x,y))(v_0(y) + 1) P(\theta_0, x, dy)\right)\nonumber\\
&\leq \frac{a}{\inf\{ \frac{\kappa(s(y))}{s(y)}: s(y)\geq m)\}} c_1 \label{eq:3.7}
\end{align}
where the last inequality follows from (\ref{eq:3.5}) and Corollary 4 of \cite{glynn08}.
On the other hand,
\begin{align*}
\int_S \pi(\theta_0+h,dx) \mathbb{I}_m^c (x) \frac{P(\theta_0+h)-P(\theta_0)}{h}(\Gamma f) (x)
\triangleq \int_S \pi(\theta_0+h, dx) s_h^m(x) = \pi(\theta_0+h)s_h^m,
\end{align*}
where
\begin{align*}
|s_h^m(x) | \leq a\int_S \omega_\epsilon(x,y) (v_0(y)+1) P(\theta_0, x, dy)\mathbb{I}(s(x) < m) \leq am,
\end{align*}
so $s_h^m$ is bounded. It follows that
\begin{equation*}
\pi(\theta_0 + h) s_h^m - \pi(\theta_0)s_h^m = \pi(\theta_0 + h) (P(\theta_0+h)-P(\theta_0))(\Gamma s_h^m).
\end{equation*}
\cf{From (\ref{eq:3.4})
$$\Gamma 1 \leq a(v_0+1)$$
and hence
$$\Gamma s_h^m \leq (am) (a(v_0+1))$$
}
Note that $\left|\frac{\Gamma s_h^m}{am}\right| \leq a(v_0+1)$ (because $\left| \frac{s_h^m}{am} \right|\leq q \vee 1$),
and hence
\begin{align}
|\pi(\theta_0 + h) s_h^m - \pi(\theta_0) s_h^m|
&\leq a^2 m|h| \int_S \pi(\theta_0+h,dx) \int_S \omega_\epsilon(x,y) P(\theta_0, x,dy) (v_0(y) + 1)\nonumber\\
&\leq a^2 m|h| \int_S \pi(\theta_0 + h, dx) s(x)\nonumber\\
&\leq a^2 m|h| c_1\to 0 \label{eq:3.9}
\end{align}
as $h\to 0$. Finally,
\begin{equation*}
\int_S \pi(\theta_0, dx) s_h^m(x) = \int_S \pi(\theta_0, dx) \mathbb{I}_m^c(x)\int_S \frac{p(\theta_0+h, x, y) - p(\theta_0, x, y)}{h} P(\theta_0, x, dy) (\Gamma f) (y)
\end{equation*}
and
\begin{equation*}
\frac{p(\theta_0+h, x, y) - p(\theta_0, x, y)}{h} \to p'(\theta_0, x, y)
\end{equation*}
as $h \searrow 0$. Furthermore, $|p(\theta_0+h, x, y) - p(\theta_0, x, y) |/h \leq \omega_\epsilon(x,y)$, $(\Gamma f)(y) \leq a(v_0(y) +1 )$, and
\begin{align*}
\int_S \pi(\theta_0, dx) \int_S \omega_\epsilon(x,y) P(\theta_0, x, dy) (v_0(y) + 1) \leq \int_S \pi(\theta_0, dx) s(x) \leq c_1,
\end{align*}
so the Dominated Convergence Theorem implies that
\begin{equation}\label{eq:3.10}
\int_S \pi(\theta_0, dx) s_h^m(x) \to \int_S \pi(\theta_0, dx) \mathbb{I}_m^c(x) \cdot \int_S p'(\theta_0, x, y) P(\theta_0, x, dy) (\Gamma f) (y)
\end{equation}
as $h\searrow 0$.
If we first let $h\to 0$ and then let $m \to \infty$, (\ref{eq:3.6}) through (\ref{eq:3.10}) imply part (iv) of our theorem.
\what{
Finally, turning to the continuity of the derivative,
note that the exactly same argument as above gives
$
\alpha'(\theta_0+h) = \int_S \pi(\theta_0+h, dx) p'(\theta_0+h,x,y) \Gamma_{\theta_0+h} f(x) P(\theta_0,x,dy)
$
where $\Gamma_{\theta_0+h}f$ is the solution $g$ of the Poisson equation $g - P(\theta_0+h)g = f - \pi(\theta_0+h) f$.
Since
$$
\alpha'(\theta_0+h) - \alpha'(\theta_0)
=
\left(\alpha'(\theta_0+h) - \frac{\alpha((\theta_0+h)+(-h)) - \alpha(\theta_0+h)}{-h}\right) - \left(\alpha'(\theta_0) - \frac{\alpha(\theta_0+h)-\alpha(\theta_0)}{h}\right),
$$
and we have seen that the second term vanishes as $h\to 0$, we are done if we show that the first term also vanishes.
Similarly as in (\ref{eq:3.3a}),
$
\alpha(\theta_0) - \alpha(\theta_0+h) = \pi(\theta_0) (P(\theta_0)-P(\theta_0+h)) \Gamma_{\theta_0+h}f
$.
Therefore,
\begin{align}
&\alpha'(\theta_0+h) - \frac{\alpha((\theta_0+h)+(-h)) - \alpha(\theta_0+h)}{-h}
=
\alpha'(\theta_0+h) + \frac{\alpha(\theta_0) - \alpha(\theta_0+h)}{h}
\nonumber
\\
&
=
\int_S \pi(\theta_0+h, dx) p'(\theta_0+h,x,y) \Gamma_{\theta_0+h} f(x) P(\theta_0,x,dy)
\nonumber
\\
&
\qquad
+
\int_S \pi(\theta_0, dx) \frac{p(\theta_0,x,y) - p(\theta_0+h,x,y)}{h} \Gamma_{\theta_0+h} f(x) P(\theta_0,x,dy)
\nonumber
\\
&
=
\int_S \big(\pi(\theta_0+h, dx) - \pi(\theta_0, dx)\big) p'(\theta_0+h,x,y) \Gamma_{\theta_0+h} f(x) P(\theta_0,x,dy)
\label{eq:first_term_pf_t41}
\\
&
\qquad
+
\int_S \pi(\theta_0, dx) \left(p'(\theta_0+h,x,y)+\frac{p(\theta_0,x,y) - p(\theta_0+h,x,y)}{h}\right) \Gamma_{\theta_0+h} f(x) P(\theta_0,x,dy).
\label{eq:second_term_pf_t41}
\end{align}
Upon a perusal of the proof of Theorem 2.3 of \cite{glynn96}, one can see that the uniform majorization condition A5 and the uniform Lyapunov inequality (\ref{eq:3.4}) implies $|\Gamma_{\theta_0+h} f(x)| \leq a(v_0(x)+1)$ with the same constant $a$ as in (iii). One can prove that
(\ref{eq:first_term_pf_t41}) vanishes as $h\to 0$ by the same argument as (\ref{eq:3.7}) and (\ref{eq:3.9}).
On the other hand,
(\ref{eq:second_term_pf_t41}) vanishes by the continuous differentiability condition A4 of $p$ and the dominated convergence along with (\ref{eq:3.5}).
}
\end{proof}
As for Theorem~\ref{thm:1}, the proof also establishes a computable bound on $|\alpha'(\theta_0)|$, namely $|\alpha'(\theta_0)|\leq \rvin{a} c_1$ \rvin{where $a$ is the constant in (iii)}.
\cf{maybe we want to explain how to compute $a$ or at least give a formula?}
Also, as in Section~\ref{sec:random_horizon}, we can further simplify the condition when $X$ is the solution to the stochastic recursion (\ref{eqA}), in which the parameter $\theta$ affects only the distribution $Z_1$. When $p(\cdot,z)$ is continuously differentiable, (\ref{eq:3.5}) may be simplified as
\begin{equation}
(P(\theta)v_1) (x) \leq v_1(x) - \kappa\left(\mathbf{E} ^{\theta_0}\Big(1\vee\sup_{|\theta-\theta_0|<\epsilon}|p'(\theta,Z_1) |\Big)\big(v_0(r(x,Z_1))+1\big)p(\theta,Z_1)\right) + c_1 \mathbb{I}(x\in A).
\end{equation}
With A5, (\ref{eq:3.4}), and (\ref{bound:sup_v_0}) also in force, this ensures the differentiability of $\alpha(\cdot)$ at $\theta_0$, with $\alpha'(\theta_0)$ given by
\begin{equation}
\alpha'(\theta_0) = \int_S \pi(\theta_0, dx) \mathbf{E}^{\theta_0} (\Gamma f) (r(x,Z_1))p'(\theta_0,Z_1).
\end{equation}
A useful example on which to illustrate the above theory (and an important model in its own right) is that of the waiting time sequence $W=(W_n: n\geq 0)$ for the single-server G/G/1 queue, with first come first serve queue discipline. Let $V_n$ be the arrival time for the $\rvin{n}$\textsuperscript{th} customer, and let $\chi_{n+1}$ be the inter-arrival time that elapses between the arrival of the $\rvin{n}$\textsuperscript{th} and $(n+1)$\textsuperscript{st} customer. If $W_n$ is the waiting time (exclusive of service) for customer $n$, the $W_n$'s satisfy the stochastic recursion
\begin{equation}
W_{n+1} = [W_n + V_n - \chi_{n+1}]^+
\end{equation}
for $n\geq 0$, where $[x]^+ \triangleq \max(x,0).$ Assume that the $V_n$'s are iid, independent of the $\chi_n$'s (which are also assumed iid). Then, $W$ is a Markov chain taking values in $S = [0,\infty)$. It is well known that $W$ is a positive recurrent Harris chain if $\mathbf{E} V_0 < \mathbf{E} \chi_1$, and that $\mathbf{E} V_0^{p+1}\rvin{<\infty}$ is then a necessary and sufficient condition for guaranteeing the finiteness of $\pi f_p$, where $f_p(x) = x^p$ (with $p>0$);
see, for example, \cite{kiefer1956}.
This suggests that it then typically will be the case that the $p$\textsuperscript{th} moment should be differentiable when $\mathbf{E} V_0 ^{p+1}< \infty$.
We consider this problem in the special case in which the service times are finite mean Pareto random variables (rv's), and $\theta$ influences the scale parameter of the Pareto distribution. In other words, we consider the setting in which
\begin{equation*}
P^\theta(V_0 > v) = (1+\theta v)^{-\alpha}
\end{equation*}
for $\alpha > 1$.
\cf{$\mathbf{E}^\theta v = \frac{1}{\theta}\frac{1}{\alpha-1}$}
In this case, the density of $V_0$ under $P^\theta$ is given by $\theta h_V(\theta v),$ where $h_V(v) = \alpha(1+v)^{-\alpha -1}$, so that
\begin{equation*}
p(\theta, v) = \left(\frac{\theta}{\theta_0}\right) \left(\frac{1+\theta v}{1+\theta_0 v}\right)^{-\alpha-1}
\end{equation*}
and
\begin{equation*}
p'(\theta, v) = p(\theta, v) \left(\frac{1}{\theta} - (\alpha + 1)\frac{v}{(1 + \theta v)}\right).
\end{equation*}
Note that both the density $p$ and its derivative (with respect to $\theta$) are bounded functions. Furthermore, the rv $p'(\theta_0, V_i)$ has mean zero under $P^{\theta_0}$. For any $c>0$, the set $A=[0,c]$ is easily seen to satisfy condition A5, and A4 is trivially verified (with $\omega_\epsilon(\cdot)$ bounded). Then, if $v_0(x) = a_1 x^{p+1}$, $v_1(x) = a_2 x^{r+2}$, and $\kappa(x)=x^{\frac{1+r}{1+p}}$ (with $r>p$ and $a_1$, $a_2$ chosen suitably), we see that (\ref{eq:3.4}), (\ref{eq:3.5}), and (\ref{bound:sup_v_0}) all hold, guaranteeing the differentiability of $\pi(\theta)f_p$ (according to Theorem~\ref{thm:3}).
For example, to verify (\ref{eq:3.4}), we note that
\begin{equation*}
x^{-p}\big[(P(\theta)v_0)(x) - v_0(x)\big] = a_1 x \mathbf{E} ^{\theta_0} \left(\left[1+\frac{1}{x} \left(\frac{\theta_0}{\theta} V_0 - \chi_1\right)\right]^+\right)^{p+1} - a_1 x.
\end{equation*}
Observe that as $x \to \infty$,
\begin{align*}
&x f_{p+1}\left(\left[1+\frac 1 x \left(\frac{\theta_0}{\theta} V_0 - \chi_1\right)\right]^+\right) - x\\
&= x \left(f_{p+1}(1) + f_{p+1}'(1)\left(\frac{1}{x}\right)\left(\frac{\theta_0}{\theta} V_0 - \chi_1\right)\right) - x + o(1) \qquad a.s.\\
&= \rvin{(}p\rvin{+1)}\left(\frac{\theta_0}{\theta} V_0 - \chi_1\right) + o(1) \qquad a.s.
\end{align*}
\rvout{as $x \to \infty$,} where $o(1)$ represents a function $k(x)$ such that $k(x) \to 0$ as $x \to \infty$ \rvin{uniformly in a neighborhood of $\theta_0$}.
In addition, note that for $p>0$ and $x>0$, the mean value theorem implies that $f_{p+1}(1+x) = f_{p+1}(1) + f'_{p+1}(1+\xi)x$ for some $\xi \in [0,x]$, so that $f_{p+1}(1+x) = f_{p+1}(1) + (p+1) (1+\xi)^{p} x \leq 1 + (p+1)(1+x)^px$. Consequently,
\begin{align*}
\hcancel{
&\left|\
x\left(
\left[
1+\frac{1}{x}
\left(
\frac{\theta_0}{\theta} V_0 - \chi_1
\right)
\right]^+
\right)^{p+1} - x\
\right|
\
\\
}
x\left(
\left[
1+\frac{1}{x}
\left(
\frac{\theta_0}{\theta} V_0 - \chi_1
\right)
\right]^+
\right)^{p+1} - x\
\
\\
&\leq x
\left(
1+\frac 1 x\frac{\theta_0}{\theta} V_0
\right)^{p+1} - x\\
&\leq x\left(1 + (p+1) \left(1 + \frac{1}{x}\frac{\theta}{\theta_0}V_0\right)^p \frac{1}{x}\frac{\theta}{\theta_0} V_0\right) - x\\
&\hcancel{
\leq (p+1) \left(1 + \frac{1}{x}\frac{\theta}{\theta_0}V_0\right)^{p+1}
}\\
&\leq (p+1) \left(1 + \frac{1}{x}\frac{\theta}{\theta_0}V_0\right)^{p}\frac{\theta}{\theta_0}V_0
\end{align*}
\cf{
\rvin{
and
\begin{align*}
&
x-
x\left(
\left[
1+\frac{1}{x}
\left(
\frac{\theta_0}{\theta} V_0 - \chi_1
\right)
\right]^+
\right)^{p+1}
\\
&\leq x - x\left(
\left[
1-\frac{\chi_1}{x}
\right]^+
\right)^{p+1} \\
&\leq x - x\left(
1-\frac{\chi_1}{x}
\right)^{p+1}
\end{align*}
}
}
\cf{For $p \in (-1,0)$, mean value theorem implies
$f_{p+1}(1+x) = f_{p+1}(1) + f'_{p+1}(1+\xi)x$ for some $\xi \in [0,x]$, so that $f_{p+1}(1+x) = f_{p+1}(1) + (p+1) (1+\xi)^{p} x \leq 1 + (p+1)x$. Consequently,
\begin{align*}
&
x\left(
\left[
1+\frac{1}{x}
\left(
\frac{\theta_0}{\theta} V_0 - \chi_1
\right)
\right]^+
\right)^{p+1} - x\
\\
&\leq x
\left(
1+\frac 1 x\frac{\theta_0}{\theta} V_0
\right)^{p+1} - x\\
&\leq x\left(1 + (p+1) \frac{1}{x}\frac{\theta}{\theta_0} V_0\right) - x\\
&\leq (p+1)\frac{\theta}{\theta_0}V_0
\end{align*}
and
\begin{align*}
&
x-
x\left(
\left[
1+\frac{1}{x}
\left(
\frac{\theta_0}{\theta} V_0 - \chi_1
\right)
\right]^+
\right)^{p+1}
\\
&\leq x - x\left(
\left[
1-\frac{\chi_1}{x}
\right]^+
\right)^{p+1}
\leq x - x\left(
\left[
1-\frac{\chi_1}{x}
\right]^+
\right)
\leq x - x\left(
1-\frac{\chi_1}{x}
\right)
\leq \chi_1
\end{align*}
}
Since $\mathbf{E} V_0^{p+1}< \infty$, \rvout{the Dominated Convergence theorem ensures} \rvin{Fatou's lemma applies to ensure} that
\hcancel{
\begin{equation*}
\mathbf{E}^{\theta_0} \left(x f_{p+1} \left(\left[1+\frac{1}{x}\left(\frac{\theta_0}{\theta}V_0 - \chi_1\right)\right]^+\right)-x\right) \to (p+1)\mathbf{E} ^{\theta_0} \left(\frac{\theta_0}{\theta}V_0 - \chi_1\right) = (p+1)\left(\frac1 {\theta\rvin{(\alpha - 1)}} - \mathbf{E} \chi_1\right)
\end{equation*}
}
\rvin{
\begin{align*}
\limsup_{x\to\infty}\,\sup_{\theta} \mathbf{E}^{\theta_0} \left(x f_{p+1} \left(\left[1+\frac{1}{x}\left(\frac{\theta_0}{\theta}V_0 - \chi_1\right)\right]^+\right)-x\right)
&\leq (p+1)\mathbf{E} ^{\theta_0}\sup_{\theta} \left(\frac{\theta_0}{\theta}V_0 - \chi_1\right) \\
&= (p+1)\sup_{\theta}\left(\frac1 {\theta(\alpha - 1)} - \mathbf{E} \chi_1\right)
\end{align*}
}
as $x\to\infty$ (with convergence that is uniform in a neighborhood of $\theta_0$). If we choose $a_1$ so that $a_1 \rvin{(}p\rvin{+1)\sup_{\theta}}(\frac{1}{\theta\rvin{(\alpha - 1)}} - \mathbf{E} \chi_1) \leq -2$ \rvout{(uniformly in $\theta$)} and $c$ so that
\begin{equation*}
a_1\rvin{\sup_{\theta}}\mathbf{E}^{\theta_0} \left(xf_{p+1}\left(\left[1+\frac{1}{x}\left(\frac{\theta_0}{\theta}V_0 - \chi_1\right)\right]^+\right)-x\right) \leq -1
\end{equation*}
for $x\geq c$, then (\ref{eq:3.4}) is validated. A similar argument applies to (\ref{eq:3.5}), in view of the boundedness of $\omega_\epsilon(\cdot)$. Our argument therefore establishes that $\pi f_p$ is differentiable if $\mathbf{E} V_0 ^q < \infty$ for some $q>p+2$. This is not quite the ``correct'' result (in that we previously argued that $\mathbf{E} V_0^{p+1}<\infty$ should be sufficient.)
The reason that our argument fails to provide optimal condition here has to do with special random walk structure that is present in the process $W$ that is difficult for general machinery to exploit. The challenge arises at (\ref{eq:3.3a}) above. Note that the argument just provided for $W$ involves using $v_0=a_1 f_{p+1}$ as a bound on the solution $g$ to Poisson's equation for $f_p$. (As we shall see in a moment, $g$ is indeed exactly of order $x^{p+1}$). The problem is that neither $P(\theta_0+h) f_{p+1}$ nor $P(\theta_0)f_{p+1}$ in (\ref{eq:3.3a}) are integrable with respect to $\pi(\theta_0+h)$ unless $\mathbf{E} V_0 ^{p+2}<\infty$. This is what leads to the extra moment appearing in our argument for $W$ above. Thus, any argument that yields differentiability under the hypothesis $\mathbf{E} V_0^{p+1}<\infty$ must take advantage of the fact that the random walk structure of $W$ yields the integrability of $(P(\theta_0 + h) - P(\theta_0))g$ under $\mathbf{E} V_0^{p+1}<\infty$ without demanding the integrability of $P(\theta_0)g$ and $P(\theta_0+h)g$ separately.
It is shown in \cite{glynn96} that, in view of the fact that $W$ regenerates at hitting times of $\{0\}$, the solution $g$ to Poisson's equation for $f_p$ can be expressed as
\begin{equation}\label{A}
g(x) = \mathbf{E}_x ^{\theta_0} \sum_{j=0}^{\tau(0)-1} (f_p(W_j)-\pi(\theta_0)f_p),
\end{equation}
where $\tau(0) = \inf\{n\geq 1: W_n = 0\}$ is the hitting time of $\{0\}$. Let $Z_j = V_{j-1} - \chi_j$, $S_j = Z_1 + \cdots + Z_j$, (for $j\geq 1$), $\tau_x(0) = \inf\{j\geq 1: x + S_j \leq 0\}$, $\mu = \mathbf{E} Z_1$, and note that (\ref{A}) implies that
\begin{align}
(P(\theta_0 + h)g)(x) - (P(\theta_0)g)(x)
&= \mathbf{E}^{\theta_0} g(W_1) [p(\theta_0 + h, V_0)-1]\label{B}\\
&= \mathbf{E}^{\theta_0} \sum_{j=1}^{\tau_x(0)\rvin{-1}} [(x+S_j)^p - \pi(\theta_0) f_p] (p(\theta_0 + h, V_0) -1 ) \mathbb{I}(x + Z_1 > 0).\nonumber
\end{align}
But
\begin{align*}
\sum_{j=1}^{\tau_x(0)\rvin{-1}} (x+S_j)^p (p(\theta_0+h,V_0)-1)
&= x^p\sum_{j=1}^{\tau_x(0)\rvin{-1}} [(1+\frac{S_j-V_0}{x})^p+p\xi_j(x)^{p-1}\frac{V_0}{x}](p(\theta_0 + h, V_0)-1)
\end{align*}
where $\xi_j(x)$ lies between $1+S_j/x-V_0/x$ and $1+S_j/x$. It is easily argued, based on Riemann sum approximations, that
\begin{align*}
\sum_{j=1}^{\tau_x(0)\rvin{-1}} \xi_j(x)^{p-1} \frac{1}{x}
&\to \int_0^{\rvin{1/}|\mu|} (1+\mu s)^{p-1}ds \qquad a.s.\\
&=\frac{1}{|\mu|}\cdot \frac{1}{p}
\end{align*}
as $x \to \infty$. Furthermore, $p(\theta_0+h,V_0)-1$ is a mean zero rv that is independent of $(1+(S_j-V_0)/x)^p$ for $j\geq 1$ and $\mathbf{E} \tau_x(0) \sim x / |\mu|$ as $x\to \infty$ (where $a_1(x) \sim a_2(x)$ as $x\to \infty$ means that $a_1(x)/a_2(x) \to 1$ as $x \to \infty$). In view of (\ref{B}), this suggests that
\begin{equation*}
(P(\theta_0+h)g)(x) - (P(\theta_0)g)(x) \sim
\frac{x^p}{|\mu|} \mathbf{E} V_0 (p(\theta_0+h,V_0)-1)
\end{equation*}
as $x\to \infty$ (i.e., one power lower than the growth of $g$ itself). Thus, this style of argument can successfully deal with the integrability issue discussed earlier, and leads to a validation of the derivative formula (\ref{C}) for $W$ under the assumption $\mathbf{E} V_0^{p+1}<\infty$.
A rigorous statement and the remaining details of the proof can be found in the Appendix.
This differentiation result for $W$ can also be found in \cite{heidergott09}, with a different (and longer) proof, and with some steps that appear to be incomplete. (In particular, the paper asserts that $\rvin{\mathbf{E}_x ^{\theta_0} \sum_{j=0}^{\tau(0)-1} f_p(W_j)}$ is bounded \rvin{for any fixed $\theta$ and $p$, which implies that our function $g$ grows at most linearly regardless of $p$}; see p.248).
|
\section{Introduction}
Understanding the economic impact of a proposed policy is crucial for ensuring policy objectives are met without being excessively burdensome on the regulated industry. In fisheries, managers are often responsible for preventing over-fishing of common-pool fish stocks. This involves developing policies that balance biological sustainability with economic impacts to the fishing industry. To date, many tools available to managers measure economic impacts at the aggregate industry-level. These tools conceal important information on differences between the impacts felt by individual firms or by types of vessels. Sorting firms that benefit and those that are harmed can help managers understand the economic implications from the policy and which policies are expected to be equitable.
We investigate individual vessel response to fishery policy changes using a vessel and target-specific positive mathematical programming (PMP) model. This research is important for several reasons. To the best of our knowledge, there have only been three previous attempts to apply PMP modeling to fisheries, although none have been published in a peer-reviewed journal.\footnote{Niels Vestagaard [1998] Policy Model for a Regulated Industry: From Command and Control to Property Rights in a Danish Multispecies Fishery, Dissertation Chapter.}\footnote{John Walden [2006] Applying Positive Math Programming to a Fisheries Problem: Formulating the Closed Area Model Structure, Social Sciences Branch, NEFSC, Wood Hole, MA, 02543, Unpublished Manuscript.}\footnote{Kathereen Bisack and Gisele Magnusson [2009] Modifications to the Harbor Porpoise Take Reduction Plan. Final Environmental Assessment, NOAA-NMFS Northeast Region.} This provides an opportunity to formalize the PMP model structure for fisheries, which will serve as reference point in the literature and encourage further model development. Given the panel data structure available for Hawaii's longline fishery, we are able to evaluate the performance of the fishery PMP model by comparing out-of-sample predictions to observations from reference years. By calibrating a vessel and target-specific PMP model, this paper provides insights into the range of individual vessel responses to realistic policy changes. Finally, this paper develops a flexible tool for fishery managers to evaluate heterogeneous policy impacts with relatively few data requirements.
Recent research suggests that fisher heterogeneity is particularly important in the Hawaii longline fleet. Fishers have differing attitudes toward risk (\cite{nguyen2013revenue}), make entry/exit decisions depending on individual fisher characteristics (\cite{pradhan2004modeling}), and choose remuneration schemes based on owner/operator status (Nguyen and Leung 2009). The network position of individual fishers in the industry has also been shown to play an important role in determining outcomes (\cite{barnes2016brokerage}). These studies taken together largely invalidate the common modeling assumption that the Hawaii longline fleet is homogeneous and can be modeled using a representative vessel (\cite{Kasaoka1989, Kasaoka1990}).
Developing a model of individual vessel response to specific policy changes will, therefore, improve fleet-wide modeling accuracy. For managers of Hawaii's longline fishery, this has added significance given the economic prominence of Hawaii's longline fishing fleet. In 2013, the fleet landed 27,053 tons of fish and generated \$88.8 million gross revenues (\cite{WPacFIN2015}). The fleet primarily targets swordfish and tuna in the Eastern Pacific and Western and Central Pacific regions. It is the largest commercial fishing fleet by revenue in the state of Hawaii, with between 124 and 135 vessels operating from 2005 to 2013 (\cite{WPacFIN2015}).
The geographic scale and environmental effects of the fishery have led managers to implement numerous regulatory policies. The fishery is subject to gear restrictions, turtle bycatch caps, and annual catch limit restrictions. In recent years, the fishery has been forced to close a number of times after these policy limits were reached. In 2006 and 2011, the fishery targeting swordfish was closed because the turtle interaction limit was reached. In 2009, 2010, and 2015, the fishery targeting bigeye tuna in the Western and Central Pacific Ocean was closed because the catch limit had been reached. There is evidence that these closures may have had a dramatic economic impact on both producers and consumers in Hawaii (\cite{allen2006monitoring}).
This paper examines how policies impact individual vessels by calibrating a vessel and target-specific PMP model for Hawaii's longline fishing fleet. By calibrating at the vessel-specific level, we hope to capture the fleet's heterogeneous composition of vessels and heterogeneous response to policy changes. We also account for two primary fishing technologies targeting bigeye tuna and swordfish, and two policy relevant management areas for bigeye tuna, one in the Eastern Pacific Ocean (EPO) and the other in the Western Central Pacific Ocean (WCPO). In order to make our model computationally feasible, and economically tractable we make several assumptions. First, we assume that vessels are profit maximizing. We feel this assumption is appropriate when modeling a large commercial fishing fleet. Second, we assume economic, environmental, and biological conditions are stable, and base year observations are representative of the important economic relationships in the fishery. Under these assumptions, we model the fishery using an objective function that maximizes individual vessel profit subject to fleet-wide annual catch constraints. Individual model parameters are then calibrated to reproduce input and output levels from an observed base year (2012). Using the calibrated model and observed catch data from 2009 to 2013, we then examine model accuracy using out-of-sample model predictions. To demonstrate the model's usefulness to fishery managers, we evaluate the impact of changing the catch limit policies for bigeye fishing in the WCPO.
Although the first application of PMP was more than 25 years ago (\cite{kasnakoglu1988concept}), the PMP framework was formalized by Howitt in 1995. The idea was to blend mathematical programming constraints, which proved useful for modeling resource and policy constraints, with "positive" inferences based on observed input allocations and production levels from a particular base year. This approach was notably different from previous "normative" mathematical programming models (\cite{day1961recursive, mccarl1982cropping}) in that it was able to exactly reproduce observed inputs and outputs without relying on numerous "flexibility" constraints, which are an additional set of constraints added by the researcher to artificially avoid corner solutions. The general PMP framework can be specified using many structural forms of production and cost functions allowing for non-linearity and substitution between inputs, and can be easily calibrated using observations from a single year. It is both consistent with microeconomic theory, and when applied to policy analysis, is able to generate smooth responses to policy adjustments.
These desirable modeling characteristics have made the PMP approach common in agricultural economic modeling. Recent versions of regional agricultural models employing PMP include SWAP in California (\cite{howitt2012calibrating}), CAPRI in Europe (\cite{gocht2011eu}), and REAP in the US (\cite{Johansson2007}). These models are used repeatedly to evaluate regional agricultural response to policy changes. \cite{heckelei2012positive} and \cite{merel2014theory} provided comprehensive reviews of regional agricultural models currently using the PMP framework and recent developments in the PMP literature. There has also been significant work on developing the economic foundations of PMP, emphasizing accurate estimation of supply elasticities to be used as priors (\cite{merel2010exact}), structurally consistent estimation of shadow values (\cite{heckelei2003estimation}), and improved calibration methods (\cite{garnache2015calibration}).
By applying the most recent PMP framework developed by \cite{garnache2015calibration}, this paper builds on extensive literature modeling fleet dynamics of Hawaii's longline fishery using mathematical programming.\footnote{Curtis and Hicks (2000) investigated the impacts of fishery closure due to turtle interaction caps using a random utility model to account for spatial choice behavior of fishers.} The first model by \cite{Inc.1986}, later modified by \cite{Kasaoka1989, Kasaoka1990}, applied a linear programming (LP) framework to optimally allocate fishing time across fishing regions and target species to maximize fleet-wide profits. The results, however, did not accurately reproduce observed fishing behavior. \cite{Miklius1990} evaluated the LP model and concluded that this shortcoming resulted from the omission of micro-level decision-making by vessel owners and operators. To address this problem, \cite{pan2001decision} developed a two-level two-objective mathematical programming model which incorporated the behavior of fishers as well as fishery managers, including separate objectives of recreational and commercial fisheries. Their approach produced more plausible optimal solutions, but it remained unclear whether the approximated profit maximizing behavior was representative. The model also assumed that vessels within the fleet were homogenous and was, therefore, unable to capture the variation in vessel responses to changes in management. To address fleet heterogeneity in Hawaii's longline fishery, \cite{Yu2013} used an agent-based model. While the agent-based model was able to capture some of the detailed behavior of individual fishers, there remained a fair amount of discrepancy between predicted and observed performances. The agent-based approach to simulation also required significant model updating and refinement as well as specialized users to operate the software.
Our approach using the PMP framework is intended to be used by policy makers and managers, as well as academics. The vessel and target-specific PMP model is able to capture fleet heterogeneity, separate fishing technologies and regional policies, and measure the distributional effects from changes to fishery policy. It requires minimal data to calibrate, and is amenable to a wide range of resource and policy constraints including catch limits, and protected species interaction caps. It is also able to exactly reproduce base year inputs, costs, revenues, and profits for individual vessels without relying on additional constraints. For these reasons we feel it will be able to address previous modeling limitations.
This paper makes four important contributions to the literature. First, the paper adapts the PMP framework developed for agriculture to a framework that can be applied to fisheries in general. With only a handful of notable exceptions, research using PMP for fisheries policy analysis has been very limited. Second, by calibrating a vessel and target-specific PMP model, we are able to demonstrate a technique to examine the heterogeneous nature of the fishing fleet and the heterogeneous responses to specific policy changes. Previous literature on Hawaii's longline fishery has made significant progress to address fleet heterogeneity, but this paper provides a method that explicitly models individual vessels and fish targeting decisions, and requires less data and less effort to calibrate and conduct policy simulations than previous frameworks. Third, it provides a rigorous out-of-sample evaluation of the accuracy of PMP model predictions. Although PMP models have been used extensively for policy analysis, model predictions are rarely evaluated. The panel data we have on Hawaii's longline fishery enable us to make out-of-sample predictions for catch and evaluate the model's predictive accuracy. Finally, the calibrated PMP model of Hawaii's longline fishery provides a valuable tool for resource managers and policy analysts to evaluate the heterogeneous economic impacts of specific fishery policies and determine which policies are likely to encounter industry support or opposition.
\section{Data}
To calibrate the PMP model, evaluate its performance, and simulate policy outcomes, we used data from four sources. We obtained data on individual vessel input costs for 2005 from the 2005 cost and earnings survey (\cite{Pan2015}), and for 2012 from the 2012 cost and earnings survey (\cite{Pan2015a}). We obtained data on annual vessel catch from 2005-2013 using the dealer data from the State of Hawaii (\cite{Quach2015}). We obtained data on annual hooks deployed from 2005-2013 from Federal logbook data (\cite{Monitoring2015}). To evaluate out-of-sample prediction accuracy we adjusted all input and output prices to 2012 dollars using the Consumer Price Index for all urban consumers nationally. Input levels for the variable costs were then scaled relative to the number of fishing hooks deployed to enable efficient optimization during model calibration and simulations. Prices of inputs were adjusted using the inverse scaling ratio to preserve the observed expenditure for each input. We were able to match vessels across data sources using vessel name, permit number, and commercial license.
In 2012, there were 129 vessels operating in Hawaii's longline fishery. Of the 129 vessels operating, 114 were represented in the cost and earnings survey (\cite{Pan2015a}). We imputed input cost for missing vessels using random regression imputation considering gear usage, vessel catch profile, and time spent on each target as regression variables. Variable costs were then grouped into six categories: fuel, captain pay, crew pay, bait, other, and gear. We grouped fuel and oil costs under fuel, fixed captain pay and shares paid to the captain under captain pay, combined crew fixed pay and crew shares paid under crew pay, total bait costs under bait, and gear replacement cost under gear. Table \ref{tab1} shows the degree of fleet heterogeneity based on these inputs. According to the survey data, total variable costs exceeded total gross revenue for six vessels. Rather than dropping these vessels because they violated the profit maximizing assumption, we scaled their input costs such that annual profits were 0.
We then disaggregated individual vessel expenditure, catch, and revenue by three policy relevant targets: bigeye EPO, bigeye WCPO, and swordfish. The EPO and WCPO management regions are separated at 150 W longitude. Bigeye and swordfish fishing sets differ by depth, with swordfish lines set shallower than deep set bigeye lines. We used set-type and location from 2012 logbook data to calculate the proportion of total trip time spent each trip on each target. Trip target time was then aggregated by vessel over the entire year indicating how much time each vessel spent on each target for 2012. Using the dealer data from 2005-2013, we matched vessel trips to observed landings to calculate annual catch and revenue by vessel and target. Observations in the dealer data recorded daily sales. Fish sales were either recorded by individual fish or groups of fish sold together. Daily vessel revenue was calculated by multiplying pounds sold per fish, or group of fish by recorded ex vessel price per pound. The data were then aggregated on vessel and year to calculate the annual pounds of swordfish and bigeye caught, and the total value of vessel catch. These data were then used to calculate fleet-wide average price of swordfish and bigeye, vessel-specific price premium for swordfish and bigeye, and price of non-target catch representing its added value. Input expenditures for each vessel were disaggregated by target according to the proportion of time spent on each target in 2012. Table \ref{tab2} summarizes the total active fleet size, and model sample size for each target over the years 2005-2013.
\begin{sidewaystable}[]
\centering
\caption{Time series data summary of total active and modeled vessels from the 2005 to 2013 dealer data. Because some vessels fish more than one target, total vessels modeled can be less than the sum of each target.}
\label{tab2}
\resizebox{\textwidth}{!}{%
\begin{tabular}{llllll}
Year & Total Vessels Operating & Total Vessels Modeled & Vessels modeled (WCPO) & Vessels modeled (EPO) & Vessels modeled (SF) \\
2005 & 125 & 105 & 103 & 41 & 11 \\
2006 & 127 & 112 & 111 & 11 & 10 \\
2007 & 129 & 116 & 115 & 53 & 13 \\
2008 & 129 & 118 & 115 & 79 & 11 \\
2009 & 127 & 120 & 118 & 73 & 15 \\
2010 & 124 & 119 & 115 & 86 & 15 \\
2011 & 129 & 124 & 122 & 83 & 16 \\
2012 & 129 & 128 & 127 & 94 & 17 \\
2013 & 135 & 126 & 124 & 83 & 10
\end{tabular}}
\end{sidewaystable}
\section{Model specification}
The PMP framework consists of an objective function defining profit maximization and resource and policy constraints that restrict input allocation decisions. To allow for non-linearity in production and limited substitution between inputs we chose to use a generalized constant elasticity of substitution (CES) production function, and for simplicity a linear expenditure function. When paired with a CES production function, the linear expenditure function allows for smooth responses to changes in policy and resource constraints without adding more parameters to calibrate. We define subscript to index the set of 128 vessels in our sample, indexes targets EPO, WCPO, and SF, and indexes inputs for fuel, captain pay, crew pay, bait, other and gear. Given a CES specification the production function for vessel targeting is given below.
\begin{equation}
y_{i,r}=\alpha_{i,r}(\sum_j \beta_{i,j,r} x_{i,j,r}^\rho)^{\delta/\rho}.
\end{equation}
We define the scale parameter for vessel technology as $\alpha_{i,r}$, input share as $\beta_{i,j,r}$, elasticity of substitution as $\rho$, and the returns to scale coefficient as $\delta$. By relating effort to catch, the scale parameter is analogous to a vessel-specific catchability parameter in traditional fishery production models. The returns to scale coefficient is defined using a myopic definition (\cite{garnache2015calibration}) relating returns to scale to supply elasticity ($\eta$)
\begin{equation}
\delta_{myo}=\frac{\eta}{1+\eta}.
\end{equation}
Because there have been no direct estimates of supply elasticity of catch in Hawaii's longline fleet, we assume $\eta=0.5$, which lies in the range of published supply elasticity estimates for the Gulf of Mexico fishery (\cite{zhang2011estimation}). To simplify notation, we use a transformed elasticity of substitution defined as
\begin{equation}
\rho=\frac{\sigma -1}{\sigma},
\end{equation}
where the untransformed elasticity of substitution ($\sigma$) is assumed to be 0.17 for all inputs. At present, we are unable to estimate an elasticity of substitution from the data available, and the value of 0.17 allows for limited substitution between inputs, which we borrow from the agriculture literature and feel is reasonable in a fishery setting (\cite{howitt2012calibrating}). Model sensitivity analyses for these assumptions are provided in Figure \ref{figA1} and indicate our results are robust to changes in assumed parameter values.
Although our production function only models targeted catch, fisher's revenue will depend on their ability to land quality fish, and on the value of non-target but commercially valuable bycatch. To fully capture these components of revenue we model the price of swordfish and bigeye separately for each vessel. The fleet-wide average prices for swordfish and bigeye are given by $p_{i,sf}$, and $p_{i,be}$, vessel-specific price premiums for swordfish and bigeye accounting for variation in quality are given by $p_{i,sfpr}$, and $p_{i,bepr}$, and the additional values from non-targeted bycatch are given by $p_{i,nsf}$, and $p_{i,nbe}$. By adding these three components together, we specify a vessel-specific price for bigeye (BE), and swordfish (SF).
\begin{equation}
p_{i,SF}=p_{i,sf}+p_{i,sfpr}+p_{i,nsf}
\end{equation}
\begin{equation}
p_{I,EPO}=p_{i,WCPO}=p_{i,be}+p_{i,bepr}+p_{i,nbe}
\end{equation}
This specification allows us to exactly reproduce observed vessel revenue, while only modeling the production of the policy relevant targets. Implicit in this price specification we assume the price of bigeye from the EPO is the same as bigeye from the WCPO, which we feel is reasonable given they belong to the same species and are both caught throughout the year.
For simplicity, we specify a linear expenditure function. The input cost data only provides total annual costs per input, therefore we assume input prices ($c_{i,j,r}$) are 1, which implies input levels ($x_{i,j,r}$) are in dollar units. The choice set $x_{i,j,r}$ is the vector of individual vessel input levels for each target. Profit maximization is constrained by three policies. We model annual catch limits for bigeye tuna in the EPO ($ACL_{EPO}$) and WCPO ($ACL_{WCPO}$), and a total annual catch limit for swordfish ($ACL_{SF}$). Vessel heterogeneity implies that the unobserved value of catch for each constraint will vary by vessel. We therefore define the unobserved value of catch as $\mu_{i,r}$ over vessels and targets. The maximization problem is given below.
\begin{equation}
\begin{array}{rrclcl}
\displaystyle \max_{x_{i,j,r}} & \multicolumn{3}{l}{\sum_i \sum_r \left[ (p_{i,r}+ \mu_{i,r})y_{i,r} - \sum_j c_{i,j,r}x_{i,j,r} \right]} \\
\textrm{s.t.} & \sum_i y_{i,EPO} & \leq & ACL_{EPO} \\
& \sum_i y_{i,WCPO} & \leq & ACL_{WCPO} \\
& \sum_i y_{i,SF} & \leq & ACL_{SF} \\
\end{array}
\end{equation}
\section{Model calibration}
We adapted the calibration procedure developed by \cite{garnache2015calibration}. Their calibration procedure is the most recent methodological advance in the PMP literature, comprehensively addressing the criticism by \cite{heckelei2003estimation} regarding the calibration of shadow values. Rather than estimated using an LP or ad hoc measures as was done previously, all unknown parameters and the shadow values are calibrated simultaneously using the same structural forms as used in model simulations, in this case a CES production function with a linear expenditure function. \cite{garnache2015calibration} calibrated a PMP model for agriculture. In agriculture, the constrained input is typically land, however, in fisheries, production inputs can be purchased at any desired level on a common market and the constrained resource is catch. We adapted the calibration procedure to account for this difference. For each target we specified a shadow value ($\lambda_r$). We then calibrated the model by minimizing the sum of squared error between observed expenditures and model expenditures resulting from the choice variable $\lambda_r$ as specified below.
\begin{equation}
\min_\lambda \sum_i \sum_r \left[ (p_{i,r}+\lambda)\bar{q}_{i,r}\delta - \sum_j c_{i,j,r} \right]^2
\end{equation}
The objective function is subject to four sets of constraints that determine the calibration of unknown parameters. The first set of constraints requires production parameters reproduce observed output ($\bar{q}_{i,r}$) for each vessel and target.
\begin{equation}
\bar{q}_{i,r}=\alpha \left( \sum_j \beta_{i,j,r} x_{i,j,r}^\rho \right)^{\delta/\rho} , \forall i, r
\end{equation}
The second set of constraints requires the first order conditions of profit maximization hold. The first order condition will be specified for each input, vessel, and target as below.
\begin{equation}
\begin{split}
p_{i,r}\alpha_{i,r}\delta \left( \sum_j \beta_{i,j,r} x_{i,j,r}^\rho \right)^{(\delta/\rho) - 1} \beta_{i,j,r} x_{i,j,r}^{\rho-1} \\
= c_{i,j,r} - (\lambda+\mu_{i,r})\alpha \delta \left( \sum_j \beta_{i,j,r} x_{i,j,r}^\rho \right)^{(\delta/\rho)-1} \beta_{i,j,r} x_{i,j,r}^{\rho-1}, \forall i,j,r
\end{split}
\end{equation}
The third set of constraints allows us to recover the vessel and target-specific unobserved value of catch ($\mu_{i,r}$).
\begin{equation}
\begin{split}
p_{i,r}\sum_j \left[ \alpha \delta \left( \sum_j \beta_{i,j,r} x_{i,j,r}^\rho \right)^{(\delta/\rho)-1} \beta_{i,j,r} x_{i,j,r}^{\rho-1} \right] \\
= \sum_j c_{i,j,r} - (\lambda+\mu_{i,r})\alpha \delta \sum_j \left[ \left( \sum_j \beta_{i,j,r} x_{i,j,r}^\rho \right)^{(\delta/\rho)-1} \beta_{i,j,r} x_{i,j,r}^{\rho-1} \right], \forall i,j.
\end{split}
\end{equation}
Finally, our calibration procedure requires that for each vessel-target combination the sum of the input share parameters is one.
\begin{equation}
\sum_j \beta_{i,j,r} = 1, \forall i,r.
\end{equation}
\section{Calibration results}
The PMP model calibration procedure is designed to calibrate unknown parameters and constraint shadow values such that profit maximizing vessels, subject to the base year resource constraints, will optimally allocate the observed base year levels of input, generating the observed outputs and revenues, and the observed expenditures. To evaluate whether the calibration was successful, we examine the range of calibrated parameter values and the differences between the observed and the modeled input levels using the base year catch constraints in 2012.
In Table \ref{tab3}, we present the range of calibrated model parameters. The largest magnitude of variation is found in unobserved shadow prices of catch and the scale parameters. These parameters carry the most weight for modeling the heterogeneous responses of the fleet. The share parameters also show significant variation indicating the model captured a large amount of vessel heterogeneity in input expenditures. Across targets, the share parameter for fuel are consistently larger than the other inputs, which is expected given fuel is the largest single input cost. To verify the calibration procedure, we examine the differences between observed and modeled input levels for each input and each vessel's output using the base year constraints. The largest difference in input is 2.02x10-14\% and the largest difference in output is 9.53x10-6\%. Such small differences indicate that we achieve an accurate calibration of all unknown parameters, and that our model can very closely replicate the observed base year economic behavior of each vessel.
\begin{table}[]
\centering
\caption{Summary of calibrated parameters for vessels modeled vessel and target-specific PMP model. The mean and standard deviation for each target-specific parameter are given.}
\resizebox{\textwidth}{!}{%
\label{tab3}
\begin{tabular}{lllll}
Description & Symbol & WCPO & EPO & SF \\
Scale parameter &$\alpha$& 1,104.02 (418.80) & 382.38 (305.60) & 1,629.79 (496.58) \\
Shadow value &$\lambda$& -7.70 (NA) & -7.57 (NA) & -4.45 (NA) \\
Unobserved price of catch &$\mu$& 17.42 (4.65) & 24.00 (16.33) & 10.41 (2.32) \\
Share parameter for fuel &$\beta_{fuel}$& 0.42 (0.10) & 0.57 (0.24) & 0.89 (0.16) \\
Share parameter for captain pay &$\beta_{cap}$& 0.19 (0.08) & 0.14 (0.10) & 0.04 (0.07) \\
Share parameter for crew pay &$\beta_{crew}$& 0.12 (0.09) & 0.08 (0.08) & 0.02 (0.02) \\
Share parameter for bait &$\beta_{bait}$& 0.13 (0.03) & 0.10 (0.06) & 0.02 (0.04) \\
Share parameter for other &$\beta_{other}$& 0.08 (0.03) & 0.07 (0.04) & 0.02 (0.03) \\
Share parameter for gear &$\beta_{gear}$& 0.05 (0.02) & 0.04 (0.02) & 0.02 (0.02)
\end{tabular}}
\end{table}
To further verify the calibration procedure, we compare the shadow values to the observed average price per pound of fish. The shadow value on each resource constraint can be interpreted as the value of relaxing the resource constraint by one pound of either bigeye or swordfish. Taken in absolute value terms, the calibrated shadow values of -7.70, -7.57, and -4.45, representing bigeye catch in the WCPO, EPO, and swordfish catch respectively, appear to be accurately calibrated. When compared to the average observed price per pound of bigeye, and swordfish (\$7.99, \$4.30 respectively), our calibrated shadow values are within a few cents of the average observed fish prices. Although average prices and shadow values do not share the same interpretation, comparing the two does provide a useful validation of the overall calibration procedure.
\section{Prediction accuracy}
We evaluate model predictions in two ways. First, we compare predicted and observed catch from 2009 to 2013. Of the 128 vessels modeled, 126 were operating in 2013; however, going back to 2009, as few as 119 of the original 128 were previously operating (Table \ref{tab2}). For each year, we simulate the model by setting the fleet-wide catch constraint less than or equal to the total observed catch of the vessels remaining from our 2012 sample. This implies that our simulated fleet size decreases as vessels operating in 2012 are no longer observed in more distant years. To account for changes in input costs over time, we adjust the cost of fuel using U.S. number 2 diesel retail price\footnote{https://www.eia.gov/dnav/pet/hist/LeafHandler.ashx?n=PET\&s=EMD\_EPD2D\_PTE\_NUS\_DPG\&f=A} and the costs for captain pay and crew pay using annual salary data from Bureau of Labor Statistics occupational profiles for farming, fishing, and forestry occupations.\footnote{http://www.bls.gov/oes/tables.htm} Regressing the predicted revenue on observed revenue for the years 2009-2013, we examine the correlation coefficient and the amount of variation explained by our model (Figure \ref{fig1}). We find the model performs best predicting bigeye catch in the WCPO, modestly for bigeye catch in the EPO, and poorly for swordfish catch. The best out-of-sample model predictions are made for the 2011 bigeye catch in the WCPO (R-squared=0.35, correlation coefficient=0.53). For all targets, model predictions become less accurate moving further in time away from the calibrated base year. This is expected as biological stock level, individual fishing location decisions, and environmental conditions could vary substantially over this time, while our model assumes conditions remain constant. In the short-term the model makes reliable predictions of individual vessel catch for the largest target in the fishery, bigeye in the WCPO.
\begin{sidewaysfigure}
\includegraphics[width=\linewidth]{comb_pred.pdf}
\caption{Evaluation of model predictions of individual vessel catch for bigeye in the WCPO (red), bigeye in the EPO (green), and swordfish (blue) from 2009-2013. The solid line indicates the 45-degree line. The correlation coefficient and R-Squared from the linear model are given in the top-left corner of each plot. Axes are scaled so the maximum catch is 1 to prevent disclosure of confidential data.}
\label{fig1}
\end{sidewaysfigure}
Second, we evaluate the model input level predictions for each target comparing the observed input levels from the 2005 cost and earnings data to the predicted input levels simulated using our PMP model. Results are shown in Table 4. In order to compare the values, we match vessels that appear in both sets, reducing our sample to 71, 25, and 1 for the WCPO, EPO, and SF targets respectively. Results from a paired Wilcoxon test comparing the observed and predicted input expenditures show the model significantly under-predicts all inputs except gear and bait for the WCPO target. The model tends to over-predict input costs for the EPO target, and it over-predicts all inputs except fuel for the one matched vessel targeting SF. By comparing observed expenditures in 2012 (Table \ref{tab1}) to 2005 (Table \ref{tab4}), the primary source of prediction error is the large differences in the observed expenditures between 2012 and 2005. For instance, fundamental changes to the remuneration schemes over these years, including the wide-spread transition from crew shares paid to domestic crew to fixed pay for foreign crew, could account for the observed differences in crew pay and captain pay. We also observed a reduction in fuel expenditures in 2005 in the WCPO and EPO, and increase in SF, which could reflect a change in fishing grounds requiring more or less travel time than in 2012. Similar explanations could account for differences in other input expenditures predicted for each target. Gear and bait expenses, which we expect to be most closely tied to catch, generate the closest predictions and are not sensitive to changes in remuneration scheme or fishing location. Any changes to the fundamental cost structure of the fleet are expected to alter model parameter values and reduce the accuracy of forecasts. This limitation is common to all model based forecasts.
\begin{table}[]
\centering
\caption{Data summary of annual input costs in dollars for WCPO, EPO, and SF targets from the Cost and Earnings Survey in 2012.}
\label{tab1}
\begin{tabular}{llll}
Inputs & Mean WCPO (SD) & Mean EPO (SD) & Mean SF (SD) \\
Fuel & \$154,045 (62,542) & \$27,134 (31,917) & \$16,318 (44,331) \\
Captain Pay & \$75,700 (47,061) & \$13,623 (18,167) & \$6,962 (19,937) \\
Crew Pay & \$47,255 (46,103) & \$7,245 (12,246) & \$1,978 (6,192) \\
Bait & \$48,722 (17,761) & \$7,928 (8,635) & \$4,013 (10,787) \\
Other & \$31,477 (12,796) & \$5,029 (5,652) & \$3,195 (8,844) \\
Hooks & \$19,346 (8,583) & \$3,160 (3,479) & \$2,062 (5,618)
\end{tabular}
\end{table}
\begin{sidewaystable}[]
\centering
\caption{Mean observed and the median difference between observed and predicted input expenditures. Observed data came from the 2005 Cost and Earnings Survey. All values are adjusted to 2012 dollars. The median difference and p-values are from a two-sample paired Wilcoxon test.}
\label{tab4}
\resizebox{\linewidth}{!}{%
\begin{tabular}{lllllll}
& WCPO & & EPO & & SF & \\
Inputs & Mean Observed (dollars) & Median Predicted Difference (P-value) & Mean Observed & Median Predicted Difference (P-value) & Mean Observed & Median Predicted Difference (P-value) \\
& & & (dollars) & & (dollars) & \\
Fuel & 106,532 & -25,324 & 10,972 & 9,388 & 22,349 & -4,563 \\
& & (\textless0.001) & & -0.059 & & (NA) \\
Captain Pay & 84,114 & -19,017 & 7,404 & 5,789 & 11,937 & 4,229 \\
& & -0.038 & & -0.101 & & (NA) \\
Crew Pay & 56,204 & -27,395 & 5,150 & 3,523 & 9,744 & 6,471 \\
& & (\textless0.001) & & -0.022 & & (NA) \\
Bait & 39,544 & 45 & 3,627 & 5,141 & 8,312 & 7,561 \\
& & -0.984 & & -0.007 & & (NA) \\
Other & 32,259 & -5,599 & 2,635 & 2,920 & 4,785 & 10,807 \\
& & -0.011 & & -0.011 & & (NA) \\
Gear & 17,426 & -1,136 & 1,430 & 2,153 & 4,006 & 3,866 \\
& & -0.389 & & (\textless0.001) & & (NA) \\
Sample & 71 & & 25 & & 1 &
\end{tabular}}
\end{sidewaystable}
\begin{figure}
\includegraphics[width=\textwidth]{policy_effect.pdf}
\caption{Distribution of responses for individual vessels measured by the percent change from 2012 catch levels. Results from 10\% increase in annual catch constraint from 2012 are given filled black and represent increases in catch. Results from 10\% decrease in annual catch constraint from 2012 are filled red and represent decreases in catch.}
\label{fig2}
\end{figure}
\section{Policy simulations}
To demonstrate the usefulness of a vessel-specific PMP model for Hawaii's longline fishery, we examine vessel responses and impacts on individual vessel catch to changes in the annual catch limit policy. We simulate two policy changes. The first is a policy that increases the annual catch limit of bigeye in the WCPO by 10\% from the 2012 base year. The second is a policy that decreases the same catch limit by 10\% from the 2012 base year. A 10\% change in the catch limit policy is roughly in line with the agreed upon changes for bigeye in the WCPO in the next few years which will see catch limit decrease 11\% from 3,763 metric tons in 2014, to 3,345 metric tons in 2017.
The vessel-specific nature of our PMP model allows us to evaluate the distributional effects of such policy changes. We expect that individual vessels will respond to varying degrees, depending on factors such as technological efficiency and profitability, which makes them more or less sensitive to policy changes. In Figure \ref{fig2}, we present the distribution of catch responses given an increase and decrease in bigeye catch limits in the WCPO. The range of responses is large. With a 10\% increase in catch limit, we see that vessels respond by increasing catch from less than 5\% to 20\%. With a 10\% decrease in catch limit, the responses are symmetric to the 10\% increase policy. Vessels reduce catch from less than 5\% to 25\%. Given the range in policy responses, individual vessels will clearly be affected differently. Some will be highly sensitive to policy changes; most will experience moderate impacts. Understanding the distributional implications is clearly important for evaluating economic impacts of fishery policies in Hawaii's longline fishery.
\section{Conclusion}
In this paper, we have shown that the vessel and target specific PMP model of Hawaii's longline fishery reliably predicts short-term effect of policies on bigeye catch in the WCPO and EPO. Model predictions are more accurate when simulating vessel responses close to the base year, but lend some insight even at further distances. By calibrating at the vessel-specific level, we are able to identify the range of economic responses to policy changes, capturing the heterogeneous nature of Hawaii's longline fleet. This more realistically models vessel responses, as well as provides an evaluation of the distributional effects of policy changes on catch, which is important for evaluating the stability of new policies. For fishery managers, the PMP model of Hawaii's longline fishery provides a valuable tool for evaluating the economic impacts of current and potential fishery policies.
The PMP framework also provides a rich structural model with which we can study fisheries in general. Later work will address parameter instability resulting from fundamental changes to underlying economic relationships or environmental and biological conditions, and estimate target switching decisions made by fishers. We will also consider the effects of overlapping policy constraints such as turtle interaction caps, and explore the individual vessel characteristics that make certain vessels more sensitive to policy changes than others.
\printbibliography
\section{Appendix}
\begin{figure}
\includegraphics[width=\textwidth]{sens_plot.pdf}
\caption{Sensitivity analysis measuring the effect from changing assumed supply elasticity and substitution elasticity values on model prediction results from 2009-2013.}
\label{figA1}
\end{figure}
\end{document} |
\section{Introduction}
\markboth{\centerline{\it Introduction}}{\centerline{\it D.T.V.~An
and N.D.~Yen}} \setcounter{equation}{0}
Investigations on differentiability properties of the \textit{optimal value function} and of the \textit{solution map} in parametric mathematical programming are usually classified as studies on differential stability of optimization problems. Some results in this direction can be found in \cite{AnYao,AnYen,Aubin_1998,Auslender_1979,Bonnans_Shapiro_2000,Dien_Yen_1991,Gauvin_Dubeau_1982,Gauvin_Dubeau_1984,Gauvin_Tolle_1977,Gollan_1984,Mordukhovich_2006,MordukhovichEtAl_2009,Rockafellar_1982,Thibault_1991}, and the references therein.
For differentiable nonconvex programs, the works of Gauvin and Tolle \cite{Gauvin_Tolle_1977}, Gauvin and Dubeau \cite{Gauvin_Dubeau_1982}, and Lempio and Maurer \cite{Lempio_Maurer_1980} have had great impacts subsequently. The authors of the first two papers studied parametric programs in a finite-dimensional setting, while a Banach space setting was adopted in the third one. The main ideas of those papers are to use linear linearizations and a regularity condition (either the Mangasarian-Fromovitz Constraint Qualification, or the Robinson regularity condition). Formulas for computing or estimating the Dini directional derivatives, the classical directional derivative, or the Clarke generalized directional derivative and the Clarke generalized gradient of the optimal value function, when the problem data undergoes smooth perturbations, were given in \cite{Gauvin_Dubeau_1982,Gauvin_Tolle_1977,Lempio_Maurer_1980}. Gollan \cite{Gollan_1984}, Outrata~\cite{Outrata}, Penot~\cite{Penot1997}, Rockafellar \cite{Rockafellar_1982}, Thibault \cite{Thibault_1991}, and many other authors, have shown that similar results can be obtained for nondifferentiable nonconvex programs. In particular, the links of the subdifferential of the optimal value function in the contingent sense and in the \F \ sense with multipliers were pointed in~\cite{Penot1997}. Note also that, if the objective function is nonsmooth and the constraint set is described by a set-valued map, differential stability analysis can be investigated by the primal-space approach; see~\cite{Outrata} and the references therein.
For optimization problems with inclusion constraints in Banach spaces, differentiability properties of the optimal value function have been established via the dual-space approach by Mordukhovich \textit{et al.} in \cite{MordukhovichEtAl_2009}, where it is shown that the new general results imply several fundamental results which were obtained by the primal-space approach.
Differential stability for convex programs has been studied intensively in the last five decades. A formula for computing the subdifferential of the optimal value function of a standard convex mathematical programming problem with right-hand-side perturbations, called the \textit{perturbation function}, via the set of the \textit{Kuhn-Tucker vectors} (i.e., the vectors of Kuhn-Tucker coefficients; see \cite[p.~274]{Rockafellar_1970}) was given by Rockafellar \cite[Theorem~29.1]{Rockafellar_1970}. Until now, many analogues and extensions of this classical result have been given in the literature.
New results on the exact subdifferential calculation for optimal value functions involving coderivatives of constraint set mapping have been recently obtained by Mordukhovich \textit{et al.} \cite{Mordukhovich_Nam_Rector_Tran} for optimization problems in Hausdorff locally convex topological vector spaces, whose convex marginal functions are generated by arbitrary convex-graph multifunctions. Actually, these developments extend those started by Mordukhovich and Nam \cite[Sect.~2.6]{Mordukhovich_Nam2014} and \cite{Mordukhovich_Nam2015} in finite dimensions.
Recently, by using the Moreau-Rockafellar theorem and appropriate regularity conditions, An and Yao \cite{AnYao}, An and Yen \cite{AnYen} have obtained formulas for computing the subdifferential and the singular subdifferential of the optimal value function of infinite-dimensional convex optimization problems under inclusion constraints and of infinite-dimensional convex optimization problems under geometrical and functional constraints. Coderivative of the constraint multifunction, subdifferential, and singular subdifferential of the objective function are the main ingredients in those formulas.
The present paper discusses differential stability of convex programming problems in Hausdorff locally convex topological vector spaces. Among other things, we obtain formulas for computing or estimating the subdifferential and the singular subdifferential of the optimal value function via suitable multiplier sets.
Optimality conditions for convex optimization problems under inclusion constraints and for convex optimization problems under geometrical and functional constraints will be formulated too. But our main aim is to clarify the connection between the subdifferentials of the optimal value function and certain multiplier sets. Namely, by using some results from~\cite{AnYen}, we derive an upper estimate for the subdifferentials via the Lagrange multiplier sets and give an example to show that the upper estimate can be strict. Then, by defining a satisfactory multiplier set, we obtain formulas for computing the subdifferential and the singular subdifferential of the optimal value function.
As far as we understand, Theorems \ref{computing_subdifferential} and \ref{thm_singular_computing} in this paper have no analogues in the vast literature on differential stability analysis of parametric optimization problems. Here, focusing on convex problems, we are able to give exact formulas for the subdifferential in question under a minimal set of assumptions. It can be added also that the upper estimates in Theorems~\ref{outer_theorem} and~\ref{outer_theorem_singular} are based on that set of assumptions, which is minimal in some sense.
As one referee of our paper has observed that the results in the convex framework are essentially different from nonconvex ones given, e.g., in the book by Mordukhovich \cite{Mordukhovich_2006}. The main difference of the results in the present paper and those from \cite{Mordukhovich_Nam2014,Mordukhovich_Nam2015}, and \cite[Theorem~7.2]{Mordukhovich_Nam_Rector_Tran}, is that the latter ones are expressed in terms of the coderivatives of the general convex-graph mappings, while the former ones are given directly via Lagrange multipliers associated with the convex programming constraints. Note that the coderivative calculations for such constraint mappings are presented, e.g., in \cite{Mordukhovich_2006} and the convex
extremal principle established in \cite[Theorem~2.2]{Mordukhovich_Nam2017} is a main tool of \cite{Mordukhovich_Nam_Rector_Tran}.
As examples of application of theoretical results on sensitivity analysis (in particular, of exact formulas for computing derivative of the optimal value function) to practical problems, we refer to \cite[Sects.~1 and~6]{Burke}, where the authors considered perturbed linear optimization programs. The results obtained in this paper can be applied to perturbed convex optimization problems in the same manner.
The organization of the paper is as follows. Section 2 recalls some definitions from convex analysis, variational analysis, together with several auxiliary results. In Section 3, optimality conditions for convex optimization problems are obtained under suitable regularity conditions. Section 4 establishes formulas for computing and estimating the subdifferential of the optimal value function via multiplier sets. Formulas for computing and estimating the singular subdifferential of that optimal value function are given in Section 5.
\section{Preliminaries}
\markboth{\centerline{\it Preliminaries}}{\centerline{\it D.T.V.~An
and N.D.~Yen}} \setcounter{equation}{0}
Let $X$ and $Y$ be Hausdorff locally convex topological vector spaces with the topological duals denoted, respectively, by $X^*$ and $Y^*$. For a convex set $\Omega\subset X$, the \textit{normal cone} of $\Omega$ at $\bar x\in \Omega$ is given by \begin{align} N(\bar x; \Omega)=\{x^*\in X^* \mid \langle x^*, x-\bar x \rangle \leq 0, \ \, \forall x \in \Omega\}.\end{align}
Consider a function $f:X \rightarrow \overline{\mathbb{R}}$ having values in the extended real line $\overline{\mathbb{R}}=[- \infty, + \infty]$. One says that $f$ is \textit{proper} if $f(x) > - \infty$ for all $x \in X$ and if the \textit{domain} ${\rm{dom}}\, f:=\{ x \in X \mid f(x) < +\infty\}$ is nonempty. The set $ {\rm{epi}}\, f:=\{ (x, \alpha) \in X \times \mathbb{R} \mid \alpha \ge f(x)\}$ is called the \textit{epigraph} of $f$. If ${\rm{epi}}\, f$ is a convex (resp., closed) subset of $X \times {\mathbb{R}}$, $f$ is said to be a \textit{convex} (resp., {\it closed}) function.
\medskip
The {\it subdifferential} of a proper convex function $f: X\rightarrow \overline{\mathbb{R}}$ at a point $\bar x \in {\rm dom}\,f$ is defined by
\begin{align}\label{subdifferential_convex_analysis}
\partial f(\bar x)=\{x^* \in X^* \mid \langle x^*, x- \bar x \rangle \le f(x)-f(\bar x), \ \forall x \in X\}.
\end{align}
The {\it singular subdifferential} of a proper convex function $f: X\rightarrow \overline{\mathbb{R}}$ at a point $\bar x \in {\rm dom}\,f$ is given by
\begin{align}\label{singular_subdifferential_convex_analysis}
\partial^\infty f(\bar x)=\{x^* \in X^* \mid (x^*,0)\in N ( (\bar x, f(\bar x));{\rm epi}\,f)\}.
\end{align}
By convention, if $\bar x \notin {\rm dom}\,f$, then $\partial f(\bar x)=\emptyset$ and $\partial^\infty f(\bar x)=\emptyset$.
Note that $x^* \in \partial f(\bar x)$ if and only if $\langle x^*, x-\bar x \rangle -\alpha f(\bar x) \le 0$ for all $(x, \alpha) \in {\rm epi}\, f$
or, equivalently, $(x^*,-1)\in N( (\bar x, f(\bar x)); {\rm epi}\, f)$. Also, it is easy to show that $\partial\iota_\Omega (x)=N(x;\Omega)$ where $\iota_\Omega (\cdot)$ is the {\it indicator function} of a convex set $\Omega \subset X$. Recall that $\iota_\Omega (x)=0$ if $x \in \Omega$ and $ \iota_\Omega (x)=+\infty$ if $x \notin \Omega$. Interestingly, for any convex function $f$, one has $\partial^\infty f(\bar x)=N(\bar x; {\rm dom}\,f)=\partial \iota_{{\rm dom}\,f} (\bar x)$~(see \cite[Proposition~4.2]{AnYen}).
One says that a multifunction $F:X\rightrightarrows Y$ is {\it closed} (resp., {\it convex}) if ${\rm gph}\,F$ is a closed (resp., convex) set, where ${\rm gph}\,F:=\{(x,y)\in X \times Y \mid y \in F(x) \}$.
Given a convex function $\varphi: X \times Y \rightarrow \overline{\mathbb{R}}$, we denote by $\partial_x \varphi(\bar x, \bar y)$ and $\partial_y \varphi(\bar x, \bar y)$, respectively, its \textit{partial subdifferentials} in $x$ and $y$ at $(\bar x,\bar y)$. Thus, $\partial_x \varphi(\bar x, \bar y)=\partial \varphi(., \bar y)(\bar x)$ and $\partial_y \varphi(\bar x, \bar y)=\partial\varphi (\bar x, .)(\bar y)$, provided that the expressions on the right-hand-sides are well defined. It is easy to check that
\begin{align}\label{inclusion_partial}
\partial \varphi(\bar x, \bar y) \subset \partial_x \varphi(\bar x, \bar y)\times \partial_y \varphi(\bar x, \bar y).
\end{align}
Let us show that inclusion \eqref{inclusion_partial} can be strict.
\begin{example}\label{Ex1} {\rm Let $X=Y=\mathbb{R}$, $\varphi(x,y)=|x+y|$, and $\bar x=\bar y=0$. Since
$$\varphi(x,y)=|x+y|=\max\{x+y,-x-y\},$$ by applying a well known formula giving an exact expression of the subdifferential of the maximum function \cite[Theorem~3, pp.~201--202]{Ioffe_Tihomirov_1979}
we get $$\partial \varphi (\bar x, \bar y)={\rm co}\left\{(1,1)^T,\, (-1,-1)^T \right\},$$
where ${\rm co}\,\Omega$ denotes the \textit{convex hull} of $\Omega$. Since $\partial_x \varphi (\bar x, \bar y)=\partial_y \varphi (\bar x, \bar y)=[-1,1]$, we see that $\partial_x \varphi(\bar x, \bar y)\times \partial_y \varphi(\bar x, \bar y )\not\subset \partial \varphi(\bar x, \bar y).$}
\end{example}
In the sequel, we will need the following fundamental calculus rule of convex analysis.
\begin{theorem} \label{MoreauRockafellar}{\rm (The Moreau-Rockafellar Theorem) (See \cite[Theorem 0.3.3 on pp. 47--50, Theorem 1 on p.~200]{Ioffe_Tihomirov_1979})} Let $f_1,\dots,f_m$ be proper convex functions on $X$. Then
$$\partial (f_1+\dots+f_m)(x) \supset\partial f_1(x) +\dots+\partial f_m(x)$$ for all $x\in X$.
If, at a point $x^0\in {\rm dom}\,f_1\cap\dots\cap {\rm dom}\,f_m$, all the functions $f_1,\dots,f_m$, except, possibly, one are continuous, then
$$ \partial (f_1+\dots+f_m)(x) = \partial f_1(x) +\dots + \partial f_m(x)$$ for all $x\in X$.
\end{theorem}
Another version of the above Moreau-Rockafellar Theorem, which is based on a geometrical regularity condition of Aubin's type, will be used later on. Note that Aubin \cite[Theorem 4.4, p. 67]{Aubin_1998} only proved this result in a Hilbert space setting, but he observed that it is also valid in a reflexive Banach space setting. It turns out that the reflexivity of the Banach space under consideration can be omitted. A detailed proof of the following theorem can be found in~\cite{Bonnans_Shapiro_2000}.
\begin{theorem} \label{Aubin_sumrule}{\rm (See {\cite[Theorem 2.168 and Remark 2.169]{Bonnans_Shapiro_2000}})} Let $X$ be a Banach space. If $f,g: X \rightarrow \overline{\mathbb{R}}$ are proper, closed, convex functions and the regularity condition
\begin{align}\label{Regularity_condition}
0 \in\, {\rm{int}} ({\rm dom}\,f- {\rm dom}\,g)
\end{align}
holds, then for any $x \in ({\rm dom}\, f)\cap ({\rm dom}\,g)$ we have
\begin{align}\label{Sum_rule}
\partial (f + g )(x) =\partial f(x) + \partial g (x),
\end{align}
where ${\rm{int}}\,\Omega$ denotes the interior of a set $\Omega$.
\end{theorem}
By using the indicator functions of convex sets, one can easily derive from Theorem~\ref{MoreauRockafellar} the next intersection formula.
\begin{proposition}\label{intersection_formula} {\rm (See \cite[p.~205]{Ioffe_Tihomirov_1979})} Let $A_1,A_2,\dots,A_m$ be convex subsets of $X$ and $A=A_1\cap A_2\cap\dots\cap A_m$. Suppose that $A_1\cap ({\rm{int}}\, A_2)\cap\dots\cap ({\rm{int}}\, A_m) \not= \emptyset$. Then,
$$ N(x;A)=N(x;A_1)+N(x;A_2)+\dots+N(x;A_m), \quad \forall x\in X.$$
\end{proposition}
The forthcoming theorem characterizes continuity of extended-real-valued convex functions defined on Hausdorff locally convex topological vector spaces.
\begin{theorem}
{\rm(See \cite[p.~170]{Ioffe_Tihomirov_1979})}
Let $f$ be a proper convex function on a Hausdorff locally convex topological vector space $X$. Then the following assertions are equivalent:
\par {\rm{(i)}} $f$ is bounded from above on a neighborhood of a point $x\in X$;
\par{\rm{(ii)}} $f$ is continuous at a point $x\in X$;
\par {\rm{(iii)}} ${\rm{int(epi}}\, f) \not=\emptyset;$
\par {\rm{(iv)}} ${\rm{int(dom}}\, f) \not=\emptyset$ and $f$ is continuous on ${\rm{int(dom}}\, f).$
Moreover,
$$ {\rm{int(epi}}\, f)=\{ (\alpha , x) \in \Bbb{R} \times X \mid x \in {\rm{int(dom}}\, f), \alpha > f(x)\}. $$
\end{theorem}
\medskip
The following infinite-dimensional version of the Farkas lemma \cite[p.~200]{Rockafellar_1970} has been obtained by Bartl \cite{Bartl_2008}.
\begin{lemma}\label{Farkas_lemma}{\rm{(See \cite[Lemma~1]{Bartl_2008})}}
Let $W$ be a vector space over the reals. Let $A: W \rightarrow \Bbb{R}^m$ be a linear mapping and $\gamma : W \rightarrow \Bbb{R}$ be a linear functional. Suppose that $A$ is represented in the form $A=(\alpha_i)_i^m$, where each $\alpha_i:W\to \Bbb{R}$ is a linear functional (i.e.,
for each $x\in W$, $A(x)$ is a column vector whose $i-th$ component is $\alpha_i(x)$, for $i=1,\dots,m$). Then, the inequality $\gamma(x) \le 0$ is a consequence of the inequalities system
$$\alpha_1(x) \le 0, \ \alpha_2(x) \le 0,\dots, \ \alpha_m(x) \le 0$$
if and only if there exist nonnegative real numbers $\lambda_1, \lambda_2,\dots,\lambda_m \ge 0$ such that
$$\gamma=\lambda_1\alpha_1+\dots+\lambda_m \alpha_m.$$
\end{lemma}
Finally, let us recall a lemma from \cite{AnYen} which describes the normal cone of the intersection of finitely many affine hyperplanes. The proof of this result has been done by applying Lemma \ref{Farkas_lemma}.
\begin{lemma}\label{lemma2} {\rm (See {\cite[Lemma 5.2]{AnYen}})} Let $X, Y$ be Hausdorff locally convex topological vector spaces.
Let there be given vectors $(x_j^*,y_j^*)\in X^*\times Y^*$ and real numbers $\alpha_j \in \Bbb{R},\ j=1,\dots,k $. Set
$$Q_j=\left\{ (x,y)\in X\times Y \mid \langle (x_j^*,y_j^*), (x,y) \rangle = \alpha_j\right\}.$$ Then, for each $(\bar x, \bar y)\in \bigcap\limits_{j=1}^k Q_j$, it holds that
\begin{align}\label{formulanon}
N\left( (\bar x, \bar y); \bigcap\limits_{j=1}^k Q_j\right)={\rm{span}}\{ (x_j^*, y_j^*) \mid j =1,\dots,k \},
\end{align}
where ${\rm{span}}\{ (x_j^*, y_j^*) \mid j =1,\dots,k\}$ denotes the linear subspace generated by the vectors $(x_j^*, y_j^*)$, $j =1,\dots,k$.
\end{lemma}
\section{Optimality conditions}
\markboth{\centerline{\it Optimality conditions}}{\centerline{\it D.T.V.~An
and N.D.~Yen}}
Optimality conditions for convex optimization problems, which can be derived from the calculus rules of convex analysis, have been presented in many books and research papers. To make our paper self-contained and easy for reading, we are going to present systematically some optimality conditions for convex programs under inclusion constraints and for convex optimization problems under geometrical and functional constraints. Observe that these conditions lead to certain Lagrange multiplier sets which are used in our subsequent differential stability analysis of parametric convex programs.
Let $X$ and $Y$ be Hausdorff locally convex topological vector spaces. Let $\varphi: X \times Y \rightarrow \overline{\mathbb{R}}$ be a proper convex extended-real-valued function.
\subsection{Problems under inclusion constraints}
Given a convex multifunction $G: X \rightrightarrows Y$, we consider the \textit{parametric convex optimization problem under an inclusion constraint}
\begin{align*}
(P_x) \quad \quad \quad \min\{\varphi(x,y)\mid y \in G(x)\}
\end{align*}
depending on the parameter $x$. The \textit{optimal value function}
$\mu: X \rightarrow \overline{\mathbb{R}}$ of problem $(P_x)$ is
\begin{align}\label{marginalfunction}
\mu(x):= \inf \left\{\varphi (x,y)\mid y \in G(x)\right\}.
\end{align}
The usual convention $\inf \emptyset =+\infty$ forces $\mu(x)=+\infty$ for every $x \notin {\rm{dom}}\, G.$
The \textit{solution map} $M: {\rm {dom}}\, G \rightrightarrows Y $ of that problem is defined by
\begin{align}\label{solution_map}
M(x):=\{y \in G(x)\mid \mu(x)= \varphi (x,y)\}.
\end{align}
The next theorem describes some necessary and sufficient optimality conditions for $(P_x)$ at a given parameter $\bar x \in X$.
\begin{theorem}\label{MR_version} Let $\bar x \in X$. Suppose that at least one of the following regularity conditions is satisfied:
\par {\rm{(a)}} $ {\rm int }\, G(\bar x) \cap {\rm{dom}}\, \varphi(\bar x, .) \not= \emptyset,$
\par {\rm{(b)}} $\varphi (\bar x, .)$ is continuous at a point belonging to $G(\bar x).$\\
Then, one has $\bar y\in M(\bar {x})$ if and only if
\begin{align} \label{NS_condition}
0 \in \partial_y \varphi(\bar x, \bar y) + N(\bar y; G(\bar x)).
\end{align}
\end{theorem}
\begin{proof}
Consider the function $\varphi_G(y)= \varphi (\bar x, y) + \iota_{G(\bar x)}( y)$, where $\iota_{G(\bar x)}(\cdot)$ is the indicator function of the convex set $G(\bar x)$. The latter means that $\iota_{G(\bar x)}(y)=0$ for $y\in G(\bar x)$ and $\iota_{G(\bar x)}(y)=+\infty$ for $y\notin G(\bar x)$. It is clear that $\bar y \in M(\bar x)$ if and only if the function $\varphi_G$ attains its minimum at $\bar y$. Hence, by \cite[Proposition~1, p. 81]{Ioffe_Tihomirov_1979}, $\bar y \in M(\bar x)$ if and only if
\begin{align}\label{NS_condition1}
0 \in \partial \varphi_G(\bar y)=\partial\bigg(\varphi (\bar x, .)+\iota_{G(\bar x)}(.)\bigg )(\bar y).
\end{align}
Since $G(\bar x)$ is convex, $\iota_{G(\bar x)}(\cdot)$ is convex. Clearly, $\iota_{G(\bar x)}(\cdot)$ is continuous at every point belonging to $ {\rm int}\,G(\bar x)$. Thus, if the regularity condition (a) is fulfilled, then $\iota_{G(\bar x)}(\cdot)$ is continuous at a point in ${\rm dom}\, \varphi(\bar x, .)$. By Theorem~\ref{MoreauRockafellar}, from \eqref{NS_condition1} one has
\begin{align*}
0 \in \partial\bigg(\varphi (\bar x, .)+\iota_{G(\bar x)}(.)\bigg)(\bar y)&=\partial_y \varphi(\bar x, \bar y)+ \partial\iota_{G(\bar x)}(\bar y)\\
&=\partial_y \varphi(\bar x, \bar y) + N(\bar y; G(\bar x)).
\end{align*}
Consider the case where (b) holds. Since ${\rm dom}\,\iota_{G(\bar x)}(\cdot)= G(\bar x),$ $\varphi(\bar x,.)$ is continuous at a point in ${\rm dom}\,{\iota_{G(\bar x )}(\cdot)}$. Then, by Theorem \ref{MoreauRockafellar} one can obtain~\eqref{NS_condition} from~\eqref{NS_condition1}.
$\hfill\Box$
\end{proof}
The sum rule in Theorem \ref{Aubin_sumrule} allows us to get the following result.
\begin{theorem}
Let $X,Y$ be Banach spaces, $\varphi: X \times Y \rightarrow \overline{\mathbb{R}}$ a proper, closed, convex function. Suppose that $G:X \rightrightarrows Y$ is a convex multifunction, whose graph is closed. Let $\bar x \in X$ be such that the regularity condition
\begin{align}\label{Aubin_RC}
0\in {\rm int}\, \big ({\rm dom}\, \varphi (\bar x,.) - G(\bar x) \big )
\end{align}
is satisfied. Then, $\bar y \in M(\bar x)$ if and only if
\begin{align} \label{NS_condition3}
0 \in \partial_y \varphi(\bar x, \bar y) + N(\bar y; G(\bar x)).
\end{align}
\end{theorem}
\begin{proof}
The proof is similar to that of Theorem \ref{MR_version}. Namely, if the regularity condition \eqref{Aubin_RC} is fulfilled, then instead of Theorem~\ref{MoreauRockafellar} we can apply Theorem~\ref{Aubin_sumrule} to the case where $X \times Y$, $\varphi(\bar x,.)$, and ${\iota_{G(\bar x)}(\cdot)}$, respectively, play the roles of $X$, $f$, and $g$.
$\hfill\Box$
\end{proof}
\subsection{Problems under geometrical and functional constraints}
We now study optimality conditions for \textit{convex optimization problems under geometrical and functional constraints}. Consider the program
\begin{align*}(\Tilde{P}_x) \quad \ \; \min \left\{\varphi (x,y) \mid y\in C(x),\ g_i(x,y) \le 0, \ i \in I,\ h_j(x,y)=0,\ j\in J \right\} \end{align*}
depending on parameter $x$, where $g_i:X\times Y\to \mathbb{R}$, $i\in I:=\{1,\dots,m\}$, are continuous convex functions, $h_j:X\times Y\to \mathbb{R}$, $j\in J:=\{1,\dots,k\}$, are continuous affine functions, and $C(x):=\{y \in Y: (x,y) \in C \}$ with $C \subset X \times Y$ being a convex set. For each $x \in X$, we put
\begin{align}\label{constraint}
G(x)=\left\{y \in Y \mid y\in C(x),\ g(x,y) \le 0, \ h(x,y)=0\right\},
\end{align}
where $$g(x,y):=(g_1(x,y),\dots,g_m(x,y))^T,\ \; h(x,y):=(h_1(x,y),\dots,h_k(x,y))^T,$$ with $^T$ denoting matrix transposition, and the inequality $z\le w$ between two vectors in $\mathbb{R}^m$ means that every coordinate of $z$ is less than or equal to the corresponding coordinate of $w$. It is easy to show that the multifunction $G(\cdot)$ given by \eqref{constraint} is convex. Fix a point $\bar x\in X$ and recall that \begin{align}\label{C_x_bar}
C(\bar x)=\{y\in Y\mid (\bar x,y)\in C\}.
\end{align}
The next lemma describes the normal cone to a sublevel set of a convex function.
\begin{lemma}\label{lemma1} {\rm (See \cite[Proposition 2 on p.~206]{Ioffe_Tihomirov_1979})}
Let $f$ be a proper convex function on $X$, which is continuous at a point $x_0\in X$. Assume that the inequality $f(x_1)<f(x_0)=\alpha_0$ holds for some $x_1\in X$. Then,
\begin{align}\label{formula_for_normal_cone}
N(x_0;[f \le \alpha_0])=K_{\partial f(x_0)},
\end{align}
where $[f \le \alpha_0]:= \{ x \mid f(x) \le \alpha_0\}$ is a sublevel set of $f$ and
$$ K_{\partial f(x_0)}:=\{u^* \in X^* \mid u^*=\lambda x^*,\ \lambda \ge 0,\ x^*\in \partial f(x_0)\}$$ is the cone generated by the subdifferential of $f$ at $x_0$.
\end{lemma}
Optimality conditions for convex optimization problems under geometrical and functional constraints can be formulated as follows.
\begin{theorem} If $\varphi(\bar x, .)$ is continuous at a point $y^0\in {\rm int}\,C(\bar x)$, $g_i(\bar x,y^0)<0$ for all $i\in I$ and $h_j(\bar x,y^0)=0$ for all $j\in J$, then for a point $\bar y \in G(\bar x)$ to be a solution of $(\Tilde{P}_{\bar x})$, it is necessary and sufficient that there exist $\lambda_i \ge 0,$ $i\in I,$ and $\mu_j \in \Bbb{R},$ $j\in J,$ such that
\\ {\rm {(a)}} $0 \in \partial_y \varphi(\bar x,\bar y)+ \sum\limits_{i\in I} \lambda_i \partial_y g_i(\bar x, \bar y)+\sum\limits_{j\in J}\mu_j \partial_y h_j(\bar x, \bar y)+ N(\bar y;C(\bar x)); $
\\ {\rm {(b)}} $\lambda_ig_i (\bar x, \bar y)=0, \ i\in I.$
\end{theorem}
\begin{proof}
For any $\bar x \in X$, let $\bar y \in G(\bar x)$ be given arbitrarily. Note that $(\Tilde{P}_{\bar x})$ can be written in the form
$$\min \big\{\varphi(\bar x,y)\mid y \in G(\bar x)\big\}.$$
If $\varphi(\bar x, .)$ is continuous at a point $y^0$ with $y^0\in {\rm int}\,C(\bar x)$, $g_i(\bar x,y^0)<0$ for all $i\in I$, and $h_j(\bar x,y^0)=0$ for all $j\in J$, then the regularity condition (b) in Theorem~\ref{MR_version} is satisfied. Consequently, $\bar y\in M(\bar x)$ if and only if
\begin{align}\label{Fermat}
0\in \partial_y \varphi (\bar x, \bar y) +N(\bar y; G(\bar x)).
\end{align}
We now show that
\begin{equation}\label{cone_normal1}
\begin{split}
N(\bar y; G(\bar x))=\left\{\!\sum\limits_{i\in I(\bar x, \bar y)} \lambda_i \partial_y g_i(\bar x, \bar y)\!+\!\sum\limits_{j\in J}\mu_j \partial_y h_j(\bar x, \bar y)\!+\! N(\bar y;C(\bar x))\right\},
\end{split}
\end{equation}
with $I(\bar x, \bar y):=\{i \mid g_i(\bar x, \bar y)=0, \, i\in I\},$ $\lambda_i \ge 0,\, i\in I,\, \mu_j \in \Bbb{R}, \, j\in J.$
First, observe that
\begin{align}\label{normal_gph}
G(\bar x)= \left( \bigcap\limits_{i \in I} \Omega_i(\bar x) \right) \cap \left( \bigcap\limits_{j \in J} \mathcal{Q}_i(\bar x) \right)\cap C,
\end{align}
where $\Omega_i(\bar x)= \{ y\mid g_i(\bar x,y) \le 0\} (i \in I)$ and $\mathcal{Q}_j(\bar x)= \{ y\mid h_j(\bar x,y) = 0\} (j \in J)$ are convex sets. By our assumptions, we have
\begin{align*}
y^0\in \left( \bigcap\limits_{i \in I}{\rm int\,} \Omega_i(\bar x) \right) \cap \left( \bigcap\limits_{j \in J} \mathcal{Q}_i(\bar x) \right)\cap ({\rm int}\,C).
\end{align*}
Therefore, according to Proposition \ref{intersection_formula} and formula \eqref{normal_gph}, one has
\begin{align}\label{Normal_cap}
N(\bar y;G(\bar x))= \sum\limits_{i \in I} N( \bar y; \Omega_i(\bar x)) + N\left( \bar y; \bigcap\limits_{j\in J}\mathcal{Q}_j(\bar x) \right)+N(\bar y;C(\bar x)).
\end{align}
On one hand, by Lemma \ref{lemma1}, for every $i \in I(\bar x, \bar y)$ we have
\begin{align}\label{normal_Omega}
N( \bar y; \Omega_i(\bar x))= K_{\partial_y g_i(\bar x, \bar y)}{=\{\lambda_i y^*\mid \lambda_i \ge 0,\; y^*\in \partial_y g_i(\bar x, \bar y)\}}.
\end{align}
On the other hand, according to Lemma \ref{lemma2} and the fact that {$$h_j(x,y) = \langle x_j^*, x\rangle + \langle y_j^*, y\rangle - \alpha_j\quad \big ((x_j^*,y_j^*)\in X^*\times Y^*,\ \alpha_j\in\mathbb R\big),$$} we can assert that
\begin{align}\label{normal_Q}
{N\left( \bar y; \bigcap\limits_{j\in J}\mathcal{Q}_j(\bar x) \right)}={\rm{span}}\{ y^*_j\mid j\in J \}={\rm{span}}\{ \partial_y h_j(\bar x, \bar y) \mid j \in J\},
\end{align}
Combining \eqref{Normal_cap}, \eqref{normal_Omega}, and \eqref{normal_Q}, we obtain \eqref{cone_normal1}. So the assertion of the theorem is valid. $\hfill\Box$
\end{proof}
\section{Subdifferential Estimates via Multiplier Sets}
\markboth{\centerline{\it Subdifferential Estimates via Multiplier Sets}}{\centerline{\it D.T.V.~An
and N.D.~Yen}}
The following result on differential stability of convex optimization problems under geometrical and functional constraints has been obtained in \cite{AnYen}.
\begin{theorem}\label{Thm5.21} {\rm (See \cite[Theorem~5.2]{AnYen})} For every $j\in J$, suppose that
$$ h_j(x,y) = \langle (x_j^*,y_j^*), (x,y) \rangle - \alpha_j, \ \, \alpha_j \in \Bbb{R}.$$ If $\varphi$ is continuous at a point $(x^0,y^0)$ with $(x^0,y^0)\in {\rm int}\,C$, $g_i(x^0,y^0)<0,$ for all $i\in I$ and $h_j(x^0,y^0)=0,$ for all $j\in J$, then for any $\bar x \in {\rm{dom}}\, \mu$, with $\mu(\bar x)\neq -\infty$, and for any $\bar y \in M(\bar x)$ we have
\begin{align}\label{UpperEstimate1}
\partial \mu(\bar x) = \bigcup\limits_{(x^*,y^*) \in \partial \varphi(\bar x, \bar y)} \big\{x^* + \tilde{Q}^*\big\}
\end{align}
and
\begin{align}\label{UpperEstimate2}
\partial^\infty \mu(\bar x) = \bigcup\limits_{(x^*,y^*) \in \partial^\infty \varphi(\bar x, \bar y)} \big\{x^* + \tilde{Q}^*\big\},
\end{align}
where
\begin{align}\label{Q_star}
\tilde{Q}^*&:=\bigg\{u^* \in X^* \mid (u^*,-y^*) \in A+N((\bar x,\bar y);C) \bigg\}
\end{align}
with
\begin{align}\label{sum_A} A:=\sum\limits_{i\in I(\bar x, \bar y)} {\rm{cone}}\, \partial g_i(\bar x, \bar y) + {\rm{span}}\{(x_j^*,y_j^*), \,j \in J\}.
\end{align}
\end{theorem}
Our aim in this section is to derive formulas for computing or estimating the subdifferential of the optimal value function of $(\tilde{P}_x)$ through suitable multiplier sets.
\medskip
The \textit{Lagrangian function} corresponding to the parametric problem $(\Tilde{P}_x)$ is
\begin{align}
\label{Lagrangian_function}
L(x,y,\lambda,\mu):= \varphi(x,y) + \lambda^T g(x,y) + \mu^T h(x,y)+{\iota_C((x,y))},
\end{align}
where $\lambda=(\lambda_1,\lambda_2,...,\lambda_m)\in \mathbb{R}^m $ and $\mu=(\mu_1,\mu_2,...,\mu_k)\in \mathbb{R}^k.$ For each pair $(x,y)\in X\times Y$, by $\Lambda_0(x,y)$ we denote the set of all the multipliers $\lambda\in \mathbb{R}^m$ and $\mu\in \mathbb{R}^k$ with $\lambda_i\geq 0$ for all $i\in I$ and $\lambda_i=0$ for every $i\in I\setminus I(x,y)$, where $ I(x,y)=\{i\in I\mid g_i(x,y)=0\}$.
\medskip
For a parameter~$\bar x$, the \textit{Lagrangian function} corresponding to the unperturbed problem $(\Tilde{P}_{\bar x})$ is
\begin{align}
\label{Lagrangian_function2}
L(\bar x,y,\lambda,\mu)= \varphi(\bar x, y) + \lambda^T g(\bar x,y) + \mu^T h(\bar x,y)+{\iota_C((\bar x,y))}.
\end{align} Denote by $\Lambda(\bar x, \bar y)$ the \textit{Lagrange multiplier set} corresponding to an optimal solution $\bar y$ of the problem $(\Tilde{P}_{\bar x})$. Thus, $\Lambda(\bar x, \bar y)$ consists of the pairs $(\lambda, \mu)\in \mathbb{R}^m \times \mathbb{R}^k$ satisfying
\begin{equation*}\begin{cases}
0 \in \partial _y L(\bar x, \bar y, \lambda, \mu),
\\
\lambda_i g_i(\bar x, \bar y)=0, \; i=1,\dots,m,\\
\lambda_i \ge 0,\; i=1,\dots,m,
\end{cases}
\end{equation*}
where $\partial_y L(\bar x, \bar y, \lambda, \mu)$ is the subdifferential of the function $L(\bar x , ., \lambda, \mu)$ defined by \eqref{Lagrangian_function2} at $\bar y$. It is clear that ${\iota_C((\bar x,y))}={\iota_{C(\bar x)}(y)}$, where $C(\bar x)$ has been defined by \eqref{C_x_bar}.
{Based on the multiplier} set $\Lambda_0(x,y)$, the next theorem provides us with a formula for computing the subdifferential of the optimal value function $\mu(x).$
\begin{theorem}\label{computing_subdifferential}
Suppose that
$ h_j(x,y) = \langle (x_j^*,y_j^*), (x,y) \rangle - \alpha_j, \ \alpha_j \in \Bbb{R}, \ j\in J,$ and $M(\bar x)$ is nonempty for some $\bar x \in {\rm dom}\, \mu $. If $\varphi$ is continuous at a point $(x^0,y^0)\in {\rm int}\,C$, $g_i(x^0,y^0)<0$ for all $i\in I$ and $h_j(x^0,y^0)=0$ for all $j\in J$ then, for any $\bar y \in M(\bar x)$, one has
\begin{align}\label{in_estimate}
\partial \mu (\bar x)= \left\{ \bigcup\limits_{(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)} {\rm pr}_{X^*}\bigg(\partial L(\bar x, \bar y, \lambda, \mu)\cap \big(X^* \times \{0\}\big)\bigg)\right\},
\end{align}
where $\partial L(\bar x, \bar y, \lambda, \mu)$ is the subdifferential of the function $L(., ., \lambda, \mu)$ at $(\bar x,\bar y)$ and, for any $(x^*,y^*)\in X^*\times Y^*$, ${\rm pr}_{X^*}(x^*,y^*):=x^*$.
\end{theorem}
\begin{proof} {To prove the inclusion ``$\subset$" in \eqref{in_estimate}, take any $\bar x^* \in \partial \mu(\bar x)$}.
By Theorem \ref{Thm5.21}, there exist $(x^*,y^*)\in \partial\varphi (\bar x, \bar y)$ and $u^*\in\tilde{Q}^*$ such that $\bar x^*=x^*+u^*$. According to \eqref{Q_star}, the condition $u^*\in\tilde{Q}^*$ means that \begin{align}\label{basic_inclusion} (u^*,-y^*) \in N((\bar x,\bar y);C)+A,\end{align} {where $A$ is given by \eqref{sum_A}}. Adding the inclusion $(x^*,y^*)\in \partial\varphi (\bar x, \bar y)$ and that one in \eqref{basic_inclusion} yields
\begin{align*}
(x^*+u^*,0) \in (x^*, y^*)+A+N((\bar x,\bar y);C).
\end{align*} Hence,
\begin{align}\label{incl_for_bar_x_star} (\bar x^*,0)\in \partial\varphi (\bar x, \bar y) +A+N((\bar x,\bar y);C).
\end{align} For every $(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)$, the assumptions made on the functions $\varphi$, $g_i$, $h_j$, and the set $C$ allow us to apply the Moreau-Rockafellar Theorem (see Theorem~\ref{MoreauRockafellar}) to the Lagrangian function $L(x,y,\lambda,\mu)$ defined by \eqref{Lagrangian_function} to get
\begin{align}\label{sum_rule_Lagragian1}
\partial\, L(\bar x, \bar y, \lambda, \mu)= \partial \varphi(\bar x, \bar y) \!\!+\!\! \sum\limits_{i\in I(\bar x,\bar y)} \lambda_i \partial g_i(\bar x, \bar y)\!+\! \sum\limits_{j\in J}\mu_j \partial h_j(\bar x, \bar y)\!+\!N((\bar x, \bar y);C).
\end{align} Since $\partial h_j(\bar x, \bar y)=\{(x_j^*,y_j^*)\}$, from \eqref{sum_rule_Lagragian1} it follows that
\begin{align}\label{basic-equality}\partial\varphi (\bar x, \bar y) +A+N((\bar x,\bar y);C)=\bigcup\limits_{(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)} \partial L(\bar x, \bar y, \lambda, \mu).\end{align} So, \eqref{incl_for_bar_x_star} means that
\begin{align}\label{incl_for_bar_x_star_2} \bar x^*\in\bigcup\limits_{(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)} {\rm pr}_{X^*} \bigg(\partial L(\bar x, \bar y, \lambda, \mu)\cap \big(X^* \times \{0\}\big)\bigg).
\end{align}
Thus, the inclusion ``$\subset$" in \eqref{in_estimate} is valid. To obtain the reverse inclusion, fixing any $\bar x^*$ satisfying \eqref{incl_for_bar_x_star_2} we have to show that $\bar x^*\in \partial \mu(\bar x)$. As it has been noted before, \eqref{incl_for_bar_x_star_2} is equivalent to \eqref{incl_for_bar_x_star}. Select a pair $(x^*,y^*)\in\partial\varphi (\bar x, \bar y)$ satisfying
\begin{align*} (\bar x^*,0)\in (x^*,y^*)+A+N((\bar x,\bar y);C).
\end{align*}
Then, for $u^*:=\bar x^*-x^*$, one has
\begin{align*} (x^*+u^*,0)\in (x^*,y^*)+A+N((\bar x,\bar y);C).
\end{align*} Therefore, the inclusion \eqref{basic_inclusion} holds. Hence, thanks to \eqref{Q_star} and \eqref{UpperEstimate1}, the vector $\bar x^*=x^*+u^*$ belongs to $\partial \mu(\bar x)$.
The proof is complete. $\hfill\Box$
\end{proof}
{As an illustration for Theorem \ref{computing_subdifferential}, let us consider the following simple example.}
\begin{example}\label{Ex2} {\rm Let $X=Y=\mathbb R$, $C=X\times Y$, $\varphi(x,y)=|x+y|$, $m=1$, $k=0$ (no equality functional constraint), $g_1(x,y)=y$ for all $(x,y)\in X\times Y$. {Choosing $\bar x=0$, one has $M(\bar x)=\{\bar y\}$ with $\bar{y}=0$}. It is clear that $\Lambda_0(\bar x,\bar y)=[0,\infty)$ and $L(x,y,\lambda)=\varphi(x,y)+\lambda y$. As in Example \ref{Ex1}, we have $$\partial \varphi (\bar x, \bar y)={\rm co}\left\{(1,1)^T,\, (-1,-1)^T \right\}.$$ Since $\partial L(\bar x,\bar y,\lambda)=\partial \varphi (\bar x, \bar y)+\{(0,\lambda)\}$, by \eqref{in_estimate} we can compute \begin{align*}
\partial \mu (\bar x)& = \left\{ \bigcup\limits_{\lambda \in \Lambda_0(\bar x, \bar y)} {\rm pr}_{X^*}\bigg(\partial L(\bar x, \bar y, \lambda)\cap \big(X^* \times \{0\}\big)\bigg)\right\}\\
& ={\rm pr}_{X^*}\left[\bigg(\bigcup\limits_{\lambda \in \Lambda_0(\bar x, \bar y)} \partial L(\bar x, \bar y, \lambda)\bigg)\cap \big(X^* \times \{0\}\big)\right]\\
& = {\rm pr}_{X^*}\bigg\{\left[{\rm co}\left\{(1,1)^T,\, (-1,-1)^T \right\} +\Big(\{0\}\times \mathbb{R}_+\Big)\right]\cap \big(X^* \times \{0\}\big)\bigg\}\\
& = [-1,0].
\end{align*} To verify this result, observe that
\begin{align*}
\mu (x)=\inf\left\{|x+y|\mid y \leq 0\right\}=\begin{cases}
0, & {\rm if}\ x\geq 0,\\
-x, & {\rm if}\ x<0.
\end{cases}
\end{align*} So we find $\partial \mu (\bar x)= [-1,0]$, justifying \eqref{in_estimate} for the problem under consideration.
}
\end{example}
We are now in a position to establish an upper estimate for the subdifferential $\mu(.)$ at $\bar x$ by using the Lagrange multiplier set $\Lambda(\bar x, \bar y)$ corresponding to a solution $\bar y$ of $(\Tilde{P}_{\bar x})$.
\begin{theorem}\label{outer_theorem}
Under the assumptions of Theorem \ref{computing_subdifferential}, one has
\begin{align}\label{outer_estimate}
\partial \mu (\bar x) \subset \bigcup\limits_{(\lambda, \mu) \in \Lambda(\bar x, \bar y)} \partial_x L(\bar x, \bar y, \lambda,\mu),
\end{align}
where $\partial_x L(\bar x, \bar y, \lambda, \mu)$ stands for the subdifferential of $L(., \bar y, \lambda, \mu) $ at $\bar x$.
\end{theorem}
\begin{proof} Fix an arbitrary vector $\bar {x}^* \in \partial\, \mu(\bar x)$. The arguments in the first part of the proof of Theorem \ref{computing_subdifferential} show that
\eqref{incl_for_bar_x_star} and \eqref{basic-equality} are valid. Hence, we can find a vector $(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)$ such that \begin{align}
\label{incl2_for_bar_x_star} (\bar x^*,0)\in \partial L(\bar x, \bar y, \lambda, \mu).
\end{align} Using the definition of subdifferential, from \eqref{incl2_for_bar_x_star} we can deduce that
$$\langle \bar x^*,x-\bar x\rangle \leq L(x, \bar y, \lambda, \mu)- L(\bar x, \bar y, \lambda, \mu)\quad \forall x\in X$$
and
$$\langle 0,y-\bar y\rangle \leq L(\bar x,y, \lambda, \mu)- L(\bar x, \bar y, \lambda, \mu)\quad \forall y\in Y.$$
Hence,
\begin{align}
\label{two_incls}
\bar x^*\in \partial_xL(\bar x, \bar y, \lambda, \mu),\ \;
0\in \partial_y L(\bar x, \bar y, \lambda, \mu).
\end{align} Since $(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)$, one has $\lambda_ig_i(\bar x,\bar y)=0$ and $\lambda_i\geq 0$ for every $i\in I$. Therefore, the second inclusion in \eqref{two_incls} implies that $(\lambda, \mu) \in \Lambda (\bar x, \bar y)$. Then,~\eqref{outer_estimate} follows from the first inclusion in \eqref{two_incls}.$\hfill\Box$
\end{proof}
The next example shows that the inclusion in Theorem \ref{outer_theorem} can be strict.
\begin{example}{\rm Let $X=Y=\mathbb R$, $C=X\times Y$, $\varphi(x,y)=|x+y|$, $m=1$, $k=0$ (no equality functional constraint), $g_1(x,y)=y$ for all $(x,y)\in X\times Y$. Choosing $\bar x=0$, we note that $M(\bar x)=\{\bar y\}$ with $\bar{y}=0$. We have $L(x,y,\lambda)=\varphi(x,y)+\lambda y$ and \begin{align*}\Lambda(\bar x,\bar y)&=\{\lambda \ge 0\mid 0 \in \partial_y L(\bar x, \bar y, \lambda) \}\\
&=\{\lambda \ge 0\mid 0 \in [-1,1]+\lambda \}\\
&=[0,1].
\end{align*} As in Example \ref{Ex2}, one has $\partial \mu(\bar x)=[-1,0].$ We now compute the right-hand-side of \eqref{outer_estimate}. By simple computation, we obtain $\partial_x L(\bar x, \bar y, \lambda)=[-1,1]$ for all $\lambda \in \Lambda(\bar x, \bar y).$ Then $\bigcup\limits_{\lambda \in \Lambda(\bar x, \bar y)} \partial_x L(\bar x, \bar y, \lambda)=[-1,1].$ Therefore, in this example, inclusion~\eqref{outer_estimate} is strict.
}
\end{example}
\section{Computation of the singular subdifferential}
\markboth{\centerline{\it Computation of the singular subdifferential}}{\centerline{\it D.T.V.~An
and N.D.~Yen}}
First, we observe that $x\in {\rm dom}\,\mu$ if and only if
$$\mu(x)={\rm inf}\{\varphi(x,y)\mid y \in G(x) \}< \infty,$$ with $G(x)$ being given by \eqref{constraint}.
Since the strict inequality holds if and only if {there exists} $y \in G(x)$ with $(x,y)\in {\rm dom}\, \varphi$, we have
\begin{align}
\label{new_problem}
{\iota_{{\rm dom}\,\mu}(x)}={\rm inf}\{{\iota_{{\rm dom}\, \varphi}((x,y))}\mid y \in G(x) \}.
\end{align}
To compute the singular subdifferential of $\mu(.)$, let us consider the minimization problem
\begin{align*}\big(P^\infty_x\big)\quad \begin{cases} {\iota_{{\rm dom}\, \varphi}((x,y))}\to \inf & \\
{\rm subject\ to}\ \; y\in C(x),\ g_i(x,y) \le 0,\ i \in I,\ h_j(x,y)=0,\ j\in J.&
\end{cases}\end{align*}
The Lagrangian function corresponding to $(P^\infty_x)$ is
\begin{align}
\label{new_largrange_function}
\widehat{L}(x,y,\lambda,\mu):= {\iota_{{\rm dom}\, \varphi}((x,y))} + \lambda^T g(x,y) + \mu^T h(x,y)+{\iota_C((x,y))},
\end{align}
where $\lambda=(\lambda_1,\lambda_2,...,\lambda_m)\in \mathbb{R}^m,$ $\mu=(\mu_1,\mu_2,...,\mu_k)\in \mathbb{R}^k$.
\medskip
Interpreting $(P^\infty_x)$ as a problem of the form $(\widetilde P_x)$, where ${\iota_{{\rm dom}\, \varphi}((x,y))}$ plays the role of $\varphi(x,y)$, we can apply Theorem~\ref{computing_subdifferential} (resp., Theorem~\ref{outer_theorem}) to compute (resp., estimate) the singular subdifferential of $\mu(.)$ as follows.
\begin{theorem}\label{thm_singular_computing}
Under the hypotheses of Theorem \ref{computing_subdifferential}, for any $\bar y \in M(\bar x)$, one has
\begin{align}\label{singular_computing}
\partial^\infty \mu (\bar x)= \left\{ \bigcup\limits_{(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)} {\rm pr}_{X^*} \bigg(\partial\widehat{L}(\bar x, \bar y, \lambda, \mu)\cap \big(X^* \times \{0\}\big)\bigg)\right\},
\end{align}
where \begin{align}\label{singular_computing2}\partial \widehat{L}(\bar x, \bar y, \lambda, \mu)\!=\!\partial^\infty\varphi(\bar x, \bar y) \!+\! \sum\limits_{i\in I(\bar x,\bar y)} \lambda_i \partial g_i(\bar x, \bar y)\!+\! \sum\limits_{j\in J}\mu_j \partial h_j(\bar x, \bar y)\!+\!N((\bar x, \bar y);C)\end{align} is the subdifferential of the function $\widehat{L}(., ., \lambda, \mu) $ at $(\bar x,\bar y)$, provided that a pair $(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)$ has been chosen.
\end{theorem}
\begin{proof} The inclusion $\bar y \in M(\bar x)$ implies that $(\bar x,\bar y)\in {\rm dom}\, \varphi$ and $\bar y \in G(\bar x)$. So, ${\iota_{{\rm dom}\, \varphi}((\bar x,\bar y))}=0$ and $\bar y$ is a feasible point of the problem $\big(P^\infty_{\bar x}\big)$. As ${\iota_{{\rm dom}\, \varphi}((\bar x,y))}\geq 0$ for all $y\in G(\bar x)$, we can assert that $\bar y$ is a solution of~$\big(P^\infty_{\bar x}\big)$. The corresponding optimal value is ${\iota_{{\rm dom}\,\mu}(\bar x)}=0$ (see \eqref{new_problem}). Hence, by Theorem \ref{computing_subdifferential} and formula \eqref{new_problem}, we have
\begin{align*} \partial{\iota_{ {\rm dom}\,\mu}(\bar x)}
= \left\{ \bigcup\limits_{(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)} {\rm pr}_{X^*} \bigg(\partial\widehat{L}(\bar x, \bar y, \lambda, \mu)\cap \big(X^* \times \{0\}\big)\bigg)\right\}.
\end{align*} Since $\partial{\iota_{ {\rm dom}\,\mu}(\bar x)}=\partial^\infty \mu (\bar x)$, the last equality implies \eqref{singular_computing}.
For every $(\lambda, \mu) \in \Lambda_0(\bar x, \bar y)$, remembering that $h_j$, $j\in J$, are affine functions, $\varphi$ is continuous at a point $(x^0,y^0)$ with $(x^0,y^0)\in {\rm int}\,C$, $g_i(x^0,y^0)<0$ for all $i\in I$ and $h_j(x^0,y^0)=0$ for all $j\in J$, we can apply Theorem~\ref{MoreauRockafellar} to the Lagrangian function ${\widehat L}(x,y,\lambda,\mu)$ defined by \eqref{new_largrange_function} to obtain
\begin{align*}\begin{array}{rl}
& \partial \widehat{L}(\bar x, \bar y, \lambda, \mu)\\
& =\partial{\iota_{{\rm dom}\, \varphi}((\bar x,\bar y))} \!\!+\!\! \sum\limits_{i\in I(\bar x,\bar y)} \lambda_i \partial g_i(\bar x, \bar y)\!+\! \sum\limits_{j\in J}\mu_j \partial h_j(\bar x, \bar y)\!+\!N((\bar x, \bar y);C).
\end{array}\end{align*} Combining this with the equality $\partial{\iota_{{\rm dom}\, \varphi}((\bar x,\bar y))}=\partial^\infty\varphi(\bar x, \bar y)$ yields \eqref{singular_computing2}.
$\hfill\Box$
\end{proof}
\begin{remark} {\rm The result in Theorem \ref{thm_singular_computing} can be derived from formula \eqref{UpperEstimate2} by a proof analogous to that of Theorem \ref{computing_subdifferential}.}
\end{remark}
Next, denote by $\Lambda^\infty (\bar x, \bar y)$ the \textit{singular Lagrange multiplier set} corresponding to an optimal solution $\bar y$ of the problem $(P^\infty_{\bar x})$, which consists of the pairs $(\lambda, \mu)\in \mathbb{R}^m \times \mathbb{R}^k$ satisfying
\begin{equation*}\begin{cases}
0 \in \partial _y {\widehat L}(\bar x, \bar y, \lambda, \mu),
\\
\lambda_i g_i(\bar x, \bar y)=0, \; i=1,\dots,m,\\
\lambda_i \ge 0,\; i=1,\dots,m.
\end{cases}
\end{equation*}
Here $\partial_y {\widehat L}(\bar x, \bar y, \lambda, \mu)$ is the subdifferential of the function ${\widehat L}(\bar x , ., \lambda, \mu)$, with ${\widehat L}(x,y, \lambda, \mu)$ being given by \eqref{new_largrange_function}, at $\bar y$.
\begin{theorem}\label{outer_theorem_singular}
Under the assumptions of Theorem \ref{computing_subdifferential}, for any $\bar y \in M(\bar x)$, one has
\begin{align}\label{outer_estimate_singular}
\partial^\infty \mu (\bar x) \subset \bigcup\limits_{(\lambda, \mu) \in \Lambda^\infty (\bar x, \bar y)} \partial_x {\widehat L}(\bar x, \bar y, \lambda,\mu),
\end{align}
where $\partial_x {\widehat L}(\bar x, \bar y, \lambda, \mu)$ stands for the subdifferential of ${\widehat L}(., \bar y, \lambda, \mu) $ at $\bar x$.
\end{theorem}
\begin{proof} To get \eqref{outer_estimate_singular}, it suffices to apply Theorem~\ref{outer_theorem} to the parametric problem~$(P^\infty_{\bar x})$, keeping in mind that $\bar y$ is a solution of $(P^\infty_{\bar x})$. Indeed, taking account of Theorem~\ref{outer_theorem} and \eqref{outer_estimate}, one has
\begin{align*} \partial{\iota_{{\rm dom}\,\mu}(\bar x)}
\subset \bigcup\limits_{(\lambda, \mu) \in \Lambda^\infty (\bar x, \bar y)} \partial_x {\widehat L}(\bar x, \bar y, \lambda,\mu).
\end{align*} As $\partial{\iota_{ {\rm dom}\,\mu}(\bar x)}=\partial^\infty \mu (\bar x)$, this inclusion is equivalent to \eqref{outer_estimate_singular}. $\hfill\Box$
\end{proof}
\section*{Acknowledgements}
\noindent
The first author was supported by the Vietnam Institute for Advanced Study in Mathematics (VIASM) and Thai Nguyen University of Sciences. This research is funded by the National Foundation for Science and Technology Development (Vietnam) under grant number 101.01-2014.37. The authors would like to thank the two anonymous referees for their very careful readings and valuable suggestions which have helped to greatly improve the presentation.
|
\section{Introduction}
It is often said that black holes are the hydrogen atom of quantum gravity since they are systems where aspects of quantum gravity are indispensable for their understanding. If this analogy is to be pursued then studying corrections to the Bekenstein-Hawking entropy should be equivalent to going beyond the Bohr energy levels. History teaches us that the leading term can be obtained without a full understanding of the degrees of freedom but that the corrections require a detailed control of various aspects of the system. In this work, we take some steps into the subleading corrections to the microstate counting of the entropy of a class of magnetically charged asymptotically AdS$_4$ black holes.
Indeed, Strominger and Vafa have demonstrated that string theory provides a framework for the microstate counting of the Bekenstein-Hawking entropy of a class of asymptotically flat black holes \cite{Strominger:1996sh}. Moreover, Sen and collaborators have carried a successful program of understanding logarithmic corrections to various black holes \cite{Banerjee:2010qc,Banerjee:2011jp,Sen:2011ba,Sen:2012cj}.
In the context of the AdS/CFT correspondence \cite{Maldacena:1997re}, a microscopic counting of the Bekenstein-Hawking entropy of a class of black holes has recently been presented by Benini, Hristov and Zaffaroni \cite{Benini:2015eyy,Benini:2016rke}. Understanding black hole entropy in this context is particularly powerful because it does provide a practical path to a fully non-perturbative definition of quantum gravity in asymptotically AdS spacetimes. The basic premise of \cite{Benini:2015eyy} is that the topologically twisted index of ABJM, namely the supersymmetric partition function on $S^1\times S^2$ with background magnetic flux on $S^2$ \cite{Benini:2015noa}, counts the ground state degeneracy of a superconformal quantum mechanics on $S^1$, and that this counting enumerates the microstates of the dual magnetically charged BPS black hole in AdS.
It was in fact demonstrated in \cite{Benini:2015eyy} that the topologically twisted index reproduces the AdS black hole entropy at leading order in the large-$N$ expansion. Here, we wish to extend this correspondence to subleading order by examining the logarithmic corrections on both the field theory and gravity sides of the duality.
In section~\ref{Sec:FieldTheory} we start by reviewing the field theory computation of the topologically twisted index and present numerical evidence pointing to a universal $-1/2 \log N$ correction. We then turn to the gravity calculation in section~\ref{Sec:Gravity}, which first reviews the prescription and special status of logarithmic corrections at the one-loop level. We then discuss the dual calculation in the context of 11-dimensional supergravity, focusing on the contribution coming from the near horizon limit of the magnetically charged BPS black hole solutions. We also discuss the absence of potential contributions coming from the asymptotically AdS$_4$ region. In contrast with the index result, we find $-2\log N$ from the quantum gravity computation, and suggest possible reasons for this discrepancy in section~\ref{Sec:Conclusions}.
\section{The topologically twisted index beyond the large-$N$ limit}\label{Sec:FieldTheory}
The topologically twisted index for three dimensional ${\cal N}=2$ field theories was defined in \cite{Benini:2015noa} (see other related work \cite{Honda:2015yha,Closset:2015rna,Hosseini:2016tor,Hosseini:2016ume,Closset:2016arn}) by evaluating the supersymmetric partition function on $S^1\times S^2$ with a topological twist on $S^2$. When applied to the microstate counting of magnetic AdS$_4$ black holes, the index is computed for ABJM theory, and the topological twist arises from the magnetic fluxes on $S^2$ \cite{Benini:2015eyy,Benini:2016rke}. Since these black holes are constructed in the STU model truncation of four-dimensional SO(8) gauged supergravity, there are a total of four U(1) gauge fields, with corresponding charges $n_a$ satisfying the supersymmetry constraint $\sum n_a=2$.
The topologically twisted index for ABMJ theory was worked out in \cite{Benini:2015eyy}, and reduces to the evaluation of
the partition function
\begin{equation}
Z(y_a,n_a)=\prod_{a=1}^4 y_a^{-\frac{1}{2}N^2 n_a}\sum_{I\in BAE}\frac{1}{\det\mathbb{B}}
\frac{\prod_{i=1}^N x_i^N \tilde{x}_i^N\prod_{i\neq j}\left(1-\frac{x_i}{x_j}\right)\left(1-\frac{\tilde{x}_i}{\tilde{x}_j}\right)}{\prod_{i,j=1}^N\prod_{a=1,2}(\tilde{x}_j-y_ax_i)^{1-n_a}\prod_{a=3,4}(x_i-y_a\tilde{x}_j)^{1-n_a}},
\label{eq:logZ}
\end{equation}
where $y_a$ are the corresponding fugacities. The summation is over all solutions $I$ of the ``Bethe Ansatz Equations" (BAE) $e^{iB_i}=e^{i\tilde{B}_i}=1$ modulo permutations, where
\begin{align}
e^{iB_i}&=x_i^k\prod_{j=1}^N\frac{(1-y_3 \frac{\tilde{x}_j}{x_i})(1-y_4 \frac{\tilde{x}_j}{x_i})}{(1-y_1^{-1} \frac{\tilde{x}_j}{x_i})(1-y_2^{-1} \frac{\tilde{x}_j}{x_i})},\nonumber\\
e^{i\tilde{B}_j}&=\tilde{x}_j^k\prod_{i=1}^N\frac{(1-y_3 \frac{\tilde{x}_j}{x_i})(1-y_4 \frac{\tilde{x}_j}{x_i})}{(1-y_1^{-1} \frac{\tilde{x}_j}{x_i})(1-y_2^{-1} \frac{\tilde{x}_j}{x_i})}.
\end{align}
Here $k$ is the Chern-Simons level, and the two sets of variables $\{x_i\}$ and $\{\tilde x_j\}$ arise from the $U(N)_k\times U(N)_{-k}$ structure of ABJM theory. Finally, the $2N\times 2N$ matrix $\mathbb{B}$ is the Jacobian relating the $\{x_i,\tilde x_j\}$ variables to the $\{e^{iB_i},e^{i\tilde B_j}\}$ variables
\begin{equation}
\mathbb B=\begin{pmatrix}x_l\fft{\partial e^{iB_j}}{\partial x_l}&\tilde x_l\fft{\partial e^{iB_j}}{\partial\tilde x_l}\\[4pt]
x_l\fft{\partial e^{i\tilde B_j}}{\partial x_l}&\tilde x_l\fft{\partial e^{i\tilde B_j}}{\partial\tilde x_l}\end{pmatrix}.
\end{equation}
See \cite{Benini:2015eyy} for additional details.
It is convenient to introduce the chemical potentials $\Delta_a$ according to $y_a=e^{i\Delta_a}$ and furthermore perform a change of variables $x_i=e^{iu_i}$, $\tilde{x}_j=e^{i\tilde{u}_j}$. In this case, the BAE become
\begin{align}
0&=ku_i -i\sum_{j=1}^N\left[\sum_{a=3,4}\log\left(1-e^{i(\tilde{u}_j-u_i+\Delta_a)}\right)-
\sum_{a=1,2}\log\left(1-e^{i(\tilde{u}_j-u_i-\Delta_a)}\right)\right]-2\pi n_i, \nonumber \\
0&=k\tilde{u}_j -i\sum_{i=1}^N\left[\sum\limits_{a=3,4}\log\left(1-e^{i(\tilde{u}_j-u_i+\Delta_a)}\right)-
\sum_{a=1,2}\log\left(1-e^{i(\tilde{u}_j-u_i-\Delta_a)}\right)\right]-2\pi \tilde{n}_j.
\label{eq:BAE}
\end{align}
The topologically twisted index is evaluated by first solving these equations for $\{u_i,\tilde u_j\}$, and then inserting the resulting solution into the partition function (\ref{eq:logZ}). This procedure was carried out in \cite{Benini:2015eyy} in the large-$N$ limit with $k=1$ by introducing the parametrization
\begin{equation}
u_i=iN^{1/2}\, t_i +\pi-\ft12\delta v(t_i), \qquad \tilde{u}_i=iN^{1/2}\,t_i+\pi+\ft12\delta v(t_i),
\end{equation}
where we have further made use of reflection symmetry about $\pi$ along the real axis. In the large-$N$ limit, the eigenvalue distribution becomes continuous, and the set $\{t_i\}$ may be described by an eigenvalue density $\rho(t)$.
\subsection{Evaluation of the index beyond the leading order in $N$}
The leading order solution for $\rho(t)$ and $\delta v(t)$ was worked out in \cite{Benini:2015eyy}, and the resulting partition function exhibits the expected $N^{3/2}$ scaling of ABJM theory
\begin{equation}
{\rm Re}\log Z_0=-\frac{N^{3/2}}{3}\sqrt{2\Delta_1\Delta_2\Delta_3\Delta_4}\sum\limits_a\frac{n_a}{\Delta_a}.
\label{eq:logZ0}
\end{equation}
A similar result was extended to the context of asymptotically AdS$_4$ black holes with hyperbolic horizon in \cite{Cabo-Bizet:2017jsl}. We are, of course, interested in taking this solution beyond the leading order. In the ABJM context, we expect the subleading behavior of the index to have the form
\begin{equation}
{\rm Re}\log Z={\rm Re}\log Z_0+N^{1/2}f_1(\Delta_a,n_a)+\log Nf_2(\Delta_a,n_a)+f_3(\Delta_a,n_a)+\mathcal O(N^{-1/2}),
\label{eq:ABJMrlz}
\end{equation}
where the functions $f_1$, $f_2$ and $f_3$ are linear in the magnetic fluxes $n_a$. In principle, we would like to systematically extend the analysis beyond the leading order in order to obtain the analytic form of these functions. However, this appears to be a challenge, mainly due to the presence of the (left and right) tails of the eigenvalue distribution. (These tails correspond to the nearly vertical segments in figure~\ref{fig:evals}.) We thus proceed with a numerical investigation.
The main setup is to arrive at a numerical solution to the BAE (\ref{eq:BAE}) through multidimensional root finding using the leading order distribution as the starting point. We have implemented this in Mathematica using FindRoot. The solution is first obtained either with MachinePrecision or with WorkingPrecision set to 30, and further refined using WorkingPrecision set to 200 and default settings for AccuracyGoal and PrecisionGoal. Convergence to a stable solution can be a bit delicate, since the BAE is highly sensitive to the tails; if even a single eigenvalue is sufficiently displaced, then it is easy for FindRoot to fail. In most cases, we have been able to obtain numerical solutions up to $N\approx200$, although larger values of $N$ are possible with some refinement of the initial distribution. As an example, the numerical solution for the $u_i$ and $\tilde u_i$ eigenvalues for$\Delta_a=\{0.4, 0.5, 0.7,2\pi-1.6\}$ and $N=60$ is shown in figure~\ref{fig:evals}. The corresponding eigenvalue density $\rho(t)$ and function $\delta v(t)$ are shown in figure~\ref{fig:rhodv}.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{evals}
\caption{The solution to the BAE for $\Delta_a=\{0.4,0.5,0.7,2\pi-1.6\}$ and $N=60$. The solid lines correspond to the leading order expression obtained in \cite{Benini:2015eyy}.}
\label{fig:evals}
\end{figure}
\begin{figure}[t]
\centering
\hbox to \linewidth{\includegraphics[width=0.48\linewidth]{rho}\hss\includegraphics[width=0.48\linewidth]{deltav}}
\caption{The eigenvalue density $\rho(t)$ and the function $\delta v(t)$ for $\Delta_a=\{0.4,0.5,0.7,2\pi-1.6\}$ and $N=60$, compared with the leading order expression.}
\label{fig:rhodv}
\end{figure}
Once the eigenvalues are obtained, it is then simply a matter of numerically evaluating the index (\ref{eq:logZ}) on the solution to the BAE. The main challenge here is the evaluation of $\det\mathbb B$, as the Jacobian matrix is ill-conditioned. (This is why we work to high numerical precision when solving the BAE.) For a given set of chemical potentials $\Delta_a$, we compute $\log Z$ for a range of $N$. We then subtract out the leading behavior (\ref{eq:logZ0}) and decompose the residuals into a sum of four independent terms
\begin{equation}
{\rm Re}\log Z={\rm Re}\log Z_0+A+B_1n_1+B_2n_2+B_3n_3,
\end{equation}
where we have used the condition $\sum_a n_a=2$. At this stage, we then perform a linear least-squares fit of $A$ and $B_a$ to the function
\begin{equation}
f(N)=f_1N^{1/2}+f_2\log N+f_3+f_4N^{-1/2}+f_5N^{-1}+f_6N^{-3/2}.
\end{equation}
We are, of course, mainly interested in $f_2$. However, since $N$ ranges from about 50 to 200, it is important to consider the first few inverse powers of $N$ as well. (We have confirmed numerically that the first subdominant term enters at $\mathcal O(N^{1/2})$, and that in particular terms of $\mathcal O(N)$ are absent.)
\begin{table}[t]
\begin{tabular}{l|l|l||l|l|l}
$\Delta_1$&$\Delta_2$&$\Delta_3$&$f_1$&$f_2$&$f_3$\\
\hline
$\pi/2$&$\pi/2$&$\pi/2$&$3.0545$&$-0.4999$&$-3.0466$\\
\hline
$\pi/4$&$\pi/2$&$\pi/4$&$4.2215-0.0491n_1$&$-0.4996+0.0000n_1$&$-4.1710-0.2943n_1$\\
&&&$-0.1473n_2-0.0491n_3$&$+0.0000n_2+0.0000n_3$&$+0.0645n_2-0.2943n_3$\\
\hline
$0.3$&$0.4$&$0.5$&$7.9855-0.2597n_1$&$-0.4994-0.0061n_1$&$-9.8404-0.9312n_1$\\
&&&$-0.5833n_2-0.6411n_3$&$-0.0020n_2-0.0007n_3$&$-0.0293n_2+0.3739n_3$\\
\hline
$0.4$&$0.5$&$0.7$&$6.6696-0.1904n_1$&$-0.4986-0.0016n_1$&$-7.5313-0.6893n_1$\\
&&&$-0.4166n_2-0.4915n_3$&$-0.0008n_2-0.0001n_3$&$-0.1581n_2+0.2767n_3$
\end{tabular}
\caption{Numerical fit for ${\rm Re}\log Z={\rm Re}\log Z_0+f_1N^{1/2}+f_2\log N+f_3+\cdots$. The values of $N$ used in the fit range from $50$ to $N_{\rm max}$ where $N_{\rm max}=290, 150, 190, 120$ for the four cases, respectively. We made use of the fact that the index is independent of the magnetic fluxes when performing the fit for the special case ($\Delta_a=\{\pi/2,\pi/2,\pi/2,\pi/2\}$).}
\label{tbl:dat}
\end{table}
The results of the numerical fit are presented in Table~\ref{tbl:dat}. Our main result is that the numerical evidence points to the coefficient of the $\log N$ term being exactly $-1/2$. We thus have
\begin{equation}
{\rm Re}\log Z=-\frac{N^{3/2}}{3}\sqrt{2\Delta_1\Delta_2\Delta_3\Delta_4}\sum\limits_a\frac{n_a}{\Delta_a}
+N^{1/2}f_1(\Delta_a,n_a)-\fft12\log N+f_3(\Delta_a,n_a)+\mathcal O(N^{-1/2}),
\label{eq:tti}
\end{equation}
where $f_1$ and $f_3$ remain to be determined. One may wonder whether their dependence on the magnetic fluxes, $n_a$, follows the same leading order behavior, namely $\sum_a n_a/\Delta_a$. Unfortunately, examination of the table shows that this is not the case.
Although we have been unable to discern the general behavior of $f_1$, for the special case we find the approximate expression
\begin{equation}
f_1=3.0545\approx\fft{11\pi}{8\sqrt{2}}=\fft\pi{\sqrt2}\left(\fft1{24}+\fft13+1\right).
\end{equation}
We have in fact extended the special case to $k>1$. For $k\sim\mathcal O(1)$, the eigenvalue distribution retains the same features, but with appropriate scaling by $k$. Working specifically up to $k=5$ and with $N$ up to 200, we find good evidence that in this case the partition function takes the form
\begin{equation}
{\rm Re}\log Z(\Delta_a=\pi/2)=-\fft{\pi\sqrt{2k}}3N^{3/2}+\fft\pi{\sqrt{2k}}\left(\fft{k^2}{24}+\fft13+1\right)N^{1/2}-\fft12\log N+\cdots,
\end{equation}
which may be compared with the ABJM free energy on $S^3$
\begin{equation}
F_{\rm ABJM}=-\fft{\pi\sqrt{2k}}3N^{3/2}+\fft\pi{\sqrt{2k}}\left(\fft{k^2}{24}+\fft13\right)N^{1/2}-\fft14\log N+\cdots.
\end{equation}
While the leading $\mathcal O(N^{3/2})$ term is identical, the first subleading term in the topologically twisted index picks up an additional contribution. In addition, the coefficients of the log terms differ, and this suggests that the two expressions are capturing distinct features of the holographic dual. Some similarities between the free energy and the topologically twisted index were first pointed out in \cite{Hosseini:2016tor}. More generally, relations between partitions functions on $S^3$ and $S^2\times S^1$ with a topological twist have recently been discussed in \cite{Closset:2017zgf}. It would be interesting to place our concrete, subleading in $N$, results within that more formal approach.
\subsection{Perilous $1/N$ expansion}
While the numerical evidence for $-1/2\log N$ appears compelling, ideally this ought to be backed up by an analytical expansion in the large-$N$ limit. Such an expansion would naturally shed light on the $f_1$ coefficient as well. However, as mentioned above, the tails make it difficult to maintain a systematic treatment of the $1/N$ expansion. In particular, the tails occur when the eigenvalues $\{u_i,\tilde u_j\}$ satisfy
\begin{equation}
\tilde u_i-u_i\approx\pm\Delta_a\qquad\Rightarrow\qquad\delta v(t_i)\approx\pm\Delta_a.
\end{equation}
In this case, the logs in the BAE, (\ref{eq:BAE}), for $j$ near $i$ are evaluated near zero. The resulting large logs cause apparently subleading terms to become important, and hence mixes up orders in the superficial $1/N$ expansion, as already noted in \cite{Benini:2015eyy}.
The leading order partition function may be obtained by properly accounting for the large logs, and we suspect that a careful treatment would allow the computation to be extended to higher orders. However, this remains a technical challenge, as can be seen from the following illustration. In the large-$N$ limit, it is natural to focus on the eigenvalue density $\rho(t)$ and the function $\delta v(t)$. In the formal large-$N$ expansion, both functions are considered to be $\mathcal O(1)$, which is consistent with the plots in figure~\ref{fig:rhodv}. However, their leading-order slopes are discontinuous where the left and right tails meet the inner interval. This gives rise to a $\delta$-function divergence when working with their second derivatives. While the divergence is unimportant at leading order, it presents difficulties at higher order.
Of course, as can be seen in figure~\ref{fig:rhodv}, the actual solution does not have discontinuous slope. As an estimate, we first note that the range where $\rho(t)$ changes slope is of $\mathcal O(1/\sqrt N)$. As a result, $\rho''(t)\sim\mathcal O(\sqrt N)$ near the transition points, and a similar estimate can be made for $\delta v(t)$. While this avoids the $\delta$-function divergences, it nevertheless mixes up orders in the formal large-$N$ expansion. Furthermore, it is not just the second derivative, but all higher derivatives as well that become important, even when considering just the first subleading correction to the index.
\section{One-loop quantum supergravity}\label{Sec:Gravity}
Based on our numerical evidence, we conjecture that the topologically twisted index has a universal logarithmic correction given by $-1/2\log N$, in contrast with the ABJM free energy that has the factor $-1/4\log N$. In the latter case, the field theory result was reproduced by a one-loop supergravity computation in \cite{Bhattacharyya:2012ye}. In particular, the standard AdS$_4$/CFT$_3$ correspondence relates ABJM theory on $S^3$ to M-theory on global AdS$_4\times S^7/\mathbb Z_k$ \cite{Aharony:2008ug}. The logarithmic term then originates purely from a ghost two-form zero mode contribution on AdS$_4$.
In the present case, however, we take ABJM theory on $S^2\times S^1$ with a topological twist generated by background magnetic flux. This topological twist relevantly deforms the ABJM theory to flow toward a superconformal quantum mechanics on $S^1$. Holographically, such an RG flow can be thought of as a Euclidean asymptotically AdS$_4$ BPS magnetic black hole, interpolating between the asymptotically AdS$_4$ region and an AdS$_2\times$ S$^2$ near horizon region. The solution can be embedded into 11-dimensional supergravity \cite{Cvetic:1999xp}, and such an embedding makes it also natural to consider the quantum correction from an 11-dimensional point of view.
We are thus interested in computing the one-loop correction to the supersymmetric partition function in the BPS black hole background that interpolates between asymptotic AdS$_4\times S^7$ and AdS$_2\times M_9$ near the horizon, where $M_9$ is a $S^7$ bundle over $S^2$. As a simplification, however, we assume a decoupling limit exists, so that we can focus mainly on the AdS$_2\times M_9$ near horizon geometry. Alternatively, corrections to the black hole entropy may be considered via the quantum entropy function in the near horizon geometry proposed in \cite{Sen:2008vm}. For extremal black hole with no electric charge, the quantum entropy function reduces to the partition function of 11-dimensional supergravity compactified in the near horizon geometry, and we are again led to AdS$_2\times M_9$.
In the computation of one-loop corrections to the partition function, we focus on the logarithmic term, as such a term, in odd dimensional spaces, arises purely from zero modes (see \cite{Vassilevich:2003xt} and \cite{Bhattacharyya:2012ye} for a review). The effect of zero modes on the logarithmic term can be naturally divided into two parts: the subtraction of zero modes from the trace of the heat kernel to make the heat kernel well defined, and the integration over zero modes in the path integral. Those two parts can be summarized schematically, for a given kinetic operator $D$ of a physical field, as
\begin{equation}
\Delta F_D= (-1)^D(\beta_D-1)n_D^0\log L,
\end{equation}
where $\beta_D$ encodes the integration over zero modes in the path integral, and $-1$ is due to the subtraction in the heat kernel. We use $(-1)^D$ to distinguish bosonic/fermionic contributions. The treatment for ghosts is slightly different, and they are considered separately as in \cite{Bhattacharyya:2012ye}. In summary, the total logarithmic correction is given by
\begin{equation}
\Delta F= \sum_{\{D\}}(-1)^D(\beta_D-1)n_D^0\log L+ \Delta F_{\mathrm{Ghost}},
\end{equation}
where the summation is over physical fields.
For completeness, we shall first summarize the fields that have non-trivial zero modes $n_D^0$ in AdS$_2$ and their $\beta_D$, although they are quite standard and well known in the literature (see for example, the appendix of \cite{Sen:2012cj}). We then compute the logarithmic correction from the physical sector and the ghost sector of 11-dimensional supergravity in the near horizon geometry AdS$_2\times M_9$.
\subsection{The number and scalings of zero modes}
The spectrum of a kinetic operator on a non-compact space, such as AdS$_2$, typically consists of two parts: a continuous part due to the non-compactness of the space, and possibly a discrete part that contains a countably infinite number of eigenfunctions with zero eigenvalue. The continuous part of the trace of the heat kernel in the case of AdS$_N$ is well defined, whereas the zero modes from the discrete part, if any, should be subtracted from the heat kernel. The formal sum that counts the number of zero modes in the compact case is divergent when the space is non-compact,
\begin{equation}\label{n0}
n_0=\sum_j \int \sqrt{g}d^2x |\phi_j(x)|^2,
\end{equation}
where $\phi_j(x)$'s are normalized to 1. Thus computing $n_0$ requires regularization. For symmetric spaces $G/H$, $n_0$ can be evaluated by working out explicit eigenfunctions, exchanging the sum and integral, and using a regularized volume as in \cite{Sen:2012cj,Bhattacharyya:2012ye}.
Here, we present another way of computing $n_0$ using the general theorem in \cite{CAMPORESI199457}. The number of zero modes can be associated with the formal degree of the discrete series representation of $G$ corresponding to the given field, which occurs when $G$ has a maximal torus that is compact. For AdS$_N=SO(N,1)/SO(N)$, they occur when $N$ is even, and they can be labeled in terms of the highest weight label $(\sigma, n_0)$, where $\sigma=(n_\frac{N-2}{2}, n_\frac{N-4}{2}, \dots, n_1)$, with $n_\frac{N-2}{2}> n_\frac{N-4}{2}>\dots > n_1>|n_0|$. Any vector bundle over AdS$_N$ can be labeled by an irreducible representation of $SO(N)$ (or $\text{Spin}(N)$) in terms of highest weight labels $\tau=(h_\frac{N}{2}, h_\frac{N-2}{2}, \dots, h_1)$, and in order to determine the number of zero modes for a given field, one looks for the branching condition
\begin{equation}\label{branching}
\frac{1}{2}<|n_0|\leq|h_1|\leq n_1 \leq \dots \leq n_{\frac{N-2}{2}} \leq h_{\frac{N}{2}}.
\end{equation}
The number of zero modes is the sum of all degrees $P(\sigma, n_0)$ of discrete series representations $(\sigma, n_0)$ that satisfies the branching condition, up to a normalization factor that only depends on the dimension:
\begin{equation}\label{zeromodes}
n_0^{\tau}=\frac{\text{Vol}(\mathrm{AdS}_N)}{c_N}\sum_{(\sigma,n_0)} P(\sigma, n_0).
\end{equation}
For AdS$_2$, $P(n_0)=n_0-\frac{1}{2}$ and $c_N=2\pi$, and a field is labeled by a single highest weight label which is its spin. (General expressions for $c_{N}$ and $P(\sigma,n_0)$ can be found in section~6 of \cite{CAMPORESI199457}.) The branching condition, (\ref{branching}), implies that fields with spin greater than $\frac{1}{2}$ have zero modes, i.e. one-form, gravitino, and graviton fields. Moreover, using (\ref{zeromodes}), one has
\begin{align}
n^0_{g}&=2\times \frac{(-2\pi)}{2\pi}\left(2-\frac{1}{2}\right)=-3,\nonumber\\
n^0_{\psi}&=2\times \frac{(-2\pi)}{2\pi}\left(\frac{3}{2}-\frac{1}{2}\right)=-2,\nonumber\\
n^0_A&=2\times \frac{(-2\pi)}{2\pi}\left(1-\frac{1}{2}\right)=-1,
\end{align}
where $n^0_{g}$, $n^0_{\psi}$, $n^0_{A}$ are respectively the number of zero modes of a graviton, a gravitino and a one form. We also used the fact that the regularized volume of AdS$_2$ is $-2\pi$. These values, of course, coincide with the direct evaluation performed in \cite{Sen:2012cj}.
The logarithmic part of the integration over zero modes in the path integral can be obtained by dimensional analysis. Given a kinetic operator $\mathcal{O}$, the path integral over zero modes is given by
\begin{equation}\label{intzeromodes}
\int Df|_{\text{zero modes}} \exp\left(-\int d^dx\sqrt{g} f \mathcal{O} f\right)=\int Df|_{\text{zero modes}}\sim L^{\beta_{\mathcal{O}}n_{\mathcal{O}}^0},
\end{equation}
through which we define $\beta_{\mathcal{O}}$ for an operator $\mathcal{O}$. To obtain the logarithmic correction, it is enough to find the $L$ dependence of (\ref{intzeromodes}), which amounts to finding the $L$ dependence in the path integral measure. In the case of Euclidean AdS$_{2N}$, all such zero modes arise due to a non-normalizable gauge parameter $\lambda$, where $f=G\lambda$ with $G$ representing the infinitesimal gauge transformation. For example, let $g_{\mu\nu}=L^2g^{(0)}_{\mu\nu}$. The path integral measure of a $p$-form in $d$ dimensions is normalized as
\begin{equation}
\int D A_{[p]} \exp\left(-L^{d-2p}\int d^dx \sqrt{g^{(0)}} g^{(0)\mu_1\nu_1 }g^{(0)\mu_2\nu_2 }\dots g^{(0)\mu_p\nu_p }A_{\mu_1\dots \mu_p}A_{\nu_1\dots \nu_p}\right)=1.
\end{equation}
Therefore, the correctly normalized measure is
\begin{equation}
D(L^{\frac{d-2p}{2}}d \lambda_{[p-1]}),
\end{equation}
where $\lambda_{[p-1]}$ is a non-normalizable $(p-1)$-form gauge parameter, and has no $L$ dependence. Such a measure gives $L^{(d-2p)/{2}}$ per zero mode, and therefore contributes as $L^{(d-2p)n_p^0/{2}}$ in the path integral. Thus $\beta_{A_{[p]}}=(d-2p)/{2}$ in $d$ dimensions. One can carry out similar computations for other fields, paying particular attention to the possible $L$ dependence of the gauge parameter, as in \cite{Sen:2012cj} and \cite{Banerjee:2010qc}. One then finds
\begin{equation}
\beta_g= \frac{D}{2}, \qquad \beta_{\psi_{\mu}} = D-1, \qquad \beta_{A_{[p]}}=\frac{d-2p}{2}.
\end{equation}
\subsection{The logarithmic corrections}
The 11-dimensional $\mathcal{N}=1$ gravitational multiplet consists of ($g_{\mu\nu}, \psi_{\mu}, C_{\mu\nu\rho}$). The fluctuation of the metric to the lowest order can be summarized as
\begin{equation}
h_{\mu\nu}(x,y)=
\begin{cases}
h_{\alpha \beta}(x)\phi(y), \\
h_{\alpha i}=\sum_a A_{\alpha}^a(x)K^a_{i}(y),\\
\phi(x)h_{ij}(y),
\end{cases}
\end{equation}
where we use $(x^{\alpha}, y^i)$ to denote AdS$_2$ and $M_9$ coordinates, respectively, and $K^{a i}(y)\partial_i$ is a killing vector of $M_9$. The graviton zero modes therefore contribute in two ways: a graviton in AdS$_2$, and gauge fields corresponding to Killing vectors of $M_9$.
From the near horizon geometry in \cite{Benini:2015eyy} one can read off the metric on $M_9$
\begin{equation}\label{nearhorizon}
ds^2_9=\Delta^{\frac{2}{3}}ds^2_{S^2}+\frac{4}{\Delta^{\frac{1}{3}}}\sum_{i=1}^4 \frac{1}{X_i}\left(d\mu_i^2+\mu_i^2(d\psi_i+\frac{n_i}{2}\cos\theta d\phi)^2\right),
\end{equation}
where we denote the coordinates on $S^2$ by $(\theta, \phi)$, $X_i$'s are constant with $\prod X_i=1$, $\Delta=\sum_{i=1}^4X_i \mu_i^2$, and $\sum_{i=1}^4\mu_i^2=1$. The metric, (\ref{nearhorizon}), suggests the following seven Killing vectors:
\begin{align}
&\Bigl\{\cos\phi\partial_{\theta}-\cot\theta \sin \phi \partial_{\phi}+\sum_{j}\frac{n_j}{2}\frac{\sin \phi}{\sin \theta}\partial_{\psi_j},~
-\sin\phi \partial_{\theta}-\cot\theta\cos\phi\partial_{\phi}+\sum_j \frac{n_j}{2}\frac{\cos\phi}{\sin\theta }\partial_{\psi_j},~
\partial_{\phi}\Bigr\},\nonumber\\
&\Bigl\{\partial_{\psi_i}\Bigr\},
\end{align}
where $i=1,2,3,4$, and the Killing vectors span the algebra of the isometry group $SU(2)\times U(1)^4$. Thus the logarithmic correction due to the 11-dimensional graviton is given by
\begin{equation}\label{graviton}
\Delta F_h= (\beta_h-1)(n_g^0+7n_{A}^0)\log L= \left(\frac{11}{2}-1\right)[(-3) \times 1 + (-1)\times 7]\log L=-45 \log L.
\end{equation}
A gravitino $\psi_{\mu}$ can either be an AdS$_2$ gravitino and a spin-1/2 fermion on $M_9$, or vice versa. Ideally one would find the number of killing spinors of $M_9$. Nevertheless, it is more convenient to reduce to four-dimensions first. In this case, the $\mathcal{N}=2$ gravitational multiplet contains two gravitinos, which further decompose to two gravitinos on AdS$_2$. As the number of gravitinos only concerns the number of supersymmetries that are preserved, it should be the same no matter whether one works directly in 11 dimensions, or through a reduction to four dimensions. Thus, the contribution due to the gravitino is given by
\begin{equation}\label{gravitino}
\Delta F_{\psi}=-\sum (\beta_{\psi}-1)n_{\psi}^0\log L=-(10-1)[(-2)\times 2]\log L=36\log L,
\end{equation}
where the minus sign is assigned as it is Grassmann odd.
The fluctuation of a 11 dimensional 3 form can be summarized as
\begin{equation}\label{3form}
C_3(x,y)=
\begin{cases}
B_{3}(y),\\
A_{1}(x)\wedge B_2(y),\\
A_{2}(x)\wedge B_1(y),
\end{cases}
\end{equation}
where the subscript represents the rank of the form, $A(x)$ represents a form on AdS$_2$ and $B(y)$ a form on $M_9$. Note for $M_9$ the Betti numbers $b_1=0$ and $b_2=1$.
Therefore the contribution from the 3-form, from the first line in (\ref{3form}), is
\begin{equation}\label{threeform}
\Delta F_{C}=(\beta_C-1)n^0_C\log L=\left (\frac{5}{2}-1\right)[ (-1)\times 1]\log L=-\frac{3}{2} \log L.
\end{equation}
We now turn to the treatment for ghosts, which requires special care. We therefore compute them separately, and we only concern ourselves with ghosts that give rise to AdS$_2$ zero modes. Therefore only the ghosts for the graviton, which gives a vector ghost $c_\mu$, and the ghosts for the 3-form are considered. The BRST quantization of supergravity generally provides a kinetic term $c^*_\mu(-g^{\mu\nu}\Box -R^{\mu\nu})c_\nu$ with other off diagonal terms that are lower triangular, which do not change the eigenvalues of the kinetic operator on $c_\nu$. In our case, $R_{\mu\nu}$ is never zero, and therefore the graviton ghosts are not relevant to the logarithmic correction.
The general action for quantizing a $p$-form $A_p$ requires a set of $(p-j+1)$-form ghost fields, with $j=2, 3, \dots, p+1$, and the ghost is Grassmann even if $j$ is odd and Grassmann odd if $j$ is even \cite{Siegel:1980jj,Copeland:1984qk}. Although for the $(p-j+1)$-form, the Laplacian operator $(\Delta_{p-j+1})^{j}$ in the computation of the heat kernel requires an extra $j-1$ removal of the zero modes, the integration over the zero modes is unchanged. That results, as in Eq.~(3.4) of \cite{Bhattacharyya:2012ye}, is
\begin{equation}
\Delta F_{\text{Ghost}}=\sum_j (-1)^j(\beta_{A_{p-j}}-j-1)n_{A_{p-j}}^0\log L.
\end{equation}
Note for our case that $b_1$ of $M_9$ is zero. Therefore the only non-vanishing term is $p=3$, $j=2$, which gives
\begin{equation}\label{ghost}
\Delta F_{\text{Ghost}}=-\frac{3}{2}\log L.
\end{equation}
Finally, adding the contributions (\ref{graviton}), (\ref{gravitino}), (\ref{threeform}) and (\ref{ghost}) leads to the total logarithmic correction
\begin{equation}\label{result}
\Delta F=\left(-45+36-\frac{3}{2}-\frac{3}{2}\right)\log L=-12\log L\sim -2\log N,
\end{equation}
where in the last equality we used the AdS/CFT dictionary $N\sim L^6$, and neglected $L$ independent terms. We note that this result does not match with the logarithmic term of the topologically twisted index, (\ref{eq:tti}), which instead has coefficient $-1/2$.
We finish this section by addressing a very natural question. In our computation we have focused exclusively on the near horizon geometry. Given that the black holes we are discussing are asymptotically AdS$_4$, are there contributions that come precisely from the asymptotic region? After all, the computation of \cite{Bhattacharyya:2012ye} obtained logarithmic corrections on the gravity side by studying quantum supergravity on AdS$_4\times S^7$ and found that the entire contribution comes from a two-form zero mode in AdS$_4$. The result of \cite{Bhattacharyya:2012ye} perfectly matches field theory results corresponding to the free energy of ABJM on $S^3$. Our case, however, pertains to a computation of ABJM on $S^2\times S^1$. In an elucidating discussion about boundary modes presented in \cite{Larsen:2015aia}, the authors considered global aspects of AdS$_4$ with $S^3$ and $S^2\times S^1$ boundary conditions. In particular, they established that the Euler number depends on these boundary conditions and is, respectively, $\chi=1$ and $\chi=0$. This result indicates the existence of a two-form zero mode in the case of $S^3$ boundary conditions which is precisely the two-form responsible for the successful match with the field theory free energy. It also indicates the absence of the corresponding two-form zero mode for $S^2\times S^1$ boundary conditions. Moreover, the crucial use of $S^3$ boundary conditions in the explicit construction of the non-trivial two forms \cite{CAMPORESI199457,10.2307/2042193,Donnelly1981}, also supports our claim.
Therefore, at least to this level of scrutiny, there is no contribution coming from the asymptotically AdS$_4$ region. It will, of course, be interesting to develop a systematic approach to dealing with asymptotically AdS contributions in the framework of holographic renormalization.
\section{Discussion}\label{Sec:Conclusions}
Given the disagreement in the computations, we shall discuss some of our underlying assumptions. On the field theory side, the topologically twisted index reproduces the Bekenstein-Hawking entropy of AdS black holes at leading order in the large-$N$ expansion \cite{Benini:2015eyy,Cabo-Bizet:2017jsl}. It is thus tempting to expect that the index provides an complete microstate description at all orders. To explore this possibility, we have performed a numerical investigation of the topologically twisted index and obtained a logarithmic correction of $-1/2\log N$. It is this term that we have attempted to reproduce by computing a one-loop partition function on the supergravity side of the duality.
While AdS/CFT suggests that the corresponding one-loop partition function ought to be computed in the full magnetic AdS$_4$ black hole background, we made a decoupling approximation and focused instead on the AdS$_2\times S^2$ near horizon region. Given the 11-dimensional supergravity origin, only zero modes contribute to the logarithmic term, and we find instead the term $-2\log N$ from the bulk computation. Of course, this treatment of separate near horizon and asymptotic regions is somewhat {\it ad hoc}, and it is entirely possible that agreement would be restored if we instead worked in the full geometry. Although one may expect the degrees of freedom responsible for the entropy to be close to the horizon, it would be important to perform a more careful investigation to see whether or not this is truly the case.
Of course, this all assumes that the asymptotically AdS$_4$ black hole partition function is the correct object to match with the topologically twisted index. An alternative framework is to compare with the quantum entropy function \cite{Sen:2008yk,Sen:2008vm}. However, this approach is technically similar and gives the same result of $-2\log N$. In this case, the disagreement suggests that the quantum entropy formalism needs to be modified for the case of asymptotically AdS black holes.
Another point that deserves discussion is the role of extremization of the index in order to match with the entropy on the gravity side \cite{Benini:2015eyy}. Our field theory computation was performed for arbitrary magnetic charges and potentials. Extremization, at leading order, relates the potentials, $\Delta_a$, to the magnetic charges $n_a$. We believe that the $\log N$ coefficient does not change at the level that we discuss because the leading order relation from extremization is independent of $N$.
More generally, the field theory index is clearly grand canonical while the gravity side is microcanonical. One can think of the extremization procedure of \cite{Benini:2015eyy} as performing a Legendre transform where the main contribution comes from electrically neutral configurations. Although it is plausible that under certain conditions a partition function equals the entropy, the equivalence of the topologically twisted index with the entropy of magnetically charged black holes deserves a more systematic discussion within the framework of holographic renormalization.
Let us now discuss a number of other directions that would be nice to explore. One natural question is motivated by the universality of the result of \cite{Bhattacharyya:2012ye}. Indeed, a large class of field theory partition functions on $S^3$ has a $-1/4\log N$ correction for matter Chern-Simons theories of various types \cite{Fuji:2011km,Marino:2011eh}. On the gravity side of the correspondence, the universality of this result relies on the logarithmic term being given strictly by a two-form zero mode in AdS$_4$; it is thus independent of the Sasaki-Einstein $X_7$ manifold where the supergravity is defined \cite{Bhattacharyya:2012ye}. Despite the current disagreement, it would be interesting to entertain a similar universality argument for the correction we find here, namely $-1/2\log N$. Although we do not fully understand the gravity side, it is clear that the answer we provide relies on very general properties of $M_9$ such as the number of Killing vectors and the second Betti number. The underlying question behind this direction rests on the various hints \cite{Hosseini:2016tor,Closset:2017zgf} for connections between the partition function on $S^3$ and the topologically twisted index in $S^2\times S^1$ that we have now investigated beyond the large-$N$ limit.
Although we have proposed an {\it ad hoc} treatment for the asymptotically AdS$_4$ region, the more general question of how to systematically apply the principles of holographic renormalization for the computation of logarithmic corrections to black hole entropy remains open. A better understanding, perhaps including generalizations of the quantum entropy function, is clearly needed to tackle the plethora of solutions that are intrinsically four-dimensional, and in particular, more general solutions of ${\cal N}=2$ gauged supergravity.
A more challenging question is: Can one obtain the full logarithmic correction to the entropy, and not just the $\log N$ coefficient? One possibility is to tackle the theory directly in four dimensions. In this case the heat kernel, being in an even dimensional space, contributes in a more complicated way. A similar technical problem appears in the 't~Hooft limit where the gravity dual theory lives on AdS$_4\times \mathbb{CP}^3$. It is worth pointing out an added difficulty in the case of the magnetically charged black holes we are considering. For asymptotically flat black holes, a typical practice is to consider particular $N$-correlated scalings of the charges; this allows for the computation of corrections in various regimes. However, generic scalings of the charges are not allowed in our case because the charges are constrained, for example, by $\sum n_i=2$. Alternatively, one could attempt a full supergravity localization following the work \cite{Dabholkar:2014wpa} and the more recent effort in \cite{Nian:2017hac}.
Of course, it is worth noting that the first subleading correction to the topologically twisted index occurs at $\mathcal O(N^{1/2})$. In principle, it would be useful to obtain an analytic expression for this correction, which we denoted $f_1(\Delta_a,n_a)$ in (\ref{eq:ABJMrlz}). On the gravity side, this term presumably originates from higher-derivative corrections to the Wald entropy. While we have been as yet unable to find the analytic form of $f_1$, it may be possible to do so with additional numerical work.
Finally, it would be interesting to discuss other asymptotically AdS gravity configurations forming AdS/CFT dual pairs. For example, we may consider black strings in AdS$_5$ that are dual to topologically twisted four-dimensional field theories \cite{Benini:2013cda}. The topologically twisted index for the dual four-dimensional field theories on $S^2\times T^2$ has been constructed in \cite{Benini:2016hjo,Honda:2015yha} and its high temperature limit has recently been discussed in \cite{Hosseini:2016cyf}. It seems also possible, with the new insight provided recently in \cite{Hosseini:2017mds}, to return to the question of microstate counting for the asymptotically AdS$_5$ black holes more generally.
\acknowledgments
We are thankful to F. Benini, A. Cabo-Bizet, A. Charles, C. Closset, A. Gnecchi, A. Grassi, S. M. Hosseini, K. Hristov, U. Kol, F. Larsen, M. Mari\~no, I. Papadimitriou, G. Silva and S. Vandoren for various discussions on closely related topics. VR and LPZ thank the ICTP and CERN, respectively, for warm hospitality. LPZ is particularly thankful to Ashoke Sen for various generous discussions on the subject of logarithmic corrections to black hole entropy. While we were completing this work, we became aware that R.\ Gupta, I.\ Jeon and S.\ Lal have independently obtained the logarithmic correction from the QEF. This work is partially supported by the US Department of Energy under Grant No.\ DE-SC0007859 and No.\ DE-SC0017808.
\bibliographystyle{JHEP}
|
\section{Introduction}
At the beginning of the millenium, Lander {\em et al.} \cite{Lander} published a draft sequence of the human genome
and reported that more than 50\% of the human genome consists of repeated substrings \cite{Lander}.
There are two types of common repeats: {\em interspersed} repeats and {\em tandem} repeats.
Interspersed repeats are caused by transposons
when a segment of DNA is copied and pasted into new positions of the genome.
In contrast, tandem repeats are caused by slipped-strand mispairings \cite{Mundy}, and
they occur
when a pattern of one or more nucleotides is repeated and the repetitions are
adjacent to each other. For example, consider the word {\tt AGTAGTCTGC}. The substring {\tt AGTAGT} is a tandem repeat, and we say that
{\tt AGTAGTCTGC} is generated from {\tt AGTCTGC} by a {\em tandem duplication} of length three.
Tandem repeats are believed to be the cause of several genetic disorders \cite{Usdin,Sutherland,Fondon}
and this motivated the study of tandem duplication mechanisms in a variety of contexts.
\begin{itemize}
\item \textbf{Formal languages}:
Leopold {\em et al.} \cite{Leupold1}, \cite{Leupold2}
defined the {\em unbounded duplication language} and the {\em $k$-bounded duplication language} to be the
set of words generated by seed word under
tandem duplications of unbounded length and tandem duplications of length up to $k$, respectively.
They investigated certain decidability problems involving unbounded duplication languages and
showed that all $k$-bounded duplication languages are context free.
Furthermore, they showed that $k$-bounded duplication language is always regular for any binary seed and $k \ge 1$.
On the other hand, the $k$-bounded duplication language is not regular
for any \textit{square-free} seed word over an alphabet of at least three letters and $k \ge 4$.
More recently, Jain {\em et al.} \cite{Jain2} completed this characterization and
proved that $k$-bounded duplication languages are regular for $k \le 3$.
\item \textbf{Information theory}: Farnoud {\em et al.} \cite{Farnoud} introduced the concept of {\em capacity} to
determine average information content of a $k$-bounded duplication language.
Later, Jain {\em et al.} \cite{Jain2} introduced the notion of {\em expressiveness} to
measure a language's capability to generate words that contain certain desired substrings.
A complete characterization of fully expressive $k$-bounded duplication languages was
provided by Jain {\em et al.} for all alphabet sizes and all $k$.
\item \textbf{Codes correcting tandem duplications}:
Motivated by applications that store data in living organisms \cite{Arita,Heider,Liss},
Jain {\em et al.} \cite{Jain} proposed the study of codes that correct tandem duplications to improve the reliability of data storage.
They investigated various types of tandem duplications and provided optimal code construction in the case
where duplication length is at most two.
\end{itemize}
In this paper, we study the last problem and investigate algorithms associated with these codes.
In particular, given two words ${\bf{x}}$ and ${\bf{y}}$, we look for efficient algorithms that answer the following question:
When are the words ${\bf{x}}$ and ${\bf{y}}$ \textit{confusable} under tandem repeats?
In other words, are there two sequences of tandem duplications such that the resulting words ${\bf{x'}}$ and ${\bf{y'}}$ are equal?
Interestingly, as we demonstrate, even for small duplication lengths, the solutions to this question are nontrivial.
This is surprising as efficient algorithms are well known for analogous questions in other problems in
string matching \cite{Navarro:2001}.
Before we give an account of these results, we introduce some necessary notations and provide a formal statement of our problem.
\section{Preliminaries}
Let $\Sigma_q=\{0,1,\cdots,q-1\}$ be an alphabet of $q\ge 2$ symbols.
For a positive integer $n$, let $\Sigma_q^n$ denote the set of all strings or words of length $n$ over $\Sigma_q$,
and let $\Sigma_q^*$ denote the set of all finite words over $\Sigma_q$, or
the {\em Kleene closure} of $\Sigma_q$.
Given two words ${{\boldsymbol{x}}}, {{\boldsymbol y}} \in \Sigma_q^*$, we denote their concatenation by ${{\boldsymbol{x}}{\boldsymbol y}}$.
We state the {\em tandem duplication} rules.
For {nonnegative} integers $k\le n$ and $i\le n-k$, we define $T_{i,k} : \Sigma_q^{n} \to \Sigma_q^{n+k}$ such that
\[
T_{i,k}({{\boldsymbol{x}}}) = {\boldsymbol u}{\boldsymbol v}\bv{\boldsymbol{w}}, \mbox{~where~} {\boldsymbol{x}}={\boldsymbol u}{\boldsymbol v}{\boldsymbol{w}},\, |{\boldsymbol u}|=i,\, |{\boldsymbol v}|=k.
\]
If a finite sequence of tandem duplications of length $k$ is performed to obtain ${{\boldsymbol y}}$ from ${{\boldsymbol{x}}}$,
then we say that ${{\boldsymbol y}}$ is a $k$-\textit{descendant} of ${{\boldsymbol{x}}}$, or
${{\boldsymbol{x}}}$ is a $k$-\textit{ancestor} of ${{\boldsymbol y}}$, and denote this relation by ${{\boldsymbol{x}}} \xRightarrow[k]{*} {{\boldsymbol y}}$.
Formally, ${{\boldsymbol{x}}} \xRightarrow[k]{*} {{\boldsymbol y}}$ means that for some positive $t$, there exist $t$ non-negative integers $i_1,i_2, \ldots i_t$ such that ${{\boldsymbol y}}= T_{i_t,k}\circ T_{i_{t-1},k}\circ \cdots \circ T_{i_1,k}({{\boldsymbol{x}}})$.
Given a sequence ${\boldsymbol{x}}$ and integer $k$, we consider the set of words that may be obtained from ${\boldsymbol{x}}$
via a finite number of tandem duplications of length $k$. We define
the $k$-{\em descendant cone} of ${\boldsymbol{x}}$ to be the set of all $k$-descendants of ${\boldsymbol{x}}$
and denote this cone by $D_k^*({\boldsymbol{x}})$. In other words,
\[D_k^*({{\boldsymbol{x}}})\triangleq \left\{{{\boldsymbol y}} \in \Sigma_q^* \mid {{\boldsymbol{x}}} \xRightarrow[k]{*} {\boldsymbol y} \right\}.\]
Our work studies tandem duplications whose lengths are upper bounded by an integer $k$
and we extend the previous definitions.
Formally, if a finite sequence of tandem duplications of length up to $k$ is performed to obtain ${{\boldsymbol y}}$ from ${{\boldsymbol{x}}}$,
then we say that ${{\boldsymbol y}}$ is a ${\le}k$-{\em descendant} of ${{\boldsymbol{x}}}$ and
denote this relation by ${{\boldsymbol{x}}} \xRightarrow[\le k]{*} y$.
We have analogous definitions of ${\le}k$-{\em descendant}, ${\le}k$-{\em ancestor},
and ${\le}k$-{\em descendant cone} $D_{\le k}^*({\boldsymbol{x}})$.
\begin{example}
Consider ${{\boldsymbol{x}}}=01210$ over $\Sigma_3$.
Since $T_{1,3}({{\boldsymbol{x}}}) = 01211210$ and $T_{0,2}(01211210)=0101211210$,
we write $01210 \xRightarrow[3]{*} 01211210$ and
$01210 \xRightarrow[\le 3]{*} 0101211210$.
Alternatively, we have that
$01211210 \in D_3^*({{\boldsymbol{x}}})$, and
$0101211210 \in D_{\le 3}^*({{\boldsymbol{x}}})$.
\end{example}
\subsection{Problem Formulation}
Motivated by applications that store data on living organisms, Jain {\em et al.} \cite{Jain}
looked at the ${\le}k$-descendant cones of a pair of words and
asked whether the two cones have a nontrivial intersection.
Specifically, we introduce the notion of confusability.
\begin{definition}[Confusability] Two words ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$, are said to be $k$-\textit{confusable} if
$D_{k}^*({{\boldsymbol{x}}}) \cap D_{k}^*({{\boldsymbol y}}) \neq \varnothing.$
Similarly, they are said to be ${\le}k$-\textit{confusable} if
$D_{\le k}^*({{\boldsymbol{x}}}) \cap D_{\le k}^*({{\boldsymbol y}}) \neq \varnothing.$
\end{definition}
To design error-correcting codes that store information in the DNA of living organisms,
Jain {\em et al.} then proposed the use of codewords that are not pairwise confusable.
\begin{definition}[${\le}k$-Tandem-Duplication Codes]
A subset ${\cal C} \subseteq \Sigma_q^n$ is a ${\le}k$-{\em tandem-duplication code} if
for all ${{\boldsymbol{x}}},{{\boldsymbol y}} \in {\cal C}$ and ${{\boldsymbol{x}}} \ne {{\boldsymbol y}}$, we have that ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are not ${\le}k$-confusable.
We say that ${\cal C}$ is an $(n,{\le}k;q)$-tandem-duplication code or $(n,{\le}k;q)$-TD code.
\end{definition}
Therefore, to determine if a set of words is a tandem-duplication code,
we need to verify that all pairs of distinct words are not confusable.
Hence, we state our problem of interest.
\vspace{2mm}
\hspace{40mm}{\sc Confusability Problem}
\hspace{40mm}{\bf Instance}: Two words ${\boldsymbol{x}}$ and ${\boldsymbol y}$ over $\Sigma_q$, and an integer $k$
\hspace{40mm}{\bf Question}: Are ${\boldsymbol{x}}$ and ${\boldsymbol y}$ ${\le}k$-confusable?
\vspace{2mm}
While the confusability problem is a natural question, efficient algorithms are only known for the case where $k\in\{1,2\}$.
We review these results in the next subsection.
\subsection{Previous Work}\label{sec:previous}
\noindent{\bf Confusability}.
We summarize known efficient algorithms that determine whether two words are $k$-confusable for all $k$, and
whether they are ${\le}k$-confusable for $k\in\{1,2\}$.
We then highlight why these methods cannot be extended for the case $k\ge 3$.
Crucial to the algorithms is the concept of irreducible words and roots.
\begin{definition} A word ${{\boldsymbol{x}}}$ is said to be $k$-\textit{irreducible} if ${{\boldsymbol{x}}}$ cannot be deduplicated into shorter words with deduplication of length $k$. In other words, if ${\boldsymbol y}\xRightarrow[k]{*} {\boldsymbol{x}}$, then ${\boldsymbol y}={\boldsymbol{x}}$.
The set of $k$-irreducible $q$-ary words is denoted by ${\rm Irr}_{k}(q)$ and the set of $k$-irreducible words of length $n$ is denoted by ${\rm Irr}_{k}(n,q)$.
The $k$-ancestors of ${{\boldsymbol{x}}} \in \Sigma_q^*$ that are $k$-irreducible words are called the $k$-\textit{roots} of
${{\boldsymbol{x}}}$. The set of $k$-roots of ${\boldsymbol{x}}$ is denoted by $R_{k}({{\boldsymbol{x}}})$.
In a similar fashion, we have the definitions of ${\le}k$-{\em irreducible} words and ${\le}k$-roots of ${\boldsymbol{x}}$, and
the notation ${\rm Irr}_{\le k}(q)$, ${\rm Irr}_{\le k}(n,q)$, and $R_{\le k}({\boldsymbol{x}})$.
\end{definition}
\begin{example}
For alphabet $\Sigma_3$, we have $R_{\le 3}(01012012) = \{012\}$, and $R_{\le 2} (012012) = \{012012\}$.
On the other hand, $R_{\le 4}(012101212) = \{012, 0121012\}$.
\end{example}
As we see, it is possible for a word to have more than one root.
Jain {\em et al.} determined when a word has exactly one root and
provided efficient algorithms to compute this unique root.
\begin{proposition}[Jain {\em et al.}\cite{Jain}]\label{prop:jain1}
For any ${{\boldsymbol{x}}} \in \Sigma_q^*$, we have that
$|R_{k}({{\boldsymbol{x}}})|=1$ for all $k$, and $|R_{\le k}({{\boldsymbol{x}}})|=1$ for all $k\in\{1,2,3\}$.
Furthermore, there exist linear-time algorithms to compute $R_{k}({{\boldsymbol{x}}})$ for all $k$.
\end{proposition}
One may easily derive linear-time algorithms to compute $R_{\le k}({{\boldsymbol{x}}})$ for all for $k\in\{1,2,3\}$.
For completeness, we formally describe these algorithms in Appendix \ref{app:root}.
As it turns out, for certain cases, determining the confusability of two words
is equivalent to computing the roots for the words.
\newpage
\begin{proposition}[Jain {\em et al.} \cite{Jain}]\label{prop:jain2}
For all ${{\boldsymbol{x}}},{{\boldsymbol y}} \in \Sigma_q^*$, we have that
\begin{enumerate}[(i)]
\item ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are $k$-confusable if and only if $R_{k}({{\boldsymbol{x}}})=R_{k}({{\boldsymbol y}})$ for all $k$;
\item ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are ${\le}k$-confusable if and only if $R_{\le k}({{\boldsymbol{x}}})=R_{\le k}({{\boldsymbol y}})$ for $k\in\{1,2\}$.
\item If $R_{\le 3}({{\boldsymbol{x}}})\ne R_{\le 3}({{\boldsymbol y}})$, then ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are not ${\le}3$-confusable.
\end{enumerate}
\end{proposition}
In other words, when $k\in\{1,2\}$, to determine ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are $\le k$-confusable,
it is both necessary and sufficient to compute the ${\le}k$-roots of ${\boldsymbol{x}}$ and ${\boldsymbol y}$.
Therefore, applying Proposition \ref{prop:jain1} and the algorithms in Appendix \ref{app:root},
we are able to determine whether two words are ${\le}k$-confusable in linear time for $k\in\{1,2\}$.
Unfortunately, when $k=3$, it is no longer sufficient to compute the ${\le}3$-roots of ${\boldsymbol{x}}$ and ${\boldsymbol y}$.
Specifically, even though $R_{\le 3}({{\boldsymbol{x}}})=R_{\le 3}({\bf y})$,
it is possible that ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are not ${\le}3$-confusable.
We illustrate this in the next example.
\begin{example}\label{confuse}
Consider ${{\boldsymbol{x}}}=012012$ and ${{\boldsymbol y}}=011112$ over $\Sigma_3$.
The words have the same root as $R_{\le 3}({{\boldsymbol{x}}})=R_{\le 3}({\bf y}) = \{012\} $.
However, ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are not ${\le}3$-confusable because any ${\le}3$-descendant of ${{\boldsymbol{x}}}$ has a 2 to the left of a 0,
whereas any ${\le}3$-descendant of ${{\boldsymbol y}}$ does not.
\end{example}
Therefore, the next smallest open case is where $k=3$.
A polynomial-time algorithm is implied by the results of Leupold {\em et al.} \cite{Leupold1} and Jain {\em et al.} \cite{Jain2}.
\begin{proposition}[Leupold {\em et al.} \cite{Leupold1}, Jain {\em et al.} \cite{Jain2}]\label{prop:regular}
Let $k\in\{2,3\}$. Let ${{\boldsymbol{x}}} \in \Sigma_q^*$.
Then $D_{\le k}^{*}({{\boldsymbol{x}}})$ is a regular language.
Furthermore, if $|{{\boldsymbol{x}}}|=m$, then the deterministic finite automaton that generates $D_{\le k}^{*}({{\boldsymbol{x}}})$ has $O(m)$ vertices and $O(m)$ edges. The number of vertices and edges is independent of the alphabet size.
\end{proposition}
Therefore, for two words ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ with $|{{\boldsymbol{x}}}|=m$ and $|{{\boldsymbol y}}|=n$,
the language $D_{\le 3}^{*}({{\boldsymbol{x}}}) \cap D_{\le 3}^{*}({{\boldsymbol y}})$ is regular and
the corresponding finite automaton has $O(mn)$ vertices and $O(mn)$ edges.
Hence, to determine if ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable,
it suffices to determine if the language $D_{\le 3}^{*}({{\boldsymbol{x}}}) \cap D_{\le 3}^{*}({{\boldsymbol y}})$ is nonempty.
The latter can be done in $O(mn)$ time.
We improve this running time by providing an algorithm
that runs in $O(\max\{m,n\})$ time.
\vspace{2mm}
\noindent{\bf Code Construction}. We are interested in determining the maximum possible size of an $(n,{\le}k;q )$-TD code.
An $(n,{\le}k;q )$-TD code that achieves this maximum is said to be {\em optimal}.
Motivated by Proposition \ref{prop:jain2}, Jain {\em et al.} used irreducible words to construct tandem-duplication codes.
\begin{construction}\label{code:jain}
For $k\in\{2,3\}$ and $n \ge k$. An $(n,{\le}k; q)$-TD code ${\cal C}_I(n, {\le}k;q)$ is given by
\begin{equation*}
{\cal C}_I(n,{\le}k;3)=\bigcup_{i=1}^{n} \left\{\xi_{n-i}({{\boldsymbol{x}}})\mid {{\boldsymbol{x}}} \in {\rm Irr}_{\le k}(i,q)\right\}.
\end{equation*}
Here, $\xi_{i}({{\boldsymbol{x}}})={\boldsymbol{x}} z^i$, where $z$ is the last symbol of ${\boldsymbol{x}}$.
In other words, $\xi_{i}({{\boldsymbol{x}}})$ is the sequence ${{\boldsymbol{x}}}$ with its last symbol $z$ repeated $i$ more times.
Furthermore, the size of ${\cal C}_I(n,{\le}k;q)$ is $\sum_{i=1}^n \left|{\rm Irr}_{\le k}(i,q)\right|$.
\end{construction}
It then follows from Proposition \ref{prop:jain2} that ${\cal C}_I(n,{\le}2;q)$ is an optimal $(n,{\le}2;q)$-TD code.
Unfortunately, the TD code ${\cal C}_I(n,{\le}3;q)$ is not an optimal for $n\ge 6$,
and we illustrate this via the following example.
\begin{example}
Consider $n=6$, $q=3$. The code ${\cal C}_I(6,\le 3;3)$ from Construction~\ref{code:jain} is given by
\begin{align*}
{\cal C}_I(6,{\le}3;q)
& = \{aaa aaa\mid a\in\Sigma_3 \} \cup \{abb bbb\mid a, b\in\Sigma_3, a\ne b \}\\
& \cup \{aba aaa,abc ccc\mid a, b,c \in\Sigma_3, a\ne b, b\ne c, a\ne c \}\\
& \cup \{abac cc, abca aa, abcb bb \mid a, b, c\in\Sigma_3, a\ne b, b\ne c, a\ne c \}\\
& \cup \{aba caa, aba cbb, abc abb, abc acc, abc baa \mid a, b,c\in\Sigma_3, a\ne b, b\ne c, a\ne c \}\\
& \cup \{aba cab, aba cba, aba cbc, abc aba, abc acb, abc bab, abc bac \mid a, b,c\in\Sigma_3, a\ne b, b\ne c, a\ne c \}.
\end{align*}
Hence, $|{\cal C}_I(6,{\le}3;3)|=111$.
However, we may remove from ${\cal C}_I(6,{\le}3;3)$
the six codewords $\{abc ccc \mid a, b,c \in\Sigma_3, a\ne b, b\ne c, a\ne c \}$ and
augment the code with twelve more codewords $\{abcabc, abbbbc \mid a, b,c \in\Sigma_3, a\ne b, b\ne c, a\ne c \}$.
Then we can check that the new code has size $117$, and
we later verify using Proposition \ref{prop:irr2-upper}
that the new code is indeed optimal.
For lengths at least six, we construct TD codes with strictly larger sizes.
\end{example}
\noindent{\bf Upper Bound}. We also study upper bounds on the size of an optimal $(n,{\le}3;q)$-TD code.
By definition, an $(n,{\le}3;q)$-TD code is also an $(n,{\le}2;q)$-TD code.
Since an optimal $(n,{\le}2;q)$-TD code is provided by Construction~\ref{code:jain},
we have the following upper bound on the size of an optimal $(n,{\le}3;q)$-TD code.
\begin{proposition}\label{prop:irr2-upper}
The size of an $(n,{\le}3;q)$-TD code is at most $\sum_{i=1}^n \left|{\rm Irr}_{\le 2}(i,q)\right|$.
\end{proposition}
Proposition \ref{prop:irr2-upper} implies that Construction~\ref{code:jain} is tight for $k=3$ and $n\le 5$.
Using a combinatorial characterization implied by our algorithm,
we improve this upper bound for longer lengths in Section \ref{sec:codes}.
\subsection{Our Contributions}
\begin{itemize}
\item {\bf Confusability}. In Section \ref{sec:confusable},
we present sufficient and necessary conditions for two words to be ${\le}3$-confusable and
propose a linear-time algorithm to solve the ${\le}3$-confusability problem.
\item {\bf Estimates on code sizes}. Using insights gained from Section \ref{sec:confusable},
we study the size of tandem-duplication codes in Section \ref{sec:codes}.
We first improve the upper bound given by Proposition \ref{prop:irr2-upper} and
then construct codes with larger sizes as compared to those given by Construction \ref{code:jain}.
We also provide certain explicit constructions and recursive constructions for tandem-duplication codes.
Furthermore, we determine the sizes of optimal ${\le}3$-tandem-duplication codes
for lengths up to {twenty}.
\end{itemize}
\section{Determining ${\le}3$-Confusability}
\label{sec:confusable}
We derive a linear-time algorithm to determine the ${\le}3$-confusability of two words ${\boldsymbol{x}}$ and ${\boldsymbol y}$.
For the sake of simplicity, we omit the use of ``${\le}3$'' and assume confusable, descendant and root to mean ${\le}3$-confusable, ${\le}3$-descendant and ${\le}3$-root, respectively.
Prior to stating the algorithm, we make some technical observations on tandem duplications of length at most three.
The proofs of the lemmas are deferred to Appendix \ref{app:order}.
The first lemma states that given a sequence of tandem duplications, we may reorder the tandem duplications
such that the lengths of the repeats are nonincreasing.
\begin{lemma}\label{lem:reorder}
Let $k_1<k_2\le 3$ and ${\boldsymbol{x}}\in \Sigma_q^*$.
Suppose that $T_{i_2,k_2}\circ T_{i_1,k_1}({\boldsymbol{x}})={{\boldsymbol{x}}'}$ for some $i_1,i_2$.
Then there exist integers $3\ge \ell_1\ge \ell_2\ge \cdots\ge \ell_t$ and $j_1,j_2,\ldots, j_t$
such that
\[
T_{j_t,\ell_t}\circ T_{j_{t-1},\ell_{t-1}}\circ \cdots \circ T_{j_1,\ell_1}({{\boldsymbol{x}}})={\boldsymbol{x}}'.
\]
\end{lemma}
The next lemma states that we may assume the duplications of length two and three
are performed on segments whose symbols are distinct.
This is because if a duplication is performed on a segment whose symbols are not distinct,
we may find an equivalent sequence of duplications of strictly shorter lengths.
\begin{lemma}\label{lem:distinct}
Let $k\in \{2,3\}$. Suppose $T_{i,k}({\boldsymbol{x}})={\boldsymbol u}{\boldsymbol v}\bv{\boldsymbol{w}}$, where ${\boldsymbol{x}}={\boldsymbol u}{\boldsymbol v}{\boldsymbol{w}}$, $|{\boldsymbol u}|=i$ and $|{\boldsymbol v}|=k$.
If the symbols in ${\boldsymbol v}$ are not pairwise distinct, then
there exist integers $k> \ell_1\ge \ell_2$ and $j_1,j_2$
such that
\[
T_{j_2,\ell_2}\circ T_{j_1,\ell_1}({{\boldsymbol{x}}})=T_{i,k}({\boldsymbol{x}}).
\]
\end{lemma}
Henceforth, we use the notation ${{\boldsymbol{x}}} \xRightarrow[k_d]{*} {{\boldsymbol y}}$
to denote that there is a sequence of tandem duplications of length $k$ of {\em distinct} symbols from ${{\boldsymbol{x}}}$ to ${{\boldsymbol y}}$.
It follows from definition that ${{\boldsymbol{x}}} \xRightarrow[1]{*} {{\boldsymbol y}}$
is equivalent to ${{\boldsymbol{x}}} \xRightarrow[1_d]{*} {{\boldsymbol y}}$.
The next lemma is immediate from Lemmas \ref{lem:reorder} and \ref{lem:distinct}.
\begin{lemma}\label{lem:ordering}
Suppose that ${{\boldsymbol{x}}} \xRightarrow[\le3]{*} {{\boldsymbol{x}}'}$.
Then there exist ${{\boldsymbol{x}}_1}$ and ${{\boldsymbol{x}}_2}$ such that
${{\boldsymbol{x}}} \xRightarrow[3_d]{*} {{\boldsymbol{x}}_2} \xRightarrow[2_d]{*} {{\boldsymbol{x}}_1} \xRightarrow[1_d]{*} {{\boldsymbol{x}}'}.$
Furthermore, ${\boldsymbol{x}}_1$ and ${\boldsymbol{x}}_2$ are uniquely determined by
${\boldsymbol{x}}_1=R_{1}({\boldsymbol{x}}')$ and ${\boldsymbol{x}}_2=R_{\le 2}({\boldsymbol{x}}')$.
\end{lemma}
Equipped with this lemma, we state the following theorem that provides a
necessary and sufficiency condition for two words to be confusable.
\begin{theorem}
\label{confusable}
Two words ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are ${\le}3$-confusable if and only if there exist ${{\boldsymbol{x}}'}$ and ${{\boldsymbol y}'}$ such that
\[{{\boldsymbol{x}}} \xRightarrow[3_d]{*} {{\boldsymbol{x}}'},~{{\boldsymbol y}} \xRightarrow[3_d]{*} {{\boldsymbol y}'}
\mbox{ and }R_{\le 2}({{\boldsymbol{x}}'})=R_{\le 2}({{\boldsymbol y}'}).\]
\end{theorem}
\begin{proof}
Suppose that ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable.
Then there exists ${{\boldsymbol{z}}}$ such that
${{\boldsymbol{x}}} \xRightarrow[\le 3]{*} {{\boldsymbol{z}}}$ and ${{\boldsymbol y}} \xRightarrow[\le 3]{*} {{\boldsymbol{z}}}$.
By Lemma~\ref{lem:ordering}, there exist ${{\boldsymbol{x}}_1}, {{\boldsymbol{x}}_2}, {{\boldsymbol y}_1}, {{\boldsymbol y}_2}$ such that
${{\boldsymbol{x}}} \xRightarrow[3_d]{*} {{\boldsymbol{x}}_2} \xRightarrow[2_d]{*} {{\boldsymbol{x}}_1} \xRightarrow[1_d]{*} {{\boldsymbol{z}}}$ and
${{\boldsymbol y}} \xRightarrow[3_d]{*} {{\boldsymbol y}_2} \xRightarrow[2_d]{*} {{\boldsymbol y}_1} \xRightarrow[1_d]{*} {{\boldsymbol{z}}}$.
Hence we set ${{\boldsymbol{x}}'}={{\boldsymbol{x}}_2}$ and ${{\boldsymbol y}'}={{\boldsymbol y}_2}$. Note that by Proposition~\ref{prop:jain2}, $R_{\le 2}({{\boldsymbol{x}}'})=R_{\le 2}({{\boldsymbol y}'})$ since
${\boldsymbol{x}}'$ and ${\boldsymbol y}'$ are ${\le}2$-confusable.
Conversely, suppose that there exist ${{\boldsymbol{x}}'}$ and ${{\boldsymbol y}'}$ such that
${{\boldsymbol{x}}} \xRightarrow[3_d]{*} {{\boldsymbol{x}}'}$, ${{\boldsymbol y}} \xRightarrow[3_d]{*} {{\boldsymbol y}'}$, and $R_{\le 2}({{\boldsymbol{x}}'})=R_{\le 2}({{\boldsymbol y}'})$.
Since ${{\boldsymbol{x}}'}$ and ${{\boldsymbol y}'}$ have a common ${\le}2$-root, it follows from Proposition~\ref{prop:jain2} that ${{\boldsymbol{x}}'}$ and ${{\boldsymbol y}'}$ are ${\le}2$-confusable. Thus, ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable.
\end{proof}
In conclusion, for two words ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ to be confusable,
it is equivalent to checking if there are tandem duplications of length three that make
their descendants ${{\boldsymbol{x}}'}$ and ${{\boldsymbol y}'}$ have the same ${\le}2$-root.
We make use of this important fact in the next section.
\subsection{Strategy behind Algorithm}
We outline our strategy to determine the confusability of two words ${\boldsymbol{x}}$ and ${\boldsymbol y}$. Recall from Proposition \ref{prop:jain2}(iii), if ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ have different roots,
then they are not confusable.
Hence, our first step is to determine $R_{\le 3}({\boldsymbol{x}})$ and $R_{\le 3}({\boldsymbol y})$ and see if the roots are equal.
If the roots are not equal, we immediately conclude that ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are not confusable.
Therefore, the nontrivial task is to determine confusability when $R_{\le 3}({\boldsymbol{x}})=R_{\le 3}({\boldsymbol y})$.
Henceforth, we assume that the {\em common root of ${\boldsymbol{x}}$ and ${\boldsymbol y}$ is ${\boldsymbol{r}}$}.
Suppose that ${\boldsymbol{r}}$ contains at most two distinct symbols.
Then from Lemma \ref{lem:ordering},
we have that ${\boldsymbol{r}}\xRightarrow[\le 2]{*}{\boldsymbol{x}}$ and ${\boldsymbol{r}}\xRightarrow[\le 2]{*}{\boldsymbol y}$.
In other words, ${\boldsymbol{r}}$ is also the ${\le}2$-root of both ${\boldsymbol{x}}$ and ${\boldsymbol y}$.
Applying Proposition \ref{prop:jain2}, we have that ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are ${\le}2$-confusable and so, ${\le}3$-confusable.
Therefore, it remains to consider the case where ${\boldsymbol{r}}$ contains {\em at least three distinct} symbols.
Given ${\boldsymbol{r}}$, ${\boldsymbol{x}}$ and ${\boldsymbol y}$, our next step is to compute certain proper prefixes ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})$ and ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol y})$ of ${\boldsymbol{x}}$ and ${\boldsymbol y}$, respectively, such that {\em under certain conditions} the following holds:
\[ \left({\boldsymbol{x}} \mbox{ and } {\boldsymbol y} \mbox{ are confusable}\right)
\mbox{ if and only if }
\left({\boldsymbol{x}}\setminus{\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})\mbox{ and } {\boldsymbol y}\setminus{\rm *Pref}({\boldsymbol{r}},{\boldsymbol y}) \mbox{ are confusable}\right)\,.\]
Here, ${\boldsymbol{x}}\setminus {\boldsymbol{z}}$ denotes the word we obtain by removing the prefix ${\boldsymbol{z}}$ from ${\boldsymbol{x}}$.
Since the words on the righthand side are strictly shorter than the words on the lefthand side, we may repeat this process for the shorter words until the common root of the words has less than three distinct symbols.
To define the prefixes ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})$ and ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol y})$, we introduce more notation.
Now, ${\boldsymbol{r}}$ contains at least three distinct symbols.
Since ${\boldsymbol{r}}\in{\rm Irr}_{\le 3}(q)$,
we easily check that either $r_1r_2r_3$ or $r_2r_3r_4$ is a substring with three distinct symbols.
Define ${\rm main}({\boldsymbol{r}})$ to be the first substring in ${\boldsymbol{r}}$ with three distinct symbols and
we define the {\em region of ${\boldsymbol{r}}$}, denoted by ${\rm Reg}({\boldsymbol{r}})$, to be a certain prefix of ${\boldsymbol{r}}$ according to the following rule.
If $r_1=r_3$, we set
\[
{\rm main}({{\boldsymbol{r}}}) \triangleq r_2 r_3 r_4, \mbox{ and }
{\rm Reg}({{\boldsymbol{r}}}) \triangleq
\begin{cases}
r_1 r_2 r_3 r_4=r_1r_2r_1r_4, &\text{if } r_2 \neq r_5, \\
r_1 r_2 r_3 r_4 r_5=r_1r_2r_1r_4r_2, &\text{if } r_2 = r_5, r_3 \neq r_6, \\
r_1 r_2 r_3 r_4 r_5 r_6=r_1r_2r_1r_4r_2r_1, &\text{otherwise.}
\end{cases}
\]
If $r_1 \neq r_3$, we set
\[
{\rm main}({{\boldsymbol{r}}}) \triangleq r_1 r_2 r_3, \mbox{ and }
{\rm Reg}({{\boldsymbol{r}}}) \triangleq
\begin{cases}
r_1 r_2 r_3, &\text{if } r_1 \neq r_4, \\
r_1 r_2 r_3 r_4=r_1 r_2 r_3r_1, &\text{if } r_1 = r_4, r_2 \neq r_5, \\
r_1 r_2 r_3 r_4 r_5=r_1 r_2 r_3r_1r_2, &\text{otherwise.}
\end{cases}
\]
Intuitively, the region ${\rm Reg}({\boldsymbol{r}})$ of ${\boldsymbol{r}}$ is defined as above so that the following properties of the region hold.
\begin{lemma}\label{lem:reg}
Suppose the root ${\boldsymbol{r}}$ has three distinct symbols. Define ${\rm Reg}({\boldsymbol{r}})$ as above.
Then ${\rm Reg}({\boldsymbol{r}})$ can be written as ${\boldsymbol{w}}(abc)^\ell ab$, where $a,b,c$ are distinct symbols, $\ell\in\{0,1\}$ and
${\boldsymbol{w}}$ is prefix of length at most three over the alphabet $\{a,b,c\}$.
Furthermore, if ${\rm Reg}({\boldsymbol{r}})\xRightarrow[3_d]{*} {{\boldsymbol{z}}}$, then ${\boldsymbol{z}}$ can be written as ${\boldsymbol{w}}(abc)^m ab$
with $m\ge \ell$. Also, the symbol in ${\boldsymbol{r}}$ after the prefix ${\rm Reg}({\boldsymbol{r}})$ is not $c$.
\end{lemma}
\begin{example}
Suppose that ${{\boldsymbol{r}}}=010201$. Then ${\rm main}({{\boldsymbol{r}}})=102$ and ${\rm Reg}({{\boldsymbol{r}}})=0102$.
Indeed, we can write ${\rm Reg}({{\boldsymbol{r}}})$ as ${\boldsymbol{w}}(abc)^\ell ab$ with ${\boldsymbol{w}}=01$, $\ell=0$, and $abc=021$.
We also check that the next symbol after ${\rm Reg}({{\boldsymbol{r}}})$ in ${{\boldsymbol{r}}}$ is $0$, which is not $c=1$.
\end{example}
Given ${\rm Reg}({\boldsymbol{r}})={\boldsymbol{w}}(abc)^\ell ab$, we consider the longest prefix of ${{\boldsymbol{x}}}$ that is generated from ${\rm Reg}({\boldsymbol{r}})$ through a sequence of tandem duplications of length at most three.
We call this prefix the {\em extended region of ${\boldsymbol{r}}$ in ${\boldsymbol{x}}$} and
denote it with ${\rm Ext}({\rm Reg}({\boldsymbol{r}}),{{\boldsymbol{x}}})$.
Finally, we define our special prefix ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})$.
Suppose that ${\rm Ext}({\rm Reg}({\boldsymbol{r}}),{{\boldsymbol{x}}})=p_1p_2\cdots p_t={\boldsymbol p}$
and $p_i$ is the last appearance of $a$ in ${\boldsymbol p}$.
The {\em special prefix} ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})$ is given by $p_1p_2\cdots p_{i-1}$.
\begin{example} Let ${{\boldsymbol{r}}}=010201$ and ${{\boldsymbol{x}}}=01102021020120111$.
Recall that ${\rm Reg}({{\boldsymbol{r}}})=0102$ and $a=0$. Then
\[{\rm Ext}(0102,{{\boldsymbol{x}}})=0110202102 \mbox{ and } {\rm*Pref}({\boldsymbol{r}}, {\boldsymbol{x}})=01102021.\]
\end{example}
We summarize the relationships of ${\boldsymbol{r}}$, ${\boldsymbol{x}}$, ${\rm Reg}({{\boldsymbol{r}}})$, ${\rm Ext}({\rm Reg}({\boldsymbol{r}}),{{\boldsymbol{x}}})$, and ${\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})$ in Figure \ref{fig:prefixes} and
we are ready to state the main characterization theorem.
\begin{figure}[!t]
\begin{center}
\vspace{-2mm}
\includegraphics[width=15cm]{Prefixes.png}
\end{center}
\caption{Prefixes Used to Determine Confusability}
\label{fig:prefixes}
\end{figure}
\begin{theorem}\label{thm:main}
Suppose that ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are two words such that ${{\boldsymbol{r}}}=R_{\le 3}({{\boldsymbol{x}}})=R_{\le3}({{\boldsymbol y}})$ and
${\boldsymbol{r}}$ contains at least three distinct symbols.
Set
\begin{align*}
{\boldsymbol p}&={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol{x}}}),&
{\boldsymbol q}&={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol y}})\\
{\boldsymbol u}&={{\boldsymbol{x}}}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol{x}}})&
{\boldsymbol v}&={{\boldsymbol y}}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol y}}).
\end{align*}
Then ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable if and only if ${{\boldsymbol p}}$ and ${{\boldsymbol q}}$ are confusable, and ${{\boldsymbol u}}$ and ${{\boldsymbol v}}$ are confusable.
\end{theorem}
\begin{proof}
From Lemma \ref{lem:reg}, we may set ${\rm Reg}({\boldsymbol{r}})={\boldsymbol{w}}(abc)^\ell ab$.
Since ${\rm Reg}({\boldsymbol{r}})\xRightarrow[\le 3]{*}{\boldsymbol p}$, Lemma \ref{lem:ordering}
implies that ${\rm Reg}({\boldsymbol{r}})\xRightarrow[3_d]{*}R_{\le 2}({\boldsymbol p})$. So,
$R_{\le 2}({\boldsymbol p})={\boldsymbol{w}}(abc)^mab$ for some $m\ge \ell$, and $R_{\le 2}({\boldsymbol q})={\boldsymbol{w}}(abc)^jab$ for some $j\ge \ell$.
Suppose that ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable. By Theorem~\ref{confusable}, there exist $3_d$-descendants of ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ namely ${\bf x'}$ and ${\bf y'}$, such that
$R_{\le 2}({{\boldsymbol{x}}'})=R_{\le 2}({{\boldsymbol y}'})$.
Set
\begin{align*}
{\boldsymbol p}'&={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol{x}}'}),&
{\boldsymbol q}'&={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol y}'})\\
{\boldsymbol u}'&={{\boldsymbol{x}}'}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol{x}}'})&
{\boldsymbol v}'&={{\boldsymbol y}'}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol y}'}).
\end{align*}
Since the overlapping region between ${\boldsymbol p}$ and ${\boldsymbol u}$ must be of the form
$ab\cdots b$, each duplication of distinct triplets from ${{\boldsymbol{x}}}$ to ${{\boldsymbol{x}}'}$ must be either entirely in ${\boldsymbol p}$ or in ${\boldsymbol u}$. In other words,
${\boldsymbol p} \xRightarrow[3_d]{*} {\boldsymbol p}'$ and ${\boldsymbol u} \xRightarrow[3_d]{*} {\boldsymbol u}'$, and likewise ${\boldsymbol q} \xRightarrow[3_d]{*} {\boldsymbol q}'$ and ${\boldsymbol v} \xRightarrow[3_d]{*} {\boldsymbol v}'$.
Looking at the overlapping region between ${\boldsymbol p}'$ and ${\boldsymbol u}'$,
we can observe that $R_{\le 2}({{\boldsymbol{x}}'}) = R_{\le 2}({\boldsymbol p}' {\boldsymbol u}')$.
Then ${\rm Reg}({\boldsymbol{r}})\xRightarrow[\le 3]{*}{\boldsymbol p}'$ and Lemma \ref{lem:ordering}
imply that ${\rm Reg}({\boldsymbol{r}})\xRightarrow[3_d]{*}R_{\le 2}({\boldsymbol p})$ and so,
$R_{\le 2}({\boldsymbol p}')={\boldsymbol{w}}(abc)^iab$ for some $i\ge m$.
Next, ${\boldsymbol{w}}(abc)^i R_{\le 2}({\boldsymbol u}')$ is a ${\le}2$-irreducible word,
because $R_{\le 2}({\boldsymbol u}')$ starts with $ab$
and by definition, ${\boldsymbol{w}}(abc)^i$ ends with $c$ but not $cac$.
Therefore, ${\boldsymbol{w}}(abc)^i R_{\le 2}({\boldsymbol u}')$ is the ${\le}2$-root of ${\boldsymbol{x}}'$, or,
$R_{\le 2}({{\boldsymbol{x}}'})={\boldsymbol{w}}(abc)^i R_{\le 2}({\boldsymbol u}')$ and likewise $R_{\le 2}({{\boldsymbol y}'})={\boldsymbol{w}}(abc)^h R_{\le 2}({\boldsymbol v}')$ for some $h\ge j.$
Since the first symbol of $R_{\le 2}({{\boldsymbol{x}}'}\setminus {\boldsymbol p}')$ and
$R_{\le 2}({{\boldsymbol y}'}\setminus {\boldsymbol q}')$ cannot be $c$,
neither $R_{\le 2}({\boldsymbol u}')$ nor $R_{\le 2}({\boldsymbol v}')$ start with $abc$.
Therefore, since $R_{\le 2}({{\boldsymbol{x}}'})=R_{\le 2}({{\boldsymbol y}'})$,
we can deduce that $h=i$ and $R_{\le 2}({\boldsymbol u}')=R_{\le 2}({\boldsymbol v}')$. Since $h=i$, we have $R_{\le 2}({\boldsymbol p}')=R_{\le 2}({\boldsymbol q}')$. Together with ${\boldsymbol p} \xRightarrow[3_d]{*} {\boldsymbol p}'$, ${\boldsymbol q} \xRightarrow[3_d]{*} {\boldsymbol q}'$, by Theorem \ref{confusable}, we can have ${\boldsymbol p}$ and ${\boldsymbol q}$ are confusable.
Since ${\boldsymbol u} \xRightarrow[3_d]{*} {\boldsymbol u}'$, ${\boldsymbol v} \xRightarrow[3_d]{*} {\boldsymbol v}'$, and $R_{\le 2}({\boldsymbol u}')=R_{\le 2}({\boldsymbol v}')$, we can also conclude by Theorem \ref{confusable} that ${\boldsymbol u}$ and ${\boldsymbol v}$ are confusable.
Suppose that ${{\boldsymbol p}}$ and ${{\boldsymbol q}}$ are confusable, and ${\boldsymbol u}$ and ${\boldsymbol v}$ are confusable. Hence ${{\boldsymbol p}}{\boldsymbol u}$ and ${{\boldsymbol q}}{\boldsymbol v}$ are confusable. Note that ${{\boldsymbol p}}{\boldsymbol u}$ is a descendant of ${{\boldsymbol{x}}}$ and ${{\boldsymbol q}}{\boldsymbol v}$ is a descendant of ${{\boldsymbol y}}$, hence ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are also confusable.
\end{proof}
We can efficiently determine whether the prefixes ${\boldsymbol p}$ and ${\boldsymbol q}$ are confusable.
When the substring ${\rm main}({\boldsymbol{r}})$ appears the same number of times in $R_{\le 2}({\boldsymbol p})$ as in $R_{\le 2}({\boldsymbol q})$,
we may assume that $R_{\le 2}({\boldsymbol p})={\boldsymbol{w}}(abc)^mab=R_{\le 2}({\boldsymbol q})$.
Hence, ${{\boldsymbol p}}$ and ${{\boldsymbol q}}$ are confusable by Proposition \ref{prop:jain2}.
In particular, we have Corollary \ref{cor:samecount}.
Consider ${\boldsymbol{z}}=abc$, where $a$, $b$, $c$ are three distinct symbols.
Let ${\rm Count}({{\boldsymbol{z}}},{{\boldsymbol{x}}})$ denote the number of times ${{\boldsymbol{z}}}$ appears as a substring in ${{\boldsymbol{x}}}$,
and ${\rm Count}({\rm rot}({\boldsymbol{z}}),{{\boldsymbol{x}}})$ denote the number of times $abc$, $bca$, or $cab$ appears as a substring in ${{\boldsymbol{x}}}$.
\begin{algorithm}[h!]\footnotesize
\caption{{\sc Confuse}(${\boldsymbol{x}},{\boldsymbol y}$)}
\label{alg:confuse}
\begin{algorithmic}[1]
\small
\Require ${\boldsymbol{x}}, {\boldsymbol y} \in \Sigma_q^*$
\Ensure {\tt true} if ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable, {\tt false} otherwise.
\State ${\boldsymbol{r}}\gets R_{\le 3}({{\boldsymbol{x}}})$
\State ${\boldsymbol{r}}'\gets R_{\le 3}({{\boldsymbol y}})$
\If{$\bf r \neq \bf r'$}
\State \Return{{\tt false}}
\Else
\State \Return{{\sc Confuse-Recursion}(${\boldsymbol{x}},{\boldsymbol y},{\boldsymbol{r}}$)}
\EndIf
\Statex
\Procedure {Confuse-Recursion}{${\boldsymbol{x}}$,${\boldsymbol y}$,${\boldsymbol{r}}$}
\If{${\boldsymbol{r}}$ has at most two distinct symbols}
\State \Return {\tt true}
\Else
\State{Compute ${\rm main}({\boldsymbol{r}})$ and ${\rm Reg}({\boldsymbol{r}})$}
\State{${\boldsymbol p}\gets {\rm Ext}({\rm Reg}({\boldsymbol{r}}),{\boldsymbol{x}})$, \hspace{20mm}${\boldsymbol q}\gets {\rm Ext}({\rm Reg}({\boldsymbol{r}}),{\boldsymbol y})$}
\State{${\boldsymbol u}\gets {\boldsymbol{x}}\setminus{\rm *Pref}({\boldsymbol{r}},{\boldsymbol{x}})$, \hspace{19mm} ${\boldsymbol v}\gets {\boldsymbol y}\setminus{\rm *Pref}({\boldsymbol{r}},{\boldsymbol y})$}
\State{$C_{\boldsymbol p}\gets {\rm Count}({\rm main}({{\boldsymbol{r}}}),R_{\le_2}(,{\boldsymbol p}))$, \hspace{2mm}
$C_{\boldsymbol q}\gets {\rm Count}({\rm main}({{\boldsymbol{r}}}),R_{\le_2}({\boldsymbol q}))$}
\State{${\boldsymbol{r}}^*\gets {\boldsymbol{r}}\setminus {\boldsymbol{w}}(abc)^\ell$, where ${\rm Reg}({\boldsymbol{r}})={\boldsymbol{w}}(abc)^\ell ab$}
\If{$C_{\boldsymbol p}=C_{\boldsymbol q}$}
\State
\Return {{\sc Confuse-Recursion}(${\boldsymbol u},{\boldsymbol v},{\boldsymbol{r}}^*$)}
\ElsIf {$C_{\boldsymbol p}<C_{\boldsymbol q}$ and ${\rm Count}({\rm rot}({\rm main}({{\boldsymbol{r}}})),{\boldsymbol p})>0$}
\State
\Return {{\sc Confuse-Recursion}(${\boldsymbol u},{\boldsymbol v},{\boldsymbol{r}}^*$)}
\ElsIf {$C_{\boldsymbol p}>C_{\boldsymbol q}$ and ${\rm Count}({\rm rot}({\rm main}({{\boldsymbol{r}}})),{\boldsymbol q})>0$}
\State
\Return {{\sc Confuse-Recursion}(${\boldsymbol u},{\boldsymbol v},{\boldsymbol{r}}^*$)}
\Else
\State \Return {\tt false}
\EndIf
\EndIf
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{corollary}\label{cor:samecount}
Suppose that ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are two words such that ${{\boldsymbol{r}}}=R_{\le 3}({{\boldsymbol{x}}})=R_{\le3}({{\boldsymbol y}})$ and
${\boldsymbol{r}}$ contains at least three distinct symbols.
Set
${\boldsymbol p}={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol{x}}})$ and ${\boldsymbol q}={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol y}})$.
Suppose further that
${\rm Count}({\rm main}({{\boldsymbol{r}}}),R_{\le2}({\boldsymbol p}))={\rm Count}({\rm main}({{\boldsymbol{r}}}),R_{\le 2}({\boldsymbol q}))$.
Then ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable if and only if
${{\boldsymbol{x}}}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol{x}}})$ and ${{\boldsymbol y}}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol y}})$ are confusable.
\end{corollary}
On the other hand, when the number of appearances is not equal, we may modify the proof
to obtain the following corollary.
A detailed proof of the corollary is deferred to Appendix \ref{app:cor1}.
\begin{corollary}\label{cor:main}
Suppose that ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are two words such that ${{\boldsymbol{r}}}=R_{\le 3}({{\boldsymbol{x}}})=R_{\le3}({{\boldsymbol y}})$ and
${\boldsymbol{r}}$ contains at least three distinct symbols.
Set
${\boldsymbol p}={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol{x}}})$ and ${\boldsymbol q}={\rm Ext}({\rm Reg}({{\boldsymbol{r}}}),{{\boldsymbol y}})$.
Suppose further that
${\rm Count}({\rm main}({{\boldsymbol{r}}}),R_{\le2}({\boldsymbol p}))<{\rm Count}({\rm main}({{\boldsymbol{r}}}),R_{\le 2}({\boldsymbol q}))$.
Then ${{\boldsymbol{x}}}$ and ${{\boldsymbol y}}$ are confusable if and only if
${\rm Count}({\rm rot}({\rm main}({{\boldsymbol{r}}})),{\boldsymbol p})>0$ and ${{\boldsymbol{x}}}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol{x}}})$ and ${{\boldsymbol y}}\setminus{\rm *Pref}({{\boldsymbol{r}}},{{\boldsymbol y}})$ are confusable.
\end{corollary}
\subsection{Confusability Algorithm}\label{sec:implementation}
Based on the ideas in the previous subsection,
we formally describe our algorithm {\sc Confuse} in Algorithm \ref{alg:confuse}
and demonstrate that the running time is linear.
In the following analysis, we assume the definitions in Algorithm \ref{alg:confuse}.
\begin{itemize}
\setlength\itemsep{2mm}
\item \underline{${\rm main}({\boldsymbol{r}})$ and ${\rm Reg}({\boldsymbol{r}})$ can be computed in constant time}. This is clear from definition.
\item \underline{The prefixes ${\boldsymbol p}$ and ${\boldsymbol q}$ can be computed in time linear in $|{\boldsymbol p}|$ and $|{\boldsymbol q}|$, respectively}.
From Proposition \ref{prop:regular}, we have that $D^*_{\le 3}({\rm Reg}({\boldsymbol{r}}))$
is a regular language.
Since the length of ${\rm Reg}({\boldsymbol{r}})$ is at most six, the finite automaton generating the language has a constant number of vertices.
Hence, the longest prefixes of ${\boldsymbol{x}}$ and ${\boldsymbol y}$ that are accepted by the finite automaton are precisely ${\boldsymbol p}$ and ${\boldsymbol q}$, respectively.
In Appendix \ref{app:prefixes}, we show that the sum of the lengths of all prefixes that are computed in all recursive calls is at most $3(|{\boldsymbol{x}}|+|{\boldsymbol y}|)$.
Therefore, the computation of all prefixes takes linear time.
\item \underline{In the recursive calls, the root ${\boldsymbol{r}}^*$ can be computed in constant time}. As mentioned earlier, the first step is to compute the roots of ${\boldsymbol{x}}$ and ${\boldsymbol y}$ in linear time.
When the roots are equal to ${\boldsymbol{r}}$ and certain conditions are met,
we make the recursive call {\sc Confuse-Recursion}.
If the roots are recomputed for the shorter words ${\boldsymbol u}$ and ${\boldsymbol v}$, then the overall running time for {\sc Confuse} is quadratic.
To avoid this, we observe that the roots for the shorter words are always equal and
in fact, the common root ${\boldsymbol{r}}^*$ of ${\boldsymbol u}$ and ${\boldsymbol v}$ may be easily inferred from ${\boldsymbol{r}}$.
More precisely, if ${\rm Reg}({\boldsymbol{r}})={\boldsymbol{w}}(abc)^\ell ab$, then ${\boldsymbol{r}}^*= {\boldsymbol{r}}\setminus {\boldsymbol{w}}(abc)^\ell$.
\item \underline{In the recursive calls, determining ${\boldsymbol{r}}$ has at most two symbols can be done in constant time}.
If ${\boldsymbol{r}}$ has length at least four, then ${\boldsymbol{r}}$ necessarily has at least three distinct symbols.
Otherwise, it contradicts the fact that ${\boldsymbol{r}}$ is irreducible.
Therefore, it suffices to check if the first four symbols ${\boldsymbol{r}}$ contain three distinct symbols.
\end{itemize}
In summary, from the above observations,
we conclude that the running time is $O(\max\{|{\boldsymbol{x}}|,|{\boldsymbol y}|\})$.
\section{${\le}3$-Tandem-Duplication Codes}\label{sec:codes}
We use insights gained from the previous section to construct $(n,{\le}3;q)$-TD codes. Motivated by the concept of roots, we consider a ${\le}3$-irreducible word ${\boldsymbol{r}}$
and we say that a $(n,{\le}3;q)$-TD code $\cal C$ is an $(n,{\le}3;{\boldsymbol{r}})$-TD code if
all words in $\cal C$ belong to $D_{\le 3}^*({\boldsymbol{r}})$.
For ${\boldsymbol{r}}\in{\rm Irr}_{\le3}(i,q)$ with $i\le n$,
suppose that ${\cal C}(n,{\boldsymbol{r}})$ is an $(n,{\le}3;{\boldsymbol{r}})$-TD code.
Then Proposition \ref{prop:jain2} implies that
$\bigcup_{{\boldsymbol{r}}\in{\rm Irr}_{\le3}(i,q),\, i\le n}{\cal C}(n,{\boldsymbol{r}})$ is an $(n,{\le}3;q)$-TD code.
Trivially, $\{\xi_{n-i}({{\boldsymbol{r}}})\}$ is an $(n,{\le}3;{\boldsymbol{r}})$-TD code for all ${{\boldsymbol{x}}} \in {\rm Irr}_{\le 3}(i,q)$.
Taking the union of these codes, we recover Construction \ref{code:jain}.
Therefore, in the rest of this section, our objective is
to provide estimates on the size of an optimal $(n,{\le}3;{\boldsymbol{r}})$-TD code for fixed ${\le}3$-irreducible word ${\boldsymbol{r}}$.
To simplify our discussion, we focus on the case where $q=3$ and let
$T(n)$ and $T(n,{\boldsymbol{r}})$ to denote the sizes of an optimal $(n,{\le}3;3)$-TD code
and an optimal $(n,{\le}3;{\boldsymbol{r}})$-TD code, respectively.
Pick ${\boldsymbol{x}}\in D_{{\le}3}^*({\boldsymbol{r}})$. Using concepts from Section \ref{sec:confusable},
we define the following combinatorial characterisation of ${\boldsymbol{x}}$.
Set ${\boldsymbol{x}}_1={\boldsymbol{x}}$ and ${\boldsymbol{r}}_1={\boldsymbol{r}}$.
For $i\ge 2$, set
${\boldsymbol{x}}_i={\boldsymbol{x}}_{i-1}\setminus *{\rm Pref}({\boldsymbol{r}}_{i-1},{\boldsymbol{x}}_{i-1}) \mbox{ and } {\boldsymbol{r}}_i= R_{{\le}3}({\boldsymbol{x}}_i).$
We terminate this recursion when ${\boldsymbol{r}}_{m+1}$ has less than three distinct symbols.
If ${\boldsymbol{r}}_{m}$ is the last root to have three distinct symbols, we say that ${\boldsymbol{r}}$ has $m$ {\em regions} and
for $1\le i\le m$, define
\[ {\boldsymbol p}_i={\rm Ext}({\rm Reg}({\boldsymbol{r}}_i),{\boldsymbol{x}}_i),\, {\boldsymbol t}_i={\rm main}({\boldsymbol{r}}_i),\, c_i={\rm Count}({\boldsymbol t}_i,R_{{\le}2}({\boldsymbol p}_i)), \mbox{ and }
\delta_i=
\begin{cases}
+, &\mbox{if }{\rm Count}({\rm rot}({\boldsymbol t}_i),{\boldsymbol p}_i)>0,\\
-, &\mbox{otherwise.}
\end{cases}\]
For the word ${\boldsymbol{x}}$, we define its {\em label} by
${\rm Label}({\boldsymbol{x}})=({\boldsymbol{r}},(c_1,\delta_1),(c_2,\delta_2),\ldots,(c_m,\delta_m))$.
\begin{example}
Consider ${\boldsymbol{r}}=01210$. Then ${\boldsymbol{r}}$ has two regions with ${\boldsymbol t}_1=012$ and ${\boldsymbol t}_2=210$.
Consider also the words $01210210$, $01201210$, and $01111210$ that belongs to $D_{{\le}3}^*(01210)$.
Then their labels are as follow.
\begin{align*}
{\rm Label}(01210210)&=(01210, (1,+),(2,+)),\\
{\rm Label}(01201210)&=(01210, (2,+),(1,+)),\\
{\rm Label}(01112110)&=(01210, (1,-),(1,-)).
\end{align*}
Then it follows from Theorem \ref{thm:main}, Corollaries \ref{cor:samecount} and \ref{cor:main} that
$01210210$ and $01210210$ are confusable, while
$01210210$ and $01112110$ are not confusable.
\end{example}
More generally, the following proposition is immediate from
Theorem \ref{thm:main}, Corollaries \ref{cor:samecount} and \ref{cor:main}.
\begin{proposition}\label{prop:label}
Let ${\boldsymbol{x}}$ and ${\boldsymbol y}$ belong to $D_{{\le}3}^*({\boldsymbol{r}})$.
Suppose that
\begin{align*}
{\rm Label}({\boldsymbol{x}})&\triangleq\left({\boldsymbol{r}}, \left(c_1^{({\boldsymbol{x}})},\delta_1^{({\boldsymbol{x}})}\right), \left(c_2^{({\boldsymbol{x}})},\delta_2^{({\boldsymbol{x}})}\right), \ldots, \left(c_m^{({\boldsymbol{x}})},\delta_m^{({\boldsymbol{x}})}\right)\right),\\
{\rm Label}({\boldsymbol y})&\triangleq\left({\boldsymbol{r}}, \left(c_1^{({\boldsymbol y})},\delta_1^{({\boldsymbol y})}\right), \left(c_2^{({\boldsymbol y})},\delta_2^{({\boldsymbol y})}\right), \ldots, \left(c_m^{({\boldsymbol y})},\delta_m^{({\boldsymbol y})}\right)\right).
\end{align*}
Then ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are not ${\le}3$-confusable if and only if
for some $1\le i\le m$,
\[ \left(c_i^{({\boldsymbol{x}})}<c_i^{({\boldsymbol y})} \mbox{ and } \delta_i^{({\boldsymbol{x}})}=-\right)
\mbox{ or }
\left(c_i^{({\boldsymbol{x}})}>c_i^{({\boldsymbol y})} \mbox{ and } \delta_i^{({\boldsymbol y})}=-\right).\]
\end{proposition}
Immediate from Proposition \ref{prop:label} are certain sufficient conditions for two words to be confusable.
\begin{corollary}\label{cor:label}
Let ${\boldsymbol{x}}, {\boldsymbol y}\in D_{{\le}3}^*({\boldsymbol{r}})$ and their labels ${\rm Label}({\boldsymbol{x}})$ and ${\rm Label}({\boldsymbol y})$
be as defined in Proposition \ref{prop:label}.
\begin{enumerate}[(i)]
\item If $c_i^{({\boldsymbol{x}})}=c_i^{({\boldsymbol y})}$ for all $i$, then ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are ${\le}3$-confusable.
\item If $\delta_i^{({\boldsymbol{x}})}=\delta_i^{({\boldsymbol y})}=+$ for all $i$, then ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are ${\le}3$-confusable.
\end{enumerate}
\end{corollary}
From Corollary \ref{cor:label}, we use the number of integer solutions to certain equations as
an upper bound for $T(n,{\boldsymbol{r}})$. As this combinatorial argument is fairly technical,
we state the upper bound and defer the proof to Appendix \ref{app:upper}.
\begin{proposition}\label{prop:upper}
Let $i\le n$. Suppose that ${\boldsymbol{r}}\in {\rm Irr}_{{\le}3}(i,3)$ has $m$ regions.
Then
\[ T(n,{\boldsymbol{r}})\le U(n,i,m)\triangleq
\begin{cases}
\binom{(n-i)/3+m}{m}-\binom{(n-i)/3+m-1}{m-1}+1, & \mbox{if 3 divides $n-i$},\\
\binom{\floor{(n-i)/3}+m}{m}, &\mbox{otherwise.}
\end{cases}
\]
\end{proposition}
Immediate from Proposition \ref{prop:upper} is that $T(|{\boldsymbol{r}}|,{\boldsymbol{r}})=T(|{\boldsymbol{r}}|+1,{\boldsymbol{r}})=T(|{\boldsymbol{r}}|+2,{\boldsymbol{r}})=1$.
Next, we provide lower bounds for $T(n,{\boldsymbol{r}})$ by constructing $(n,{\le}3;{\boldsymbol{r}})$-TD codes.
We first show that there exists an $(n,{\le}3;{\boldsymbol{r}})$-TD code of size two.
\begin{construction}\label{code:2}
Let ${\boldsymbol{r}}\in{\rm Irr}_{\le 3}(i,3)$ for some $i\ge 4$. Set
\[{\cal C}_{2}\triangleq
\begin{cases}
\{r_1r_2r_3r_1r_2r_3r_4\cdots r_i,\hspace{4mm} r_1r_2r_2r_3r_3r_4r_4\cdots r_i\},
&\mbox{if $r_1\ne r_3$,}\\
\{r_1r_2r_1r_4r_2r_1r_4r_5\cdots r_i,\hspace{1mm} r_1r_2r_1r_1r_4r_4r_5r_5\cdots r_i\},
&\mbox{if $r_1= r_3$ and $i\ge 5$}\\
\{r_1r_2r_1r_4r_2r_1r_4,\hspace{12mm} r_1r_2r_1r_1r_4r_4r_4\},
&\mbox{if $r_1= r_3$ and $i=4$}\\
\end{cases}
\]
Then ${\cal C}_{2}$ is an $(i+3,{\le}3;{\boldsymbol{r}})$-TD code of size two.
By Proposition \ref{prop:upper}, ${\mathcal C}_2$ is optimal.
\end{construction}
\begin{proof}Let ${\cal C}_2=\{{\boldsymbol{x}}, {\boldsymbol y}\}$ and the labels of the words be as given in Proposition \ref{prop:label}.
Observe that in all cases,
$\left(c_1^{({\boldsymbol{x}})},\delta_1^{(x)}\right)=(2,+)$ and $\left(c_1^{({\boldsymbol y})},\delta_1^{(y)}\right)=(1,-)$.
Proposition \ref{prop:label} then implies that ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are not confusable.
\end{proof}
Following the above construction, we have a improvement to Construction \ref{code:jain}.
\begin{corollary}For $n\ge 6$, we have that
\[T(n)\ge \sum_{\substack{1\le i\le 3, {\rm~or}\\ n-2\le i\le n}} {\rm Irr}_{\le 3}(i,3)
+2\sum_{i=4}^{n-3}{\rm Irr}_{\le 3}(i,3).\]
\end{corollary}
Proposition \ref{prop:label} suggests that we examine code constructions according to the number of regions of ${\boldsymbol{r}}$.
When $|{\boldsymbol{r}}|\le 2$, the number of regions is zero and hence $T(n,{\boldsymbol{r}})=1$.
We consider the smallest nontrivial case where ${\boldsymbol{r}}$ has exactly one region.
Without loss of generality, assume that ${\boldsymbol{r}}\in R\triangleq\{012,0120,01201,1012,10120,101201$, $0121,01202,012010,10121,101202,1012010\}$, and so, ${\boldsymbol t}_1=012$.
In this case, we have the following construction whose proof is deferred to Appendix \ref{app:codes}.
\begin{construction}\label{code:oneregion}
For ${\boldsymbol{r}}\in R$ and $\ell\ge 1$, define the words ${\boldsymbol{x}}({\boldsymbol{r}},\ell)$ and ${\boldsymbol{z}}({\boldsymbol{r}},\ell)$ with the following rule.
\vspace{1mm}
\begin{center}
\begin{tabular}{|l|l|l||l|l|l|}
\hline
${\boldsymbol{r}}$ & ${\boldsymbol{x}}({\boldsymbol{r}},\ell)$ & ${\boldsymbol{z}}({\boldsymbol{r}},\ell)$ &
${\boldsymbol{r}}$ & ${\boldsymbol{x}}({\boldsymbol{r}},\ell)$ & ${\boldsymbol{z}}({\boldsymbol{r}},\ell)$ \\ \hline
$012$ & $0(112200)^{\ell-1} 112$ & $(012)^\ell$ &
$1012$ & $10(112200)^{\ell-1} 112$ & $1(012)^\ell$ \\ \hline
$0120$ & $0(112200)^{\ell-1} 11220$ & $(012)^\ell0$ &
$10120$ & $10(112200)^{\ell-1} 11220$ & $1(012)^\ell0$ \\ \hline
$01201$ & $0(112200)^{\ell-1} 1122001$ & $(012)^\ell01$ &
$101201$ & $10(112200)^{\ell-1} 1122001$ & $1(012)^\ell01$ \\ \hline
$0121$ & $0(112200)^{\ell-1} 1121$ & $(012)^\ell1$ &
$10121$ & $10(112200)^{\ell-1} 1121$ & $1(012)^\ell1$ \\ \hline
$01202$ & $0(112200)^{\ell-1} 112202$ & $(012)^\ell02$ &
$101202$ & $10(112200)^{\ell-1} 112202$ & $1(012)^\ell02$ \\ \hline
$012010$ & $0(112200)^{\ell-1} 11220010$ & $(012)^\ell010$ &
$1012010$ & $10(112200)^{\ell-1} 11220010$ & $1(012)^\ell010$ \\ \hline
\end{tabular}
\end{center}
For ${\boldsymbol{r}}\in R$ and $n\ge |{\boldsymbol{x}}({\boldsymbol{r}},2)|$, set $\ell_z=\floor{(n-|{\boldsymbol{r}}|)/3}+1$ and
\[{\cal C}_{{\boldsymbol{r}}}(n)
\triangleq \left\{\xi_{n-|{\boldsymbol{x}}({\boldsymbol{r}},\ell)|}({\boldsymbol{x}}({\boldsymbol{r}},\ell)):|{\boldsymbol{x}}({\boldsymbol{r}},\ell)| \le n\right\}
\cup \left\{\xi_{n-|{\boldsymbol{z}}({\boldsymbol{r}},\ell_z)|} ({\boldsymbol{z}}({\boldsymbol{r}},\ell_z)) \right\}.
\]
Then ${\cal C}_{{\boldsymbol{r}}}(n)$ is an $(n,{\le}3;{\boldsymbol{r}})$-TD code.
Furthermore, ${\cal C}_{{\boldsymbol{r}}}(n)$ in optimal.
Therefore, if we set $n_2=|{\boldsymbol{x}}(r,2)|$,
we have
\[ T({\boldsymbol{r}},n) =
\begin{cases}
\floor{\frac{n-n_2}{6}}+3 &\mbox{if $n\ge n_2$},\\
2, &\mbox{if $|{\boldsymbol{r}}|+3\le n< n_2$},\\
1, &\mbox{otherwise.}
\end{cases}
\]
\end{construction}
Hence, the value of $T(n,{\boldsymbol{r}})$ is completely determined whenever ${\boldsymbol{r}}$ has at most one region.
Combining this result with Proposition \ref{prop:upper}, we have the following theorem.
\begin{theorem}\label{thm:upper}
Let $I(i,m)$ denote the number of irreducible words in ${\rm Irr}_{{\le}3}(i,3)$ with exactly $m$ regions, and
$U(n,i,m)$ be as defined in Proposition \ref{prop:upper}.
Then
\begin{equation}
T(n)\le \sum_{{\boldsymbol{r}} \in R} T(n,{\boldsymbol{r}}) +\sum_{i=5}^n \sum_{m=2}^{i} I(i,m)U(n,i,m).\label{eq:upper}
\end{equation}
\end{theorem}
In addition to the above constructions,
we construct tandem-duplication codes for small lengths by searching for them exhaustively.
Specifically, fix ${\boldsymbol{r}}\in{\rm Irr}_{\le 3}(i,3)$ and $n\ge i$, we construct the graph ${\cal G}(n,{\boldsymbol{r}})$,
whose vertices correspond to the set of all labels of descendants of ${\boldsymbol{r}}$ of length $n$, or
$\{{\boldsymbol L}: {\boldsymbol L}={\rm Label}({\boldsymbol{x}}) \mbox{ and } {\boldsymbol{x}}\in \Sigma_3^n\cap D^*_{\le 3}({\boldsymbol{r}})\}$.
Two vertices or labels ${\boldsymbol L}_1$ and ${\boldsymbol L}_2$ are connected in ${\cal G}(n,{\boldsymbol{r}})$ if and only if
the words whose labels are ${\boldsymbol L}_1$ and ${\boldsymbol L}_2$ are not ${\le}3$-confusable.
Hence, a clique of size $M$ in the graph ${\cal G}(n,{\boldsymbol{r}})$
correspond to an $(n,{\le}3;{\boldsymbol{r}})$-TD code of size $M$ and
$T(n,{\boldsymbol{r}})$ is the maximum size of a clique in ${\cal G}(n,{\boldsymbol{r}})$.
We use the exact algorithm {\tt MaxCliqueDyn} \cite{Konc} to
determine the maximum size of the clique in these graph ${\cal G}(n,{\boldsymbol{r}})$ for ${n\le 20}$.
Since $T(n)=\sum_{{\boldsymbol{r}}\in{\rm Irr}_{\le3}(i,3),\, i\le n}T(n,{\boldsymbol{r}})$,
we tabulate the results $T(n)$ in Table \ref{table:tn}.
\begin{remark}\hfill
\begin{itemize}
\item Such a method to compute $T(n,{\boldsymbol{r}})$ is only possible because
we have developed the algorithm to determine confusability in Section \ref{sec:confusable}.
Prior to this work, the necessary conditions for ${\le}3$-confusability was not known and
hence, methods to compute $T(n,{\boldsymbol{r}})$ were not available.
\item For a fixed length and root, even though the set of descendants is huge, the set of all labels are significantly smaller. Hence, the task of computing maximum cliques remains feasible. More concretely, when $n=20$,
the order of the largest graph ${\cal G}(20,{\boldsymbol{r}})$ is {366},
despite the fact that the average size%
\footnote{The number of ${\le}3$-irreducible ternary words of length at most twenty is 27687.}
of $\Sigma_3^{20}\cap D^*_{\le 3}({\boldsymbol{r}})$ is $3^{20}/27687\approx 125935$.
\end{itemize}
\end{remark}
Finally, we develop recursive constructions in the following proposition,
whose proof is deferred to Appendix \ref{app:codes}.
\begin{proposition}\label{prop:recursive}
Let ${\boldsymbol{r}}=r_1r_2\cdots r_i\in{\rm Irr}_{\le 3}(i,3)$. Then the following holds.
\[
\small
T(n,{\boldsymbol{r}})\ge
\begin{cases}
T(n-1,{\boldsymbol{r}}\setminus r_1), &
\mbox{if $r_1= r_3$},\\
\max\{2T(n-4,{\boldsymbol{r}}\setminus r_1), 3T(n-8,{\boldsymbol{r}}\setminus r_1)\}, &
\mbox{if $r_1\ne r_3$, $r_1\ne r_4$},\\
\max\{2T(n-5,{\boldsymbol{r}}\setminus r_1r_2), 3T(n-10,{\boldsymbol{r}}\setminus r_1r_2)\}, &
\mbox{if $r_1\ne r_3$, $r_1=r_4$, $r_2\ne r_5$},\\
\max\{2T(n-6,{\boldsymbol{r}}\setminus r_1r_2r_3), 3T(n-12,{\boldsymbol{r}}\setminus r_1r_2r_3)\}, &
\mbox{if $r_1\ne r_3$, $r_1=r_4$, $r_2= r_5$}.\\
\end{cases}
\]
Furthermore, $T(n,{\boldsymbol{r}})\ge T(n-1,{\boldsymbol{r}})$ and $T(n,{\boldsymbol{r}})=T(n,{\boldsymbol{r}}^R)$,
where ${\boldsymbol{z}}^R$ denotes the reverse of word ${\boldsymbol{z}}$.
\end{proposition}
{Using Proposition \ref{prop:recursive} with
Constructions \ref{code:2} and \ref{code:oneregion}
and the values computed by {\tt MaxCliqueDyn},
we derive lower bounds for $T(n)$ for {$21\le n\le 30$}.
The results are summarized in Table \ref{table:tn}.
In addition to the lower bounds for the code size $T(n)$,
we also compare the upper bounds in Proposition \ref{prop:irr2-upper} and \eqref{eq:upper}.
Observe that \eqref{eq:upper} is tight up to lengths at most ten and
the constructions in this paper improve the rates%
\footnote{The rate of a code ${\mathcal C}$ of length $n$ is given by $(\log_3 |{\mathcal C}|)/n$.}
for Construction \ref{code:jain} by as much as {6.74\%}.
\begin{table}
\caption{Estimates and Exact Values for $T(n)$}
\label{table:tn}
\centering
\begin{tabular}{|c|r|r|r|r||c|r|r|r|r|}
\hline
Length &
{Constr. \ref{code:jain}} &
This paper &
Eq \eqref{eq:upper} &
Prop. \ref{prop:irr2-upper}&
Length &
{Constr. \ref{code:jain}} &
This paper &
Eq \eqref{eq:upper} &
Prop. \ref{prop:irr2-upper} \\
\hline
\input{Table.tex}
\hline
\end{tabular}
\vspace{1mm}
Optimal values of $T(n)$ are highlighted in {\bf bold}.
\vspace{-2mm}
\end{table}
\section{Discussion}
We studied the problem of determining the ${\le}k$-confusability of two words.
Combining the results of this paper, we have linear-time algorithms to solve the confusability problem for $k\in\{1,2,3\}$.
It remains open whether there exist efficient algorithms to determine ${\le}4$-confusability.
One key obstacle is the fact that Proposition \ref{prop:jain2}(iii) does not hold for $k=4$.
In particular, there exists ${\boldsymbol{x}}$ and ${\boldsymbol y}$ such that $R_{\le 4}({\boldsymbol{x}})\ne R_{\le 4}({\boldsymbol y})$,
but ${\boldsymbol{x}}$ and ${\boldsymbol y}$ are ${\le}4$-confusable.
An example is provided by Jain {\em et al.}\cite{Jain},
where ${\boldsymbol{x}}=012$ and ${\boldsymbol y}=0121012$ belongs to ${\rm Irr}_{\le4}(3)$, but
$012101212$ is a common descendant of ${\boldsymbol{x}}$ and ${\boldsymbol y}$.
Another approach is to consider the intersection of the descendant cones
$D_{\le 4}^*({\boldsymbol{x}})\cap D_{\le 4}({\boldsymbol y})$ as formal languages.
Unfortunately, Leupold {\em et al.}\cite{Leupold1} demonstrated that
while the language $D_{\le 4}^*({\boldsymbol{x}})$ is context free for all ${\boldsymbol{x}}$, the language is not regular in general.
However, given two context free languages $L_1$ and $L_2$,
determining whether $L_1\cap L_2$ is empty is an undecidable problem
(see, for example, Sipser \cite[Exercise 5.32]{Sipser2006}). While this does not imply that the confusability problem is undecidable,
we nevertheless conjecture that it is undecidable for $k\ge 4$.
\section*{Acknowledgement}
We are grateful for an anonymous reviewer who proposed the definition of label given in Section \ref{sec:codes}.
This definition has led to a more succint presentation and a spectrum of new results, including Proposition \ref{prop:upper}
and the computation of the size of optimal codes of lengths up to twenty.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The VLT-FLAMES Tarantula Survey (VFTS, \cite[Evans et al. 2011]{evans11}) is an ESO large programme that has obtained mid-to-high resolution multi-epoch optical
spectra of hundreds of O-type stars in 30\,Doradus, the largest H\,\textsc{ii} region in the Local Group, located in the Large Magellanic Cloud. VFTS was developed aiming at the study of
rotation,
binarity and wind properties of an unprecedented number of massive stars within the same star-forming region. Within this project, a huge effort has been made in the study of O-type stars,
including the
analysis of multiplicity (\cite[Sana et al. 2013]{sana13}), spectral classifications (\cite[Walborn et al. 2014]{w14}) and the distribution of rotational velocities
(\cite[Ram\'irez-Agudelo et al. 2013, 2015]{oscar13,oscar15}). These studies are essential for the estimation of stellar parameters and chemical abundances of the complete O-type sample, which will
allow to investigate fundamental questions in stellar and cluster evolution. To date, the quantitative study of the O-type sample is about to be complete (see
\cite[Sab\'in-Sanjuli\'an et al. 2014, 2017]{cssj14,cssj17}; \cite[Ram\'irez-Agudelo et al. 2017]{oscar17}), as well as the determination of nitrogen abundances (\cite[Grin et al. 2016]{grin16},
Sim\'on-D\'iaz et al. in prep.)
In this work, we study the properties of a sample of O stars close to the ZAMS, those with luminosity classes IV and V.
\section{Observations and data sample}
We have used spectroscopic data obtained by means of the Medusa mode of the FLAMES spectrograph at the VLT (Paranal, Chile). In addition, $I$-band images obtained with the $HST$ (GO12499, P.~I.:
D.~J.~Lennon) were
utilized to evaluate possible contamination in the Medusa fibres.
We have selected a sample of 105 likely single and unevolved O stars based on the multiplicity properties derived by \cite[Sana et al. (2013)]{sana13} and the spectral classifications
by \cite[Walborn et al. (2014)]{w14}. Our sample includes O stars with luminosity classes V and IV, as well as Vz, V-III (uncertain classification) and III-IV (interpolation between
III and IV).
\section{Quantitative spectroscopic analysis} \label{sec:qsa}
The quantitative analysis of the spectroscopic data for our sample of O dwarfs was performed by means of the IACOB Grid-Based Automatic Tool (\textsc{iacob-gbat}, see
\cite[Sim\'on-D\'iaz et al. 2011]{ssd11}, \cite[Sab\'in-Sanjuli\'an et al. 2014]{cssj14}). \textsc{iacob-gbat} is based on a grid of $\sim$190\,000 precomputed \textsc{fastwind} stellar atmosphere
models (\cite[Santolaya-Rey
et al. 1997]{sr97},
\cite[Puls et al. 2005]{puls05}) and a $\chi^2$ algorithm that compares observed and synthetic H and He line profiles.
Absolute magnitudes in the $V$-band calculated by \cite[Ma\'iz-Apell\'aniz et al. (2014)]{jma14} and projected rotational velocities by \cite[Ram\'irez-Agudelo et al. (2013)]{oscar13} were used.
We estimated mean values and uncertainties for effective temperature T$_{\rm eff}$, surface gravity $g$, helium abundance $Y(He)$, stellar radius $R$, luminosity $L$, spectroscopic mass $M_{\rm sp}$
and wind-strength $Q$-parameter (defined as $Q$\,=\,$\dot{M}(R\,v_{\infty})^{-3/2}$, where
$\dot{M}$ is the mass-loss rate and $v_{\infty}$ the wind terminal velocity). Evolutionary masses ($M_{\rm ev}$) were obtained by means of the Bayesian tool \textsc{bonnsai} (\cite[Schneider et al.
2014]{bonnsai}),
which used
evolutionary models by \cite[Brott et al. (2011)]{brott11} and \cite[K\"ohler et al. (2015)]{kohler15}.
Most of the stars in our sample showed strong nebular contamination, but only 11 critical cases had to be analyzed using nitrogen lines (HHeN). This diagnostic was also used for 9 cases
with too weak or
inexistent He\,\textsc{i} lines. We could only provide upper limits for log\,$Q$ for about 70\% of the sample due to thin winds. In addition, possible/confirmed contamination in fibre was present in
several stars, which could have altered the determination of physical parameters. A few cases were found with too low helium abundances and too high gravities, a possible indication of undetected
binarity.
\section{Spectral calibration}
Figure~\ref{fig:calibration} represents our T$_{\rm eff}$ scale as a function of spectral types. We perform a linear fit to the points excluding O2 stars, since they show indications of
undetected binarity and/or contamination in fibre. When comparing with the most recent and complete spectral calibration for O dwarfs in the LMC before this work (\cite[Rivero-Gonz\'alez et al.
2012a,b]{rivero12a,rivero12b}, who used HHeN diagnostics to analyse optical spectra of 25 stars, including 16 dwarfs), we found that there is an excellent agreement between both scales for spectral
types later than O4. In the earliest regime, Rivero-Gonz\'alez et al. utilized a quadratic fit obtaining hotter temperatures than ours, although the estimates for our O2 stars agree with their
calibration. However, we cannot reach a firm conclusion about the necessity of changing the slope in the O2-O4 regime due to the small number of observed stars and their very likely
binarity/multiplicity. A larger sample of O-type stars earlier than O4, as well as information on their possible composite nature are necessary.
\begin{figure}[h!!!]
\begin{center}
\includegraphics[scale=0.5,angle=90]{calib_spt_teff_sigmas.eps}
\caption{Effective temperature calibration compared to that derived by \cite[Rivero-Gonz\'alez et al. (2012a,b)]{rivero12a,rivero12b}. For each spectral type, the dark grey circles and bars
represent the
mean value and standard deviation of our T$_{\rm eff}$ estimates. Our linear fit is represented by the black line, and the
grey zones indicate the associated uncertainties. O2 stars have been excluded from the fit due to their possible contamination in fibre and/or undetected binarity. }
\label{fig:calibration}
\end{center}
\end{figure}
\section{Mass discrepancy }
The mass discrepancy problem consists on a lack of agreement between spectroscopic and evolutionary masses. Since its identification by \cite[Herrero et al.
(1992)]{h92}, several explanations have been proposed but it remains unresolved due to the diversity of results in different environments.
We compare our derived spectroscopic and evolutionary masses in Figure~\ref{fig:mass_discrep}. We find a certain trend in the distribution: for masses
below 20~M$_{\odot}$, we can only find stars with $M_{\rm ev}$\,$>$\,$M_{\rm sp}$, i.e., a positive discrepancy, while for higher masses both positive and negative discrepancies are present.
Nevertheless, this trend could be explained by a selection bias in our sample. All the O dwarfs in this work have masses systematically above 15~M$_{\odot}$ (see indication in
Fig.~\ref{fig:mass_discrep}). As a consequence, all stars with a negative mass discrepancy and masses below 15~M$_{\odot}$ are not present. To investigate this effect and reach a clearer conclusion
about the mass discrepancy in the VFTS sample, results from the ongoing quantitative analysis of the early B-dwarfs should be included.
\begin{figure}[h!!!]
\begin{center}
\includegraphics[scale=0.5,angle=90,trim=0mm 0mm 0mm 140mm,clip]{ratios_mass_discrep_v2.eps}
\caption{Logarithmic ratio of evolutionary masses from
\textsc{bonnsai} ($M_{\rm ev}(B)$) to spectroscopic masses ($M_{\rm
sp}$). Grey filled symbols correspond to stars analyzed with nitrogen lines and no indications of binarity or contamination in fibre. Open symbols
are those stars with
possible undetected binarity or contamination in fibre. Typical error bars are given in the upper-right corner. The dashed
line indicates points where $M_{\rm ev}(\rm B)$\,=\,15\,M$_{\odot}$. O2 stars are excluded.}
\label{fig:mass_discrep}
\end{center}
\end{figure}
\section{Wind properties}
\begin{figure}[h!!!]
\begin{center}
\includegraphics[scale=0.5,angle=90,trim=0mm 0mm 0mm 0mm,clip]{dmom.eps}
\caption{Wind-momentum\,--\,Luminosity Relationship (WLR) for our sample compared with theoretical predictions from \cite[Vink et al. (2001)]{vink01} for the metallicities of the Milky Way and the
Magellanic Clouds. Open symbols represent stars with upper limits on log\,$Q$ (hence $D_{\rm mom}$). Diamonds correspond to stars analyzed using nitrogen lines.
Typical uncertainties are shown in the
upper-left corner, and the vertical dashed line indicates log\,$L/L_{\odot}$\,=\,5.1.}
\label{fig:wind}
\end{center}
\end{figure}
In this study, we estimated the logarithm of the wind-strength $Q$-parameter (see Sect.~\ref{sec:qsa}) of the sample stars. If terminal velocities were available, log\,$Q$ could be used to estimate
mass-loss rates. Nevertheless, we lack UV spectroscopic data to derive directly terminal velocities. For this reason, we have followed the approach used by \cite[Mokiem et al.
(2007a,b)]{mokiem07a,mokiem07b} to estimate the escape velocity $v_{esc} = 2gR(1-\Gamma)^{1/2}$ and the terminal velocity $v_{\infty}/v_{esc} = 2.65\,Z^{0.13}$, where $\Gamma$ is the
Eddington factor and $Z$ the metallicity corresponding to the LMC. The wind momentum was calculated via $D_{\rm mom}$\,=\,$\dot{M}v_{\infty}R^{1/2}$, and is represented
in Figure~\ref{fig:wind} as a function of stellar luminosity.
Two different behaviours are present in the distribution:
\begin{itemize}
\item At luminosities higher than log\,$L/L_{\odot}$\,$\sim$\,5.1, winds were constrained for most of the stars. Most of them show consistency with the linear trend predicted by \cite[Vink et
al. (2001)]{vink01} within the error bars. However, a large dispersion is found, showing some stars above the theoretical WLR for Galactic metallicity and some others below the prediction for the
SMC. The
first case could be explained by unclumped wind models and large uncertainties on the estimated terminal velocities, but no suitable explanation can be found for the second one.
\item At luminosities lower than log\,$L/L_{\odot}$\,$\sim$\,5.1, only upper limits could be given for log\,$Q$ and therefore $D_{\rm mom}$. To constrain such thin
winds, UV spectroscopy is necessary.
\end{itemize}
|
\section{Introduction}
Normative texts, or ``contracts'', are documents which describe the permissions, obligations, and prohibitions of different parties over a set of actions.
They also include descriptions of the penalties which must be paid when the main norms of the document are violated.
We frequently encounter such texts in the form of
privacy policies,
software licenses,
workflow descriptions,
regulations, and
service-level agreements.
Despite being written for human consumption and thus expressed in natural language,
these kinds of documents are typically long and difficult to follow, making them hard to analyse manually.
\begin{quote}
\itshape
What commitments am I agreeing to make?\\
Can my information be shared with third parties?\\
Is it possible to avoid providing credit card details?
\end{quote}
These are the kinds of questions about a contract that we may want answered, both as users and as authors.
Using text-based search to find such answers can be tedious and unreliable, for example when clauses cross-reference each other and when exceptions are present.
Our goal is to bring formal methods to this kind of natural language analysis,
packaged as a tool which is usable by non-experts
and which requires as little understanding of the underlying technologies as possible.
We do this by first modelling these documents using a suitable formalism,
which then makes them amenable to verification using standard techniques.
This includes answering queries based on a syntactic traversal of the model,
as well as using model-checking to verify temporal properties,
by converting the model to a timed automata representation.
Each of the components required for this have been described individually in previous works by the current authors~\cite{CGS16conpar,CPS14cnl,CS17jlamp}.
This paper presents a new web-based tool for the analysis of normative texts in English,
bringing each of these components together into a single interface.
The rest of the paper continues as follows.
\autoref{sec:tool-intro} introduces the structure of the tool and gives an overview of the workflow it establishes.
In \autoref{sec:example} we then present the running example which will be used throughout the paper to demonstrate the use of our tool.
\autoref{sec:front-end} describes extraction and building a model through post-editing of tabular data,
including verbalisation of an existing model using controlled natural language.
\autoref{sec:back-end} describes the analysis which can be performed on the model, by using a set of query templates which can be customised based on the current model.
The software architecture is then summarised in \autoref{sec:architecture}.
\autoref{sec:related-work} takes a look at some related work in this area, and
we conclude with a summary of the benefits and limitations of our approach in \autoref{sec:conclusion}.
\section{The \textit{Contract Verifier}\ tool}
\label{sec:tool-intro}
The main contribution of this work is the \textit{Contract Verifier}, a web-based tool for modelling and analysing normative texts in English.
\autoref{fig:system-overview} shows an overview of the workflow which our tool covers,
summarised in the following steps:
\begin{enumerate}
\item Users start with an English text which they wish to build a model out of.
\item The text is submitted to an \textbf{extraction} phase which attempts to automatically extract the clauses from the text, each of which concerning at least an agent, action and modality.
\item The results of the extraction are shown to the user in a tabular format where each cell is editable. At this point, the user must check the results of the extraction and \textbf{post-edit} the clauses which are not completely correct.
\item Once the user is happy with the model in tabular format, this is \textbf{converted} into an actual model, which is internally represented in an XML format.
\item From here, the model can be verbalised in a controlled natural language and viewed in a compact formal notation.
\item The user then performs \textbf{analysis} by selecting the queries which should be run against the model. Queries are presented as English sentence templates, which may include slots for specifying relevant arguments.
\item The \textbf{queries} are then submitted to the server which computes the \textbf{results} and displays them to the user beside each respective query.
\item If the answer to a query is not as expected, this could point to either:
\begin{compactenum}
\item a problem with the contract model, in which case the user may go back and edit the model and repeat the steps as above; or
\item a problem with the original normative document, where the user may then modify the text and perform the analysis again, or simply leave the original formulation as it is (e.g., the analysis might detect a lack of a deadline associated with a given obligation, but this might be considered desirable by the user).
\end{compactenum}
\end{enumerate}
A demonstration of the \textit{Contract Verifier}\ is openly available on the web at \url{http://remu.grammaticalframework.org/contracts/verifier/}.
Upon visiting this URL, users will be asked to log in.
New users may create an account so that their work can be saved under their profile.
Alternatively, one may log in with a guest account, using username \verb|guest@demo| and password \verb|contract|.
\begin{figure}[t]
\includegraphics[width=\textwidth]{system-overview.pdf}
\caption{Overview of the contract analysis workflow established by the \textit{Contract Verifier}.}
\label{fig:system-overview}
\end{figure}
\section{Running example}
\label{sec:example}
To show how our tool is used in practice, we pick a running example of a normative text
describing the rules of a university course (see \autoref{fig:case-study-text}).
This example is based on courses held at our own department,
covering the requirements for passing the course and the deadlines for the submission and grading of assignments.
It has been chosen as an example because it is concise yet contains a variety of temporal constraints and dependencies between clauses.
The text itself has been written by ourselves.
Before publishing these rules as an official course description,
the authors (teachers/administrators) may wish to ensure that all the requirements for passing the course are enforced
and that the rules are consistent.
Once published, end users of these rules (students) may wish to query the parts which are relevant to them
or work out any flexibilities in their deadlines.
The \textit{Contract Verifier}\ tool aims to meet the needs of both of these groups.
\begin{figure}[t]%
\begin{framed}%
\setdefaultleftmargin{4mm}{}{}{}{}{}%
\begin{asparablank}
\item Students need to register for the course before the registration deadline, one week after the course has started.
\item To pass the course, a student must pass both the assignment and the exam.
\item The deadline for the first assignment submission is on day 10.
\item Graders should correct an assignment within one week of it being submitted.
\item If the submission is not accepted, the student will have until the final deadline on day 25 to re-submit it.
\item The exam will be held on day 60.
\item Registered students must sign up for the exam latest 2 weeks before it is held.
\end{asparablank}%
\end{framed}%
\vspace{-1em}%
\caption{%
Example of a normative text describing the rules involved in the running of a university course.
The course is assumed to start on day 0.
This text is used as the running example throughout this paper.%
}%
\label{fig:case-study-text}%
\end{figure}
\section{Building a contract model}
\label{sec:front-end}
\subsection{Extraction from English}
\label{sec:extraction}
The first step towards building a formal model of our natural language contract is to process the text in order to see what clause information can be extracted automatically.
We have built such a tool, named \textsf{ConPar}~\cite{CGS16conpar}, which can extract partial contract models from English normative texts.
\textsf{ConPar}{} is a natural language processing (NLP) tool, which uses the Stanford Parser~\cite{Klein03} to obtain dependency trees for each sentence in the input text.
These trees are then processed by \textsf{ConPar}{}, which uses the dependency representation to attempt to extract information related to:
\begin{compactenum}[(i)]
\item subject (agent)
\item verb and object (action)
\item modality (obligation, permission, prohibition)
\item temporal and non-temporal conditions.
\end{compactenum}
In addition, \textsf{ConPar}{} will also try to identify refinement clauses, where a single input sentence may translate to multiple sub-clauses joined together using a connective such as conjunction, choice or sequence.
The output of the \textsf{ConPar}{} tool is a tabular representation of the extracted data, which is produced in a tab-separated value (TSV) format.
This is a simple format which can be easily loaded in any spreadsheet or table-editing software.
In this representation, each row corresponds to a clause while the columns indicate the various components, as listed above.
The result of passing our example through this tool is shown in \autoref{fig:conpar-output}.
\begin{figure}[t!]
\includegraphics[width=\textwidth]{tsv.png}
\caption{%
Screenshot showing the output from passing our normative text through the extraction tool
(image has been cropped to improve readability).
Each row indicates a clause in the model.
Note how the second English sentence has been refined into multiple sub-clauses.
The O in the modalities column stands for \emph{obligation}, while D stands for \emph{declaration}.
}
\label{fig:conpar-output}
\end{figure}
While quite a lot of clause information has been correctly extracted by the tool, there are still some errors which need to corrected manually by the user.
The use of a tabular format for displaying the output of the extraction phase facilitates this post-editing.
All cells in the table are editable, and rows can be added and deleted as needed.
While a little knowledge of this tabular format is required by the user, we nonetheless believe it is an effective way of preparing the extracted information before the model itself is created in the next step.
\subsection{Conversion to contract model}
\label{sec:conversion}
After post-editing the extracted clause information, this can be converted into a formal contract model.
The formalism used for representing contracts is based on the deontic modalities of \textbf{obligation}, \textbf{permission} and \textbf{prohibition} of agents over actions.
It includes constructs for refining clauses by conjunction, choice and sequence, and includes the possibility to specify reparations for when a clause is violated.
Clauses can be constrained by temporal restrictions or guarded based on the status of other clauses.
This formalism, which is based on \textit{C-O Diagrams}~\cite{MCD+10mvs}, is described fully in~\cite{CS17jlamp}.
This conversion step is implemented as a straight-forward script which takes the TSV representation as input and produces a contract model file.
We refer to the format of this file as COML, which is an XML-based format for storing contract models in our formalism.
Once the conversion is complete, the COML file is stored on the server and the user can view it using three different representations simultaneously:
\begin{compactenum}[(i)]
\item post-edited text,
\item controlled natural language (CNL), and
\item a compact formal notation (\textit{C-O Diagram}{} Shorthand or CODSH).
\end{compactenum}
These are shown in \autoref{fig:model-representations} and explained in the sections below.
\begin{figure}[t!]
\includegraphics[width=\textwidth]{model-representations.png}
\caption{%
Screenshot showing the different output representations of the contract model
(image has been cropped to improve readability).
Top: the post-edited input text;
bottom-left: controlled natural language;
bottom-right: compact formal notation.}
\label{fig:model-representations}
\end{figure}
\subsection{Post-edited text}
\label{sec:post-editing}
This text corresponds to the initial column from the tabular view in the extraction phase.
While this originally comes from the input text, the individual sentences may be edited during the post-editing phase.
This means that the text displayed here is not necessarily the same as the original input text.
\subsection{Verbalisation using CNL}
\label{sec:cnl}
Given a contract model in our formalism, we have developed a method for linearising it as a phrase in a \emph{controlled natural language} (CNL)~\cite{wyner+09cnl}.
A CNL is a reduced version of a natural language (NL) which has limited syntax and vocabulary, making it in fact a formal language and thus expressible using a grammar.
CNLs are often used as interfaces for formal languages which are human-friendly, yet still unambiguous and well-defined.
The CNL designed for our contract formalism is decribed in previous work~\cite{CPS14cnl}.
We use the Grammatical Framework (GF)~\cite{Ranta2011gf} for defining the grammar for our CNL and converting to and from our internal formal representation.
This also includes the possibility of building a contract model directly using the CNL (rather than using the extraction step), however we do not cover this input method in the present work.
The CNL representation may resemble the original NL text in some ways, however it is characterised by less variation in the expressions used and by certain structural features such as labels before each clause.
The generation of the CNL representation requires that subjects, verbs and objects are present in the lexicon.
The lexicon used is a large-scale English dictionary containing over 64,000 entries,
but if the contract contains terms which are not present in this lexicon then the generation to CNL will fail.
This failure however will not affect the usability of the rest of the tool.
\subsection{Compact formal notation}
\label{sec:codsh}
In addition to the post-edited text and the CNL, we also display a view of the model in a formal syntax.
We refer to this notation as \textit{C-O Diagram}{} Shorthand (CODSH).
It is designed mostly for developers who understand the formal structure of the contract model but would like a more condensed representation than the COML format.
This can be helpful when debugging.
In particular, this notation reveals the names which are automatically assigned to each clause in the conversion phase.
\section{Analysis}
\label{sec:back-end}
Once the model has been built, we can perform analysis on the contract by running queries against it.
The user is presented with a list of query templates, as shown in \autoref{fig:queries}.
Each query may have slots for parameters which the user should provide; these are either names of clauses, agents or actions.
The possible completions for these slots are extracted automatically from the contract model.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{queries.png}
\caption{Available query templates which users may execute against the contract model,
with drop-downs for specifying agents and actions extracted from the model.
Queries 1--6 are \emph{syntactic}, while
queries 7--10 are \emph{semantic}.}
\label{fig:queries}
\end{figure}
Internally, queries provided by our tool are computed in one of two ways,
either through syntactic filtering or via conversion to timed automata and using the \textsc{Uppaal}{} verification tool.
Both of these techniques are described below.
\subsection{Syntactic}
\label{sec:syntactic}
Our tool currently provides six different syntactic queries
(items 1--6 in \autoref{fig:queries}).
These queries are \emph{syntactic} in the sense that each can be solved by traversing the contract model and filtering out the corresponding clauses which match the query.
As an example, consider the following query:
\begin{quote}\itshape
What are the obligations of \emph{[agent]}?
\end{quote}
This query is internally encoded as the following conjunction of predicates:
\begin{quote}
$\mathit{isObl} \land \mathit{agentOf}([agent])$
\end{quote}
where $[agent]$ is replaced with a concrete agent name chosen by the user.
The solution to this query is computed via a Haskell function which takes the contract model as a Haskell term together with the query as a set of predicate functions, and returns a list of matching clauses.
The final output is given as a natural language sentence listing the actions corresponding to the matching clauses.
For the example, asking for all the obligations for \emph{student} will give the output below:
\begin{quote}\itshape
The following are obligations of student:
\begin{compactitem}
\item register for course
\item submit assignment
\item sign up for exam
\item pass exam
\end{compactitem}
\end{quote}
The way the result is phrased will vary based on the query, as well as the number of items in the result:
up to two results are inlined, while three or more are given as bullets.
This is to make the response more natural for the user.
\subsection{Semantic}
\label{sec:semantic}
Our tool also includes four semantic queries
(items 7--10 in \autoref{fig:queries}).
We use the term \emph{semantic} to refer to those queries which cannot be answered simply by looking at the structure of the model.
Consider the following example:
\begin{quote}\itshape
The \emph{[agent]} must \emph{[action]} before time \emph{[number]}.
\end{quote}
Determining this must take into consideration the operational behaviour of a contract model,
including when actions are performed, how new clauses are enabled and others expire.
Processing semantic queries is achieved through using model checking techniques,
by first converting a contract model into a network of timed automata~\cite{AD94tta}
and then using the \textsc{Uppaal}{} tool~\cite{LPY97uppaal} to verify temporal properties against the translated model.
This idea was introduced for \textit{C-O Diagrams}{} in~\cite{DCM+14svn}; the details of our own translation can be found in~\cite{CS17jlamp}.
By using verification, we are able to quantify over all possible sequences of events with respect to the contract.
This approach requires that the query itself is encoded as a property in a temporal logic which the model checker can process.
In the case of \textsc{Uppaal}{}, the property specification language is a subset of TCTL~\cite{BDL06uppaal}.
The example query above is encoded as the following \textsc{Uppaal}{} property:
\begin{align*}
\TCTLalways &\mathit{allComplete}() \implies\\
&\big(\mathit{isDone}([agent]\_[action]) \land t_0 - Clocks\big[[agent]\_[action]\big] < [number]\big)
\end{align*}
Here, $\mathit{allComplete}$ and $\mathit{isDone}$ are helper functions included in the translated \textsc{Uppaal}{} system
which allow the status of clauses and actions to be queried from within the timed properties.
$t_0$ is a never-reset clock representing global time, while the $Clocks$ array contains a clock for each action in the system,
which is reset when that action is performed.
The expression $t_0 - Clocks\big[a\big]$ thus gives the absolute time at which action $a$ was completed.
The property is then verified using \textsc{Uppaal}, which will return a result of \textbf{satisfied} or \textbf{not satisfied}.
In cases where a symbolic trace is produced as part of the verification, this is parsed by our tool in order to provide a meaningful abstraction of it.
In this processing step, we pick out the actions performed in the trace along with their time stamps and present these as part of the result to the user.
For example, when running the query below on our contract example:
\begin{quote}\itshape
The \emph{student} must \emph{register for course} before time \emph{5}.
\end{quote}
we get the following result in the tool:
\begin{quote}\itshape
NOT Satisfied\\
The property is violated by the following action sequence:\\
- student register for course at time 6\\
- student submit assignment at time 6\\
- ...
\end{quote}
where the remainder of the trace contains the other obliged actions at time 6 or later.
This is in fact as we expect; the contract states that students have up to~7~days to register for the course,
and thus it is not true that they must have registered before day 5 in order to satisfy the entire contract.
If we change the time value in the query from 5 to 7, then the result returned is \emph{Satisfied}.
\section{Software architecture}
\label{sec:architecture}
The \textit{Contract Verifier}\ tool is implemented as a PHP web application, using a MySQL database for storing user accounts and the query templates available in the system.
Contract models in COML and \textsc{Uppaal}-XML format are saved as files on the server.
No server-side framework is used.
The client-side interface is based on the AdminLTE Control Panel Template\footnote{\url{https://almsaeedstudio.com/}},
and the tabular editing interface makes use of the editableTable jQuery library\footnote{\url{http://mindmup.github.io/editable-table/}}.
The \conpar{} extraction tool is written in Java, primarily because it uses the Stanford parser which is also implemented in Java.
The core of our system is written in Haskell,
using algebraic data types to define the structure of a contract model.
The conversion from TSV and translation to NTA are thus also written as Haskell functions to and from this data type.
The linearisation of a contract model to CNL uses the GF runtime, which is a standalone application.
Similarly, executing semantic queries requires running \textsc{Uppaal}{} as an external process.
Because of the variety of languages and programs used in our tool chain,
we provide a convenient layer over these components in the form of a small server application which provides these separate functionalities as individual web services.
This modular approach allows the web application providing the user interface to consume each component via a web API,
removing limitations on implementation language and hosting requirements, and allowing a clean separation of concerns between front-end and back-end.
\autoref{tab:web-service-api} shows a summary of the API covering all the web services provided by the server.
This API is fully documented and publicly accessible at the following URL: \url{http://remu.grammaticalframework.org:5446/}.
The server itself is also implemented in Haskell.
\begin{table}[t]
\caption{API of services provided in the web server, showing the relevant URL path and input/output formats for each service.
The various formats are: TSV (tab separated values),
COML (Contract-Oriented XML),
CODSH (\textit{C-O Diagram}\ Shorthand),
CNL (Controlled Natural Language),
and \textsc{Uppaal}-compatible XML.}
\label{tab:web-service-api}
\centering
\bgroup
\def1.1{1.1}
\setlength\tabcolsep{0.35em}
\begin{tabular}{llll}
\textbf{Path} & \textbf{Description} & \textbf{Request} & \textbf{Response} \\
\hline
\texttt{/nl/tsv} & Clause extraction (\conpar) & English text & TSV \\
\texttt{/tsv/coml} & Convert TSV to COML & TSV & COML \\
\texttt{/coml/codsh} & Show contract in shorthand & COML & CODSH \\
\texttt{/coml/cnl} & Verbalise contract using CNL & COML & CNL \\
\texttt{/coml/syntactic} & Execute syntactic query & Query + COML & Clause names \\
\texttt{/coml/uppaal} & Translate contract to NTA & COML & UPPAAL XML \\
\end{tabular}
\egroup
\end{table}
\section{Related work}
\label{sec:related-work}
AnaCon~\cite{ACS13fca} is a similar framework for the analysis of contracts, based on the contract logic \ensuremath{\mathcal{CL}}~\cite{PS07fle,PS12ddl},
which allows for the detection of contradictory clauses in normative texts using the CLAN tool~\cite{FPS09clan}.
By comparison, the underlying logical formalism we use, based on \textit{C-O Diagrams}, is more expressive than \ensuremath{\mathcal{CL}}\ as it includes temporal constraints, cross-referencing of clauses and more.
Besides this, our translation into \textsc{Uppaal}{} allows for checking more general properties, not only normative conflicts.
In addition, their interface for specifying contracts is purely CNL-based and there is no extraction tool from English as in our case.
Information extraction from natural language is of course a field within itself,
and even in the domain of contractual documents in English there are many other works with similar goals.
The extraction task in our work can be seen as similar to that of Wyner \& Peters~\cite{WP11rer},
who present a system for identifying and extracting rules from legal texts using the Stanford parser and other NLP tools within the GATE system.
Their approach is somewhat more general, producing as output an annotated version of the original text,
whereas we are targeting a specific, well-defined output format.
Other similar works include that of Cheng et al.~\cite{cheng+09ield},
who combine surface-level methods like tagging and named entity recognition (NER) with hand-crafted semantic analysis rules,
and Mercatali et al.~\cite{mercatali+05uml} who extract the hierarchical structure of the documents into a UML-based format using shallow syntactic chunks.
One crucial aspect in any work targeting formal analysis of natural language documents is the confidence in the extraction from a source document to the target formal language.
In our tool, the result of the extraction is usually incomplete and some amount of manual post-editing is always generally required.
Azzopardi et al.~\cite{AGP16ric} handle this incompleteness using a deontic-based logic including {\em unknowns}, representing the fact that some parts have not been fully parsed.
Furthermore, the same authors present in~\cite{AGP16inl} a tool to support legal document drafting in the form of a plugin for Microsoft Word.
Though the final objective of their work diverges from ours,
we are both concerned with the translation of natural language documents into a formal language by using an intermediate CNL.
The main difference is that they target a more abstract formal language (very much like \ensuremath{\mathcal{CL}}), and as a consequence their CNL is also different.
Our formalism allows for richer representations not present in their language (e.g.~real time constraints and cross-references).
Additionally,
they do not target complex analysis of contracts (they only provide a normative conflict resolution algorithm), and
we do not provide assistance in the contract drafting process.
There is considerable work in modelling normative documents using representations other than logic-like formalisms such as our own.
LegalRuleML \cite{athan+15legalruleml} is a rule interchange format for the legal domain,
allowing the contents of legal texts to be structured in a machine-readable format.
The format aims to enable modelling and reasoning, allowing users to evaluate and compare legal arguments using tools customised for this format.
A similar project with a broader scope is MetaLex \cite{boer+08metalex}, an open XML interchange format for legal and legislative resources.
Its goal is to enable public administrations to link legal information between various levels of authority and different countries and languages, improving transparency and accessibility of legal content.
Semantics of Business Vocabulary and Business Rules (SBVR) \cite{SVBR1.3} uses a CNL to provide a fixed vocabulary and syntactic rules for expressing the terminology, facts, and rules of business documents of various kinds.
This allows the structure and operational controls of a business to have natural and accessible descriptions,
while still being representable in predicate logic and convertible to machine-executable form.
We note that none of the works mentioned above present a single tool for end-to-end document analysis,
starting from a natural language text and finally allowing for rich syntactic and semantic queries, as in the case of our tool.
\section{Conclusion}
\label{sec:conclusion}
In this paper we have presented \textit{Contract Verifier}, a web-based tool for analysing normative documents written in English.
The tool brings together a number of different components
packaged together as a user-friendly application.
We demonstrate a typical workflow through the system, starting with an English text, extracting a contract model from it, and executing different kinds of queries against it.
Each of the components used by our tool is implemented as a standalone module, with a web-based API exposing each module as a web service.
This allows individual modules to be replaced, or new interfaces to be built, without having to make changes to the entire system.
An important feature of the \textit{Contract Verifier}\ tool is the level of automation it provides: everything except the post-editing of the extracted model (and of course choosing the queries to be performed) is automatic.
For example, the names of clauses, agents and actions are automatically extracted from the contract so the user can select them using drop-down menus when making queries.
Also, each clause
is given a unique identifier as well as a {\em clock} that is reset when that clause is activated.
Though these are mainly intended for internal use when performing semantic queries,
users may even use them explicitly in the post-editing phase to encode relative timing constraints.
We see this tool as a successful implementation of a user interface for bringing together various separate components and providing a clear workflow for analysing normative texts in English.
That being said, it is as such a proof-of-concept tool and has not undergone any extensive usability testing or application in real-world scenarios.
We conclude here with a critical look at the shortcomings and limitations of the current work.
\subsection*{Evaluation}
Our goal is to make the task of analysing a normative document easier and more reliable than if one were to do it completely manually.
Measuring whether we achieve this, and to what extent, requires proper assessment.
Some evaluation has already been carried out for the natural language extraction part of the workflow (the \textsf{ConPar}{} tool),
measuring the accuracy of tool for extracting a correct model from a normative document.
By calculating precision and recall, $F_1$ scores of 0.49 to 0.86 were obtained for the test set of 4 documents~\cite{CGS16conpar}.
However, we currently do not have a thorough evaluation of the complete \textit{Contract Verifier}\ workflow as presented here.
This would take the form of an empirical study comparing document analysis using our tool with a purely manual approach, measuring
the amount of post-editing required to build a correct model,
the time required to formulate a query and obtain a result, and
the overall ease of use of the tool in a qualitative sense.
While we consider a study of this kind important future work.
\subsection*{Limitations}
\paragraph{Extraction}
The extraction phase relies on dependency trees and thus takes a syntactic-level approach to parsing.
While a fair deal of information can be extracted in this way from simpler sentences,
a deeper understanding of a phrase often involves using related or opposite concepts which cannot be determined without more elaborate processing on the semantic level.
In addition, we assume that each input sentence translates into one or more clauses, and have no support for detecting when a phrase should actually modify an existing clause instead.
\paragraph{Modelling}
Our tool uses a tabular interface to help make the task of modelling user-friendly,
but some understanding of the underlying formal language and its semantics is necessary in order to work efficiently with it.
For example,
our formalism is essentially \emph{action-based}, where clauses prescribe what an agent should or should not \emph{do}.
However, empirically we have found that normative documents often describe what should or should not \emph{be}, i.e. referring to state-of-affairs.
While these can often be paraphrased to fit into our formalism, this is a non-trivial task which currently must be done completely by the user.
\paragraph{Verification}
When it comes to running queries, those which are syntactic can quickly be answered by an algorithm which is linear in the size of the model.
For semantic queries however, the conversion to timed automata means that the state-space explosion problem typical of model checking is a potential problem.
Certain optimisations made during the translation process could improve this somewhat.
For instance, our generated NTA contain many parallel synchronising automata as a result of the modularity of our translation,
and some {\em ad hoc} heuristics could likely be used to reduce the number of automata or the need for certain synchronisations,
thus improving the performance.
That said, this is ultimately a theoretical problem which we cannot avoid altogether.
\subsection*{Scalability}
\paragraph{Extraction}
We are limited here by the speed of the \conpar\ tool, which itself uses the Stanford dependency parser.
While parse time is related to the length of the input sentence, in our tests based on the test data from~\cite{CGS16conpar}
we have found that parsing and extracting a single sentence takes on average roughly half a second.%
\footnote{\label{fn:test-machine} Tests carried out on a dual-core MacBook Air from 2013.}
\paragraph{Post-editing}
The tabular interface for editing the extracted clauses has no concrete limits in terms of the number of clauses it can handle.
What it does lack is support for managing a document's internal hierarchy (sections and sub-sections), which is a common feature of normative texts.
\paragraph{Analysis}
As discussed above, the running of semantic queries is the biggest barrier to the scalability of the system due to the state-space explosion problem.
For our small example here, a query requiring a search of the entire search space requires a few milliseconds to complete,
but this performance may degrade drastically as the model size increases.
\paragraph{Queries}
Our current implementation only offers a limited number of syntactic and semantic queries.
New query templates can easily be added without any theoretical constraints.
However, the user interface may need to be updated to help users navigate and possibly search through a long list of queries.
\\[12pt]
In summary, the \textit{Contract Verifier}\ tool in its current state can handle documents of essentially any size in what concerns the extraction of a formal representation from a natural language text, its post-editing, and the execution of syntactic queries.
This however could be improved by adding an extra layer of document hierarchy management to the tool,
allowing the task to be segmented into smaller sub-parts.
In what concerns semantic queries, the tool can realistically only handle smaller individual contracts containing tens of clauses.
This is essentially due to our choice to translate contract models into \textsc{Uppaal}~timed automata.
An alternative here could be to use SAT solvers, or some other verification technology, to process the kind of semantic queries we perform.
In any case, we believe \textit{Contract Verifier}\ is an important step towards a rich analysis of normative texts, even if the analysis were to be restricted to just syntactic queries.
\bibliographystyle{splncs03}
|
\section{Supplementary}
\subsection{Hyperparameter space
}
\label{sec:space}
The space of hyperparameters (mostly, learning rates for different parts of the network) we explored in those runs was built as follows:
\begin{enumerate}[label={\arabic*)}]
\item on SVHN-MNIST we found reasonably working vectors of hyper-parameters for each method essentially by cross-validation on test
\item then we combined all these vectors into a set $H$ and then inflated $H$ by adding many combinations of parts from these vectors and other similar and in-between vectors; for example, if for two models two optimal sets of parameters included (1e-6, 1e-3) for the first model and (1-4, 1e-1) for the second one, we would inflate H by adding all nine combinations of ([1e-6, 1e-5 1e-4], [1e-3, 1e-2, 1e-1]) and varying model-specific parameters such as regularization parameters
\item then we conducted experiments with combinations of pairs [vector from $H$, model]
\end{enumerate}
\subsection{Implementation details}
\label{sec:supl}
During domain adaptation experiments, discriminators of all models that have a trainable discriminator (ADDA, WGAN, Dual) were trained for five epochs prior to the beginning of actual adaptation, which improved performance for all models. Kernel density estimates on Figure \ref{fig:real_acc_dist} were obtained with a fixed $0.03$ bandwidth and a linear boundary correction method. All algorithms took greyscale images resized to $28\times28$ as input. Restrictions on $\alpha$ were imposed by adding an error term $+\lambda_1 | \sum \alpha_A - \sum \alpha_B |$ and positivity by $-\lambda_2|\min(0, \alpha)|$. Hyperparameters for the Dual method can be set so that restrictions above hold by the end of adaptation. To implement the dual algorithm in mini-batch fashion we extracted slices of $\alpha_*$ that correspond to points processed in this batch on each iteration. Search space was built as explained above, learning rates and weight decay for updating target network and discriminator were as follows:
\begin{itemize}
\item for target network update we used Adam with lr in [1e-6, 1e-5, 1e-4] and beta in [0.5, 0.9]
\item for discriminator updates we used SGD with lr in [1e-6, 1e-5, 1e-7] or Adam with learning rates in [2e-4, 1e-6, 1e-5] and beta in [0.5, 0.9]
\item discriminator weight decay parameter was chosen from [2.5e-05, 0]
\end{itemize}
resulting into 90 possible combinations + model specific parameters.
\subsection{Validation heuristics}
Considering runs such that the following holds for different values of $\lambda$ did not change results significantly.
\[\frac{d(P, Q^{(0)}) - d(P, Q^{(T)})}{\max_t d(P, Q^{(t)}) - \min_t d(P, Q^{(t)})} >\lambda.\]
\label{sec:valdd}
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth, trim={5cm 0 4.5cm 0}, clip]{figs/figure_1-5.pdf} \\
\end{figure}
\section{Conclusion}
We presented an adversarial objective that does not lead to a min-max problem. We proposed using the dual of the discriminator objective to improve the stability of distribution alignment, showed its connection to MMD, and presented quantitative results for alignment of toy datasets and unsupervised domain adaptation results on real-image classification datasets. Our results suggest that the proposed dual optimization objective is indeed better suited for learning with gradient descent than the saddle point objective that naturally arises from the original primal formulation of adversarial alignment. Further attempts to use duality to reformulate other notions of statistical distances in adversarial settings as computationally feasible minimization problems may be promising.
\section{Acknowledgments}
This work was supported by NSF Award \#1723379 and NSF CAREER Award \#1559558.
\section{Discussion and Future Work}
While exact dual exists only in the logistic discriminator case, when one can use the duality and solve the inner problem in closed form, we want to stress that our paper presents a more general framework for alignment that can be extended to other classes of functions. More specifically, one can rewrite the quadratic form in kernel logistic regression \eqref{eq:dual_d_q} as a Frobenius inner product of a kernel matrix $Q$ with a symmetric rank 1 alignment matrix $S$ (outer product of alpha with itself).
\begin{equation*}
d(A, B) = \min_{0\leq \alpha_i \leq 1} ~ \frac{1}{2\lambda} \langle \alpha \alpha^T, Q \rangle_F + H(\alpha) = \min_{S_\alpha} ~ \langle S_\alpha, Q \rangle_F + H(S_\alpha)
\end{equation*}
The kernel matrix specifies distances between points and $S$ chooses pairs that minimize the total distance. This way the problem reduces to “maximizing the maximum agreement between the alignment and similarity matrices” that in turn might be seen as replacing “adversity” in the original problem with “cooperation” in the dual maximization problem. In our paper, S is a rank 1 matrix, but we could choose different alignment matrix parameterizations and a corresponding regularizer that would correspond to having a neural network discriminator in the adversarial problem or a Wasserstein distance in Earth Mover's Distance form. The resulting problem is not dual to minimization of any existing adversarial distances, but exploits same underlying principle of “iteratively-reweighted alignment matrix fitting” discussed in this paper.
We assert that in order to understand the basic properties of the resulting formulation, an in-depth discussion of the well-studied logistic case is no less important than the discussion involving complicated deep models, which deserves a paper of its own. This paper proposes a more stable “cooperative” problem reformulation rather than a new adversarial objective as many recent papers do.
\section{Experiments}
\textbf{Synthetic Distribution Matching}
We first test the performance of our proposed approach on a synthetic point cloud matching problem. The data consists of two clouds of points on a two-dimensional plane and the goal is to match points from one cloud with points from the other. There are no restrictions on the transformation of the target point cloud, so $M_{\theta}$ includes all possible transformations, and is therefore parameterized by the point coordinates themselves, so the coordinates themselves were updated on each gradient step. We minimized the logistic adversarial distance in primal space by solving the corresponding min-max problem in Eq~\eqref{eq:primal_adv_problem} and compared this to maximizing the proposed negative adversarial distance given by the dual of the logistic classifier in Eq~\eqref{eq:dual_d_q} and the corresponding kernelized logistic classifier with a Gaussian kernel.
As expected, the optimization of distances given by the dual versions of domain classifiers (linear and kernelized) worked considerably better than the same distance given by a linear classifier in the primal form. More specifically, the results in the primal case were very sensitive to the choice of learning rate. In general, the resulting decent iterations for the saddle point problem did not converge to a single solution, whereas both dual versions successfully converged to solutions that matched the two clouds of points both visually and in terms of means and covariances, as presented in \figref{fig:synth_behaviour}.
\begin{figure}[t]
\centering
\vspace{-15pt}
\def0.23\textwidth{0.18\textwidth}
\setlength{\tabcolsep}{2pt}
\renewcommand{\arraystretch}{1.5}
\setlength{\belowcaptionskip}{-8pt}
\def0.22{0.22}
\begin{tabular}{cccc}
& Objective & Discriminator Accuracy & $||\hat \Sigma_A(t) - \hat \Sigma_B ||^2$
\\
\begin{turn}{90}\quad \quad Primal \end{turn} & \includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/primal/negloss.png} &
\includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/primal/acc.png} &
\includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/primal/cov_dist.png}
\\
\begin{turn}{90}\quad \quad Dual \end{turn}
& \includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/dual_linear/negloss.png} &
\includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/dual_linear/acc.png} &
\includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/dual_linear/cov_dist.png}
\\
\begin{turn}{90} \ \ \ Dual Kernel \end{turn}
& \includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/dual_kern/negloss.png} &
\includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/dual_kern/acc.png} &
\includegraphics[width=0.22\textwidth, trim={0 0 0px 3px}, clip]{figs/synt_cropped/dual_kern/cov_dist.png}
\\
& Iteration & Iteration & Iteration
\end{tabular}
\caption{ Convergence analysis for the synthetic point alignment problem. \textbf{Top row:} The procedure of adversarial alignment in the primal discriminator weight space for points from $A$ and $B$ lying in a two-dimensional plane does not converge to a single solution and oscillates over time. \textbf{Middle~row:} In contrast, the proposed dual minimization adversarial objective steadily converges to an optimum. \textbf{Bottom row:} We observe similar behavior for the kernelized version. Even though the accuracy of the learned discriminators (column two) drifts over time in all cases, the distance between the covariance matrices (column three) and means (not shown in the figure) decreased in all cases; however, in both cases marked as "dual" iterative procedure converged to a single stable solution that indeed corresponds to the desired point cloud alignment. See Figure~\ref{fig:synth_frames} for an illustration. Note that absolute values of objectives are not comparable, and therefore not shown on plots above.
}
\label{fig:synth_behaviour}
\end{figure}
We suggest one more intuitive explanation of why the dual procedure might work better, in addition to the fact that optimization problems are just inherently easier than saddle point problems. The descision boundary of the classifier in the dual space is defined implicitly through a weighted average of observed data points, so when these data points move, the decision boundary moves with them. If points move too rapidly and the discriminator explicitly parametrizes the decision boundary, weights of the discriminator may change drastically to keep up with moved points, leading to the overall instability of the training procedure. In support of this hypothesis, we observed interesting patterns in the behavior of the linear primal discriminator: when point clouds become sufficiently aligned, the decision boundary starts "spinning" around these clouds, slightly pushing them in corresponding directions. In contrast, both dual classifiers end up gradually converging to solutions that assigned each point with 0.5 probability of corresponding to either of two domains.
\begin{figure}[t]
\centering
\vspace{-15pt}
\def0.23\textwidth{0.23\textwidth}
\setlength{\tabcolsep}{2pt}
\renewcommand{\arraystretch}{1}
\setlength{\belowcaptionskip}{-4pt}
\begin{tabular}{ccccc}
\begin{turn}{90}\quad \ \ \ Primal \end{turn}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/primal/frame-0.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/primal/frame-94.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/primal/frame-199.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/primal/frame-163.png}
\\
\begin{turn}{90}\quad \quad Dual \end{turn}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_linear/frame-1.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_linear/frame-3.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_linear/frame-16.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_linear/frame-139.png}
\\
\begin{turn}{90}\ \ Dual Kernel \end{turn}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_kern/frame-0.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_kern/frame-24.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_kern/frame-204.png}
& \includegraphics[width=0.23\textwidth]{figs/synt_cropped/dual_kern/frame-284.png}
\\
& $t_0$ & \multicolumn{2}{c}{$\dots$ } & $t_T$
\end{tabular}
\caption{ (Best viewed in color) When trained on a point cloud matching task, the primal approach leads to an unstable solution that makes the decision boundary \textit{spin} around data points when they are almost aligned, whereas both the linear and kernel dual approaches lead to stable solutions that gradually assign 0.5 probability of belonging to either $A$ or $B$ to all points, which is exactly the desired behaviour. Yellow and blue points are the original point clouds, red points correspond to positions of yellow points after transformation $M_{\theta}$.}
\label{fig:synth_frames}
\end{figure}
\section{Introduction}
Adversarial methods have recently become a popular choice for learning distributions of high-dimensional data.
The key idea is to learn a parametric representation of a distribution by aligning it with the empirical distribution of interest according to a distance given by a discriminative model. At the same time, the discriminative model is trained to differentiate between true and artificially obtained samples.
Generative Adversarial Networks (GANs) that use neural networks to both discriminate samples and parameterize a learned distribution, have achieved particularly impressive results in many applications such as generative modeling of images \cite{goodfellow_nips14,denton2015deep,radford2015unsupervised}, image super-resolution~\cite{LedigTHCATTWS16}, and image-to-image translation \cite{isola2016image}. Adversarial matching of empirical distributions has also shown promise for aligning train and test data distributions n scenarios involving domain shift \cite{ganin_icml15,tzeng_iccv15,tzeng_cvpr17}.
However, GANs and related models have proved to be extremely difficult to optimize. It has been widely reported that training GANs is a tricky process that often diverges and requires very careful parameter initialization and tuning. Arjovsky and Bottou~\cite{arjovsky2017towards} have recently identified several theoretical problems with loss functions used by GANs, and have analyzed how they contribute to instability and saturation during training.
In this paper, we focus on one of the major barriers to stable optimization of adversarial methods, namely their min-max nature. Adversarial methods seek to match the generated and real distributions by minimizing some notion of statistical distance between the two, which is often defined as a maximal difference between values of certain test (\textit{witness}) functions that could differentiate these distributions. More specifically in the case of GANs, the distance is usually considered to be equal to the likelihood of the best neural network classifier that discriminates between the distributions, assigning "real or generated?" labels to the input points. This way, in order to align these distributions one has to minimize this maximum likelihood w.r.t. the parameters of the learned or aligned distribution.
Unfortunately, solving min-max problems using gradient descent is inherently very difficult.
Below we use a simple example to demonstrate that different flavors of gradient descent are very unstable when it comes to solving problems of this kind.
To address this issue, we explore the possibility of replacing the maximization part of the adversarial alignment problem with a dual minimization problem for linear and kernelized linear discriminators. The resulting dual problem turns out to be much easier to solve via gradient descent. Moreover, we make connections between our formulation and existing objectives such as the Maximum Mean Discrepancy (MMD) \cite{Gretton2012}. We show that it is strongly related to the iteratively reweighted empirical estimator of MMD.
We first evaluate how well our dual method can handle a point alignment problem on a low-dimensional synthetic dataset. Then, we compare its performance with the analogous primal method on a real-image
domain adaptation problem using the
Street View House Numbers (SVHN) and MNIST domain adaptation dataset pair.
Here the goal is to align the feature distributions produced by the network on the two datasets so that a classifier trained to label digits on SVHN does not loose accuracy on MNIST due to the domain shift.
In both cases, we show that our proposed dual formulation of the adversarial distance
often shows improvement over time, whereas using the primal formulation results in drifting objective values and often does not converge to a solution.
Our contributions can be summarized as follows:
\begin{itemize}
\vspace{-7px}
\item we explore the dual formulation of the adversarial alignment objective for linear and kernelized linear discriminators and how they relate to the Maximum Mean Discrepancy;
\vspace{-3px}
\item we demonstrate experimentally on both synthetic and real datasets that the resulting objective leads to more stable convergence and better alignment quality;
\vspace{-3px}
\item we apply this idea to a domain adaptation scenario and show that the stability of reaching high target classification accuracy is also positively impacted by the dual formulation.
\end{itemize}
\section{A Motivating Example}
\begin{wrapfigure}{R}{0.4\textwidth}
\begin{center}
\vspace{-10pt}
\includegraphics[width=0.4\textwidth, trim={2cm 1cm 2cm 2.2cm},clip]{figs/hyper_surface.png}
\end{center}
\caption{Gradient descent fails to solve the saddle point problem $\min_x\max_y xy$. Red line presents a trajectory of the gradient descent if vector field $g(x, y)$ is used at each iteration. Blue lines are examples of vectors from this vector field. \vspace{-1cm}}
\label{fig:xy}
\end{wrapfigure}
We start with a well-known \cite{goodfellow2016nips} motivating example of a simple min-max problem to show that, even in this basic case, gradient descent might fail dramatically.
Let us consider the simplest min-max problem with a unique solution: finding a saddle point of a hyperbolic surface. Given the function $f(x, y) = xy$, our problem is to solve $\min_x \max_y f(x, y)$, which has a unique solution at $(0,0)$. Suppose that we want to apply gradient descent to solve this problem.
The intuitive analog of the gradient vector that we might consider using in the update rule is defined by the vector field $g(x, y) = (x, -y)$. However, at any given point the vector $g(x, y)$ will be tangent to a closed circular trajectory, thus following this trajectory would never lead to the true solution $(0, 0)$.
One can observe the trajectory produced by the update rule $x_{t+1} = x_t + \alpha g(x, y)$ applied to the problem above in \figref{fig:xy}. Neither block coordinate descent nor various learning rate schedules can significantly improve the performance of the gradient descent on this problem.
We want to mention that while there is a huge body of work on using alternative descent schemes for convex-concave saddle point optimization, including, but not limited to different variants of mirror descent, such as Nesterov's dual averaging \cite{nesterov2009primal} and mirror prox \cite{nemirovski2004prox}, authors are not aware of any successful attempts to use it in the context of adversarial distribution alignment. Some techniques developed for solving continuous games such as fictitious play were successfully adopted by \citet{salimans_arxiv16}.
\section{Approach}
We first propose a new formulation of the adversarial objective for distribution alignment problems. Then we apply this approach to the domain adaptation scenario in Section~\ref{sec:da}.
Suppose that we are given a finite set of points $A$ sampled from the distribution $p$, and a finite set of points $B$ sampled from the distribution $q$, and our goal is to match $q$ with $p$ by aligning $B$ with $A$.
More specifically, we aim to learn a matching function $M_{\theta}(B)$ that maps $B$ to be as close as possible to $A$ by minimizing some empirical estimate of a statistical distance $d(\cdot, \cdot)$ between them where $\theta$ are parameters of the matching function: $\theta^* = \argmin_{\theta} d(A, M_{\theta}(B))$.
Let us denote $B'_{\theta}=M_{\theta}(B)$ or just $B'$ in contexts where dependence on $\theta$ is not important. The regular adversarial approach obtains the distance function by finding the best classifier $D_w(x)$ with parameters $w$ that discriminates points $x\in A$ from points $x\in B'$ and considers the distance between $A$ and $B'$ to be equal to the likelihood of this classifier. A higher likelihood of separating $A$ from $B'$ means that $A$ is far from $B'$. This can be any form of hypothesis in general, and is often chosen to be a linear classifier \cite{tzeng_arxiv15} or a multi-layer neural network \cite{goodfellow_nips14}. In this work, we use the class of linear classifiers, specifically, logistic regression in its primal and dual formulations. The solution can also be kernelized to obtain nonlinear discriminators.
We will define the distance between distributions to be equal to the maximum likelihood of the logistic classifier parametrized by $w$:
\begin{gather*}
d(A,B') = \max_{w} \sum_{x_i \in A} \log(\sigma (w^T x_i)) + \sum_{x_j \in B'} \log(1 - \sigma(w^T x_j)) - \frac{\lambda}{2} w^T w
\end{gather*}
We can equivalently re-write this expression as:
\begin{gather}
C_\theta = \{(x_i, y_i) \ : \ x_i \in A \cup B'_\theta \ , \ y_i = 1 \ \text{ if } \ x_i \in A \ \text{ else } \ \text{--}1 \} \nonumber \\
\label{eq:primal_adv_problem}
\min_{\theta} d(A,B'_{\theta}) = \min_{\theta} \max_{w, b} \sum_{x_i,y_i \in C_{\theta}} \log (\sigma(y_i (w^T x_i + b))) - \frac{\lambda}{2} w^T w
\end{gather}
The duality derivation \cite{jaakkola1999probabilistic, minka2003comparison} follows from the fact that the log-sigmoid has a sharp upper-bound
\begin{gather*}
\log(\sigma(u)) \leq \alpha^T u + H(\alpha) \ , \ \ \alpha_i \in [0,1] \\
H(\alpha) = \alpha^T\log\alpha + (1-\alpha)^T\log(1-\alpha)
\end{gather*}
thus we can upper-bound the distance as
\begin{gather*}
d(A,B'_{\theta}) = \min_{0 \leq \alpha \leq 1} \max_{w, b} \sum_{x_i,y_i \in C_{\theta}} \alpha_iy_i(w^Tx_i+b) + H(\alpha) - \frac{\lambda}{2} w^T w
\end{gather*}
where the dual variable $\alpha_i$ corresponds to the weight of the point. Higher weight means that the point is contributing more to the decision hyperplane. Optimal value of alpha attains this upper bound.
The $w$ that maximizes the inner expression can be computed in a closed form, ${w^* = \frac{1}{\lambda} (\sum_j x_j y_j \alpha_j)}$. Optimally of bias requires $\sum_i \alpha_i y_i = 0$. By substituting $w^*$ we obtain a minimization problem:
\begin{gather}
\label{eq:dual_d}
d(A,B'_{\theta}) = \min_{0\leq \alpha_i \leq 1} ~~\frac{1}{2\lambda} \sum_{ij} \alpha_i\alpha_j (y_i x_i)^T(y_j x_j) + H(\alpha) = \min_{0\leq \alpha_i \leq 1} ~ \frac{1}{2\lambda} \alpha^T Q \alpha + H(\alpha) = \\
\label{eq:dual_d_q}
= \min_{0\leq \alpha_i \leq 1 / \lambda} ~\frac{1}{2} \alpha_A^T Q_{AA}\alpha_A +
\frac{1}{2} \alpha_B^T Q_{BB}\alpha_B - \alpha_A^T Q_{AB} \alpha_B + H(\alpha_A) + H(\alpha_B) \\
\text{s.t. } || \alpha_A ||_1 = || \alpha_B ||_1 \nonumber
\end{gather}
The Eq. \eqref{eq:dual_d_q} is obtained by splitting the summation into blocks that include samples only from $A$, only from $B$, and from both $A$ and $B$. For example, matrix $Q_{AB}$ consists of pairwise similarities between points from $A$ and $B$, and is equal to the dot product between corresponding data points in the linear case. The factor of two in front of the cross term comes from the fact that off-diagonal blocks in the quadratic form are equal. The constraint on alpha sums comes from splitting optimality conditions on the bias into two term. We will denote the resulting objective as $d_D(\alpha, A, B)$.
The above expression gives us a tight upper bound on the likelihood of the discriminator. Thus, by minimizing this upper bound, we can minimize the likelihood itself, as in the original loss, and therefore minimize the distance between the distributions:
\begin{equation}
\theta^*,\alpha^* = \argmin_{\theta,\alpha \in \mathcal A} d_D(\alpha, A, M_{\theta}(B))
\end{equation}
Note that the overall problem has changed from an unconstrained saddle point problem to a smooth constrained minimization problem, which ultimately converges when gradient descent has a properly chosen learning rate, whereas the descent iterations for the saddle point problem are not guaranteed to converge at all.
The resulting smooth optimization problem consists of minimization over $\alpha$ to improve classification scores and over $\theta$ to move points towards the decision boundary. Next section provides more intuition behind the resulting iterative procedure.
\subsection{Relationship to MMD}
In this section, we show that our dual formulation of the adversarial objective has an interesting relationship to another popular alignment objective. The integral probability metric between distributions $p$ and $q$ with a given function family $\mathcal H$ is defined as
\begin{equation*}
d(p, q) = \sup_{f \in \mathcal H} \Big| {\mathbb E_{x \sim p} f(x) - \mathbb E_{x \sim q} f(x) \Big|}.
\end{equation*}
It was shown to have a closed form solution and a corresponding closed form empirical estimator if $\mathcal H$ is a unit ball in reproducing kernel Hilbert space with the reproducing kernel $k(x, y)$ and is commonly referred to as Maximum Mean Discrepancy~\cite{Gretton2012}:
\begin{gather*}
d(p, q) = \frac{1}{2} \mathbb E_{p \times p} k(x, x') + \frac{1}{2} \mathbb E_{q \times q} k(y, y') - \mathbb E_{p \times q} k(x, y)
\\ d(A, B) = \frac{1}{2|A|}\sum_{i, j \in A} k(x, x') + \frac{1}{2|B|}\sum_{i, j \in B} k(y, y') - \frac{1}{|A||B|} \sum_{A \times B} k(x, y).
\end{gather*}
From the definition, it is essentially the distance between means of vectors from $p$ and $q$ embedded into the corresponding RKHS. The resulting empirical estimator combines average inner and outer similarities between samples from the two distributions and goes to zero as the number of samples increases if $p=q$.
Note that if sample weights in Eq.~\eqref{eq:dual_d_q} are constant and equal across all samples, so $\alpha_i = c$, then the dual distance introduced above becomes exactly an empirical estimate of the MMD plus the constant from the entropic regularizer. Thus, the adversarial logistic distance introduced in Eq.~\eqref{eq:dual_d_q} can be viewed as an \textit{iteratively reweighted} empirical estimator of the MMD distance. Intuitively, what this means is that the optimization procedure consists of two alternating minimization steps: (1)~find the best sample weights assignment by changing $\alpha$ so that the regularized weighted MMD is minimized, and then (2) use a fixed $\alpha$ to minimize the resulting \textit{weighted} MMD distance by changing the matching function $M_\theta$. This makes the resulting procedure similar to the Iteratively Reweighted Least Squares Algorithm \cite{green1984iteratively} for logistic regression. An interesting observation here is that it turns out that high weights in this iterative procedure are given to the most mutually close subsets of $A$ and $B'$, where closeness is measured in terms of Maximum Mean Discrepancy. These happen to be exactly support vectors of the corresponding optimal domain classifier. Therefore, the procedure described above essentially brings sets of the support vectors of the optimal domain classifier from different domains closer together.
We note that the computational complexity of a single gradient step of the proposed method grows quadratically with the size of the dataset because of the kernelization step. However, our batched GPU implementation of the method performed on par with MMD and outperformed primal methods, probably because inference in modern neural networks requires so many dot products that a batch~size~$\times$~batch~size multiplication is negligibly cheap compared to the rest of the network with modern highly parallel computing architectures.
\section{Application to Distribution Alignment in Domain Adaptation}\label{sec:da}
We now show how the above formulation can be applied to distribution alignment for the specific problem of unsupervised domain adaptation. In this scenario, we train our classifier in a supervised fashion on some domain A and have to update it to perform well on a different domain B without using any labeled samples from the latter. Common examples include adapting to a camera with different image quality or to different weather conditions.
More rigorously, we assume that there exist two distinct distributions on $\mathcal X \times \mathcal Y$: a source distribution $P_S(X, Y)$ and a target distribution $P_T(X, Y)$. We assume that we observe a finite number of labeled samples from the source distribution~${D_S \subset [\mathcal X \times \mathcal Y ]^n \sim P_S(X, Y)}$ and a finite number of unlabeled samples from the target distribution~${D_T \subset [\mathcal X]^m \sim P_T(X)}$. Our goal then is to find a labeling function $f: \mathcal X \to \mathcal Y$ from a hypothesis space~$\mathcal F$ that minimizes target risk ${\mathcal R}_{T}$, even though we only have labels for samples from source.
\begin{equation*}
{\mathcal R}_{T}(f) = \mathbb E_{(x, y) \sim P_T} L(f(x), y) \leq {\mathcal R}_{S}(f) + d(P_T, P_S) + C(V, n)
\end{equation*}
\citet{Ben-David2007} showed that, under mild restrictions on probability distributions, the target risk is upper-bounded by the sum of three terms: (1) the source risk, (2) the complexity term involving the dataset size and the VC-dimensionality of $\mathcal F$, and (3) the \textit{discrepancy} between source and target distributions. Thus, in order to make the target risk closer to the source risk, we need to minimize the discrepancy between distributions. They define discrepancy as a supremum of differences in measures across all events in a given $\sigma$-algebra: \ $d(p, q) = \sup_{A \in \Sigma} |p(A) - q(A)|$. Estimation of the indicated expression is hard in practice, therefore it is usually replaced with more computationally feasible statistical distances. The total variation between two distributions and the Kolmogorov-Smirnov test are closely related to the discrepancy definition above and are also often considered to be too strong to be useful, especially in higher dimensions.
Our approach can be applied directly to this scenario if the discrepancy is replaced with an adversarial objective that uses a logistic regression domain classifier. In Section~\ref{sec:daexp}, we consider an instance of this problem where the main task is classification and the hypothesis space corresponds to multi-layer neural networks.
We compare the standard min-max formulation of the adversarial objective in Eq.~\eqref{eq:primal_adv_problem} with our min-min formulation in Eq.~\eqref{eq:dual_d_q}, and report the accuracy of the resulting classifier on the target domain.
\section{Related Work}
There has been a long line of work on unsupervised generative machine learning, a review of which is beyond the scope of this work. Recently, adversarial methods for learning generative neural networks have achieved popularity due to their ability to effectively model high dimensional data, e.g., generating very realistic looking images. These include the original Generative Adversarial Networks (GANs)~\citep{goodfellow_nips14} and follow up work proposing improved formulations such as Wasserstein GANs~\cite{arjovsky_arxiv17} and Conditional GANs~\cite{mirza_arxiv14}.
Related ideas have been proposed for unsupervised domain adaptation. Neural domain adaptation methods seek to improve the performance of a classifier network on a target distribution that is different from the original training distribution by introducing an additional objective that minimizes the difference between representations learned for source and target data. Some models align feature representations across domains by minimizing the distance between first or second order feature space statistics~\cite{tzeng_arxiv15,long_icml15,Sun2016}. When adversarial objectives are used for domain adaptation, a domain classifier is trained to distinguish between the generated source and target representations, both using the standard minimax objective~\cite{ganin_icml15}, as well as alternative losses~\cite{tzeng_iccv15,tzeng_cvpr17}. Instead of aligning distributions in feature space, several models perform generative alignment in pixel-space, e.g., using Coupled GANs~\cite{Liu2016} or conditional GANs~\cite{Bousmalis2016}. These models adapt by ``hallucinating'' what the raw training images might look like in the target test domain. A related line of work uses adversarial training to translate images in one domain to the ``style'' of a different domain to create artistic effects~\cite{Mathieu2016}. Several recent works perform the alignment in an unsupervised way without aligned image pairs~\cite{Zhu2017,Onsiderations2017}. Other objectives for distribution matching that have been proposed in the literature, including Maximum Mean Discrepancy~\cite{Gretton2012}, f-discrepancy \cite{Nowozin2016} and others, have also been used for generative modeling~\cite{Dziugaite2015,Li2015}. A single step of our iterative reweighting procedure is similar to instance reweighting methods that were theoretically and empirically shown to improve accuracy in the presence of domain shift. For example, \citet{huang2007correcting} used sample reweighting that minimized empirical MMD between populations to plug it as instance-weights in weighted classification loss, whereas \citet{pmlr-v28-gong13} did that to choose points for a series of independent auxiliary tasks, so no \textit{iterative} reweighting was performed in both cases.
In general, most statistical distances used for distribution alignment fall into one of two categories: they are either f-divergences (e.g. GAN objective, KL-divergence), or integral probability metrics (IPMs) that are differences in expected values of a test function at samples from different distributions maximized over some function family (e.g. Maximum Mean Discrepancy, Wasserstein distance).
In this work we specifically consider the logistic adversarial objective (f-divergence), show that it is useful to optimize its dual, and present a relation between this adversarial dual objective and MMD, another statistical distance with a test function from reproducing kernel Hilbert space (RKHS).
|
\section{Introduction}
A (binary) code $C$ of length $n$ is a vector subspace of
$\mathbb{F}_2^n$, where $\mathbb{F}_2$ denotes the finite field of order $2$.
The {\em dual} code $C^{\perp}$ of $C$ is defined as
$
C^{\perp}=
\{x \in \mathbb{F}_2^n \mid x \cdot y = 0 \text{ for all } y \in C\},
$
where $x \cdot y$ is the standard inner product.
A code $C$ is called
{\em self-dual} if $C = C^{\perp}$.
Self-dual codes are divided into two classes.
A self-dual code $C$ is {\em doubly even} if all
codewords of $C$ have weight divisible by four, and {\em
singly even} if there is at least one codeword of weight $\equiv 2
\pmod 4$.
Let $C$ be a singly even self-dual code and
let $C_0$ denote the
subcode of codewords having weight $\equiv0\pmod4$.
Then $C_0$ is a subcode of codimension $1$.
The {\em shadow} $S$ of $C$ is defined to be $C_0^\perp \setminus C$.
Shadows for self-dual codes were introduced by Conway and
Sloane~\cite{C-S}
in order to
derive new upper bounds for the minimum weight of
singly even self-dual codes.
By considering shadows, Rains~\cite{Rains} showed that
the minimum weight $d$ of a self-dual code of length $n$
is bounded by
$d \le 4 \lfloor{\frac {n}{24}} \rfloor + 6$
if $n \equiv 22 \pmod {24}$,
$d \le 4 \lfloor{\frac {n}{24}} \rfloor + 4$
otherwise.
A self-dual code meeting the bound is called {\em extremal}.
Let $C$ be a singly even self-dual code of
length $n$ with shadow $S$.
Let $d(S)$ denote the minimum weight of $S$.
We say that $C$ is a code with {\em minimal shadow}
if $r = d(S)$, where
$r=4,1,2$ and $3$ if $n \equiv 0,2,4$ and $6 \pmod 8$,
respectively.
The concept of self-dual codes with minimal shadow was introduced
in~\cite{performance}.
In that paper, different types of self-dual codes with the same
parameters were compared with regard to the decoding error probability.
In~\cite{BV}, the connection between singly even self-dual
codes with minimal shadow of some lengths, combinatorial designs and
secret sharing schemes was considered.
It was shown in~\cite{BW} that there is no extremal singly even
self-dual code with minimal shadow
for lengths $24m+2$, $24m+4$,
$24m+6$, $24m+10$ and $24m+22$.
In~\cite{BV}, it was shown that the weight enumerator of a
(non-extremal) singly even self-dual $[24m+2, 12m+1, 4m+2]$ code
with minimal shadow is uniquely determined
for each positive integer $m$. These motivate us to study singly even
self-dual codes with minimal shadow having minimum weight two less than
the hypothetical extremal case.
The main aim of this paper is to investigate singly even
self-dual codes with minimal shadow having minimum weight
$4m+2$ for the lengths $24m+2$,
$24m+4$ and $24m+10$.
We show that the weight enumerator of a singly even self-dual
code with minimal shadow having minimum weight $4m+2$
is uniquely determined for lengths $24m+4$ and $24m+10$.
For lengths $24m+2$, $24m+4$ and $24m+10$,
nonnegativity of the coefficients of weight enumerators shows
that there is no such code for $m$ sufficiently large.
We also show that the uniqueness of the weight enumerator
fails for the parameters $[24m+6,12m+3,4m+2]$ and
$[24m+22,12m+11,4m+4]$.
The paper is organized as follows.
In Section~\ref{sec:2},
we review the results given by Rains~\cite{Rains}.
In Section~\ref{sec:24m+2}, we show
that there is no singly even self-dual $[24m+2,12m+1,4m+2]$
code with minimal shadow for $m \ge 155$.
In Sections~\ref{sec:24m+4} and~\ref{sec:24m+10},
for parameters $[24m+4,12m+2,4m+2]$ and $[24m+10,12m+5,4m+2]$,
we show that there is no
singly even self-dual code with minimal shadow for $m \ge 156$
and for $m \ge 160$, respectively.
Finally, in Section~\ref{sec:rem}, we
demonstrate that the weight enumerator of a singly even self-dual
code with minimal shadow is not uniquely determined
for parameters $[24m+6,12m+3,4m+2]$ and
$[24m+22,12m+11,4m+4]$.
All computer calculations in this paper
were done with the help of
the algebra software {\sc Magma}~\cite{Magma} and
the mathematical softwares {\sc Maple} and {\sc Mathematica}.
\section{Preliminaries}\label{sec:2}
Let $C$ be a singly even self-dual code of
length $n$ with shadow $S$.
Write $n=24m+8l+2r$, where $m$ is an integer, $l\in\{0,1,2\}$ and
$r\in\{0,1,2,3\}$.
The weight enumerators $W_C(y)$ and $W_S(y)$ of $C$ and $S$
are given by (\cite[(10), (11)]{C-S})
\begin{align}\label{eq:WC}
W_C(y) &= \sum_{i=0}^{12m+4l+r}a_i y^{2i}
=
\sum_{j=0}^{3m+l}
c_j(1+y^2)^{12m+4l+r-4j}(y^2(1-y^2)^2)^j,
\\
\label{eq:WS}
W_S(y) &= \sum_{i=0}^{6m+2l}b_i y^{4i+r}
= \sum_{j=0}^{3m+l}(-1)^j c_j2^{12m+4l+r-6j}y^{12m+4l+r-4j}(1-y^4)^{2j},
\end{align}
respectively, for suitable integers $c_j$.
Let
\begin{equation}\label{4a8}
(1+y^2)^{n/2-4j}(y^2(1-y^2)^2)^j=
\sum_{i=0}^{12m+4l+r} \alpha'_{i,j} y^{2i}\quad(0\leq j\leq 3m+l).
\end{equation}
Then
\begin{equation}\label{4b0}
\alpha_{i,j}'=\begin{cases}
0&\text{if $0\leq i<j\leq 3m+l$,}\\
1&\text{if $0\leq i=j\leq 3m+l$.}
\end{cases}
\end{equation}
This implies that the $(3m+l+1)\times(3m+l+1)$ matrix $[\alpha'_{i,j}]$
is invertible, since it is unitriangular.
Let $[\alpha_{i,j}]$ be its inverse matrix.
Then by~\eqref{4b0}, we have
\begin{equation}\label{4b1}
\alpha_{i,j}=\begin{cases}
0&\text{if $0\leq i<j\leq 3m+l$,}\\
1&\text{if $0\leq i=j\leq 3m+l$,}
\end{cases}
\end{equation}
and
\begin{equation}\label{4by}
y^{2i}=\sum_{j=0}^{3m+l}\alpha_{j,i}(1+y^2)^{n/2-4j}(y^2(1-y^2)^2)^j
\quad(0\leq i\leq 3m+l)
\end{equation}
by~\eqref{4a8}.
By~\eqref{eq:WC}, \eqref{4b1} and~\eqref{4by}, we obtain
\begin{equation}\label{4a10}
c_i=\sum_{j=0}^{i} \alpha_{i,j} a_j.
\end{equation}
\begin{lem}\label{lem:alpha}
For $1\leq i\leq 3m+l$, we have
\begin{equation}\label{eq:alpha}
\alpha_{i,0}=
-\frac{n}{2i}
\sum_{\substack{0\leq t\leq n/2+1-6i\\
t+i\text{ is odd}}}
(-1)^t\binom{\frac{n}{2}+1-6i}{t}
\binom{\frac{n-7i-t-1}{2}}{\frac{i-t-1}{2}}.
\end{equation}
\end{lem}
\begin{proof}
For $1 \le i$,
\begin{align*}
\alpha_{i,0}=&
-\frac{n}{2i}[\text{coeff.\ of } y^{i-1} \text{ in }
(1+y)^{-n/2-1+4i}(1-y)^{-2i}],
\end{align*}
\cite{Rains}.
Since
\begin{align*}
&(1+y)^{-n/2-1+4i}(1-y)^{-2i}\\
&=(1-y^2)^{-n/2-1+4i}(1-y)^{n/2+1-6i} \\
&=(1-y^2)^{-n/2-1+4i}
\sum_{t=0}^{n/2+1-6i}(-1)^t\binom{\frac{n}{2}+1-6i}{t}y^t,
\end{align*}
we have
\begin{align*}
\alpha_{i,0}&=
-\frac{n}{2i}
\sum_{t=0}^{n/2+1-6i}(-1)^t\binom{\frac{n}{2}+1-6i}{t}
[\text{coeff.\ of } y^{i-1} \text{ in }
(1-y^2)^{-n/2-1+4i} y^t]
\\&=
-\frac{n}{2i}
\sum_{\substack{0\leq t\leq n/2+1-6i\\
t+i\text{ is odd}}}
(-1)^t\binom{\frac{n}{2}+1-6i}{t}
(-1)^{(i-t-1)/2}\binom{-\frac{n}{2}-1+4i}{\frac{i-t-1}{2}}.
\end{align*}
The result follows by applying the formula
\[(-1)^j\binom{-n}{j}=\binom{n+j-1}{j}.\]
\end{proof}
Write
\begin{equation*}\label{4c8}
(-1)^j2^{n/2-6j}y^{n/2-4j}(1-y^4)^{2j}=
\sum_{i=0}^{6m+2l} \beta'_{i,j} y^{4i+r}
\quad(0\leq j\leq 3m+l).
\end{equation*}
Since $n/2-4j=4(3m+l-j)+r$, we have
\begin{equation*}\label{4d0}
\beta_{i,j}'=\begin{cases}
0&\text{if $i<3m+l-j$,}\\
(-1)^j 2^{n/2-6j}&\text{if $i=3m+l-j$.}
\end{cases}
\end{equation*}
This implies that the $(3m+l+1)\times(3m+l+1)$ matrix $[\beta'_{i,3m+l-j}]$
is invertible, since it is lower triangular such that
the diagonal elements are not zeros. Thus, the matrix
$[\beta'_{i,j}]$ is also invertible.
Let $[\beta_{i,j}]$ be its inverse matrix. Then
\begin{equation}\label{4c9}
y^{4i+r}=\sum_{j=0}^{3m+l} \beta_{j,i}(-1)^j 2^{n/2-6j}y^{n/2-4j}(1-y^4)^{2j}
\quad(0\leq i\leq 3m+l).
\end{equation}
Moreover,
$[\beta_{3m+l-i,j}]$ is the inverse of the lower triangular matrix
$[\beta'_{i,3m+l-j}]$, and so lower triangular as well, and
\[\beta_{3m+l-j,j}={\beta'}^{-1}_{j,3m+l-j}.\]
Thus
\begin{equation}\label{4d1}
\beta_{i,j}=\begin{cases}
0&\text{if $i>3m+l-j$,}\\
(-1)^{3m+l-j} 2^{6(3m+l-j)-n/2}&\text{if $i=3m+l-j$.}
\end{cases}
\end{equation}
By~\eqref{eq:WS}, \eqref{4c9} and~\eqref{4d1}, we obtain
\begin{equation}\label{4c10}
c_i=\sum_{j=0}^{3m+l-i} \beta_{i,j} b_j.
\end{equation}
\begin{lem}[Rains~\cite{Rains}]\label{lem:6b}
For $1\leq i\leq3m+l$ and $0\leq j\leq3m+l$ with $i+j\leq 3m+l$, we have
\begin{equation}\label{eq:beta}
\beta_{i,j}=(-1)^i2^{-n/2+6i}\frac{3m+l-j}{i}\binom{3m+l+i-j-1}{3m+l-i-j}.
\end{equation}
\end{lem}
From~\eqref{4a10} and~\eqref{4c10}, we have
\begin{equation}\label{eq:ci}
c_i=\sum_{j=0}^{i} \alpha_{i,j} a_j=\sum_{j=0}^{3m+l-i} \beta_{i,j} b_j.
\end{equation}
Now let $C$ be a singly even self-dual $[24m+8l+2r,12m+4l+r,4m+2]$ code
with minimal shadow.
Suppose that $(l,r) \in \{(0,1),(0,2),(1,1)\}$.
Since the minimum weight of $C$ is $4m+2$, we have
\begin{equation}\label{eq:ai}
a_0=1,a_1=a_2=\cdots = a_{2m}=0.
\end{equation}
Since the minimum weight of the shadow is $1$ or $2$, we have
\begin{equation}\label{eq:bi}
\begin{cases}
b_0=1 & \text{ if } m=1, \\
b_0=1, b_1=b_2= \cdots =b_{m-1}=0 & \text{ if } m \ge 2.
\end{cases}
\end{equation}
From~\eqref{eq:ci}, \eqref{eq:ai} and~\eqref{eq:bi},
we have
\begin{equation}\label{eq:c_i-all}
c_i=
\begin{cases}
\alpha_{i,0} & \text{ if }i=0,1,\ldots,2m, \\
\beta_{i,0} & \text{ if }i=2m+l+1,2m+l+2,\ldots,3m+l.
\end{cases}
\end{equation}
Suppose that $l=0$.
From~\eqref{eq:ci}, \eqref{eq:ai},
\eqref{eq:bi} and~\eqref{eq:c_i-all}, we obtain
\begin{align}\label{eq:1}
c_{2m}=&\alpha_{2m,0}
=\beta_{2m,0} + \beta_{2m,m}b_m,
\\
c_{2m-1} =&
\alpha_{2m-1,0}
=\beta_{2m-1,0} + \beta_{2m-1,m}b_m +\beta_{2m-1,m+1}b_{m+1}.
\label{eq:2}
\end{align}
Suppose that $l=1$.
From~\eqref{eq:ci}, \eqref{eq:ai},
\eqref{eq:bi} and~\eqref{eq:c_i-all}, we obtain
\begin{align}\label{eq:3}
c_{2m}=&\alpha_{2m,0}=\beta_{2m,0}+\beta_{2m,m}b_{m}+\beta_{2m,m+1}b_{m+1},
\\
c_{2m+1}
=&\alpha_{2m+1,0}+\alpha_{2m+1,2m+1}a_{2m+1}
=\beta_{2m+1,0}+\beta_{2m+1,m}b_m.
\label{eq:4}
\end{align}
\section{Singly even
self-dual $[24m+2,12m+1,4m+2]$ codes with minimal shadow}\label{sec:24m+2}
It was shown in~\cite{BV} that
the weight enumerator of a singly even self-dual $[24m+2,12m+1,4m+2]$
code with minimal shadow is uniquely determined for each length.
In this section, we show that
there is no singly even self-dual $[24m+2,12m+1,4m+2]$ code
with minimal shadow for $m \ge 155$.
Suppose that $m \ge 1$.
Let $C$ be a singly even self-dual $[24m+2,12m+1,4m+2]$ code
with minimal shadow.
The weight enumerators of $C$ and its shadow $S$ are written
as in~\eqref{eq:WC} and~\eqref{eq:WS}, respectively.
From~\eqref{eq:alpha},
\begin{align*}
\alpha_{2m,0}=
\frac{12m+1}{m}\binom{5m}{m-1}.
\end{align*}
From~\eqref{4d1},
\[
\beta_{2m,m}=\frac{1}{2}.
\]
From~\eqref{eq:beta},
\begin{align*}
\beta_{2m,0}
=\frac{3}{2^2}\binom{5m-1}{m}
=\frac{3(4m+1)}{5m}\binom{5m}{m-1}.
\end{align*}
From~\eqref{eq:1},
\begin{align*}
b_m=\frac{\alpha_{2m,0}-\beta_{2m,0}}{\beta_{2m,m}}
=\frac{4(24m+1)}{5m}\binom{5m}{m-1}.
\end{align*}
\begin{rem}
Unfortunately, $b_m$ was incorrectly reported
in~\cite{ZMFG}. The correct formula for $b_m$ is given in~\cite{ZMFG2}.
We showed that $b_m$ is always a positive integer (see~\cite{ZMFG2}).
\end{rem}
From~\eqref{eq:alpha},
\begin{align*}
\alpha_{2m-1,0}=&
-\frac{12m+1}{2m-1}
\left(
\binom{5m+4}{m-1}
+28\binom{5m+3}{m-2}
+70\binom{5m+2}{m-3}
\right.
\\ &
\left.
+28\binom{5m+1}{m-4}
+\binom{5m}{m-5}
\right)
\\ =&
-
\frac{8(12m+1)(376m^3- 4m^2 + 5m +1)}
{(4m+2)(4m+3)(4m+4)(4m+5)}
\binom{5m}{m-1}.
\end{align*}
From~\eqref{4d1},
\[
\beta_{2m-1,m+1}=-\frac{1}{2^7}.
\]
From~\eqref{eq:beta},
\begin{align*}
\beta_{2m-1,0}
=&-\frac{1}{2^{7}}\frac{3m}{2m-1}\binom{5m-2}{m+1}
=-\frac{1}{2^{4}}
\frac{3(4m-1)(4m+1)}{5(5m-1)(m+1)}\binom{5m}{m-1},\\
\beta_{2m-1,m}
=&-\frac{m}{2^{5}}.
\end{align*}
From~\eqref{eq:2},
\begin{align*}
b_{m+1} =&\frac{\alpha_{2m-1,0}- \beta_{2m-1,0} -
\beta_{2m-1,m}b_m}{\beta_{2m-1,m+1}}
\\
=&-\frac{64(24m+1)f(m)}
{(5m-1)(4m+2)(4m+3)(4m+4)(4m+5)}\binom{5m}{m-1},
\end{align*}
where
\[
f(m)=64m^5-14816m^4+2812m^3+46m^2-14m+1.
\]
\begin{thm}
All coefficients in the weight enumerators of
a singly even self-dual $[24m+2,12m+1,4m+2]$ code
and its shadow are nonnegative integers
if and only if $1 \le m \le 154$.
In particular,
for $m \ge 155$, there is no
singly even self-dual $[24m+2,12m+1,4m+2]$
code with minimal shadow.
\end{thm}
\begin{proof}
We verified that the equation
$f(m)=0$
has three solutions consisting of real numbers and
the largest solution is in the interval $(231,232)$.
Thus, $b_{m+1}$ is negative for $m \ge 232$.
Using~\eqref{eq:WC} and~\eqref{eq:WS},
we determined numerically
the weight enumerators of
a singly even self-dual $[24m+2,12m+1,4m+2]$
code with minimal shadow and its shadow
for $m \le 231$.
The theorem follows from this calculation.
\end{proof}
\section{Singly even
self-dual $[24m+4,12m+2,4m+2]$ codes with minimal shadow}\label{sec:24m+4}
\begin{prop}
The weight enumerator of a singly even self-dual $[24m+4,12m+2,4m+2]$
code with minimal shadow is uniquely determined for each length.
\end{prop}
\begin{proof}
The weight enumerator of a singly even self-dual
$[4,2,2]$ code is uniquely determined.
Suppose that $m \ge 1$.
Let $C$ be a singly even self-dual $[24m+4,12m+2,4m+2]$ code
with minimal shadow.
The weight enumerators of $C$ and its shadow $S$ are written
as using~\eqref{eq:WC} and~\eqref{eq:WS}.
Since $\alpha_{i,0}$ $(i=0,1,\ldots,2m)$ and
$\beta_{i,0}$ $(i=2m+1,2m+2,\ldots,3m)$ are calculated by
\eqref{eq:alpha} and~\eqref{eq:beta}, respectively,
from~\eqref{eq:c_i-all},
$c_i$ $(i=0,1,\ldots,3m)$ depends only on $m$.
This means that the weight enumerator of $C$ is uniquely
determined for each length.
\end{proof}
From~\eqref{eq:alpha},
\begin{align*}
\alpha_{2m,0}=&
\frac{6m+1}{m}\left(3\binom{5m+1}{m-1}+\binom{5m}{m-2}\right)
\\=&
\frac{(6m+1)(8m+1)}{m(2m+1)}\binom{5m}{m-1}.
\end{align*}
From~\eqref{4d1},
\[
\beta_{2m,m}=\frac{1}{2^2}.
\]
From~\eqref{eq:beta},
\begin{align*}
\beta_{2m,0}
=\frac{1}{2^2}\frac{3}{2}\binom{5m-1}{m}
=\frac{3(4m+1)}{10m}\binom{5m}{m-1}.
\end{align*}
From~\eqref{eq:1},
\begin{align*}
b_m=&\frac{\alpha_{2m,0}-\beta_{2m,0}}{\beta_{2m,m}}
=
\frac{2(12m+1)(38m+7)}{5m(2m+1)}\binom{5m}{m-1}.
\end{align*}
\begin{rem}
Unfortunately, $b_m$ was incorrectly reported
in~\cite{ZMFG}. The correct formula for $b_m$ is given in~\cite{ZMFG2}.
We showed that $b_m$ is always a positive integer (see~\cite{ZMFG2}).
\end{rem}
From~\eqref{eq:alpha},
\begin{align*}
\alpha_{2m-1,0}=&
-\frac{12m+2}{2m-1}
\left(
\binom{5m+5}{m-1}
+36\binom{5m+4}{m-2}
+126\binom{5m+3}{m-3}\right.
\\ &
\left.
+84\binom{5m+2}{m-4}
+9\binom{5m+1}{m-5}
\right)
\\=&
-\frac{16 (5m+1)(6m+1)(8m+1)(68m^2-m+3)}
{(4m+2)(4m+3)(4m+4)(4m+5)(4m+6)}\binom{5m}{m-1}.
\end{align*}
From~\eqref{4d1},
\[
\beta_{2m-1,m+1}=-\frac{1}{2^8}.
\]
From~\eqref{eq:beta},
\begin{align*}
\beta_{2m-1,0}
=&-\frac{1}{2^{8}}\frac{3m}{2m-1}\binom{5m-2}{m+1}
=-\frac{1}{2^{5}}
\frac{3(4m-1)(4m+1)}{5(5m-1)(m+1)}\binom{5m}{m-1}, \\
\beta_{2m-1,m}
=&-\frac{m}{2^6}.
\end{align*}
From~\eqref{eq:2},
\begin{align*}
b_{m+1} =&\frac{\alpha_{2m-1,0}- \beta_{2m-1,0} -
\beta_{2m-1,m}b_m}{\beta_{2m-1,m+1}}
\\
=&
-\frac{128(12m+1)f(m)}
{(5m-1)(4m+2)(4m+3)(4m+4)(4m+5)(4m+6)}
\binom{5m}{m-1},
\end{align*}
where
\[
f(m) =
1216 m^6-212096 m^5-33020 m^4+5440 m^3+1171 m^2+88 m+6.
\]
\begin{thm}
All coefficients in
the weight enumerators of
a singly even self-dual $[24m+4,12m+2,4m+2]$ code
and its shadow are nonnegative integers
if and only if $1 \le m \le 155$.
In particular, for $m \ge 156$, there is no
singly even self-dual $[24m+4,12m+2,4m+2]$
code with minimal shadow.
\end{thm}
\begin{proof}
We verified that the equation
$f(m)=0$
has two solutions consisting of real numbers and
the largest solution is in the interval $(174,175)$.
Thus, $b_{m+1}$ is negative for $m \ge 175$.
Using~\eqref{eq:WC} and~\eqref{eq:WS},
we determined numerically
the weight enumerators of
a singly even self-dual $[24m+4,12m+2,4m+2]$
code with minimal shadow and its shadow
for $m \le 174$.
The theorem follows from this calculation.
\end{proof}
\section{Singly even
self-dual $[24m+10,12m+5,4m+2]$ codes with minimal shadow}\label{sec:24m+10}
\begin{lem}[Harada~\cite{H60}]
\label{lem:H}
Suppose that $n \equiv 2 \pmod 8$.
Let $C$ be a singly even self-dual $[n,n/2,d]$ code
with minimal shadow.
If $d \equiv 2 \pmod 4$, then $a_{d/2} =b_{(d-2)/4}$.
\end{lem}
As a consequence, the weight enumerator of a singly even
self-dual $[58,29,10]$ code with minimal shadow
was uniquely determined in~\cite{H60}.
\begin{prop}
The weight enumerator of a singly even self-dual
$[24m+10,12m+5,4m+2]$ code with minimal shadow
is uniquely determined for each length.
\end{prop}
\begin{proof}
The weight enumerator of a singly even self-dual
$[10,5,2]$ code with minimal shadow is uniquely determined.
Suppose that $m \ge 1$.
Let $C$ be a singly even self-dual $[24m+10,12m+5,4m+2]$ code
with minimal shadow.
The weight enumerators of $C$ and its shadow $S$ are written
as in~\eqref{eq:WC} and~\eqref{eq:WS}, respectively.
Since $\alpha_{i,0}$ $(i=0,1,\ldots,2m)$ and
$\beta_{i,0}$ $(i=2m+2,2m+3,\ldots,3m+1)$ are calculated
by~\eqref{eq:alpha} and~\eqref{eq:beta}, respectively,
from~\eqref{eq:c_i-all},
$c_i$ $(i=0,1,\ldots,2m,2m+2,\ldots,3m+1)$ depends only on $m$.
From~\eqref{4b1} and~\eqref{4d1}, we have
\[
\alpha_{2m+1,2m+1}=1 \text{ and }
\beta_{2m+1,m}=-2,
\]
respectively.
By Lemma~\ref{lem:H}, it holds that $a_{2m+1}=b_m$.
From~\eqref{eq:4}, we obtain
\begin{equation}\label{eq:a2m+1}
a_{2m+1}=\frac{\beta_{2m+1,0}-\alpha_{2m+1,0}}{3}.
\end{equation}
Therefore, from~\eqref{eq:4},
$c_{2m+1}$ depends only on $m$.
This means that the weight enumerator of $C$ is uniquely
determined for each length.
\end{proof}
From~\eqref{eq:alpha},
we have
\begin{align*}
\alpha_{2m+1,0}=&
-\frac{12m+5}{2m+1}\binom{5m+1}{m}.
\end{align*}
From~\eqref{eq:beta}, we have
\begin{align*}
\beta_{2m+1,0}
= -2 \frac{3m+1}{2m+1}\binom{5m+1}{m}.
\end{align*}
Since $a_{2m+1}=b_m$,
from~\eqref{eq:a2m+1}, we have
\[
b_m=\binom{5m+1}{m}=\frac{5m+1}{4m+1}\binom{5m}{m}.
\]
From~\eqref{eq:alpha},
\begin{align*}
\alpha_{2m,0}=&
\frac{12m+5}{2m}\left(
6\binom{5m+4}{m-1}
+20\binom{5m+3}{m-2}
+6\binom{5m+2}{m-3}
\right)
\\
=&
\frac{4(12m+5)(5m+1)(5m+2)(32m^2 +19m+3)}{
(4m+1)(4m+2)(4m+3)(4m+4)(4m+5)}\binom{5m}{m}.
\end{align*}
From~\eqref{4d1},
\[
\beta_{2m,m+1}=\frac{1}{2^5}.
\]
From~\eqref{eq:beta},
\begin{align*}
\beta_{2m,0}
=&\frac{1}{2^5}\frac{3m+1}{2m} \binom{5m}{m+1}
=\frac{1}{2^4}\frac{3m+1}{m+1} \binom{5m}{m},
\\
\beta_{2m,m}
=&\frac{1}{2^5} \frac{2m+1}{2m} 4m
=\frac{2m+1}{2^4}.
\end{align*}
From~\eqref{eq:3},
\begin{align*}
b_{m+1}=&\frac{\alpha_{2m,0}-\beta_{2m,0}-\beta_{2m,m}b_{m}}{\beta_{2m,m+1}}\\
=&
-\frac{16(5m+2)f(m)}
{(4m+1)(4m+2)(4m+3)(4m+4)(4m+5)}
\binom{5m}{m},
\end{align*}
where
\[
f(m)=
64m^5- 15040m^4- 18036m^3 - 7924m^2 - 1511m -105.
\]
\begin{thm}
All coefficients in
the weight enumerators of
a singly even self-dual $[24m+10,12m+5,4m+2]$ code
and its shadow are nonnegative integers
if and only if $1 \le m \le 159$.
In particular, for $m \ge 160$, there is no
singly even self-dual $[24m+10,12m+5,4m+2]$
code with minimal shadow.
\end{thm}
\begin{proof}
We verified that the equation
$f(m)=0$
has three solutions consisting of real numbers and
the largest solution is in the interval $(236,237)$.
Thus, $b_{m+1}$ is negative for $m \ge 237$.
Using~\eqref{eq:WC} and~\eqref{eq:WS},
we determined numerically
the weight enumerators of
a singly even self-dual $[24m+10,12m+5,4m+2]$
code with minimal shadow and its shadow
for $m \le 236$.
The theorem follows from this calculation.
\end{proof}
\section{Remaining cases}\label{sec:rem}
For the remaining cases, we demonstrate that
the weight enumerator of a singly even self-dual
code with minimal shadow is not uniquely determined.
\subsection{Singly even
self-dual $[24m+6,12m+3,4m+2]$ codes with minimal shadow}
Using~\eqref{eq:WC} and~\eqref{eq:WS},
the possible weight enumerators of
a singly even self-dual $[30,15,6]$ code with minimal shadow
and its shadow are given by
\begin{align*}
&
1+( 35 - 8 \beta) y^6
+( 345 + 24 \beta) y^8
+ 1848 y^{10}
+ \cdots,
\\&
\beta y^3
+( 240 - 6 \beta) y^7
+( 6720 + 15 \beta) y^{11}
+ \cdots,
\end{align*}
respectively, where $\beta$ is an integer with $1 \le \beta \le 4$.
It is known that there is
a singly even self-dual $[30,15,6]$ code with minimal shadow
for $\beta \in \{1,2\}$ (see~\cite{C-S}).
Using~\eqref{eq:WC} and~\eqref{eq:WS},
the possible weight enumerators of
a singly even self-dual $[54,27,10]$ code with minimal shadow
and its shadow are given by
\begin{align*}
&
1
+( 351 - 8 \beta) y^{10}
+( 5543 + 24 \beta) y^{12}
+( 43884 + 32 \beta) y^{14}
+ \cdots,
\\&
y^3
+(- 12 + \beta) y^{7 }
+( 2874 - 10 \beta) y^{11}
+( 258404 + 45 \beta) y^{15}
+ \cdots,
\end{align*}
respectively, where $\beta$ is an integer with $12 \le \beta \le 43$.
It is known that there is
a singly even self-dual $[54,27,10]$ code with minimal shadow
for $\beta \in \{12,13,\ldots,20,21,22,24,26\}$ (see~\cite{YL14}).
\subsection{Singly even
self-dual $[24m+22,12m+11,4m+4]$ codes with minimal shadow}
Using~\eqref{eq:WC} and~\eqref{eq:WS},
the possible weight enumerators of
a singly even self-dual $[22,11,4]$ code with minimal shadow
and its shadow are given by
\begin{align*}
&
1
+ 2 \beta y^4
+( 77 - 2 \beta) y^6
+( 330 - 6 \beta) y^8
+( 616 + 6 \beta) y^{10}
+ \cdots,
\\&
\beta y^3
+( 352 - 4 \beta) y^7
+( 1344 + 6 \beta) y^{11}
+ \cdots,
\end{align*}
respectively, where $\beta$ is an integer with $1 \le \beta \le 38$.
It is known that there is a
singly even self-dual $[22,11,4]$ code with minimal shadow
for $\beta \in \{2,4,6,8,10,14\}$ (see~\cite{PS75}).
Using~\eqref{eq:WC} and~\eqref{eq:WS},
the possible weight enumerators of
a singly even self-dual $[46,23,8]$ code with minimal shadow
and its shadow are given by
\begin{align*}
&
1
+ 2 \beta y^8
+ (884 - 2 \beta) y^{10}
+ (10556 - 14 \beta) y^{12}
+ (54621 + 14 \beta) y^{14}
+ \cdots,
\\&
y^3
+(- 10 + \beta) y^7
+( 6669 - 8 \beta) y^{11}
+( 242760 + 28 \beta) y^{15}
+ \cdots,
\end{align*}
respectively, where $\beta$ is an integer with $10 \le \beta \le 442$.
Let $C_{46}$ be the code with generator matrix
$\left[
\begin{array}{cc}
I_{23} & R
\end{array}
\right]$,
where $I_{23}$ denotes the identity matrix of order $23$ and
$R$ is the $23 \times 23$ circulant matrix with
first row
\[
(01011101011100000111110).
\]
We verified that $C_{46}$ is a singly even self-dual
$[46,23,8]$ code.
By considering self-dual neighbors of $C_{46}$,
we found singly even self-dual
$[46,23,8]$ codes $N_{46,i}$ with minimal shadow $(i=1,2,\ldots,10)$.
These codes are constructed as
$\langle C_{46} \cap \langle x \rangle^\perp, x \rangle$,
where the supports $\supp(x)$ of $x$ are listed in Table~\ref{Tab:nei}.
The values $\beta$ in the weight enumerators of $N_{46,i}$
are also listed in the table.
\begin{table}[thb]
\caption{Singly even self-dual $[46,23,8]$ codes $N_{46,i}$ with
minimal shadows}
\label{Tab:nei}
\begin{center}
{\small
\begin{tabular}{c|l|c}
\noalign{\hrule height0.8pt}
Codes & \multicolumn{1}{c|}{$\supp(x)$} & $\beta$\\
\hline
$N_{46,1}$ & $\{1,24,26,27,29,30,31,32,33,34,36,37,42,43,45,46\}$&36\\
$N_{46,2}$ & $\{1,27,28,31,33,35,36,37,42,43,45,46\}$&42\\
$N_{46,3}$ & $\{10,11,20,27,29,34,38,41,42,45\}$&44\\
$N_{46,4}$ & $\{5,6,25,29,30,32,33,36,40,41,44,45\}$&46\\
$N_{46,5}$ & $\{1,23,28,29,30,31,32,37,40,41,44,45\}$&48\\
$N_{46,6}$ & $\{1,26,27,28,30,32,35,36,37,42,43,45\}$&50\\
$N_{46,7}$ & $\{2,3,24,25,26,28,29,33,34,36,37,41,42,44\}$&52\\
$N_{46,8}$ & $\{1,25,28,29,32,33,34,36,38,42,43,45\}$&54\\
$N_{46,9}$ & $\{1,23,24,27,30,36,40,41,44,45\}$&56\\
$N_{46,10}$& $\{1,2,25,29,30,33,35,38,44,46\}$&58\\
\noalign{\hrule height0.8pt}
\end{tabular}
}
\end{center}
\end{table}
The possible weight enumerators of
a singly even self-dual $[70,35,12]$ code with minimal shadow
and its shadow are given by
\begin{align*}
&
1
+ 2 \beta y^{12}
+( 9682 - 2 \beta) y^{14}
+( 173063 - 22 \beta) y^{16}
+ \cdots,
\\&
y^3
+(- 104 + \beta )y^{11}
+( 88480 - 12 \beta) y^{15}
+ \cdots,
\end{align*}
respectively, where $\beta$ is an integer
$104 \le \beta \le 4841$~\cite{H70}.
It is known that there is a
singly even self-dual $[70,35,12]$ code with minimal shadow
for many different $\beta$~\cite[p.~1191]{YLGI}.
\bigskip
\noindent
{\bf Acknowledgment.}
The first author
is supported by Grant DN 02/2/13.12.2016 of the
Bulgarian National Science Fund.
The second author
is supported by JSPS KAKENHI Grant Number 15H03633.
|
\subsection{Please Capitalize the First Letter of Each Notional Word in Subsection Title}
\section{Introduction}
The supernova remnant (SNR) G15.4+0.2, also called G15.42+0.18 (G15 hereafter), has a break-out morphology toward its south and a northern shell with a radius of about 6$^{\prime} $\ in 1420~MHz. The spectral index of G15 is $-$0.62 $\pm$ 0.03 from 330 to 4800~MHz \citep{Supan2015A&A...576A..81S, Sun2011A&A...536A..83S} with its peak flux density of 5.6~Jy at 1~GHz. G15 has a morphological correspondence with the TeV $\gamma$-ray source HESS J1818-154 \citep{Adrian-Martinez2011arXiv1112.0478A}. The high energy radiation possibly originates from a pulsar wind nebula (PWN) although the pulsar has not been detected. Multiwavelength studies have been performed to explore their properties (e.g. \citealt{Abramowski2014A&A...562A..40H, Castelletti2013A&A...557L..15C})
Distance is a key to further study this TeV SNR. A near distance for the remnant G15 was suggested by \citet{Castelletti2013A&A...557L..15C} by analyzing a noisy \textsc{Hi} absorption spectrum to the remnant. We make a further effort to obtain its reliable distance by taking advantage of new-released \textsc{Hi}/OH/Recombination line survey (THOR, \citealt{Beuther2016A&A...595A..32B}), the archive Southern Galactic Plane Survey (SGPS, \citealt{McClure-Griffiths2005ApJS..158..178M}), the Very Large Array Galactic Plane Survey (VGPS, \citealt{Stil2006AJ....132.1158S}), and the Galactic Ring Survey (GRS, \citealt{Jackson2006ApJS..163..145J}) data.
The SGPS has a resolution of 100$^{\prime\prime} $\ and a sensitivity of less than 1~mJy~beam$^{-1}$ and were observed by The Australia Telescope Compact Array (ATCA) and Parkes telescopes. The SGPS survey has a reliable absolute flux scale by including the Parkes data. The continuum data of SGPS near G15 region is currently not available. The THOR survey observed \textsc{Hi}, four OH, 19 H$_{\alpha}$ recombination lines, and the L-band continuum covering the northern Galactic plane (15~$<~l~<$~67$^{\circ}$ and $|b|$~$<$~1$^{\circ}$). The THOR data has an angular resolution about 20$^{\prime\prime}$. We convert the unit of Jy~beam$^{-1}$ in the THOR data to Kelvin (K) using factors of 1536.20 for the cube data. The absolute scale of the surface brightness in the THOR data is not reliable due to lacking zero baseline. However, the relative scale is good enough for us to do \textsc{Hi} absorption analysis. In addition, we adopt the $^{13}$CO data from the GRS to confirm the reality of the absorption peaks, because $^{13}$CO is a tracer of cold and dense H$_2$ molecular cloud.
\section{Spectra analysis}
\label{sec:spec}
We use well-tested methods in order to build a reliable \textsc{Hi} absorption spectrum to G15 as possible \citep{Tian2007A&A...474..541T, Tian2011ApJ...729L..15T, Zhu2013ApJ...775...95Z}.
We improve the absorption spectrum extraction by two steps. We choose a background region which is an annulus around the source region. Our background is adjacent to the source region, which makes that the background spectrum in the source direction. We remove six point-like sources from both the continuum map and the data cube to avoid contamination (see Fig.~\ref{fig:g15_img}). We show the \textsc{Hi} emission, the \textsc{Hi} absorption, and the $^{13}$CO emission spectra of G15 (see Fig.~\ref{fig:sgps} and~\ref{fig:thor}). We extract the \textsc{Hi} emission spectra from our source and background regions and then calculate the difference between the two spectra. This difference spectrum is actually proportional to the absorption spectrum, which is used to do absorption analysis. We extract the $^{13}$CO spectrum in our selected background region from the GRS data and compare with our absorption spectrum.
The maximum velocity \textsc{Hi} absorption feature appears at about 140~km s$^{-1}$ (see Fig.~\ref{fig:sgps} and~\ref{fig:thor}). The $^{13}$CO has data between -5 km s$^{-1}$ and 135 km s$^{-1}$. There is, in fact, a $^{13}$CO emission feature at 135 km s$^{-1}$ which seems corresponding with the \textsc{Hi} maximum velocity absorption feature. The rms of 0.009~K is obtained by averaging no $^{13}$CO emission-line velocity ranges from -5 to 10~km s$^{-1}$ and from 60 to 120~km s$^{-1}$. The $^{13}$CO peak temperature at 135~km s$^{-1}$ is 0.047~K, which is a 5$\sigma$ detection. This supports the \textsc{Hi} absorption feature at about 140~km s$^{-1}$ is real. The velocity of 140~km s$^{-1}$ corresponds to a distance of about 8~kpc based on the Galactic rotation model with $\Theta_{\odot}$ = 220~km s$^{-1}$ and R$_{\odot}$ = 8.5~kpc \citep{Fich1989ApJ...342..272F}. Thus, the lower limit distance of G15 is 8~kpc.
The \textsc{Hi} emission at a velocity of about 95~km s$^{-1}$ seems no responsible absorption, which means G15 is likely located in front of this \textsc{Hi} cloud. This velocity corresponds to a far side distance of about 10.5~kpc. The absorption peak at the velocity of about -20~km s$^{-1}$ is possibly caused by \textsc{Hi} emission fluctuation so likely not a real feature (see the \textsc{Hi} emission spectra in Figs.~\ref{fig:sgps} and~\ref{fig:thor}). Similar fake absorption feature is often seen in faint SNRs (e.g. \citealt{Leahy2008A&A...480L..25L}).
\section{Results and brief discussion}
We obtain a new distance of 9.3~$\pm$~1.3~kpc for G15. This is consistent with the 10 $\pm$ 3~kpc derived from the surface brightness vs. diameter relation of Galactic SNRs \citep{Adrian-Martinez2011arXiv1112.0478A}.
\citet{Castelletti2013A&A...557L..15C} showed an \textsc{Hi} absorption spectrum and suggested a smaller distance for the remnant. However, we do not find a clue from their paper to reproduce their spectrum, so we believe our careful measurement is more reliable.
The \textsc{Hi} shell of G15 is 16.5~pc in radius (for 6\farcm1 arcmin) at a distance of 9.3~kpc. The PWN inside G15 has a total energy of less than 1.2 $\times$ 10$^{50}$~erg scaled from the value derived in \citet{ Abramowski2014A&A...562A..40H}.
We revise the parameters of G15 using the shell-type model of SNR in \citet{Sturner1997ApJ...490..619S}. We assume the progenitor supernova of G15 has an explosion energy of 10$^{51}$~erg, a mass of 8~M$_{\odot}$, a magnetic field strength of 1~$\mu$G, and an interstellar medium (ISM) density of 7 hydrogen atoms per cube centimeter, so we get the expansion velocity of about 700~km s$^{-1}$, an age of 11~kyr, and a total atomic gas mass that
forms the G15 shell of 3.3~$\times$~10$^3$~M$_{\odot}$. The estimated parameters may have a 50 percent uncertainty due to the large variance of the interstellar medium density.
\section*{Acknowledgements}
We thank Dr. Wenwu Tian for his initial questioning the SNR G15.4+0.1's distance issue and helpful comments during our writing the paper. We acknowledge support from the NSFC (11473038, 11273025). This publication makes use of molecular line data from the Boston University-FCRAO Galactic Ring Survey (GRS).
\begin{figure}
\includegraphics[width=0.8\textwidth]{ms0117fig1.eps}
\caption{The image of G15 at 1420~MHz from the VGPS data. The black dashed circle shows the source region (`ON' direction) with its center at (g, l) = (15.43$^\circ$, 0.17$^\circ$) and a radius of 6\farcm1. The region between the white solid circle and the white dashed circle is the background region (`OFF' direction). The black circles show point-like sources from the THOR continuum image, which are excluded from our analysis. The black solid ellipse shows a bright source in the SGPS \textsc{Hi} channel maps, which is also flagged in our spectra analysis. The gray contour levels are 20.5, 21.0, and 21.2 Kelvin.}
\label{fig:g15_img}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\textwidth]{ms0117fig2.eps}
\caption{The \textsc{Hi} emission, \textsc{Hi} absorption, and the $^{13}$CO spectra of G15 from SGPS and GRS data. The bottom panel shows the relationship between the radial velocity and the distance.}
\label{fig:sgps}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\textwidth]{ms0117fig3.eps}
\caption{The \textsc{Hi} emission, \textsc{Hi} absorption, and the $^{13}$CO spectra of G15 from THOR and GRS data.}
\label{fig:thor}
\end{figure}
\label{lastpage}
\bibliographystyle{raa}
|
\section{#1}}
\def\rlap 1\mkern4mu{\rm l}{\rlap 1\mkern4mu{\rm l}}
\def\coeff#1#2{\relax{\textstyle {#1 \over #2}}\displaystyle}
\def\displaystyle{\displaystyle}
\def\overline{N}{\overline{N}}
\def\overline{P}{\overline{P}}
\def\overline{Q}{\overline{Q}}
\def\mathbb{C}{\mathbb{C}}
\def\mathbb{P}{\mathbb{P}}
\def\mathbb{R}{\mathbb{R}}
\def\mathbb{Z}{\mathbb{Z}}
\def\Pi{\Pi}
\def{\cals A}{{\cal A}}
\def{\cal B}{{\cal B}}
\def{\cal D}{{\cal D}}
\def{\cal F}{{\cal F}}
\def{\cal I}{{\cal I}}
\def{\cal K}{{\cal K}}
\def{\cal M}{{\cal M}}
\def{\cal N}{{\cal N}}
\def{\cal O}{{\cal O}}
\def{\cal P}{{\cal P}}
\def{\cal Q}{{\cal Q}}
\def{\cal R}{{\cal R}}
\def{\cal S}{{\cal S}}
\def{\cal T}{{\cal T}}
\def{\mathcal V}{{\mathcal V}}
\def\Neql#1{{\cal N}\!=\!{#1}}
\def\deql#1{d\!=\!{#1}}
\def\epsilon{\epsilon}
\def\delta{\delta}
\def={=}
\def\epsilon{\epsilon}
\def\delta{\delta}
\def={=}
\def\Omega{\Omega}
\def\lambda{\lambda}
\def\langle{\langle}
\def\rangle{\rangle}
\def\alpha{\alpha}
\def{\rm Tr}\,{{\rm Tr}\,}
\def\mathbb{V}{\mathbb{V}}
\def\mathbb{X}{\mathbb{X}}
\def\mathbb{T}{\mathbb{T}}
\def{\cal C}{{\cal C}}
\def\mathbb{C}{\mathbb{C}}
\def\mathbb{R}{\mathbb{R}}
\def\mathbb{P}{\mathbb{P}}
\def{\cal W}{{\cal W}}
\def{\cal L}{{\cal L}}
\def{\bar\zeta}{{\bar\zeta}}
\def\overline\zeta {}{\overline\zeta {}}
\def{\omega}{{\omega}}
\def\leavevmode\hbox{$\rm {}^{TM}$}{\leavevmode\hbox{$\rm {}^{TM}$}}
\defz{z}
\def{\bar z}{{\bar z}}
\def{\zeta_2}{{\zeta_2}}
\def{\bar\zeta_2}{{\bar\zeta_2}}
\def{\zeta_3}{{\zeta_3}}
\def{\bar\zeta_3}{{\bar\zeta_3}}
\def\zeta_{23}{\zeta_{23}}
\def\bar\zeta_{23}{\bar\zeta_{23}}
\def{\rm GL}{{\rm GL}}
\def{\rm SL}{{\rm SL}}
\def\CC{\rm H}{\mathbb{C}{\rm H}}
\defi\mspace{1mu}\CC{\rm H}{i\mspace{1mu}\mathbb{C}{\rm H}}
\def{\rm diag}{{\rm diag}}
\def{\rm O}{{\rm O}}
\def{\rm SO}{{\rm SO}}
\def{\rm U}{{\rm U}}
\def{\rm SU}{{\rm SU}}
\def{\rm SU}{{\rm SU}}
\def{\rm Sp}{{\rm Sp}}
\def\frak{sl}{\frak{sl}}
\def\mathfrak{sl}{\mathfrak{sl}}
\def\frak{so}{\frak{so}}
\def\frak{su}{\frak{su}}
\def\frak{gl}{\frak{gl}}
\def\frak{g}{\frak{g}}
\def\frak{h}{\frak{h}}
\def\frak{a}{\frak{a}}
\def\frak{t}{\frak{t}}
\def\frak{n}{\frak{n}}
\def\frak{k}{\frak{k}}
\defs_{\lambda/2}{s_{\lambda/2}}
\defc_{\lambda/2}{c_{\lambda/2}}
\def{s_{\lambda}}{{s_{\lambda}}}
\def{c_{\lambda}}{{c_{\lambda}}}
\def\cmnt#1{{\color{blue} \noindent [[\,\small KP: #1\,]]}}
\def\beq#1{{\bf \small {(#1)}}}
\def\bfs#1{{\boldsymbol #1}}
\def\cals#1{\mathcal{#1}}
\def{\cals A}{{\cals A}}
\def{\cals E}{{\cals E}}
\def\beqs#1{{\bf (#1)}}
\def\overset{_{\phantom{.}\circ}}{e}{}{\overset{_{\phantom{.}\circ}}{e}{}}
\def\overset{_{\phantom{.}\circ}}{D}{}{\overset{_{\phantom{.}\circ}}{D}{}}
\def\overset{_{\phantom{.}\circ}}{E}{}{\overset{_{\phantom{.}\circ}}{E}{}}
\def\overset{_{\circ}}{\Pi}{}{\overset{_{\circ}}{\Pi}{}}
\newcommand{\begin{equation}} \newcommand{\ee}{\end{equation}}{\begin{equation}} \newcommand{\ee}{\end{equation}}
\newcommand{\begin{equation} \begin{aligned}} \newcommand{\eea}{\end{aligned} \end{equation}}{\begin{equation} \begin{aligned}} \newcommand{\eea}{\end{aligned} \end{equation}}
\newcommand{\begin{multline}} \newcommand{\emu}{\end{multline}}{\begin{multline}} \newcommand{\emu}{\end{multline}}
\topmargin=-0.4in \oddsidemargin=-0.2in
\textheight=8.8in \textwidth=6.8in
\renewcommand{\baselinestretch}{1.2}
\newcommand{\textit{e.g.}}{\textit{e.g.}}
\newcommand{\textit{E.g.}}{\textit{E.g.}}
\newcommand{\textit{et al.}}{\textit{et al.}}
\newcommand{\textit{i.e.}}{\textit{i.e.}}
\newcommand{\textit{\`a la}}{\textit{\`a la}}
\newcommand{\textit{etc}\dots}{\textit{etc}\dots}
\newcommand\equ[1] {\begin{equation}#1\end{equation}}
\newcommand\equn[1] {\begin{equation*}#1\end{equation*}}
\newcommand\eqs[1] {\begin{align}#1\end{align}}
\newcommand\eqss[1] {\begin{align}\begin{split}#1\end{split}\end{align}}
\newcommand\eqsn[1] {\begin{align*}#1\end{align*}}
\renewcommand\( {\left(}
\renewcommand\) {\right)}
\newcommand{\mat}[1]{\begin{pmatrix} #1 \end{pmatrix}}
\newcommand{d\mathrm{vol}}{d\mathrm{vol}}
\newcommand{\mathrm{vol}}{\mathrm{vol}}
\usepackage[cmtip,all]{xy}
\newcommand{\xymatrix{{}\ar@{~>}[r]&{}}}{\xymatrix{{}\ar@{~>}[r]&{}}}
\begin{document}
\begin{titlepage}
\medskip
\begin{center}
{\Large \bf A Universal Counting of Black Hole Microstates in AdS$_4$}
\bigskip
\bigskip
\bigskip
\bigskip
\bigskip
{\bf Francesco Azzurli,${}^{(1)}$ Nikolay Bobev,${}^{(2)}$ P. Marcos Crichigno,${}^{(3)}$ Vincent S. Min,${}^{(2)}$ and Alberto Zaffaroni${}^{(1)}$ \\ }
\bigskip
\bigskip
${}^{(1)}$ Dipartimento di Fisica, Universit\`a di Milano-Bicocca, I-20126 Milano, Italy\\
INFN, sezione di Milano-Bicocca, I-20126 Milano, Italy\\
\vskip 5mm
${}^{(2)}$
Instituut voor Theoretische Fysica, KU Leuven \\
Celestijnenlaan 200D, B-3001 Leuven, Belgium
\vskip 5mm
${}^{(3)}$ Institute for Theoretical Physics, University of Amsterdam,\\
Science Park 904, Postbus 94485, 1090 GL, Amsterdam, The Netherlands
\vskip 5mm
\texttt{<EMAIL>,<EMAIL>,<EMAIL>,\\~vincent.<EMAIL>,<EMAIL>} \\
\end{center}
\bigskip
\bigskip
\begin{abstract}
\noindent
\end{abstract}
\noindent
Many three-dimensional $\mathcal{N}=2$ SCFTs admit a universal partial topological twist when placed on hyperbolic Riemann surfaces. We exploit this fact to derive a universal formula which relates the planar limit of the topologically twisted index of these SCFTs and their three-sphere partition function. We then utilize this to account for the entropy of a large class of supersymmetric asymptotically AdS$_4$ magnetically charged black holes in M-theory and massive type IIA string theory. In this context we also discuss novel AdS$_2$ solutions of eleven-dimensional supergravity which describe the near horizon region of large new families of supersymmetric black holes arising from M2-branes wrapping Riemann surfaces.
\end{titlepage}
\setcounter{tocdepth}{2}
\tableofcontents
\newpage
\section{Introduction}
String theory offers a unique vantage point into the physics of black holes. One the one hand it gives rise to supergravity theories in ten and eleven dimensions as its low energy limits and these theories admit a plethora of classical black hole (or more generally black brane) backgrounds. These black holes are sourced by the massive strings and branes, the microscopic degrees of freedom in the theory. On the other hand one can understand the low energy physics of these branes in terms of the quantum field theory living on their worldvolume. This dichotomy has been the source of many advances in our understanding of black holes, as well as being ultimately responsible for the concrete realization of the holographic principle in string theory. A famous example of this success story is the microscopic counting of the entropy for a class of asymptotically flat black holes in string theory \cite{Strominger:1996sh} (see also \cite{Sen:1995in} for earlier important work). However, this perspective on black holes has not been put to good use in the context of asymptotically AdS black holes. This state of affairs changed recently with the results in \cite{Benini:2015eyy}. The authors of \cite{Benini:2015eyy} employed a certain supersymmetric index, $Z_{S^1\times \Sigma_{\mathfrak{g}}}$, defined in \cite{Benini:2015noa} to account for the entropy of a large class of supersymmetric asymptotically AdS$_4$ black holes arising from M2-branes wrapped on Riemann surfaces. These results were later extended to dyonic black holes in \cite{Benini:2016rke}. See also \cite{Hosseini:2016tor,Hosseini:2016ume,Benini:2016hjo,Closset:2016arn,Hosseini:2016cyf,Cabo-Bizet:2017jsl} for further recent developments.
Our goal in this work is to build upon this recent success in a number of ways. First, we show that there is a universal expression for the topologically twisted index in the large $N$ limit for a large class of $\mathcal{N}=2$ three-dimensional SCFTs.\footnote{The conditions on the $\mathcal{N}=2$ theory are discussed in detail in Section \ref{sec:fieldtheory}.} Namely, we find that the index is related to the three-sphere free energy, $F_{S^3}$, of the three-dimensional theory through the simple relation
\begin{equation}\label{eq:univintro}
\log Z_{S^1\times \Sigma_{\mathfrak{g}}} = (\mathfrak{g}-1) \, F_{S^3}\;.
\end{equation}
The universality of this relation stems from a universal partial topological twist that is used in the definition of the twisted index. In general the index depends on a set of background fluxes for the global symmetries of the $\mathcal{N}=2$ SCFT as well as a set of complex fugacities. A given choice of Riemann surface and background fluxes represents a particular topological twist of the three-dimensional theory on $\Sigma_\frak{g}$. It was argued in \cite{Benini:2015eyy} that for a given such choice of twist one obtains the degeneracy of vacua after extremizing the index as a function of the complex fugacities, and this extremal value appears in \eqref{eq:univintro}. The universal topological twist, for which the relation in \eqref{eq:univintro} is valid, is singled out by choosing the background fluxes on $\Sigma_\frak{g}$ such that the only non-zero flux is in the direction of the exact superconformal R-symmetry of the three-dimensional $\mathcal{N}=2$ SCFT.
Such a simple universal relation in QFT should have an equally simple bulk realization for three-dimensional $\mathcal{N}=2$ SCFTs with a holographic dual. Indeed, we show that there is a simple black hole solution with a hyperbolic horizon in minimal four-dimensional gauged supergravity which provides the holographic realization of the relation \eqref{eq:univintro}. In fact, the entropy of this black hole, originally found in \cite{Romans:1991nq} and later studied further in \cite{Caldarelli:1998hg}, in the large $N$ limit is simply given by the twisted index and thus we arrive at a microscopic understanding of the entropy for this simple universal black hole. An important point in our story is the fact that this simple black hole solution can be embedded in eleven-dimensional and massive type IIA supergravity in infinitely many ways. This provides further evidence that the relation \eqref{eq:univintro} holds (at least in the planar limit) for a large class of three-dimensional $\mathcal{N}=2$ SCFTs.
In the course of our analysis, we also clarify the relation proposed in \cite{Benini:2015eyy} between the topologically twisted index and the entropy of the black hole. According to the holographic dictionary the logarithm of the partition function of the CFT in the large $N$ limit should correspond to the on-shell action of the dual gravitational background. We show that, for the universal black hole, the on-shell action indeed coincides with (minus) the entropy. This computation is subtle and it should be done by considering a non-extremal deformation of the black hole and carefully
taking the extremal limit.
The string and M-theory embedding of the universal black hole can be interpreted as arising from D2- or M2-branes wrapped on $\Sigma_\frak{g}$. As pointed out in \cite{Bershadsky:1995qy} these wrapped branes constructions implement the partial topological twist for the QFT which describes the low-energy dynamics of the branes. For a large number of branes one has to work in the supergravity limit in which one typically finds a black hole solution with a near horizon AdS$_2\times \Sigma_\frak{g}$ geometry. This holographic construction was first employed for D3- and M5-branes in \cite{Maldacena:2000mw} and later generalized to M2-branes in \cite{Gauntlett:2001qs}.\footnote{For a review on branes wrapped on calibrated cycles and further references see \cite{Gauntlett:2003di}.} This prompts us to seek generalization of the simple black hole solutions in M-theory. Due to the complicated nature of the BPS supergravity equations in eleven-dimensional supergravity we focus only on finding explicit solutions for the AdS$_2$ near horizon region of the black hole. Employing this approach we find a large class of analytic explicit AdS$_2$ solutions of the eleven-dimensional supersymmetry equations with non-trivial internal fluxes. The full black hole solution for which these backgrounds are a near-horizon limit should correspond to turning on background magnetic fluxes for the global non-R symmetry in the dual CFT. The entropy of the black hole should then be captured by the twisted topological index with the given background global symmetry fluxes. While we do not establish fully this holographic picture we provide some evidence for its consistency. We should point out that the explicit AdS$_2$ solutions which we find fall into the general classification of such backgrounds in M-theory discussed in \cite{Kim:2006qu,Gauntlett:2007ts}. Other examples of AdS$_2$ vacua in eleven-dimensional supergravity of the class discussed here can be found in \cite{Gauntlett:2006ns,Kim:2012ek,Donos:2012sy,Kim:2013xza}.
We begin our story in the next section with a discussion of a universal partial topological twist that can be applied to three-dimensional $\mathcal{N}=2$ SCFTs and we argue how for a large class of these twisted theories one can calculate the topologically twisted index of \cite{Benini:2015noa} in the planar limit. We then proceed in Section \ref{sec:A simple 4d black hole} with a holographic description of the RG flow induced by the topological twist, which is realized by a black hole in four-dimensional minimal gauged supergravity. In addition, we show how to embed this black hole in various ways in M-theory and massive type IIA string theory. In Section \ref{AdS2Hor} we present a large class of AdS$_2$ vacua of eleven-dimensional supergravity which can be viewed as near horizon geometries of black holes constructed out of wrapped M2-branes. We conclude with some comments and a discussion on interesting avenues for future research in Section \ref{sec:discussion}. In the three appendices we collect various technical details used in the main text. In Appendix \ref{app:fieldIIA} we provide some details on the calculation of the topologically twisted index for three-dimensional $\mathcal{N}=2$ SCFTs with massive IIA holographic duals. Appendices \ref{sec:MIIA} and \ref{Appendix11D} contain details on the construction of our massive IIA and eleven-dimensional supergravity solutions.
\textit{Note added:} While we were preparing this manuscript the papers \cite{Guarino:2017pkw,Guarino:2017eag} appeared. In them the authors find a supersymmetric AdS$_4$ black hole solution, with a near horizon AdS$_2\times \Sigma_\frak{g}$ geometry, in the four-dimensional maximal $ISO(7)$ gauged supergravity. We believe that upon an uplift to massive IIA supergravity this black hole is the same as the universal black hole discussed in Section \ref{sec:upMIIA} below. Soon after this paper appeared on the arXiv other supersymmetric AdS$_4$ black hole solutions in massive type IIA supergravity were studied in \cite{Hosseini:2017fjo,Benini:2017oxt} and their microstates were counted using the topologically twisted index.
\section{A universal twist of 3d $\mathcal{N}=2$ SCFTs}
\label{sec:fieldtheory}
We consider $\mathcal{N}=2$ superconformal field theories compactified on a Riemann surface $\Sigma_\frak{g}$ of genus $\frak{g}$ with a topological twist. This is implemented by turning on a non-trivial background for the R-symmetry of the theory. More precisely, there is a magnetic flux on $\Sigma_\frak{g}$ for the background gauge field $A_R$ coupled to the R-symmetry current such that $\int_{\Sigma_\frak{g}} d A_R=2\pi (1- \frak{g})$. This condition ensures that the R-symmetry background field precisely cancels the spin connection on $\Sigma_\frak{g}$ and the resulting theory preserves two real supercharges. In general, since the R-symmetry can mix with global symmetries, the choice of $A_R$ is not unique and there is a family of different twists parametrized by the freedom to turn on magnetic fluxes along the Cartan subgroup of the continuous global flavor symmetry group.\footnote{By a flavor symmetry here we mean any global symmetry of the $\mathcal{N}=2$ SCFT which is not the superconformal $U(1)$ R-symmetry. Later on we will make a distinction between different types of non-R symmetries.} On the other hand, in a superconformal field theory there is an exact R-symmetry that is singled out by the superconformal algebra and can be fixed uniquely by performing $F$-maximization \cite{Jafferis:2010un,Closset:2012vg}. Following the terminology in \cite{Benini:2015bwz,BC}, we refer to a partial topological twist of the three-dimensional theory on $\Sigma_\frak{g}$ along the exact superconformal R-symmetry as a \textit{universal twist}.
The degeneracy of ground states of the compactified theory after this partial topological twist can be extracted from the topologically twisted index $Z$, which is defined as the twisted supersymmetric partition function on $\Sigma_\frak{g}\times S^1$ \cite{Benini:2015noa,Benini:2016hjo,Closset:2016arn}. Let us briefly summarize some of the salient features of the topologically twisted index.
The index depends on a set of integer magnetic fluxes $\mathfrak{n}= \frac{1}{2\pi}\int_{\Sigma_\frak{g}} F^{{\rm flav}}$ for the Cartan generators of the flavor symmetry group,
parameterizing the inequivalent twists. A convenient parameterization for the fluxes is the following. We can assign
a magnetic flux $\mathfrak{n}_I$ to each chiral field $\Phi_I$ in the theory with the constraint that, for each term $W_a$ in the superpotential,\footnote{In this paper we restrict to three-dimensional $\mathcal{N}=2$ SCFTs which admit a Lagrangian description, however we suspect that the universal twist relation \eqref{entropyfree} below is valid more generally.}
\begin{equation}\label{fluxes}
\sum_{I\in W_a} \frak{n}_I = 2(1-\frak{g}) \, .
\end{equation}
Since the superpotential has R-charge $2$, this condition ensures that $\int_{\Sigma_\frak{g}} d A_R=2\pi (1-\frak{g})$ and supersymmetry is preserved. The Dirac quantization condition further restricts the flux parameters $\frak{n}_I$ to be
integer.
The index also depends on a set of complex fugacities $y$ for the flavor symmetries. We can again assign a complex number $y_I$ to each chiral field $\Phi_I$ in the theory with the constraint that, for each term $W_a$ in the superpotential,
\begin{equation}\label{fugacities}
\prod_{I\in W_a} y_I = 1 \, .
\end{equation}
It will be important also to consider the complexified chemical potentials $\Delta_I$, defined by $y_I \equiv e^{i \Delta_I}$. Notice that the chemical potentials are periodic, $\Delta_I\sim \Delta_I +2 \pi$. Therefore (\ref{fugacities}) becomes
\begin{equation}\label{chemical0}
\sum_{I\in W_a} \Delta_I \in 2 \pi \, \mathbb{Z} \,.
\end{equation}
Using this periodicity we can always choose $0\le {\rm Re} \, \Delta_I \le 2\pi$. With this at hand, (\ref{chemical0}) implies that $ \sum_{I\in W_a} \Delta_I \ne 0$ unless all $\Delta_I=0$. This will be important in the discussion below.
The topologically twisted index can be evaluated by localization and reduced to a matrix model \cite{Benini:2015noa}. The large $N$ limit of the matrix model has been analyzed in \cite{Benini:2015eyy} for the ABJM theory \cite{Aharony:2008ug} and generalized to other classes of quiver gauge theories with M-theory or massive type IIA duals in \cite{Hosseini:2016tor,Hosseini:2016ume}.
The results of this analysis is surprisingly simple. One finds a consistent large $N$ solution of the matrix model only when
\begin{equation}\label{chemical}
\sum_{I\in W_a} \Delta_I = 2 \pi \;,
\end{equation}
%
for each term $W_a$ in the superpotential. Under this condition, the logarithm of the topologically twisted index is given by\footnote{The formula appears in \cite{Hosseini:2016tor} only for $\Sigma_\frak{g}=S^2$. The generalization to arbitrary $\frak{g}$ is straightforward and discussed in details for ABJM in \cite{Benini:2016hjo}: the general rule is simply $\log Z_{\frak{g}}= (1-\frak{g}) \log Z_{\frak{g}=0} (\frak{n}_I/(1-\frak{g}))\,$.}
%
\begin{equation}\label{index}
\log Z (\Delta_I,\frak{n}_I) = (1-\frak{g}) i \left ( \frac{2}{\pi} {\mathcal V} (\Delta_I) + \sum_I \left [ \( \frac{\frak{n}_I}{1-\frak{g}} - \frac{\Delta_I}{\pi} \) \frac{\partial {\mathcal V}}{\partial\Delta_I} \right ] \right )\;,
\end{equation}
where the function ${\mathcal V}$ was called Bethe potential in \cite{Benini:2015noa} and is the Yang-Yang function of an associated integrable system \cite{Nekrasov:2014xaa,Closset:2016arn}.\footnote{What we call $ {\mathcal V} (\Delta_I)$ here is the extremal value with respect to the eigenvalues $u_i$ of the Bethe functional ${\mathcal V} [\Delta_I, u_i]$, defined in \cite{Benini:2015eyy,Hosseini:2016tor,Hosseini:2016ume}.} Quite remarkably, in the large $N$ limit, the Bethe potential ${\mathcal V}$ is related to the free energy on $S^3$ of the three-dimensional $\mathcal{N}=2$ theory \cite{Hosseini:2016tor} through the simple identity
\begin{equation}
\label{indexS3}
- \frac{2 i}{\pi} {\mathcal V}(\Delta_I) = F_{S^3} \left (\frac{\Delta_I}{\pi}\right ) \, .
\end{equation}
This relation might look puzzling at first sight and deserves some comments. Recall that the free energy $F_{S^3}$ is a function of a set of trial R-charges that parameterize a family of supersymmetric Lagrangians on $S^3$ \cite{Jafferis:2010un,Hama:2010av}. The importance of this functional is that its extremization gives the exact R-charges of the theory \cite{Jafferis:2010un}. In \eqref{indexS3}, ${\mathcal V}$ is a function of chemical potentials for the flavor symmetries while $F_{S^3}$ is a function of R-charges. However, although the $\Delta_I$ parameterize flavor symmetries in the three-dimensional theory on $\Sigma_\frak{g}\times S^1$, the relation \eqref{chemical}
ensures that $\Delta_I/\pi$ can be consistently identified with a set of R-charges for the theory on $S^3$.
There is a subtlety that arises when computing the topologically twisted index or the three-sphere free energy. In three-dimensional $\mathcal{N}=2$ SCFTs there are finite counterterms which affect the imaginary part of the complex function $F_{S^3}$. These are given by Chern-Simons terms with purely imaginary coefficients for the background gauge fields that couple to conserved currents, see \cite{Closset:2012vg,Closset:2012vp} for a detailed discussion. Moreover, the imaginary part of $\log Z$ is only defined modulo $2\pi$ and is effectively $O(1)$ in the large $N$ limit. The upshot of this discussion is that the physically unambiguous quantity in the large $N$ limit are the real parts of the topologically twisted index and the free energy on $S^3$.
There is an additional important point in the story. To obtain the degeneracy of vacua of the compactified theory for a given choice of the flux parameters, $\frak{n}_I$, one has to extremize the function $Z (\Delta_I,\frak{n}_I)$ with respect to the fugacities $\Delta_I$ \cite{Benini:2015eyy}. This prescription is analogous to the extremization principles that exist in four \cite{Intriligator:2003jj}, three \cite{Jafferis:2010un}, and two-dimensional \cite{Benini:2012cz,Benini:2013cda} SCFTs with an Abelian R-symmetry.
After this short introduction to the topologically twisted index we are ready to discuss the universal twist. This is obtained by choosing fluxes $\frak{n}_I$ proportional to the exact UV R-charges $\bar \Delta_I$ and these, as we already mentioned, can be found by extremizing $F_{S^3}$ \cite{Jafferis:2010un}. From the identification (\ref{indexS3}) it is clear that the Bethe potential ${\mathcal V}$ is also extremized at the values $\bar \Delta_I$. Given the normalizations (\ref{fluxes}) and (\ref{chemical}), we find that the universal twist is determined by
\begin{equation}
\label{ident}
\frac{\bar \frak{n}_I}{1-\frak{g}} = \frac{\bar \Delta_I}{\pi} \, .
\end{equation}
It is easy to see that, for this choice of fluxes, $ \log Z$ in (\ref{index}) is also extremized at $\Delta_I=\bar\Delta_I$:
\begin{equation}
\frac{\partial \log Z}{\partial \Delta_I} \Big |_{\bar \Delta_I} =0 \, ,
\end{equation}
since $\partial {\mathcal V} (\bar \Delta_I) =0$.\footnote{We note in passing that if one imposes $\frac{ \frak{n}_I}{1-\frak{g}} = \frac{\Delta_I}{\pi}$ before extremizing $F_{S^3}$ then the second term in \eqref{index} vanishes. Thus, after using \eqref{indexS3} we find that the extremization of $F_{S^3}$ is the same as the extremization of the topologically twisted index.} The value at the extremum reads
\begin{equation}\label{entropyfree}
\log Z (\bar \Delta_I,\bar\frak{n}_I) = (\frak{g}-1) F_{S^3} \left (\frac{\bar \Delta_I}{\pi}\right ) \, ,
\end{equation}
where we made use of \eqref{indexS3} and \eqref{ident}.
Equation \eqref{entropyfree} amounts to a universal relation between the value of the index of the universal twist of a 3d $\mathcal{N}=2$ SCFT and the free energy on $S^3$ of the same superconformal theory in the planar limit. For theories with a weakly coupled holographic dual this identity should translate into a universal relation between the entropy of some universal black hole solution and the AdS$_4$ supersymmetric free energy. As we discuss in detail in section~\ref{sec:A simple 4d black hole} this expectation indeed bears out for the case of hyperbolic Riemann surfaces, i.e. for $\frak{g}>1$.
It is worth pointing out that the universal relation in \eqref{entropyfree} is the three-dimensional analog of the universal relation among central charges established in \cite{Benini:2015bwz} for twisted compactifications of four-dimensional ${\cal N}=1$ SCFTs on $\mathbb{R}^2\times \Sigma_{\frak{g}}$.\footnote{Similar relations can be established for SCFTs with a continuous R-symmetry in various dimensions and with different amount of supersymmetry \cite{BC}.} A notable difference is that for four-dimensional SCFTs the universal relation can be established at finite $N$. It would be most interesting to study subleading, i.e. non-planar, corrections to the universal relation in \eqref{entropyfree}.\footnote{It is interesting to note that similar relations between $F_{S^{3}}$ and partition functions on $S^{1}\times \Sigma_{\frak{g}}$ with line operator insertions were discussed in \cite{Closset:2017zgf}.}
\subsection{M-theory and massive type IIA models}
The derivation of (\ref{entropyfree}) is based on the large $N$ identities (\ref{index}) and (\ref{indexS3}), which in turn can be established for a large class of Yang-Mills-Chern-Simons theories with fundamental and bi-fundamental chiral fields with M-theory or massive type IIA duals. We now discuss the class of theories for which these identities are valid.
Consider first superconformal theories dual to M-theory on AdS$_4\times$ SE$_7$, where SE$_7$ is a Sasaki-Einstein manifold. Many quivers describing such theories have been proposed in the literature. Most of them are obtained by dimensionally reducing a ``parent'' four-dimensional quiver gauge theory with bi-fundamentals and adjoints with an AdS$_5\times$ SE$_5$ dual, and then adding Chern-Simons terms and flavoring with fundamentals. In such theories, the sum of all Chern-Simons levels is zero, $\sum_a k_a=0$. Holography predicts that the $S^3$ free energy and the twisted index of such theories scales as $N^{3/2}$ for $N\gg k_a$. The large $N$ behavior of the $S^3$ free energy has been computed in \cite{Jafferis:2011zi} and successfully compared with the holographic predictions only for a particular class of quivers. In particular, for the method in \cite{Jafferis:2011zi} to work, the bi-fundamental fields must transform in a real representation of the gauge group and the total number of fundamentals must be equal to the total number of anti-fundamentals. It turns out that, under the same conditions, the topologically twisted index scales like $N^{3/2}$ and the identities (\ref{index}) and (\ref{indexS3}) are valid \cite{Hosseini:2016tor}.
This particular class of quivers include all the vector-like examples in \cite{Hanany:2008fj, Hanany:2008cd, Martelli:2008si} and many of the flavored theories in \cite{Gaiotto:2009tk,Benini:2009qs}. In particular, the latter includes the dual of AdS$_4\times Q^{1,1,1}/\mathbb{Z}_k$ and AdS$_4\times N^{0,1,0}/\mathbb{Z}_k$. The conditions are also satisfied for the $\mathcal{N}=3$ necklace and $\widehat D$- and $\widehat E$-type quivers \cite{Jafferis:2008qz,Herzog:2010hf,Gulotta:2011vp,Crichigno:2012sk,Crichigno:2017rqg}, as well as the quiver for the non-toric manifold $V^{5,2}$ discussed in \cite{Martelli:2009ga}. The evaluation of the index in the large $N$ limit for most of these examples is given in \cite{Hosseini:2016ume}, where the identities (\ref{index}) and (\ref{indexS3}) are also verified by explicit computation. For the ``chiral'' theories discussed in \cite{Hanany:2008fj, Hanany:2008cd, Martelli:2008si}, on the other hand, it is not known how to properly take the large $N$ limit in the matrix model to obtain the correct scaling predicted by holography. This applies both for the topologically twisted index and for the $S^3$ partition function. This class of quivers includes interesting models, like the quiver for $M^{1,1,1}$ proposed in \cite{Martelli:2008si} and further studied in \cite{Benini:2011cma}.
Consider now superconformal theories dual to warped AdS$_4\times Y_6$ flux vacua of massive type IIA. A well-known example is the $\mathcal{N}=2$ $U(N)$ gauge theory with three adjoint
multiplets and a Chern-Simons coupling $k$ described in \cite{Guarino:2015jca}. It corresponds to
an internal manifold $Y_6$ with the topology of $S^6$. This has been generalised in \cite{Fluder:2015eoa} to the case where $Y_6$ is an $S^2$ fibration over a general K\"ahler-Einstein manifold KE$_4$.
The dual field theory is obtained by considering the four-dimensional theory dual to AdS$_5\times$ SE$_5$, where SE$_5$ is the five-dimensional Sasaki-Einstein
with local base KE$_4$, reducing it to three dimensions and adding a Chern-Simons term with level $k$ for all gauge groups.\footnote{The original example in \cite{Guarino:2015jca} has KE$_4=\Bbb{CP}_2$, SE$_5=S^5$, and the superconformal theory is obtained by reducing $\mathcal{N}=4$ SYM to three-dimensions and adding a Chern-Simons coupling with level $k$. The solution in \cite{Fluder:2015eoa}
is obtained by replacing $\Bbb{CP}_2$ with more general KE$_4$ manifolds.} The large $N$ limit at fixed $k$ of the $S^3$ free energy has been computed in \cite{Guarino:2015jca} and \cite{Fluder:2015eoa} and shown to scale as $N^{5/3}$, as predicted by holography. The large $N$ limit of the topologically twisted index
has been computed in \cite{Hosseini:2016tor}. The identities (\ref{index}) and (\ref{indexS3}) also hold for massive type IIA. The explicit derivation was not reported in \cite{Hosseini:2016tor}
and is given in Appendix \ref{app:fieldIIA}. Notice that for massive type IIA quivers there is no need to restrict to vector-like models.
The discussion above shows that there is a large number of three-dimensional superconformal theories with M-theory or massive type IIA duals for which relation (\ref{entropyfree}) formally holds.
However, it is important to notice that not all of them really admit a universal twist since we need to restrict ourselves to $\mathcal{N}=2$ SCFTs with rational R-charges. Indeed, since $\frak{n}_I$ and $\frak{g}$ are integers, the relation in (\ref{ident}) implies that the exact R-charge of the fields must be rational. This slightly restricts the class of theories where we can perform the universal twist. However, we can still find infinitely many models for which the R-charges are rational and the universal twist exists. Since $\mathcal{N}=3$ theories necessarily have rational R-charges this applies to the $\mathcal{N}=3$ necklace quivers \cite{Jafferis:2008qz,Herzog:2010hf} as well as the $\widehat D$- and $\widehat E$-type quivers \cite{Gulotta:2011vp,Crichigno:2012sk,Crichigno:2017rqg}, and the $\mathcal{N}=3$ quiver for $N^{0,1,0}$ \cite{Gaiotto:2009tk,Imamura:2011uj,Cheon:2011th}. In addition, one can check that the $\mathcal{N}=2$ quivers for $Q^{1,1,1}$ \cite{Benini:2009qs} and $V^{5,2}$ \cite{Martelli:2009ga} in M-theory have rational R-charges. The same holds for the theory in \cite{Guarino:2015jca}
and some of its generalizations in massive type IIA.\footnote{The R-charges for M-theory vacua can be computed using volume minimization \cite{Martelli:2006yb} and for massive type IIA
by $a$-maximization \cite{Intriligator:2003jj}. The result is generically irrational. However, we can easily find special classes of SE$_7$ or SE$_5$ where the result is rational.}
\section{A simple 4d black hole}
\label{sec:A simple 4d black hole}
Here we provide the holographic description of the universal twisted compactification of 3d ${\cal N}=2$ SCFTs discussed in the previous section. As we shall show, this corresponds to the supersymmetric magnetically charged AdS$_{4}$ black hole of \cite{Romans:1991nq,Caldarelli:1998hg}, thus providing the appropriate field theory interpretation of this solution. We also review the known uplift of this solution to M-theory and provide a new uplift to massive IIA supergravity.
The appropriate supergravity is 4d minimal ${\cal N} = 2$ gauged supergravity \cite{Freedman:1976aw}, with eight supercharges and bosonic content the graviton and an $SO(2)$ gauge field, dual to the stress energy tensor and R-symmetry current, respectively. The bosonic action reads\footnote{Here we follow the conventions in \cite{Benini:2015eyy} and truncate the ${\cal N}=8$ supergravity to the minimal one by setting $L_a=1$ ($\phi_{ij}=0$), $A_a=A$, and set the coupling constant $g=1/\sqrt{2}$ which in turn amounts to setting the radius of the AdS$_{4}$ vacuum to one.}
\begin{equation}\label{eq:4daction}
I= \frac{1}{16\pi G_N^{(4)}} \int d^4x \sqrt{-g} \(R +6 -\frac14 F^{2}\) \,,
\end{equation}
with $G_N^{(4)}$ the 4d Newton constant. We have chosen the cosmological constant such that the AdS$_4$ vacuum of the theory has $R_{\text{AdS}_{4}}=-12$ and $L_{\text{AdS}_{4}}=1$. The magnetically charged black hole solution of \cite{Romans:1991nq,Caldarelli:1998hg} preserves two supercharges and is given by\footnote{A generalization to include rotation while maintaining supersymmetry was also found in these references.}
\eqss{ \label{univ}
ds_4^2 &= -\left ( \rho -\frac{1}{2\rho}\right )^2 dt^2 + \left ( \rho -\frac{1}{2\rho}\right )^{-2} d\rho^2 + \rho^2\, ds_{\Bbb H^2}^2 \,, \\
F&= \frac{dx_1 \wedge dx_2}{x_2^2} \,,
}
where $ds_{\Bbb H^2}^2 = \frac{1}{x_2^2}\left(dx_1^2+dx_2^2\right) $ is the local constant-curvature metric on a Riemann surface $\Sigma_\frak{g}$ of genus $\frak{g}>1$,\footnote{As discussed in \cite{Caldarelli:1998hg}, there is no supersymmetric static black hole solution with $\frak{g}=0,1$.} normalized such that $R_{\Bbb H^2}=-2$. Using the Gauss-Bonnet theorem the volume of the Riemann surface is then ${\rm vol}(\Sigma_\frak{g})=4 \pi(\frak{g}-1)$. Dirac quantization of the flux requires $\frac{1}{2\pi}\int_{\Sigma_{\frak{g}}}F\in \Bbb Z$, which holds for any genus $\frak{g}$. We note the solution has a fixed magnetic charge, set by supersymmetry. The entropy of this extremal black hole is given in terms of the horizon area by the standard Bekenstein-Hawking formula
\begin{equation}\label{eq:SBH}
S_\text{BH} = \frac{\text{Area}}{4 G_N^{(4)}} = \frac{(\frak{g}-1)\,\pi}{2 G_N^{(4)}} \, .
\end{equation}
As will soon become clear the large $N$ limit of the topologically twisted index reproduces exactly this entropy. However, there is a slight subtlety in this story. According to the standard holographic dictionary the logarithm of the partition function of the CFT (in the appropriate large $N$ limit) should correspond to a properly regularized on-shell action of the dual gravitational background, rather than the black hole entropy. In the next section we clarify this relation, showing that the black hole entropy in \eqref{eq:SBH} is indeed closely related to the on-shell action and thus to the topologically twisted index.
\subsection{On-shell action}
We are interested in calculating the value of the Euclideanized action \eqref{eq:4daction}, evaluated on-shell for the solution \eqref{univ}. This action is divergent unless properly regularized by counterterms, following the standard holographic renormalization prescription, which we carry out explicitly next. As we show, this is intimately related to the entropy of the black hole. The regularized action we consider is given by:\footnote{See for example Section 4.4 of \cite{Martelli:2012sz} as a reference for the counterterms. With respect to their normalizations we have $F^\text{there} = F^\text{here}/2$.}
\begin{equation}\label{Ireg}
\begin{split}
I_\text{Eucl} =\,& I_\text{Einst+Max} + I_\text{ct+bdry} \,, \\
I_\text{Einst+Max} =\,&-\frac{1}{16\pi G_N^{(4)}} \int d^4 x \sqrt{g} \left(R+6-\frac{1}{4} F^2 \right) \\
I_\text{ct+bdry} = \,&\frac{1}{8\pi G_N^{(4)}} \int d^3x \sqrt{\gamma} \left(2+\frac{1}{2}R(\gamma)-K\right) \, ,
\end{split}
\end{equation}
where $\gamma$ is the induced metric on the boundary, defined by the radial cutoff $\rho=r$, and $K$ is the trace of the extrinsic curvature of this boundary metric. Taking $r\rightarrow \infty$ leads to a divergence in $I_\text{Einst+Max}$, which is cancelled by $I_\text{ct+bdry}$, where we have collected the appropriate counterterms as well as boundary terms necessary for a well-defined variational principle. An important additional subtlety arises, however, in the explicit evaluation of $I_\text{Eucl}$ for the extremal solution \eqref{univ}. Indeed, it is easy to see that this integral is naively not well defined, as the integrand of $I_\text{Eucl}$ for this solution vanishes, while the integration over Euclidean time $\int_{0}^{\infty} d\tau$ leads to infinity as a consequence of the solution being extremal and thus $T=0$. To obtain the correct finite result we thus consider a non-extremal deformation of the solution, compute $I_\text{Eucl}$ for the non-extremal solution and take the extremal limit at the end.
There are two non-extremal deformations. One amounts to allowing for a generic magnetic charge $Q$ under the graviphoton, and a second to adding a mass $\eta$. This non-extremal generalization was discussed in \cite{Romans:1991nq,Caldarelli:1998hg} and the solution reads
\eqss{\label{eq:nonextr}
ds^2 &= -V(\rho) \, dt^2 +\frac{1}{V(\rho)} \, d\rho^2 + \rho^2 \, ds_{\Bbb H^2}^2 \, ,\\
F &= 2Q \,\frac{dx_1 \wedge dx_2}{x_{2}^{2}}\,,
}
where
\begin{equation}
V(\rho) = -1-\frac{2\eta}{\rho} + \frac{Q^2}{\rho^2}+\rho^2 \, .
\end{equation}
The extremal solution is recovered for $Q\rightarrow 1/2$ and $\eta\rightarrow 0$. The horizon radius $\rho_0$ is obtained by solving the quartic equation $V(\rho_0)=0$. The temperature $T$ of the black hole can be obtained by requiring the Euclidean metric to be free of conical singularities, which gives
\begin{equation}
T = \frac{\left|V'(\rho_0)\right|}{4\pi} = \frac{1}{2\pi \rho_0} \left|\rho_0^2 + \frac{\eta}{\rho_0}-\frac{Q^2}{\rho_0^2}\right|\, .
\end{equation}
In the extremal limit $\rho_0 \rightarrow 1/\sqrt{2}$ and $T\rightarrow0$.
Evaluating the on-shell action \eqref{Ireg} for general values of $Q$ and $\eta$ we find
\begin{equation}\label{actionnonext}
I_\text{Eucl} = \frac{(\frak{g}-1)}{2 G_N^{(4)}}\frac{\beta}{\rho_0} \left(Q^2-\rho_0^4+\eta \rho_0\right) \, ,
\end{equation}
where $\beta=1/T$ is the Euclidean time periodicity. Taking the extremal limit of this expression gives the finite answer
\begin{equation}
I_\text{extr}= - \frac{ (\frak{g}-1)\,\pi}{2 G_N^{(4)}} + \mathcal{O}\left(\left(Q-1/2\right)^{1/2}\right)+\mathcal{O}\left(\eta^{1/2}\right) \, .
\end{equation}
Comparing this to the entropy \eqref{eq:SBH} of the extremal black hole, we thus have
\begin{equation}\label{eq:ISBH}
I_\text{extr}=-S_\text{BH} \, .
\end{equation}
On the other hand, the holographic dictionary relates the gravitational on-shell action to the partition function of the dual CFT as $I=-\log Z$, leading to the satisfying expression
\begin{equation}\label{eq:ident}
S_\text{BH}=\log Z \,.
\end{equation}
We have thus shown that the black hole entropy can be identified with the topologically twisted index of the dual CFT to leading order in $N$. This relation was argued to hold more generally for a class of black holes in non-minimal gauged supergravity in \cite{Benini:2015eyy}; it would be interesting to establish this explicitly by generalizing the computation above to this larger class of black holes. It would also be of interest to study these relations to subleading orders in $N$.
Finally, let us recall that the renormalized on-shell action for Euclidean AdS$_4$ with an $S^3$ boundary is given by \cite{Emparan:1999pm}\footnote{With slight abuse of notation we define the CFT free energy as $F=-\log Z=I$.}
\begin{equation}
F_{S^3} = \frac{\pi }{2 G_N^{(4)}} \, ,
\end{equation}
which using \eqref{eq:SBH} leads to the relation
\begin{equation}\label{eq:SBHFS3}
S_\text{BH} = (\frak{g}-1) \, F_{S^3} \, .
\end{equation}
With the identification \eqref{eq:ident} one recognizes this result as the holographic analog of the universal field theory relation \eqref{entropyfree}. Thus, the result \eqref{eq:SBHFS3} provides strong evidence that the AdS$_{4}$ black hole \eqref{univ} describes the RG flow from 3d ${\cal N}=2$ SCFTs with a universal topological twist on $\Sigma_{\frak{g}}$ to superconformal quantum mechanical theories with two supercharges.
We emphasize that although the black hole solution \eqref{univ} is derived as a supersymmetric solution of minimal gauged supergravity, it is also a solution to non-minimal gauged supergravity, with the additional vector and hyper multiplet bosonic fields set to zero.\footnote{This can be checked, for instance, by setting $\mathfrak n_{a}=\frac{\kappa}{2}\,, \vec \phi=0\,, L_{a}=1$ in the BPS equations (A.27) in \cite{Benini:2015eyy}.} The fact that ``freezing'' the vector multiplets to zero is consistent corresponds to the property of universality of the topological twist \eqref{ident}. In other words, as the vector multiplet scalars in the gauged supergravity are not sourced along the flow from AdS$_4$ to AdS$_2$ realized by the black hole \eqref{univ}, there is no mixing between R-symmetry and flavor symmetry along the flow and the R-symmetry in the IR coincides with the one in the UV. This is consistent with the field theory discussion of Section~\ref{sec:fieldtheory}.
\subsection{Uplift to M-theory}
\label{sec:MTheoryUplift}
Here we review the uplift of the universal solution in \eqref{univ} to eleven-dimensional supergravity, which was carried out in \cite{Gauntlett:2007ma}. The metric and four-form read\footnote{Compared to Equation (2.3) in \cite{Gauntlett:2007ma} we have introduced an overall scale $L$.}
\eqss{\label{redAnsatz11d}
ds_{11}^2 &= L^2\left (\frac{1}{4}\, ds_4^2 + ds_7^2\right )\,, \\
ds_7^2 &= \(d\psi + \sigma + \frac{1}{4}\, A\)^2 + ds_{\rm{KE}}^2\,, \\
G_4 &= L^3\left ( \frac{3}{8}\, {\rm vol}_4 -\frac{1}{4} *_4 F \wedge J \right )\,,
}
where $ds_7|_{A=0}$ is a seven-dimensional Sasaki-Einstein with $R^{\rm{SE}_{7}}_{ij}= 6 g^{\rm{SE}_{7}}_{ij}$, $ds_{\rm{ KE}}$ is a six-dimensional K\"ahler-Einstein space with $R^{\rm{KE}}_{ij}= 8g^{\rm{KE}}_{ij}$ and K\"ahler form $J$, with $d\sigma=2 J$, $F=dA=\text{vol}_{\mathbb{H}^2}$, $ds_{4}^{2}$ is given in \eqref{univ}, and $*_4$ is with respect to this four-dimensional metric. This solution of eleven-dimensional supergravity can be interpreted as the backreaction of $N$ M2-branes wrapping a supersymmetric cycle in a Calabi-Yau five-fold. This point of view was discussed in \cite{Gauntlett:2001qs}, following the ideas in \cite{Maldacena:2000mw}, for the case where the Sasaki-Einstein manifold is $S^7$ and thus the six-dimensional K\"ahler-Einstein space is $\mathbb{CP}^3$.
In order for this local supergravity solution to extend to a well defined M-theory background, some of its parameters should be properly quantized. The quantization condition on $G_{4}$ reads
\equ{
N=\frac{1}{(2\pi \ell_{11})^{6}}\int_{Y_{7}}\ast_{11}\, G_{4}\,,
}
where $N$ is an integer determining the number of M2-branes and $\ell_{11}$ is the Planck length in eleven dimensions. This translates into the quantization of the AdS$_4$ length scale in terms of the Planck length,\footnote{Upon a reduction to 10d type IIA supergravity the 11d Planck length is related to the string length and coupling constant via the equation $\ell_{ 11}=l_{s} g_{s}^{1/3}$.}
\equ{\label{quant L ads2}
L=\pi \ell_{11}\,\(\frac{32N}{3\mathrm{vol}(Y_{7})}\)^{1/6}\,.
}
The four-dimensional Newton constant is given by
\equ{\label{GN4}
\dfrac{1}{G_{N}^{(4)}}=\dfrac{L^{7}\, \mathrm{vol}(Y_{7})}{ G_{N}^{(11)}}\,, \qquad 16\pi G_{N}^{(11)}= (2\pi)^{8}\ell_{11}^{9}\,.
}
\subsubsection{Entropy}
Using \eqref{quant L ads2} and \eqref{GN4} we compute the properly normalized horizon area, obtaining the black hole entropy:
\equ{\label{univM2}
S_{\rm{BH}}=(\frak{g}-1)\,\sqrt{\frac{2\pi^{6}}{27\mathrm{vol}(Y_{7})}}\,N^{3/2}\,.
}
This is consistent with \eqref{eq:SBHFS3} and the well-known expression for the holographic free energy on $S^3$ \cite{Herzog:2010hf,Drukker:2010nc}
\begin{equation}
F_{S^{3}}=\sqrt{\frac{2\pi^{6}}{27\mathrm{vol}(Y_{7})}}\,\,N^{3/2}\;.
\end{equation}
It is instructive to unpack this equation for a couple of examples of well-known three-dimensional SCFTs. For the special case of the ABJM theory, i.e., $Y_{7}=S^{7}/\Bbb Z_{k}$, the free energy on $S^{3}$ is given by $F_{S^{3}}=\frac{\pi\sqrt 2}{3}k^{1/2}N^{3/2}$ \cite{Drukker:2010nc} and thus
\equ{
S_{\rm BH}^{\rm ABJM}= (\frak{g}-1)\, \frac{\pi\sqrt 2}{3}k^{1/2}N^{3/2}\,.
}
This result is correctly reproduced by the topologically twisted index for the ABJM theory. This follows from the general discussion in Section \ref{sec:fieldtheory}, which is model independent.
It is nonetheless instructive to see explicitly how this works. The topologically twisted index for ABJM is given by \cite{Benini:2015eyy,Benini:2016rke}
\equ{\label{topindexABJM}
\log Z^{\rm ABJM}= -\frac{\sqrt 2}{3}k^{1/2}N^{3/2} \sqrt{\Delta_1\Delta_2\Delta_3\Delta_4} \, \sum_{I=1}^4 \frac{\frak{n}_I}{\Delta_I} \, ,
}
where $\Delta_I,\frak{n}_I$ are, respectively, the chemical potentials and fluxes associated to the four chiral bi-fundamental fields of ABJM, subject to the constraints $\sum_{I=1}^4 \Delta_I=2\pi$ and $\sum_I \frak{n}_I= 2(1-\frak{g})$.\footnote{See formula (28) in \cite{Benini:2016rke}, where the notations are: $u_I=\Delta_I$ and $\mathfrak{p}_I=-\frak{n}_I$ and the case $k=1$ is considered. For $k>1$ there could be ambiguities related to other saddle points, but we still expect (\ref{topindexABJM}) to hold.} The exact R-symmetry corresponds to $\Delta_I=\pi/2$ and the universal twist to $\frak{n}_I=(1-\frak{g})/2$. We see that $\frak{g}$ must be odd. We obtain
$\log Z^{\rm ABJM} = S_{\rm BH}^{\rm ABJM}$, as expected.
As another example, we can consider the Sasaki-Einstein manifold $Y_{7}=Q^{1,1,1}/\Bbb Z_{k}$ whose dual SCFT was discussed in \cite{Benini:2009qs}. The free energy on $S^{3}$ is given by $F_{S^{3}}=\frac{4\pi}{3\sqrt{3}}k^{1/2}N^{3/2}$ \cite{Jafferis:2011zi}. Then, from \eqref{eq:SBHFS3} we have
\equ{
S_{\rm BH}^{Q^{1,1,1}}= (\frak{g}-1)\,\frac{4\pi}{3\sqrt{3}}k^{1/2}N^{3/2} \, .
}
The dual field theory is a flavored version of ABJM and the exact R-charge of the fields $A_i$ is $2/3$ and that of $B_i$ is $1/3$ \cite{Jafferis:2011zi}. Quantization of fluxes then requires $\frak{g}-1$ to be an integer multiple of $3$. The topologically twisted index was computed in \cite{Hosseini:2016ume} and it is easy to check that $\log Z^{ Q^{1,1,1}} = S_{\rm BH}^{Q^{1,1,1}}$, again in agreement with
the general result in Section \ref{sec:fieldtheory}.\footnote{The free energy on $S^{3}$ is given by formula (6.15) in \cite{Jafferis:2011zi} and the twisted index for $\frak{g}=0$ by formula (5.47) in \cite{Hosseini:2016ume}. The generalization to general genus $\frak{g}>0$ is obtained by $\log Z_{\frak{g}}= (1-\frak{g}) \log Z_{\frak{g}=0} (\frak{n}_i/(1-\frak{g}))$ \cite{Benini:2016hjo}. As discussed in \cite{Jafferis:2011zi}, the exact R-charge of
the fields $A_i$ is $\Delta_{A_i}/\pi=2/3$ and that of the fields $B_i$ is $\Delta_{A_i}/\pi=1/3$, while $\Delta_m$=0. The fluxes for the universal twist are then $n_{A_i}=2(1-\frak{g})/3$, $n_{B_i}=(1-\frak{g})/3$ and $\mathfrak{t}+\mathfrak{\tilde t}=0$. From formula (5.47) in \cite{Hosseini:2016ume} we find $\log Z^{Q^{1,1,1}} = (\frak{g}-1)\frac{4\pi}{3\sqrt{3}}k^{1/2}N^{3/2}$.}
Combining the results in \cite{Jafferis:2011zi} and \cite{Hosseini:2016ume} one can check many other examples, all in agreement with the general result (\ref{entropyfree}).
As discussed in Section \ref{sec:fieldtheory}, the universal twist is not only possible for theories with an M-theory dual, but also for field theories with massive IIA duals with an $N^{5/3}$ scaling of the free energy. We discuss this next.
\subsection{Uplift to massive IIA}
\label{sec:upMIIA}
Here we discuss new black hole solutions in massive IIA supergravity, obtained by uplifts of the 4d solution \eqref{univ}. This can be done by using the formulae of \cite{Guarino:2015vca,Varela:2015uca,Guarino:2015qaa}, where the uplift of the SU$(3)$-invariant sector of 4d ${\cal N}=8$ supergravity with ISO$(7)$ gauging is given. The bosonic content of the 4d SU$(3)$-invariant sector is the graviton $e_{\mu}^{a}$, 6 scalars $\varphi, \chi, \phi, a,\zeta, \tilde \zeta$, 2 electric vectors $A^{0},A^{1}$ and their magnetic duals $\tilde A_{0},\tilde A_{1}$, 3 two-forms $B^{0},B_{1},B_{2}$ and 2 three-forms $C^{0},C^{1}$. One should keep in mind that these are not all independent as, e.g., the field strengths of $C^{0},C^{1}$ can be dualized into functions of the scalar fields, see \cite{Guarino:2015qaa} for more details.
Since we are interested in solutions that asymptote to the $\mathcal{N}=2$ supersymmetric AdS$_{4}$ vacuum of the theory described in \cite{Guarino:2015jca}, and we expect the scalars to be set to constant values, we set them to the values of the AdS$_4$ vacuum solution, provided in \eqref{eq:scalarIIA}. The electric potential $A^{1}$ is identified (up to a normalization, which we fix by using the massive IIA equations of motion) with the gauge field $A$ of the minimal four-dimensional theory in \eqref{eq:4daction}. With this at hand we can use the uplift formulae of \cite{Guarino:2015vca,Varela:2015uca,Guarino:2015qaa}, reproduced in \eqref{eq:uplift1}, to obtain explicit expressions for the bosonic fields of the ten-dimensional massive IIA supergravity, i.e., the metric, the dilaton, $\hat{\phi}$, as well as the two-, three-, and four-form fluxes: $\hat{F}_2$, $\hat{H}_3$, and $\hat{F}_4$, respectively. The metric and dilaton read
\eqss{\label{flowIIAmetric}
ds^{2}_{10}&=L^{2}(\cos (2\alpha)+3)^{1/2}(\cos(2\alpha)+5)^{1/8}\(ds^{2}_{4}+ds^{2}_{6}\)\,,\\
ds^{2}_{6}&=\frac32 d\alpha^{2}+\frac{6 \sin ^2(\alpha )}{\cos (2 \alpha )+3} \, ds^{2}_{\Bbb{CP}^{2}}+\frac{9 \sin ^2(\alpha )}{\cos (2 \alpha )+5}\,( d\psi + \sigma + g A^1)^{2}\,,\\
e^{\hat{\phi}} &= 2^{1/4} g^{5/6} m^{-5/6} \, \frac{(\cos (2 \alpha )+5)^{3/4}}{\cos (2 \alpha )+3}\,,
}
where $ds^{2}_{4}$ is the black hole metric \eqref{univ}, $ds^{2}_{\Bbb{CP}^{2}}$ is the standard Einstein metric on ${\mathbb{CP}^2}$ (see \eqref{CP2}), $A^1 = \frac{1}{3g} \frac{dx_1}{x_2} $ is the connection on the Riemann surface, and
\equ{\label{LmassiveIIA}
L^{2}= 3^{-1} 2^{-5/8}g^{-25/12}m^{1/12}\,.
}
We note that \eqref{flowIIAmetric} can be obtained from the AdS$_{4}$ vacuum solution of \cite{Guarino:2015jca} by the simple replacement $ds^{2}_{\text{AdS}_{4}}\to ds^{2}_{4}$ and $d\psi\to d\psi+gA^1$. The massive IIA form fields, however, are not so easily obtained and require more work. The basic point is that in order to determine these one must find a further truncation of the remaining $SU(3)$-invariant fields consistent with the duality transformations of \cite{Guarino:2015qaa} and the equations of motion, which due to the connection on the Riemann surface is non-trivial. We discuss this in detail in Appendix~\ref{sec:AppUplift}. Here we simply present the final answer:
\begin{equation}\label{eq:uplift2A}
\begin{split}
\hat{F}_2 = & \, \left(\frac{m}{g}\right)^{2/3} \Bigg(-\frac{4 \sin ^2(\alpha ) \cos (\alpha )}{g (\cos (2 \alpha )+3) (\cos (2 \alpha )+5)}\, J - \frac{3 \sin (\alpha ) (\cos (2 \alpha )-3) }{g (\cos (2 \alpha )+5)^2}\, \eta \wedge d\alpha \\
& + \frac{3 \cos (\alpha )}{\cos (2 \alpha )+5} H^{1}_{(2)} -\frac{1}{2} \sqrt{3}\cos (\alpha ) \ast_4 H^{1}_{(2)} \Bigg) \, , \\
\hat{H}_3 = & \, \frac{1}{m} \left(\frac{m}{g}\right)^{2/3} \Bigg( \frac{8 \sin ^3(\alpha )}{g (\cos (2 \alpha )+3)^2}\, d\alpha \wedge J + \frac{1}{2} \sqrt{3} \sin (\alpha ) \, d\alpha \wedge \ast_4 H^{1}_{(2)}\Bigg) \, ,\\
\hat{F}_4 =& \, \frac{1}{g^2} \left(\frac{m}{g}\right)^{1/3} \Bigg( \frac{\text{vol}_4}{\sqrt{3}g} +\frac{\sin ^4(\alpha ) (3 \cos (2 \alpha )+7)}{g (\cos (2 \alpha )+3)^2}\, J\wedge J \\
& -\frac{3 \sin ^3(\alpha ) \cos (\alpha ) (\cos (2 \alpha )+9) }{g (\cos (2 \alpha )+3) (\cos (2 \alpha )+5)} \, \eta \wedge d\alpha \wedge J +\frac{3}{4} \sin (\alpha ) \cos (\alpha ) \, \eta \wedge d\alpha \wedge H^{1}_{(2)} \\
& -\frac{3 \sin ^2(\alpha ) \cos ^2(\alpha )}{\cos (2 \alpha )+3} \, H^{1}_{(2)}\wedge J + \frac{3 \sqrt{3} \sin (2 \alpha ) }{4 (\cos (2 \alpha )+5)} \, \eta \wedge d\alpha \wedge \ast_4 H^{1}_{(2)} -\frac{\sqrt{3} \sin ^2(\alpha )}{\cos (2 \alpha )+3} \ast_4 H^{1}_{(2)} \wedge J \Bigg) \, ,
\end{split}
\end{equation}
where $J=\frac12 d\sigma$ is the K\"ahler form on $\mathbb{CP}^2$, $H^{1}_{(2)}\equiv dA^1$, $\text{vol}_4$ is the volume form of $ds_4$, and $\ast_{4}$ is the hodge dual with respect to $ds_4$. We have explicitly checked that \eqref{flowIIAmetric} and \eqref{eq:uplift2A} indeed satisfy the equations of motion of massive IIA supergravity.\footnote{These equations can be found in Equation (A.5) of \cite{Guarino:2015vca}.} At large $\rho$ the solution asymptotes to the $\mathcal{N}=2$ AdS$_4$ solution of \cite{Guarino:2015jca} and thus \eqref{flowIIAmetric} and \eqref{eq:uplift2A} describe the twisted compactification of the corresponding three-dimensional field theory dual, consisting of a $U(N)$ Chern-Simons theory at level $k$ with three adjoints superfields $\Phi_{i}$.
Based on our field theory analysis we actually expect \eqref{flowIIAmetric} and \eqref{eq:uplift2A} to be part of a more general class of solutions, where the $\mathbb{CP}^2$ is replaced by a general K\"{a}hler-Einstein base. As an example, we have explicitly checked that replacing
\begin{equation}
\begin{split}
ds^2_{\mathbb{CP}^2} &\rightarrow ds^2_{\Bbb{P}^{1}\times \Bbb{P}^{1}} = \frac{1}{6} \left(d\theta_1^2+\sin^2 \theta_1 \, d\phi_1^2 + d\theta_2^2+\sin^2 \theta_2 \, d\phi_2^2 \right) \, , \\
\sigma &\rightarrow \frac{1}{3} \left( \cos \theta_1 \,d\phi_1 + \cos \theta_2 \, d\phi_2 \right) \,,
\end{split}
\end{equation}
in \eqref{flowIIAmetric} (along with suitable replacements of $J$ and $\eta$ in \eqref{eq:uplift2A}) is also a solution of the equations of motion.
More generally, replacing $\mathbb{CP}^2$ by a generic K\"ahler-Einstein base ${\cal B}$ leads to a large family of new massive IIA black holes, which are asymptotic to the AdS$_{4}$ solutions described in \cite{Fluder:2015eoa}. As discussed there, the 3d field theory duals are obtained from a 4d parent field theory with an AdS$_{5}\times Y_{5}$ dual, with ${\cal B}$ corresponding to the K\"ahler-Eistein base of the Sasaki-Einstein $Y_{5}$, i.e.,
\equ{ds^{2}_{Y^{5}}=\eta^{2}+ds^{2}_{{\cal B}}
\,.}
Thus, the black hole solutions reported here describe the compactifications of these 3d SCFTs on a Riemann surface, twisted by the exact superconformal R-symmetry.
To ensure that the local supergravity solution given in \eqref{flowIIAmetric} and \eqref{eq:uplift2A} extends to a well defined string theory background, it should be properly quantized.\footnote{In the case of $ds^{2}_{4}= ds^{2}_{ \rm{AdS}_{4}}$ this was carried out in \cite{Fluder:2015eoa} for a general base ${\cal B}$. The quantization for the black hole solution here is very similar and we provide it here for completeness.} The four-form quantization condition in massive IIA reads\footnote{See for example Equation (4.11) in \cite{Varela:2015uca}.}
\equ{\label{quantconditionmIIA}
N=-\frac{1}{(2\pi \ell_s)^5}\int_{M_6}\(e^{\frac12 \hat \phi} \ast_4 \hat F_{(4)}+\hat B_{(2)}\wedge d\hat A_{(3)}+\frac16 \, m\, \hat B_{(2)}\wedge \hat B_{(2)}\wedge \hat B_{(2)}\,\),
}
where $N$ is an integer determining the number of D2-branes, and the potentials $\hat{B}_{(2)}, \hat{A}_{(3)}$ can be found in \cite{Varela:2015uca}. We can evaluate this for the solution in \eqref{flowIIAmetric} and \eqref{eq:uplift2A} for a general base ${\cal B}$.
Using the explicit expression for the background fluxes presented in Appendix \ref{sec:AppUplift} we find
\eqs{ \nonumber
N &= \frac{1}{(2\pi \ell_s)^5} \int_{M_{6}} \frac{\sin ^5(\alpha ) (132 \cos (2 \alpha )+7 \cos (4 \alpha )+213)}{4 g^5 (\cos (2 \alpha )+3)^3} \,d\alpha \wedge J \wedge J \wedge d\psi \\ \label{eq:N}
&= \frac{1}{(2\pi \ell_s)^5} \frac{16}{3g^{5}} \, \mathrm{vol}(Y_{5}) \, ,
}
where we integrated over $0\leq\alpha\leq\pi$ and used the identity $\frac12 J\wedge J\wedge d\psi=d\mathrm{vol} (Y_{5})$. Using
\eqref{LmassiveIIA} this translates into a quantization condition on the asymptotic AdS$_{4}$ and the near-horizon AdS$_{2}$ radii:
\equ{
L_{\text{AdS}_2}=\frac 12 L_{\text{AdS}_4}=\frac{\pi \ell_{s}n^{1/24}}{2^{55/48}3^{7/24}}\(\frac{N}{\mathrm{vol}(Y_{5})}\)^{5/24}\,,
}
where we have used that asymptotically $ds^{2}_{4}\to ds^{2}_{ \rm{AdS}_{4}}$, while close to the horizon $ds^{2}_{4}\to \frac14 (ds^{2}_{ \rm{AdS}_{2}}+2ds^{2}_{\Sigma_{\frak{g}}})$ and defined $n\equiv2\pi \ell_s m$.
Note that the quantization condition on $L_{\text{AdS}_4}$ coincides with the one obtained in \cite{Fluder:2015eoa}. This is expected since our solution is asymptotic to the solutions discussed there. Similarly, the four-dimensional Newton constant is given by\footnote{For a solution of the form
$ds^{2}_{10}=e^{2\lambda}L^{2}\(ds^{2}_{4}+ds^{2}_{6}\)$, the effective four-dimensional Newton constant is given by $1/G_N^{(4)}=(1/G_N^{(10)})\,L^{6}\int d^{6}x \sqrt g_{6}e^{8\lambda}$, with $16\pi G_{N}^{(10)}=(2\pi)^{7}\ell_{s}^{8}$, following the conventions used in \cite{Guarino:2015jca}. Note in particular that in \cite{Guarino:2015jca} they set $g_s=1$.
}
\equ{
\frac{1}{G_N^{(4)}}=\frac{6^{5/2}L_{\text{AdS}_{4}}^{6}\mathrm{vol}(Y_{5})}{5\ell_{s}^{8}\pi^{6}}\,.
}
\subsubsection{Entropy}
Putting all this together, we can compute the properly normalized horizon area and obtain the following general formula for the black hole entropy:
\equ{\label{genEntropyIIA}
S_{\rm BH}= (\frak{g}-1)\,\frac{2^{1/3}\,3^{1/6}\,\pi^{3}}{5\,\mathrm{vol}(Y_{5})^{2/3}}\;n^{1/3}\,N^{5/3}\,.
}
Let us specialize this to the two simple cases considered above, namely $\Bbb{CP}^{2}$ and $\Bbb{P}^{1}\times\Bbb{P}^{1}$, with corresponding $Y_{5}$'s equal to $S^{5}$ and $Y^{1,0}$, respectively. We recall that for a $U(N)^{G}$ gauge theory, where each factor has the same Chern-Simons level $k$, the parameter $n$ is related to the Chern-Simons levels by $n=Gk$ \cite{Gaiotto:2009mv}.
For $Y_{5}=S^{5}$ we have $\mathrm{vol}(Y_{5})=\pi^{3}$ and $n=k$ and \eqref{genEntropyIIA} reads
\equ{\label{entropyS5}
S_{\rm BH}^{S^{5}}=(\frak{g}-1)\, \pi \, 2^{1/3}3^{1/6}5^{-1}N^{5/3}k^{1/3}\,.
}
The dual three-dimensional field theory is a $U(N)$ Chern-Simons theory with three adjoints superfields $\Phi_i$ and
superpotential $W= \Phi_1[\Phi_2,\Phi_3]$ \cite{Jafferis:2011zi}. The index is explicitly computed in Appendix \ref{app:fieldIIA}; it is given by (\ref{index theorem:attractor2}) and (\ref{IIAN=4}), where $\Delta_1+\Delta_2+\Delta_3=2\pi$ and $\frak{n}_1+\frak{n}_2+\frak{n}_3=2(1-\frak{g})$. The exact R-symmetry corresponds to $\Delta_i=2\pi/3$ and the universal twist to $\frak{n}_i= 2(1-\frak{g})/3$. Quantization of fluxes requires $\frak{g} -1$ to be an integer multiple of $3$. We then see that the index is given by
\begin{equation}
\log Z = (\frak{g} -1) \, \pi \, 2^{1/3}3^{1/6}5^{-1}N^{5/3}k^{1/3} ( 1- i/\sqrt{3})\,,
\end{equation}
while the free energy on $S^3$ is given by equation (\ref{indexS32}),
\begin{equation}
F_{S^3} =\pi\, 2^{1/3}3^{1/6}5^{-1}N^{5/3}k^{1/3} ( 1- i/\sqrt{3})\, .
\end{equation}
As expected this leads to $\log Z= (\frak{g} -1) F_{S^3}$, in agreement with our general argument in Section \ref{sec:fieldtheory}. We also note that both $\log Z$ and $F_{S^3}$ are complex but as discussed above Equation \eqref{ident} we should focus on the real part. We then arrive at the following, by now familiar, relation between the black hole entropy and the SCFT partition functions; $S_{\rm BH}={\rm Re} \log Z= (\frak{g} -1) {\rm Re} F_{S^3}$.
Similarly, for $Y^{5}=Y^{1,0}$ one has $\mathrm{vol}(Y_{5})=\frac{16\pi^{3}}{27}$, $n=2k$ and \eqref{genEntropyIIA} reads
\equ{\label{entropyY10}
S_{\rm BH}^{Y^{1,0}}=(\frak{g}-1)\, \pi \, 2^{-7/3}3^{13/6}5^{-1}N^{5/3}k^{1/3}\,.
}
The three-dimensional field theory dual is a $U(N)\times U(N)$ gauge theory with four bifundamentals $\Phi_{I}$ with the same quiver and superpotential as the 4d $\mathcal{N}=1$ Klebanov-Witten theory \cite{Klebanov:1998hh}. In this case the exact R-symmetry corresponds to $\Delta_I=\pi/2$ and the universal twist to $\frak{n}_I= (1-\frak{g})/2$, which requires $\frak{g}$ to be odd. The topologically twisted index, computed explicitly in Appendix \ref{app:fieldIIA}, correctly reproduces the entropy in \eqref{entropyY10}.
\subsubsection{Checking supersymmetry}
As a final consistency check of the supergravity discussion of these new massive IIA black hole solutions here we show explicitly that the solution in \eqref{flowIIAmetric} and \eqref{eq:uplift2A} preserves two supercharges. The supersymmetry variations of the fermionic fields are given in \cite{Romans:1985tz} and read\footnote{The conventions of this reference are related to ours by $\phi\leftrightarrow-\frac{1}{2}\hat \phi, F_{mnpq}\leftrightarrow \frac{1}{2}\hat F_{mnpq},G_{pqr}\leftrightarrow \hat H_{pqr}, B_{pq} \leftrightarrow \hat B_{pq}$.}
\begin{equation}\label{eq:var}
\begin{split}
\delta \psi_{\mu}=&\,D_{\mu}\epsilon-\frac{1}{32}m \, e^{\frac54 \hat \phi}\Gamma_{\mu}\epsilon -\frac{1}{64} e^{\frac34 \hat \phi}\hat F_{\nu\rho}(\Gamma_{\mu}^{\;\;\nu \rho}-14 \delta_{\mu}^{\;\;\nu}\Gamma^{\rho})\Gamma_{11}\epsilon\\
&+\frac{1}{96}e^{-\frac12 \hat \phi}\hat H_{\nu\rho\lambda}(\Gamma_{\mu}^{\;\;\nu\rho\lambda}-9 \delta_{\mu}^{\;\;\nu}\Gamma^{\rho\lambda})\Gamma_{11}\epsilon+\frac{1}{256}e^{\frac14 \hat \phi}\hat F_{\nu\rho\lambda\tau}(\Gamma_{\mu}^{\;\;\nu\rho\lambda\tau}-\frac{20}{3} \delta_{\mu}^{\;\;\nu}\Gamma^{\rho\lambda\tau})\epsilon\,,\\
\delta \lambda=&\, -\frac{1}{2\sqrt 2}(\partial_{\mu}\hat\phi) \, \Gamma^{\mu}\epsilon-\frac{5}{8\, \sqrt2}\, m\, e^{\frac54 \hat \phi}\epsilon+\frac{3}{16\, \sqrt2}\, e^{\frac34 \hat \phi}\hat F_{\mu\nu}\Gamma^{\mu\nu}\Gamma_{11}\epsilon\\
&+\frac{1}{24 \, \sqrt 2}\, e^{-\frac12 \hat \phi} \hat H_{\mu\nu\rho}\Gamma^{\mu\nu\rho}\Gamma_{11}\epsilon-\frac{1}{192 \, \sqrt 2}\, e^{\frac14 \hat \phi} \hat F_{\mu\nu\rho\lambda}\Gamma^{\mu\nu\rho\lambda}\epsilon\,.
\end{split}
\end{equation}
Plugging in the background \eqref{flowIIAmetric} and \eqref{eq:uplift2A} and setting these variations to zero we obtain a set of differential equations for $\epsilon$, with the following solution (see Appendix \ref{sec:AppSUSYIIA} for details):
\begin{equation}
\epsilon = |g_{tt}|^{1/4} e^{\frac{3}{2}\Gamma^{67}\psi}\mathcal{R}(\theta_1,\theta_2) \eta \, ,
\end{equation}
where $g_{tt}$ the time-time component of the 10d metric in \eqref{flowIIAmetric}. The quantity $\mathcal{R}(\theta_1,\theta_2)$ is an $\alpha$-dependent ``rotation operator,'' defined as
\begin{equation}\label{eq:rotations}
\begin{split}
\mathcal{R}(\theta_1,\theta_2) &\equiv \mathcal{R}_{29}(\theta_1) \mathcal{R}_{1234}(\theta_2) \,; \\
\mathcal{R}_{29}(\theta_1) &\equiv \left(\cos\left(\theta_1/2\right)-\sin\left(\theta_1/2\right)\Gamma^{29}\right) \,, \\
\mathcal{R}_{1234}(\theta_2) &\equiv \left(\cos\left(\theta_2/2\right)-\sin\left(\theta_2/2\right)\Gamma^{1234}\right) \,,
\end{split}
\end{equation}
where the functions $\theta_{1,2}=\theta_{1,2}(\alpha)$ are given in \eqref{eq:angles}, and $\eta$ is a constant ten-dimensional spinor obeying the projection conditions
\equ{\label{projeta}
\Pi_{1}\eta=\Pi_{2}\eta=\Pi_{3}\eta=\Pi_{4}\eta=0\,.
}
Here we have defined the projectors
\eqsn{
\Pi_1 &\equiv \frac{1}{2}\left(1-\Gamma^{5678} \right) \,, && \Pi_2 \equiv \frac{1}{2}\left(1+\Gamma^{3467} \right) \,, \\
\Pi_3 &\equiv \frac{1}{2}\left(1+\Gamma^{134}\right) \,, &&
\Pi_4 \equiv \frac{1}{2}\left(1+\Gamma^{67910}\right) \,.
}
Using \eqref{eq:rotations} one may alternatively express \eqref{projeta} as projection conditions on $\epsilon$:
\equ{
\Pi_{1}\epsilon=\Pi_{2}\epsilon=\widehat \Pi_{3}\epsilon=\widehat \Pi_{4}\epsilon=0\,,
}
where we defined the ``rotated projectors,''
\eqs{
\widehat \Pi_3 \equiv \mathcal{R}(\theta_1,\theta_2) \,\Pi_3 \, \mathcal{R}^{-1}(\theta_1,\theta_2) \,, &&
\widehat \Pi_4 \equiv \mathcal{R}(\theta_1,\theta_2) \, \Pi_4 \, \mathcal{R}^{-1}(\theta_1,\theta_2) \,.
}
Similar rotated projectors have appeared in \cite{Pilch:2015dwa,Pilch:2015vha}, albeit in a somewhat different class of supergravity solutions with internal fluxes. Note that the $\Pi_3$ projector can be understood as a dielectric deformation of the standard D2-brane projector due to the internal fluxes. The four projection conditions in \eqref{projeta} determine that the black hole solution of interest preserves 2 out of the 32 supercharges of the massive IIA supergravity.\footnote{As usual here we discuss only the Poincar\'e type supercharges. For a supersymmetric background with an AdS factor in the metric the number of supercharges is doubled due to the presence of superconformal supercharges.} Note this is half the number of supercharges of the AdS$_4$ solution of \cite{Guarino:2015jca}, since the $\Pi_2$ projector is absent in that case.
\section{AdS$_2$ horizons in 11d}
\label{AdS2Hor}
\global\long\def\mathrm{vol}{\mathrm{vol}}
The discussion thus far has been restricted to the so-called universal twist of 3d $\mathcal{N}=2$ SCFTs and its holographic dual description in terms of a magnetically charged black hole in AdS$_4$. As emphasized, the universal twist is characterized by the fact that the background magnetic flux in the CFT is only along the direction of the 3d superconformal R-symmetry. Most 3d $\mathcal{N}=2$ SCFTs with holographic duals, however, admit continuous flavor symmetries and it is natural to study the behavior of these theories in the presence of magnetic fluxes for these global symmetries, as well as their holographic duals. In this section we begin exploring this question for a class of 3d $\mathcal{N}=2$ SCFTs arising from M2-branes placed at the tip of a $CY_4$ conical singularity in M-theory. It is well-known that upon backreaction of the branes this setup leads to a supersymmetric Freund-Rubin vacuum of M-theory of the form AdS$_4\times X_7$ where $X_7$ is a Sasaki-Einstein manifold which is the base of the $CY_4$ cone. We have already mentioned a few of these spaces above, for example $M^{1,1,1}$, $Q^{1,1,1}$, and $V^{5,2}$. Many other examples of such Sasaki-Einstein manifolds with explicit metrics are known in the literature (see for example \cite{Gauntlett:2004hh} and references thereof). In addition to the omnipresent Reeb vector, which exists on all such spaces and is dual to the superconformal R-symmetry, these examples of Sasaki-Einstein manifolds typically have additional isometries, which are dual to mesonic flavor symmetries in the dual CFT. In addition, if the seven-dimensional manifold has non-trivial two-cycles (and thus by Poincar\'e duality five-cycles) the dual CFT has continuous baryonic symmetries. After placing the 3d SCFT on $S^1\times \Sigma_\frak{g}$ we can turn on background magnetic fluxes for these mesonic and baryonic flavor symmetries while still preserving supersymmetry.\footnote{From the point of view of the M2-brane picture this corresponds to wrapping the M2-branes on $\Sigma_\frak{g}$ which then by the general result in \cite{Bershadsky:1995qy} automatically implements the topological twist and fibers the $CY_4$ over the Riemann surface.} The holographic dual description of this setup should be given by a magnetically charged supersymmetric black hole which preserves (at least) two supercharges and is asymptotic to the original AdS$_4\times X_7$ background. The near horizon limit of this black hole should be given by a warped product of the form AdS$_2\times_w M_9$ where the manifold $M_9$ is a fibration of the Sasaki-Einstein space $X_7$ over $\Sigma_\frak{g}$. Our goal in this section is to construct explicit examples of such warped AdS$_2$ supersymmetric vacua of eleven-dimensional supergravity.
A few comments on our strategy to attack this technically challenging problem are in order. A classification of warped AdS$_2$ solutions of 11d supergravity was given in \cite{Kim:2006qu,Gauntlett:2007ts}. The result of these papers is that the manifold $M_9$ should be given as a $U(1)$ bundle over an eight-dimensional K\"ahler manifold. The metric on this eight-dimensional space should obey a certain fourth-order non-linear PDE.\footnote{It is curious to note that such K\"ahler manifolds have appeared recently in a seemingly unrelated context in \cite{Ang:2016tvp}.} Finding explicit expressions for such metrics is a challenging problem to tackle (see however \cite{Gauntlett:2006ns,MacConamhna:2006nb,Donos:2008ug} for some examples) so we employ a different strategy, which closely follows the one in \cite{Benini:2015bwz} where many explicit AdS$_3$ vacua of type IIB supergravity with $(0,2)$ supersymmetry were constructed. To be more concrete we propose an Ansatz for the metric and four-form flux, $G_4$ of the 11d theory and impose the vanishing of the gravitino supersymmetry variation as well as the equations of motion and the Bianchi identity for $G_{4}$. We show that all solutions within our Ansatz are determined by a single function satisfying a nonlinear ODE. This nonlinear equation admits quartic polynomial solutions which lead to a rich class of supergravity solutions. This class contains the near horizon AdS$_2$ region of the universal solution \eqref{redAnsatz11d} as a special case.
After this brief introduction let us proceed with the construction of our solutions. The eleven-dimensional space is of the form AdS$_{2}\times_{w}M_9$ where $M_9$ is a nine-dimensional manifold given by a seven-dimensional space $M_7$ fibered over a Riemann surface $\Sigma_{\frak{g}}$:
\[
\begin{array}{cccc}
M_7 & \hookrightarrow & M_9&\\
& & \downarrow & .\\
& & \Sigma_{\mathfrak{g}}&
\end{array}
\]
We will take $\Sigma_{\frak{g}}$ to be compact (i.e. with no punctures) and have a constant-curvature metric,\footnote{In principle, this assumption could be relaxed but the analysis is more involved. In addition the results of \cite{Anderson:2011cz} suggest that all the interesting physics in the IR is captured by the constant-curvature metric.} given by
\equ{
ds_{\Sigma_{\mathfrak{g}}}^{2}=e^{2h(x_{1},x_{2})}\left(dx_{1}^{2}+dx_{2}^{2}\right),
}
with
\equ{\label{defh}
h(x_{1},x_{2})=\begin{cases}
-\log\frac{1+x_{1}^{2}+x_{2}^{2}}{2} & \mathfrak{g}=0\\
\frac{1}{2}\log2\pi & \mathfrak{g}=1\,.\\
-\log x_{2} & \mathfrak{g}>1
\end{cases}
}
The Ansatz for $M_{7}$ will be modelled on a large class of seven-dimensional
Sasaki-Einstein manifolds found and studied in \cite{Gauntlett:2004hh,Martelli:2008rt},
denoted by $Y^{p,q}(\mathcal{B})$, where $\mathcal{B}$, in our case,
is a four-dimensional K\"{a}hler-Einstein manifold that can be either
$\mathbb{CP}^{1}\times\mathbb{CP}^{1}$ or $\mathbb{CP}^{2}$, upon
which the full manifold is constructed. $M_7$ is then spanned by the four coordinates on $\mathcal{B}$
and $(y,\beta,\psi)$, the latter two being angles. In particular,
$\psi$ is the angle associated to the Reeb vector of $M_{7}$. Thus, we employ an eleven-dimensional metric Ansatz which has explicit AdS$_2$, $\Sigma_\frak{g}$ and $\mathcal{B}$ factors as well as two $U(1)$ isometry directions, $\beta$ and $\psi$, which are fibered over $\Sigma_\frak{g}$ and $\mathcal{B}$. With these assumptions we arrive at the following general metric Ansatz:
\begin{equation}
ds_{11}^{2}=f_{1}^{2}\,ds_{\text{AdS}_{2}}^{2}+f_{2}^{2}\,ds_{\Sigma_{\mathfrak{g}}}^{2}+f_{3}^{2}\,dy^{2}+f_{4}^{2}\,(D\beta)^{2}+f_{5}^{2}\,ds_{{\cal B}}^{2}+f_{7}^{2}\,(D\psi)^{2},\label{AdS2Hor_metricAnsatz}
\end{equation}
where $f_{n}=f_{n}(y)$ are functions of the coordinate $y$ only, and
\begin{equation}
\label{defDpsibeta}
D\beta\equiv d\beta+cA+bA_{{\cal B}}\,, \qquad D\psi\equiv d\psi+aA_{{\cal B}}+\tilde{c}A-f_{8}(y)D\beta\,,
\end{equation}
with $A=A_{1}(x_{1},x_{2})\,dx_{1}+A_{2}(x_{1},x_{2})\,dx_{2}$ a connection on the Riemann surface and $A_{\mathcal{B}}$ a connection on ${\cal B}$, satisfying,
\begin{equation}
dA=\mathrm{vol}_{\Sigma_{\mathfrak{g}}}\,, \qquad dA_{\mathcal{B}}=J_{\mathcal{B}}\,,
\label{AdS2Hor_ConnectionRels}
\end{equation}
with $J_{\mathcal{B}}$ the K\"ahler form on ${\cal B}$. The most general Ansatz for the 4-form compatible with the symmetries of the problem reads
\begin{alignat}{1}
G_{4} =&\, f_1^2 \mathrm{vol}_{\text{AdS}_{2}}\wedge\left(f_2^2 g_{1}\mathrm{vol}_{\Sigma_{\mathfrak{g}}}+f_3 f_4 g_{2}dy\wedge D\beta+f_4 f_7 g_{3}D\beta\wedge D\psi+ f_3 f_7 g_{4}dy\wedge D\psi+f_5^2 g_{5}J_{\mathcal{B}}\right)\nonumber \\
& +f_2^2 \mathrm{vol}_{\Sigma_{\mathfrak{g}}}\wedge\left( f_3 f_4 i_{1}dy\wedge D\beta + f_5^2 i_{2}J_{\mathcal{B}} + f_4 f_7 i_{4}D\beta\wedge D\psi + f_3 f_7 i_{5}dy\wedge D\psi\right)\label{AdS2Hor_4FormAnsatz}\\
& +f_5^2 J_{\mathcal{B}}\wedge\left(f_3 f_4 l_{1}dy\wedge D\beta+f_5^2 l_{2} J_{\mathcal{B}} + f_4 f_7 l_{3}D\beta\wedge D\psi+ f_3 f_7 l_{4}dy\wedge D\psi\right)\,,\nonumber
\end{alignat}
where $g_{n}=g_{n}(y),i_{n}=i_{n}(y)$, and $l_{n}=l_{n}(y)$ are functions of the coordinate $y$ only.
Before we describe the solutions, let us discuss the interpretation of various terms in the Ansatz. Since the Reeb vector $\partial_{\psi}$ corresponds to the $U(1)_{R}$ R-symmetry of the 3d ${\cal N}=2$ field theory dual, the $\tilde c A$ term in $D\psi$ thus corresponds to the part of the topological twist along the three-dimensional superconformal R-symmetry. The isometry $\partial_{\beta}$, on the other hand, is dual to a flavor symmetry and thus the $cA$ term in $D\beta$ corresponds to flavor flux in the dual field theory (which should be set to zero in the special case of the universal twist). The parameters $a,b$ specify the structure of the fibration over ${\cal B}$. From the black hole perspective, since the reduction along $D\psi$ and $D\beta$ leads to a graviphoton and a vector field, respectively, the coefficients $c,\tilde c$ are related to the magnetic charges of the black hole under these two gauge fields. Other charges may arise in four dimensions from the reduction of the 4-form, i.e., Betti multiplets. These arise from the terms proportional to elements of $H^{2}(M_7)$
and $\mathrm{vol}_{\Sigma_{\mathfrak{g}}}$ and correspond to a background for a baryonic symmetry in the dual picture; we discuss a class of examples of this in section~\ref{AdS2Hor-BaryonicFluxes}. It is worth pointing out that the map between the parameters in our supergravity Ansatz and quantities in the dual CFT could in general be complicated by a possible mixing between the three-dimensional superconformal R-symmetry and flavor and baryonic symmetries as one flows to the IR. Thus, one should take the field theory interpretation of the supergravity parameters with a healthy dose of skepticism.
Plugging the Ansatz \eqref{AdS2Hor_metricAnsatz} and \eqref{AdS2Hor_4FormAnsatz} into the BPS equations and the equations of motion and Bianchi identity for the four-form: $dG_{4}=d\ast_{11} G_{4}+ \frac{1}{2}G_4\wedge G_4=0$, one obtains a set of differential equations, analyzed in detail in Appendix~\ref{Appendix11D}. As discussed there it is convenient to introduce a new basis of functions $F_{i}(y)$, related to the original functions $f_{i}(y)$ by \eqref{AdS2Hor_FBasis}. Here we give the final result of our analysis, which gives a metric
\begin{alignat}{1}
ds_{11}^{2} =&\,\(\frac{F_{2}^{2}F_{3}^{4}}{F_{1}^{2}}\)^{1/3}\,ds_{\text{AdS}_{2}}^{2}+\(\frac{F_{1}F_{2}^{2}}{F_{3}^{2}}\)^{1/3}\,ds_{\Sigma_{\mathfrak{g}}}^{2}+K\frac{\(F_{1}F_{2}^{2}F_{3}^{4}\)^{1/3}}{F_{5}}\,dy^{2}\nonumber \\
& +KF_{5}\(\frac{F_{1}}{F_{2}^{4}F_{3}^{8}}\)^{1/3}\,(D\beta)^{2}+\(\frac{F_{1}F_{3}}{F_{2}}\)^{1/3}\,ds_{{\cal B}}^{2}\label{AdS2Hor_AltAnsatzMain}\\
& +4\omega_{1}^{2}\(\frac{F_{2}^{2}F_{3}^{4}}{F_{1}^{2}}\)^{1/3}\left(d\psi+aA_{{\cal B}}+\tilde{c}A-\frac{F_{6}}{F_{2}F_{3}^{2}}\,D\beta\right)^{2}\,\nonumber
\end{alignat}
and 4-form
\begin{alignat}{1}
G_{4} =&\,\mathrm{vol}_{\text{AdS}_{2}}\wedge\left[\frac{2F_{1}F_{2}-2\kappa F_{2}F_{3}^{2}+cF_{5}'}{2F_{1}}\,\mathrm{vol}_{\Sigma_{\mathfrak{g}}}+\left(2\omega_{1}\frac{F_{2}F_{3}^{2}}{F_{1}}\left(\frac{F_{6}}{F_{2}F_{3}^{2}}\right)'+K\right)\,dy\wedge D\beta\right.\nonumber\\
& \left.-2\omega_{1}\frac{\left(cKF_{3}^{2}+2bKF_{2}F_{3}\right)F_{1}-F_{2}F_{3}^{2}F_{1}'}{F_{1}^{2}}\,dy\wedge D\psi+\frac{2F_{1}F_{3}-4qF_{2}F_{3}^{2}+bF_{5}'}{2F_{1}}\, J_{\mathcal{B}}\right]\label{AdS2Hor_G4Solution}\\
& +\frac{F_{2}\mathcal{I}_{0}}{F_{3}^{2}}\left[\mathrm{vol}_{\Sigma_{\mathfrak{g}}}\wedge\left(-2dy\wedge D\beta+J_{\mathcal{B}}\right)+J_{\mathcal{B}}\wedge\left(dy\wedge D\beta-J_{\mathcal{B}}\right)\right]\,,\nonumber
\end{alignat}
where $\omega_{1}$ and ${\cal I}_{0}$ are integration constants, the functions $F_{2},F_{3}$ are linear in $y$ and $F_{1},F_{6}$ are given in terms of $F_{2},F_{3}$ and the function $F_{5}$ by:
\begin{equation}
\begin{split}F_{1} & =-\frac{F_{5}''}{2K}+\kappa F_{3}^{2}+4qF_{2}F_{3}\,,\\
F_{2} & =cKy+S_{2}\,,\\
F_{3} & =bKy+S_{3}\,,\\
F_{6} & =\frac{F_{5}'-4\omega_{2}F_{2}F_{3}^{2}}{4\omega_{1}}\,,
\end{split}
\label{AdS2Hor_FSummary}
\end{equation}
with $\{S_{2},S_{3},\omega_{2},K\}$ integration constants\footnote{In Appendix~\ref{Appendix11D-Singularities} we show that actually $\omega_{2}$ can be always
set to zero without loss of generality. Furthermore, $\omega_1$ and $K$ can be set to convenient non-zero values.} and $q$ is related to the Ricci curvature of $\mathcal{B}$ by:
\equ{
q=\frac{R_{\mathcal{B}}}{8}\,.
}
The entire class of solutions is thus characterized by the single function $F_{5}$ which, as shown in Appendix~\ref{Appendix11D}, satisfies the second-order nonlinear ODE given in $\eqref{AdS2Hor_F5Eq}$ coming from the Bianchi identity for $G_{4}$, which for $b\neq 0$ reads
\eqss{
\frac{1}{KF_{2}F_{3}^{2}}\left[-F_{5}'^{2}+2F_{5}\left(F_{5}''-2K\left(4qF_{2}F_{3}+\kappa F_{3}^{2}\right)\right)\right]-\frac{4\mathcal{I}_{0}^{2}\left(3bcKy+bS_{2}+2cS_{3}\right)}{b^{3}KF_{3}^{2}}\\
+\frac{16}{3}Kqy^{2}\left(b\kappa Ky+cKqy+3\kappa S_{3}+3qS_{2}\right)+p_{0}+p_{1}y=0\,,\label{AdS2Hor_F5EqMain}
}
with $p_{0},p_{1}$ two more integration constants.
Finally, there are some constraints on $a,b,c,\tilde{c}$:
\equ{\label{eq:caconstr}
c\, \omega_{2}+\tilde{c}\, \omega_{1} =\frac{\kappa}{2}\,,\qquad
b\, \omega_{2}+a\, \omega_{1} =q\,.
}
Thus, the most general solution within our Ansatz is characterized by solutions to the nonlinear second-order ODE \eqref{AdS2Hor_F5EqMain}.\footnote{For $b=0$ the ODE follows from the general Equation \eqref{AdS2Hor_F5Eq}.} Although we have not found the most general solution, one can show that the most general {\it polynomial} solution is at most quartic, i.e.,
\equ{\label{eq:F5pol}
F_{5}(y)=\sum_{n=0}^{4}\alpha_{n}\, y^{n}\,.
}
Plugging this into \eqref{AdS2Hor_F5EqMain} leads to a set of algebraic equations for the coefficients $\alpha_{n}$ and the other integration constants specifying the solution, given in \eqref{constalphas}.
Solutions to this set of algebraic constraints leads to a rich family of local AdS$_{2}$ solutions. As we discuss next, it contains the universal solution as a special case, as well as interesting generalizations of it by the addition of mesonic and baryonic flavor fluxes.
\subsection{Universal twist\label{AdS2Hor-UniversalTwist}}
The universal solution is characterized by the absence of flavor fluxes.
To this end we will set $c=\mathcal{I}_0=0$.
We will also assume $b\neq 0$, leaving the case $b=0$ to the next section, as this case admits a more general solution.
One can show that there are no solutions for $\kappa=0$, but the values $\kappa=\pm1$ are allowed by the constraints.
We will now discuss the solution for $\kappa=-1$ and discuss the case $\kappa=1$ in Section \ref{sec:AdS2xS3xT6}.
Plugging in the polynomial Ansatz \eqref{eq:F5pol} for $F_5$ into the Bianchi constraint \eqref{AdS2Hor_F5EqMain} fixes $\{p_0,p_1,\alpha_1,\alpha_2,\alpha_3,\alpha_4\}$, leaving $\alpha_0$ undetermined.
There are in fact two solutions to the Bianchi constraint, but only one obeys the metric positivity constraint $F_1>0$.
The parameters $a,\tilde{c}$ are fixed by \eqref{eq:caconstr} to be
\[
a=\frac{q}{\omega_1}\,,\qquad\tilde{c}=\frac{\kappa}{2\omega_1}\,.
\]
The dependence of the solution on the parameters $\omega_1,\omega_2,K$ and $S_3$ can be removed by rescaling and shifting $y,\beta$ and $\psi$.
It is convenient to define $S_2 \equiv L^3/32$ and to perform a rescaling $\beta'=2\beta,\psi'=(\omega_1/2)\psi $
so that now $D\beta'=d\beta'+2cA+2bA_{{\cal B}}$, and define the polynomial
\[
Q_{2}(x)\equiv3x^{2}+2x+1\,.
\]
The metric \eqref{AdS2Hor_AltAnsatzMain} can then be written as
\eqs{\nonumber
ds_{11}^{2} =&\,L^{2}\left[\frac{1}{4}\left(\frac{1}{4}ds_{\text{AdS}_2}^{2}+\frac{1}{2} ds^2_{\Sigma_g}\right)+\frac{3}{4}\frac{\left(1-c_{Y^{p,q}}y\right)^{2}}{a_{Y^{p,q}}-y^{2}Q_{2}(1-c_{Y^{p,q}}y)}\,dy^{2}\right.\\
& +\frac{q^{2}}{48}\frac{a_{Y^{p,q}}-y^2\, Q_{2}(1-c_{Y^{p,q}}y)}{\left(1-c_{Y^{p,q}}y\right)^{2}}(D\beta')^{2}+\frac{q}{4}\left(1-c_{Y^{p,q}}y\right)ds_{{\cal B}}^{2}\\ \nonumber
& \left.+\left(d\psi'+\frac{q}{2}A_{{\cal B}}-\frac{1}{4}A+\frac{q}{4}yD\beta'\right)^{2}\right],
}
where, inspired by the notation in \cite{Gauntlett:2004hh}, we defined $a_{Y^{p,q}}$ to be a constant related to $\alpha_{0}$ and $c_{Y^{p,q}}=-b$. This metric has exactly the form of the near horizon limit of the universal metric in \eqref{redAnsatz11d}.
The seven-dimensional manifold parametrized by $x,\mathcal{B},\psi',\beta'$ at a fixed point on the Riemann surface, coincides with the seven-manifold described in \cite{Gauntlett:2004hh}.\footnote{In their notation we have $\Lambda=8, \lambda=2q$.} The corresponding 4-form reads
\begin{equation}
G_4 = \left(\frac{L}{4}\right)^3 \mathrm{vol}_{\text{AdS}_{2}}\wedge \left(3 \text{vol}_{\Sigma_g} + q \, dx \wedge D\beta'+2 q\, (1-c_{Y^{p,q}}) J_\mathcal{B} \right)\, .
\end{equation}
Deformations of this solution with $c\neq0$, which we do not analyze in full detail here, should thus describe the twisted compactification of the corresponding SCFTs with flavor flux.
\subsection{AdS$_2 \times S^3 \times T^6$}\label{sec:AdS2xS3xT6}
As mentioned in Section \ref{AdS2Hor-UniversalTwist}, there is another solution within our Ansatz with $b\neq 0$, $\kappa=1$ and $c=0$, which reads
\begin{equation}
ds^2 = L^2 \left( ds^2_{\text{AdS}_2}+ ds^2_{S^2} + \frac{3}{4} \frac{x^2\,dx^2}{C+x^3}+\frac{4}{3} \frac{C+x^3}{x^2}\, q^2\, (d\beta'+A_{\mathcal{B}})^2 + x \, q \, ds^2_{\mathcal{B}}+(d\psi'+A)^2 \right) \, ,
\end{equation}
where $x$ is related to $y$ by a shift and rescaling, $\beta'= \beta/b$, $\psi'= 2 (\omega_1 \psi-q \beta')$, and $C$ is a constant. The flux for this solution is given by
\begin{equation}
G_4 = L^3 \, \text{vol}_{\text{AdS}_2} \wedge \left( dx \wedge d\beta'+ x q J_\mathcal{B}\right) \, .
\end{equation}
For $\mathcal{B}=\mathbb{CP}^2$ and $C=0$, the solution becomes the well-known AdS$_2\times S^3\times T^6$ background of 11d supergravity. To see this note that the $S^{2}$ combines with $(d\psi'+A)^2$ into an $S^3$, and $(D\beta')^2$ combines with $ds^2_{\mathbb{CP}^2}$ into an $S^5$.
Then, with the coordinate transformation $x=\rho^2$, we recognize a six-dimensional metric which is simply $\mathbb{R}^6$, written as a cone over $S^5$. We can then periodically identify the coordinates on this $\mathbb{R}^6$ to obtain $T^6$.\footnote{For ${\cal B}=\Bbb P^{1}\times \Bbb P^{1}$ and $C=0$ one obtains a cone over the conifold $Y^{1,0}$. The resulting six-dimensional manifold cannot be made compact and regular.} For $C\neq 0$, instead, the internal manifold is singular and we thus discard the solution.
\subsection{Baryonic fluxes\label{AdS2Hor-BaryonicFluxes}}
As mentioned above, in the special case $b=0$ there is a more general solution to the constraints imposed by Bianchi, even while keeping $c=0$ (i.e., no mesonic fluxes). Note that setting $b=0$ and using the results in \cite{Gauntlett:2004hh} amounts to limiting the discussion to the Sasaki-Einstein manifolds $Q^{1,1,1}$ and $M^{1,1,1}$, and orbifolds thereof. As discussed in detail below, these solutions are characterized by the presence of 4-form flux through non-trivial cycles in $M_9$. This, in four dimensions, means that the corresponding black hole is charged under the Betti multiplets and thus the dual field theory is twisted by baryonic flux, in addition to R-symmetry flux. These charges may be of electric or magnetic type, depending on whether the flux is through AdS$_2$ or $\Sigma_{\mathfrak{g}}$. Here we analyze the general case in which they are both present, leaving to the following paragraphs a more detailed presentation of the purely electric and purely magnetic solutions.\footnote{Black holes charged under Betti multiplets have been numerically found in \cite{Halmagyi:2013sla} in four-dimensional supergravity in the particular case of the homogeneous manifolds $Q^{1,1,1}$ and $M^{1,1,1}$ using the consistent truncation in \cite{Cassani:2012pj}.}
A general result that follows when we set $b=c=0$, is that the functions $f_1$ and $f_2$ are constant.\footnote{This follows from the constraint \eqref{AdS2Hor_F5Eq}, which implies that $F_{5}$ is at most a polynomial of
degree two, \eqref{AdS2Hor_FSummary} and the definitions \eqref{AdS2Hor_FBasis}. } Let us denote these by
\equ{\label{AdS2Hor_BaryonicProof1}
f_{1}\equiv L\,,\qquad f_{2} \equiv \sqrt u \, L\,,
}
with $u,L>0$. Using \eqref{AdS2Hor_FSummary} and the definitions \eqref{AdS2Hor_FBasis} gives the constraint on $\alpha_2$:
\begin{equation}
F_{5}''=2K\left(\left(\kappa-u\right)S_{3}^{2}+4qS_{2}S_{3}\right)=2\alpha_{2}\,.\label{AdS2Hor_BaryonicProof2}
\end{equation}
On the other hand, the parameters $\alpha_0$ and $\alpha_1$ are not constrained, but we will assume that $\alpha_0 \alpha_2<0$.
Then, the part of the metric involving $y$ and $\beta$, namely
\[
K\frac{uL^{2}S_{3}^{2}}{F_{5}}dy^{2}+KF_{5}\frac{1}{uL^{4}S_{3}^{2}}d\beta^{2}\;,
\]
can be recast, via rescaling and shifts of the coordinates $y$ and $\beta$, into the metric of a two-sphere with a certain radius $L\sqrt{v}$:
\[
vL^2\left(\frac{1}{1-y^2}dy^{2}+(1-y^2)d\beta^{2}\right).
\]
This choice of coordinates is equivalent to setting
\begin{equation}
K=vL^{3},\quad\alpha_{2}=-S_{2}S_{3}^{2}\,,\quad\alpha_{0}=-\alpha_{2}\,,\quad\alpha_{1}=0\,.
\label{AdS2Hor_BaryonicProof4}
\end{equation}
Combining \eqref{AdS2Hor_BaryonicProof4} with \eqref{AdS2Hor_BaryonicProof2}, we can determine $S_{3}$
in terms of the other parameters:
\begin{equation}
S_{3}=-\frac{4qS_{2}K}{S_{2}+K\left(\kappa-u\right)}\,.\label{AdS2Hor_BaryonicProof3}
\end{equation}
It is convenient to write $\mathcal{I}_0=wS_{3}^{2}$, with $w$ a real parameter. Then, the general solution reads
\begin{eqnarray}
\label{AdS2Hor_BaryonicGenSol} \nonumber
&&ds_{11}^{2} =\, L^{2}\left(ds_{\text{AdS}_{2}}^{2}+u\,ds_{\Sigma_{\mathfrak{g}}}^{2}+v\,ds_{S^{2}}^{2}+\frac{4quv}{-u+v(u-\kappa)}ds_{{\cal B}}^{2}+\left(d\psi+2q\, A_{{\cal B}}+\kappa \, A+y\, d\beta\right)^{2}\right)\,,\\
&&G_{4} =\,L^{3}\,\mathrm{vol}_{\text{AdS}_{2}}\wedge\left(\left(u-\kappa\right)\mathrm{vol}_{\Sigma_{\mathfrak{g}}}+\left(v-1\right)dy\wedge d\beta+2q\frac{u+v(u+\kappa)}{-u+v\left(u-\kappa\right)}\,J_{{\cal B}}\right)\\ \nonumber
&&\qquad \quad +\, uwL^{3}\left[\mathrm{vol}_{\Sigma_{\mathfrak{g}}}\wedge\left(-2dy\wedge D\beta+J_{\mathcal{B}}\right)+J_{\mathcal{B}}\wedge\left(dy\wedge D\beta-J_{\mathcal{B}}\right)\right]\,.
\end{eqnarray}
The three parameters $u,v,w$ are not independent, but are constrained by the Bianchi identity to satisfy:
\begin{equation}
\frac{4q^{2}\left(-3\kappa^{2}v^{2}+2\kappa u(v-1)v+u^{2}(v-1)(v+3)\right)+3w^{2}(\kappa v-uv+u)^{2}}{(\kappa v-uv+u)^{4}}=0\,.
\label{AdS2Hor_UVWConstraint}
\end{equation}
We thus obtain a two-parameter family of deformations of the near-horizon geometry of the universal black hole associated to either the $Q^{1,1,1}$ or $M^{1,1,1}$ 3d SCFT, according to the choice of ${\cal B}$. The black hole entropy reads:
\equ{\label{entropy baryonic}
S_{\text{BH}}=\frac{u}{8\pi}\,\mathrm{vol}_{\Sigma_{\frak{g}}} \, F_{S^{3}}\,.
}
\subsubsection{Electric baryonic charges}
In this section, we consider a purely electric baryonic flux and therefore set $w=0$, to eliminate any magnetic baryonic fluxes that could arise from the terms proportional to $\mathrm{vol}_{\Sigma_{\mathfrak{g}}}$ in the third line of \eqref{AdS2Hor_BaryonicGenSol}. Setting $w=0$ in the constraint \eqref{AdS2Hor_UVWConstraint} we find that the resulting metric is acceptable only for $\kappa=-1$ and the solution reads:
\begin{equation}
\begin{split}ds_{11}^{2} & =L^{2}\left(ds_{\text{AdS}_{2}}^{2}+u\,ds_{\Sigma_{\mathfrak{g}}}^{2}+v\,ds_{S^{2}}^{2}+\frac{4quv}{v-u+uv}\, ds_{{\cal B}}^{2}+\left(d\psi+2qA_{{\cal B}}- A+y\, d\beta\right)^{2}\right)\,,\\
G_{4} & =L^{3}\,\mathrm{vol}_{\text{AdS}_{2}}\wedge\left(\left(u+1\right)\mathrm{vol}_{\Sigma_{\mathfrak{g}}}+\left(v-1\right)dy\wedge d\beta+2q\,\frac{u-v+uv}{v-u+uv}\,J_{{\cal B}}\right)\,.
\end{split}
\label{AdS2Hor_BaryonicEFluxSol}
\end{equation}
The corresponding entropy, given by \eqref{entropy baryonic}, reads
\equ{
S_{\text{BH}}=\frac{u}{2}\,(\frak{g}-1)\, F_{S^{3}}\,.
}
We note that in the special case $u=v=2$ the solution reduces to the universal twist of $M^{1,1,1}$ or $Q^{1,1,1}$. The latter was first derived in \cite{Donos:2012sy} (see Equation (3.25) there). To better understand the physical interpretation of the deformation consider the reduction of this eleven-dimensional background to four dimensions. The terms in the 4-form proportional
to $H^{2}(M_7)$ give rise to four-dimensional Betti vector multiplets.
Since $G_{4}$ in \eqref{AdS2Hor_BaryonicEFluxSol} is proportional to $\mathrm{vol}_{\text{AdS}_{2}}$, the black hole
is {\it electrically} charged under this vector field, as measured by the flux of $\ast_{11}\,G_{4}$ through
non-trivial 7-cycles.
Let us take for example $Q^{1,1,1}$. This space is a fibration of $\psi$ over the product of three two-spheres, which we denote by $\mathcal{S}_{1},\mathcal{S}_{2},\mathcal{S}_{3}$. The generators of $H^{2}(Q^{1,1,1})$
are then given by
\begin{equation}
h_{1}=\mathrm{vol}_{\mathcal{S}_{1}}-\mathrm{vol}_{\mathcal{S}_{2}}\,,\qquad h_{2}=\mathrm{vol}_{\mathcal{S}_{2}}-\mathrm{vol}_{\mathcal{S}_{3}}\,,
\label{AdS2Hor_Q111H2}
\end{equation}
since\footnote{Here by $D\psi$ we mean $d\psi+A_{\text{KE}_{6}}$. }
\[
\mathrm{vol}_{\mathcal{S}_{1}}+\mathrm{vol}_{\mathcal{S}_{2}}+\mathrm{vol}_{\mathcal{S}_{3}}=d(D\psi)
\]
is trivial. Define now the 7-cycles $C_{1},C_{2},C_{3}$ as
the submanifolds obtained by fixing a point in AdS$_{2}$ and $\mathcal{S}_{1},\mathcal{S}_{2},\mathcal{S}_{3}$, respectively. Cohomology suggests that the
non-trivial 7-cycles on which we should integrate $\ast_{11}\,G_{4}$ are
\[
H_{1}=C_{1}-C_{2}\,,\qquad H_{2}=C_{2}-C_{3}\,.
\]
Denoting the sphere spanned by $(y,\beta)$ by $\mathcal{S}_{1}$, it is easy to see that only the Betti multiplet associated to $H_{1}$ has a non-vanishing
flux,\footnote{\label{footbary}Since $b_{2}(Q^{1,1,1})=2$ one should expect a more general solution with two electric baryonic fluxes. The fact that we only see one flux here can be understood as follows. Recall that in our Ansatz \eqref{AdS2Hor_metricAnsatz}, \eqref{AdS2Hor_4FormAnsatz} we have assumed the base ${\cal B}$ is a K\"ahler-Einstein manifold. This implies that $\mathrm{vol}_{\mathcal{S}_{2}}$ and $\mathrm{vol}_{\mathcal{S}_{3}}$ cannot appear in an arbitrary combination, but instead are always added so as to give $J_{\mathcal{B}}$. If one relaxes the Einstein condition on ${\cal B}$, the resulting equations are almost identical to the ones presented here and their solutions allow for both baryonic fluxes to be present at once. In contrast, since $b_{2}(M^{1,1,1})=1$ our Ansatz captures the most general baryonic solution in that case.} due to the presence of $J_{\mathcal{B}}=\mathrm{vol}_{\mathcal{S}_{2}}+\mathrm{vol}_{\mathcal{S}_{3}}$
in $G_{4}$:
\begin{alignat}{1}
\mathcal{F}_{1} & \equiv\int_{H_{1}}\ast_{11}\,G_{4}=-64\, \pi^{3}\, (\mathfrak{g}-1)\, \Delta_{\psi}\, L^{6}\, \frac{u v (u (v-1)-v) (u (v-3)+v)}{(u (v-1)+v)^2}\,\label{AdS2Hor_ElFlux},\\
\mathcal{F}_{2} & \equiv\int_{H_{2}}\ast_{11}\, G_{4}=0\,\nonumber,
\end{alignat}
where $\Delta_{\psi}$ is the period of $\psi$. Since Betti multiplets are dual to baryonic symmetries in the field theory, for general values of the parameter $v>0$ (or $u>0$), the solution describes the twisted compactification with baryonic fluxes. In the special case $u=v=2$ one has $\mathcal{F}_{1}=0$ and there is no baryonic flux. Indeed, the 4-form can be written as
\[
G_{4}=L^{3}\,\mathrm{vol}_{\text{AdS}_{2}}\wedge\left(4\, \mathrm{vol}_{\Sigma_{\mathfrak{g}}}+d\,(D\psi)\right)\,,
\]
and only a trivial cocycle of $M_7$ appears.
The solution described here is reminiscent of the AdS$_3$ solutions of type IIB supergravity presented in Section~2.2 of \cite{Benini:2015bwz} (see also \cite{Amariti:2016mnz} for a five-dimensional perspective on this solution), describing the twisted compactification of the Klebanov-Witten theory (and its $\Bbb Z_{p}$ orbifolds) with magnetic baryonic flux.
\subsubsection{Magnetic baryonic charges}
\label{subsubsec:mag}
Magnetic baryonic charges can be obtained by a non-zero $w$. For concreteness, we present the case $\mathcal{B}=\mathbb{CP}^{1}\times\mathbb{CP}^{1}$ (the $\mathbb{CP}^{2}$ case is analogous) and we analyze the solution with a purely magnetic baryonic flux and thus impose that $\eqref{AdS2Hor_ElFlux}$ vanishes. Combining this constraint with \eqref{AdS2Hor_UVWConstraint} determines $u$ and $v$ in terms of $w$. In contrast to the purely electric case, there are solutions for all $\kappa=\{0,\pm1\}$.
For $\kappa=-1£ì$ we find
\begin{equation}
\begin{split}ds_{11}^{2} & =\, L^{2}\left(ds_{\text{AdS}_{2}}^{2}+\left(3w^2+2\right)\, ds_{\Sigma_{\mathfrak{g}}}^{2}+\frac{3w^2+2}{3w^2+1}\, ds_{S^{2}}^{2}+\left(3w^2+2\right)ds_{{\cal B}}^{2}+D\psi^{2}\right)\,,\\
G_{4} & =\,L^{3}\,\mathrm{vol}_{\text{AdS}_{2}}\wedge\left(\left(3w^2+3\right)\mathrm{vol}_{\Sigma_{\mathfrak{g}}}+\frac{1}{3w^2+1}\,dy\wedge D\beta+\left(3w^2+1\right)J_{\mathcal{B}}\right)\\
& + L^{3}\left(3w^2+2\right)w\left[\mathrm{vol}_{\Sigma_{\mathfrak{g}}}\wedge\left(-2dy\wedge d\beta+J_{\mathcal{B}}\right)+J_{\mathcal{B}}\wedge\left(dy\wedge d\beta-J_{\mathcal{B}}\right)\right]\,.
\end{split}
\label{AdS2Hor_BaryonicMFluxSol}
\end{equation}
The universal solution is recovered for $w=0$. For any other value the baryonic charges are given by the integral of $G_4$ through the non-trivial 4-cycles of $M_9$, whose form is suggested by $\eqref{AdS2Hor_Q111H2}$ and schematically reads
\[
\tilde{H}_1 \equiv \Sigma_{\mathfrak{g}}\times\mathcal{S}_1-\Sigma_{\mathfrak{g}}\times\mathcal{S}_2\,,\qquad \tilde{H}_2 \equiv \Sigma_{\mathfrak{g}}\times\mathcal{S}_2-\Sigma_{\mathfrak{g}}\times\mathcal{S}_3\,.
\]
The result for the fluxes is
\begin{alignat}{1}
\mathcal{\tilde{F}}_{1} & \equiv\int_{\tilde{H}_{1}}G_{4}=-48\,|\mathfrak{g}-1|\,\pi^{2}\,L^{3}\,\left(3w^2+2\right)w\,\label{AdS2Hor_MagFlux},\\
\mathcal{\tilde{F}}_{2} & \equiv\int_{\tilde{H}_{2}}G_{4}=0\,.\nonumber
\end{alignat}
The entropy reads
\equ{
S_{\text{BH}}= \frac12 \,(3 \,w^2+2) \, (\frak{g}-1)\, F_{S^{3}}\,.
}
For $\kappa=1$ and $|w|>1$, we find
\begin{equation}
\begin{split}ds_{11}^{2} & =L^{2}\left(ds_{\text{AdS}_{2}}^{2}+\frac{w^{2}+2}{w^{2}-1}\,ds_{S^2}^{2}+(w^{2}+2)\,ds_{S^{2}}^{2}+(w^{2}+2)\,ds_{{\cal B}}^{2}+D\psi^{2}\right)\,,\\
G_{4} & =L^{3}\, \mathrm{vol}_{\text{AdS}_{2}}\wedge\left(\frac{3}{w^{2}-1}\,\mathrm{vol}_{S^2}+\frac{3}{w^{2}-1}\,d\theta_{1}\wedge d\phi_{1}+(w^{2}+1)\,J_{{\cal B}}\right)\\
& +w\,\frac{w^{2}+2}{w^{2}-1}\,L^{3}\,\left[\,\mathrm{vol}_{S^2}\wedge\left(-2\,dy\wedge d\beta+J_{\mathcal{B}}\right)+J_{\mathcal{B}}\wedge\left(dy\wedge d\beta-J_{\mathcal{B}}\right)\right]\,.
\end{split}
\label{AdS2Hor_BaryonicMFluxSolAlt}
\end{equation}
The resulting magnetic flux along $\tilde{H}_1$ is
\begin{equation}
\mathcal{\tilde{F}}_{1} =-48\pi^{2}L^{3}\, w\, \frac{w^{2}+2}{w^{2}-1}\,,
\end{equation}
and the entropy is given by
\begin{equation}
S_{\text{BH}}=\frac{1}{2}\,\frac{w^2+2}{w^2-1}\,F_{S^{3}}\,.
\end{equation}
For the special value $w=\sqrt{2}$ we obtain a solution analogous to the one presented in \cite{Donos:2012sy} (see Equation (3.16) there), where the other Betti multiplet was turned on. As discussed in Footnote \ref{footbary}, the absence of the second Betti multiplet is a consequence of our Ansatz.
Finally, for $\kappa=0$ the solution can be expressed in terms of $u$ alone and reads\footnote{The constraint \eqref{AdS2Hor_UVWConstraint} has two acceptable solutions leading to the two signs in the last line of \eqref{AdS2Hor_MagBarTorusSol}. This choice of sign amounts to the sign of the magnetic baryonic charge.}
\begin{equation}
\label{AdS2Hor_MagBarTorusSol}
\begin{split}ds_{11}^{2} & =L^{2}\left(ds_{\text{AdS}_{2}}^{2}+u\,ds_{T^{2}}^{2}+3\,ds_{S^{2}}^{2}+3\,ds_{{\cal B}}^{2}+\left(d\psi+A_{{\cal B}}+y\,d\beta\right)^{2}\right)\,,\\
G_{4} & =L^{3}\,\mathrm{vol}_{\text{AdS}_{2}}\wedge\left(u\,\mathrm{vol}_{T^{2}}+2\,dy\wedge d\beta+2J_{{\cal B}}\right)\\
&\quad \pm u\,L^{3}\,\left[\,\mathrm{vol}_{T^{2}}\wedge\left(-2\,dy\wedge D\beta+J_{\mathcal{B}}\right)+J_{\mathcal{B}}\wedge\left(dy\wedge D\beta-J_{\mathcal{B}}\right)\,\right]\,,
\end{split}
\end{equation}
By fixing $u$ one recovers a solution analogous to the one found in \cite{Donos:2012sy} (see Equation (2.17) there). The entropy and baryonic flux are given by
\begin{equation}
S_{\text{BH}}=\frac{1}{4}\,u\, F_{S^{3}}\,, \qquad \mathcal{\tilde{F}}_{1}=-48\pi^{2}L^{3}\,u\,.
\end{equation}
\subsection{Field theory interpretation\label{FTHor-BaryonicFluxes}}
It would be very interesting to reproduce the entropy of the black hole horizons we found above through a field theory computation. The dual three-dimensional SCFTs for some specific $Y^{p,q}(\mathcal{B})$ have been studied in \cite{Martelli:2008si,Benini:2009qs,Benini:2011cma}. The dual of $M^{1,1,1}/\mathbb{Z}_k= Y^{2k,3k}(\mathbb{CP}^2)$ is a chiral theory and,
as we already mentioned, the large $N$ limit of the three-sphere partition function as well as the topologically twisted index are not very well understood. The quiver for
$Q^{1,1,1}/\mathbb{Z}_k= Y^{k,k}(\mathbb{CP}^1 \times \mathbb{CP}^1)$ on the other hand is vector-like \cite{Benini:2009qs} and thus we have calculational control over the large $N$ limit of the localization calculations.
Any comparison of the black hole entropy for the near-horizon geometries discussed above with the field theory would require a proper understanding
of the cycles of the internal manifold and of the quantization of fluxes. The solution with no mesonic flavor and only baryonic magnetic flux presented in Section \ref{subsubsec:mag} should be the simplest case study. However, for reasons that are still elusive, the fugacity and flux parameters associated with baryonic symmetries seem to disappear in the large $N$ limit of the twisted index, as is evident from the results in \cite{Hosseini:2016tor,Hosseini:2016ume}.
The same phenomenon has been noticed for the $S^3$ free energy in \cite{Jafferis:2011zi}. These are all very interesting open problems that we leave for future work.
\section{Discussion}\label{sec:discussion}
In this paper we provided overwhelming evidence for the existence of a universal RG flow across dimensions between partially topologically twisted three-dimensional $\mathcal{N}=2$ SCFTs on $S^1\times \Sigma_\frak{g}$ and one-dimensional quantum mechanical theories with two supercharges. This flow results in a universal relation between the topologically twisted index and the three-sphere partition function of the $\mathcal{N}=2$ SCFT, given by \eqref{eq:univintro}. If the three-dimensional SCFT admits a weakly-coupled supergravity dual description this universal RG flow has a simple holographic description in terms of a supersymmetric black hole solution of minimal four-dimensional gauged supergravity. This solution can then be embedded into string or M-theory in infinitely many distinct ways, which distinguish the different features of the $\mathcal{N}=2$ SCFTs.
There are a number of interesting open questions that stem from our results. The most obvious one is to understand in detail the plethora of AdS$_2$ solutions in M-theory presented in Section \ref{AdS2Hor}. The natural guess is that many of these solutions correspond to twisted compactifications of three-dimensional $\mathcal{N}=2$ SCFTs with background flavor and baryonic fluxes. The three-dimensional theories should arise from M2-branes probing a $CY_4$ singularity in M-theory. The topological twist should then be realized by wrapping these branes on $\Sigma_\frak{g}$ and the AdS$_2$ vacua we find should describe the low-energy behavior of this system upon backreaction. Clearly, these expectations await an explicit confirmation which hinges on a more detailed understanding of the global properties of the AdS$_2$ backgrounds in Section \ref{AdS2Hor}. A particularly puzzling feature is that in supergravity the background flux for baryonic $U(1)$ symmetries affects the details of the AdS$_2$ vacuum and thus the black hole entropy. On the other hand, it seems that such baryonic magnetic fluxes do not change the large $N$ limit of the topologically twisted index. It would be most interesting to resolve this apparent puzzle.
Another interesting avenue for generalizing our results is to incorporate electric charges into the supersymmetric black hole solutions we studied. While one can show that there is no dyonic generalization of the universal twist, it is clear that the more general supergravity solutions in Section \ref{AdS2Hor} should admit dyonic counterparts.\footnote{These dyonic solutions should be additional new examples of the type of dyonic black holes studied in the four-dimensional gauged supergravity literature. See for example \cite{Cacciatori:2009iz,DallAgata:2010ejj,Hristov:2010ri,Halmagyi:2013qoa} and reference thereof.} The entropy of these dyonic black hole near-horizon geometries can then be accounted for by a generalization of the topologically twisted index which incorporates both electric and magnetic flavor fluxes. This was pursued successfully in \cite{Benini:2016rke} for supersymetric dyonic black holes asymptotic to AdS$_4\times S^7$ which are dual to the ABJM theory on $S^1\times \Sigma_\frak{g}$ deformed by magnetic fluxes on $\Sigma_\frak{g}$ and Wilson lines along $S^1$.
We have limited the discussion in this work to the large $N$ limit of the topologically twisted index which, for theories with a holographic dual, is captured by the supergravity approximation of string or M-theory. It is of great interest to go beyond this limit, both in the field theory as well as in the gravitational analysis. On the field theory side it is natural to ask whether some remnant of the universal twist relation in \eqref{eq:univintro} survives beyond the large $N$ approximation. There might be reasons to be cautiously optimistic, given that similar universal relations were derived in \cite{Benini:2015bwz,BC} for the conformal anomaly coefficients of even-dimensional SCFTs related by RG flows across dimensions. On the gravity side the problem is equally challenging. One has to find subleading (in $N$) corrections to the entropy of the universal black hole presented in Section \ref{sec:A simple 4d black hole}. Perhaps the methods employed in \cite{Sen:2011ba} can be useful in finding these corrections. It is worth noting that a similar question was addressed successfully in \cite{Bhattacharyya:2012ye} for the subleading corrections to the $S^{3}$ partition function for three-dimensional $\mathcal{N}=2$ theories with an AdS$_4$ dual in M-theory. Resolving this question for the class of asymptotically AdS$_4$ black holes studied here is bound to teach us important lessons about holography and the quantum structure of black holes.
\bigskip
\bigskip
\bigskip
\bigskip
\bigskip
\bigskip
\bigskip
\noindent \textbf{ Acknowledgements }
\medskip
\noindent We would like to thank Francesco Benini, Davide Cassani, Fri\dh rik Gautason, Adolfo Guarino, Seyed Morteza Hosseini, Dario Martelli, Noppadol Mekareeya, Krzysztof Pilch, James Sparks, and Phil Szepietowski for interesting discussions. FA is partially supported by INFN. The work of NB is supported in part by the starting grant BOF/STG/14/032 from KU Leuven and by an Odysseus grant G0F9516N from the FWO. The work of VSM is supported by a doctoral fellowship from the Fund for Scientific Research - Flanders (FWO) and in part by the ERC grant 616732-HoloQosmos. NB and VSM are also supported by the KU Leuven C1 grant ZKD1118 C16/16/005, by the Belgian Federal Science Policy Office through the Inter-University Attraction Pole P7/37, and by the COST Action MP1210 The String Theory Universe. PMC is supported by Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO) via a Vidi grant. The work of PMC is part of the Delta ITP consortium, a program of the NWO that is funded by the Dutch Ministry of Education, Culture and Science (OCW). AZ is partially supported by the INFN and ERC-STG grant 637844-HBQFTNCER. NB and AZ would like to thank KIAS, Seoul for warm hospitality in the beginning stages of this project.
\begin{appendices}
|
\section{Introduction}
Accurate clip-level video classification, utilising a rich vocabulary of sophisticated terms, remains a challenging problem.
One of the contributing factors is the complexity and ambiguity of the interrelations between linguistic terms and the actual audio-visual content of the video. For example, while a "travel" video can depict any location with any accompanying sound, it is the {\em intent of the producer} or even the {\em perception of the viewer} that makes it a "travel" video, as opposed to a "news" or "real estate" clip. Hence true {\em understanding} of the video's meaning is called for, and not mere {\em recognition} of the constituent locations, objects or sounds.
The recent Kaggle competition entitled "Google Cloud \& YouTube-8M Video Understanding Challenge" provides a unique platform to benchmark existing methods and to develop new approaches to video analysis and classification. It is based around the {\bf YouTube-8M (v.2)} dataset, which contains approximately 7 million individual videos, corresponding to almost half a million hours (50 years!), annotated with a rich vocabulary of 4716 labels \cite{45619}. The challenge for participants was to develop classification algorithms which accurately assign video-level labels.
Given the complexity of the video understanding task, where humans are known to use diverse clues, we hypothesise that a successful solution must efficiently combine different expert models. We pose two important questions: (i) How do we construct such diverse models and how to combine them?, and (ii) do we need to individually train and combine discrete models or can we simply train a very large/flexible DNN to obtain a fully trained end-to-end solution? The first question clearly links to ensemble-based classifiers, where significant body of prior work demonstrates that diversity is important. However, do we know all the different ways to promote diversity in DNN architectures? On the second question, our analysis shows that training a single network results in sub-optimal solutions as compared to an ensamble.
In the following section we briefly review the state-of-the-art in video labelling and ensemble-based classifiers. We then introduce the Kaggle competition, including data-sets, performance measures and the additional features engineered and evaluated by the Yeti team. Next, in Section \ref{sec:dnns}, we describe the different forms of DNNs that were employed and quote the baseline performance of individual DNNs trained on different features. Section \ref{sec:exp} demonstrates that further gains in performance can be achieved by promoting diversification of DNNs during training by adjusting dropout rates, different architectures and - surprisingly - using over-fitted DNNs. We then provide analysis on the link between diversity of the DNNs in the final {\em Yeti} ensemble, performance gains in Section \ref{sec:diverse_analysis}, and conclude in Section \ref{sec:conclusions}.
\section{Related Work}
We first overview some existing approaches to video classification before discussing ensemble-based classifiers. Ng et al. \cite{7299101} introduced two methods which aggregate frame-level features into video-level predictions: Long short-term memory (LSTM) and Feature pooling. Fernando et al. \cite{7458903} proposed a novel rank-based pooling method that captures the latent structure of the video sequence data.
Karpathy et al. \cite{6909619} investigated several methods for fusing information across temporal domain and introduced Multiresolution CNNs for efficient video-classification.
Wu et al. \cite{Wu} developed a multi-stream architecture to model short-term motion, spatial and audio information respectively. LSTM are then used to capture long-term temporal dynamics.
DNNs are known to provide significant improvement in performance over traditional classifiers across a wide range of datasets. However, it was also found that further significant gains can be achieved by constructing ensembles of DNNs. One example is the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) \cite{ILSVRC15}. Here, improvements up to 5\% were achieved over individual DNN performance (e.g. GoogLeNet\cite{googlenet}) by using ensembles of existing networks. Furthermore, all the top entries in this challenge employed ensembles of some form.
One of the key reasons for such a large improvement was found to be due to the diversity present across different base classifiers (i.e. different classifiers specialise to different data or label subsets)\cite{Hansen,Krogh95}. An increase in diversity of classifiers of equal performance will usually increase the ensemble performance. There are numerous methods for achieving this: random initialisation of the same models, or data modification using Bagging \cite{Breiman} or Boosting \cite{Boosting} processes. Recently, work was carried out on end-to-end training of an ensemble based on diversity-aware loss functions. Chen et al. \cite{NCL} proposed to use Negative Correlation Learning for promoting diversity in an ensemble of DNNs, where a penalty term based on the covariance of classifier outputs is added to an existing loss function. An alternative was proposed by Lee et al \cite{Mheads} based on the approach of Multiple Choice Learning (MCL) \cite{MCL}. Here, DNNs are trained based on a loss function that uses the final prediction chosen from an individual DNN with the lowest independent loss value.
\section{Youtube-8M Kaggle competition}
\label{sec:data}
The complete Youtube-8M dataset consists of approximately 7 million Youtube videos, each approximately 2-5 minutes in length, with at least 1000 views each. There are 4716 possible classes for each video, given in a multi-label form. For the Kaggle challenge, we were provided with 6.3 million labelled videos (i.e. each video was associated with a 4716 binary vector for labels). For test purposes, approximately 700K unlabelled videos were provided. The resulting class test predictions from our trained models were uploaded to the Kaggle website for evaluation.
The evaluation measure used is called `GAP20'. This is essentially the mean average precision of the top-20 ranked predictions across all examples. To calculate its value, the top-20 predictions (and their corresponding ground-truth labels) are extracted for each test video. The set of top-20 predictions for all videos is concatenated into a long list of predictions. A similar process is performed for the corresponding ground-truth labels. Both lists are then sorted according to the confidence prediction values and mean average precision is calculated on the resulting list.
Below we present different features used for classification; the first two (FF, MF) were provided by Google, the remaining ones were computed by our team. For some features, we also quote the performance as a rough guide of the usefulness of the individual feature: this was computed by 4k-4k-4k DNN with dropout $0.4$.
\begin{itemize}
\item{\bf Frame-level Deep Features (FL)}\\
In the Kaggle challenge, the raw frames (image data) of the videos were not provided. Instead, each video in the dataset was decoded at 1 frame-per-second up to first 300 seconds and then passed through an Inception-v3 network \cite{44903}. The ReLu activation values of the last hidden layer formed a frame-level representation (2048 dimensions), which was subsequently reduced to 1024 dimensions using a PCA transformation with whitening. Similar processing was performed on the audio stream, resulting in an additional 128-dimensional audio feature vector. Video and audio features are concatenated to yield a frame feature vector of 1152 dimensions. The set of frame-level deep features extracted for a video $I$ is denoted as $\mathcal X^{I}=\{x_{t}\in \mathbb{R}^{d}, t=1...T\}$.
\end{itemize}
The extracted features are then aggreagted using state-of-the-art aggregation methods: Mean aggregation, Mean + Standard Deviation aggregation, ROI-Pooling \cite{Radenovic2016CNNIR}, VLAD, Fisher Vectors \cite{Jegou12PAMI}, RVD \cite{RVD} \cite{BRVD} \cite{RVD_PAMI} and BoW.
\begin{itemize}
\item{\bf Video-Level Mean Features (MF)}\\
Google also provided the mean feature $\mu^I$ for each video, which was obtained by averaging frame-level features across the time dimension.
Our reference performance for MF feature is $81.94\%$, but it can peak at $82.55\%$ with a 12k-12k-12k network and dropout of $0.4$.
\item{\bf Video-Level Mean Features + Standard Deviation (MF+STD)}\\
We extract the standard deviation feature $\sigma^I$ from each video. The signature $\sigma^I$ is L2-normalised and concatenated with mean feature $\mu^I$ to form a 2304-Dim representation $\phi^I=[\mu^I;\sigma^I]$.
\item{\bf Region of Interest pooling (ROI)}\\
The ROI-pooling based descriptor, proposed by Tolias et al \cite{Radenovic2016CNNIR}, is a global image representation that achieves state-of-the-art performance in image retrieval and classification. We compute a new video-level representation using the ROI-pooling approach. More precisely, the frame-level features are max-pooled across 10 temporal-scale overlapping regions, obtained from a rigid-grid covering the frame-level features, producing a single signature per region. These region-level signature are independently L2-normalised, PCA transformed and subsequently whitened. The transformed vectors are then sum-aggregated and finally L2-normalised. The dimensionality of final video-level representation is 1152. The ROI-based training architecture is presented in Fig. \ref{GDES}(b); it achieves $82.34\%$ with the 12k-12k-12k net.
\item{\bf Fisher Vectors, RVD and VLAD pooling (FV, RVD, VLAD)}\\
We encode the frame level features using the classical Fisher Vectors, RVD and VLAD approaches. Fisher Vector encoding aggregates local features based on the Fisher Kernel framework while VLAD is simplified version of Fisher vectors. The detailed experimental results show that the mean pooling achieves significantly better classification accuracy than FV, RVD and VLAD approaches (81.94\% vs 81.3\%, 80.8\% and 80.4\%) [4k-4k-4k network].
\item{\bf BoW pooling (BoW)}\\
We compute the BoW representation of the frame-level video features, using 2k and 10k BOW representations. We compute BoW features by first applying K-Means clustering across the frame-level deep features with either 2k or 10k clusters, and then calculating the number of frames in each cluster for each video. Finally, we L1-normalize this BoW vector to remove the effect of video length on the features. The base BoW performance is $78.1\%$ with the 4k-4k-4k net.
\end{itemize}
\section{DNN-based Multi-Label Classifiers}
\label{sec:dnns}
This section describes the base neural network architectures that we used for multi-label predictions on this dataset.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{ROI.pdf}
\caption[\bf{CNN architectures}]{(a) Mean features CNN (b) ROI features CNN (c) Audio Visual fusion CNN.}
\label{GDES}
\end{figure}
\subsection{Fully Connected NN Architecture}
For our work, we use a 3-hidden layer fully connected neural network, with layers FC6, FC7 and FC8. The size of the input layer is dependent on the feature vectors chosen. These will be described in more detail in the sections below.
The activation function for all hidden units is ReLU. Additionally, we also employ dropout on each hidden layer.
The number of hidden units and dropout value will be detailed in Section \ref{sec:exp}.
The output layer is again a fully connected layer, with a total of 4716 output units, one for each class. In order to provide class-prediction probability values, each output unit uses a sigmoid activation function.
Since this challenge is a multi-label problem, we used the binary cross entropy loss function for training the DNNs. We have chosen the Adam optimization algorithm \cite{adam} for training, with a learning rate of $1e-4$. Using the learning rate of $1e-3$ as in the original paper led to NNs getting stuck in local minima very early in training.
All the DNNs were trained to convergence, and the number of epochs required to achieve this ranged from 15 to 150 depending on the hyper-parameter settings chosen, as detailed in Section \ref{sec:exp}.
\subsection{Audio Visual Fusion}
This method comprises of two stages: Firstly the audio and visual features networks are trained separately to minimise the classification loss and then these two NNs are integrated in a fusion network consisting of two fully-connected layers.
1) Training audio and video networks: We first train the audio and video networks individually. This is conducted by connecting their features to three fully connected layers similar to FC6, FC7 and FC8, respectively. The size of all FC layers is 4096. Each FC layer is followed by a ReLu and a dropout layer. The output of the FC8 layer is passed through another fully connected layer FC9 which computes the predictions and finally updates the network parameters to minimise the cross entropy loss over the training data.
2) Training fusion networks: After training the audio and video networks, we discard their FC9 layers and connect their FC8 layers to the fusion network shown in Fig. \ref{GDES}(c). In this way, 4096-Dim audio and 4096-Dim video features are concatenated to form a 8192-Dim representation as an input to the fusion network. This fusion network contains two fully connected layers of size 8192, followed by a fully connected prediction layer and cross-entropy optimisation.
The model based on Audio-Visual Fusion achieved $82.28\%$, but added significant diversity to our ensemble.
\subsection{Ensemble of DNNs}
The test predictions from multiple DNNs that were trained separately with different architectures, input features and hyper-parameters can be combined together by averaging them. We have found such ensembles often provide significant improvements over the performance of individual DNNs. The details of the diversification process and ensemble construction is presented in Section \ref{sec:exp}.
\section{Diversification of DNNs}
\label{sec:exp}
It was found that the performance of individual models (architectures and input features) could be significantly improved when they were combined together into a DNN ensemble. However, in order to achieve these gains, it was necessary to build diverse DNNs. In this section, we describe a number of approaches that we have attempted that were mostly successful in achieving the aim of diversification of DNNs. These range from using different dropouts, hidden unit counts, use of overfitted models and segmented frame-level features.
\subsection{Sizes of Hidden Layers}
For our experiments, we have considered the use of the following number of units for each hidden layers: 4096, 8192, 10240, 12288 and 16384. All the hidden layers within a model were set to have the same number of hidden units, as we did not see substantial gains by varying the hidden layer size within a model.
\subsection{Dropout Sizes}
In the process of training different DNNs, a number of different dropout values were used: 0.0 (no dropout), 0.25, 0.3, 0.4 and 0.5. As expected, we have found that the higher the dropout, the larger the number of epochs required for convergence to be achieved.
\subsection{Use of Overfitted DNNs}
We have also used models that are overfitting. We have found that individual DNNs will have a validation GAP20 score that peaks after a certain number of training epochs (usually around 40-50 for large networks of ${>}$8K units). If training continues, we find that the validation GAP20 score will steadily decrease. This implies that the model is overfitting to the training data. Existing practise often discards these models and use the model with the best validation score.
Counter-intuitively, using large network models that have overfitted was found to give a {\em larger} performance improvement to the ensemble of classifiers. This is despite individual validation GAP scores that are less than its peak GAP score many epochs prior.
\subsection{Using Different Training Subsets}
Finally, we have explored how using different training subsets for building similar architectures can influence the final performance of the ensemble. To this end, we first trained a DNN ensemble using DNNs with the above different hidden units, dropouts and feature vectors (ROI, video mean, video mean + std. dev.) using the training dataset and validation set (except last 100K validation data) provided by Google.
We then produced another set of training data and 100K validation data that was split differently to the above. Next, DNNs with 4K-4K-4K, 4K-8K-8K, 8K-8K-8K, 10K-10K-10K, 12K-12K-12K, 14K-14K-14K and 16K-16K-16K architectures were trained using a separate training set for the video-mean features. These were used to form a separate DNN ensemble. We have found that this also provides improvement when outputs of both ensembles are linearly combined together.
\subsection{Diversification-based Loss Function}
\label{sec:divbased}
Recently, there has been work on performing end-to-end learning of multiple DNN output layers that promote diversity \cite{Opitz2017} for the task of multi-class classification. Here, a multi-output layer DNN was proposed. The final output label is the class with the maximum votes from all the output layers. In order to learn this DNN, a ``diversity-aware'' loss function was proposed. This was a linear combination of the MSE error with the sum of cross-entropy of outputs for the different layers. The aim was not only to have each output layer minimise classification error, but to also provide classification outputs that are different from other output layers.
We have attempted to use a similar approach to sequentially train diversity-aware DNNs. In order to achieve this, we first train a single 3-layer fully connected DNN as described above. Our ensemble is initialised using this single DNN. The outputs for the training data of this ensemble is then recorded for subsequent use.
In order to further add new DNNs into our ensemble, we wish to learn DNNs that minimise labelling errors {\em and} produce outputs that are different to that of the ensemble.
Learning the next DNNs was performed by proposing a loss function that accounts for multi-label classification accuracy and is also diversity-aware. For the multi-label classification, we have used the binary cross entropy method. For the diversity awareness, we use the negative of cross entropy between the current DNN and the ensemble output. The final loss function is a linear combination between the above two losses, with a combination parameter of $\lambda = 0.3$ for diversity and $0.7$ for multi-label accuracy. The new DNN is then added into the ensemble set and this step repeated a number of times until a pre-defined ensemble size is reached (here we chose 4).
\section{Experimental Results}
In this section, we shall provide results from the individual models. We also show how a significantly improved GAP20 score can be achieved by combining the individual classifiers into an ensemble. We further achieve improvement by performing linear combination on ensembles trained using different training sets.
\subsection{Performance of Individual Features}
In this section, we detail the baseline performances of DNNs that were trained on the different features described above.
\begin{table}[h!]
\centering
\begin{tabular}{|c c c c c|}
\hline
Features & Architecture & DropOut & GAP20 & Peak \\
\hline \hline
Mean & 4k-4k-4k & 0.25 & 81.94 & 81.94 \\
\hline
Mean & 4k-4k-4k & 0.3 & 82.01 & 82.01 \\
\hline
Mean & 4k-4k-4k & 0.4 & 82.08 & 82.08\\
\hline
Mean & 8k-8k-8k & 0.4 & 82.48 & 82.48\\
\hline
Mean & 12k-12k-12k & 0.4 & 82.31 & 82.55\\
\hline
ROI & 8k-8k-8k & 0.4 & 82.25 & 82.25\\
\hline
ROI & 12k-12k-12k & 0.4 & 82.12 & 82.34\\
\hline
Fusion & 8k-8k-8k & 0.4 & 82.28 & 82.28\\
\hline
Mean+sd & 8k-8k-8k & 0.3 & 82.1 & 82.1\\
\hline
Mean+sd & 10k-10k-10k & 0.4 & 82.54 & 82.54\\
\hline
Mean+sd & 12k-12k-12k & 0.4 & 82.41 & 82.6\\
\hline
\end{tabular}
\bigskip
\caption{Table showing the GAP performance of different architectures for the first ensemble, features and dropout settings. Shown are two GAP scores, one at the last epoch (GAP20) and another at the epoch where the peak GAP score was achieved. }
\label{table:indiv_perf}
\end{table}
It can be observed from the Table \ref{table:indiv_perf} that GAP20 score increases as we increase the dropouts percentage, keeping the rest of the network hyperparameters the same. Also the deeper 8k-8k-8k architecture performs significantly better than 4k-4k-4k using dropout 0.4. Furthermore, adding second order statistics (standard deviation features) to mean features increases the GAP20 from 82.0\% to 82.1\%. The ROI and Fusion CNNs performs marginally less that Mean CNNs. However, all the architectures presented add value to the overall performance.
\begin{table}[h!]
\centering
\begin{tabular}{|c c c c c|}
\hline
Features & Architecture & DropOut & GAP20 & Peak \\
\hline \hline
Mean & 4k-4k-4k & 0.25 & 81.87 & 81.87 \\
\hline
Mean & 4k-4k-4k & 0.3 & 81.92 & 81.92 \\
\hline
Mean & 4k-4k-4k & 0.4 & 82.01 & 82.01\\
\hline
Mean & 8k-8k-8k & 0.4 & 82.15 & 82.48\\
\hline
Mean & 10k-10k-10k & 0.4 & 82.18 & 82.28\\
\hline
Mean & 12k-12k-12k & 0.4 & 82.11 & 82.38\\
\hline
Mean & 14k-14k-14k & 0.4 & 82.16 & 82.24\\
\hline
Mean & 16k-16k-16k & 0.4 & 82.20 & 82.41\\
\hline
Mean+sd & 10k-10k-10k & 0.4 & 82.39 & 82.43\\
\hline
Mean+sd & 12k-12k-12k & 0.4 & 82.37 & 82.45\\
\hline
\end{tabular}
\bigskip
\caption{Table showing the GAP performance of different architectures for the second ensemble, trained with a different train-validation split, input features and dropout settings. Shown are two GAP scores, one at the last epoch (GAP20) and another at the epoch where the peak GAP score was achieved. }
\label{table:indiv_perf2}
\end{table}
\subsection{Performance of DNN Ensemble}
We have found that the overall GAP20 performance of the ensemble $E1$ formed in Table \ref{table:indiv_perf} was 83.884\% and ensemble $E2$ from Table \ref{table:indiv_perf2} was 83.634\% on the Kaggle leaderboard. When combined together, we have found potential improvements in the linearly weighted predictions from both ensembles, with a weighting of $\alpha \in (0,1)$ for one ensemble and $1-\alpha$ for the other ensemble. The results can be seen in Fig. \ref{fig:weight_res}. The optimal GAP20 score achieved is 83.96\% on the Kaggle leaderboard using the formula $0.65\times E1+0.35\times E2$.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{fig1.pdf}
\caption{ This figure shows different linear combination values for combining the two ensembles trained with different training-validation splits.}
\label{fig:weight_res}
\end{figure}
An interesting discovery was that the use of overfitted DNNs can improve the generalisation performance when incorporated into an ensemble. We have found that for large DNNs, (8K and above hidden units), when models trained up to later epochs (100+) were used, the validation error of the ensemble further decreases. This is despite the {\em increase} of validation error in the individual models. We have found that the use of overfitted models resulted in an average of 0.671\% improvement in the ensemble GAP20 score, compared with 0.579\% when using peak-validation GAP models. One hypothesis is that the overfitted models are overfitting to different video and label subsets. This in turn promotes diversity across different DNNs used, which results in better generalisation of the ensemble.
The ensemble that was trained using the sequential addition with the diversity aware loss function (Section \ref{sec:divbased}) did not yield any improvement over a simple average of randomly initialised and different architecture DNNs. We found that a 4-DNN ensemble (8K-8K-8K DNN) of learnt this way yielded a GAP score of 82.15\% and this did not improve by adding more DNNs.
\section{DNN Ensemble Diversity Analysis}
\label{sec:diverse_analysis}
It is generally agreed that greater output diversity of member classifiers in an ensemble result in improved performance. Unfortunately, the measurement of diversity is not straightforward, and at present, a generally accepted formulation does not exist \cite{Kuncheva}. Here, at least 10 different measures of diversity were found.
For our purposes, first suppose there are $M$ classifiers in our ensemble. There are 2 diversity measures that are relevant to our analysis. The first is based on the Pearsons correlation coefficient and the second based on Generalised Diversity Measure.
\begin{figure}[!h]
\centering
\includegraphics[width=0.8\columnwidth]{gap_scatter.pdf} \\
(a)\\
\includegraphics[width=0.8\columnwidth]{gap_heatmap_3.pdf} \\
(b)\\
\includegraphics[width=0.8\columnwidth]{divergence_heatmap.pdf} \\
(c)
\caption[]{a) A scatter plot showing the improvement in GAP score as a function of the models' diversity. b) shows the gap improvement (in \%) for different DNN pairs and c) shows the corresponding diversity score. In b),c) the type of DNN is identified as $\langle $feature$\rangle$ $\langle$ hid units$\rangle$ $\langle$ droupout$\rangle$, where for feature: M is mean, S is std. dev. and R means ROI. }
\label{fig:gap_corr}
\end{figure}
\subsection{Correlation-based Diversity Analysis}
The first analysis is based on Pearsons correlation coefficient defined as:
\[
R_{ij} = \frac{C_{ij}}{\sigma_i\sigma_j }
\]
where $i,j \in \{1,2,...,M\}$ and for classifiers indexed by $i$ and $j$, $R_{ij}$ is their correlation coefficient, $C_{ij}$ represents the covariance between these 2 classifiers and $\sigma_i,\sigma_j$ their respective output prediction standard deviations. Next, we find that one measure of diversity is: $1 - R_{ij}$, where if the correlation is minimal (i.e. 0), diversity is maximal and vice versa.
This method has the advantage of not requiring the classifier outputs to be binary, as is the case here.
When applied to the output predictions of the different classifiers in our ensembles, we find that a lower correlation is indicative of a greater improvement in the GAP20 score. This can be seen in Fig. \ref{fig:gap_corr}a). As shown there, we find that a a higher diversity score is highly correlated with an magnitude of improvement in the ensemble GAP score. Additional detail on the divergence scores and corresponding GAP20 improvement between pairs of DNNs can be seen in their respective heatmaps can be seen in Fig. \ref{fig:gap_corr}b,c.
Additionally, we can also use the diversity score to analyse the performance of overfitted models. This can be seen in Table \ref{table:overfit}. Here, we observe that when we allow a model to overfit past its highest validation score, this leads to an increase in diversity with other models. By ensembling overfitted models with lower individual scores, we actually observe that whilst this is detrimental to a single model's performance, it provides better improvement when incorporated into an ensemble
\begin{table}[h!]
\centering
\begin{tabular}{|c| c | c|}
\hline
& GAP Improvement & Diversity Score \\
\hline \hline
Peak Model & 0.579\% & 0.0322 \\
\hline
Overfitted Model& 0.671\% & 0.0340 \\ \hline
\end{tabular}
\bigskip
\caption{ Table showing the GAP improvement and diversity score for ensembles that use models with peak validation GAP20 or overfitted models with suboptimal GAP20 validation scores. }
\label{table:overfit}
\end{table}
\subsection{Generalised Diversity Measure-based Analysis}
Our second analysis is inspired by the Generalised Diversity Measure proposed by Partridge et al. \cite{Partridge}. In this measure, the authors propose that maximum diversity exists between two classifiers if, given an example, an error made by one classifier is accompanied by a correct classification of another classifier.
In order to obtain more insight into the improvements of classifier addition into an ensemble, we propose to analyse the performance of classifiers using ``wrong example sets''.
Consider that each class has two sets of video examples, $N^+$ number of positive videos (label 1) and $N^-$ number of negative videos (label 0). Let these sets of videos be defined as $X^+ = \{x^+_1, x^+_2,...,x^+_{N^+}\}$ and $X^-= \{x^-_1, x^-_2,...,x^-_{N^-}\}$ respectively. Now, suppose we are given a classifier $h$, which can be an ensemble or single DNN. Correspondingly, the predictions given by $h$ on the different video sets are: $Y_h^+ = \{y^+_{h,1}, y^+_{h,2},...,y^+_{h,N^+}\}$ and $Y_h^- = \{y^-_{h,1}, y^-_{h,2},...,y^-_{h,N^-}\}$.
We can now extract the set of videos that are considered ``wrong'' with respect to some threshold $\theta \in (0,1)$:
\begin{eqnarray*}
\varepsilon^+_{h,\theta} & = & \{i \in \{1,2,..., N^+\} : 1 - y^+_{h,i} \geq \theta\}\\
\varepsilon^-_{h,\theta} & = & \{i \in \{1,2,..., N^+\} : y^-_{h,i} \geq \theta\}
\end{eqnarray*}
The final set of ``wrong examples'' for classifier $h$ is:
\begin{equation}
\varepsilon_{h,\theta} = \varepsilon^+_{h,\theta} \cup \varepsilon^-_{h,\theta}
\label{eq:wrong_ex}
\end{equation}
We can now use the Eq. \ref{eq:wrong_ex} to analyse the effect of combining all of these classifiers together into an ensemble. In particular, we would like to discover if individual classifiers produce errors for {\em different} videos. If this were the case, when the classifiers are combined together, the erroneous predictions of individual classifiers can potentially be diluted by correct predictions from other classifiers.
To achieve this, first suppose we have an ensemble of $M$ classifiers: $H = h_1, h_2, ..., h_M$, and we assume that the errors for each classifiers are approximately equal. Next, the wrong-example sets are extracted using Eq. \ref{eq:wrong_ex} for each classifier, giving: $\varepsilon_{H,\theta} = \{\varepsilon_{h_1,\theta}, \varepsilon_{h_1,\theta}, ..., \varepsilon_{h_M,\theta}\}$.
Now, consider the intersection of the sets in $\varepsilon_{H,\theta}$:
\[
\Upsilon_{H,\theta} = \bigcap^M_{i=1} \varepsilon_{h_i,\theta}
\]
The set of examples that fall into $\Upsilon_{H,\theta}$ are those that {\em all} the classifiers in the ensemble gave wrong predictions (w.r.t $\theta$) for. As such, this ensemble will not improve the predictions for any example in $\Upsilon_{H,\theta}$. Nonetheless, we find that the size of the set $\Upsilon_{H,\theta}$ either decreases or remains unchanged as we add new classifiers into the ensemble $H$.
Additionally, we find that the union of sets in $\varepsilon_{H,\theta}$
\[
\Upsilon'_{H,\theta} = \bigcup^M_{i=1} \varepsilon_{h_i,\theta}
\]
represent the total unique videos that were wrongly classified by at least one classifier. However, examples in $\Upsilon'_{H,\theta}$ that are {\em not} in $\Upsilon_{H,\theta}$ will have an overall improved prediction in the ensemble.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{union_intersection.pdf}
\caption{ This graph shows the size of the sets representing the intersection of ``extremely wrong'' examples ($\theta = 0.9$) of individual classifiers in an ensemble, as more classifiers are added. Shown are also the size of the union of wrong examples that at least one classifier in the ensemble got significantly wrong. }
\label{fig:union_int}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=\columnwidth]{1theta0_9_hist.pdf}\\
(a)\\
\includegraphics[width=\columnwidth]{2theta0_9_hist.pdf}\\
(b)
\caption{ Shown here is how adding additional classifiers that are diverse into an ensemble diffuses the severity of wrong predictions. For clarity we have provided a zoomed-in view of prediction value histograms for wrong examples associated with very low predictions (+ve examples) (a) or very high predictions (-ve) examples (b). Shown are the histograms of examples with wrong predictions for two ensembles, one with 6 DNNs, and later when 5 more DNNs have been added. }
\label{fig:diffuse}
\end{figure}
Fig. \ref{fig:union_int} shows the size of $\Upsilon_{H,\theta}$ and $\Upsilon'_{H,\theta}$ as more classifiers are added to the ensembled used for this challenge, where $\theta = 0.9$. These represents extreme wrongly labelled videos. As such, these examples would have the greatest impact on decreasing the final GAP20 score. If these extreme mislabelling is due only to a small number of classifiers, then the ensemble should improve on their predictions (by means of accurate labelling from other classifiers).
Furthermore, if the above phenomena is occurring, we expect to see the intersection of ``wrong examples'' sets from individual classifiers decrease in size as we add more classifiers into the ensemble. This can indeed be seen in Fig. \ref{fig:union_int}. Here, we find that the number of examples that are wrongly labelled by all the individual classifiers in the ensemble steadily decreases as the ensemble size increases. This indicates that the individual classifiers of the ensemble each label different subsets of videos wrongly, suggesting that diversity is present. This in turn is results in a steady increase in the GAP score.
An additional confirmation of the diversity is that the size of the union of wrong example sets is increasing. The classifiers in the ensemble all have approximately the same accuracy. That means their wrong example sets will be approximately the same size. Thus, their union will only expand in size if these examples are different.
Finally, we present results where we ``track'' the movement of extremely wrong predictions as we expand the ensemble size. We start by identifying the example videos that have prediction errors greater than $\theta = 0.9$. A histogram of their prediction scores is then built. We then obtain their predictions after a number of DNNs have been added in, and construct an updated histogram. The result of this is shown in Fig. \ref{fig:diffuse}. Here, the baseline-ensemble of 6 DNNs misclassified many examples and classes (approx. 42K), as can be observed in the blue histograms. However, after having added 5 additional DNNs that were found to be diverse, we find that many examples have smaller error scores, as shown in the green histograms. This will in turn result in the entries corresponding to these wrong predictions migrate further down the final sorted GAP list, thus improving the final GAP20 score.
\section{Conclusions}
\label{sec:conclusions}
In this paper we have investigated factors controlling DNN diversity in the context of the "Google Cloud and YouTube-8M Video Understanding Challenge". We have shown that diversity can be cultivated by using DNN different architectures. Surprisingly, we have also discovered diversity can be achieved through some unexpected means, such as model over-fitting and dropout variations. We have presented details of our overall solution to the video understanding problem, which ranked \#7 in the Kaggle competition (Yeti team - gold medal).
\section*{Acknowledgements}
The work in this paper was partially funded by Innovate UK under the iTravel project (Ref: 102811). |
\section{Introduction}
The single-phase flow of incompressible, miscible fluids in a porous medium, arising in the modelling of enhanced oil recovery is described by a coupled system of non-linear elliptic-parabolic equations on the pressure and concentration of invading solvent, sometimes referred to as the Peaceman model as derived by Peaceman in \cite{peaceman1962numerical} (not to be confused with the Peaceman model of wells). Given its complexity, the behaviour of this system can only be fully understood through numerical approximations. These approximations must account for the specificities of flows in porous media: discontinuous data, non-Cartesian (and possibly non-conforming) grids, etc.
In this paper, we design an arbitrary-order numerical scheme for the Peaceman model. The stability of
the approximation is established, and numerical results are provided. Based on these, we provide advice
on how to choose the various parameters of the scheme (spatial order of approximation, time-stepping method, etc.)
to optimise the accuracy of the result and the overall computational cost.
Existence of a weak solution to this miscible flow model has been first established in \cite{feng1995existence}, and
then extended in \cite{chen1999mathematical} to include gravity effects and various boundary conditions.
In practice, the molecular diffusion is much smaller than the dispersion effect (driven by the Darcy viscosity),
and often neglected in numerical simulation. In that case, the parabolic equation takes on a degenerate form,
which makes the analysis of the model even more complex. \cite{az04} establishes the existence of a solution
in the case of a vanishing molecular diffusion and regular source terms. Given the scale of the reservoir
and the well bores, it is customary in simulations to consider wells concentrated on measures (Dirac measures
in 2D, measures along lines in 3D). The corresponding mathematical analysis has been carried out in \cite{droniou2014miscible} for non-zero molecular diffusion and in \cite{DT16} for vanishing molecular diffusion.
A variety of numerical schemes have been considered, with or without convergence
analysis, for the Peaceman model. Finite-element based methods coupled with a modified method of characteristic
for the advective part of the flow, are applied in \cite{ewing1983simulation} and
analysed in \cite{ERW84}. Another method of characteristics, the Eulerian--Lagrangian Adjoint Method,
is coupled with $\mathbb{RT}_k$ mixed finite elements (for the pressure) and $\mathbb{Q}_k$ finite elements (for the
concentration) in \cite{wang2000approximation,wang2008optimal}. A method combining $\mathbb{RT}_k$
finite elements and discontinuous Galerkin (dG) schemes is analysed in \cite{DGPeaceman}. Optimal error estimates for conforming
$\mathbb{P}_k$ finite elements on triangles are established in \cite{GalerkinPeaceman}, based
on maximal regularity properties of the continuous model.
Finite-element based methods are natural
and well known, but suffer from restrictions on the mesh geometries, which must be conforming and essentially
made of triangles or squares. Other schemes have been considered to deal with meshes with generic geometries,
as encountered in geophysical applications. In the last few
years, schemes have been developed to be applicable on such generic grids. They are mostly based
on finite volume techniques, which have the advantage of providing conservative approximations
of the Darcy fluxes \cite{droniou2014finite}, that can be used in the discretisation of the advective terms.
In \cite{droniou2007convergence},
the mixed finite volume (MFV) of \cite{droniou2006mixed} is adapted to the Peaceman model, its convergence analysed
and numerical tests are provided; as shown in \cite{DEGH09}, this method can be embedded
in a larger family, the hybrid mimetic method (HMM) family, that also contains the SUSHI scheme
of \cite{EYM10} and the mixed-hybrid mimetic finite difference methods of \cite{Brezzi.Lipnikov.ea:05}.
Discrete duality finite volume (DDFV) methods are considered in \cite{DDFV1,DDFV2}.
HMM and DDFV are finite volume schemes with first-order approximation properties, and rely for the
miscible displacement model on upwinding to stabilise
the advective terms; this raises the concern of an over-diffusion of the transition layer between
the invading solvent and the residing oil.
In this work, we develop an arbitrary-order numerical scheme for the Peaceman model, which is applicable
on generic grids. The scheme is an adaptation of the hybrid high-order (HHO) method, initially developed
for stationary diffusion PDEs in \cite{di2014arbitrary,di2015hybrid} and then extended to stationary
advection--diffusion--reaction models in \cite{droniou2015discontinuous}. The HHO method can be seen as
a higher-order extension of the HMM method, and is very close to virtual element methods \cite{Beirao-da-Veiga.Brezzi.ea:13}, to non-conforming mimetic finite difference methods \cite{Lipnikov-Manzini:2014} and to hybridizable dG methods \cite{cockburn2016bridging}.
The initial degrees of freedom of the HHO method are scalar valued polynomials of arbitrary order $k$ on the cells and faces of the mesh. The cell degrees of freedom can however be eliminated by a local
static condensation procedure, and only the face degrees of freedom remain coupled, in a way that
is however highly parallelisable. The HHO scheme is built on a collection of high-order local reconstruction operators that mimic the quantities present in the weak formulation of the continuous equation. Our executive summary is as follows:
\begin{itemize}
\item There is a real advantage in going for a higher order method.
The choice $k=0$ leads to strong grid effects, that are mostly eliminated by taking
$k=1$. This choice $k=1$ seems to be optimal in terms of accuracy vs. computational
cost; the choices $k=2,3,\ldots$ increase the computational cost with only minor further improvements of the
accuracy.
\item A Crank-Nicolson or second order backward differentiation formula (BDF) time-stepping is sufficient to obtain good results; high-order BDF tend to become
unstable unless the time step is reduced a lot, and do not lead to perceptible improvement (even considering
higher order spatial approximation, that is $k\ge 2$).
\item The specific mesh geometry is mostly irrelevant to the quality of the numerical approximation,
which mostly seems to only depend on the number of faces of the mesh (which is expected, the face
unknowns being the main unknowns in the HHO method).
\end{itemize}
Let us conclude by describing the organisation of the paper. In the next section, we describe
the continuous miscible displacement model, both in strong and weak form. The scheme is described
in Section \ref{sec:scheme}, starting from the time stepping, designed in a classical way to decouple the pressure and concentration equations. As explained above, the HHO method is built on local polynomial
spaces and reconstruction operators; these are respectively described in Sections \ref{sec:discretisation.spaces} and \ref{sec:reconstruction.operators}. The discretised pressure equation
is then presented. In order to discretise the convection terms appearing in
the concentration equation, cell Darcy velocities and corresponding fluxes have
to be designed from the numerical solution of the pressure equation. The reconstruction
of this velocity and fluxes is described in Section \ref{sec:discrete.fluxes}.
We note that this reconstruction has to be carefully performed to preserve the
scheme stability; in particular, this implies discretising the pressure equation at
an order twice the order chosen for the concentration equation. The numerical approximation
of the latter is described in Section \ref{sec:concentration.equation}.
Existence, uniqueness and stability results for our complete scheme are stated in Theorem \ref{thm:stability_of_crank_nicolson} at the end of Section
\ref{sec:scheme}.
Numerical tests are provided in Section \ref{sec:tests}. We analyse in particular
the effect (in terms of cost as well as efficiency)
of varying the spatial degree of the method and of having to use distorted polygonal meshes.
This analysis is done on test-cases involving homogeneous or discontinuous permeability,
with various mesh geometries and by considering both the general eye-ball quality of the solution (compared
with the expected behaviour), as well as quantitative assessments based
on the variation, with respect to the polynomial degrees, of the amount of oil recovered after 10 years. The executive summary above is backed up by the extensive numerical results in
this section.
A short conclusion is given in Section \ref{sec:conclusion}, and the proof of the existence,
uniqueness and stability result is given in the first appendix, Section \ref{sec:appen}. These proofs
show in particular the importance of choosing, in the discretisation
of the concentration equation, Darcy velocity and fluxes adapted to the discretisation
of the pressure equation. A second appendix, Section \ref{appen:implementation}, describes
the practical implementation of the scheme, and provides a link to the code
we developed for the numerical tests.
\section{The continuous model}\label{sec:continuous}
We introduce the following notations that will be used to describe the model, and then present the aforementioned system.
\begin{center}
\begin{tabular}{ l l }
$ d \in \{2,3\}$ & the number of dimensions considered in the model, \\
$\Omega \subset \R^d$ & a bounded Lipschitz domain representing the reservoir, \\
$(0, t_f) \subset \R$ & the time interval on which we consider the problem, \\
$p : (0,t_f) \times \Omega \to \R$ & the pressure in the mixture, \\
$\BU : (0,t_f) \times \Omega \to \R^d$ & the Darcy velocity of the fluid, \\
$c : (0,t_f) \times \Omega \to \R$ & the concentration of the invading solvent in the reservoir, \\
$\hat{c} : (0,t_f) \times \Omega \to \R$ & the concentration of solvent as it is injected, \\
$\Phi : \Omega \to \R$ & the porosity of the medium, \\
$\mu : [0,1] \to \R$ & the viscosity of the fluid mixture at a given concentration, \\
$\BK : \Omega \to \R^{d\times d}$ & the absolute permeability tensor of the medium, \\
$\BD : \Omega \times \R^d \to \R^{d\times d}$ & the diffusion-dispersion tensor of the medium, \\
$q^+ : (0,t_f) \times \Omega \to \R$ & the source term corresponding to the injection well, \\
$q^- : (0,t_f) \times \Omega \to \R$ & the source term corresponding to the production well.
\end{tabular}
\end{center}
\medskip
Taking the effects of gravity to be negligible, the model reads:
\begin{subequations}\label{scheme:all.eqn}
\begin{align}
&\begin{cases} \label{eqn:peaceman_pressure}
\divergence (\BU) = q^+ - q^- & \text{in } (0,t_f) \times \Omega , \\[1ex]
\BU = -\cfrac{\BK(x)}{\mu(c)}\nabla p & \text{in } (0,t_f) \times \Omega ,
\end{cases} \\[1ex]
&\Phi(x)\frac{\partial c}{\partial t} - \divergence(\BD(x,\BU)\nabla c - c\BU) + q^-c= q^+\hat{c} \qquad \text{in } (0,t_f) \times \Omega . \label{peaceman_concentration}
\end{align}
This system is comprised of two very natural parts. The pressure equation \eqref{eqn:peaceman_pressure} is an anisotropic diffusion equation with diffusivity $-\frac{\BK}{\mu(c)}$. For simplicity of notation, we may instead write
\begin{equation} \label{eqn:permeability-viscosity-ratio}
\bkappa = \cfrac{\BK(x)}{\mu(c)}.
\end{equation}
The concentration equation \eqref{peaceman_concentration} describes the convection of the fluid mixture via an advection-diffusion-reaction equation with diffusivity $\BD$ and advective velocity corresponding to the Darcy velocity of the fluid mixture $\BU$. We will now briefly summarise the models used for the data. Following \cite{peaceman1966improved}, Peaceman derived the diffusion-dispersion tensor $\BD$ to be
\begin{equation} \label{eqn:diffusion_dispersion_tensor}
\BD(x,\BU) = \Phi(x)\left(d_m \eye + |\BU|\left(d_l E(\BU) + d_t(\eye - E(\BU)) \right) \right),
\end{equation}
where $d_m$ is the molecular diffusion coefficient, $d_l$ and $d_t$ are the longitudinal and transverse dispersion coefficients, and $E(\BU)$ is an orthogonal projection in the direction of the Darcy velocity, given by the outer-product
\begin{equation}
E(\BU) = \frac{\BU \otimes \BU}{|\BU|^2} = \left[\frac{\BU_i \BU_j}{|\BU|^2}\right]_{1\leq i,j \leq d} .
\end{equation}
Physical experiments reveal that the longitudinal dispersion $d_l$ is far stronger than the transverse dispersion $d_t$, and that the molecular dispersion $d_m$ is negligible in comparison \cred{\cite{WLELQ00}}.
The viscosity of the mixture is determined by the mixing rule given in \cite{koval1963method}:
\begin{equation} \label{eqn:viscosity}
\mu(c) = \mu(0)\left(1 + (M^\frac{1}{4} - 1)c\right)^{-4} \qquad c \in [0,1],
\end{equation}
where $\mu(0)$ is the viscosity of the oil and $M$ is the mobility ratio between the oil and the injected solvent, given by $M = \frac{\mu(0)}{\mu(1)}$.
In order to maintain a balance of mass in the domain, the boundary of the reservoir $\partial \Omega$ is taken to be impermeable. Consequently, we include, denoting by $\nv$ the unit exterior normal to $\partial\Omega$, homogeneous no flow Neumann boundary conditions:
\begin{equation}
\begin{cases}
\BU\cdot \nv = 0 & \text{on } (0,t_f)\times \partial \Omega, \\
D \nabla c\cdot \nv = 0 & \text{on } (0,t_f)\times \partial \Omega.
\end{cases}
\end{equation}
Additionally, in order to satisfy the no flow boundary conditions and maintain mass, we must further impose that our injection and production source terms are compatible:
\begin{equation} \label{eqn:source_compatability}
\int_\Omega q^+(t,x)dx = \int_\Omega q^-(t,x)dx \qquad \text{in } (0, t_f).
\end{equation}
We must also prescribe an initial condition:
\begin{equation}
c(0,x) = c_0(x).
\end{equation}
In practice we take the initial concentration in the well to be $0$ everywhere. It is also usual to take $\hat{c}$, the concentration at the injection well to be $1$.
Lastly, since the pressure is only defined up to an arbitrary constant, we normalise $p$ by the following condition:
\begin{equation} \label{eqn:pressure_normalization}
\int_\Omega p(\cdot,x)dx = 0 \qquad \text{in } (0,t_f).
\end{equation}
\end{subequations}
\subsection{Weak formulation}\label{sec:weak.formulation}
Noting that the geometry of a typical oil reservoir will contain many geological layers of varying porosity and permeability, it is very important to take into account the fact that these quantities can not be assumed to be smooth and continuous everywhere. The following assumptions given in \cite{droniou2007convergence} are reasonable and allow us to devise both analytically sound and physically acceptable solutions.
\begin{subequations} \label{eqn:regularity_assumptions}
\begin{equation}
\begin{gathered} \label{eqn:permeability_assumptions}
\Phi \in L^\infty(\Omega) \text{ such that there exists $\Phi_* > 0$ satisfying} \\[-0.3em]
\text{ $\Phi_* \leq \Phi \leq \Phi_*^{-1}$ almost everywhere in $\Omega$},
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
q^+, q^- \in L^\infty (0, t_f; L^2(\Omega)) \text{ are non-negative and compatible, i.e.} \\
\int_{\Omega} q^+(\cdot, x)dx = \int_\Omega q^-(\cdot, x) dx \text{ for almost every $t\in(0, t_f)$},
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
\BK : \Omega \to M_d(\R) \text{ is a bounded matrix-valued function} \\[-0.3em]
\text{admitting symmetric, uniformly coercive values, that is, }\\[-0.3em] \exists \alpha_K > 0 \text{ s.t. } \BK(x)\xi \cdot \xi \geq \alpha_K |\xi|^2
\text{ for almost every $x\in\Omega$ and all}\\[-0.3em]
\xi \in \R^d,\ \exists \Lambda_K > 0 \text{ s.t. }
|\BK(x)| \leq \Lambda_K \text{ for almost every $x\in\Omega$,}
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
\BD : \Omega \times \R^d \to M_d(\R) \text{ is given by \eqref{eqn:diffusion_dispersion_tensor} with
$d_l, d_m, d_t > 0$}.
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
\mu \in \scriptC(\R) \text{ is positively bounded, that is } 0 \leq a \leq \mu(c) \leq b \text{ for} \\[-0.3em]
\text{positive constants $a, b$ for all $c \in \R$. This is clearly satisfied by \eqref{eqn:viscosity}},
\end{gathered}
\end{equation}
\begin{equation}\begin{aligned}
&\hat{c} \in L^\infty((0,t_f)\times \Omega) \text{ satisfies } 0 \leq \hat{c} \leq 1 \text{ almost everywhere},\\
&c_0 \in L^\infty(\Omega) \text { satisfies } 0 \leq c_0 \leq 1 \text{ almost everywhere}.
\end{aligned}
\end{equation}
\end{subequations}
Under the regularity assumptions \eqref{eqn:regularity_assumptions}, the existence of a weak solution to \eqref{eqn:peaceman_pressure}--\eqref{peaceman_concentration} is established in \cite{feng1995existence,chen1999mathematical}.
The HHO scheme presented in Section \ref{sec:scheme} consists in discretising the following
equations, satisfied by this weak solution $(p, \BU, c)$:
\begin{equation}
\label{eqn:weak_spaces}
\left\{\begin{split}
&p \in L^\infty(0,t_f;H^1(\Omega))\,,\;\BU \in L^\infty(0,t_f;L^2(\Omega))^d, \\
&c \in \scriptC([0,t_f];L^2(\Omega)) \cap L^2(0,t_f;H^1(\Omega))\mbox{ with }\\
& \Phi \partial_t c \in L^2(0,t_f;(W^{1,4}(\Omega))')\mbox{ and }c(0) = c_0,
\end{split}\right.
\end{equation}
\begin{equation} \label{weak_pressure_equation}
\left\{
\begin{split}
&\text{For almost every } t \in (0,t_f), \text{ for all } \varphi \in H^1(\Omega), \\[0.5ex]
&- \int_\Omega \BU(t,\cdot)\cdot \nabla \varphi = \int_\Omega \left(q^+(t,\cdot) - q^-(t,\cdot) \right)\varphi,
\end{split}
\right.
\end{equation}
and
\begin{equation} \label{weak_concentration_equation}
\left\{
\begin{split}
&\text{For almost every } t \in (0,t_f), \text{ for all } \varphi \in W^{1,4}(\Omega), \\[0.5ex]
&\langle \Phi \partial_t c(t),\varphi\rangle_{(W^{1,4})',W^{1,4}} + \int_\Omega \BD(\cdot,\BU(t, \cdot))\nabla c(t, \cdot) \cdot \nabla \varphi \\
& \qquad- \int_\Omega c(t, \cdot) \BU(t, \cdot) \cdot \nabla \varphi + \int_\Omega q^-(t, \cdot)c(t, \cdot)\varphi = \int_\Omega q^+(t, \cdot)\hat{c}(t)\varphi.
\end{split}
\right.
\end{equation}
\section{Numerical scheme}\label{sec:scheme}
The scheme for the pressure equation consists in the standard HHO method for variable diffusion problems \cite{di2015hybrid}, taking into account the coupling of the pressure with the concentration equation. Adapting \cite{cockburn2016bridging}, we derive conservative discrete version of the Darcy velocity and its fluxes that are required to discretise the advective
term $\divergence(\BU c)$ in the concentration equation. The spatial terms in this equation are discretised by
using the HHO method for linear advection--diffusion--reaction \cite{droniou2015discontinuous}, incorporating the aforementioned discrete Darcy velocity and fluxes. The time stepping presented here is based on the Crank-Nicolson
method, due to its strong stability properties \cite{leveque2007finite} (we also tested BDF time steppings).
\subsection{Time stepping}
We define our time-stepping as follows, let $N \in \N$ be the number of time-steps to be taken and let
\begin{equation}
\Delta t = \frac{t_f}{N}, \qquad t^n = n\Delta t, \qquad n = 0,1,...,N.
\end{equation}
Let us denote the pressure, concentration and Darcy velocity at time-step $n$ by
$p^n = p(t^n, \cdot)$, $c^n = c(t^n, \cdot)$ and $\BU^n = \BU(t^n, \cdot) = -\bkappa(c^n, \cdot)\nabla p^n$.
The Crank-Nicolson time stepping of \eqref{peaceman_concentration} consists in writing
\begin{equation} \label{eqn:crank_nicolson_continuous_concentration}
\begin{split}
\Phi(x)\frac{c^{n+1} - c^n}{\Delta t} &- \divergence(D(x,\BU^\hs)\nabla c^\hs - c^\hs\BU^\hs) + q^-c^\hs \\&= q^+\hat{c}(t^\hs),
\end{split}
\end{equation}
where the intermediate time-stepped values are defined by
\begin{equation} \label{eqn:half_time_concentration}
\xi^\hs = \frac{\xi^{n+1} + \xi^n}{2} \qquad\mbox{($\xi=t$, $c$ or $\BU$)}.
\end{equation}
The problem data, eg.\ the well terms $q^+$ and $q^-$ are evaluated at $t^\hs$.
The formulation \eqref{eqn:crank_nicolson_continuous_concentration} leads to the following
direct relation between $c^n$ and $c^\hs$:
\begin{equation}
\label{eqn:crank_nicolson_half_step_continuous_concentration}
\begin{split}
\Phi(x)\frac{2(c^\hs - c^n)}{\Delta t} &- \divergence(D(x,\BU^\hs)\nabla c^\hs - c^\hs\BU^\hs) \\& + q^-c^\hs = q^+\hat{c}(t^\hs).
\end{split}
\end{equation}
We note that this formulation is equivalent to performing a half time-step with an implicit Euler scheme to obtain $c^\hs$ and then linearly extrapolating $c^n$ with $c^\hs$ to obtain $c^{n+1}$ using \eqref{eqn:half_time_concentration}.
\subsection{Discretisation spaces}\label{sec:discretisation.spaces}
Let us briefly introduce the notion of a mesh and define the polynomial spaces central to the scheme.
\begin{definition} \label{definition_mesh}
A mesh $\scriptM$ of $\Omega \subset \R^d$ is a collection of cells and faces $(\scriptT, \scriptF)$ where:
\begin{enumerate}
\item $\scriptT$ (The cells or control volumes) is a subdivision of $\Omega$ into a disjoint family of open, non-empty polygons (or polyhedra in higher dimensions). Formally,
\begin{equation}
\bigcup_{T\in\scriptT} \overline{T} = \overline{\Omega}.
\end{equation}
\item $\scriptF$ (The faces or edges) is a disjoint family of non-empty affine subsets of $\Omega$, with positive $(d-1)$ dimensional measure such that for each cell $T \in \scriptT$, there exists $\scriptF_T \subset \scriptF$ where
\begin{equation}
\bigcup_{F \in \scriptF_T}\overline{F} = \partial T,
\end{equation}
such that each face $F \in \scriptF$ borders exactly one or two cells.
\end{enumerate}
\end{definition}
Additionally, we may employ the following notation
\begin{center}
\begin{tabular}{ l l }
$\scriptF_T \subset \scriptF$ & the faces that border the cell $T$, \\
$\scriptF_b \subset \scriptF$ & the faces on the boundary of the domain $\Omega$, \\
$|T|$ or $|F|$ & the Lebesgue measure of the cell $T$ or face $F$, \\
$\overline{x}_T$ & the centre of mass of the cell $T$, \\
$\nv_{TF}$ & a normal vector to the face $F$ facing outward from the cell $T$, \\
$h_T$ or $h_F$ & the diameter of a cell $T$ or a face $F$, \\
$h = \max_{T\in\scriptT} h_T$ & the maximum diameter of any cell.
\end{tabular}
\end{center}
Take $\scriptM_h = (\scriptT_h, \scriptF_h)$ a generic polygonal/polytopal mesh of $\Omega$ as above.
The degrees of freedom of the scheme are scalar valued polynomials on the cells $T \in \scriptT_h$ and the faces $F \in \scriptF_h$. No continuity conditions between cells and faces, or
between cells and cells,
are imposed on the degrees of freedom. Selecting an integer $m \geq 0$, the following notation will help us describe the polynomial spaces; here, $K$ is a set of dimension $l$ (that is,
the affine space spanned by $K$ has dimension $l$).
\begin{center}
\begin{tabular}{ l l }
$\Poly^m(K)$ & the space of $l$-variate polynomials of degree $\leq m$\\
$\nabla \Poly^m(K)$ & the space of functions $\{ \nabla u : u \in \Poly^m(K) \}$, \\
$\pi^m_K : L^2(K) \to \Poly^m(K)$ & the $L^2$ orthogonal projector onto $\Poly^m(K)$,
\end{tabular}
\end{center}
where the $L^2$ orthogonal projector is given by:
\begin{equation} \label{eqn:projection_definition}
\pi^m_K(u) = v \in \Poly^m(K) \text{ is such that, for all } w \in \Poly^m(K),\int_K (u - v)w = 0.
\end{equation}
The spaces of degrees of freedom are then given as follows.
\begin{definition} Let $\scriptM_h$ be a mesh.
For each cell $T \in \scriptT_h$, the space of local degrees of freedom on $T$ is defined by
\begin{equation}
\dof_T^m = \Poly^m(T) \times \left\{ \bigtimes_{F\in\scriptF_T} \Poly^m(F) \right\}.
\end{equation}
The space of global degrees of freedom on the mesh is defined as
\begin{equation}
\dof^m_h = \left\{ \bigtimes_{T\in\scriptT_h} \Poly^m(T) \right\} \times \left\{ \bigtimes_{F\in\scriptF_h} \Poly^m(F) \right\}.
\end{equation}
\end{definition}
For a given $k\ge 0$, these spaces will be used to approximate both the pressure (with $m=2k$)
and the concentration (with $m=k$).
An arbitrary set of global degrees of freedom is denoted by
\begin{equation}
\du_h = ((\pu_T)_{T\in\scriptT_h}, (\pu_F)_{F\in\scriptF_h}) \in \dof^m_h,
\end{equation}
and, similarly, a set of local degrees of freedom is
\begin{equation}
\du_{T} = (\pu_T, (\pu_F)_{F\in\scriptF_T}) \in \dof^m_T.
\end{equation}
For a given $\du_h\in \dof^m_h$, we define $\pu_h\in L^2(\Omega)$ as the piecewise polynomial
function given by $(\pu_h)_{|T}=\pu_T$ for all $T\in \scriptT_h$.
The space of degrees of freedom with zero average is then
\begin{equation}
\dof^m_{h,*} = \left\{ \du_h \in \dof^m_h\,:\, \int_{\Omega} \pu_h = 0 \right\}.
\end{equation}
\subsection{Reconstruction operators}\label{sec:reconstruction.operators}
We now introduce the local reconstruction operators which are central to the scheme. The cornerstone of the HHO method is a high-order local gradient reconstruction operator based on the cell and face polynomial degrees of freedom.
\begin{definition} \label{def:local_reconstruction}
Let $\bLambda$ be a bounded, real, symmetric, coercive tensor-valued function on $\Omega$. Take a cell $T \in \scriptT_h$ of $\scriptM$. The local reconstruction operator $\ro_{T,\bLambda}^{m+1} : \dof_T^m \to \Poly^{m+1}(T)$ is defined such that, for a bundle of local degrees of freedom $\du_{T} = (\pu_T, (\pu_F)_{F\in\scriptF_T}) \in \dof_T^m$ and any test function $w \in \Poly^{m+1}(T)$,
\begin{equation} \label{eqn:gradient_reconstruction}
\lipd{T}{\bLambda\nabla\ro_{T,\bLambda}^{m+1} \du_{T}}{\nabla w} = \lipd{T}{\bLambda \nabla \pu_T}{\nabla w} + \sum_{F\in\scriptF_T} \lip{F}{(\pu_F - \pu_T)}{\nabla w \cdot (\bLambda \nv_{TF})},
\end{equation}
and
\begin{equation} \label{eqn:local_reconstruction}
\int_T \ro_{T,\bLambda}^{m+1} \du_{T} = \int_T \pu_T.
\end{equation}
\end{definition}
In addition to the local reconstruction operator, we define the local high-order correction operator.
\begin{definition}
The high-order correction operator $\RO_{T,\bLambda}^{m+1} : \dof_T^m \to \Poly^{m+1}(T)$ is defined such that
\begin{equation} \label{high_order_potential}
\RO_{T,\bLambda}^{m+1} \du_{T} = \pu_T + (\ro^{m+1}_{T,\bLambda} \du_{T} - \pi^m_T\ro_{T,\bLambda}^{m+1} \du_{T}).
\end{equation}
We remark that the bracketed term of \eqref{high_order_potential} is orthogonal to $\Poly^m(T)$ by construction, and hence this operator can be seen as adding a high-order orthogonal correction to the cell unknown $\pu_T\in \Poly^m(T)$.
\end{definition}
\subsection{The pressure equation}\label{sec:pressure.equation}
Let us fix an integer $k\ge 0$. The HHO scheme for \eqref{scheme:all.eqn} consists in discretising
the pressure in $\dof^{2k}_h$ and the concentration in $\dof^k_h$. The choice of
an order $2k$, instead of $k$, for the pressure is driven by stability considerations,
which are made clear in the proof of Theorem \ref{thm:stability_of_crank_nicolson} (see Remark \ref{rem:2k.flux}).
To write the scheme on the pressure, we introduce the following local bilinear forms:
$\dbf_{T,\bLambda} : \dof^{2k}_T \times \dof^{2k}_T \to \R$ and $\stab_{\bLambda,T} : \dof^{2k}_T \times \dof^{2k}_T \to \R$, defined by
\begin{align} \label{eqn:pressure_diffusion_local_bilinear_form}
\dbf_{T,\bLambda}(\du_{T}, \tf_{T}) &= \lipd{T}{\bLambda \nabla\ro_{T, \bLambda}^{2k+1} \du_{T}}{\nabla\ro_{T, \bLambda}^{2k+1} \tf_{T}} + s_{\bLambda,T}(\du_{T}, \tf_{T}), \\
\stab_{\bLambda,T}(\du_{T}, \tf_{T}) &= \sum_{F\in\scriptF_T} \cfrac{{\bLambda_{TF}}}{h_F}\lip{F}{\pi^{2k}_F(\pu_F - \RO^{2k+1}_{T,\bLambda} \du_{T})}{\pi^{2k}_F(\pf_F - \RO^{2k+1}_{T,\bLambda} \tf_{T})}, \label{eqn:diffusion_stabilisation}
\end{align}
where
\begin{equation}\label{def:LambdaF}
{\bLambda_{TF}} = \|\nv_{TF} \cdot {\bLambda_{|T}} \nv_{TF} \|_{L^\infty(F)}
\end{equation}
is a controlling factor for the size of $\bLambda$ (in cell $T$) across the face $F$. The first term of $\dbf_{T,\bLambda}$ can readily be recognised as a discrete analogue of the weak diffusive terms of the weak formulation. The function $\stab_{\bLambda,T}$ is the diffusive stabilisation term whose purpose is to enforce a least-squares penalty between face unknowns, and the projection of the high-order correction of the cell unknown. This is required to ensure that the cell and face unknowns are related, and to ensure that the global bilinear form defined in \eqref{eqn:global_pressure_diffusion_bilinear_form} below is symmetric positive definite on $\dof^{2k}_{h,*}$ (see Lemma \ref{lemma:diffusive_bilinear_form_is_positive}).
Finally we can define the global pressure bilinear form that mimics the weak pressure equation \eqref{weak_pressure_equation}. We denote by $\dbf_{h,\bLambda} : \dof^{2k}_h \times \dof^{2k}_h \to \R$, the global bilinear form such that
\begin{equation} \label{eqn:global_pressure_diffusion_bilinear_form}
\dbf_{h,\bLambda}(\du_h,\tf_h) = \sum_{T\in\scriptT_h} \dbf_{T,\bLambda}(\du_{T}, \tf_{T}).
\end{equation}
We also write the linear functional $l^{p,\hs}_h : \dof^{2k}_h \to \R$ such that
\begin{equation} \label{eqn:global_diffusion_linear_functional}
l^{p,\hs}_h(\tf_h) = \lip{\Omega}{(q^+(t^\hs,\cdot)-q^-(t^\hs,\cdot))}{\pf_h},
\end{equation}
which mimics the right-hand side of the weak pressure equation.
Recalling that the pressure equation has diffusion tensor $\bLambda = \bkappa(c)$ given by \eqref{eqn:permeability-viscosity-ratio}, the discrete pressure $\dpdof_h^\hs \in \dof^{2k}_{h,*}$ on $[t^n,t^{n+1})$ is computed by solving
\begin{equation} \label{eqn:discrete_pressure_equation}
\dbf_{h,\bkappa^{\hs}}(\dpdof_h^\hs, \tf_h) = l^{p,\hs}_h(\tf_h)\qquad \forall \tf_h\in \dof^{2k}_{h,*}
\end{equation}
where $\bkappa^{\hs}=\bkappa(\tilde{\pc}^\hs)$ with
$\tilde{\pc}^{\hs}$ the following extrapolation of the concentration at time $t^\hs$:
\begin{equation}\label{eqn:concentration_extrapolation}
\tilde{\pc}^{\hs} = \frac{3}{2} \pc^n - \frac{1}{2} \pc^{n-1}
\end{equation}
(we take $\pc_T^{-1}=\pc_T^0=\pi^k_T c_0$ for all $T\in\scriptT_h$). Note that this choice of extrapolation decouples the
pressure equation from the concentration equation (see Algorithm \ref{algo1} below).
Of course, \eqref{eqn:discrete_pressure_equation} only defines the pressure up to an additive
constant, so we normalise by imposing
\begin{equation}\label{eqn:discrete_pressure_equation:norm}
\int_\Omega \pp_h^\hs=0.
\end{equation}
\subsection{The discrete fluxes}\label{sec:discrete.fluxes}
As can be readily seen from the concentration equation \eqref{peaceman_concentration}, the distribution of the concentration depends heavily on the Darcy velocity of the invading fluid as defined by the pressure equation \eqref{eqn:peaceman_pressure}. In the HHO framework, the discretisation of advective
terms is done by using both the velocity in each cell and its fluxes through the faces \cite{droniou2015discontinuous}.
This velocity and fluxes must be properly chosen to ensure the numerical stability of the
discretised concentration equation -- the fact that $\divergence (\BU) = q^+ - q^-$ at the continuous
level is what ensures stability estimates on the continuous concentration, and this must be
mimicked at the discrete level.
For simplicity of notation, we let here $\dpdof_h=\dpdof_h^\hs$ and $\bkappa=\bkappa^\hs$.
Define $\bkappa_{TF}$ from $\bkappa$ by \eqref{def:LambdaF}.
Following \cite[Section 3.1]{cockburn2016bridging},
the numerical flux $\fv_{TF}$, outward from cell $T$ through the face $F$, is given by:
\begin{equation} \label{eqn:discrete_fluxes}
\fv_{TF} = -\bkappa \nabla \ro^{2k+1}_{T,\bkappa} \dpdof_{T} \cdot \nv_{TF} + \frac{{\bkappa_{TF}}}{h_F}\br^{2k,\dagger}_{\partial T}\left(\pi^{2k}_F (\RO^{2k+1}_{T,\bkappa} \dpdof_{T} - \pp_F)\right),
\end{equation}
where, setting $\dof^{2k}_{\partial T}=\bigtimes_{F\in\scriptF_T} \Poly^{2k}(F)$
(identified with a subspace of $L^2(\partial T)$),
$\br^{2k,\dagger}_{\partial T}$ is the adjoint, for the $L^2(\partial T)$ inner product,
of the operator $\br^{2k}_{\partial T} : \dof^{2k}_{\partial T} \to \dof^{2k}_{\partial T}$ defined by:
for all $\tf_{\partial T}=(\pf_F)_{F\in\scriptF_T}\in \dof^{2k}_{\partial T}$,
\begin{equation}
\br^{2k}_{\partial T}(\tf_{\partial T}) = \left(\pi^{2k}_{F}[\pf_F - \ro^{2k+1}_{T,\bkappa} (0,\tf_{\partial T}) + \pi^{2k}_T \ro^{2k+1}_{T,\bkappa} (0,\tf_{\partial T})]\right)_{F\in\scriptF_T}.
\end{equation}
The first term of \eqref{eqn:discrete_fluxes} can be seen as the ``naive'' discrete flux that we would obtain if we simply substituted the discrete pressure into the definition of the continuous fluxes. The second term can thus be thought of as a discrete conservative correction to the ``naive'' flux.
The discrete Darcy velocity in cell $T$ is then given by
\begin{equation} \label{eqn:darcy_velocity_volumetric_reconstruction}
\RU_T = -\bkappa \nabla r^{2k+1}_{T,\bkappa} \dpdof_{T}.
\end{equation}
Finally, it is important to note that the discrete fluxes and Darcy velocity satisfy the following local conservation condition from \cite{cockburn2016bridging}.
\begin{theorem} \label{thm:conservation_of_the_discrete_fluxes}
Let $T \in \scriptT_h$ and let $\dpdof_h$ be the solution to \eqref{eqn:discrete_pressure_equation}. Then for any $\tf_{T} \in \dof^{2k}_T$, the discrete Darcy velocity and fluxes \eqref{eqn:discrete_fluxes} satisfy
\begin{equation} \label{eqn:conservation_of_the_discrete_fluxes}
\dbf_{T,\bkappa}(\dpdof_{T}, \tf_{T}) = -\lipd{T}{\RU_T}{\nabla \pf_T} + \sum_{F\in\scriptF_T}\lip{F}{\fv_{TF}}{(\pf_T - \pf_F)}.
\end{equation}
\end{theorem}
\subsection{Concentration equation}\label{sec:concentration.equation}
The discrete concentration equation is formulated in terms of a stationary advection--diffusion--reaction equation as in \cite{droniou2015discontinuous},
with source and reaction terms incorporating the time-stepping. We recall that, for the concentration
equation, the relevant discrete space is $\dof^k_h$.
We consider $\RU=((\RU_T)_{T\in\scriptT_h},(U_{TF})_{T\in\scriptT_h, F\in\scriptF_T})$ the Darcy velocity and fluxes
reconstructed above at time $t^\hs$.
The following discrete advective derivative $\ad_{T,\RU}^k \dv_{T}$ is designed to be a high-order approximation to the continuous quantity $\BU \cdot \nabla v$ on the cell $T$.
\begin{definition} \label{def:advective_derivative}
For all $T \in \scriptT_h$, the discrete advective derivative $\ad_{T,\RU}^k : \dof_T^k \to \Poly^k(T)$ is such
that, for any $\dv_{T} \in \dof^k_T$ and any test function $w \in \Poly^k(T)$,
\begin{equation}
\lip{T}{(\ad_{T,\RU}^k \dv_{T})}{w} = \lip{T}{(\RU_T \cdot \nabla \pv_T)}{w} + \sum_{F \in \scriptF_T} \lip{F}{\fv_{TF} (\pv_F - \pv_T)}{w}.
\end{equation}
\end{definition}
Recalling Definition \eqref{eqn:diffusion_dispersion_tensor}, we define the discrete reconstructed diffusion tensor $\dD$ for each cell $T \in \scriptT_h$, for any $x \in T$ by
\begin{equation} \label{eqn:diffusion_tensor_reconstruction}
\dD(x) = \Phi(x)(d_m \eye + |\RU_T(x)|(d_l E(\RU_T(x)) + d_t (\eye - E(\RU_T(x)))).
\end{equation}
The local advection--reaction bilinear form $\dbf_{T,R, \RU} : \dof^k_h \times \dof^k_h \to \R$ is then defined as
\begin{equation} \label{eqn:local_advection_bilinear_form}
\dbf_{T,R, \RU}(\du_{T}, \tf_{T}) = -\lip{T}{\pu_T}{(\ad_{T,\RU}^k \tf_{T})} + \lip{T}{R\pu_T}{\pf_T} + \stab_{T,\RU}(\du_{T}, \tf_{T}),
\end{equation}
where $R$ are the reaction terms, encompassing the time-stepping, given by
\begin{equation}\label{eqn:concentration_reaction_terms}
R = \frac{2 \Phi}{\Delta t} + q^-(t^\hs,\cdot).
\end{equation}
Furthermore, $\stab_{T,\RU} : \dof^k_T \times \dof^k_T \to \R$ is the advective stabilisation term given by
\begin{equation} \label{eqn:advective_stabilisation_form}
\stab_{T,\RU}(\du_{T}, \tf_{T}) = \sum_{F \in \scriptF_T} \lip{F}{[\fv_{TF}]^-(\pu_F - \pu_T)}{(\pf_F - \pf_T)}
\end{equation}
where $[\fv_{TF}]^-=\max(0,-\fv_{TF})$.
The global advection-reaction bilinear form $\dbf_{h, R, \RU} : \dof^k_h \times \dof^k_h \to \R$ is defined such that
\begin{equation} \label{eqn:global_advection_bilinear_form}
\dbf_{h,R, \RU}(\du_h, \tf_h) = \sum_{T \in \scriptT_h} \dbf_{T,R, \RU}(\du_{T}, \tf_{T}).
\end{equation}
Combining the diffusion and advection--reaction bilinear forms, the complete bilinear form for advection--diffusion--reaction is $\dbf_{h,\dD, R,\RU} : \dof^k_h \times \dof^k_h \to \R$ such that
\[
\dbf_{h,\dD, R, \RU}(\du_h, \tf_h) = \dbf_{h,\dD}(\du_h, \tf_h) + \dbf_{h,R, \RU}(\du_h, \tf_h),
\]
and the global linear functional $l^{c,\hs}_h : \dof^k_h \to \R$ is
\begin{equation} \label{eqn:concentration_rhs_linear_form}
l^{c,\hs}_h(\tf_h) = \int_\Omega \left(q^+(t^\hs,\cdot) \hat{c}(t^\hs,\cdot) + \frac{2\Phi}{\Delta t}\pc^n_h\right) \pf_h,
\end{equation}
The discrete concentration scheme then consists in seeking $\dc^\hs_h\in\dof^k_h$ such that
\begin{equation} \label{eqn:discrete_advection_diffusion_equation}
\dbf_{h,\dD, R,\RU}(\dc^\hs_h, \tf_h) = l^{c,\hs}_h(\tf_h),\qquad
\forall \tf_h \in \dof^k_h,
\end{equation}
then extrapolating to obtain $\dc^{n+1}_h$. We recall that $\dD$ and $\RU$ are computed
from $\dpdof^\hs_h$, and therefore depend on $n$.
\medskip
The full scheme is summarised in Algorithm \ref{algo1}.
\begin{algorithm}[!h]
\caption{Complete scheme for the pressure--concentration \label{algo1}}
\begin{algorithmic}[1]
\State \emph{Set $\pc_T^{-1}=\pc_T^0=\pi^k_T c_0$ for all $T\in\scriptT_h$}
\For{$n=0$ to $N-1$}
\State Compute $\dpdof^\hs_h$ by \eqref{eqn:discrete_pressure_equation}--\eqref{eqn:discrete_pressure_equation:norm},
from $\pc_h^{n-1}$ and $\pc_h^n$.
\State Compute $\RU$ by \eqref{eqn:discrete_fluxes} and \eqref{eqn:darcy_velocity_volumetric_reconstruction} with $\dpdof_h=\dpdof_h^\hs$.
\State Compute $\dc_h^\hs$ by \eqref{eqn:discrete_advection_diffusion_equation} and set $\dc_h^{n+1}=2\dc_h^\hs-\dc_h^n$.
\EndFor
\end{algorithmic}
\end{algorithm}
At each iteration, $(\dpdof^\hs_h,\dc_h^\hs)$ are computed by solving two
decoupled linear equations, one corresponding to an HHO scheme for a pure diffusion
equation, the other one to an HHO scheme for a diffusion--advection--reaction equation.
Essential questions when designing a numerical scheme are the existence, uniqueness and stability
of its solution; here, stability is understood as the grid size and time steps go to zero. The following
theorem brings an answer to these questions. It states the existence and uniqueness of the
solution to the scheme, and provides a bound on the cell unknowns corresponding to the concentration.
We note that this bound is, of course, uniform with respect to the grid size, but also with respect to
other important parameters, in particular the molecular, longitudinal and transverse dispersion coefficients
$d_m$, $d_l$ and $d_t$. A stability with respect to $d_m$ is all the more essential since this coefficient
tends to be taken equal to $0$ in numerical tests \cite{wang2000approximation,droniou2007convergence}.
\begin{theorem}[Existence, uniqueness and estimates for the discrete solution] \label{thm:stability_of_crank_nicolson}
Let $\scriptM = (\scriptT_h, \scriptF_h)$ be a mesh of $\Omega$ and take $N \geq 1$. If $(q^+, q^-, \BK, \BD, \Phi, \hat{c}, c_0)$ are data satisfying \eqref{eqn:regularity_assumptions}, then there exists a unique solution $(\dpdof_h, \dc_h)$ to the iterative scheme described in Algorithm \ref{algo1}.
Morever, for all $n=0,\ldots,N$,
\begin{equation}\label{est:cN}
\|\pc^{n}_h \|^2_{L^2(\Omega)} \leq {\frac{e^2}{\Phi_*^2}\left( \| \pc^0_h \|^2_{L^2(\Omega)}
+2t_f^2 \|q^+\|^2_{L^\infty(0,t_f;L^2(\Omega))}\right)}.
\end{equation}
\end{theorem}
The proof of Theorem \ref{thm:stability_of_crank_nicolson} is provided in the appendix.
\section{Tests}\label{sec:tests}
In this section, we illustrate the results obtained from the HHO scheme for the simulation of miscible fluid displacement in an oil reservoir. Some of the following test cases first appeared in \cite{wang2000optimal} for the ELLAM-MFEM method and have been subsequently applied in \cite{droniou2007convergence} for the MFV scheme. In every simulation here, we use the spatial domain $\Omega = (0,1000)^2$ measured in ft$^2$ and consider the time period $[0, 3600]$ (approximately $10$ years) measured in days. The injection and production source terms are Dirac masses, approximated as usual by a piecewise constant function on the relevant mesh cell. The injection well is located at $(1000,1000)$ with an injection rate of $30$ ft$^2$/day. The production well is correspondingly located at $(0,0)$ with a production rate of $30$ ft$^2$/day. We always take the injected concentration $\hat{c} = 1$ with an initial condition given by $c_0(x) = 0$. The viscosity of the oil is given by $\mu(0) = 1.0$ cp and the mobility ratio is $M=41$ (see \eqref{eqn:viscosity}).
We assume that molecular diffusion is negligible, setting $d_m = 0.0$ \fspd\ contrasting with the dispersion effects $d_l = 50.0$ \fspd\ and $d_t = 5.0$ \fspd.
The porosity of the medium is taken to be a constant $\Phi(x) = 0.1$. For each test we present surface and contour plots of the concentration $c$, the principle quantity of interest.
All tests were ran on a laptop with processor Intel i7-4710MQ 2.5Ghz, 6MB Cache
and 16GB of RAM at 1600Mhz.
\begin{remark}
Taking $d_m > 0$ is required to meet the regularity assumptions \eqref{eqn:regularity_assumptions} but physically unrealistic due to the fact that the magnitude of the molecular diffusion present in miscible fluid flow is negligibly small compared to the dispersive effects. We will present results taking $d_m = 0$ to demonstrate the suitability of the scheme to real-world parameters. We recall that the stability result of Theorem \ref{thm:stability_of_crank_nicolson} is
independent of $d_m$, and is therefore uniformly valid up to the limit $d_m\to 0$.
\end{remark}
Tables \ref{tab:mesh_parameters} and \ref{tab:mesh_parameters2} give the numbers of edges and sizes (maximum ratio of area to perimeter across all cells) for each mesh used in the following tests. The four mesh varieties are shown in Figure \ref{fig:meshes}. The triangular, Cartesian and Kershaw meshes were first introduced in the FVCA5 benchmark \cite{herbin2008benchmark} as mesh families 1, 2 and 4.1 respectively. The hexagonal-dominant mesh was used in \cite{BDM10,di2015extension}.
\begin{figure}[h]
\centering
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[scale=0.15]{meshes/mesh1_2}
\caption{Triangular}
\end{subfigure
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[scale=0.15]{meshes/mesh2_2}
\caption{Cartesian}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[scale=0.155]{meshes/mesh4_1}
\caption{Kershaw} \label{fig:kershaw_mesh}
\end{subfigure
\begin{subfigure}{.24\textwidth}
\centering
\hspace{5px}\input{img/meshes/pi6_tiltedhexagonal_2.tex}
\caption{Hexagonal} \label{fig:tilted_hexagonal_mesh}
\end{subfigure}
\caption{The four varieties of meshes. Figures obtained from \cite{di2014arbitrary}.}
\label{fig:meshes}
\end{figure}
\begin{center}
\begin{table}[!h]
\begin{tabular}{| c | c | c | c | c |}
\hline
\multicolumn{5}{|c|}{Number of edges}\\\hline
& Triangular & Cartesian & Kershaw & Hexagonal \\\hline
Mesh 1 & 92 & 40 & 612 & 62 \\\hline
Mesh 2 & 352 & 144 & 2380 & 220 \\\hline
Mesh 3 & 1376 & 544 & 5304 & 824 \\\hline
Mesh 4 & 5540 & 2112 & 9384 & 3184 \\\hline
Mesh 5 & 21632 & 8320 & 14620 & 12512 \\\hline
\end{tabular}\\
\begin{tabular}{| c | c | c | c | c |}
\hline
\multicolumn{5}{|c|}{Mesh size}\\\hline
& Triangular & Cartesian & Kershaw & Hexagonal \\\hline
Mesh 1 & 31.8 & 62.5 & 16.2 & 70.6 \\\hline
Mesh 2 & 15.9 & 31.2 & 8.96 & 36.7 \\\hline
Mesh 3 & 7.95 & 15.6 & 6.12 & 18.5 \\\hline
Mesh 4 & 3.98 & 7.81 & 4.64 & 9.26 \\\hline
Mesh 5 & 1.99 & 3.91 & 3.73 & 4.63 \\\hline
\end{tabular}
\caption{Mesh parameters for the meshes used in homogeneous permeability tests.}
\label{tab:mesh_parameters}
\end{table}
\begin{table}
\begin{tabular}{| c | c | c |}
\hline
\multicolumn{3}{|c|}{Number of edges}\\\hline
& Triangular & Cartesian \\\hline
Mesh 1 & 545 & 144 \\\hline
Mesh 2 & 2140 & 840 \\\hline
Mesh 3 & 4785 & 1860 \\\hline
Mesh 4 & 8480 & 3280 \\\hline
Mesh 5 & 13225 & 5100 \\\hline
\end{tabular}\\
\begin{tabular}{| c | c | c |}
\hline
\multicolumn{3}{|c|}{Mesh size}\\\hline
& Triangular & Cartesian \\\hline
Mesh 1 & 12.72 & 31.25 \\\hline
Mesh 2 & 6.36 & 12.5 \\\hline
Mesh 3 & 4.24 & 8.33 \\\hline
Mesh 4 & 3.18 & 6.25 \\\hline
Mesh 5 & 2.54 & 5.00 \\\hline
\end{tabular}
\caption{Mesh parameters for the meshes used in discontinuous permeability tests.}
\label{tab:mesh_parameters2}
\end{table}
\end{center}
\subsection{Numerical results} \label{sec:numres}
In these first four tests, we use a polynomial degree $k = 1$ for the spatial discretisation.
The time-step is $\Delta t = 18$ days ($N \approx 200$ time-steps over $10$ years).
\begin{test} \label{test:peaceman2}\rm
We take a homogeneous permeability tensor $\BK = 80\BI$ uniformly over the domain. We experiment on a $32\times 32$ Cartesian mesh (Cartesian Mesh 4 in Table \ref{tab:mesh_parameters}.) The mobility within the solvent saturated regions caused by the large adverse mobility ratio $M$ and the lack of molecular diffusion should result in the front of the injected fluid progressing most rapidly along the diagonal between the injection and production wells. These effects are seen on the surface and contour plots in Figure \ref{fig:peaceman_test2} on the Cartesian mesh and imply that the flow is indeed strongest along the diagonal direction as expected. This effect is well studied in the literature and is referred to as the macroscopic fingering phenomenon \cite{ewing1983mathematics}. Notably, our results at $t = 3$ with $k \geq 1$ are far more realistic than those given in \cite{droniou2007convergence}, which suffer from fluid progressing much too rapidly along the boundary of the domain. The scheme in \cite{droniou2007convergence} corresponds to a variant of the HHO method with $k=0$; our own tests with $k=0$ reproduced similar results as in this reference (see, e.g., Figures \ref{fig:compare_ks_t3}\textsc{(a)} and \ref{fig:compare_ks}\textsc{(a)}).
\end{test}
\begin{figure}[h]
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test2/time3_surface_k1_CN}
\caption{Surface plot at $t = 3$ years}
\end{subfigure
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=5.18cm,height=5cm]{peaceman/test2/time3_contour_k1_CN}
\caption{Contour plot at $t = 3$ years}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test2/time10_surface_k1_CN}
\caption{Surface plot at $t = 10$ years}
\end{subfigure
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=5.18cm,height=5cm]{peaceman/test2/time10_contour_k1_CN}
\caption{Contour plot at $t = 10$ years}
\end{subfigure}
\caption{Concentration of the invading solvent in Test \ref{test:peaceman2} with $k=1$ and $\Delta t = 18$ on Cartesian Mesh 2 with constant permeability.}
\label{fig:peaceman_test2}
\end{figure}
\begin{test}\rm
We retain the parameters of Test \ref{test:peaceman2} and use a hexagonal mesh in place of the Cartesian mesh (Hexagonal Mesh 4 in Table \ref{tab:mesh_parameters}). The fingering phenomenon is also observed in the results, which are shown in Figure \ref{fig:peaceman_test2hexa}, where we notice that there is a slight bias along the opposite diagonal at $t = 3$. This bias
is expected, given that the mesh is skewed in this direction (see Figure \ref{fig:meshes}), but it remains
rather small and does not seem to impact much the final result at $t=10$ years.
\end{test}
\begin{figure}[h]
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test2_hexa/time3_surface_k1_CN}
\caption{Surface plot at $t = 3$ years}
\end{subfigure
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=5.18cm,height=5cm]{peaceman/test2_hexa/time3_contour_k1_CN}
\caption{Contour plot at $t = 3$ years}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test2_hexa/time10_surface_k1_CN}
\caption{Surface plot at $t = 10$ years}
\end{subfigure
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=5.18cm,height=5cm]{peaceman/test2_hexa/time10_contour_k1_CN}
\caption{Contour plot at $t = 10$ years}
\end{subfigure}
\caption{Concentration of the invading solvent in Test \ref{test:peaceman2} with $k=1$ and $\Delta t = 18$ on Hexagonal Mesh 2 with constant permeability.}
\label{fig:peaceman_test2hexa}
\end{figure}
\begin{test} \label{test:peaceman4}\rm
For this test, we take a discontinuous permeability tensor $\BK = 80\BI$ except on the four subdomains $(200,400)\times(200,400)$, $(600,800)\times(200,400)$, $(200,400)\times(600,800)$ and $(600,800)\times(600,800)$ where instead $\BK = 20\BI$ (see Figure \ref{fig:heterogeneous}). We use a $40 \times 40$ Cartesian for this test in order to ensure that the regions of discontinuity are aligned with the edges (Cartesian Mesh 4 in Table \ref{tab:mesh_parameters2}). The results shown in Figure \ref{fig:peaceman_test4} are of great interest to us as they depict very different behaviour to those presented in \cite{droniou2007convergence}. Notably in \cite{droniou2007convergence}, by $t = 10$ years, the invading fluid has yet to subsume the two low permeability regions along the main diagonal. However, the HHO scheme depicts both blocks almost entirely saturated by $t = 10$ years. This implies the presence of a significantly higher amount of dispersion in the solution compared to that produced by the MFV scheme, which suggests that low order schemes may underestimate the amount of diffusion described by the model. Finally, we note that the region saturated with solvent is larger when the permeability is inhomogeneous compared to Test \ref{test:peaceman2} where it was not. This is another common phenomenon that has been well observed \cite{droniou2007convergence,wang2000approximation}.
\end{test}
\begin{figure}[h]
\centering
\input{fig-heterogeneous.pdf_t}
\caption{Permeability tensor for Tests \ref{test:peaceman4} and \ref{test:peaceman5}. \label{fig:heterogeneous}}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4/time3_surface_k1_CN}
\caption{Surface plot at $t = 3$ years}
\end{subfigure
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4/time3_contour_k1_CN}
\caption{Contour plot at $t = 3$ years}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4/time10_surface_k1_CN}
\caption{Surface plot at $t = 10$ years}
\end{subfigure
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4/time10_contour_k1_CN}
\caption{Contour plot at $t = 10$ years}
\end{subfigure}
\caption{Concentration of the invading solvent in Test \ref{test:peaceman4} with $k=1$ and $\Delta t = 18$ on a $40 \times 40$ Cartesian mesh with a discontinuous permeability.}
\label{fig:peaceman_test4}
\end{figure}
\begin{test}\label{test:peaceman5}\rm
We retain the parameters of Test \ref{test:peaceman4} and use a triangular mesh in place of the Cartesian mesh. The mesh is a $10 \times 10$ grid of the triangular pattern depicted in Figure \ref{fig:meshes} (Triangular Mesh 2 in Table \ref{tab:mesh_parameters2}), which ensures its alignment with the permeability discontinuities. The results can be seen in Figure \ref{fig:peaceman_test4_tri}, where we observe the same general distribution as in the Cartesian case, although the solvent seems slightly more dispersed. For such hybrid methods as the HHO method,
it is well known that the main unknowns are the edge-based unknowns (see in particular Section \ref{sec:comp.cost});
it is therefore expected that, for a comparable mesh size, a mesh with fewer edges will perform slightly worse
than a mesh with more edges. The Cartesian mesh used in Test \ref{test:peaceman4} has 1600 cells, 3280 edges
and a size of $6.25$; the triangular mesh used here has 1400 cells, 2060 edges and
a size of $6.36$.
\end{test}
\begin{figure}[h]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4_tri/time3_surface_k1_CN}
\caption{Surface plot at $t = 3$ years}
\end{subfigure
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4_tri/time3_contour_k1_CN}
\caption{Contour plot at $t = 3$ years}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4_tri/time10_surface_k1_CN}
\caption{Surface plot at $t = 10$ years}
\end{subfigure
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/test4_tri/time10_contour_k1_CN}
\caption{Contour plot at $t = 10$ years}
\end{subfigure}
\caption{Concentration of the invading solvent in Test \ref{test:peaceman4} with $k=1$ and $\Delta t = 18$ on a Triangular mesh with a discontinuous permeability.}
\label{fig:peaceman_test4_tri}
\end{figure}
\subsection{Comparison of higher order parameters} \label{sec:numres.comp}
Although the scheme is arbitrary order in space and can easily be extended to arbitrary order in time (by substituting the Crank-Nicolson time-stepping for a higher-order scheme such as backward differentiation), we argue that the $k=1$ scheme provides the best balance of accuracy and computational efficiency. In order to demonstrate that this is the case, we present the following tests.
\begin{test}\rm
To contrast the different solutions, we compare the principal quantity of interest to us, which is the total volume of oil recovered after ten years, as measured by the integral
\begin{equation*}
\int_{\Omega} \Phi(x) \pc_h^N(x).
\end{equation*}
For the $k = 1$ scheme as presented in Test \ref{test:peaceman2}, the total of volume of oil recovered can be measured to be $65.798\%$ of the total volume of the reservoir.
For the high-order tests, we range $k$ from $0$ to $3$, and replace the Crank-Nicolson time-stepping scheme with a high order backward differentiation formula of order $4$ in order to minimise the contribution of the temporal error. Additionally, in order to mitigate the majority of the extrapolation error produced by the pressure estimate when extrapolating $\pc_h$ (Equation \eqref{eqn:concentration_extrapolation}), we take a reduced time-step of $\Delta t = 7.2$ days (approximately $N = 500$ steps.)
Figure \ref{fig:oil_recovery_convergence} depicts the total recovery volume on each family of meshes in Table \ref{tab:mesh_parameters} with various polynomial degrees $k$. The $k = 0$ scheme is shown to perform quite poorly, producing results that are well out-of-line with the rest of the schemes. All of the other schemes however quickly converge to a similar estimate as the mesh size is refined, which is consistently within $1\%$ of the estimate produced by Test \ref{test:peaceman2} with only $k=1$, a Crank-Nicolson time-stepping, and $\Delta t=18$ days. We note that the scheme behaves well even on distorted meshes (the Kershaw mesh) until $k = 3$, where the linear system becomes too difficult to solve (Figure \ref{fig:oil_recovery_convergence} (c)). This solvability issue of HHO with high degrees
(remember that for $k=3$, the pressure equation is approximated with an order $2k=6$) on severely distorted
meshes has already been noticed even for the Poisson problem, and might be a consequence of rounding errors \cite{Dprivate}. A way to mitigate this poor conditioning of the system matrix on skewed meshes is
to change the local basis functions by applying a Gram--Schmidt orthonormalization process,
see \cite{BBCDPT12}. We however did not explore this option here as, in our experiments, the quality of the results do not significantly improve when using orders higher than $k = 1$ or $2$, and the computational cost increases drastically.
Finally, we also compare the high order schemes on the discontinuous permeability tensor of Test \ref{test:peaceman4}, using otherwise the same parameters. We use the mesh families, described in Table \ref{tab:mesh_parameters2}, whose edges are aligned with the discontinuities depicted in Figure \ref{fig:heterogeneous}. The results in Figure \ref{fig:oil_recovery_convergence_discontinuous} show that we obtain similar convergence pat\-terns to those of the homogeneous permeability tests. The $k=0$ scheme still produces results that are well out of line with the rest, while all $k \geq 1$ schemes convergence to a similar value as the mesh is refined.
\end{test}
\begin{figure}[h]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/convergence/triangular}
\caption{Recovery on triangular meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/convergence/square}
\caption{Recovery on Cartesian meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/convergence/kershaw}
\caption{Recovery on Kershaw meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/convergence/hexagonal}
\caption{Recovery on hexagonal meshes}
\end{subfigure}
\caption{Total percentage of oil recovered from the reservoir with a uniform permeability,
after $10$ years and for various polynomial degrees and mesh sizes. To minimise the temporal error contribution, a high order backward difference time-stepping scheme with $\Delta t = 7.2$ is
used.}
\label{fig:oil_recovery_convergence}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/convergence/discontinuous_triangular}
\caption{Recovery on triangular meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/convergence/discontinuous_square}
\caption{Recovery on Cartesian meshes}
\end{subfigure}
\caption{Total percentage of oil recovered from the reservoir with a discontinuous permeability tensor, after $10$ years for various polynomial degrees and mesh sizes.
To minimise the temporal error contribution, a high order backward difference time-stepping scheme with $\Delta t = 7.2$ is
used}
\label{fig:oil_recovery_convergence_discontinuous}
\end{figure}
\begin{test}\rm
The quality of the numerical approximations for various values of $k$ can also be observed visually. Depicted in Figures \ref{fig:compare_ks_t3} and \ref{fig:compare_ks} are the contour plots for the solution to Test \ref{test:peaceman2} using $k=0,1,2,3$ at time $t=3$ and $t=10$ respectively. The low-order solution using $k=0$ suffers from obvious grid effects in which the fluid mixture is progressing too rapidly along and clinging to the boundary of the domain. A similar effect is present in the results of the MFV scheme of \cite{droniou2007convergence}, suggesting that this artefact is a result of the low order of the scheme. Moving to a higher order scheme, even just $k = 1$, remedies this effect and shows the solvent mixture progressing in a physically realistic pattern.
The results of Figures \ref{fig:oil_recovery_convergence}, \ref{fig:oil_recovery_convergence_discontinuous}, \ref{fig:compare_ks_t3} and \ref{fig:compare_ks}
also show that there is little advantage in selecting a spatial order $k\ge 2$, since the results for
these higher order are qualitatively and quantitatively similar to those obtained with $k=1$.
\end{test}
\begin{figure}[h]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/t3_order0}
\caption{$k = 0$}
\end{subfigure
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/t3_order1}
\caption{$k = 1$}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/t3_order2}
\caption{$k=2$}
\end{subfigure
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/t3_order3}
\caption{$k=3$}
\end{subfigure}
\caption{Comparison of the quality of the numerical approximation for $k = 0,1,2,3$ on a Cartesian mesh at time $t = 3$ years. For $k = 0$, the fluid mixture clings to and overly rapidly progresses along the reservoir boundary. Moreover, the expected fingering effect is not visible. These issues are remedied by already selecting $k=1$, and further increases of the order does
not noticeably impact the solution (at least visually).}
\label{fig:compare_ks_t3}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/order0}
\caption{$k = 0$}
\end{subfigure
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/order1}
\caption{$k = 1$}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/order2}
\caption{$k=2$}
\end{subfigure
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{peaceman/compare_order/square/order3}
\caption{$k=3$}
\end{subfigure}
\caption{Comparison of the quality of the numerical approximation for $k = 0,1,2,3$ on a Cartesian mesh at time $t = 10$ years.}
\label{fig:compare_ks}
\end{figure}
\subsection{Computational cost}\label{sec:comp.cost}
High-order versions of the HHO scheme have been shown to produce very reliable results. This increased
accuracy however obviously comes with a higher computational cost than low-order schemes. The HHO
method has been designed to allow for a static condensation of the cell unknowns: by local Gaussian elimination, the systems \eqref{eqn:discrete_pressure_equation} and \eqref{eqn:discrete_advection_diffusion_equation}
can be expressed in terms of the face unknowns, resulting in
systems on the face unknowns only and with the same sparsity structure as the original equations.
On a given mesh $\scriptM_h = (\scriptT_h, \scriptF_h)$ with polynomials of degree $m$, the
global system is therefore of size
\[
\binom{m + d - 1}{d - 1}|\scriptF_h|,
\]
and has an $\mathcal O(m^{d-1})$ growth with respect to the polynomial degree.
Recall that the pressure is solved at a degree $m=2k$.
The cost of the time discretisations can be considered independent since it is clearly linear in the number of time-steps $N$.
Figure \ref{fig:peaceman_computational_time} illustrates the relative costs of the methods for various degrees $k$. Depicted are the average times taken per step using $N = 100$ time-steps for the data given in Test \ref{test:peaceman2} on the mesh families in Table \ref{tab:mesh_parameters}. We emphasise that our implementation is not optimised for high performance, and that the tests are performed on a personal computer. These measurements are not intended to give an absolute estimate of the cost, only a comparison of running times of the various order schemes (this comparison is valid since all tests were done on the same computer). It is expected that, even if the times vary from one computer to the other, the relative positions of the curves corresponding to various $k$ will be similar to those in Figure \ref{fig:peaceman_computational_time}.
As predicted, the running times begin to grow very rapidly for fine meshes with high-order $k$. Since the number of degrees of freedom of the scheme is directly tied to the number of faces in the mesh, the execution time can be seen to be larger for meshes with a greater number of faces (see Table \ref{tab:mesh_parameters}).
Combined with the qualitative and quantitative results in Sections \ref{sec:numres} and \ref{sec:numres.comp},
these relative running times further supports our argument that the $k = 1$ scheme may be the best balance of accuracy and speed.
\begin{figure}[h]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{runtime/timing_triangle}
\caption{Execution time on triangular meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{runtime/timing_square}
\caption{Execution time on Cartesian meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{runtime/timing_kershaw}
\caption{Execution time on Kershaw meshes}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=0.3]{runtime/timing_hexagon}
\caption{Execution time on hexagonal meshes}
\end{subfigure}
\caption{The average execution time for one time-step of Test \ref{test:peaceman2} on various meshes.}
\label{fig:peaceman_computational_time}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
We designed and implemented an arbitrary-order scheme for a miscible incompressible
flow model used in tertiary oil recovery. The scheme was based on the Hybrid High-Order method,
and is applicable on any kind of polygonal or polyhedral meshes.
To ensure the stability of the numerical approximation, the pressure equation has to be approximated
with an order twice as large as the order used for the concentration equation,
and special care must be taken in reconstructing advective Darcy fluxes from the approximate
pressure.
We produced several numerical tests on classical 2D test cases encountered in the literature. These
tests show that the best balanced of accuracy vs. computational cost is obtained by selecting a spatial order $k=1$ and
a Crank-Nicolson time stepping. The results are stable with respect to the geometry of the meshes, in the sense
that the total recovered oil is similar for all meshes and $k\ge 1$. Selecting an order $k=1$
fixes grid effects that are present with $k=0$ (as in low-order finite volume methods).
\section{Appendix: existence and stability of the solution to the scheme}\label{sec:appen}
Let $\scriptM_h = (\scriptT_h, \scriptF_h)$ be a mesh. As in \cite{di2011mathematical}, we take
a matching simplicial subdivision $\mathcal I_h$ of $\scriptM_h$ and $\varrho>0$
such that for any simplex $S\in\mathcal I_h$ of diameter $h_S$ and inradius $r_S$, $\varrho h_S\le r_S$, and
for all $T\in\scriptT_h$ and all $S\in\mathcal I_h$ such that $S\subset T$, $\varrho h_T \le h_S$.
In the following, $A\lesssim B$ means that $A\le CB$ for some $C$ depending only on $\Omega$ and $\varrho$
(not on $h$).
The following lemma states a stability result for the discrete elliptic bilinear form.
A proof can be found for $\Lambda={\rm Id}$ in \cite{di2014arbitrary}, and a sketch for
extending this to generic $\Lambda$ is given in \cite{di2015hybrid}.
\begin{lemma} \label{lemma:diffusive_bilinear_form_is_positive}
Let $\bLambda$ be a bounded, symmetric, uniformly coercive tensor-valued function on $\Omega$. Then
for any $\tf_h \in \dof^k_h$, the discrete diffusive bilinear form $\dbf_{h,\bLambda}$ \eqref{eqn:global_pressure_diffusion_bilinear_form} satisfies
\begin{equation} \label{eqn:diffusive_bilinear_form_seminorm_inequality}
\dbf_{h,\bLambda}(\tf_h, \tf_h) \gtrsim \sum_{T\in\scriptT_h} \| \bLambda^\frac{1}{2} \nabla \pf_T \|^2_{L^2(T)} + \sum_{T\in\scriptT_h} \sum_{F\in\scriptF_T} \frac{{\bLambda_{TF}}}{h_F} \|\pf_F - \pf_T \|^2_{L^2(F)}.
\end{equation}
\end{lemma}
The following lemma is the key ingredient in proving the stability of the solution
to the scheme. It holds true because of the specific choice of reconstructed Darcy velocity,
chosen to be compatible with the discretisation of the concentration equation.
\begin{lemma} \label{lemma:advective_bilinear_form_is_positive}
Let $R$ be given by \eqref{eqn:concentration_reaction_terms} and $((\RU_T)_{T\in\scriptT_h}, (\fv_{TF})_{T\in\scriptT_h, F\in\scriptF_T})$ be the reconstructed Darcy velocity \eqref{eqn:discrete_fluxes}--\eqref{eqn:darcy_velocity_volumetric_reconstruction} corresponding to a solution $\dpdof_h^\hs$ of the discrete pressure equation \eqref{eqn:discrete_pressure_equation}. Then for any $\tf_h \in \dof^k_h$, the discrete advection--reaction bilinear form \eqref{eqn:global_advection_bilinear_form} satisfies
\begin{equation} \label{eqn:advective_bilinear_form_seminorm_inequality}
\dbf_{h,R,\RU}(\tf_h, \tf_h) \geq \int_\Omega \frac{2\Phi}{\Delta t} \pf^2_h + \sum_{T\in\scriptT_h} \sum_{F\in\scriptF_T} \frac{1}{2}\int_F |\fv_{TF}| (\pf_T - \pf_F)^2.
\end{equation}
\end{lemma}
\begin{proof}
{\scshape Proof}:
Equation \eqref{eqn:local_advection_bilinear_form} and \eqref{eqn:global_advection_bilinear_form} give
\begin{equation}
\dbf_{h,R, \RU}(\tf_h, \tf_h) = \sum_{T\in\scriptT_h} \left\{ -\lip{T}{\pf_T}{(\ad_{T,\RU}^k \tf_{T})} + \lip{T}{R\pf_T^2}{} + \stab^-_{\RU,T}(\tf_{T}, \tf_{T}) \right\} .
\end{equation}
Expanding via the definitions of the advective derivative $\ad_{T,\RU}^k$ (Definition \ref{def:advective_derivative}), the advection stabilisation $\stab^-_{\RU,T}$ \eqref{eqn:advective_stabilisation_form} and the reaction terms $R$
\eqref{eqn:concentration_reaction_terms} and collecting cell and face terms, we find
\begin{equation} \label{eqn:advection_stability_terms}
\begin{split}
\dbf_{h,R, \RU}&(\tf_h, \tf_h) =\sum_{T\in\scriptT_h} \left\{ \lip{T}{\frac{2\Phi}{\Delta t}}{\pf_T^2} - \lip{T}{(\RU_T \cdot \nabla \pf_T)}{\pf_T} + \lip{T}{q^-(t^\hs,\cdot) \pf_T^2}{} \right\} \\& + \sum_{T\in\scriptT_h} \sum_{F \in \scriptF_T} \left\{ -\lip{F}{\fv_{TF} (\pf_F - \pf_T)}{\pf_T} + \lip{F}{[\fv_{TF}]^-\left(\pf_F - \pf_T\right)^2}{} \right\}.
\end{split}
\end{equation}
Considering the second cell term of \eqref{eqn:advection_stability_terms}, we notice that
\begin{equation} \label{eqn:stability_gradient_factor}
-(\RU_T \cdot \nabla \pf_T)\pf_T = -\RU_T \cdot \nabla \left(\frac{1}{2}\pf_T^2 \right).
\end{equation}
Next we make use of the identity $(a-b)a = a^2 - ab = \frac{1}{2}(a^2 - b^2) + \frac{1}{2}(a-b)^2$
to rewrite the first face term of \eqref{eqn:advection_stability_terms} as
\begin{equation} \label{eqn:stability_flux_inequality}
\begin{split}
-\lip{F}{(\fv_{TF} (\pf_F - \pf_T))}{\pf_T} &= \frac{1}{2} \lip{F}{\fv_{TF} (\pf_T^2 - \pf_F^2)}{} + \frac{1}{2} \lip{F}{\fv_{TF} (\pf_T - \pf_F)^2}{}.
\end{split}
\end{equation}
Inspired by \eqref{eqn:stability_gradient_factor} and \eqref{eqn:stability_flux_inequality}, we set for any cell $T\in\scriptT_h$,
\begin{equation}\label{def:wtilde}
\widetilde{\tf}_{T} = \frac{1}{2} \left( \pf_T^2, \left( \pf_F^2 \right)_{F\in\scriptF_T} \right) \in \dof^{2k}_T,
\end{equation}
and use the conservation of the fluxes (Theorem \ref{thm:conservation_of_the_discrete_fluxes}) with $\widetilde{\tf}_T$
instead of $\tf_T$ to write
\begin{equation} \label{eqn:stability_conservation_term}
\begin{split}
- \frac{1}{2}&\lipd{T}{\RU_T}{\nabla \left(\pf_T^2 \right)} + \frac{1}{2} \sum_{F\in\scriptF_T} \lip{F}{\fv_{TF} (\pf_T^2 - \pf_F^2)}{} \\
&= - \lipd{T}{\RU_T}{\nabla \widetilde{\pf}_T } + \sum_{F\in\scriptF_T} \lip{F}{\fv_{TF} (\widetilde{\pf}_T - \widetilde{\pf}_F)}{} = \dbf_{T,\bkappa^\hs} (\dpdof_{T}^\hs, \widetilde{\tf}_{T}).
\end{split}
\end{equation}
Summing over the cells and using the fact that $\dpdof_{T}^\hs$ solves the discrete pressure equation \eqref{eqn:discrete_pressure_equation}, we deduce that
\begin{equation} \label{eqn:stability_conservation_term2}
\begin{split}
- \frac{1}{2} \sum_{T\in\scriptT_h} &\lipd{T}{\RU_T}{\nabla \left(\pf_T^2 \right)} + \frac{1}{2} \sum_{T\in\scriptT_h} \sum_{F\in\scriptF_T} \lip{F}{\fv_{TF} (\pf_T^2 - \pf_F^2)}{} \\
&= l_h^{p,\hs}(\widetilde{\tf}_h)
= \sum_{T\in\scriptT_h} \frac{1}{2} \lip{T}{(q^+(t^\hs, \cdot) - q^-(t^\hs, \cdot))}{\pf_T^2}.
\end{split}
\end{equation}
Gathering the results from \eqref{eqn:stability_flux_inequality}--\eqref{eqn:stability_conservation_term2} and substituting into \eqref{eqn:advection_stability_terms}, we have
\begin{align}
&\dbf_{h,R, \RU}(\tf_h, \tf_h) \nonumber\\
&= \sum_{T\in\scriptT_h} \left\{ \lip{T}{\frac{2 \Phi}{\Delta t}}{\pf_T^2} + \frac{1}{2} \lip{T}{(q^+(t^\hs, \cdot) - q^-(t^\hs, \cdot))}{\pf_T^2} + \lip{T}{q^-(t^\hs,\cdot) \pf_T^2}{} \right\} \nonumber \\
&+ \sum_{T\in\scriptT_h} \sum_{F \in \scriptF_T} \left\{\frac{1}{2} \lip{F}{\fv_{TF} (\pf_T - \pf_F)^2}{} + \lip{F}{[\fv_{TF}]^-\left(\pf_F - \pf_T\right)^2}{} \right\}.
\label{eqn:stability_after_conservation}
\end{align}
Combining the second and third cell terms of \eqref{eqn:stability_after_conservation}, and using $\frac{1}{2} \fv_{TF} + [\fv_{TF}]^- = \frac{1}{2} |\fv_{TF}|$ and the non-negativity of $q^+$ and $q^-$, we find
\begin{equation*}
\begin{split}
\dbf_{h,R, \RU}(\tf_h, \tf_h) ={}& \sum_{T\in\scriptT_h} \left\{ \lip{T}{\frac{2 \Phi}{\Delta t}}{\pf_T^2} + \frac{1}{2} \lip{T}{(q^+(t^\hs, \cdot) + q^-(t^\hs, \cdot))}{\pf_T^2} \right\} \\
&+ \sum_{T\in\scriptT_h} \sum_{F \in \scriptF_T} \frac{1}{2} \lip{F}{|\fv_{TF}| (\pf_T - \pf_F)^2}{}\\
\geq{}& \lip{\Omega}{\frac{2\Phi}{\Delta t}}{} \pf_h^2 + \sum_{T\in\scriptT_h} \sum_{F \in \scriptF_T} \frac{1}{2} \lip{F}{|\fv_{TF}| (\pf_T - \pf_F)^2}{}.
\end{split}
\end{equation*}
The proof is complete.
\end{proof}
\begin{remark}[Order $2k$ on the pressure, and choice of the Darcy fluxes]\label{rem:2k.flux}
The reason for discretising the pressure equation with an HHO scheme of order $2k$, instead of $k$,
is found in \eqref{eqn:stability_conservation_term}. Obtaining this relation requires
the usage of $\widetilde{\tf}_T$, defined by \eqref{def:wtilde} and belonging to $\dof^{2k}_T$,
into \eqref{eqn:conservation_of_the_discrete_fluxes}.
Equation \eqref{eqn:stability_conservation_term} is an essential component of the
stability proof, and it also justifies our choice of Darcy flux and volumetric velocity \eqref{eqn:discrete_fluxes}
and \eqref{eqn:darcy_velocity_volumetric_reconstruction}.
\end{remark}
We are now ready to prove the existence, uniqueness and stability of the solution to the scheme.
\begin{proof}[Proof of Theorem \ref{thm:stability_of_crank_nicolson}]
Let us first assume that we have a solution $(\dpdof_h,\dc_h)$ to the scheme, and
let us prove the \emph{a priori} estimate \eqref{est:cN}.
By \eqref{eqn:discrete_advection_diffusion_equation}, we have for all $\tf_h \in \dof^k_h$
\begin{equation}
\dbf_{h,\dD, R, \RU}(\dch_h, \tf_h) = l^{c,\hs}_h(\tf_h).
\end{equation}
Select $\tf_h = \dch_h$ as the test function and expand by the definitions of the discrete linear forms to write
\begin{equation*
\dbf_{h,\dD}(\dch_h, \dch_h) + \dbf_{h,R, \RU}(\dch_h, \dch_h) = l^{c,\hs}_h(\dch_h).
\end{equation*}
Using Lemma \ref{lemma:diffusive_bilinear_form_is_positive} (with $\bLambda=\dD$)
and Lemma \ref{lemma:advective_bilinear_form_is_positive},
the definition \eqref{eqn:concentration_rhs_linear_form} of $l^{c,\hs}$ then yields
\begin{equation}\label{est:for_existence}
\begin{split}
\lip{\Omega}{\frac{2\Phi (\pch_h)^2}{\Delta t}}{}
+\sum_{T\in\scriptT_h}\sum_{F\in\scriptF_h} \frac{\beta}{h_F}\|\pc_F^\hs-\pc_T^\hs \|_{L^2(F)}^2\\
\leq \lip{\Omega}{\left(q^+(t^\hs, \cdot) \hat{c}(t^\hs, \cdot) + \frac{2\Phi}{\Delta t}\pc_h^n\right)}{\pch_h}
\end{split}
\end{equation}
where $\beta>0$ is a coercivity constant of $\dD$ ($\beta$ depends on $\Phi$, $d_m$, $d_l$ and $d_t$).
Gathering the time-stepping terms together and dropping the second term in the left-hand side, we then write
\begin{equation} \label{eqn:stability_inequality}
\lip{\Omega}{\frac{2\Phi (\pch_h - \pc^n_h)}{\Delta t}}{\pch_h} \leq \lip{\Omega}{\left(q^+(t^\hs, \cdot) \hat{c}(t^\hs, \cdot)\right)}{\pch_h}.
\end{equation}
Recalling the definition of the half time-stepped concentration \eqref{eqn:half_time_concentration}, we easily deduce
\[
\lip{\Omega}{\frac{2\Phi (\pch_h - \pc^n_h)}{\Delta t}}{\pch_h}
= \lip{\Omega}{\Phi \frac{(\pc^{n+1}_h)^2 - (\pc^n_h)^2}{2\Delta t}}{}.
\]
Hence, using the Cauchy--Schwarz and Young's inequalities in the right-hand side of \eqref{eqn:stability_inequality} yield,
for any $\varepsilon>0$,
\[
\begin{split}
\lip{\Omega}{\Phi \frac{(\pc^{n+1}_h)^2 - (\pc^n_h)^2}{2\Delta t}}{} \leq{}& \| q^+(t^\hs, \cdot) \hat{c}(t^\hs, \cdot) \|_{L^2(\Omega)} \|\pch_h \|_{L^2(\Omega)}\\
\leq{}& \frac{1}{2\varepsilon}\| q^+(t^\hs, \cdot) \hat{c}(t^\hs, \cdot) \|_{L^2(\Omega)}^2
+ \frac{\varepsilon}{2}\|\pch_h \|_{L^2(\Omega)}^2.
\end{split}
\]
Summing over the time steps $n=0,\ldots,N-1$, the sum telescopes in the left-hand side. Using $|\widehat{c}|\le 1$
and, by convexity of the square function,
\[
(\pch_h)^2=\left(\frac{\pc_h^n+\pc_h^{n+1}}{2}\right)^2\le \frac{(\pc_h^n)^2+(\pc_h^{n+1})^2}{2}
\]
we infer
\[
\begin{split}
\lip{\Omega}{\Phi \frac{(\pc^{N}_h)^2 - (\pc^0_h)^2}{2\Delta t}}{}
\le{}& \frac{N}{2\varepsilon}\| q^+\|_{L^\infty(0,t_f;L^2(\Omega))}^2 \\
& + \frac{\varepsilon}{4}\sum_{n=0}^{N-1}\left(\|\pc_h^n \|_{L^2(\Omega)}^2+\|\pc_h^{n+1} \|_{L^2(\Omega)}^2\right)\\
\le{}& \frac{N}{2\varepsilon}\| q^+\|_{L^\infty(0,t_f;L^2(\Omega))}^2
+ \frac{\varepsilon}{2}\sum_{n=0}^{N}\|\pc_h^n \|_{L^2(\Omega)}^2.
\end{split}
\]
Applying the boundedness of $\Phi$ \eqref{eqn:permeability_assumptions},
\[
\int_\Omega{\Phi_* \frac{(\pc^{N}_h)^2}{2\Delta t}} \leq \int_\Omega{\Phi_*^{-1} \frac{(\pc^0_h)^2}{2\Delta t}} + \frac{N}{2\varepsilon} \|q^+\|^2_{L^\infty(0,t_f;L^2(\Omega))} + \frac{\varepsilon}{2} \sum_{n=0}^{N}\int_\Omega (\pc^n_h)^2.
\]
Multiplying both sides by $2\Delta t/\Phi_*$ and recalling that $N \Delta t = t_f$ yields
\begin{equation} \label{eqn:stability_bounds3}
\int_\Omega (\pc^{N}_h)^2 \leq \int_\Omega \frac{(\pc^0_h)^2}{\Phi_*^2} + \frac{t_f}{\varepsilon\Phi_* } \|q^+\|^2_{L^\infty(0,t_f;L^2(\Omega))} + \frac{\varepsilon t_f }{\Phi_*N} \sum_{n=0}^{N} \int_\Omega (\pc^n_h)^2.
\end{equation}
Take $\varepsilon=\frac{\Phi_*}{2t_f}$, so that $\frac{\varepsilon t_f}{\Phi_* N}=\frac{1}{2N}<1$.
Applying the Gronwall inequality of \cite[Lemma 5.1]{HR90} yields
\[
\| \pc^N_h\|_{L^2(\Omega)}^2\le \exp\left(\frac{1}{2N}\sum_{n=0}^N \frac{2N}{2N-1}\right)
\left(\frac{\|\pc^0_h\|_{L^2(\Omega)}^2}{\Phi_*^2} + \frac{t_f}{\varepsilon\Phi_* } \|q^+\|^2_{L^\infty(0,t_f;L^2(\Omega))}\right).
\]
The proof of \eqref{est:cN} is complete since $\frac{t_f}{\varepsilon\Phi_* }=\frac{2t_f^2}{\Phi_*^2}$
and $\frac{1}{2N}\sum_{n=0}^N \frac{2N}{2N-1}=\frac{N+1}{2N-1}\le 2$. The estimate was obtained
for $\pc_h^N$ but the same reasoning shows that it holds for $\pc_h^n$ for all $n=0,\ldots,N$.
\medskip
The existence and uniqueness of $(\dpdof_h,\dc_h)$ follows easily. At each iteration of Algorithm \ref{algo1},
$\dpdof_h^\hs$ is sought as a solution of the linear system \eqref{eqn:discrete_pressure_equation}--\eqref{eqn:discrete_pressure_equation:norm}.
If $l^{p,\hs}_h=0$, plugging $\tf_h=\dpdof_h^\hs$ in \eqref{eqn:discrete_pressure_equation} and
using Lemma \ref{lemma:diffusive_bilinear_form_is_positive} shows that the only possible solution to
this linear system is zero
(by \eqref{eqn:diffusive_bilinear_form_seminorm_inequality}, all cell unknowns must be constant and, working from
neighbour to neighbour, equal to all face unknowns and to all other cell unknowns; then
\eqref{eqn:discrete_pressure_equation:norm} fixes this constant uniform value to zero).
Hence, the matrix corresponding to \eqref{eqn:discrete_pressure_equation}--\eqref{eqn:discrete_pressure_equation:norm} is invertible, which means that this system has a unique solution $\dpdof_h^\hs$ at each time step.
After the pressure is fixed, $\dc_h^\hs$ is sought as a solution to the linear
equation \eqref{eqn:discrete_advection_diffusion_equation}. If $l^{c,\hs}_h=0$,
the right-hand side of \eqref{est:for_existence} vanishes, which shows that all cell and face
degrees of freedom are equal to 0.
Hence, the matrix of \eqref{eqn:discrete_advection_diffusion_equation} has a trivial
kernel, which shows the existence and uniqueness of $\dc_h^\hs$ solution to this equation.
\end{proof}
\section{Appendix: Implementation of the Scheme}\label{appen:implementation}
We present here algorithms for computing the local operators that define the numerical scheme. A fully functional implementation of the scheme in \texttt{C++} along with all of the tests present in Section \ref{sec:tests} can be found at
\centerline{
\href{https://github.com/DanielLiamAnderson/hho-peaceman}{\texttt{https://github.com/DanielLiamAnderson/hho-peaceman}}.}
The code for handling the mesh is a preliminary version of what later became the \texttt{DiSk++} library by Cicuttin et al. \cite{CICUTTIN2017}.
\subsection{A basis for the function spaces}
In order to realise the algorithms for solving the pressure and concentration equations, we first need to express our function spaces and test functions concretely. We recall the space $\Poly^m(K)$ of polynomials of degree $\le m$ over the domain $K$, and decompose its elements in terms of the following basis functions. Denote by $(\bar{x}_K, \bar{y}_K)$ the centre of mass of $K$, and by $h_K$ its diameter. The basis functions for 2D elements $T \in \scriptT_h$ are given by
\begin{equation} \label{eqn:polynomial_basis}
\varphi^T_{r,s} : T \to \R, \qquad \varphi^T_{r,s}(x,y) = \left( \frac{x - \bar{x}_T}{h_T} \right)^{r} \left( \frac{y - \bar{y}_T}{h_T} \right)^{s},
\end{equation}
for $r,s \geq 0$ and $r + s \leq m$. The basis functions for 1D elements $F \in \scriptF_h$ are given by
\begin{equation}
\varphi^F_r : F \to \R, \qquad \varphi^F_r(x,y) = \left( \frac{(x-\bar{x}_F)(x_0 - \bar{x}_F)}{h_F^2} + \frac{(y-\bar{y}_F)(y_0 - \bar{y}_F)}{h_F^2} \right)^r,
\end{equation}
for $0 \leq r \leq m$ where $(x_0,y_0)$ is one of the endpoints of $F$. We then take, for each cell $T \in \scriptT_h$ and face $F \in \scriptF_h$, the set of all basis functions covering the entire mesh, where each function is extended to $\Omega$ by defining $\varphi = 0$ outside its initial domain:
\begin{equation}
\scriptB^m = \left(\bigcup_{T \in \scriptT_h} \{\varphi^T_{r,s}\}_{r+s\leq m} \right) \bigcup \left( \bigcup_{F \in \scriptF_h} \{\varphi^F_{r}\}_{r \leq m} \right).
\end{equation}
The algorithms for solving the pressure and concentration equations then become square linear systems by evaluating the discrete equations \eqref{eqn:discrete_pressure_equation} and \eqref{eqn:discrete_advection_diffusion_equation} at all basis functions $\tf_h \in \scriptB^m$. Since each basis function is only non-zero in a single element of the meshed domain, the resulting linear systems will be sparse. The following notation are referred to throughout the implementation.
\begin{center}
\begin{tabular}{ l l }
$\scriptB^m$ & the set of all basis functions over mesh elements up to degree $m$, \\
$\scriptB^{m}_K$ & the basis functions on the cell or face $K$ of degree up to $m$, \\
$\scriptB^{m,1}_K$ & the basis functions on the cell or face $K$ of degree at least $1$, up to $m$, \\
$\scriptB^{m}_{\underline{T}}$ & the basis functions on the cell $T$ and all adjacent faces of degree up to $m$.
\end{tabular}
\end{center}
\begin{remark}
The space $\scriptB^{m,1}_K$ is useful when considering gradients of high-order basis functions, since gradients of degree zero basis functions are identically zero.
\end{remark}
\subsection{Numerical quadrature}
The assembly of the scheme matrices requires the numerical integration of products of arbitrary order polynomial basis functions. To ensure no loss of accuracy or stability, sufficiently accurate numerical quadrature rules must be used. Most of the integrals that we are required to evaluate consist of the product of two or three degree $m$ polynomials. Based on this observation, for polynomial degrees of freedom of order $m$, we employ numerical quadrature schemes that are exact for polynomials of degree up to $3m$. To integrate cell polynomials, we use the numerical quadrature schemes introduced by Dunavant in \cite{dunavant1985high}. The Dunavant quadrature rules provide exact integrals for fixed degree polynomial functions on triangular domains so we will split each cell of the mesh into triangular sub-elements such that each face of the cell corresponds to one sub-element. Quadrature for edge polynomials is performed using standard Gaussian quadrature rules for one-dimensional domains.
\begin{figure}[H]
\centering
\begin{tikzpicture}
\draw (0,0) -- (1,2);
\draw (1,2) -- (5,2);
\draw (5,2) -- (6,-1);
\draw (6,-1) -- (3,-2);
\draw (3,-2) -- (0,0);
\draw[fill] (2.9,0.2) circle [radius=0.1];
\node [below left] at (2.9,0.2) {$\bar{x}_T$};
\draw [decorate,decoration={brace,amplitude=10pt}]
(1,2.1) -- (5,2.1) node [black,midway,yshift=0.6cm]
{\footnotesize $F$};
\draw (3, 1.2) node {$K_{TF}$};
\draw [dashed] (0,0) -- (2.9,0.2);
\draw [dashed] (1,2) -- (2.9,0.2);
\draw [dashed] (5,2) -- (2.9,0.2);
\draw [dashed] (6,-1) -- (2.9,0.2);
\draw [dashed] (3,-2) -- (2.9,0.2);
\end{tikzpicture}
\caption{A cell split into triangular sub-elements to facilitate numerical quadrature.} \label{fig:split_cell_into_subelements}
\end{figure}
In the algorithms that follow, $K_{TF}$ stands for the triangular sub-element of the cell $T$ adjacent to the edge $F \in \scriptF_T$ as depicted in Figure \ref{fig:split_cell_into_subelements}.
\subsection{Assembly of the local operators for the pressure scheme}
The majority of the integrals in \eqref{eqn:discrete_pressure_equation} will be zero, so the local operators $\dbf_{T,\bLambda}$ can be efficiently assembled for each cell and then combined to form a sparse matrix for the scheme. These local constructions are outlined in the following algorithms. Algorithm \ref{algo:local_reconstruction} demonstrates the computation of local gradient reconstruction operator $\nabla \ro^{m+1}_{T, \bLambda}$ by inverting a high-order mass matrix of the basis gradient functions. The local gradient reconstruction is used in the assembly of the local diffusion operator $\dbf_{T,\bLambda}$ and in the reconstruction of the fluxes from the solution to the pressure equation. Algorithm \ref{algo:local_diffusion_bilinear_form} outlines the assembly of the local diffusion operator. The key step is the computation of the projections from the high-order correction, which is accomplished by computing a mass matrix of mixed high and low-order basis functions. Finally, Algorithm \ref{algo:local_load_vector} presents the implementation of the pressure source term, i.e.\ the right-hand side of the equation.
\subsection{Computation of the numerical fluxes}
We now present a simple algorithm to compute the numerical fluxes $\fv_{TF}$, that does not require us to
compute the whole right-hand side of \eqref{eqn:discrete_fluxes}. Let $T \in \scriptT_h$ and use the conservativity of the fluxes (Theorem \ref{thm:conservation_of_the_discrete_fluxes}) with a cell-absent test function $\tf_{T}=(0,\tf_F) \in \{0\}\times\dof^{2k}_{\partial T}$ to obtain
\begin{equation}\label{bal.eq.algo}
\dbf_{T,\bkappa}(\dpdof_{T}, \tf_{T}) = - \sum_{F\in\scriptF_T} (\fv_{TF}, \pf_F)_{L^2(F)}.
\end{equation}
Select an edge $F \in \scriptF_T$, then write the flux $\fv_{TF}$ in terms of the basis $\scriptB^{2k}_F$ like so
\begin{equation}\label{eqn:computed_fluxes}
\fv_{TF} = \sum_{i=0}^{|\scriptB^{2k}_F| - 1} \lambda^F_i \varphi^F_i.
\end{equation}
The balance equation \eqref{bal.eq.algo} now reads
\begin{equation} \label{eqn:balance_of_basis_functions}
\dbf_{T,\bkappa}(\dpdof_T, \tf_T)
= - \sum_{i=0}^{|\scriptB^{2k}_F| - 1} \lambda_i^F (\varphi^F_i, \pf_F)_{L^2(F)},
\end{equation}
which results in a square linear system for $\{\lambda^F_i\}_i$ in terms of the basis functions ${\scriptB^{2k}_F}$. The construction of the numerical fluxes is shown in Algorithm \ref{algo:flux_reconstruction}.
\subsection{Assembly of the local operators for the concentration scheme}
The computation of the advective derivative is similar to that of the local gradient reconstruction. Algorithm \ref{algo:local_advective_derivative} shows how to efficiently compute $\ad_{T,\RU}^k$ for each cell $T \in \scriptT_h$ by solving a local problem involving the mass matrix of the cell. The local advection-reaction operator is computed as shown in Algorithm \ref{algo:local_advection_reaction_bilinear_form} by assembling together the advective derivatives of the local test functions combined with a mass-reaction matrix that accounts for the value of the reaction terms. Lastly, we present Algorithm \ref{algo:concentration_local_load_vector}, which computes the right-hand side of the concentration equation.
\vfill
\thanks{\textbf{Acknowledgement}: this research was supported by the Australian Government through the Australian Research Council's Discovery Projects funding scheme (pro\-ject number DP170100605).
The authors would also like to thank Daniele Di Pietro and Matteo Cicuttin for giving us access
to the {\sf hho} software platform\footnote{\emph{Agence pour la Protection des Programmes} deposit number IDDN.FR.001.220005.000.S.P.2016.000.10800}, which served as an invaluable starting point for our implementation
and whose development was funded by Agence Nationale de la Recherche project HHOMM ANR-15-CE40-0005.
}
\begin{algorithm}[h]
\caption{Computation of the local gradient reconstruction matrix\label{algo:local_reconstruction}}
\begin{algorithmic}[1]
\State {Set $M_{\varphi_i, \varphi_j}=0$ for all $\varphi_i, \varphi_j \in \scriptB^{m+1,1}_{T}$}
\State {Set $B_{\varphi_i, \varphi_j}=0$ for all $\varphi_i \in \scriptB^{m+1,1}_{T}, \varphi_j \in \scriptB^{m}_{\underline{T}}$}
\For{each edge $F \in F_T$}
\LineComment{Compute volumetric terms corresponding to the sub-element of $F$}
\For{each quadrature node $(x, w_x)$ of $K_{TF}$}
\For{each basis function $\varphi_i \in \scriptB^{m+1,1}_T$}
\For{each basis function $\varphi_j \in \scriptB^{m+1,1}_T$}
\State $M_{\varphi_i, \varphi_j} \leftarrow M_{\varphi_i, \varphi_j} + w_x (\bLambda(x) \nabla \varphi_i(x)) \cdot \nabla \varphi_j(x)$
\EndFor
\For{each basis function $\varphi_j \in \scriptB^{m}_T$}
\State $B_{\varphi_i, \varphi_j} \leftarrow B_{\varphi_i, \varphi_j} + w_x (\bLambda(x) \nabla \varphi_i(x)) \cdot \nabla \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\LineComment{Compute edge terms}
\For{each quadrature node $(x, w_x)$ of $F$}
\For{each basis function $\varphi_i \in \scriptB^{m+1,1}_T$}
\For{each basis function $\varphi_j \in \scriptB^{m}_F$}
\State $B_{\varphi_i, \varphi_j} \leftarrow B_{\varphi_i, \varphi_j} + w_x \nabla \varphi_i(x) \cdot (\bLambda(x) \nv_{TF}) \varphi_j(x)$
\EndFor
\For{each basis function $\varphi_j \in \scriptB^{m}_T$}
\State $B_{\varphi_i, \varphi_j} \leftarrow B_{\varphi_i, \varphi_j} - w_x \nabla \varphi_i(x) \cdot (\bLambda(x) \nv_{TF}) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\EndFor
\State {Set $G = M^{-1} B$} \Comment{Solve for the gradient reconstruction}
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Assembly of the local diffusion matrix\label{algo:local_diffusion_bilinear_form}}
\begin{algorithmic}[1]
\LineComment{Compute mass matrices}
\State {Set $M^{TT}_{\varphi_i,\varphi_j} = 0$ for all $\varphi_i, \varphi_j \in \scriptB^{m+1}_T$}
\State {Set $M^{TF}_{\varphi_i,\varphi_j} = 0$ for all $\varphi_i \in \scriptB^{m}_F,\ \varphi_j \in \scriptB^{m+1}_T$ for all edges $F \in \scriptF_T$}
\State {Set $M^{FF}_{\varphi_i,\varphi_j} = 0$ for all $\varphi_i, \varphi_j \in \scriptB^{m}_F$ for all edges $F \in \scriptF_T$}
\For{each edge $F \in F_T$}
\LineComment{Compute cell-on-cell mass matrix}
\For{each quadrature node $(x, w_x)$ of $K_{TF}$}
\For{each basis function $\varphi_i \in \scriptB^{m+1}_T$}
\For{each basis function $\varphi_j \in \scriptB^{m+1}_T$}
\State $M^{TT}_{\varphi_i, \varphi_j} \leftarrow M^{TT}_{\varphi_i, \varphi_j} + w_x \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\LineComment{Compute cell-on-edge and edge-on-edge mass matrices}
\For{each quadrature node $(x, w_x)$ of $F$}
\For{each basis function $\varphi_i \in \scriptB^{m}_F$}
\For{each basis function $\varphi_j \in \scriptB^{m+1}_T$}
\State $M^{TF}_{\varphi_i, \varphi_j} \leftarrow M^{TF}_{\varphi_i, \varphi_j} + w_x \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\For{each basis function $\varphi_i \in \scriptB^{m}_F$}
\For{each basis function $\varphi_j \in \scriptB^{m}_F$}
\State $M^{FF}_{\varphi_i, \varphi_j} \leftarrow M^{FF}_{\varphi_i, \varphi_j} + w_x \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\EndFor
\LineComment{Compute the volumetric term}
\State {Compute the gradient reconstruction $G$ (Algorithm \ref{algo:local_reconstruction})}
\State {Set $A = B^{tr} G$} \Comment{$B^{tr}$ is the transpose of $B$ from Algorithm \ref{algo:local_reconstruction}}
\LineComment{Compute the local reconstruction cell projection matrix}
\State {Set $M^{TT, m}_{\varphi_i, \varphi_j} = M^{TT}_{\varphi_i, \varphi_j}$ for all $\varphi_i, \varphi_j \in \scriptB^{m}_T$}
\State {Set $M^{TT, m + 1}_{\varphi_i, \varphi_j} = M^{TT}_{\varphi_i, \varphi_j}$ for all $\varphi_i \in \scriptB^{m}_T, \varphi_j \in \scriptB^{m+1,1}_T$}
\State {Solve $M^{TT, m} P_T = M^{TT, m+1}$ for $P_T$}
\LineComment{Compute the edge terms}
\For{each edge $F \in F_T$}
\LineComment{Compute the edge projection matrix}
\State {Set $P_F = (M^{FF})^{-1}$}
\State {Set $M^{TF,m}_{\varphi_i, \varphi_j} = M^{TF}_{\varphi_i, \varphi_j}$ for all $\varphi_i \in \scriptB^{m}_F, \varphi_j \in \scriptB^{m}_T$}
\State {Set $M^{TF,m+1}_{\varphi_i, \varphi_j} = M^{TF}_{\varphi_i, \varphi_j}$ for all $\varphi_i \in \scriptB^{m}_F, \varphi_j \in \scriptB^{m+1,1}_T$}
\LineComment{Compute projections}
\State {Set $B_F = P_F M^{TF,m+1} G - I_F$} \Comment{$I_F=$ identity matrix on the edge terms}
\State {Set $B_T = P_F M^{TF,m} (I_T - P_T)$} \Comment{$I_T=$ identity matrix on the cell terms}
\State {Set $B_{RF} = B_F + B_T$}
\LineComment{Compute local stabilisation terms}
\State $A \leftarrow A + \frac{\bLambda_{TF}}{h_F} B_{RF}^{tr} M^{FF} B_{RF}$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Assembly of the pressure source vector\label{algo:local_load_vector}}
\begin{algorithmic}[1]
\State Set $b_{\varphi_i} = 0$ for all $\varphi_i \in \scriptB^{2k}_{\underline{T}}$
\For{each edge $F \in F_T$}
\For{each quadrature node $(x, w_x)$ of $K_{TF}$}
\For{each basis function $\varphi_i \in \scriptB^{2k}_T$}
\State $b_{\varphi_i} \leftarrow b_{\varphi_i} + w_x \varphi_i(x) (q^+(t^\hs, x) - q^-(t^\hs, x))$
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Computation of the local fluxes\label{algo:flux_reconstruction}}
\begin{algorithmic}[1]
\State {Compute the local diffusion operator matrix $A$ with $m = 2k$ (Algorithm \ref{algo:local_diffusion_bilinear_form})}
\For{each edge $F \in \scriptF_T$}
\State Set $\alpha_{\varphi_i} = 0$ for all $\varphi_i \in \scriptB^{2k}_F$
\For{each basis function $\varphi_i \in \scriptB^{2k}_F$}
\State $\alpha_{\varphi_i} \leftarrow - (\dpdof_T \cdot A_{:, \varphi_i})$ \Comment{$A_{:,\varphi_i}$ is the column of $A$ corresponding to $\varphi_i$}
\EndFor
\LineComment{Build the Gram matrix}
\State {Set $G_{\varphi_i, \varphi_j} = 0$ for all $\varphi_i, \varphi_j \in \scriptB^{2k}_F$}
\For{each quadrature node $(x, w_x)$ on $F$}
\For{each basis function $\varphi_i \in \scriptB^{2k}_F$}
\For{each basis function $\varphi_j \in \scriptB^{2k}_F$}
\State $G_{\varphi_i, \varphi_j} \leftarrow G_{\varphi_i, \varphi_j} + w_x \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\State {Set $\lambda^F = G^{-1} \alpha$}
\EndFor
\State {Compute $\fv_{TF}$ in terms of $\lambda^F$ as in \eqref{eqn:computed_fluxes}}
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Assembly of the local advective-reactive derivative matrix\label{algo:local_advective_derivative}}
\begin{algorithmic}[1]
\State {Set $M_{\varphi_i, \varphi_j} = 0$ for all $\varphi_i, \varphi_j \in \scriptB^k_T$}
\State {Set $B_{\varphi_i, \varphi_j} = 0$ for all $\varphi_i \in \scriptB^k_T, \varphi_j \in \scriptB^k_{\underline{T}}$}
\For{each edge $F \in \scriptF_T$}
\LineComment{Compute the volumetric terms}
\For{each quadrature node $(x, w_x)$ in $K_{TF}$}
\For{each basis function $\varphi_i \in \scriptB^k_T$}
\For{each basis function $\varphi_j \in \scriptB^k_T$}
\State $M_{\varphi_i, \varphi_j} \leftarrow M_{\varphi_i, \varphi_j} + w_x \varphi_i(x) \varphi_j(x)$
\State $B_{\varphi_i, \varphi_j} \leftarrow B_{\varphi_i, \varphi_j} + w_x (\fv_{TF}(x) \cdot \nabla \varphi_i(x)) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\LineComment{Compute the edge terms}
\For{each quadrature node $(x, w_x)$ on $F$}
\For{each basis function $\varphi_i \in \scriptB^k_T$}
\For{each basis function $\varphi_j \in \scriptB^k_F$}
\State $B_{\varphi_i, \varphi_j} \leftarrow B_{\varphi_i, \varphi_j} + w_x (\fv_{TF}(x) \cdot \nv_{TF}) \varphi_i(x) \varphi_j(x)$
\EndFor
\For{each basis function $\varphi_j \in \scriptB^k_T$}
\State $B_{\varphi_i, \varphi_j} \leftarrow B_{\varphi_i, \varphi_j} - w_x (\fv_{TF}(x) \cdot \nv_{TF}) \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\EndFor
\LineComment{Solve for the advective-reaction derivative}
\State {Set $G = M^{-1}B$}
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Assembly of the local advection-reaction matrix\label{algo:local_advection_reaction_bilinear_form}}
\begin{algorithmic}[1]
\State {Set $A_{\varphi_i,\varphi_j} = 0$ for all $\varphi_i,\varphi_j \in \scriptB^k_{\underline{T}}$}
\State {Compute the advective-reactive derivative $G$ (Algorithm \ref{algo:local_advective_derivative})}
\For{each edge $F \in \scriptF_T$}
\LineComment{Compute the volumetric terms}
\For{each quadrature node $(x, w_x)$ in $K_{TF}$}
\For{each basis function $\varphi_i \in \scriptB^k_T$}
\For{each basis function $\varphi_j \in \scriptB^k_T$}
\State $A_{\varphi_i, \varphi_j} \leftarrow A_{\varphi_i, \varphi_j} + w_x \varphi_i(x) \varphi_j(x) \mu(x)$
\EndFor
\EndFor
\For{each basis function $\varphi_i \in \scriptB^k_{\underline{T}}$}
\LineComment{Compute $\ad_{T,\RU}^k \tf_{\underline{T}}$}
\State {Set $v = 0$}
\For{each basis function $\varphi_j \in \scriptB^k_T$}
\State $v \leftarrow v + \varphi_j(x) G_{\varphi_j, \varphi_i}$
\EndFor
\For{each basis function $\varphi_j \in \scriptB^k_T$}
\State $A_{\varphi_i, \varphi_j} \leftarrow A_{\varphi_i, \varphi_j} - w_x \varphi_j(x) v$
\EndFor
\EndFor
\EndFor
\LineComment{Compute the edge terms}
\For{each quadrature node $(x, w_x)$ on $F$}
\For{each basis function $\varphi_i \in \scriptB^k_T$}
\For{each basis function $\varphi_j \in \scriptB^k_F$}
\State $A_{\varphi_i, \varphi_j} \leftarrow A_{\varphi_i, \varphi_j} - w_x [\fv_{TF}(x) \cdot \nv_{TF}]^- \varphi_i(x) \varphi_j(x)$
\State $A_{\varphi_j, \varphi_i} \leftarrow A_{\varphi_j, \varphi_i} - w_x [\fv_{TF}(x) \cdot \nv_{TF}]^- \varphi_i(x) \varphi_j(x)$
\EndFor
\For{each basis function $\varphi_j \in \scriptB^k_T$}
\State $A_{\varphi_i, \varphi_j} \leftarrow A_{\varphi_i, \varphi_j} + w_x [\fv_{TF}(x) \cdot \nv_{TF}]^- \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\For{each basis function $\varphi_i \in \scriptB^k_F$}
\For{each basis function $\varphi_j \in \scriptB^k_F$}
\State $A_{\varphi_i, \varphi_j} \leftarrow A_{\varphi_i, \varphi_j} + w_x [\fv_{TF}(x) \cdot \nv_{TF}]^- \varphi_i(x) \varphi_j(x)$
\EndFor
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Assembly of the concentration source vector\label{algo:concentration_local_load_vector}}
\begin{algorithmic}[1]
\State Set $\bb_{\varphi_i} = 0$ for all $\varphi_i \in \scriptB^{k}_{\underline{T}}$
\For{each edge $F \in F_T$}
\For{each quadrature node $(x, w_x)$ of $K_{TF}$}
\For{each basis function $\varphi_i \in \scriptB^{k}_T$}
\State $b_{\varphi_i} \leftarrow b_{\varphi_i} + w_x \varphi_i(x) (q^+(t^\hs, x)\hat{c}(t^\hs, x) + \frac{2\Phi}{\Delta t} \pc^n_h(x))$
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\clearpage
\bibliographystyle{abbrv}
|
Subsets and Splits